A Simple Temporal Information Matching Mechanism for Entity Alignment between Temporal Knowledge Graphs

Proceedings of the 29th International Conference on Computational Linguistics, pages 2075–2086 October 12–17, 2022. 2075 A Simple Temporal Information Matching Mechanism for Entity Alignment Between Temporal Knowledge Graphs Li Cai1,2, Xin Mao1, Meirong Ma3, Hao Yuan3, Jianchao Zhu3, Man Lan1 ∗ 1School of Computer Science and Technology, East China Normal University 2College of Computer Science and Technology, Guizhou University 3Shanghai Transsion Co., Ltd caili2020stu@gmail.com, xmao@stu.ecnu.edu.cn {meirong.ma, hao.yuan, jianchao.zhu}@transsion.com mlan@cs.ecnu.edu.cn Abstract Entity alignment (EA) aims to find entities in different knowledge graphs (KGs) that refer to the same object in the real world. Recent stud- ies incorporate temporal information to aug- ment the representations of KGs. The exist- ing methods for EA between temporal KGs (TKGs) utilize a time-aware attention mech- anism to incorporate relational and temporal information into entity embeddings. The ap- proaches outperform the previous methods by using temporal information. However, we be- lieve that it is not necessary to learn the em- beddings of temporal information in KGs since most TKGs have uniform temporal representa- tions. Therefore, we propose a simple graph neural network (GNN) model combined with a temporal information matching mechanism, which achieves better performance with less time and fewer parameters. Furthermore, since alignment seeds are difficult to label in real- world applications, we also propose a method to generate unsupervised alignment seeds via the temporal information of TKG. Extensive experiments on public datasets indicate that our supervised method significantly outperforms the previous methods and the unsupervised one has competitive performance. 1 Introduction Knowledge graphs (KGs) describe facts of the real world in a structured form of triples (h, r, t), where h represents the head entity, r represents the rela- tion, t represents the tail entity. KGs have drawn great research attention from the academia (Lin et al., 2021; Ji et al., 2022) and have been widely used to enhance downstream applications such as question answering (Saxena et al., 2020; Qiu et al., 2020) and recommendation systems (Anelli et al., 2021; Zhou et al., 2020). Recently, Wikidata (Vrandecic and Krötzsch, 2014) and YOGO2 (Hoffart et al., 2013) add tem- poral information to the relations to represent the ∗*Corresponding author �������������� �������� ������������ ������������ �������������� ���� ��������� ���� ����������� �������� ��������� ������� ������ ����� ���� ��������� ���� ���� ����� ��� ��������� ��� ������ ��������� ��� ���� ���� ��������� ��� Figure 1: One sample of EA between TKGs. TKG1 is a sub-graph in YAGO. TKG2 is a sub-graph in WIKI, where the entity names corresponding to the items are as follows: Q39444 (Ronaldinho), Q7156 (FC Barcelona), Q166177 (Ballon d’Or), Q1543 (A.C.Milan), Q59207 (Ronaldo), P54 (member of sports team), P166(award received). KGs more accurately. In temporal knowledge graphs (TKGs), triples are extended to quadruples as (h, r, t, τ), where τ represents the timestamps. Figure 1 is an example of entity alignment (EA) be- tween TKGs, where nodes represent entities, edges include relational and temporal information, and the quadruple (Ronaldinho, playsfor, A.C.Milan, 2008-2010) represents one of the facts. EA seeks to find the same entities in the real …

Content truncated to 500 words