MCD-ARC: Multi-Modal Contrastive Domain-Adaptive Requirements Classification for Software Engineering

Abstract

Software requirements classification into Functional Requirements (FR) and Non-Functional Requirements (NFR) is a foundational task in automated software engineering, serving as a prerequisite for estimation, architectural design, and quality assurance. Despite the proliferation of machine learning approaches, current methods face three critical limitations: (1) a reliance on single-modal textual features that ignore the rich structural and semantic context of requirements; (2) poor generalization across diverse application domains (e.g., Healthcare vs. Finance) due to distributional shifts; and (3) performance degradation caused by severe class imbalance in real-world datasets. This paper presents MCD-ARC (Multi-Modal Contrastive Domain-Adaptive Requirements Classification), a comprehensive deep learning framework designed to overcome these challenges. MCD-ARC integrates a novel Multi-Modal Fusion Layer combining BERT-based textual embeddings with dependency-graph structural features and statistical complexity metrics. To ensure robustness, we introduce a customized Contrastive Self-Supervised Learning objective that maximizes the mutual information between semantically similar requirements while pushing apart dissimilar ones, leveraging a curriculum-based data augmentation strategy. Furthermore, we employ Adversarial Domain Adaptation via a Gradient Reversal Layer (GRL) to learn domain-invariant feature representations, allowing the model to generalize effectively to unseen projects. Extensive experiments on a large-scale dataset of 6,086 requirements across five distinct domains demonstrate that MCD-ARC achieves state-of-the-art performance, recording 95.8% Accuracy and 95.4% F1-Score. Ablation studies confirm the significant contribution of each component, with multi-modal fusion and domain adaptation providing gains of 3.6% and 2.6% respectively over strong baselines.