Datasets:

Modalities:
Text
Formats:
json
ArXiv:
Libraries:
Datasets
pandas
License:
icaswell commited on
Commit
c950fd0
·
verified ·
1 Parent(s): abb012f

Add note on data schemata

Browse files
Files changed (1) hide show
  1. README.md +98 -7
README.md CHANGED
@@ -2632,16 +2632,106 @@ translations into 221 Low-Resource Languages, for the purpose of training
2632
  translation models, and otherwise increasing the representations of said
2633
  languages in NLP and technology.
2634
 
2635
- Please read the [SMOL Paper](...) and the [GATITOS Paper](https://arxiv.org/abs/2303.15265) for a much more thorough description!
 
 
2636
 
2637
  There are three resources in this directory:
2638
 
2639
- * **GATITOS:** token-level translations into 172 languages
2640
- * **SmolSent:** sentence-level translations into 81 languages
2641
  * **SmolDoc:** document-level translations into 100 languages
 
 
2642
  * **SmolDoc-factuality-annotations:** factuality annotations and rationales
2643
  for 661 documents from `SmolDoc`
2644
-
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2645
  ## Notes on the GATITOS multilingual Lexicon
2646
 
2647
  The GATITOS (Google's Additional Translations Into Tail-languages: Often Short)
@@ -2649,11 +2739,12 @@ dataset is a high-quality, multi-way parallel dataset of tokens and short
2649
  phrases, intended for training and improving machine translation models.
2650
  Experiments on this dataset and Panlex focusing on unsupervised translation in
2651
  a 208-language model can be found in
2652
- [BiLex Rx: Lexical Data Augmentation for Massively Multilingual Machine Translation](https://arxiv.org/abs/2303.15265).
2653
 
2654
  This dataset consists in 4,000 English segments (4,500 tokens) that have been
2655
- translated into each of 173 languages, 170 of which are low-resource, and three
2656
- of which are mid-high resource (es, fr, hi). All translations were made
 
2657
  directly from English, with the exception of Aymara, which was translated from
2658
  the Spanish.
2659
 
 
2632
  translation models, and otherwise increasing the representations of said
2633
  languages in NLP and technology.
2634
 
2635
+ Please read the [SMOL Paper](https://arxiv.org/abs/2502.12301) and the
2636
+ [GATITOS Paper](https://arxiv.org/abs/2303.15265) for a much more
2637
+ thorough description!
2638
 
2639
  There are three resources in this directory:
2640
 
 
 
