基于关系记忆与路径信息的多跳知识图谱问答算法  

Multi-hop Knowledge Graph Question Answering Algorithm Based on Relational Memory and Path Information

作　　者：孟令鑫 才华[1] 付强[2] 易亚希 刘广文[1] 张晨洁[1] MENG Lingxin;CAI Hua;FU Qiang;YI Yaxi;LIU Guangwen;ZHANG Chenjie(School of Electronic and Information Engineering,Changchun University of Science and Technology,Changchun 130022,China;Institute of Space Ophotoelectronics Technology,Changchun University of Science and Technology,Changchun 130022,China)

机构地区：[1]长春理工大学电子信息工程学院,长春130022 [2]长春理工大学空间光电技术研究所,长春130022

出　　处：《吉林大学学报（理学版）》2024年第6期1391-1400,共10页Journal of Jilin University:Science Edition

基　　金：国家自然科学基金重大项目(批准号:61890963);国家自然基金联合基金(批准号:U2341226);吉林省人才专项基金(批准号:20240602015RC);2023年度西安市飞行器光学成像与测量技术重点实验室开放基金(批准号:2023-13);吉林省科技厅科技发展计划项目(批准号:20240302089GX).

摘　　要：针对自然语言处理领域中不完整知识图谱导致实体关联膨胀,进而需进行额外推理和推断使答案的推导过程变得更复杂的问题,提出一种结合关系记忆与路径信息的知识图谱问答算法RMP-KGQA.该算法利用关系记忆网络解决问题与知识图谱映射空间不一致的问题,利用其路径信息丰富评分函数,显著提高了智能问答检索系统的准确性和鲁棒性.实验结果表明,在基准数据集WebQSP和WebQSP-50上,RMP-KGQA的准确率分别比EmbedKGQA提升了2.8,2.4个百分点.消融实验进一步验证了关系记忆感知和路径信息在模型中的关键作用.因此,RMP-KGQA是一种解决复杂环境下多跳知识图谱问答问题的有效方法.Aiming at the problem that in the field of natural language processing,incomplete knowledge graphs led to the entity association expansion,which required additional inference and reasoning to make the derivation process of answers more complex,we proposed a knowledge graph question answering algorithm RMP-KGQA that combined relational memory and path information.The algorithm used a relational memory network to solve the problem of inconsistency between the problem and the knowledge graph mapping space,and enriched the scoring function with its path information,significantly enhancing the accuracy and robustness of the intelligent question answering retrieval system.The experimental results show that on the WebQSP and WebQSP-50 benchmark datasets,the accuracy of RMP-KGQA increases by 2.8 and 2.4 percentage points respectively compared to EmbedKGQA.Ablation experiments further verify the key roles of relational memory perception and path information in the model.Therefore,RMP-KGQA is an effective method for solving multi-hop knowledge graph question answering problems in complex environments.

关 键 词：知识图谱问答 知识图谱 知识图谱嵌入 关系记忆网络 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]