A Two-Phase NLP-Driven Deep Learning Framework for Automated Software Requirements Classification and User Story Generation in Agile Development

This research proposes a novel two-phase Natural Language Processing-driven deep learning framework for the automation of software requirement classification and the generation of actionable user stories from functional requirements. The framework tackles some of the critical challenges in the requirements engineering process, which would help minimize human errors, increase efficiency, and smooth the Agile Software Development Life Cycle (SDLC). It contains two core models: a classification model and a user story generation model. It relies on the PURE annotated dataset and implements three methods: FastText, Long Short-Term Memory (LSTM), and a BERT-based (Bidirectional Encoder Representations from Transformers) model. The fine-tuned BERT model performed best, with F1-scores of 0.92 for Functional Requirements (F) and 0.64 for Non-Functional Requirements (NFR), settling the effectiveness of the model on automated classification tasks. The user story generation model, built on top of the T5 (Text-to-Text Transfer Transformer), showed excellent results in transforming functional requirements into coherent and contextually aligned user stories. For instance, the functional requirement “The system must allow users to reset their password via email” was transformed into the user story with 89% alignment to the acceptance criteria. During training, the model showed a stable improvement in performance: training loss decreased from 10.12 to 0.22, and validation loss dropped from 0.68 to 0.21, which shows good learning and generalization capabilities. The results prove that the model is able to generate user stories with high relevance and accuracy; hence, it is strongly applicable to real-world software development processes. This research provides a more robust AI-driven solution to the classification of requirements and generation of user stories and contributes much to enhancing the automation and efficiency of software engineering practices.