Knowledge Graph Completion using RAG and Improved Structural Information

Semantic and structural information are essential in Knowledge graph completion. We propose KRIS (Knowledge Graph Completion using Retrieval-Augmented Generation and Improved Structural Information), a method based on Retrieval-Augmented Generation (RAG) with enhanced structural information. KRIS uses semantic embeddings and RAG to enrich triple semantic information, reducing the hallucination problem, especially for domain-specific knowledge graphs. To address the weakness of semantic-embedding-based methods in leveraging structural information, we conduct joint training via supervised contrastive learning and structural information learning. Our model effectively utilizes both types of information. Experiments on public datasets show that KRIS significantly outperforms baseline models in prediction accuracy.