大语言模型提示词迭代的低成本查询词理解方法OA

To address the high cost and development barriers associated with query term understanding in natural language un-derstanding tasks,this paper proposes a low-cost query term understanding method based on iterative prompting of Large Lan-guage Models(LLMs).Aiming at the query classification task in search engines,the method uses LLMs to construct initial prompt words and guide the large model to optimize the prompt words through multiple rounds,achieving automated high-quality annotation of large-scale training data.Then,through knowledge distillation,the capabilities are transferred to small models to meet the application requirements of real-time performance and resource-constrained scenarios.The experimental results show that,under the optimal prompt ensemble,the LLMs achieves a precision of 97.5％and an accuracy of 93.0％in query classifica-tion.This method significantly reduces reliance on manual annotation,cutting the labeling cost for 70000 samples from 21000 RMB(manual labeling)to only 30.24 RMB(LLMs-based labeling),greatly improving both data annotation efficiency and model development efficiency.Further,through knowledge distillation,the capabilities of large models are transferred to small models,maintaining high classification performance while significantly enhancing the deployment flexibility of the models.Therefore,this method provides a feasible,economical and efficient technical path for text classification tasks in low-resource environments.