Not Just Single Reasoning! A Hierarchical Network for Document-level Biomedical Relation Extraction with Multi-Granularity Learning

Document-level biomedical relation extraction (Bio-DocRE) involves identifying relations between entities distributed across multiple sentences in biomedical literature. Most existing approaches construct a document graph structure to model entity-level relation maps and achieve local implicit reasoning at the entity-pair granularity. However, these methods have two main drawbacks. On the one hand, mention-pair granularity is crucial for modeling entity-level relation map and global reasoning but is often ignored in most approaches. On the other hand, these approaches primarily achieve local reasoning of cross-sentence relation instances with the interaction of sentence or document nodes in the graph. Such an approach has difficulty modeling global reasoning and cannot completely address the sub-optimal performance caused by the long dependency issue. To tackle these issues, we propose a novel framework based on hierarchical multi-granularity learning and reasoning (HMGLR) for Bio-DocRE. Experimental results on two widely used biomedical datasets demonstrate that our model achieves new state-of-the-art performance.