Improving Biomedical Entity Linking with Cross-Entity Interaction

Biomedical entity linking (EL) is the task of mentions in a biomedical document to corresponding entities knowledge base (KB). The challenge EL lies leveraging mention context select most appropriate among possible candidates. Although some models achieve competitive results by retrieving candidate and then exploiting re-rank them, these re-ranking concatenate with one at time. They lack fine-grained interaction candidates, potentially cannot handle ambiguous when facing candidates both high lexical similarity. We cope this issue using model based on prompt tuning, which represents all once, letting comparison attend each other. also propose KB-enhanced self-supervised pretraining strategy. Instead large-scale data previous work, we use masked language modeling synonyms from KB. Our method achieves state-of-the-art 3 datasets: NCBI disease, BC5CDR COMETA, showing effectiveness cross-entity Code available https://github.com/HITsz-TMG/Prompt-BioEL.