PIE Dataset Card for "abstrct"

This is a PyTorch-IE wrapper for the AbstRCT dataset (paper and data repository). Since the AbstRCT dataset is published in the BRAT standoff format, this dataset builder is based on the PyTorch-IE brat dataset loading script.

Therefore, the abstrct dataset as described here follows the data structure from the PIE brat dataset card.

Dataset Summary

A novel corpus of healthcare texts (i.e., RCT abstracts on various diseases) from the MEDLINE database, which are annotated with argumentative components (i.e., MajorClaim, Claim, and Premise) and relations (i.e., Support, Attack, and Partial-attack), in order to support clinicians' daily tasks in information finding and evidence-based reasoning for decision making.

Usage

from pie_datasets import load_dataset
from pie_datasets.builders.brat import BratDocumentWithMergedSpans
from pytorch_ie.documents import TextDocumentWithLabeledSpansAndBinaryRelations

# load default version
dataset = load_dataset("pie/abstrct")
assert isinstance(dataset["neoplasm_train"][0], BratDocumentWithMergedSpans)

# if required, normalize the document type (see section Document Converters below)
dataset_converted = dataset.to_document_type("pytorch_ie.documents.TextDocumentWithLabeledSpansAndBinaryRelations")
assert isinstance(dataset_converted["neoplasm_train"][0], TextDocumentWithLabeledSpansAndBinaryRelations)

# get first relation in the first document
doc = dataset_converted["neoplasm_train"][0]
print(doc.binary_relations[0])
# BinaryRelation(head=LabeledSpan(start=1769, end=1945, label='Claim', score=1.0), tail=LabeledSpan(start=1, end=162, label='MajorClaim', score=1.0), label='Support', score=1.0)
print(doc.binary_relations[0].resolve())
# ('Support', (('Claim', 'Treatment with mitoxantrone plus prednisone was associated with greater and longer-lasting improvement in several HQL domains and symptoms than treatment with prednisone alone.'), ('MajorClaim', 'A combination of mitoxantrone plus prednisone is preferable to prednisone alone for reduction of pain in men with metastatic, hormone-resistant, prostate cancer.')))

Supported Tasks and Leaderboards

• Tasks: Argumentation Mining, Component Identification, Boundary Detection, Relation Identification, Link Prediction
• Leaderboard: More Information Needed

Languages

The language in the dataset is English (in the medical/healthcare domain).

Dataset Variants

The abstrct dataset comes in a single version (default) with BratDocumentWithMergedSpans as document type. Note, that this in contrast to the base brat dataset, where the document type for the default variant is BratDocument. The reason is that the AbstRCT dataset has already been published with only single-fragment spans. Without any need to merge fragments, the document type BratDocumentWithMergedSpans is easier to handle for most of the task modules.

Data Schema

See PIE-Brat Data Schema.

Document Converters

The dataset provides document converters for the following target document types:

• pytorch_ie.documents.TextDocumentWithLabeledSpansAndBinaryRelations
  • LabeledSpans, converted from BratDocumentWithMergedSpans's spans
    • labels: MajorClaim, Claim, Premise
  • BinraryRelations, converted from BratDocumentWithMergedSpans's relations
    • labels: Support, Partial-Attack, Attack

See here for the document type definitions.

Data Splits

Diseease-based Split	neoplasm	glaucoma	mixed
No.of document - _train - _dev - _test	350 50 100	100	100

Important Note:

• mixed_test contains 20 abstracts on the following diseases: glaucoma, neoplasm, diabetes, hypertension, hepatitis.
• 31 out of 40 abstracts in mixed_test overlap with abstracts in neoplasm_test and glaucoma_test.

Label Descriptions and Statistics

In this section, we describe labels according to Mayer et al. (2020), as well as our label counts on 669 abstracts.

Unfortunately, the number we report does not correspond to what Mayer et al. reported in their paper (see Table 1, p. 2109). Morio et al. (2022; 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).

Components

Components	Count	Percentage
MajorClaim	129	3 %
Claim	1282	30.2 %
Premise	2842	66.8 %

• MajorClaim are more general/concluding claim's, which is supported by more specific claims
• Claim is a concluding statement made by the author about the outcome of the study. Claims only points to other claims.
• 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.

