KGE-CL: Contrastive Learning of Tensor Decomposition Based Knowledge Graph Embeddings

Proceedings of the 29th International Conference on Computational Linguistics, pages 2598–2607 October 12–17, 2022. 2598 KGE-CL: Contrastive Learning of Tensor Decomposition Based Knowledge Graph Embeddings Zhiping Luo1,4∗, Wentao Xu1,4∗, Weiqing Liu2, Jiang Bian2, Jian Yin3,4†, and Tie-Yan Liu2 1 School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China 2 Microsoft Research Asia, Beijing, China 3 School of Artificial Intelligence, Sun Yat-sen University, Zhuhai, China 4 Guangdong Key Laboratory of Big Data Analysis and Processing, Guangzhou, China {luozhp7@mail2,xuwt6@mail2,issjyin@mail}.sysu.edu.cn {weiqing.liu, jiang.bian, tyliu}@microsoft.com Abstract Learning the embeddings of knowledge graphs (KG) is vital in artificial intelligence, and can benefit various downstream applications, such as recommendation and question answering. In recent years, many research efforts have been proposed for knowledge graph embed- ding (KGE). However, most previous KGE methods ignore the semantic similarity be- tween the related entities and entity-relation couples in different triples since they sepa- rately optimize each triple with the scoring function. To address this problem, we pro- pose a simple yet efficient contrastive learn- ing framework for tensor decomposition based (TDB) KGE, which can shorten the seman- tic distance of the related entities and entity- relation couples in different triples and thus im- prove the performance of KGE. We evaluate our proposed method on three standard KGE datasets: WN18RR, FB15k-237 and YAGO3- 10. Our method can yield some new state-of- the-art results, achieving 51.2% MRR, 46.8% Hits@1 on the WN18RR dataset, 37.8% MRR, 28.6% Hits@1 on FB15k-237 dataset, and 59.1% MRR, 51.8% Hits@1 on the YAGO3- 10 dataset. Source codes and data of this paper can be found at https://github.com/ Wentao-Xu/KGE-CL. 1 Introduction The knowledge graph (KG) stores a vast number of human knowledge in the format of triples. A triple (h, r, t) in a KG contains a head entity h, a tail entity t, and a relation r between h and t. The knowledge graph embedding (KGE) aims to project the massive interconnected entities and relations in a KG into vectors or matrices, which can preserve the semantic information of the triples. Learning the embeddings of KG can benefit various down- stream artificial intelligence applications, such as ∗The first two authors contributed equally. †Corresponding author. question answering (Huang et al., 2019), machine reading comprehension (Yang and Mitchell, 2017), image classification (Marino et al., 2016), and per- sonalized recommendation (Wang et al., 2018). In general, most of the KGE methods would de- fine a scoring function f(hi, rj, tk), and the train- ing target of KGE is maximizing the score of a true triple (hi, rj, tk) and minimizing the score of a false tripe (hi, rj, tx). In this way, the trained embeddings of entities and relations in KG can pre- serve the intrinsic semantics of a true triple. We can mainly divide the existing KGE methods into two categories. The first category is the distance based (DB) methods, which use the Minkowski dis- tance as scoring function to measure a triple’s plau- sibility, include the TransE (Bordes et al., 2013), TransH (Wang et al., 2014), …

Content truncated to 500 words