Coherent temporal logical reasoning via fusing multi-faced information for link forecasting over temporal knowledge graphs

Link forecasting over Temporal knowledge graphs (TKGs) aims to predict the unknown facts in future timestamps, which has gained increasing attention due to its significant practical value. Logical reasoning plays a pivotal role in this task by achieving explainable reasoning through extracting and applying temporal logical rules. However, existing logical reasoning methods are challenged in sufficiently and effectively utilizing the underlying information (e.g., relation dependencies and textual semantics) present in TKGs. In this paper, we propose a two-stage framework, FM-CTRL, to Fuse the Multi-faced information for Coherent TempoRal Logical reasoning over TKGs, which comprehensively considers the rich semantic information among entities, relations, and timestamps to learn reliable logical rules. In the first stage, we construct a temporal relation structure graph (TRSG) according to the structural dependencies between relations and introduce a temporal reverse search algorithm to extract temporal logical paths based on the TRSG. To expedite the search, we introduce a time-fusion search graph (TFSG) to facilitate real-time temporal path search without requiring additional post-processing of timestamps. In the second stage, we introduce a pre-trained language model and a time sequence encoder to mine the textual semantics and temporal periodic information from the paths. By fusing both pieces of information, accurate logical correlations can be captured to generate reliable logical rules. Furthermore, we design a learnable logical gated network to allow our framework to fuse additional useful information, such as frequency information of facts. To comprehensively evaluate the performance of our method, we propose three new datasets corresponding to transductive, inductive, and few-shot scenarios. The experimental results on six benchmark datasets prove the superiority of our method with significant improvements over state-of-the-art baselines. Our code and the proposed datasets are available at https://github.com/Sakura-del/FM-CTRL.