Lightweight lexical augmentation for robust transformer-based student feedback classification

Transformer-based models such as RoBERTa achieve state-of-the-art performance in text classification, yet their effectiveness is constrained in educational settings where annotated student feedback is limited, heterogeneous, and institution-specific. This study examines the extent to which a lightweight lexical augmentation strategy can enhance model robustness under such low-resource conditions. A minimal reversible perturbation, character-level inversion of a single randomly selected word per sentence is applied to expand a student feedback dataset without relying on external corpora or generative models. RoBERTa is fine-tuned under two conditions: using real data alone and using a combination of real and synthetically augmented samples. Empirical evaluation demonstrates statistically significant improvements with augmentation. Classification accuracy increases by 3.2%, the F1-score improves by 0.032, and final training loss decreases by approximately 32% (p < 0.05 via paired bootstrap resampling). Confusion matrix analysis further reveals reductions in both false positives (–32.5%) and false negatives (–10.3%), indicating improved sensitivity and specificity in capturing subtle sentiment cues common in educational feedback. These findings show that controlled lexical perturbation despite its simplicity can regularize transformer models effectively, improving generalization without adding computational burden or semantic drift. The approach provides a practical and reproducible pathway for institutions seeking scalable solutions for automated feedback analysis, adaptive learning analytics, and student support systems in data-constrained environments.