Deep learning for non-functional requirements: a Convolutional Neural Network approach

The Requirements Engineering (ER) phase plays a critical role in software development, as any shortcomings during this stage can lead to project failure. Analysts rely on Requirements Specification (RS) to define a comprehensive list of quality requirements. The process of requirements classification, within RS, involves assigning each requirement to its respective class, presenting analysts with the challenge of accurate categorization. This research focuses on enhancing the classification of non-functional requirements (NFR) using a Convolutional Neural Network (CNN). The study also emphasizes the significance of preprocessing techniques, the implementation of sampling strategies, and the incorporation of pre-trained word embeddings such as Fasttext, Glove, and Word2vec. Evaluation of the proposed approach is performed using metrics like Recall, Precision, and F1, resulting in an average performance improvement of up to 30% compared to related work. Additionally, the model is assessed concerning its utilization of pre-trained word embeddings through ANOVA analysis, providing valuable insights into its effectiveness. This study aims to demonstrate the utility of CNNs and pre-trained word embeddings in the classification of NFRs, offering valuable contributions to the field of Requirements Engineering and enhancing the overall software development process.