(Mayer et al. 2020, p.2110)

Relations

Relations	Count	Percentage
support: Support	2289	87 %
attack: Partial-Attack	275	10.4 %
attack: Attack	69	2.6 %

• Support: All statements or observations justifying the proposition of the target component
• 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
• Attack: A component is attacking another one, if it is
  • i) contradicting the proposition of the target component, or
  • ii) undercutting its implicit assumption of significance constraints
• Premise can only be connected to either Claim or another Premise
• Claim's can only point to other Claim's
• There might be more than one outgoing and/or incoming relation . In rare case, there is no relation to another component at all.

(Mayer et al. 2020, p.2110)

Example

Collected Statistics after Document Conversion

We use the script evaluate_documents.py from PyTorch-IE-Hydra-Template to generate these statistics. After checking out that code, the statistics and plots can be generated by the command:

python src/evaluate_documents.py dataset=abstrct_base metric=METRIC

where a METRIC is called according to the available metric configs in config/metric/METRIC (see metrics).

This also requires to have the following dataset config in configs/dataset/abstrct_base.yaml of this dataset within the repo directory:

_target_: src.utils.execute_pipeline
input:
  _target_: pie_datasets.DatasetDict.load_dataset
  path: pie/abstrct
  revision: 277dc703fd78614635e86fe57c636b54931538b2

For token based metrics, this uses bert-base-uncased from transformer.AutoTokenizer (see AutoTokenizer, and bert-based-uncased to tokenize text in TextDocumentWithLabeledSpansAndBinaryRelations (see document type).

Relation argument (outer) token distance per label

The distance is measured from the first token of the first argumentative unit to the last token of the last unit, a.k.a. outer distance.

We collect the following statistics: number of documents in the split (no. doc), no. of relations (len), mean of token distance (mean), standard deviation of the distance (std), minimum outer distance (min), and maximum outer distance (max). We also present histograms in the collapsible, showing the distribution of these relation distances (x-axis; and unit-counts in y-axis), accordingly.

Command

python src/evaluate_documents.py dataset=abstrct_base metric=relation_argument_token_distances

neoplasm_train (350 documents)

	len	max	mean	min	std
ALL	2836	511	132.903	17	80.869
Attack	72	346	89.639	29	75.554
Partial-Attack	338	324	59.024	17	42.773
Support	2426	511	144.481	26	79.187

Histogram (split: neoplasm_train, 350 documents)

neoplasm_dev (50 documents)

	len	max	mean	min	std
ALL	438	625	146.393	24	98.788
Attack	16	200	90.375	26	62.628
Partial-Attack	50	240	72.04	24	47.685
Support	372	625	158.796	34	99.922

Histogram (split: neoplasm_dev, 50 documents)

neoplasm_test (100 documents)

	len	max	mean	min	std
ALL	848	459	126.731	22	75.363
Attack	32	390	115.688	22	97.262
Partial-Attack	88	205	56.955	24	34.534
Support	728	459	135.651	33	73.365

Histogram (split: neoplasm_test, 100 documents)

glaucoma_test (100 documents)

	len	max	mean	min	std
ALL	734	488	159.166	26	83.885
Attack	14	177	89	47	40.171
Partial-Attack	52	259	74	26	51.239
Support	668	488	167.266	38	82.222

Histogram (split: glaucoma_test, 100 documents)

mixed_test (100 documents)

	len	max	mean	min	std
ALL	658	459	145.067	23	77.921
Attack	6	411	164	34	174.736
Partial-Attack	42	259	65.762	23	62.426
Support	610	459	150.341	35	74.273

Histogram (split: mixed_test, 100 documents)

Span lengths (tokens)

The span length is measured from the first token of the first argumentative unit to the last token of the particular unit.

We collect the following statistics: number of documents in the split (no. doc), no. of spans (len), mean of number of tokens in a span (mean), standard deviation of the number of tokens (std), minimum tokens in a span (min), and maximum tokens in a span (max). We also present histograms in the collapsible, showing the distribution of these token-numbers (x-axis; and unit-counts in y-axis), accordingly.

Command

python src/evaluate_documents.py dataset=abstrct_base metric=span_lengths_tokens

statistics	neoplasm_train	neoplasm_dev	neoplasm_test	glaucoma_test	mixed_test
no. doc	350	50	100	100	100
len	2267	326	686	594	600
mean	34.303	37.135	32.566	38.997	38.507
std	22.425	29.941	20.264	22.604	24.036
min	5	5	6	6	7
max	250	288	182	169	159

Histogram (split: neoplasm_train, 350 documents)

Histogram (split: neoplasm_dev, 50 documents)

Histogram (split: neoplasm_test, 100 documents)

Histogram (split: glucoma_test, 100 documents)

Histogram (split: mixed_test, 100 documents)

Token length (tokens)

The token length is measured from the first token of the document to the last one.

