Deep Learning Methods for Software Requirement Classification: A Performance Study on the Pure Dataset

Requirement Engineering (RE) is a fundamental phase in software development, involving the classification of software requirements into Functional Requirements (FRs) and Non-Functional Requirements (NFRs). FRs define the system’s core functionalities, whereas NFRs address quality attributes such as performance, security, and usability. Accurate classification of requirements is essential for ensuring software meets user needs, yet manual classification is labor-intensive and prone to errors. This challenge highlights the necessity of automated approaches. Traditional Machine Learning (ML) methods rely on handcrafted features, limiting their ability to understand contextual relationships within textual requirements fully. In contrast, Deep Learning (DL) techniques, particularly transformer-based models, can effectively capture complex dependencies in natural language. This study investigates the DL-based classification of software requirements using the PURE repository, constructing a dataset of 4656 requirements (2550 FRs and 2106 NFRs). We evaluated eight DL models using accuracy, precision, recall, and F1-score, with Bidirectional Encoder Representations from Transformers (BERTs) achieving the best performance (F1-score: 88.71%, Accuracy: 89.74%), surpassing traditional ML and other DL models. While BERT achieved better accuracy, its computational demands, increased memory usage, and longer inference time reduce its practicality in resource-constrained environments. Traditional DL models provide a good balance between accuracy and resource efficiency, while ML models, though less accurate, are ideal for lightweight applications. Additionally, ablation studies revealed that preprocessing techniques such as stopword removal and lemmatization resulted in minor, statistically insignificant improvements.