Integrating Relation Dependences and Textual Semantics for Coherent Logical Reasoning over Temporal Knowledge Graph

Qing Li, Guanzhong Wu, kaiwen wei

Keywords: Temporal knowledge graph, Knowledge graph, Multi-hop logical rules, Link forecasting, Inductive reasoning

Abstract:

Temporal knowledge graphs (TKGs) reflect the evolution patterns of facts, which can be summarized as logical rules and applied to forecast future facts. However, existing logical reasoning methods on TKGs face two limitations: 1) A lack of efficient strategies for extracting logical paths. 2) Insufficient utilization of structural and textual information. To bridge these gaps, we propose CoLR, a two-stage framework that mines relation dependencies and textual semantics for Coherent Logical Reasoning over TKGs. In the first stage, we construct a temporal relation structure graph (TRSG) composed of relations and cohesion weights between them. Besides, we define a novel time-fusion search graph (TFSG) along with TRSG to facilitate efficient and reliable temporal path searching. In the second stage,
the textual content and timestamp sequences from these paths undergo encoding via a pre-trained language model and a time sequence encoder to accurately capture potential logical rules. Additionally, for quadruplets missing paths, historical edges sampled based on relation cohesion are used as supplements. Given the limitations of existing benchmark datasets in evaluating accuracy, generalization, and robustness, we construct three new datasets tailored to transductive, inductive, and few-shot scenarios, respectively. These datasets, combined with four real-world datasets, are employed to evaluate our model comprehensively. Experimental results demonstrate that our approach significantly outperforms existing methods across all three scenarios. Our code is available at https://anonymous.4open.science/r/CoLR-0839

Primary Area: learning on graphs and other geometries & topologies

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3651