Improving inference via rich path information and logic rules for document-level relation extraction

作     者：Su, Huizhe Xie, Shaorong Yu, Hang Yuan, Changsen Wang, Xinzhi Luo, Xiangfeng 

作者机构：Shanghai Univ Sch Comp Engn & Sci Shanghai 200444 Peoples R China Beijing Inst Technol Sch Comp Sci & Technol Beijing 100876 Peoples R China 

出 版 物：《KNOWLEDGE AND INFORMATION SYSTEMS》 (Knowl. Inf. Systems. Syst.)

年 卷 期：2025年第67卷第5期

页      面：4207-4231页

核心收录：

学科分类：0711[理学-系统科学] 07[理学] 08[工学] 070105[理学-运筹学与控制论] 081101[工学-控制理论与控制工程] 0701[理学-数学] 071101[理学-系统理论] 0811[工学-控制科学与工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：National Key Research and Development Program of China [2021YFC3300602] Outstanding Academic Leader Project of Shanghai [20XD1401700] National Natural Science Foundation of China Shanghai Sailing Program [22YF1413600] 

主　　题：Document-level relation extraction Logical reason Inter-sentence inference Entity co-occurrence graph Rule-based inference 

摘      要：The task of document-level relation extraction is the classification of the relations between pairs of entities within a document. The identification of relations between inter-sentence entity pairs is one of the main challenges, as there may be no direct connections between the entities. Previous researches have attempted to address this challenge by leveraging path information between entities in the graph to predict their relations. However, these methods ignore the insufficiency of relations information in the existing paths and the absence of connecting paths between entity pairs. In this paper, we propose an effective inference model that enhances inter-sentence reasoning at both the document and global levels. Our model enhances path information by aggregating features from various sources along the logical reasoning paths between entities within each document. Additionally, the model learns relational inference rules from large graphs created from multiple documents and applies these rules to enrich existing relational knowledge. The experimental results indicate that our model outperforms existing models on three widely used benchmark datasets. Moreover, further analysis highlights that our model is especially effective in document-level relation extraction, particularly for inter-sentence relation extraction.