Improving biomedical entity linking with generative relevance feedback

Area of Science:

• Biomedical informatics
• Natural Language Processing
• Knowledge Discovery

Background:

• Biomedical Entity Linking (BEL) is crucial for mapping text mentions to standardized identifiers.
• Current BEL systems face limitations in recall due to suboptimal candidate retrieval.
• This restricts the overall effectiveness of biomedical text normalization.

Purpose of the Study:

• To systematically evaluate Generative Relevance Feedback (GRF) for improving candidate retrieval in BEL.
• To assess GRF's impact on direct linking prediction and cascading normalization pipelines.
• To analyze GRF's sensitivity to different LLMs, feedback types, and integration strategies.

Main Methods:

• Implemented GRF leveraging large language models (LLMs) for zero-shot mention enrichment.
• Evaluated GRF in direct linking prediction and candidate generation scenarios.
• Conducted experiments across eight corpora and four biomedical knowledge bases.

Main Results:

• GRF significantly improved both accuracy and recall in BEL candidate retrieval.
• Enhanced performance increased the upper bound for normalization.
• Demonstrated GRF's effectiveness across diverse biomedical datasets and knowledge bases.

Conclusions:

• GRF offers an efficient and model-agnostic solution for enhancing BEL.
• GRF has the potential to be a key component in advancing biomedical entity linking.
• The study provides a systematic evaluation and reproducible code for GRF in BEL.