Architecture of a System for Extracting Software Architecture from Requirements Specifications

Abstract

 The design of software architecture remains a critical and complex task, often requiring expert knowledge, significant time investment, and the ability to interpret evolving requirements. This paper presents Archescape, a prototype system that supports the automated generation of software architecture from structured requirements using a hybrid approach that integrates machine learning algorithms with generative language models, specifically ChatGPT. The proposed system guides users through an interactive interface to collect detailed project requirements and produces architectural suggestions based on both algorithmic analysis and natural language generation. The system aims to assist both experienced architects seeking alternative perspectives and less experienced developers in need of guidance. Experimental results show that combining machine learning with large language models yields more adaptive, context-aware, and user-friendly architectural solutions than using either technique in isolation. Additionally, the platform supports educational purposes by enhancing understanding of the link between requirements and architecture. Limitations and future improvements, including architecture validation and domain specific tuning, are discussed. The results demonstrate the potential of AI-assisted tools to streamline the architectural design process and improve communication between technical and non-technical stakeholders.