An assessment of partial annotation learning for biomedical entity recognition

Partial annotation learning is useful for entity recognition when there are missing entity annotations. In our work, we systematically study partial annotation learning methods for biomedical entity recognition over different simulated scenarios for missing entity annotations. We harmonize 15 biomedical NER corpora encompassing five entity types to serve as golden standard. To explore the effectiveness of partial annotation learning methods, we compare two commonly used partial annotation learning models with the state-of-the-art biomedical entity recognition model PubMedBERT tagger. Our experiments show that partial annotation learning methods can effectively learn from biomedical corpora with even significant fractions of entity annotations missing, suggesting further work in this direction would be promising.