Exploring Knowledge from Knowledge Graphs and Large Language Models for Link Prediction

Knowledge Graphs, with their ability to represent and manage large volumes, provide high-quality structured knowledge for a variety of downstream tasks, but are often limited by incompleteness. An important application of knowledge inference technology is knowledge graph completion (KGC), which improves the completeness of the graph by filling in missing links. The main methods are embedding-based and language model-based. The language model-based approach uses pre-trained language understanding to predict unknown facts in the knowledge graph. Currently, large language models (LLMs) have proven their excellent performance on inference tasks. In order to tap the potential of LLMs in the field of knowledge graph completion, this paper will carry out the link prediction work of knowledge graph based on large language model. The specific work includes: (1) we effectively extract the neighbor-hood information of head entity h from knowledge graph, and determine its representation and order; (2) we design prompt templates to extract useful information from LLMs for link prediction tasks. This includes relevant semantic information about the head entity h, the relation r, and incomplete triple queries. (3) we propose a link prediction method for LLMs, which effectively combines the internal information of LLMs with the neighborhood information of knowledge graph. Moreover, our experiments show that the proposed method achieves state-of-the-art results on the Wikidata5M dataset. We also find that models with relatively small tweaks, such as Qwen1.5-7B, outperform the recent GPT-3.5.