Accurate classification of software requirements, distinguishing between functional and non-functional aspects, is crucial for developing reliable and efficient software systems. However, existing methods often struggle with insufficient semantic understanding and managing diverse software requirements. In this study, we introduce an innovative framework named EnsCL-CatBoost (Ensembled Contrastive Learning with CatBoost) to address these challenges by enhancing classification accuracy and robustness. Our method uses a weighted ensemble of Doc2Vec, Word2Vec, and FastText, leveraging their strengths for richer semantic representation. Unlike conventional strategies, we incorporate contrastive learning with the InfoNCE loss function, boosting discriminative power by clustering similar samples and distancing dissimilar ones, thus enhancing robustness and model generalization. To tackle class imbalance, we integrate SMOTE-Tomek links into the embedding process, achieving balanced class distribution before classification. Additionally, our evaluation extends beyond test datasets to new, unlabeled datasets, demonstrating practical applicability in real-world scenarios. We compare our framework’s performance with five machine learning classifiers which we trained using traditional embedding techniques. The results show that our method significantly outperforms others, achieving 94% accuracy. Our research transparently presents tool-based results, highlighting the transformative potential of automation in software requirement classification and setting a new benchmark for practical deployment in diverse environments, paving the way for future research in the field.

The software development process mostly depends on accurately identifying both essential and optional features. Initially, user needs are typically expressed in free-form language, requiring significant time and human resources to translate these into clear functional and non-functional requirements. To address this challenge, various machine learning (ML) methods have been explored to automate the understanding of these requirements, aiming to reduce time and human effort. However, existing techniques often struggle with complex instructions and large-scale projects. In our study, we introduce an innovative approach known as the Functional and Non-functional Requirements Classifier (FNRC). By combining the traditional random forest algorithm with the Accuracy Sliding Window (ASW) technique, we develop optimal sub-ensembles that surpass the initial classifier’s accuracy while using fewer trees. Experimental results demonstrate that our FNRC methodology performs robustly across different datasets, achieving a balanced Precision of 75% on the PROMISE dataset and an impressive Recall of 85% on the CCHIT dataset. Both datasets consistently maintain an F-measure around 64%, highlighting FNRC’s ability to effectively balance precision and recall in diverse scenarios. These findings contribute to more accurate and efficient software development processes, increasing the probability of achieving successful project outcomes.

In the domain of software engineering, the accurate and effective classification of requirements is of paramount importance. Proper classifications of these requirements enable developers to create robust and error-free software solution. Traditional methods of user requirements classification face the issue of the reliance on manual processes, which are time-consuming, labor-intensive, and prone to human error. The limitations of traditional methods underscore the need for more automated, scalable, and robust approaches to user requirements classification in order to meet the demands of modern software development practices. To improve the classification process, we employed a Deep Learning (DL) methodology termed the Bidirectional Long Short-Term Memory (Bi-LSTM) model to conduct feature extraction, after which we merged these feature vectors into ML classifiers. Our research methodology is structured around a five-step process. Initially, the textual input is tokenized and converted to lowercase. Subsequently, we eliminate all punctuation. The pre-processed text is then subjected to a Bi-LSTM (Bidirectional Long Short-Term Memory) model for feature vector extraction. After that, this feature vector is fed into different classifiers such as Medium KNN, Cubic KNN, Linear SVM, Quadratic SVM, and Cubic SVM and obtained an accuracy of around 99.60% to 99.80% on a publicly available dataset of requirements.