Explainable Reasoning over Temporal Knowledge Graphs by Pre-trained Language Model

Guanzhong Wu

Abstract: Temporal knowledge graph reasoning (TKGR) has been considered as a crucial task for modeling evolving knowledge, aiming to infer the unknown connections between entities at specific times. Traditional TKGR methods try to aggregate structural information between entities and evolve representations of entities over distinct snapshots, while some other methods attempt to extract temporal logic rules from historical interactions. However, these methods fail to address the continuously emerging unseen entities over time and ignore the historical dependencies between entities and relations. To overcome these limitations, we propose a novel method, termed TPNet, which introduces historical information completion strategy (HICS) and pre-trained language model (PLM) to conduct explainable inductive reasoning over TKGs. Specifically, TPNet extracts reliable temporal logical paths from historical subgraphs using a temporal-correlated search strategy. For unseen entities, we utilize HICS to sample or generate paths to supplement their historical information. Besides, a PLM and a time-aware encoder are introduced to jointly encode the temporal paths, thereby comprehensively capturing dependencies between entities and relations. Moreover, the semantic similarity between the query quadruples and the extracted paths is evaluated to simultaneously optimize the representations of entities and relations. Extensive experiments on entity and relation prediction tasks are conducted to evaluate the performance of TPNet. The experimental results on four benchmark datasets demonstrate the superiority of TPNet over state-of-the-art TKGR methods, achieving improvements of 14.35%, 23.08%, 6.75% and 5.38% on MRR, respectively.