Partial Annotation Learning for Biomedical Entity Recognition

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Biology

Background:

• Biomedical Named Entity Recognition (BioNER) is crucial for research.
• High-quality annotated data for BioNER is scarce and costly.
• Existing methods struggle with the unlabeled entity problem due to missing annotations.

Purpose of the Study:

• To systematically evaluate partial annotation learning for BioNER.
• To propose and validate a novel partial annotation learning model, TS-PubMedBERT-Partial-CRF.
• To address the performance degradation caused by missing entity annotations in BioNER.

Main Methods:

• Conducted a comprehensive evaluation across simulated missing entity annotation scenarios.
• Developed and tested the TS-PubMedBERT-Partial-CRF model.
• Standardized 16 BioNER corpora with five distinct entity types for benchmarking.
• Compared performance against state-of-the-art partial and full annotation models.

Main Results:

• Partial annotation learning methods effectively handle missing entity annotations in BioNER.
• The proposed TS-PubMedBERT-Partial-CRF model outperformed existing partial annotation models.
• Our model achieved a 38% higher F1-score than the PubMedBERT tagger under high missing entity rates.
• Entity mention recall demonstrated competitive alignment with fully annotated datasets.

Conclusions:

• Partial annotation learning is a viable strategy for BioNER with incomplete data.
• The TS-PubMedBERT-Partial-CRF model offers a significant advancement for BioNER.
• This approach mitigates the impact of missing annotations, improving model performance and recall.