Retrieve Then Rerank: An End-to-End Learning Paradigm for Biomedical Entity Linking

Area of Science:

• Biomedical Natural Language Processing
• Bioinformatics
• Computational Biology

Background:

• Biomedical entity linking is crucial for NLP, connecting concepts to knowledge bases.
• Current multistage methods are accurate but slow and prone to errors.

Purpose of the Study:

• To develop an efficient and effective end-to-end entity linking paradigm.
• To overcome the limitations of slow inference and error propagation in existing methods.

Main Methods:

• Proposed ER2, an End-to-End entity linking paradigm using a Retrieval-Rerank framework.
• Implemented a lightweight reranker for joint candidate retrieval, mention detection, and reranking.
• Utilized knowledge distillation from a cross-encoder teacher model to enhance reranker performance.

Main Results:

• ER2 demonstrated efficiency and effectiveness on end-to-end entity linking benchmarks.
• Achieved competitive performance compared to state-of-the-art methods.
• The proposed method is nearly 10 times faster than previous approaches.

Conclusions:

• ER2 offers a significant advancement in linking mentions to knowledge base entities.
• Facilitates downstream applications like automatic diagnosis and drug-drug interaction prediction.
• Enhances the practical utility of biomedical NLP in personalized medicine and related fields.