BELHD: improving biomedical entity linking with homonym disambiguation.

Area of Science:

• Biomedical informatics
• Computational biology
• Natural language processing

Background:

• Biomedical entity linking (BEL) grounds mentions to knowledge bases (KBs).
• Neural name-based BEL methods excel but fail with homonyms (multiple entities with the same name).
• Homonyms significantly impact BEL performance in large KBs like UMLS and NCBI Gene.

Purpose of the Study:

• To address the challenge of homonyms in neural name-based BEL.
• To introduce a novel method, BELHD (Biomedical Entity Linking with Homonym Disambiguation), for improved entity linking accuracy.

Main Methods:

• BELHD extends the BioSyn model with two key strategies.
• KB preprocessing: Expanding homonyms with unique disambiguating strings.
• Candidate sharing: Incorporating similar mentions within a document to enhance training signals.

Main Results:

• BELHD achieved state-of-the-art performance across 10 corpora and 5 entity types.
• Significant improvement in recall@1, averaging 4.55pp across tested datasets.
• Demonstrated superior results in 6 out of 10 corpora compared to existing neural BEL approaches.

Conclusions:

• BELHD effectively resolves homonyms in biomedical entity linking.
• The KB preprocessing component is model-agnostic and can enhance other neural BEL methods.
• BELHD represents a significant advancement for accurate biomedical entity recognition and linking.