Adaptive Data Augmentation Techniques for Software Requirements Classification Using Deep Learning

Software requirements classification is a critical yet time-consuming task in requirements engineering. Existing datasets suffer from severe class imbalance, degrading classifier performance on minority requirement types. This paper proposes an adaptive data augmentation framework that leverages TinyLlama 1.1B, a lightweight large language model, to generate synthetic requirements for underrepresented classes using few-shot prompting, balancing all 13 requirement types to a uniform target size. Three deep learning models are fine‑tuned and evaluated on the augmented dataset: BERT (bert‑base‑uncased), GPT‑2, and a novel hybrid RoBERTa + BiGRU architecture that combines RoBERTa contextual embeddings with a two‑layer bidirectional GRU and masked mean pooling for enhanced sequential representation. All models are assessed using accuracy, macro F1‑score, per‑class precision and recall, and confusion matrices. Explainability is provided through Layer Integrated Gradients for BERT and SHAP force plots for GPT‑2 and RoBERTa + BiGRU, making predictions interpretable. The best‑performing model is deployed as a local Flask web application enabling real‑time requirement classification. Experimental results demonstrate that adaptive augmentation significantly improves classification performance across all minority requirement categories.

Keywords: software requirements classification, data augmentation, TinyLlama, BERT, GPT‑2, RoBERTa, bidirectional GRU, natural language processing, deep learning, SHAP, explainability, requirements engineering