RotateCT: Knowledge Graph Embedding by Rotation and Coordinate Transformation in Complex Space

Proceedings of the 29th International Conference on Computational Linguistics, pages 4918–4932 October 12–17, 2022. 4918 RotateCT: Knowledge Graph Embedding by Rotation and Coordinate Transformation in Complex Space Yao Dong1,2, Lei Wang1∗, Ji Xiang1, Xiaobo Guo1, Yuqiang Xie1,2 1Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China 2School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China {dongyao,wanglei,xiangji,guoxiaobo,xieyuqiang}@iie.ac.cn Abstract Knowledge graph embedding, which aims to learn representations of entities and relations in knowledge graphs, finds applications in var- ious downstream tasks. The key to success of knowledge graph embedding models are the ability to model relation patterns including symmetry/antisymmetry, inversion, commuta- tive composition and non-commutative com- position. Although existing methods fail in modeling the non-commutative composition patterns, several approaches support this pat- tern by modeling beyond Euclidean space and complex space. Nevertheless, expanding to complicated spaces such as quaternion can eas- ily lead to a substantial increase in the amount of parameters, which greatly reduces the com- putational efficiency. In this paper, we propose a new knowledge graph embedding method called RotateCT, which first transforms the co- ordinates of each entity, and then represents each relation as a rotation from head entity to tail entity in complex space. By design , Ro- tateCT can infer the non-commutative compo- sition patterns and improve the computational efficiency. Experiments on multiple datasets empirically show that RotateCT outperforms most state-of-the-art methods on link predic- tion and path query answering. 1 Introduction Knowledge graphs (KGs) contain structured facts of the real world. Real-world large-scale KGs such as WordNet (Miller, 1995), YAGO (Suchanek et al., 2007), Freebase (Bollacker et al., 2008) and Nell (Mitchell et al., 2018) have been applied to strengthen the performance of several downstream tasks including recommender systems (Zhang et al., 2016), question answering (Hao et al., 2017), con- versation generation (Zhou et al., 2018), relation extraction (Vashishth et al., 2018) and machine translation (Zhao et al., 2020). Generally, KGs ∗Corresponding author. Sam Mary Charlie spouse mother John Rachel Figure 1: An example in real world. Sam’s mother’s spouse is John, i.e., Sam’s father; Sam’s spouse’s mother is Rachel, i.e., Sam’s mother-in-law. "mother" and "spouse" form a non-commutative composition pattern. have the following features: large-scale and incom- plete. Therefore, to exploit the semantic informa- tion in KGs, predicting missing links based on the existing facts has gained growing interest in recent years. This task can be divided into two categories according to the length of the path: link prediction and path query answering. Link prediction focuses on single-hop reasoning (e.g., answering the path s → r → ?, where r is a relation), while path query answering (PQA) focuses on multi-hop reasoning (e.g., answering the path query s → path → ?, where path contains multiple relations). A popular approach for link prediction and PQA is knowledge graph embedding (KGE), which en- codes each element in KG into a continuous low- dimensional vector space. The performance of KGE methods greatly relys on the ability of model- ing and inferring relation patterns …

Content truncated to 500 words