An Effective and Efficient Entity Alignment Decoding Algorithm via Third-Order Tensor Isomorphism

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5888 - 5898 May 22-27, 2022 c⃝2022 Association for Computational Linguistics An Effective and Efficient Entity Alignment Decoding Algorithm via Third-Order Tensor Isomorphism Xin Mao1∗, Meirong Ma2, Hao Yuan2, Jianchao Zhu2, Zongyu Wang3, Rui Xie3, Wei Wu3, Man Lan1∗ 1School of Computer Science and Technology, East China Normal University 2Transsion Group, 3Meituan Group xmao@stu.ecnu.edu.cn,mlan@cs.ecnu.edu.cn {meirong.ma,hao.yuan,jianchao.zhu}@transsion.com {wangzongyu02,rui.xie,wuwei30}@meituan.com Abstract Entity alignment (EA) aims to find the equiva- lent entity pairs between KGs, which is a cru- cial step for integrating multi-source KGs. For a long time, most researchers have regarded EA as a pure graph representation learning task and focused on improving graph encoders while paying little attention to the decoding process. In this paper, we propose an effective and effi- cient EA Decoding Algorithm via Third-order Tensor Isomorphism (DATTI). Specifically, we derive two sets of isomorphism equations: (1) Adjacency tensor isomorphism equations and (2) Gramian tensor isomorphism equations. By combining these equations, DATTI could ef- fectively utilize the adjacency and inner cor- relation isomorphisms of KGs to enhance the decoding process of EA. Extensive experiments on public datasets indicate that our decoding algorithm can deliver significant performance improvements even on the most advanced EA methods, while the extra required time is less than 3 seconds. 1 Introduction Knowledge graphs (KGs) illustrate the relations be- tween real-world entities—e.g., objects, situations, or concepts—and usually are stored in the form of triples (subject, relation, object). Over recent years, a large number of KGs have been constructed to provide structural knowledge to facilitate down- stream applications, such as recommendation sys- tems (Cao et al., 2019) and question-answering systems (Zhao et al., 2020). Most KGs are independently extracted from dif- ferent languages or domains. Thus, these KGs usu- ally hold unique information individually but also have some shared parts. Integrating these cross- lingual / domain KGs could provide a broader view for users, especially for the minority language users who usually suffer from lacking language resources. As shown in Figure 1, entity alignment (EA) aims [Real Name] [Member] [Friend] [Enemy] Red Skull Captain America Avengers Steven Rogers Winter Soldier [真名] [朋友] [宿敌] 红骷髅 美国队长 史蒂夫罗 杰斯 冬日战士 [成员] Figure 1: An example of cross-lingual entity alignment. to find the equivalent entity pairs between KGs, which is a crucial step for integrating KGs. Existing EA methods are built on the same core premise: equivalent entity pairs between KGs have similar neighborhood structures (i.e., isomor- phism). Therefore, most existing EA methods (Wang et al., 2018; Sun et al., 2020b; Mao et al., 2020) could be abstracted into the same architec- ture (as shown in Figure 2): encoding the structural information of KGs into a low-dimensional vector space by Siamese graph encoders and then mapping equivalent entity pairs into the proximate space by alignment loss functions. For a long time, most researchers have regarded EA as a graph representation learning task and focused on improving graph encoders. Starting from the simplest graph …

Content truncated to 500 words