ArneBinder
commited on
Commit
·
277dc70
1
Parent(s):
fd681a0
Upload 3 files
Browse files- README.md +224 -0
- abstrct.py +38 -0
- requirements.txt +1 -0
README.md
ADDED
@@ -0,0 +1,224 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# PIE Dataset Card for "abstrct"
|
2 |
+
|
3 |
+
This is a [PyTorch-IE](https://github.com/ChristophAlt/pytorch-ie) wrapper for the AbstRCT dataset ([paper](https://ebooks.iospress.nl/publication/55129) and [data repository](https://gitlab.com/tomaye/abstrct)). Since the AbstRCT dataset is published in the [BRAT standoff format](https://brat.nlplab.org/standoff.html), this dataset builder is based on the [PyTorch-IE brat dataset loading script](https://huggingface.co/datasets/pie/brat).
|
4 |
+
|
5 |
+
Therefore, the `abstrct` dataset as described here follows the data structure from the [PIE brat dataset card](https://huggingface.co/datasets/pie/brat).
|
6 |
+
|
7 |
+
### Dataset Summary
|
8 |
+
|
9 |
+
A novel corpus of healthcare texts (i.e., RCT abstracts on various diseases) from the MEDLINE database, which
|
10 |
+
are annotated with argumentative components (i.e., `MajorClaim`, `Claim`, and `Premise`) and relations (i.e., `Support`, `Attack`, and `Partial-attack`),
|
11 |
+
in order to support clinicians' daily tasks in information finding and evidence-based reasoning for decision making.
|
12 |
+
|
13 |
+
### Supported Tasks and Leaderboards
|
14 |
+
|
15 |
+
- **Tasks**: Argumentation Mining, Component Identification, Boundary Detection, Relation Identification, Link Prediction
|
16 |
+
- **Leaderboard:** [More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
|
17 |
+
|
18 |
+
### Languages
|
19 |
+
|
20 |
+
The language in the dataset is English (in the medical/healthcare domain).
|
21 |
+
|
22 |
+
### Dataset Variants
|
23 |
+
|
24 |
+
The `abstrct` dataset comes in a single version (`default`) with `BratDocumentWithMergedSpans` as document type. Note,
|
25 |
+
that this in contrast to the base `brat` dataset, where the document type for the `default` variant is `BratDocument`.
|
26 |
+
The reason is that the AbstRCT dataset has already been published with only single-fragment spans.
|
27 |
+
Without any need to merge fragments, the document type `BratDocumentWithMergedSpans` is easier to handle for most of the task modules.
|
28 |
+
|
29 |
+
### Data Schema
|
30 |
+
|
31 |
+
See [PIE-Brat Data Schema](https://huggingface.co/datasets/pie/brat#data-schema).
|
32 |
+
|
33 |
+
### Usage
|
34 |
+
|
35 |
+
```python
|
36 |
+
from pie_datasets import load_dataset, builders
|
37 |
+
|
38 |
+
# load default version
|
39 |
+
datasets = load_dataset("pie/abstrct")
|
40 |
+
doc = datasets["neoplasm_train"][0]
|
41 |
+
assert isinstance(doc, builders.brat.BratDocumentWithMergedSpans)
|
42 |
+
```
|
43 |
+
|
44 |
+
### Document Converters
|
45 |
+
|
46 |
+
The dataset provides document converters for the following target document types:
|
47 |
+
|
48 |
+
- `pytorch_ie.documents.TextDocumentWithLabeledSpansAndBinaryRelations`
|
49 |
+
- `LabeledSpans`, converted from `BratDocumentWithMergedSpans`'s `spans`
|
50 |
+
- labels: `MajorClaim`, `Claim`, `Premise`
|
51 |
+
- `BinraryRelations`, converted from `BratDocumentWithMergedSpans`'s `relations`
|
52 |
+
- labels: `Support`, `Partial-Attack`, `Attack`
|
53 |
+
|
54 |
+
See [here](https://github.com/ChristophAlt/pytorch-ie/blob/main/src/pytorch_ie/documents.py) for the document type
|
55 |
+
definitions.
