Fine-Grained Mention-Level Analysis of Biomedical Entity Linking Models

Biomedical Entity Linking (BEL) is essential for structuring knowledge from biomedical texts, yet global evaluation metrics often obscure systematic model weaknesses. We propose a fine-grained evaluation framework that analyzes performance across interpretable mention-level characteristics, including length, lexical variation, synonymy, homonymy, and training frequency. Using the BELB benchmark, we apply this analysis to neural and rule-based systems. Our results show that performance degradation stems from mention-level difficulty, with consistent drops across characteristics that reflect limited training coverage, and expose model weaknesses beyond aggregate scores in a unified benchmark setting.