BELXTR: Biomedical Entity Linking via Contextualized Token Retrieval

BELXTR utilizes a multi-vector late interaction architecture to capture token-level matching information instead of compressing mentions into single vectors. Applying task-specific training and active query expansion across ten corpora, the model achieves an average 5pp improvement in recall@1 over existing state-of-the-art embedding-based biomedical entity linking methods.