基于柯西-施瓦茨不等式的知识图谱稠密表示方法  

Dense Representation Method of Knowledge Graph Based on Cauchy-Schwarz Inequality

作　　者：林霞 王聪 李敏[1] 李俊华 LIN Xia;WANG Cong;LI Min;LI Jun-hua(College of Computer Science,Sichuan Normal University,Chengdu 610101,China;Department of Computer Science and Technology,Sichuan Police College,Luzhou 646000,China)

机构地区：[1]四川师范大学计算机科学学院,成都610101 [2]四川警察学院计算机科学与技术系,四川泸州646000

出　　处：《小型微型计算机系统》2023年第2期300-306,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61602331)资助;四川省重点实验室开放课题项目(NDSMS201606)资助;四川省教育厅重点项目(17ZA0322)资助;四川省教育厅科研项目(17ZB0361)资助.

摘　　要：知识图谱稠密表示将形态各异的知识转化为结构化的实值向量,是当前有效的知识表示方式之一,广泛应用于知识计算和知识推理.在知识图谱稠密表示发展过程中,现有表示方法采用随机梯度下降法优化知识向量,导致知识嵌入精确度低.为此,本文提出了基于柯西-施瓦茨不等式的知识图谱稠密表示方法,将知识嵌入到低维稠密的向量空间.首先借助基于翻译操作的嵌入学习模型,利用势能函数表示实体和关系在向量空间中的距离,然后基于柯西-施瓦茨不等式,以极小化势能函数,最后在低维向量空间中最优化知识图谱的实体和关系向量.在基于数据集FB15k和WN18的对比实验中,度量标准hits@10和hits@1均得到了提升,证明了该方法提高了知识图谱稠密表示的准确性.The dense knowledge graph represents the transformation of different forms of knowledge into structured real-valued vectors.It is one of the current effective knowledge representation methods and is widely used in knowledge calculation and knowledge reasoning.In the development process of the dense representation of the knowledge graph,the existing representation method uses the stochastic gradient descent method to optimize the knowledge vector.This method leads to low accuracy of knowledge embedding.For this reason,this paper proposes a dense representation method of knowledge graph based on Cauchy-Schwarz inequality,which embeds knowledge into a low-dimensional dense vector space.First,with the help of the embedded learning model based on translation operation,the potential energy function is used to express the distance between entities and relationships in the vector space.Then the method is based on the Cauchy-Schwarz inequality to minimize the potential energy function.Finally,the entity and relationship vectors of the knowledge graph are optimized in the low-dimensional vector space.In the comparative experiments based on the data sets FB15k and WN18,it is shown that the metrics hits@10 and hits@1 have been improved,which proves that the method improves the accuracy of the dense representation of the knowledge graph.

关 键 词：知识图谱 稠密表示 翻译操作 势能函数 柯西-施瓦茨不等式 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]