Enhancing Zero-Shot Knowledge Graph Relation Prediction through Large Language Models and Contrastive Learning

Knowledge graph completion (KGC) aims to predict missing facts in a knowledge graph (KG), with zero-shot relation prediction remaining a significant challenge due to limited descriptive information. In this paper, we propose ROCKGC, a framework that leverages large language models (LLMs) to generate enriched relation descriptions and integrates contrastive learning with a conditional variational autoencoder (C-VAE) to enhance relational separation in the embedding space. By combining LLM-generated semantic representations with C-VAE and contrastive loss, ROCKGC effectively captures nuanced relational differences and improves zero-shot learning performance. Extensive experiments on benchmark datasets show that our method outperforms state-of-the-art approaches, achieving superior accuracy and robustness in predicting unseen relations.