bibtex long conference paper
Appeared in: Findings of the Association for Computational Linguistics: EACL 2027 (EACL 2027)
Abstract:
Most biomedical entity linking (BEL) systems decompose the task into two stages, first detecting entity spans and then linking them. This demands costly span annotation, is brittle on overlapping mentions, and is unnecessary when an application only needs the concepts a document expresses, not their locations. We formulate BEL as mention-free retrieval, mapping text directly to concepts with no span detection or boundary supervision. Standard retrieval represents each input as a single vector and scores concepts by dot product, which is known to be incapable of capturing certain relevance relations at any embedding dimension. We overcome this by adapting the Hypencoder architecture (previously applied in text ranking) to generate a small text-conditioned network that scores concepts directly, and present HypLink, its first application to entity linking. Trained on MedMentions, HypLink surpasses single-vector retrieval and the closest prior mention-free entity tagger, particularly on recall, comes within 0.025 R-Precision of a mention-based pipeline given gold entity spans while detecting no spans itself, and retains a decisive recall lead even when the inventory expands to the full UMLS.
BibTeX @inproceedings{elsherbeny:eacl2027-hyplink, author = {El Sherbeny, Hana and MacAvaney, Sean and Lever, Jake}, title = {HypLink: Text-conditioned Retrieval for Mention-free Biomedical Entity Linking}, booktitle = {Findings of the Association for Computational Linguistics: EACL}, year = {2027} }