Enhancing Multi-Class Classification of Non-Functional Requirements Using a BERT-DBN Hybrid Model

Abstract

Background: Software requirements classification is essential to group Non-Functional Requirements (NFR) into several aspects, such as security, usability, performance, and operability. The main challenges in NFR classification are data limitations, text complexity, and high generalization needs. Objective: This research seeks to create a classification model using a hybrid of BERT and DBN, optimize hyperparameters, and improve data representation. Methods: A BERT and DBN-based approach is used, where DBN enhances BERT's ability to extract hierarchical features. Bayesian Optimization determines the optimal hyperparameters and data augmentation is applied to enrich the dataset variation. The model is tested on the PROMISE dataset consisting of 625 data. Results: The BERT-DBN model achieves 95% accuracy on the baseline configuration and 94% on the extensive configuration, better than the previous model, BERT-CNN. The model shows stability without any indication of overfitting. Conclusion: The combination of data augmentation, hyperparameter optimization, and DBN's ability to capture hierarchical patterns improves the accuracy of NFR classification, making it more effective than existing methods, and is expected to enhance text-based classification for software requirements.

Keywords

• Natural Language Processing
• BERT-DBN
• Augmentation Data
• Bayesian Optimization
• Non Functional Requirements