MVG-DMA: A multi-view graph with dynamic multi-head attention-based transformer encoder for software requirements classification

Software requirements classification is most important for software quality assurance. Existing graph-based methods fail to capture the syntactic, semantic, and contextual depth of requirement sentences. While recent approaches leverage Graph Neural Networks (GNNs), they suffer from three limitations: rigid grammar-based edge weights, term-frequency representations that ignore context, and a single-view bias that considers only one type of linguistic relationship. To overcome these challenges, we propose MVG-DMA, a novel multi-view graph framework that introduces three paradigm shifts: (1) from static dependency parsing to dynamic graph construction module that builds a dependency graph with neural dependency weight prediction for edge weights, while simultaneously constructs context aware semantic graphs with syntactic neighborhoods and positional relationships, replacing shallow term statistic. Both the dependency and semantic graphs are constructed through parallel batch processing. (2) The framework further integrates Principal Component Analysis (PCA) based dimensionality reduction for sentence transformers to maximize semantic retention while reducing computational complexity. (3) From single view to multi-view synthesis, our system utilizes a view-specific DCMHA Transformer Encoder, which enables adaptive feature-level fusion through residual connections between transformer-encoded and graph-structured representations. (4) The view-level fusion and classification module combines the vectorized weighted pooling with attention-based scoring to generate semantically dense sentence representations, which are then processed through view fusion and a classifier to predict requirement categories. Experimental evaluations on benchmark datasets PROMISE, NFR-SO, and NFR-Review show that the proposed system reduces the gap between sequential and graph-based NLP. It provides a reliable and scalable solution for requirements classification with broader implications for legal document classification, medical text mining, and other specialised NLP tasks.