We collect the following statistics: number of documents in the split (no. doc), mean of document token-length (mean), standard deviation of the length (std), minimum number of tokens in a document (min), and maximum number of tokens in a document (max). We also present histograms in the collapsible, showing the distribution of these token lengths (x-axis; and unit-counts in y-axis), accordingly.

Command

python src/evaluate_documents.py dataset=abstrct_base metric=count_text_tokens

statistics	neoplasm_train	neoplasm_dev	neoplasm_test	glaucoma_test	mixed_test
no. doc	350	50	100	100	100
mean	447.291	481.66	442.79	456.78	450.29
std	91.266	116.239	89.692	115.535	87.002
min	301	329	292	212	268
max	843	952	776	1022	776

Histogram (split: neoplasm_train, 350 documents)

Histogram (split: neoplasm_dev, 50 documents)

Histogram (split: neoplasm_test, 100 documents)

Histogram (split: glucoma_test, 100 documents)

Histogram (split: mixed_test, 100 documents)

Dataset Creation

Curation Rationale

"[D]espite its natural employment in healthcare applications, only few approaches have applied AM methods to this kind of text, and their contribution is limited to the detection of argument components, disregarding the more complex phase of predicting the relations among them. In addition, no huge annotated dataset for AM is available for the healthcare domain (p. 2108)...to support clinicians in decision making or in (semi)-automatically filling evidence tables for systematic reviews in evidence-based medicine. (p. 2114)"

Source Data

MEDLINE database

Initial Data Collection and Normalization

Extended from the previous dataset in Mayer et al. 2018, 500 medical abstract from randomized controlled trials (RCTs) were retrieved directly from PubMed by searching for titles or abstracts containing the disease name.

(See the definition of RCT in the authors' guideline (Section 1.2) and US National Library of Medicine)

Who are the source language producers?

[More Information Needed]

Annotations

Annotation process

"An expert in the medical domain (a pharmacist) validated the annotation guidelines before starting the annotation process." (p. 2110)

"Annotation was started after a training phase, where amongst others the component boundaries were topic of discussion. Gold labels were set after a reconciliation phase, during which the annotators tried to reach an agreement. While the number of annotators vary for the two annotation phases (component and relation annotation).

On the annotation of argument components, "IAA among the three annotators has been calculated on 30 abstracts, resulting in a Fleiss’ kappa of 0.72 for argumentative components and 0.68 for the more fine-grained distinction between claims and evidence." (p. 2109)

On the annotation of argumentative relation, "IAA has been calculated on 30 abstracts annotated in parallel by three annotators, resulting in a Fleiss’ kappa of 0.62. The annotation of the remaining abstracts was carried out by one of the above mentioned annotators." (p. 2110)

See the Annotation Guideline for more information on definitions and annotated samples.

Who are the annotators?

Two annotators with background in computational linguistics. No information was given on the third annotator.

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

"These [intelligent] systems apply to clinical trials, clinical guidelines, and electronic health records, and their solutions range from the automated detection of PICO elements 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 quantity of data available for the different diseases and treatments, the exact information they necessitate and to present this information in a structured way, easy to be (possibly semi-automatically) analyzed...Given its aptness to automatically detect in text those argumentative structures that are at the basis of evidence-based reasoning applications, AM represents a potential valuable contribution in the healthcare domain." (p. 2108)

"We expect that our work will have a large impact for clinicians as it is a crucial step towards AI supported clinical deliberation at a large scale." (p. 2114)

Discussion of Biases

[More Information Needed]

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

• License: the AbstRCT dataset is released under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License
• Funding: This work is partly funded by the French government labelled PIA program under its IDEX UCA JEDI project (ANR-15-IDEX-0001). This work has been supported by the French government, through the 3IA Cote d’Azur Investments in the Future project managed by the National Research Agency (ANR) with the reference number ANR19-P3IA-0002

Citation Information

@inproceedings{mayer2020ecai,
  author    = {Tobias Mayer and
               Elena Cabrio and
               Serena Villata},
  title     = {Transformer-Based Argument Mining for Healthcare Applications},
  booktitle = {{ECAI} 2020 - 24th European Conference on Artificial Intelligence},
  series    = {Frontiers in Artificial Intelligence and Applications},
  volume    = {325},
  pages     = {2108--2115},
  publisher = {{IOS} Press},
  year      = {2020},
}

Contributions

Thanks to @ArneBinder and @idalr for adding this dataset.