Learning Latent Relations for Temporal Knowledge Graph Reasoning

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 12617–12631 July 9-14, 2023 ©2023 Association for Computational Linguistics Learning Latent Relations for Temporal Knowledge Graph Reasoning Mengqi Zhang1,2 , Yuwei Xia3,4 , Qiang Liu1,2 , Shu Wu1,2∗, Liang Wang1,2 1School of Artificial Intelligence, University of Chinese Academy of Sciences 2Center for Research on Intelligent Perception and Computing State Key Laboratory of Multimodal Artificial Intelligence Systems Institute of Automation, Chinese Academy of Sciences 3Institute of Information Engineering, Chinese Academy of Sciences 4School of Cyber Security, University of Chinese Academy of Sciences mengqi.zhang@cripac.ia.ac.cn,xiayuwei@iie.ac.cn, {qiang.liu,shu.wu,wangliang}@nlpr.ia.ac.cn Abstract Temporal Knowledge Graph (TKG) reasoning aims to predict future facts based on histori- cal data. However, due to the limitations in construction tools and data sources, many im- portant associations between entities may be omitted in TKG. We refer to these missing as- sociations as latent relations. Most of the exist- ing methods have some drawbacks in explicitly capturing intra-time latent relations between co-occurring entities and inter-time latent rela- tions between entities that appear at different times. To tackle these problems, we propose a novel Latent relations Learning method for TKG reasoning, namely L2TKG. Specifically, we first utilize a Structural Encoder (SE) to obtain representations of entities at each times- tamp. We then design a Latent Relations Learn- ing (LRL) module to mine and exploit the intra- and inter-time latent relations. Finally, we ex- tract the temporal representations from the out- put of SE and LRL for entity prediction. Exten- sive experiments on four datasets demonstrate the effectiveness of L2TKG. 1 Introduction Temporal knowledge graphs (TKGs) play a vital role in capturing temporal facts in the real world. Each fact in a TKG is represented as a quadru- ple (s, r, o, t), such as (Obama, run for, president, 2012). Reasoning over TKGs involves two primary settings: interpolation and extrapolation. In recent years, there has been significant interest in the ex- trapolation setting due to its practical value in event prediction (Deng et al., 2020), question answering (Mavromatis et al., 2022), and other applications. In the extrapolation setting, the objective is to pre- dict facts that occur at a time t with t > tn, based on the historical information available in the TKG from t0 to tn. *To whom correspondence should be addressed. (withdrawn) (complain) (go to) (leave) USA Taliban Border Area President Ghani Pakistan Afghan people 2021-05 2021-08 (Inter-time latent relation) (Intra-time latent relation) Afghanistan Afghanistan Afghanistan (will be governed by) ? 2021-09 … … (has occupied) Figure 1: An example of reasoning over TKG. The gray dotted lines indicate two types of latent relations Most extrapolation models utilize the temporal and structural information available in the TKG for reasoning. For example, RE-NET (Jin et al., 2020a) and RE-GCN (Li et al., 2021) incorporate recur- rent neural networks and graph neural networks to capture the temporal and structural dependencies of historical TKG sequences. Additionally, xERTE (Han et al., 2021a) and TITer (Sun et al., 2021) de- velop …

Content truncated to 500 words