Entity Fusion Contrastive Inference Network for Biomedical Document Relation Extraction

In recent years, the field of biomedical information has experienced remarkable growth. Consequently, the extraction of semantic relationships between biological entities from unstructured biomedical documents has gained increasing significance. Recent research has often employed sequential or graph models to predict relationships among biological entities in scientific articles. However, these models may not fully harness contextual information, resulting in the absence of entity reference details that can influence relationship judgments. In this paper, we introduce the EFCI model: Entity Fusion Contrastive Inference Network. Comprising an Entity Information Exchange Fusion module and a Contrast Enhanced Inference module. This model facilitates the interaction of essential information from the contextual context of both the head and tail entities through the information exchange fusion module. It consolidates this information into a feature matrix and subsequently employs the contrast enhancement inference module to capture implicit dependency relationships between entity pairs. This expansion extends the coverage of relational triples compared to prior studies. Additionally, the model enhances its inference capabilities and effectively addresses the issue of imbalanced label distribution in biomedical literature. Our comprehensive experiments demonstrate significant performance improvements of our model compared to the baseline model, showcasing its competitive advantage across two biomedical datasets: BIORED and CDR.