计算机应用

• 人工智能与仿真 •    下一篇

基于关系记忆的胶囊网络知识图谱嵌入模型

陈恒1,2,王思懿1,李正光1,李冠宇2,刘鑫1   

  1. 1.大连外国语大学 语言智能研究中心,辽宁 大连 116044;
    2.大连海事大学 信息科学技术学院,辽宁 大连 116026
  • 收稿日期:2021-05-12 修回日期:2021-09-15 发布日期:2021-09-15 出版日期:2021-09-22
  • 通讯作者: 陈恒

Capsule network knowledge graph embedding model based on relational memory

  • Received:2021-05-12 Revised:2021-09-15 Online:2021-09-15 Published:2021-09-22

摘要: 作为一种语义知识库,知识图谱使用结构化三元组的形式存储真实世界的实体及其内在关系。为了推理知识图谱中缺失的真实三元组,考虑关系记忆网络较强的三元组表征能力和胶囊网络强大的特征处理能力,提出一种基于关系记忆的胶囊网络知识图谱嵌入模型。首先,通过编码实体和关系之间的潜在依赖关系和部分重要信息并形成编码嵌入向量;然后,与过滤器卷积以生成不同的特征图,再重组为对应的胶囊;最后,通过压缩函数和动态路由指定从父胶囊到子胶囊的连接,根据子胶囊与权重内积的得分判断当前三元组的可信度。链接预测实验表明,与CapsE相比,本文模型在WN18RR数据集上,倒数平均排名(MRR)和Hit@1评价指标,分别提高了3个百分点和7个百分点;在FB15K-237数据集上,分别提高了4个百分点和2个百分点。

关键词: 知识图谱, 关系记忆网络, 胶囊网络, 知识图谱嵌入, 动态路由

Abstract: As a semantic knowledge base, the knowledge graph uses structured triples to store real-world entities and their internal relationships. In order to infer the missing real triples in the knowledge graph, and consider the strong triple representation ability of the relational memory network and the powerful feature processing ability of the capsule network, a knowledge graph embedding model of the capsule network based on the relational memory network was proposed. First, the embedding vectors were obtained by encoding the potential dependencies and some important information between entities and relationships; then, the embedding vectors were convolved with the filter to generate different feature maps, from which the corresponding capsules was extracted; finally, the connection from the parent capsule to the child capsule was specified through the squash function and dynamic routing, and the confidence coefficient of the current triple was estimated by the inner product score of the child capsule and the weights. Link prediction experiments show that compared with CapsE, Mean-Reciprocal-Rank (MRR) and Hit@1 evaluation indicators of the proposed model on the WN18RR data set are increased by 3 percentage points and 7 percentage points respectively, and on the FB15K-237 data set, increased by 4 percentage points and 2 percentage points respectively. Experiments prove that the proposed model can more accurately infer the relationship between the head entity and the tail entity.

Key words: knowledge graph, relational memory network, capsule network, knowledge graph embedding, dynamic routing

中图分类号: