Attention Based Document-level Relation Extraction with None Class Ranking Loss

• DOI: 10.24963/ijcai.2024/726
• Publication Date: 2024-07-26T14:28:11Z
• Authors: Xiaolong Xu; Chenbin Li; Haolong Xiang; Lianyong Qi; Xuyun Zhang; Wanchun Dou
• Keywords: Relationship extraction

Through document-level relation extraction (RE), the analysis of global between entities in text is feasible, and more comprehensive accurate semantic information can be obtained. In RE, model needs to infer implicit relations two different sentences. To obtain information, existing methods mainly focus on exploring entity representations. However, they ignore correlations indivisibility relations, contexts. Furthermore, current only independently estimate cases predefined ignoring case "no relation'', which results poor prediction. address above issues, we propose a RE method based attention mechanisms, considers relation''. Specifically, our approach leverages graph multi-head networks capture among entities, contexts, respectively. addition, novel multi-label loss function that promotes large margins label confidence scores each class none employed improve prediction performance. Extensive experiments conducted benchmarking datasets demonstrate proposed outperforms state-of-the-art baselines with higher accuracy.