Optimising Requirements Elicitation: An Iterative Rule-Based and LLM-Driven Approach

Natural Language (NL) serves as the primary medium of communication for eliciting requirements from stakeholders throughout the system development lifecycle. However, the inherent freedom and flexibility of NL risk introducing ambiguities or omissions, leading to a commensurate increase in development costs. To address these challenges, we propose a tool designed to automate the interviews conducted during requirements elicitation. Specifically, we present an application wherein a Large Language Model (LLM) executes targeted queries and proposes revisions to requirement formulations to ensure adherence to established quality rules derived from Requirements Engineering (RE) literature. Our contribution consists in the application of diverse prompt engineering techniques — such as few-shot prompting — to systematise requirement quality assessment, mitigate LLM hallucinations, and structure the interview process. This approach enhances the likelihood of converging upon a final formulation that preserves the user's intent while ensuring adherence to an optimal format. While this paper presents two running examples as a proof of concept for the effectiveness of the proposed approach, we intend to undertake more systematic and quantitative evaluations in future work.