Explainable attention based BiLSTM-SVM for Software Requirement Classification: Integrating Generative Artificial Intelligence

The human classification of software requirements, a difficult subjective process prone to inconsistencies and classification errors, is a major difficulty in software requirement specification (SRS). With an increasing number of requirements, manual classification becomes unfeasible. To address this issue, this work proposes a customized attention layer based BiLSTM-SVM hybrid architecture to implement automation in the categorization of software requirement. The model integrates a pre-trained GPT-Neo model to overcome the limitation of small real-world dataset by augmenting the data. The augmented dataset expands to 6,783 functional and non-functional requirements, seven times larger than the Promise_exp dataset. Experimental results on the augmented dataset are promising, achieving 97% accuracy on the testing set—significantly outperforming the other datasets. These findings demonstrate the effectiveness of data augmentation using the GPT-Neo model. GPT-Neo also enables the generation of balanced datasets, ensuring equal representation of FR and NFR. Additionally, LIME is implemented to ensure transparency in the model’s decision-making process. This approach allows future researchers to experiment with more complex architectures, such as transformers, without concerns of data scarcity or overfitting.