Improving Temporal Knowledge Graph Reasoning with Hierarchical Semantic-Aware Contrastive Learning

Abstract

Temporal Knowledge Graph (TKG) reasoning seeks to infer the future evolution of incomplete facts from observed historical data. Although supervised contrastive learning has recently enhanced query representations for TKG reasoning, two critical challenges remain. First, current methods uniformly treat all negative samples, overlooking their semantic and temporal correlations. Second, these approaches do not fully exploit the hierarchical relationships between fine-grained events and higher-level event categories, thereby missing crucial event taxonomies. To address these limitations, we propose a Hierarchical Semantic-aware Contrastive Learning (HSCL) framework. Specifically, our Instance-level objective introduces a dynamic adaptive weighting mechanism that differentiates negative samples based on semantic similarity, while our Category-level objective incorporates ontology-guided clustering to represent hierarchical event semantics. This dual-level design encourages cohesive embeddings within the same event category and clear separation across different categories. Extensive experiments on four real-world benchmarks demonstrate that HSCL consistently outperforms state-of-the-art baselines( $^1$ 1 The code is available at https://github.com/AONE-NLP/TKGR-HSCL .