Biomedical entity-aware semantic role labelling via model merging

Semantic Role Labeling (SRL) is essential for extracting structured knowledge from biomedical texts, yet it remains challenging due to the complexity of biomedical terminology. This paper introduces a novel method that enhances biomedical SRL by directly integrating Named Entity Recognition (NER) knowledge using model merging. NER plays a crucial role in anchoring domain-specific arguments, allowing the model to better distinguish semantic roles in biomedical contexts. Unlike prior approaches that rely on multitask learning or syntax-aware features, our method merges pretrained SRL and NER models in a way that preserves entity knowledge while aligning with semantic role structures. Experimental results demonstrate that our approach surpasses state-of-the-art multitask and syntax-aware SRL models in both accuracy and training speed. The integration of biomedical entity knowledge proves especially effective for predicates with high proportions of entity-based arguments, highlighting the value of direct NER-to-SRL knowledge transfer in advancing biomedical semantic understanding.