Two-Stage Prompt-Based Entity Representation in Contrastive Learning for Knowledge Graph Completion

The task of Knowledge Graph Completion (KGC) focuses on inferring missing entities automatically. Existing methodologies mainly fall into two categories: structure-based and description-based approaches. Recently, contrastive learning has shown great advantage in description-based KGC. However, this approach necessitates large batch sizes for negative sampling, yet increasing batch size risks introducing false negatives. Furthermore, many contrastive-based methods underutilize the rich semantic content in textual data. To address these issues, leveraging recent research in sentence representation learning and Chain of thought (CoT) in prompt engineering, we proposes a novel description-based KGC method with two-stage prompt-based template. Simultaneously, we proposed a contrastive learning model architecture with <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{infoNCE} -\gamma$</tex> loss. Extensive experiments on three KGC benchmarks validate the superior performance of our approach in KGC.