基于互逆和对称关系补全的知识图谱数据扩展方法  

A Knowledge Graph Data Expansion Method Based on Reciprocal and Symmetric Relationship Completion

作　　者：应坚超 蒲飞 徐晨鸥 杨柏林[1] YING Jian-chao;PUFei;XU Chen-ou;YANG Bai-lin(School of Computer and Information Engineering,Zhejiang Gongshang University,Hangzhou 310018,China)

机构地区：[1]浙江工商大学计算机与信息工程学院,杭州310018

出　　处：《西南大学学报（自然科学版）》2020年第11期43-51,共9页Journal of Southwest University(Natural Science Edition)

基　　金：浙江省重点研发项目(2019C01004).

摘　　要：知识图谱表示学习方法旨在将知识图谱的实体与关系表示为低维、稠密的向量,并用于高效的语义计算,在知识图谱的构建、融合以及其他方面发挥重要作用.传统的知识图谱表示学习模型通常考虑了知识图谱中已有的事实,而忽略了知识图谱中隐藏的语义信息,使得表示学习并不能充分地表达原知识图谱的信息.目前的数据增强知识图谱表示学习模型需要借助第三方工具或者大量人工干预,增强数据的可靠性与稳定性有待加强.该文基于集合论中所提出的互逆/对称关系概念,提出了关系统计扩展方法(Relationship Statistics Expansion,RSE)方法,即通过统计的方法从现有知识图谱中获取稳定且可靠的先验知识,将其用于数据集扩展.同时,利用先验知识对互逆关系的表示模长施加约束,更加符合语义逻辑.该研究分别在WN18,FB15K,WN11,NELL-995共4个常用数据集上进行链路预测任务来评价模型效果,采用了目前主流的4个具有代表性知识图谱表示学习模型TransE,DistMult,RotatE,HAKE作为基准.结合该文提出的RSE方法后,RSE-TransE的MRR值分别提高了7.9%,12%,4.5%,7.2%;RSE-DistMult的MRR值分别提高了11.3%,5.8%,4.1%,1%;RSE-RotatE的MRR值分别提高了2.6%,6.7%,5.1%,1%;RSE-HAKE的MRR值分别提高了3.2%,4.5%,5.5%,11.3%.实验结果表明,该文提出的基于互逆和对称关系补全的知识图谱数据扩展方法可以挖掘知识图谱中隐含的语义信息,并且能显著提升传统的知识图谱表示学习模型在链路预测任务上的准确率和性能.The knowledge map representation learning method aims to represent the entities and relationships of a knowledge map as low dimensional and dense vectors,and is used for efficient semantic computing.It plays an important role in the construction,fusion and other applications of knowledge maps.The traditional knowledge map representation learning model usually considers the existing facts in the knowledge map,but ignores the semantic information hidden in it,which makes representation learning unable to fully express the information of the original knowledge map.The current data enhancement knowledge map representation learning model needs the help of third-party tools or a large number of manual interventions,so the reliability and stability of data need to be strengthened.Based on the concept of reciprocal/symmetric relation in the set theory,this paper proposes an RSE(relationship statistics expansion)method to obtain stable and reliable prior knowledge from the existing knowledge map by statistical methods,and then use it to expand the data set.At the same time,it is more consistent with the semantic logic to constrain the length of the representation module of the reciprocal relationship by using the prior knowledge.The effect of the model is evaluated through the link prediction task on three common datasets:wn18,fb15k and WN11.Four representative knowledge map representation learning models(TransE,DistMult,RotatE and HAKE)are used as the benchmark.Combined with the RSE method proposed in this paper,the MRR values of RSE-transE are increased by 7.9%,12%,4.5%and 7.2%;the MRR values of RSE-DistMult are increased by 11.3%,5.8%,4.1%and 1%;the MRR values of RSE-RotatE are increased by 2.6%,6.7%,5.1%and 1%;and the MRR values of RSE-HAKE are increased by 3.2%,4.5%,5.5%and 11.3%,respectively.The experimental results show that the proposed data expansion method based on reciprocal and symmetric relation completion can mine the semantic information hidden in the knowledge map,and can significantly improve the accuracy

关 键 词：知识图谱 表示学习 互逆关系 对称关系 统计方法 数据扩展 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]