计算机工程 ›› 2020, Vol. 46 ›› Issue (8): 21-26.doi: 10.19678/j.issn.1000-3428.0055390

• 热点与综述 • 上一篇    下一篇

改进的胶囊网络知识图谱补全方法

王维美, 史一民, 李冠宇   

  1. 大连海事大学 信息科学技术学院, 辽宁 大连 116026
  • 收稿日期:2019-07-04 修回日期:2019-09-27 发布日期:2019-10-10
  • 作者简介:王维美(1994-),女,硕士研究生,主研方向为智能信息处理;史一民,副教授、硕士;李冠宇,教授、博士。
  • 基金项目:
    国家自然科学基金(61371090,61976032);辽宁省自然科学基金(20170540232,20170540144,20180540003)。

Improved Knowledge Graph Completion Method for Capsule Network

WANG Weimei, SHI Yimin, LI Guanyu   

  1. College of Information Science and Technology, Dalian Maritime University, Dalian, Liaoning 116026, China
  • Received:2019-07-04 Revised:2019-09-27 Published:2019-10-10

摘要: 为准确表征知识图谱中实体与关系属性的关系,提出一种改进的胶囊网络知识图谱补全方法。将表示多关系数据的三元组转换为矩阵的形式与多个过滤器进行卷积,产生不同特征图并重构为相应的胶囊,每个胶囊代表一组神经元。在此基础上,通过路由操作产生维度较小的胶囊,生成连续向量并将其与权重向量做点积运算,构建评分函数用于判断三元组的正确性。采用公开数据集WN18RR、FB15K-237、FB15K分别进行链接预测和三元组分类实验,结果表明,与DistMult、ComplEx、ConvE等模型相比,该算法链接预测性能较优,与TransE、TransH、TransR等模型相比,其三元组分类准确率达到91.5%,具有显著优势。

关键词: 知识图谱, 胶囊网络, 知识图谱补全, 链接预测, 三元组分类

Abstract: In order to complete the missing relations between entities of knowledge graph,this paper proposes an improved knowledge graph completion method for capsule network.First,the triplets are presented as a 3-column matrix,which is convolved with multiple filters to produce different feature maps.Secondly,these feature maps are reconstructed into corresponding capsules,each capsule composed of a group of neurons.The capsules with a lower dimension are produced through routing operation,and then a continuous vector is generated.Finally,the vector and the weight vector are subjected to a dot product operation to construct a ranking function to determine the correctness of the given triple.Experiments for link prediction and triplet classification are carried out on the public data sets including WN18RR,FB15K-237 and FB15K.The experimental results show that the proposed algorithm outperforms DistMult,ComplEx,ConvE and other models in link prediction.Also,it outperforms TransE,TransH,TransR and other models in triple classification,increasing the accuracy to 91.5%.

Key words: knowledge graph, capsule network, knowledge graph completion, link prediction, triad classification

中图分类号: