融合语义与结构信息的知识图谱补全模型研究  

Research on the Semantic and Structure Fusion-Based Knowledge Graph Completion Model

作　　者：马志远 高颖 张强 周洪[3] 李兵 陶皖 Ma Zhiyuan;Gao Ying;Zhang Qiang;Zhou Hong;Li Bing;Tao Wan(School of Computer and Information,Anhui Polytechnic University,Wuhu 241000,China;School of Information Management,Central China Normal University,Wuhan 430079,China;National Science Library(WuHan),Chinese Academy of Science,Wuhan 430071,China;Exchange,Development&Service Center for Science&Technology Talents,the Ministry of Science&Technology,Beijing 100045,China)

机构地区：[1]安徽工程大学计算机与信息学院,芜湖241000 [2]华中师范大学信息管理学院,武汉430079 [3]中国科学院武汉文献情报中心,武汉430071 [4]科学技术部科技人才交流开发服务中心,北京100045

出　　处：《数据分析与知识发现》2024年第4期39-49,共11页Data Analysis and Knowledge Discovery

基　　金：湖北省教育厅哲学社会科学研究项目(项目编号:22Q095);安徽工程大学资助项目(项目编号:xjky2022147)的研究成果之一

摘　　要：【目的】针对知识图谱补全任务,挖掘语义与结构信息,完善知识图谱并提升质量与可靠性。【方法】提出一种融合语义与结构信息的知识图谱补全模型,通过预训练语言模型增强知识图谱内文本及上下文数据的嵌入表示,捕获实体与关系的语义信息,并构建实体-关系矩阵映射知识图谱网络结构,获取实体的邻域信息与关系约束,进一步融合潜在数据,进行模型训练并预测丢失实体,最终达成知识图谱补全任务。【结果】与基线方法性能相比,该模型的Hits@3评测指标在FB15k-237、WN18RR和UMLS数据集上分别提升0.5、0.6和0.6个百分点。【局限】受限于语言模型的基础表示能力,未能结合多模态数据进一步提升补全任务效果。【结论】该模型具有较好的补全性能,融合语义与结构信息的方式对比其他方法具有一定优势,能够较好地完成知识图谱补全任务,对知识图谱及其下游应用的发展具有重要意义。[Objective]This paper proposes a knowledge graph completion model with semantic and structural information.It improves the completion,reliability,and quality of the knowledge graph.[Methods]First,we used a pre-trained language model to enhance the knowledge graph’s embedded text and context data.Then,we captured the semantic information of entities and relationships.Third,we constructed an entity-relationship matrix to map the network structure of the knowledge graph and obtain each entity’s neighborhood information and relationship constraints.Finally,we integrated the potential data to train the model and predict the missing entity of the knowledge graph.[Results]Compared to the baseline method,the proposed model’s Hits@3 metric improved by 0.5%,0.6%and 0.6%on the FB15k-237,WN18RR and UMLS data sets,respectively.[Limitations]Due to the language models’information representation ability limits,we cannot further improve the knowledge graph by completing tasks with the help of multimodal data.[Conclusions]The proposed method can perform better for the knowledge graph completion task,promoting the knowledge graph’s development and its downstream application.

关 键 词：知识图谱补全 预训练语言模型 自然语言处理 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] G35[自动化与计算机技术—计算机科学与技术]