基于图注意力网络的时序知识图谱人机交互模型  

A Human-Computer Interaction Model over Temporal Knowledge Graphs Based on Graph Attention Networks

作　　者：于泳 乔少杰 陈金勇[3] 高林 黄江涛 刘晨旭 韩楠[5] 张桃 蔡宏果 YU Yong;QIAO Shaojie;CHEN Jinyong;GAO Lin;HUANG Jiangtao;LIU Chenxu;HAN Nan;ZHANG Tao;CAI Hongguo(School of Software Engineering,Chengdu University of Information Technology,Chengdu 610225,China;Sichuan Southwest Liansheng Communication Technology Co.,Ltd.,Yibin 644012,China;The 54th Research Institute of CETC,Shijiazhuang 050081,China;School of Computer and Information Engineering,Nanning Normal University,Nanning 530001,China;School of Management,Chengdu University of Information Technology,Chengdu 610225,China;Faculty of Artificial Intelligence and Big Data,Yibin University,Yibin 644012,China;School of Mathematics and Computer Science,Nanning Normal University,Nanning 530001,China)

机构地区：[1]成都信息工程大学软件工程学院,四川成都610225 [2]四川西南联盛通讯技术有限公司,四川宜宾644012 [3]中国电子科技集团公司第五十四研究所,河北石家庄050081 [4]南宁师范大学计算机与信息工程学院,广西南宁530001 [5]成都信息工程大学管理学院,四川成都610225 [6]宜宾学院人工智能与大数据学部,四川宜宾644012 [7]南宁师范大学数学与信息科学学院,广西南宁530001

出　　处：《无线电工程》2024年第7期1676-1686,共11页Radio Engineering

基　　金：国家自然科学基金(62272066);四川省科技计划(2021JDJQ0021,2022YFG0186,2022NSFSC0511,2023YFG0027);教育部人文社会科学研究规划基金(22YJAZH088);宜宾市引进高层次人才项目(2022YG02);成都市“揭榜挂帅”科技项目(2022-JB00-00002-GX,2021-JB00-00025-GX);成都市技术创新研发项目(重点项目)(2024-YF08-00029-GX);成都市区域科技创新合作项目(2023-YF11-00020-HZ);中国电子科技集团公司第五十四研究所高校合作课题(SKX212010057);成都海关科研项目资助(2022CK008)。

摘　　要：组织和检索信息是人机交互重点关注的话题之一。基于知识图谱(Knowledge Graph,KG)的智能问答系统通过语义解析用户问题,检索知识并回答问题,已成为一种信息检索的有效途径,是人机交互的典型应用。时序知识图谱(Temporal Knowledge Graph,TKG)问答系统通过语言模型获取问题中的实体和时间戳,并在大型TKG中检索答案。TKG问答系统包含2个挑战:①给定问题,需检索整个TKG,效率低且易受干扰项的影响;②难以捕获问题中隐含的时间词和时间顺序信息。提出一种基于图注意力网络的时间对比学习(Time Contrast Learning,TCL)模型,将源问题与替换时间词后的对比问题同时训练,使用图注意力网络更新实体邻接子图的节点特征,缩小潜在答案的检索空间。在CRONQUESTIONS数据集上进行大量实验,结果表明TCL比其他基准方法具有更好的性能,相较于最先进的基准方法在H it@1和Hits@10指标上平均提升3.44%和2.02%。Organizing and retrieving information is one of the key topics in human-computer interaction.Intelligent question-answering system over Knowledge Graph(KG)has become an effective way to retrieve information through semantic parsing of user's questions,retrieving knowledge and answering questions,which is a typical application of human-computer interaction.Question-answering system over Temporal Knowledge Graph(TKG)obtains entities and timestamps from questions through the language model and retrieves answers from a large TKG However,the question-answering system over temporal knowledge graphs contains two challenges:①Given the question,the entire TKG needs to be retrieved,which is inefficient and vulnerable to interfering items;②It is difficult to capture the time words and chronological information in the question.A Time Contrast Learning(TCL)is proposed,which trains both the source problem and the contrast problem after replacing the temporal words,and uses the graph attention network to update the node features of the entity neighbor sub-graph,in order to reduce the retrieval space for potential answers.Extensive experiments are conducted on the CRONQUESTIONS dataset,and the results show that TCL has better performance than other benchmark methods with an average i mprovement of 3.44%in the metric of Hits@1 and 2.02%in Hits@10,respectively,when compared to the state-of-the-art baseline method.

关 键 词：智能问答 时序知识图谱 图注意力网络 时间对比学习 语言模型 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]