A Conceptual Framework for the Generation of High‑Quality Software Requirements

Abstract

Requirements expressed in natural language are an essential artifact in the software development process, as all stakeholders can understand them. However, their inherent ambiguity remains a persistent challenge. To address this issue, organizations such as the Institute of Electrical and Electronics Engineers (IEEE) and the International Council on Systems Engineering (INCOSE) publish guidelines with rules that support the task of writing clearer requirements. Additionally, agile methodologies propose patterns and structures for formulating stakeholder needs in natural language, aiming to reduce ambiguity. Nevertheless, differences in stakeholders’ understanding of the requirements and how to express them correctly make the specification task even more difficult. Recently, large language models (LLMs) have emerged to enhance natural language processing tasks. These are deep learning architectures that emulate attention mechanisms similar to those used by humans. This work aims to assess the potential of LLMs in this domain. The goal is to use these models to improve the quality of software requirements, assisting analysts in the tasks of analysis and specification. The proposed framework, its architecture, key components, and their interactions are described. A conceptual test is also conducted to evaluate the framework's usefulness. Finally, the paper discusses the framework’s potential and limitations, as well as future directions for its validation and refinement.