摘要
知识图谱是真实世界三元组的结构化表示.通常,三元组被表示成(头实体,关系,尾实体)的形式.为补全知识图谱中缺失的三元组,提出一种改进的Tucker分解知识图谱补全算法.该算法利用Tucker分解将三阶张量表示的知识图谱分解成一个核心张量与每个mode上因子矩阵的乘积.通过将三阶张量分解成一个核心张量每一维度乘以一个因子矩阵的形式,利用打分函数计算每个三元组的得分,得到每个三元组正确的概率,将正确的三元组添加到知识图谱,对知识图谱进行补全.实验中,采用5个公开数据集WN18RR、FB15K-237、WN18、FB15K和NELL-995进行相关的链接预测实验.实验结果表明,在WN18RR中,平均倒数排名(Mean Reciprocal Rank)比TuckER提高3.1%,Hit@10比TuckER提高1.1%;在FB15K-237中,平均倒数排名(Mean Reciprocal Rank)比TuckER提高3.4%,Hit@3比TuckER提高1.1%;在NELL-995中,平均倒数排名(Mean Reciprocal Rank)比ConvE提高3.3%,Hit@10比ConvE提高2.1%.实验证明改进的Tucker分解算法可以有效提高三元组预测精度.
The knowledge graph is a structured representation of real world triples.Typically,triples are represented in the form of(head entities,relationships,tail entities).In order to complete the missing triples in the knowledge graph,an improved Tucker decomposition knowledge graph completion algorithm is proposed.This algorithm uses Tucker decomposition to decompose the knowledge graph represented by the third-order tensor into a product of a core tensor and a factor matrix on each mode.By decomposing the third-order tensor into a core tensor and multiplying by a factor matrix in each dimension,the scoring function is used to calculate the grade of each triple,and get their correct probability,then the correct ones are added to the knowledge graph and achieve completion.In the experiment,five open data sets WN18RR,FB15K-237,WN18,FB15K and NELL-995 are used for the related link prediction experiments.The experimental results show that in WN18RR,Mean Reciprocal Rank is 3.1% higher than TuckER,and Hit@10 is 1.1% higher than TuckER.In FB15K-237,Mean Reciprocal Rank is 3.4% higher than TuckER,Hit@3 is 1.1% higher than TuckER.In NELL-995,the Mean Reciprocal Rank is 3.3% higher than ConvE,and Hit%10 is 2.1% higher than ConvE.Experiments show that this algorithm can effectively improve the prediction accuracy of the triple.
作者
陈恒
李冠宇
祁瑞华
朱毅
郭旭
CHEN Heng;LI Guan-yu;QI Rui-hua;ZHU Yi;GUO Xu(Research Center for Language Intelligence,Dalian University of Foreign Languages,Dalian 116044,China;Faculty of Information Science&Technology,Dalian Maritime University,Dalian 116026,China)
出处
《数学的实践与认识》
北大核心
2020年第16期164-176,共13页
Mathematics in Practice and Theory
基金
国家自然科学基金项目(61976032,61371090,61806038)
国家社会科学基金一般项目(15BYY028)
辽宁省自然科学基金项目(2019-ZD-0513)
2020年辽宁省教育厅科学研究经费项目(多目标协同的知识图谱补全机理研究)
辽宁省高等学校基本科研项目(2017JYT09)
大连外国语大学研究创新团队"计算语言学与人工智能创新团队"(2016CXTD06)。
关键词
知识图谱
张量分解
知识图谱补全
链接预测
knowledge graph
tensor factorization
knowledge graph completion
link prediction
