Modeling the Correlations of Relations for Knowledge Graph Embedding

• Published in: Journal of computer science and technology Vol. 33; no. 2; pp. 323 - 334
• Main Authors: Zhu, Ji-Zhao, Jia, Yan-Tao, Xu, Jun, Qiao, Jian-Zhong, Cheng, Xue-Qi
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.03.2018; Springer; Springer Nature B.V; College of Computer Science and Engineering, Northeastern University, Shenyang 110169, China; Key Laboratory of Network Data Science and Technology, Institute of Computing Technology Chinese Academy of Sciences, Beijing 110190, China%Key Laboratory of Network Data Science and Technology, Institute of Computing Technology Chinese Academy of Sciences, Beijing 110190, China%College of Computer Science and Engineering, Northeastern University, Shenyang 110169, China
• Subjects: Artificial Intelligence; Computer Science; Correlation; Data Structures and Information Theory; Datasets; Decomposition; Embedding; Employment; Graphs; Information Systems Applications (incl.Internet); Knowledge representation; Mathematical models; Methods; Modelling; R&D; Regular Paper; Research & development; Software Engineering; Theory of Computation; Vector spaces; United States > US; California; knowledge graph embedding; low-rank; matrix decomposition
• ISSN: 1000-9000; 1860-4749
• DOI: 10.1007/s11390-018-1821-8

Knowledge graph embedding, which maps the entities and relations into low-dimensional vector spaces, has demonstrated its effectiveness in many tasks such as link prediction and relation extraction. Typical methods include TransE, TransH, and TransR. All these methods map different relations into the vector space separately and the intrinsic correlations of these relations are ignored. It is obvious that there exist some correlations among relations because different relations may connect to a common entity. For example, the triples (Steve Jobs, PlaceOfBrith, California) and (Apple Inc., Location, California) share the same entity California as their tail entity. We analyze the embedded relation matrices learned by TransE/TransH/TransR, and find that the correlations of relations do exist and they are showed as low-rank structure over the embedded relation matrix. It is natural to ask whether we can leverage these correlations to learn better embeddings for the entities and relations in a knowledge graph. In this paper, we propose to learn the embedded relation matrix by decomposing it as a product of two low-dimensional matrices, for characterizing the low-rank structure. The proposed method, called TransCoRe (Translation-Based Method via Modeling the Correlations of Relations), learns the embeddings of entities and relations with translation-based framework. Experimental results based on the benchmark datasets of WordNet and Freebase demonstrate that our method outperforms the typical baselines on link prediction and triple classification tasks.