Rethinking Logical Reasoning Design for Document-level Biomedical Relation Extraction

Document-level biomedical relation extraction (Bio-DocRE) involves automatically identifying relation facts between multiple entity pairs from a document. Logical reasoning has become a recent research focus since many relations of long-distance entity pairs in biomedical documents need to be predicted through reasoning from other existing relations. However, this task faces two crucial challenges: long dependency and evidence selection. Existing graph-based or transformer-based models independently model the internal structure of entities and ignore additional global interactions between relation triples. In addition, evidence graph reasoning plays an important role in boosting the logical reasoning between entities, but it is ignored by these methods. To tackle these issues, we propose a novel Bio-DocRE framework called EGALR based on evidence graph assisted logical reasoning. This allows EGALR to focus on important sentences while predicting implicit relations between entities with relation reasoning. Extensive experiments show that our EGALR model achieves state-of-the-art performance on two widely-used biomedical datasets, CDR and GDA.