DDPC: Dual Dynamic Presentation with Contrastive Learning for Robust Temporal Knowledge Graph Completion
Abstract: Temporal knowledge graph completion has made significant progress, but several research gaps persist. This study addresses the challenges of temporal changes by proposing DDP and DDPC, novel dual-perspective learning frameworks that integrate static and temporal knowledge using a dual-layer embedding mechanism and a contrastive learning-enhanced version, respectively. This approach effectively captures both dynamic changes and time-invariant properties of entities and relations, optimizing the completeness and accuracy of information. Additionally, a perturbation learning mechanism is introduced to enhance the model's robustness to anomalous data and noise by simulating data perturbations during training, improving adaptability and stability in changing environments. DDPC achieves state-of-the-art results on multiple standard evaluation datasets, experimentally verifying the effectiveness of the proposed theories and methods. This study contributes to advancing the field of temporal knowledge graph completion by developing an innovative framework that integrates temporal and static perspectives, enhances robustness, and undergoes rigorous evaluations.
Paper Type: Long
Research Area: Information Extraction
Research Area Keywords: knowledge base construction, graph-based methods, knowledge-augmented methods, contrastive learning, representation learning, word embeddings
Contribution Types: Model analysis & interpretability, Approaches to low-resource settings, Theory
Languages Studied: English
Submission Number: 4831
Loading