README.md

metadata

task_categories:
  - text-classification
  - text-generation
language:
  - en
tags:
  - biomedical
  - causal-relation-extraction
  - nlp

PubMedCausal: A Span-Level Annotated Corpus for Causal Relation Extraction in Biomedical Text

PubMedCausal is a comprehensive benchmark suite for causal relation extraction (CRE) and detection on biomedical text. Built from PubMed abstracts, it features 30,000 annotated rows designed to evaluate models on their ability to detect and extract causal claims as they are actually expressed in scientific literature.

• Paper: PubMedCausal: A Span-Level Annotated Corpus for Causal Relation Extraction in Biomedical Text
• Repository: https://github.com/josiahpaul07/PubMedCausal_Exp

Dataset Summary

The corpus contains 30,000 paragraph-level rows, including 3,945 causal rows and 6,491 adjudicated cause–effect pairs. Each causal relation is annotated with:

• Full-text cause and effect spans.
• Causality type: (e.g., Explicit vs. Implicit).
• Sententiality: (Intra-sentential vs. Inter-sentential).

It supports two primary tasks:

1. Causal Detection: Binary classification to identify if a text contains a causal claim.
2. Causal Extraction: Identifying specific cause-effect spans and their attributes.

Dataset Statistics

Task	Split	Train	Test
Detection	-	15,000	15,000
Extraction	Combined	1,972	1,973
Extraction	X_only (multi-pair)	1,206	1,207
Extraction	Y_only (single-pair)	766	766

Data Format

Raw Data (30k_train.json / 30k_test.json)

Each sample contains a sentence and up to 16 cause-effect pairs:

{
  "s/n": 6350,
  "Sentence": "Insulin resistance causes hyperglycemia.",
  "Cause 1": "Insulin resistance",
  "Effect 1": "hyperglycemia",
  "Sententiality 1": "Intra",
  "Causality 1": "Explicit",
  "Cause 2": "",
  ...
}

Extraction Task Format

{
  "s/n": 17500,
  "sentence": "Insulin resistance causes hyperglycemia.",
  "pairs": [
    {
      "cause": "Insulin resistance",
      "effect": "hyperglycemia",
      "sententiality": "Intra",
      "causality": "Explicit"
    }
  ],
  "num_pairs": 1
}

Citation

If you use this dataset in your research, please cite the following paper:

@article{paul2025pubmedcausal,
  title={PubMedCausal: A Span-Level Annotated Corpus for Causal Relation Extraction in Biomedical Text},
  author={Paul, Josiah and others},
  journal={arXiv preprint arXiv:2605.28363},
  year={2025}
}