LERE: A Method Based on LLMS and Enhanced Ears in Requirements Engineering

Normalized requirements permeate the entire requirements engineering and serve as a crucial foundation for the project. To improve the accuracy of requirements normalization in Chinese, in this study, we propose a method: LERE, which is based on LLMs and an enhanced EARS paradigm. The enhanced EARS is an optimized version of the original English EARS, incorporating elements from Chinese linguistic structures. This method employs a two-step processing prompt, with each step utilizing the enhanced EARS to perform structural element analysis on the input. The first step involves preprocessing the input, while the second step normalizes the output from the preprocessing stage by refining its syntactic structure and semantic expression. We have constructed a dataset consisting of 138 Chinese requirements, categorized into eight types of errors. The dataset and prompts have been made publicly available. Testing results on this dataset using Qwen2.5 and ChatGPT-4o show that the performance of LERE improved by approximately $\mathbf{2 0 \%}$ compared to baseline.