Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (15): 101-108.DOI: 10.3778/j.issn.1002-8331.2006-0421

Previous Articles     Next Articles

Knowledge Graph Completion Method for Semantic Hierarchies of Spherical Coordinate Modeling

CHEN Heng, QI Ruihua, ZHU Yi, YANG Chen, GUO Xu, WANG Weimei   

  1. 1.Research Center for Language Intelligence, Dalian University of Foreign Languages, Dalian, Liaoning 116044, China
    2.College of Information Science and Technology, Dalian Maritime University, Dalian, Liaoning 116026, China
  • Online:2021-08-01 Published:2021-07-26

球坐标建模语义分层的知识图谱补全方法

陈恒,祁瑞华,朱毅,杨晨,郭旭,王维美   

  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

Abstract:

The knowledge graph is a structured representation of real-world triples. Typically, triples are represented as (head entity, relationship, tail entity), which means that the head entity and tail entity are related to each other through a specific relationship. Aiming at the problem of sparse data widely existing in the knowledge graph, this paper proposes a knowledge graph completion method for semantic hierarchies of spherical coordinate modeling. In this paper, the spherical coordinate system is used to model entities and relationships for link prediction. Specifically, the radial coordinate aims to model entities at different levels, and the entity with a smaller radius has a higher level; the angular coordinate aims to distinguish entities at the same level, that is, entities with the same length and different angles. This method maps the entities into the spherical coordinate system, which can effectively model the semantic hierarchies that are common in the knowledge graph. In the experiment, the public data sets WN18RR, FB15K-237 and YAGO3-10 are used to carry out related link prediction experiments. Experimental results show that in WN18RR, the Mean Reciprocal Rank is 3.6% higher than RotatE, and Hit@10 is 1.9% higher than RotatE. In FB15K-237, the Mean Reciprocal Rank is improved 4.8% than ConvKB, Hit@10 is 3.5% higher than ConvKB. Experiments show that the knowledge graph completion method of spherical coordinate modeling semantic hierarchies can effectively improve the accuracy of triple prediction.

Key words: knowledge graph, semantic hierarchies, knowledge graph completion, spherical coordinate system, link prediction

摘要:

知识图谱是真实世界三元组的结构化表示。通常,三元组表示形式为(头实体,关系,尾实体),这表示头实体和尾实体通过特定关系相互联系。针对知识图谱中广泛存在的数据稀疏问题,提出一种球坐标建模语义分层的知识图谱补全方法。使用球坐标系对实体和关系进行建模表示,以进行链接预测。具体来说,半径坐标旨在对不同层级的实体进行建模,半径较小的实体级别越高;角度坐标旨在区分相同层级的实体,即模长相等而角度不同的实体。该方法将实体映射到球坐标系中,可以有效建模知识图谱中普遍存在的语义分层现象。实验中,采用公开数据集WN18RR、FB15K-237与YAGO3-10进行相关的链接预测实验。实验结果表明,在WN18RR中,平均倒数排名(Mean Reciprocal Rank)比RotatE提高3.6%,Hit@10比RotatE提高1.9%;在FB15K-237中,平均倒数排名(Mean Reciprocal Rank)比ConvKB提高4.8%,Hit@10比ConvKB提高3.5%。实验证明球坐标建模语义分层的知识图谱补全方法可以有效提高三元组预测准确度。

关键词: 知识图谱, 语义分层, 知识图谱补全, 球坐标系, 链接预测