A contrastive dual attention framework for enhancing adverse drug event relations extraction

Accurate extraction of relationships between drugs and adverse drug events (ADEs) is essential for improving patient safety. However, current approaches struggle to capture complex relationships due to limitations in contextual representation. In the n2c2 dataset, ADE-Drug instances (1107) are significantly fewer than others such as Strength-Drug (6702) or Reason-Drug (5169), creating a strong class imbalance that hinders identification. A model based on transformer encoders is used to generate contextual embeddings, incorporating a dual attention mechanism that focuses on both the entities and their clinical context. Contrastive learning refines the representation of entity pairs, enabling more precise differentiation between correct and incorrect relationships. Experimental evaluations show a general F1 score of 93.31% and 83.31% for the ADE-Drug relation, outperforming previous methods. The combination of contextual encoding, specialized attention, and contrastive discrimination effectively addresses the challenges of class imbalance and the semantic complexity of clinical language.