The Contribution of AI Techniques to Enhancing Software Requirements Quality: A Hybrid Explainable RAG Framework for Automated Software Requirements Quality Assessment

Abstract

Software requirements quality is a critical determinant of downstream project success, yet manual review of requirements specifications for ambiguity, incompleteness, inconsistency, and related defects remains labor-intensive and inconsistent across reviewers. This paper proposes a hybrid, explainable, retrieval-augmented framework that combines Sentence-BERT embeddings, FAISS-based retrieval over established requirements-engineering standards (ISO/IEC 29148, IEEE 830, SWEBOK, and a requirements-smells catalogue), a large language model for defect detection, natural-language explanation, and rewriting, and SHAP-based explainability for transparent quality scoring. We construct a Unified Requirements Quality Dataset (URQD) by merging and re-labelling three public sources - the PURE dataset, the PROMISE repository, and a requirements-smells dataset - and benchmark the proposed SBERT+XGBoost hybrid classifier against classical machine-learning baselines (SVM, Random Forest, XGBoost over TF-IDF features) and fine-tuned transformer models (BERT, RoBERTa, DeBERTa-v3). On a held-out test split, the proposed hybrid model achieves 97.31% accuracy and an F1-score of 0.9692, outperforming all TF-IDF-based baselines and improving Matthews Correlation Coefficient by 18.2 percentage points over the strongest baseline. Ten-fold stratified cross-validation confirms the improvement (mean F1 = 0.9773), and a Friedman test indicates the differences among models are statistically significant (chi-square = 25.56, p = 1.18e-05). Fine-tuned transformer models, evaluated separately, achieve the highest raw classification scores (RoBERTa: 99.82% accuracy), at substantially higher computational and data-labelling cost, motivating the retrieval-augmented hybrid design as a lighter-weight, more explainable alternative suited to practical deployment. We discuss the trade-offs among accuracy, explainability, and cost, and outline threats to validity and directions for future work.