基于历史对比学习的时序知识图谱补全  

Completion of Temporal Knowledge Graph for Historical Contrastive Learning

作　　者：许智宏 邱鹏林 王利琴 董永峰 XU Zhihong;QIU Penglin;WANG Liqin;DONG Yongfeng(School of Artificial Intelligence and Data Science,Hebei University of Technology,Tianjin 300401,China;Hebei Key Laboratory of Big Data Computing,Tianjin 300401,China;Hebei Engineering Research Center of Data-Driven Industrial Intelligence,Tianjin 300401,China)

机构地区：[1]河北工业大学人工智能与数据科学学院,天津300401 [2]河北省大数据计算重点实验室,天津300401 [3]河北省数据驱动工业智能工程研究中心,天津300401

出　　处：《计算机工程与应用》2024年第22期154-161,共8页Computer Engineering and Applications

基　　金：河北省高等学校科学技术研究项目(ZD2022082,QN2021213);河北省高等教育教学改革研究与实践项目(2022GJJG049)。

摘　　要：针对现有的时序知识图谱补全模型高度依赖历史上已经发生过的事件,对历史上未发生过的事件预测不够准确的问题,提出了一种加入时序信息的对比历史与非历史信息的时序知识图谱补全模型(completion of temporal knowledge graph for comparing historical and non-historical information,CHNH)。该模型通过BiLSTM捕捉序列中的长期依赖关系,确保准确地编码历史信息。使用RGCN进行图卷积操作,从而学习到全局的图表示。在预测过程中,针对分开编码的历史和非历史信息,采用不同的评分函数来确定预测实体对这两类信息的依赖程度。通过这种方式,模型能够更有效地补全实体和关系,提高模型的预测性能。在ICEWS18、GDELT和YAGO数据集上的实验结果表明,CHNH模型在MRR、Hits@1、Hits@3和Hits@10上普遍优于基线模型。Aiming at the problem that the existing temporal knowledge graph completion model is highly dependent on events that have occurred in history and the prediction of events that have not occurred in history is inaccurate,a comple-tion of temporal knowledge graph for comparing historical and non-historical information(CHNH)with time series infor-mation is proposed.Firstly,the model captures long-term dependencies in the sequence through BiLSTM,ensuring accu-rate encoding of historical information.Then,the graph convolution operation is performed using RGCN to learn the global graph representation.In the prediction process,different scoring functions are used for separately coded historical and non-historical information to determine the dependence degree of the prediction entity on these two types of information.In this way,the model can more effectively complete entities and relationships,improving the predictive performance of the model.Experimental results on ICEWS18,GDERT and YAGO datasets show that the CHNH model generally outper-forms the baseline model in MRR,Hits@1,Hits@3 and Hits@10.

关 键 词：时序知识图谱 历史信息 非历史信息 对比学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]