Defect Correction Method for Software Requirements Text Using Large Language Models

Text defect correction holds a pivotal role within the domain of text processing. In the face of the rapid advancement of Natural Language Processing (NLP) technologies, significant challenges persist in achieving precision and efficiency, particularly within specialized domains such as software requirements. For instance, traditional methods require labor-intensive verification, and the generative Large Language Models(LLMs) paradigm suffers from illusions and knowledge deficits. So for professional software requirement text and defect correction processing, we construct a Chinese requirement text defect corpus containing 215 Chinese requirement texts and 10,049 manually labeled sentences and propose a contextual prompt pattern based on LLMs, which mainly includes the following steps: (1) Requirements text defect detection classification using fine-tuned BERT. (2) Extracting primary key information from requirement text using fine-tuned UIE. (3) The above results are embedded as contextual information in extensible prompt templates, respectively, and defects are corrected by interacting with generative LLMs. The experimental results show that by utilizing prompt learning and in-context learning techniques for requirement text defect correction, we achieved an outstanding improvement (30.3%) in defect correction effectiveness.