A document-level relation extraction method based on dual-angle attention transfer fusion

Fuyuan Wei, Wenzhong Yang, Shengquan Liu, Chenghao Fu, Qicai Dai, Danni Chen, Xiaodan Tian, Bo Kong, Liruizhi Jia

Abstract: Document-level relation extraction has significant promise in practical applications and has captured the attention of numerous researchers. The multientity nature of document-level relation extraction shows that a task relies on multi-hop reasoning between entities and obtains entity-related information from the context. However, previous studies have neglected the potential association and dynamic fusion between the multi-hop reasoning process and contextual information. To address this problem, we propose a document-level relation extraction method based on dual-angle attention transfer fusion. First, a dual-angle attention transfer module captures the characteristics of entity pairs from two angles of attention mechanisms: entity pair<->entity pair and entity pair<->context. In this module, we propose a transfer module to construct potential associations between contextual information and multi-hop reasoning. Second, we propose a reweighted fusion module that dynamically weights and fuses different features according to their importance to each other. Finally, we conduct experiments on five public datasets (RE-DocRED, GDA, CDR, HacRED, and DocRED), demonstrating that our model significantly improves the performance of document-level relation extraction.