<![CDATA[Few-shot text classification is a challenging problem in natural language processing. Models have to generalize from a small number of labeled examples. This paper investigates the effectiveness of data augmentation methods (back-translation, paraphrasing, and noise injection) on overfitting prevention and generalization enhancement in few-shot scenario for small transformer models such as DistilBERT and DistilRoBERTa. We perform experiments on two standard benchmark datasets, DBPedia-14 and BBC News, in the 3-shot, 5-shot, and 7-shot settings. The result shows that the data augmentation substantially improves the performance of the classification. With back-translation, DistilBERT achieves 0.98 and 0.96 accuracy on DBPedia-14 and BBC News in 7-shot setting, compared to 0.88 and 0.86 accuracy without any augmentation. These results demonstrate that carefully selected augmentations can bridge the performance gap between few-shot and fully supervised learning, enabling competitive results on resource constrained hardware without the need for massive, labeled datasets.]]>
Copyrights © 2026
