Design of Question-Answering and Reasoning System Combining Large Language Models and Knowledge Graphs

Problem reasoning uses natural language processing technology to analyze natural language questions input by users, and finally generate accurate and suitable answers. At present, the performance of Q&A model has been significantly improved, and the deep semantic understanding of text can be obtained through accurate knowledge reasoning. Based on LLaMa model, this paper systematically studies Transformer architecture, normalization technology, rotary position coding and packet query attention, creatively studies the combination of LLM and KG, and constructs a mathematical model of instruction fine-tuning algorithm. It not only understands the surface meaning of text, but also uses background knowledge such as entity relationships in KG to execute instructions more accurately and intelligently. The LLM studied in this paper is combined with KG to build a question-and-answer reasoning system, which can overcome the "illusion" problem of large models. KG significantly improves the accuracy by constrains the generated results of large models with structured knowledge.