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

Abstract

Aim: Biomedical entity linking is essential in natural language processing for identifying and linking biomedical concepts to entities in a knowledge base. Current methods, which involve a multistage recognition-retrieve-read process, achieve high performance but are hindered by slow inference times and error propagation.

Methods: The authors propose ER2, an End-to-End entity linking paradigm following a Retrieval-Rerank framework. It reversely selects mentions in context and their corresponding entities based on the prior knowledge of candidate entities, enabling jointly performing candidates retrieval, mention detection, and candidates rerank in one pass via a lighten-weight reranker that models deep relevance between the context and its candidates at the embedding level. We further introduce a more powerful cross-encoder as the teacher model, thereby enhancing the rerank performance via knowledge distillation from the teacher to the student reranker.

Results: Experiments on several end-to-end entity linking benchmarks demonstrate the efficiency and effectiveness. Notably, our method achieves competitive performance compared with the previous state-of-the-art methods while being nearly 10 times faster.

Conclusions: The research has a significant reference for connecting mentions within unstructured contexts to their corresponding entities in KBs, thereby facilitating the application effect of downstream tasks such as automatic diagnosis, drug–drug interaction prediction and personalized medicine and other fields.

Keywords

biomedical entity linking / cross-encoder / end-to-end entity linking / knowledge distillation / natural language processing / retrieval-then-rerank

RIGHTS & PERMISSIONS

2025 Chinese Cochrane Center, West China Hospital of Sichuan University and John Wiley & Sons Australia, Ltd.