Extração de requisitos de software a partir de comentários de aplicativos por grandes modelos de linguagem e orquestração de agentes

Abstract

The study investigates the extraction of software requirements using LLM-based resources from user comments published in application store platforms, considering the classical challenges of Requirements Engineering, such as ambiguity, duplicated text and linguistic variation. It is well known that modern computational systems have a global reach, which magnifies cultural and semantic differences, compromising requirement quality and increasing costs, rework, and project risks. In this context, the study addresses four main questions: the feasibility of extracting requirements directly from user comments; the possibility of automating this process through software agents; the construction of an agent workflow to support Requirements Engineering; and the quality of the extracted requirements. To address these questions, our work proposes the integrated use of Natural Language Processing models, Large Language Models (LLMs), and CrewAI agent-orchestration techniques, assessing their ability to identify, structure, and validate functional and non-functional requirements. Finally, we highlight that recent advances in AI offer promising avenues for reducing manual effort, mitigating ambiguities, and enhancing the systematisation of Requirements Engineering.