Utilizing Everything in History: Modeling Relation Inference Path and Entity Structure for Temporal Knowledge Graph Reasoning
Abstract: Temporal Knowledge Graph (TKG) extrapolation fundamentally involves selecting the correct answer from all entities based on historical information. Current methods can easily eliminate most incorrect answers, narrowing the candidate pool to a tiny area called the candidate zone. However, these methods often fail to find the correct answer within this zone, primarily because the entities within the candidate zone are similar in subgraph structure or relational connectivity, causing significant interference. These methods, which either model the graph structure of entities or the paths of relationships, can only address one type of similarity. To address this issue, we propose a model called the $\textbf{R}$elation Causal Logic $\textbf{I}$nference and $\textbf{E}$ntity $\textbf{S}$tructure Learning (RIES), which consists of two modules: relation inference and entity structure. These two modules model the causal logic of relations over time and the temporal evolution of entities' subgraph structure, respectively, allowing for the differentiation of candidates similar in subgraph structure and relational connectivity. When evaluated on five commonly used public datasets, the performance of RIES surpasses that of other state-of-the-art baselines.
Paper Type: Long
Research Area: NLP Applications
Research Area Keywords: knowledge graphs;reasoning
Contribution Types: NLP engineering experiment, Publicly available software and/or pre-trained models
Languages Studied: English
Submission Number: 1936
Loading