知识图谱的增强CP分解链接预测方法  

Enhanced canonical polyadic decomposition link predictionembedding method of knowledge graph

作　　者：赵博 王宇嘉[1] 倪骥 Zhao Bo;Wang Yujia;Ni Ji(Institute of Electronic&Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201600,China)

机构地区：[1]上海工程技术大学电子电气工程学院,上海201600

出　　处：《计算机应用研究》2023年第5期1396-1401,共6页Application Research of Computers

基　　金：国家自然科学基金资助项目(61403249);科技创新2030-“新一代人工智能”重大项目(2020AAA0109300)。

摘　　要：CP分解作为知识图谱链接预测的方法之一,能够对一些包含常规数据的知识图谱进行链接预测补全。但当知识图谱存在大量稀疏数据及可逆关系时,该方法不能体现两个实体间具有的隐藏联系,无法对此类数据进行处理。为解决上述问题,提出增强CP分解方法,对三元组中前实体和后实体的两个嵌入向量分别进行学习,并在训练过程中使用概率方法生成更高质量的负例三元组,引入ELU损失函数和AMSGrad优化器,有效对可逆关系和稀疏数据进行处理。在通用数据集上的实验结果表明,所提方法可以有效提升链接预测精度,与对比模型相比取得了5%的性能提升,同时应用在汽车维修知识图谱数据集补全中,取得83.2%正确率的实体补全结果。Canonical polyadic decomposition as one of the methods for link prediction of knowledge graphs,enabling link prediction complementation of some knowledge graphs containing regular data.However,when there is a large amount of sparse data and reversible relationships in the knowledge graph,the method cannot reflect the hidden connection between two entities and cannot process such data.To address the above issues,this paper proposed an enhanced canonical polyadic decomposition method to learn the two embedding vectors of the front entity and the back entity in the triad separately,and used probabilistic methods to generate higher quality negative example triads during the training process.It introduced ELU loss function and AMSGrad optimiser to effectively process reversible relationships.Experimental results on the generic dataset show that the proposed method can effectively improve link prediction accuracy,with a 5%performance improvement compared to the comparison model,and is also applied in the complementation of the automotive repair knowledge graph dataset,achieving 83.2%correct entity complementation results.

关 键 词：知识图谱 链接预测 CP分解 知识图谱嵌入 知识图谱补全 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]