Modeling dual temporal dependencies for temporal knowledge graph reasoning on Chinese cultural heritage

The digitization of Chinese Cultural Heritage (CCH) facilitates the study of its long-term evolution, driving growing interest in Temporal Knowledge Graph Reasoning (TKGR). However, modeling the complex temporal dependencies remains a major challenge, particularly in long-cycle domains with rich contextual semantics like CCH. To this end, we propose the D ual T emporal D ependency M odel (DTDM), a novel TKGR framework that decouples and jointly models two temporal dependencies: (1) a temporal cross-relation evolution module, which uses attention mechanisms to capture dynamic interactions among relations; and (2) a temporal convolution module, which combines gated recurrent unit to capture the long-range evolution and parallel convolutions to enhance entity representations. We constructed CCH-Ceramics, the first TKG dataset focused on Chinese ceramics. Experimental results on three benchmarks demonstrate that DTDM significantly outperforms all baselines, achieving a 4.13% absolute improvement in performance on entity prediction on CCH-Ceramics.