The biomedical discourse relation bank.

Area of Science:

• Biomedical Informatics
• Computational Linguistics
• Natural Language Processing

Background:

• Identifying discourse relations (e.g., causal, contrastive) is vital for biomedical text mining.
• A corpus annotated with discourse relations is needed for developing and evaluating biomedical discourse processing methods.
• Limited resources currently exist for biomedical discourse relation annotation.

Purpose of the Study:

• To develop the Biomedical Discourse Relation Bank (BioDRB).
• To annotate explicit and implicit discourse relations in biomedical texts.
• To adapt and refine annotation guidelines from the Penn Discourse TreeBank (PDTB) for the biomedical domain.

Main Methods:

• Annotated 24 open-access full-text biomedical articles from the GENIA corpus.
• Adapted Penn Discourse TreeBank (PDTB) guidelines, introducing new conventions for sense classification.
• Achieved reliable inter-annotator agreement exceeding 80% for all annotation sub-tasks.

Main Results:

• Explicit discourse connectives reliably indicate coarse sense classification (90.9% accuracy, 0.89 F1 score).
• Performance degraded for refined sense classification (69.2% accuracy, 0.28 F1) due to data sparsity.
• Classifiers trained on PDTB performed poorly on BioDRB (54.5% accuracy, 0.57 F1), indicating domain-specific differences.

Conclusions:

• Discourse relations can be reliably annotated in biomedical text.
• Coarse sense disambiguation of explicit connectives is feasible using connectives as features.
• Domain-specific differences in connective usage necessitate specialized biomedical discourse resources.