|
56 |
+
|
57 |
+
### Data Splits
|
58 |
+
|
59 |
+
| Diseease-based Split | `neoplasm` | `glaucoma` | `mixed` |
|
60 |
+
| --------------------------------------------------------- | ----------------------: | -------------------: | -------------------: |
|
61 |
+
| No.of document <br/>- `_train`<br/>- `_dev`<br/>- `_test` | <br/>350<br/>50<br/>100 | <br/> <br/> <br/>100 | <br/> <br/> <br/>100 |
|
62 |
+
|
63 |
+
**Important Note**:
|
64 |
+
|
65 |
+
- `mixed_test` contains 20 abstracts on the following diseases: glaucoma, neoplasm, diabetes, hypertension, hepatitis.
|
66 |
+
- 31 out of 40 abstracts in `mixed_test` overlap with abstracts in `neoplasm_test` and `glaucoma_test`.
|
67 |
+
|
68 |
+
### Label Descriptions
|
69 |
+
|
70 |
+
In this section, we describe labels according to [Mayer et al. (2020)](https://ebooks.iospress.nl/publication/55129), as well as our label counts on 669 abstracts.
|
71 |
+
|
72 |
+
Unfortunately, the number we report does not correspond to what Mayer et al. reported in their paper (see Table 1, p. 2109).
|
73 |
+
Morio et al. ([2022](https://aclanthology.org/2022.tacl-1.37.pdf); p. 642, Table 1), who utilized this corpus for their AM tasks, also reported another number, claiming there were double annotation errors in the original statistic collection (see [reference](https://github.com/hitachi-nlp/graph_parser/blob/main/examples/multitask_am/README.md#qas)).
|
74 |
+
|
75 |
+
#### Components
|
76 |
+
|
77 |
+
| Components | Count | Percentage |
|
78 |
+
| ------------ | ----: | ---------: |
|
79 |
+
| `MajorClaim` | 129 | 3 % |
|
80 |
+
| `Claim` | 1282 | 30.2 % |
|
81 |
+
| `Premise` | 2842 | 66.8 % |
|
82 |
+
|
83 |
+
- `MajorClaim` are more general/concluding `claim`'s, which is supported by more specific claims
|
84 |
+
- `Claim` is a concluding statement made by the author about the outcome of the study. Claims only points to other claims.
|
85 |
+
- `Premise` (a.k.a. evidence) is an observation or measurement in the study, which supports or attacks another argument component, usually a `claim`. They are observed facts, and therefore credible without further justifications, as this is the ground truth the argumentation is based on.
|
86 |
+
|
87 |
+
(Mayer et al. 2020, p.2110)
|
88 |
+
|
89 |
+
#### Relations
|
90 |
+
|
91 |
+
| Relations | Count | Percentage |
|
92 |
+
| ------------------------ | ----: | ---------: |
|
93 |
+
| support: `Support` | 2289 | 87 % |
|
94 |
+
| attack: `Partial-Attack` | 275 | 10.4 % |
|
95 |
+
| attack: `Attack` | 69 | 2.6 % |
|
96 |
+
|
97 |
+
- `Support`: All statements or observations justifying the proposition of the target component
|
98 |
+
- `Partial-Attack`: when the source component is not in full contradiction, but weakening the target component by constraining its proposition. Usually occur between two claims
|
99 |
+
- `Attack`: A component is attacking another one, if it is
|
100 |
+
- i) contradicting the proposition of the target component, or
|
101 |
+
- ii) undercutting its implicit assumption of significance constraints
|
102 |
+
- `Premise` can only be connected to either `Claim` or another `Premise`
|
103 |
+
- `Claim`'s can only point to other `Claim`'s
|
104 |
+
- There might be more than one **outgoing** and/or **incoming relation** . In rare case, there is no relation to another component at all.
|
105 |
+
|
106 |
+
(Mayer et al. 2020, p.2110)
|
107 |
+
|
108 |
+
## Dataset Creation
|
109 |
+
|
110 |
+
### Curation Rationale
|
111 |
+
|
112 |
+
"\[D\]espite its natural employment in healthcare applications, only few approaches have applied AM methods to this kind
|
113 |
+
of text, and their contribution is limited to the detection
|
114 |
+
of argument components, disregarding the more complex phase of
|
115 |
+
predicting the relations among them. In addition, no huge annotated
|
116 |
+
dataset for AM is available for the healthcare domain (p. 2108)...to support clinicians in decision making or in (semi)-automatically
|
117 |
+
filling evidence tables for systematic reviews in evidence-based medicine. (p. 2114)"
|
118 |
+
|
119 |
+
### Source Data
|
120 |
+
|
121 |
+
[MEDLINE database](https://www.nlm.nih.gov/medline/medline_overview.html)
|
122 |
+
|
123 |
+
#### Initial Data Collection and Normalization
|
124 |
+
|
125 |
+
Extended from the previous dataset in [Mayer et al. 2018](https://webusers.i3s.unice.fr/~riveill/IADB/publications/2018-COMMA.pdf), 500 medical abstract from randomized controlled trials (RCTs) were retrieved directly from [PubMed](https://www.ncbi.nlm.nih.gov/pubmed/) by searching for titles or abstracts containing the disease name.
