Pathtune: Knowledge Graph Completion Based on Multi-Hop Path Fine-Tuning

Knowledge Graph (KG) as a structured knowledge representation often suffers from problems such as missing knowledge and knowledge errors, and needs to improve the completeness and accuracy of knowledge through Knowledge Graph Completion (KGC). In recent years, the Large Language Model (LLM) with strong semantic comprehension capability has been applied to the field of KGC to mine the semantic information implicit in text message for triples, but the structural information around the triple is not sufficiently learned, which affects the performance of triple prediction. In this regard, this paper proposes PathTune, a fine-tuning framework for the LLM based on multi-hop path instructions. By extracting the multi-hop paths in the subgraphs around the triple and transforming them into natural language text to design the multi-hop path instructions, the LLM is then guided to learn the semantic patterns of the graph structure. The results of the triple classification comparison experiments demonstrate that PathTune achieves state-of-the-art performance on the FB15K-237N dataset, with an accuracy increase of 1.79% and 0.74% in F1-score compared to the KG-LLaMA method that uses triples for Instruction Tuning (IT). The multi-hop paths in the subgraphs are helpful for reasoning about the LLM, and we find that the triple classification works best when the number of path entries is set to 2.