Automated User Story Generation in Requirements Elicitation using Fine-Tuned Large Language Models

Abstract

In software development, generating user stories from requirements elicitation interviews is a critical yet time-consuming and subjective task. Traditional methods often rely heavily on human interpretation, which can introduce bias and limit efficiency. This thesis explores the potential of Fine-tuned Large Language Models (LLMs) to automate and enhance this process. To address these challenges, LLMs were fine-tuned using supervised learning on real interview transcripts and further optimized through prompt engineering to generate well-structured user stories. The approach aimed to reduce manual effort while maintaining the accuracy and completeness of the generated user stories. A comprehensive evaluation was conducted using both quantitative and qualitative methods. Quantitative metrics included time taken for generation, carbon emissions estimation, readability scores, semantic similarity, and lexical diversity. Additionally, domain experts assessed the accuracy, completeness, and relevance of the generated user stories through structured questionnaires. Findings show that fine-tuned LLMs can produce user stories that are comparable in quality to those created by human experts, while significantly improving efficiency and consistency. The results demonstrate the practical value of LLMs in supporting requirements engineering tasks and reducing the cognitive load on analysts.