PILOT sets new SOTA for biomedical entity linking with neighborhood-aware dual reranking
PILOT uses ontology structure to link biomedical mentions with state-of-the-art accuracy
Biomedical entity linking is the task of grounding mentions in clinical and scientific text to entities in a curated knowledge base (KB) with ontological structure. It underpins critical applications such as literature-scale information extraction and electronic health record normalization, but faces tough challenges: huge KBs, highly ambiguous mentions, and corpus-specific annotation conventions. In a new arXiv preprint, researchers Yicheng Tao and Jie Liu introduce PILOT, a three-stage framework designed to tackle these issues head-on.
PILOT's first stage performs neighborhood-aware retrieval, injecting ontological structure from both the query and KB side by reformulating mentions and pooling entity embeddings. The retrieved candidate pool is then scored from two complementary views—one over surface forms and one over context—before the scores are fused together. This dual reranking approach captures both lexical and semantic signals, while the score fusion step balances their contributions. The authors report that PILOT achieves state-of-the-art performance on average across five widely-used benchmarks, while remaining efficient at inference time—a practical advantage for deployment on large-scale biomedical corpora. The work is available on arXiv under ID 2608.04144.
- PILOT is a three-stage framework: neighborhood-aware retrieval, dual reranking, and score fusion
- It injects ontological structure from both query and KB sides, using mention reformulation and entity embedding pooling
- Achieves state-of-the-art average results across five biomedical entity linking benchmarks with efficient inference
Why It Matters
More accurate biomedical entity linking improves literature mining and patient-record normalization, accelerating clinical research and healthcare data interoperability.