|
126 |
+
|
127 |
+
(See the definition of RCT in the authors' [guideline](https://gitlab.com/tomaye/abstrct/-/blob/master/AbstRCT_corpus/AnnotationGuidelines.pdf) (Section 1.2) and [US National Library of Medicine](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6235704/))
|
128 |
+
|
129 |
+
#### Who are the source language producers?
|
130 |
+
|
131 |
+
\[More Information Needed\]
|
132 |
+
|
133 |
+
### Annotations
|
134 |
+
|
135 |
+
#### Annotation process
|
136 |
+
|
137 |
+
"An expert in the medical domain (a pharmacist) validated the annotation
|
138 |
+
guidelines before starting the annotation process." (p. 2110)
|
139 |
+
|
140 |
+
"Annotation was started after a training phase, where amongst others the component boundaries were topic of discussion. Gold labels
|
141 |
+
were set after a reconciliation phase, during which the annotators
|
142 |
+
tried to reach an agreement. While the number of annotators vary for
|
143 |
+
the two annotation phases (component and relation annotation).
|
144 |
+
|
145 |
+
On the annotation of argument components, "IAA among the three annotators has been calculated
|
146 |
+
on 30 abstracts, resulting in a Fleiss’ kappa of 0.72 for argumentative
|
147 |
+
components and 0.68 for the more fine-grained distinction between
|
148 |
+
claims and evidence." (p. 2109)
|
149 |
+
|
150 |
+
On the annotation of argumentative relation, "IAA has been calculated on 30 abstracts annotated in parallel by three annotators,
|
151 |
+
resulting in a Fleiss’ kappa of
|
152 |
+
0.62. The annotation of the remaining abstracts was carried out by
|
153 |
+
one of the above mentioned annotators." (p. 2110)
|
154 |
+
|
155 |
+
See the [Annotation Guideline](https://gitlab.com/tomaye/abstrct/-/blob/master/AbstRCT_corpus/AnnotationGuidelines.pdf?ref_type=heads) for more information on definitions and annotated samples.
|
156 |
+
|
157 |
+
#### Who are the annotators?
|
158 |
+
|
159 |
+
Two annotators with background in computational linguistics. No information was given on the third annotator.
|
160 |
+
|
161 |
+
### Personal and Sensitive Information
|
162 |
+
|
163 |
+
\[More Information Needed\]
|
164 |
+
|
165 |
+
## Considerations for Using the Data
|
166 |
+
|
167 |
+
### Social Impact of Dataset
|
168 |
+
|
169 |
+
"These \[*intelligent*\] systems apply to clinical trials,
|
170 |
+
clinical guidelines, and electronic health records, and their solutions range from the automated detection of PICO elements
|
171 |
+
in health records to evidence-based reasoning for decision making. These applications highlight the need of clinicians to be supplied with frameworks able to extract, from the huge
|
172 |
+
quantity of data available for the different diseases and treatments,
|
173 |
+
the exact information they necessitate and to present this information in a structured way, easy to be (possibly semi-automatically)
|
174 |
+
analyzed...Given its aptness to automatically detect in text those
|
175 |
+
argumentative structures that are at the basis of evidence-based reasoning applications, AM represents a potential valuable contribution
|
176 |
+
in the healthcare domain." (p. 2108)
|
177 |
+
|
178 |
+
"We expect that our work will have a large impact for clinicians as it
|
179 |
+
is a crucial step towards AI supported clinical deliberation at a large
|
180 |
+
scale." (p. 2114)
|
181 |
+
|
182 |
+
### Discussion of Biases
|
183 |
+
|
184 |
+
\[More Information Needed\]
|
185 |
+
|
186 |
+
### Other Known Limitations
|
187 |
+
|
188 |
+
\[More Information Needed\]
|
189 |
+
|
190 |
+
## Additional Information
|
191 |
+
|
192 |
+
### Dataset Curators
|
193 |
+
|
194 |
+
\[More Information Needed\]
|
195 |
+
|
196 |
+
### Licensing Information
|
197 |
+
|
198 |
+
- **License**: the AbstRCT dataset is released under a [Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License](https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode)
|
199 |
+
- **Funding**: This work is partly funded by the French government labelled PIA
|
200 |
+
program under its IDEX UCA JEDI project (ANR-15-IDEX-0001).
