Transferring From Textual Entailment to Biomedical Named Entity Recognition

Area of Science:

• Computational Biology
• Natural Language Processing
• Bioinformatics

Background:

• Biomedical Named Entity Recognition (BioNER) is crucial for extracting entities like genes, proteins, and diseases from text.
• BioNER faces significant challenges due to a scarcity of high-quality, token-level labeled data, especially in specialized biomedical domains.
• Existing methods using sequential labeling models with gazetteers produce noisy labels and have limited entity coverage.

Purpose of the Study:

• To develop a gazetteer-based BioNER system from scratch with zero token-level annotations.
• To address the issue of noisy labels and limited data in building BioNER systems.
• To improve the discrimination ability of BioNER models.

Main Methods:

• Formulated the BioNER task as a Textual Entailment problem.
• Proposed Textual Entailment with Dynamic Contrastive learning (TEDC) to alleviate noisy labeling.
• Employed dynamic contrastive learning to differentiate between entities and non-entities within sentences.

Main Results:

• TEDC effectively alleviates the noisy labeling problem inherent in gazetteer-based BioNER.
• The method successfully transfers knowledge from pre-trained textual entailment models.
• Achieved state-of-the-art performance on two real-world biomedical datasets for gazetteer-based BioNER.

Conclusions:

• Textual Entailment with Dynamic Contrastive learning (TEDC) offers a robust solution for BioNER with limited labeled data.
• The proposed approach significantly enhances the accuracy and discrimination capabilities of BioNER systems.
• TEDC provides a promising direction for building effective BioNER tools from scratch.