# PubMedCausal

Causal relation extraction corpus from PubMed abstracts.

**Paper:** Adewole et al. (2025). *PubMedCausal: A Biomedical Causal Relation Extraction Corpus.*
arXiv:2605.28363

**Source:** https://github.com/josiahpaul07/PubMedCausal_Exp

## Statistics

| Split | Sentences | Causal | Non-causal | Pairs |
|-------|-----------|--------|------------|-------|
| Train | 15,000    | 1,972  | 13,028     | ~3,900 |
| Test  | 15,000    | 1,973  | 13,027     | ~2,600 |

Pairs are typed along two dimensions:
- **Causality:** Explicit / Implicit
- **Sententiality:** Intra-sentential / Inter-sentential

No countercausal annotations — all relations map to `Relation.Procausal`.

## Conversion

Data is fetched directly from GitHub at conversion time; no local download is needed.

```bash
python conversion_script.py
```

Produces three task parquets per split under this directory:

```
causality-detection/
causal-candidate-extraction/
causality-identification/
```

## Known limitation — unresolvable spans in the identification task

PubMedCausal sometimes records the **canonical / normalised form** of a span rather
than the exact surface text from the sentence. For example:

```
sentence: "...thus precluding practice and placebo effects..."
effect annotation: "precludes practice"   ← lemmatised, does not match
```

Because the identification task requires inserting `<e1>…</e1>` markers at exact
character positions, pairs where either span cannot be located verbatim are
**skipped** during conversion. The conversion script prints a warning for each
skipped pair and a summary count at the end.

This affects **1,078 train pairs and 997 test pairs** (roughly 28% of intra-sentential
pairs), predominantly Implicit-causality annotations where the annotator wrote the
inferred proposition rather than the surface string. The detection and extraction
tasks are unaffected.