Research on Aerospace Intelligent Question Answering Model Based on Large Model Technology

With the rapid development of aerospace technology in our country, news reports in related fields are increasingly common. However, the aerospace domain is characterized by vast, complex, and highly specialized data, making it difficult for the general public and researchers to obtain the information they need in a short time. To help people more conveniently acquire and understand knowledge in the aerospace field, this paper constructs a high-precision dialogue model by integrating innovative dataset construction methods, model fine-tuning techniques, and retrieval-augmented generation techniques. To address the challenges of manually constructing datasets, the study employed chain-of-thought prompting techniques combined with the GLM-4 large language model to generate high-quality question-and-answer datasets from raw data. To improve the accuracy of the model in text generation tasks in the aerospace field, the LoRA fine-tuning technique was applied to fine-tune and compare the ChatGLM3-6B and ChatGLM2-6B large language models. To resolve the “hallucination” issue that may occur in complex dialogues after model fine-tuning, retrieval-augmented generation techniques were applied to establish a dedicated local knowledge base and optimize retrieval strategies, ultimately increasing the accuracy of the question-and-answer model to 89%. Experimental results indicate that the proposed construction scheme significantly enhances the accuracy and practicality of the aerospace domain question-and-answer model. Through automated dataset construction, domain-specific model fine-tuning, and retrieval augmentation techniques, an efficient and accurate aerospace information dialogue model has been successfully built.