ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2022, Vol. 59 ›› Issue (9): 1966-1979.doi: 10.7544/issn1000-1239.20210651

• 人工智能 • 上一篇    下一篇

一种融合关系路径与实体描述信息的知识图谱表示学习方法

宁原隆1,周刚1,2,卢记仓1,杨大伟1,张田1   

  1. 1(战略支援部队信息工程大学 郑州 450001);2(数学工程与先进计算国家重点实验室(战略支援部队信息工程大学) 郑州 450001) (ningyuanlong@163.com)
  • 出版日期: 2022-09-01
  • 基金资助: 
    河南省科技攻关计划项目(192102210129)

A Representation Learning Method of Knowledge Graph Integrating Relation Path and Entity Description Information

Ning Yuanlong1, Zhou Gang1,2, Lu Jicang1, Yang Dawei1, Zhang Tian1   

  1. 1(Strategic Support Force Information Engineering University, Zhengzhou 450001);2(State Key Laboratory of Mathematical Engineering and Advanced Computing (Strategic Support Force Information Engineering University), Zhengzhou 450001)
  • Online: 2022-09-01
  • Supported by: 
    This work was supported by the Science and Technology Program of Henan Province (192102210129).

摘要: 知识图谱表示学习旨在通过学习的方法将知识图谱中的实体和关系映射到一个连续的低维向量空间而获得其向量表示.已有的知识图谱表示学习方法大多仅从三元组角度考虑实体间的单步关系,未能有效利用多步关系路径及其实体描述等重要信息,从而影响性能.针对上述问题,提出了一种融合关系路径与实体描述的知识图谱表示学习模型.首先,对知识图谱中的多步关系路径进行联合表示,将路径上的所有关系和实体相加,得到关系路径信息的表示;其次,使用BERT(bidirectional encoder representations from transformers)模型对实体描述信息进行编码,得到相对应的语义表示;最后,对知识图谱中的三元组表示、实体描述的语义表示以及关系路径的表示进行融合训练,得到融合向量表示.在FB15K,WN18,FB15K-237,WN18RR数据集上,对提出的模型和基准模型进行链接预测和三元组分类任务,实验结果表明,与现有的基准模型相比,该模型在2项任务中均具有更高的准确性,证明了方法的有效性和优越性.

关键词: 知识图谱, 表示学习, 知识推理, 关系路径, 实体描述

Abstract: The representation learning of knowledge graph aims to map entities and relationships of knowledge graph into a continuous low-dimensional vector space through the learning method to obtain its vector representation. Most existing knowledge graph representation learning methods only consider the single-step relationship between entities from the perspective of triples, and fail to effectively use important information such as multi-step relationship paths and entity descriptions, which affects performance. In response to the above problems, we propose a knowledge graph representation learning model(PDRL) that integrates relationship paths and entity descriptions. Firstly, it is to perform a joint representation on the multi-step relationship path in the knowledge graph, and obtain the representation of the relationship path information by adding all the relationships and entities on the path; secondly, use BERT model to encode entity description information to obtain the corresponding semantic representation; finally, the fusion training is performed on the triples in the knowledge graph, the semantic representation of entity description and the representation of the relationship path to obtain the fusion vector representation. On the FB15K, WN18, FB15K-237 and WN18RR data sets, the proposed model and the benchmark model are used to perform link prediction and triple classification tasks. The experimental results show that compared with the existing benchmark models, the model in this paper has higher performance in two tasks, which proves the effectiveness and superiority of this method.

Key words: knowledge graph, representation learning, knowledge reasoning, relation path, entity description

中图分类号: