Document-Level Biomedical Entity Relation Extraction Based on Semantic Segmentation

Document-level biomedical entity-relationship extraction aims to extract relationship facts from unstructured biomedical literature. Since many relationship facts span multiple sentences and involve complex entity interactions, the model needs to have strong logical reasoning capabilities. Although previous methods have achieved good performance, the fine-grained representation of entities and reasoning abilities remain relatively weak. Inspired by the semantic segmentation task in computer vision, this paper proposes the DSUNet model to enhance the model's logical reasoning ability. Specifically, DSUNet uses a pre-trained language model in the biomedical domain as an encoder to strengthen entity representation and calculates a relationship matrix between entities through contextual semantics. The semantic segmentation module is then employed to capture local and global information about entity relations, effectively capturing the contextual information of entities and the global interdependencies between triples, thereby enhancing the model's logical reasoning capability in relation extraction. Experimental results on the biomedical datasets CDR and GDA demonstrate that DSUNet exhibits excellent performance.