PubMedCausal

Intra-sentence No Polarity Signal No Strength Signal

PubMedCausal is a biomedical causal relation extraction corpus built from PubMed abstracts, typing each cause-effect pair along two dimensions: Explicit/Implicit causality and Intra-/Inter-sentential sententiality. There are no countercausal annotations — every relation converts to Relation.Causal.

Each pair's causality field (Explicit/Implicit) is already present in the raw JSON causalatee fetches but is not yet read by the converter. This is not a strength/certainty signal, though — it records whether an explicit causal connective is present in the text, not whether the causal claim itself is hedged (e.g. "may cause") or asserted with confidence markers (e.g. "I think"). It's a distinct, currently unmodeled property (connective explicitness), left undocumented as a pill on purpose to avoid conflating the two.

Overview

Domain	Biomedical
Year	2026
Sentences	30,000
Granularity	Intra-sentence

Tasks & Splits

Task	Train	Test
Causality Detection	15,000	15,000
Causal Event Candidate Extraction	1,972	1,973
Causality Identification	1,529	1,528
Causality Extraction	—	—

Data Explorer

Causality Detection

Causal Event Candidate Extraction

Causality Identification

Citation

@article{kunle-john:2026,
  title = {{{PubMedCausal}}: {{A Span-Level Annotated Corpus}} for {{Causal Relation Extraction}} in {{Biomedical Text}}},
  shorttitle = {{{PubMedCausal}}},
  author = {{Kunle-John}, Ifeoluwa and Paul, Josiah and Agbaakin, Oluwatosin and Aina, Peter and Odezuligbo, Ikenna and Anuyah, Sydney},
  year = 2026,
  journal = {CoRR},
  volume = {abs/2605.28363},
  eprint = {2605.28363},
  doi = {10.48550/ARXIV.2605.28363},
  url = {https://doi.org/10.48550/arXiv.2605.28363},
  urldate = {2026-07-01},
  archiveprefix = {arXiv}
}