Graph-LLM fusion: enhancing fact representation and logical reasoning in artificial intelligence systems

Abstract

Knowledge graphs organize and represent entity relationships through graph structures, providing a foundation for machine understanding and reasoning, but their reasoning capabilities are limited by coverage and manual rules. Large language models demonstrate strong semantic understanding and generation abilities but lack effective utilization of symbolic knowledge and interpretability. To combine the strengths of both technologies, academic and industrial communities have devoted significant effort in recent years to exploring the integration of knowledge graphs and large language models, aiming to build more powerful and interpretable AI systems. Firstly, this paper reviews the current state of research on the fusion of knowledge graphs and large language models, with a focus on the key achievements in enhancing fact representation and logical reasoning. These achievements include pre-trained language models based on knowledge graphs, knowledge graph representation learning based on large language models, and reasoning methodsthat leverage the fusion of the two approaches. Furthermore, the paper outlines the mainstream technical approaches and application scenarios of graph-model integration in the industry. Finally, future development directions of graph-model intgeration are discussed, and it is posited that the integration of these two technologies represents a crucial trend in the advancement of artificial intelligence.

Keywords

• large language model
• graph neural network
• machine reasoning
• pre-training
• attention
• graph representation learning