Abstract
Document-level biomedical relation extraction (BioDocuRE) is essential for biomedical knowledge discovery, as many factual relationships between biomedical entities span multiple sentences or even the entire document. Despite recent advances, existing approaches often overlook the comprehensive integration of external domain knowledge and fail to fully exploit the rich multi-granular structural and contextual information inherent in biomedical documents, thereby limiting their reasoning capacity and extraction accuracy for complex, long-range relations. Here, we introduce KnowFDI, a novel framework for document-level biomedical relation extraction that systematically combines local and global contextual information, explicit document structural features (including inter-sentence relations via bridge nodes), and external entity-centric domain knowledge. KnowFDI leverages a pre-trained language model for contextual encoding and employs multi-view representation learning augmented by a two-stage channel-wise attention fusion module to dynamically integrate these diverse information sources. Furthermore, descriptive knowledge from external biomedical knowledge bases is incorporated to enrich entity semantics and enhance relation inference robustness. We evaluate KnowFDI on two widely-used benchmarks, CDR and GDA. Experimental results demonstrate that KnowFDI achieves state-of-the-art performance, with significant and consistent improvements in overall and particularly inter-sentence relation extraction tasks, outperforming previous methods. Ablation studies confirm the necessity and combined efficacy of our hierarchical information fusion design and external knowledge integration, highlighting their crucial roles in deciphering complex document-level dependencies.