基于时序感知的动态知识图谱补全方法  

Temporal Aware Approach for Dynamic Knowledge Graph Completion

作　　者：李凤英[1] 范伟豪 LI Fengying;FAN Weihao(Guangxi Key Laboratory of Trusted Software,Guilin University of Electronic Technology,Guilin,Guangxi 541004,China)

机构地区：[1]桂林电子科技大学广西可信软件重点实验室,广西桂林541004

出　　处：《计算机工程与应用》2022年第15期202-209,共8页Computer Engineering and Applications

基　　金：国家自然科学基金(62062029)。

摘　　要：针对动态知识图谱的补全方法大多将时间维度内嵌于实体或关系中,将四元组降维成三元组后以静态知识图谱补全理论进行补全。静态补全方法通常只对实体关系建模,忽略了时间信息在四元组中的重要作用。同时知识库内时间表述存在稀疏性和不规则性。针对以上问题,提出了时序感知编码器和时序卷积解码器。时序感知编码器将时间维度同实体和关系嵌入为同规模向量,通过改进的图卷积神经网络实现四元组的特征提取。针对时序编码器特征提取后的四元组向量,时序卷积解码器利用卷积神经网络评估全局关系以进行链接预测。所提出的方法可以提供更精确的时间维度特征,提升补全时序图谱的性能。在ICEWS14、ICEWS05-15、Wikidata12k和YAGO11k数据集上的实验验证了提出方法的有效性,同时链接预测效果较优。Most current approaches for dynamic knowledge graph completion embed temporal dimension into entities or relations.In other words,quaternions are reduced into triples to be completed in the theory of static knowledge graph completion.However,static knowledge graph completion approaches usually focus on entities and relations,which ignore the role of temporal information in the quaternions.Meanwhile,there are sparsity and irregularity in the temporal repre-sentation within the knowledge base.To address the above problems,a temporal aware encoder and a temporal convolu-tional decoder are proposed in this paper.The temporal aware encoder embeds temporal entities,relations and times as same scale vectors and extracts features of quaternions by an improved graph convolutional networks.The temporal convolutional decoder is improved by convolutional neural networks.And it evaluates featrures of the quaternions extracted by the encoder for link prediction.Proposed approach provides more accurate features of time dimension and improves the performance of the completion in temporal knowledge graph.Experiments verify the effectiveness of the proposed work on ICEWS14,ICEWS05-15,Wikidata12k and YAGO11k datasets.Even more,the proposed approach makes a significant improvement in link prediction.

关 键 词：动态知识图谱补全 链接预测 图卷积神经网络 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]