|
201 |
+
This work has been supported by the French government, through the
|
202 |
+
3IA Cote d’Azur Investments in the Future project managed by the
|
203 |
+
National Research Agency (ANR) with the reference number ANR19-P3IA-0002
|
204 |
+
|
205 |
+
### Citation Information
|
206 |
+
|
207 |
+
```
|
208 |
+
@inproceedings{mayer2020ecai,
|
209 |
+
author = {Tobias Mayer and
|
210 |
+
Elena Cabrio and
|
211 |
+
Serena Villata},
|
212 |
+
title = {Transformer-Based Argument Mining for Healthcare Applications},
|
213 |
+
booktitle = {{ECAI} 2020 - 24th European Conference on Artificial Intelligence},
|
214 |
+
series = {Frontiers in Artificial Intelligence and Applications},
|
215 |
+
volume = {325},
|
216 |
+
pages = {2108--2115},
|
217 |
+
publisher = {{IOS} Press},
|
218 |
+
year = {2020},
|
219 |
+
}
|
220 |
+
```
|
221 |
+
|
222 |
+
### Contributions
|
223 |
+
|
224 |
+
Thanks to [@ArneBinder](https://github.com/ArneBinder) and [@idalr](https://github.com/idalr) for adding this dataset.
|
abstrct.py
ADDED
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from pytorch_ie.documents import TextDocumentWithLabeledSpansAndBinaryRelations
|
2 |
+
|
3 |
+
from pie_datasets.builders import BratBuilder, BratConfig
|
4 |
+
from pie_datasets.builders.brat import BratDocumentWithMergedSpans
|
5 |
+
|
6 |
+
URL = "https://gitlab.com/tomaye/abstrct/-/archive/master/abstrct-master.zip"
|
7 |
+
SPLIT_PATHS = {
|
8 |
+
"neoplasm_train": "abstrct-master/AbstRCT_corpus/data/train/neoplasm_train",
|
9 |
+
"neoplasm_dev": "abstrct-master/AbstRCT_corpus/data/dev/neoplasm_dev",
|
10 |
+
"neoplasm_test": "abstrct-master/AbstRCT_corpus/data/test/neoplasm_test",
|
11 |
+
"glaucoma_test": "abstrct-master/AbstRCT_corpus/data/test/glaucoma_test",
|
12 |
+
"mixed_test": "abstrct-master/AbstRCT_corpus/data/test/mixed_test",
|
13 |
+
}
|
14 |
+
|
15 |
+
|
16 |
+
class AbstRCT(BratBuilder):
|
17 |
+
BASE_DATASET_PATH = "DFKI-SLT/brat"
|
18 |
+
BASE_DATASET_REVISION = "bb8c37d84ddf2da1e691d226c55fef48fd8149b5"
|
19 |
+
|
20 |
+
BUILDER_CONFIGS = [
|
21 |
+
BratConfig(name=BratBuilder.DEFAULT_CONFIG_NAME, merge_fragmented_spans=True),
|
22 |
+
]
|
23 |
+
DOCUMENT_TYPES = {
|
24 |
+
BratBuilder.DEFAULT_CONFIG_NAME: BratDocumentWithMergedSpans,
|
25 |
+
}
|
26 |
+
|
27 |
+
# we need to add None to the list of dataset variants to support the default dataset variant
|
28 |
+
BASE_BUILDER_KWARGS_DICT = {
|
29 |
+
dataset_variant: {"url": URL, "split_paths": SPLIT_PATHS}
|
30 |
+
for dataset_variant in ["default", None]
|
31 |
+
}
|
32 |
+
|
33 |
+
DOCUMENT_CONVERTERS = {
|
34 |
+
TextDocumentWithLabeledSpansAndBinaryRelations: {
|
35 |
+
"spans": "labeled_spans",
|
36 |
+
"relations": "binary_relations",
|
37 |
+
},
|
38 |
+
}
|
requirements.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
pie-datasets>=0.4.0,<0.9.0
|