2641
  * **SmolDoc:** document-level translations into 100 languages
2642
+ * **SmolSent:** sentence-level translations into 81 languages
2643
+ * **GATITOS:** token-level translations into 172 languages
2644
  * **SmolDoc-factuality-annotations:** factuality annotations and rationales
2645
  for 661 documents from `SmolDoc`
2646
+
2647
+
2648
+ ## Data schemata
2649
+
2650
+ The schemata are pretty straightforward. Source and Target languge are
2651
+ provided in `sl` and `tl` fields. The `is_src_orig` field has a value
2652
+ of `true` if the source text was the original text, and the target field
2653
+ was translated, and avalue of `false` if the data is back-translated.
2654
+
2655
+ ### SmolDoc
2656
+
2657
+ SmolDoc provides sentence-split and aligned translations through the `srcs` and `trgs` fields. These will always have the same length.
2658
+
2659
+ ```
2660
+ {
2661
+ 'id': 'topic_587__weyiwiniwaaotiwenwy',
2662
+ 'sl': 'en',
2663
+ 'tl': 'pcm',
2664
+ 'is_src_orig': True,
2665
+ 'factuality': 'ok', # this is a story so there is no factual claim that could be wrong
2666
+ 'srcs': ['"What the hell are you doing, you idiot?!"',
2667
+ '"Excuse me?"',
2668
+ '"You cut me off! You almost made me crash!"',
2669
+ '"I\'m sorry, I didn\'t mean to. I was just trying to get around that slow-moving truck."',
2670
+ '"Well, you could have at least used your turn signal!"',
2671
+ '"I did use my turn signal!"',
2672
+ '"No, you didn\'t! You just pulled right out in front of me!"',
2673
+ '"I\'m telling you, I used my turn signal!"',
2674
+ '"Whatever. You\'re still a terrible driver."',
2675
+ '"And you\'re a jerk!"',
2676
+ '"At least I know how to drive!"',
2677
+ '"Oh, yeah? Well, I\'m a better writer than you are!"',
2678
+ '"That\'s debatable."',
2679
+ '"It\'s not debatable! I\'m Ernest Hemingway!"',
2680
+ '"Who?"',
2681
+ '"Ernest Hemingway! The greatest writer of all time!"',
2682
+ '"Never heard of him."',
2683
+ '"Well, you\'ve heard of me now!"',
2684
+ '"Yeah, I heard of you."'],
2685
+ 'trgs': ['"Wetin di hell dey do, yu idiot?!"',
2686
+ '"Ekskuse mi?"',
2687
+ '"Yu komot mi! Yu almost make mi krash!"',
2688
+ '"I dey sorry, I nor wont do am. I just dey try get around dat truk wey slow."',
2689
+ '"Well, yu for don yus yor turn sign!"',
2690
+ '"I yus mai turn sign!"',
2691
+ '"No, yu nor turn am! Yu just turn rite in front of mi!"',
2692
+ '"I dey tell yu, I yus mai turn sign!"',
2693
+ '"Wateva. Yu still bi one tribol driva."',
2694
+ '"And yu bi jerk!"',
2695
+ '"At least I sabi hau to drive!"',
2696
+ '"Oh, yeah? Well, I bi ogbonge writa pass yu!"',
2697
+ '"Wi fit dibate dat."',
2698
+ '"nortin to dibate! I bi Ernest Hemingway!"',
2699
+ '"Who?"',
2700
+ '"Ernest Hemingway! De writa of all taim wey grate pass!"',
2701
+ '"Neva hear am."',
2702
+ '"Well, yu don hear mi nau!"',
2703
+ '"Na so, I don hear yu."']
2704
+ }
2705
+ ```
2706
+
2707
+ ### SmolSent
2708
+
2709
+ SmolSent's schema differs from SmolDoc's only in that `src` and `trg` are single strings, not lists:
2710
+
2711
+ ```
2712
+ {'id': 381,
2713
+ 'sl': 'en',
2714
+ 'tl': 'ber',
2715
+ 'is_src_orig': True,
2716
+ 'src': 'Rih, a deaf former soldier, plots rebellion while married to a queer, teenage god.',
2717
+ 'trg': 'ⵔⵉⵀ, ⴷ ⴰⴷⴻⴼⵔⵉⵔ ⴰⵇⴱⵓⵔ ⴰⴻⵎⴻⵙⵍⵉ, ⵢⴻⵜⵜⵀⴻⴳⴳⵉ ⵜⴰⴴⴻⵡⵡⴰⵜ-ⵉⵙ, ⴴⴰⵙ ⴰⴽⴽⴻⵏ ⵢⴻⵣⵡⴻⴵ ⴷ ⵢⵉⵡⴻⵏ ⵢⵉⵍⵓ ⵉⵍⴻⵎⵥⵉ ⵉⵁⴻⵎⵎⵍⴻⵏ ⴰⵔⵔⴰⵛ.'
2718
+ }
2719
+ ```
2720
+
2721
+ ### GATITOS
2722
+
2723
+ GATITOS has a one-to-many schema, with each source mapping to one or more targets:
2724
+
2725
+ ```
2726
+ {
2727
+ 'sl': 'en',
2728
+ 'tl': 'aa',
2729
+ 'is_source_orig': True,
2730
+ 'src': 'how are you',
2731
+ 'trgs': ['mannah taniih?', 'anninnaay?']
2732
+ }
2733
+ ```
2734
+
2735
  ## Notes on the GATITOS multilingual Lexicon
2736
 
2737
  The GATITOS (Google's Additional Translations Into Tail-languages: Often Short)
 
2739
  phrases, intended for training and improving machine translation models.
2740
  Experiments on this dataset and Panlex focusing on unsupervised translation in
2741
  a 208-language model can be found in
2742
+ the [GATITOS paper](https://arxiv.org/abs/2303.15265).
2743
 
2744
  This dataset consists in 4,000 English segments (4,500 tokens) that have been
2745
+ translated into each of 173 languages, and as a 4001st token, the endonym.
2746
+ Of these, 170 are low-resource, and three
2747
+ are mid-high resource (es, fr, hi). All translations were made
2748
  directly from English, with the exception of Aymara, which was translated from
2749
  the Spanish.
2750