Enhanced Differential Interaction Representation for Joint Entity-Relation Extraction

Although significant advances in jointly extracting entities and relationships from text, challenges remain in enhancing semantic richness and effectively leveraging the intrinsic relationships between subtasks. Many recent approaches either share feature representations, leading to feature interference, or independently encode the two tasks but neglect their inherent interdependence. In cases where interaction exists, they often prioritize commonalities, overlooking the critical capture of distinguishing information. To overcome this limitation, this paper introduces a novel Cross-Attention Information Interaction Fusion (CAIF) module, which separates the tasks into two independent subtasks, achieving effective fusion of common information while capturing differential information. Specifically, CAIF consists of a Differential Information Extraction Module (DIEM) and a Similar Information Fusion Module (SIFM). DIEM facilitates cross-tasks differential feature extraction, while SIFM captures common information and integrates long dependencies. Extensive experiments conducted on four datasets have shown that our model outperforms other baseline models.