改进的TransH模型在知识表示与推理领域的研究  

Study on a modified TransH-based model in knowledge representation and reasoning field

作　　者：昌攀 曹扬 CHANG Pan;CAO Yang(National Engineering Laboratory for CETC Research Institute of Big Data, Guiyang 550081, China)

机构地区：[1]中电科大数据研究院国家工程实验室,贵州贵阳550081

出　　处：《广西大学学报（自然科学版）》2020年第2期321-327,共7页Journal of Guangxi University（Natural Science Edition）

基　　金：天津市科技计划项目(18ZXZNGX00370);国家自然科学基金资助项目(61572350)。

摘　　要：针对Trans系列的知识图谱表示与推理模型在训练的过程中,随机构造正负例三元组样本进行训练,没有考虑替换的实体与原实体之间存在的相似度差异度关系,导致模型无法识别实体之间的相似度,效果低下。在TransH模型的构建基础上,采用单层神经网络的非线性操作来精确刻画实体和关系之间的语义信息,同时创新性地加入了正、负三元组之间的头/尾实体之间的差异度信息,用于校正正、负三元组样本之间的联系,使模型能够辨别替换的实体与原实体间的相似度,进而提出了mTransH模型。实验证明:mTransH模型在知识图谱的链接预测任务中,提高了模型对正例样本的辨识度,从而提高知识推理的链接预测准确率。During the training of knowledge graph representation and reasoning model for Trans series,the relationship of similarity and difference between the replaced head or tail entity and the original entity are not considered when positive and negative sample triples are constructed for training,which leads to low efficiency and the model's inability to identify the similarity between entities.In this study,based on the construction of the TransH model,the nonlinear operation of single layer neural network is used to describe the semantic information between entity and relation.As an innovation,the difference between the head and tail entities of the positive and negative triples is introduced,so as to correct the relationship between the positive and negative triplet samples and to enable the mTransH model to identify the similarity between the replaced entity and the original entity,and then the mTransH model is proposed.Experimental results show that the mTransH model improves the recognition of positive samples in the link prediction task of knowledge graph,thus improving the link prediction accuracy of knowledge reasoning.

关 键 词：知识图谱 表示 推理 相似度 mTransH模型 链接预测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]