ColKGC: Collaborative Enhancement using Large Language Model for Knowledge Graph Completion

Knowledge Graph Completion (KGC) aims to address incomplete facts in Knowledge Graphs (KGs). As a mainstream approach for KGC, text-based methods are limited in performance due to insufficient knowledge caused by constraints in fine-tuning data. To incorporate broader external knowledge, some researchers explored KGC approaches based on large language models (LLMs), but general-purpose LLMs lack direct awareness of knowledge graph domain knowledge, making it challenging for them to independently perform KGC tasks. To balance the dependence of text-based methods on knowledge graph domain knowledge and LLMs on external knowledge, we propose Collaborative Enhancement using Large Language Model for KGC (ColKGC). We define the prompt template to implement the interaction between the text-based model and the LLMs, and divide the knowledge graph completion process into two stages. In the concept enhancement stage, we supplement the fine-tuning data by interacting with the prompts of the LLMs to enrich the graph domain knowledge learned by the text-based model. In completion inference stage, we adopted the iterative interaction strategy, taking the text-based model output as the prompt and taking the interaction content of the previous stage as the context to enhance the inference, which improved the LLMs' understanding of the KGC task. Our experiments demonstrate that ColKGC achieves superior results across various standard knowledge graph benchmarks. Extensive experiments also bear out ColKGC's efectiveness in generated data and the iterative interaction framework in assisting LLMs reasoning.