MetaTKG: Learning Evolutionary Meta-Knowledge for Temporal Knowledge Graph Reasoning

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7230 - 7240 December 7-11, 2022 ©2022 Association for Computational Linguistics MetaTKG: Learning Evolutionary Meta-Knowledge for Temporal Knowledge Graph Reasoning Yuwei Xia1,3 , Mengqi Zhang2,4 , Qiang Liu2,4∗, Shu Wu2,4 , Xiao-Yu Zhang1,3∗ 1Institute of Information Engineering, Chinese Academy of Sciences 2Center for Research on Intelligent Perception and Computing (CRIPAC) National Laboratory of Pattern Recognition (NLPR) Institute of Automation, Chinese Academy of Sciences 3School of Cyber Security, University of Chinese Academy of Sciences 4School of Artificial Intelligence, University of Chinese Academy of Sciences xiayuwei@iie.ac.cn, mengqi.zhang@cripac.ia.ac.cn {qiang.liu,shu.wu}@nlpr.ia.ac.cn, zhangxiaoyu@iie.ac.cn Abstract Reasoning over Temporal Knowledge Graphs (TKGs) aims to predict future facts based on given history. One of the key challenges for prediction is to learn the evolution of facts. Most existing works focus on exploring evo- lutionary information in history to obtain ef- fective temporal embeddings for entities and relations, but they ignore the variation in evolu- tion patterns of facts, which makes them strug- gle to adapt to future data with different evo- lution patterns. Moreover, new entities con- tinue to emerge along with the evolution of facts over time. Since existing models highly rely on historical information to learn embed- dings for entities, they perform poorly on such entities with little historical information. To tackle these issues, we propose a novel Tem- poral Meta-learning framework for TKG rea- soning, MetaTKG for brevity. Specifically, our method regards TKG prediction as many tem- poral meta-tasks, and utilizes the designed Tem- poral Meta-learner to learn evolutionary meta- knowledge from these meta-tasks. The pro- posed method aims to guide the backbones to learn to adapt quickly to future data and deal with entities with little historical information by the learned meta-knowledge. Specially, in temporal meta-learner, we design a Gating In- tegration module to adaptively establish tempo- ral correlations between meta-tasks. Extensive experiments on four widely-used datasets and three backbones demonstrate that our method can greatly improve the performance. 1 Introduction Temporal Knowledge Graphs (TKGs) (Boschee et al., 2015) are of great practical values as an effective way to represent the real-world time- evolving facts. In TKGs, the expression of a fact is extended from a triple to a quadruple, such as *To whom correspondence should be addressed. 1 2 3 4 Period 34 36 38 40 RE-NET RE-GCN (a) [0,50] [100,500] [1000,5000] Historical interaction numbers 30 40 50 60 0 500 1000 1500 MRR (%) Number of entities (b) Figure 1: (a) The y-axis is the performance (MRR (%)) of existing TKG models RE-NET and RE-GCN on pre- dicting facts in different time periods. (b) The left y-axis is the performance of RE-NET on predicting entities with different historical interaction numbers. And the right y-axis is the number of entities in each group. The results of both (a) and (b) are obtained based on the widely used TKG dataset ICEWS14. (Obama, run for, president, 2012). Due to the great practical values of TKGs, related research on TKG reasoning has continued to …

Content truncated to 500 words