This study evaluates whether text augmentation improves classification of internal quality-audit findings into ISO 9001:2015 clause categories under limited and imbalanced labeled data. The dataset comprised 191 cleaned findings across seven classes. A domain-aware NLP pipeline was combined with TF-IDF features and three classical classifiers: Logistic Regression, Support Vector Machine (SVM), and Random Forest. Four data settings were compared: no augmentation, SynonymAug, BackTranslation, and ContextualAug. For augmented settings, only training folds were expanded to 2x, 3x, and 5x; test folds were left unchanged. Performance was evaluated with repeated stratified group 5-fold cross-validation using Accuracy, Macro Precision, Macro Recall, and Macro F1. All augmentation methods improved on the no-augmentation baseline, although gains were modest. The strongest result within this experiment was BackTranslation + SVM at 5x, which achieved Accuracy = 0.742, Macro Precision = 0.719, Macro Recall = 0.679, and Macro F1 = 0.680. The findings provide comparative evidence that controlled augmentation can improve clause-level classification of short audit text under constrained-data conditions, but the benefit depends on both augmentation strategy and classifier. The study also outlines a practical dashboard use case for audit triage and hotspot monitoring.

Keywords

internal quality audit, text augmentation, ISO 9001:2015, text classification, machine learning, audit findings