Cyberbullying is a growing concern in digital communication, particularly in low-resource languages such as Bangla, where the lack of labeled data poses a challenge for automated detection systems. This study evaluates how data augmentation improves Bangla cyberbullying detection by addressing class imbalance in the Bangla Online Comments dataset (44,001 samples). We developed an augmented dataset by applying synonym replacement, random swap, back-translation, and paraphrasing techniques, expanding the dataset with 6,081 additional samples. A comparative evaluation is conducted using traditional machine learning models (Support Vector Machine (SVM), Random Forest, eXtreme Gradient Boosting (XGBoost)), deep learning architectures (Bidirectional Long Short-Term Memory (BiLSTM)), and transformer-based models (BanglaBERT). The experimental setup considers both binary and multiclass classification performance. Experimental results show that data augmentation improved overall multiclass accuracy by up to 2%, while substantially enhancing F1 scores for minority classes such as threat (21.88% to 69.19%) and religious (71.65% to 73.83%). Among all models, BiLSTM achieved the highest performance, reaching 80.39% accuracy in multiclass and 86.65% in binary classification. While overall accuracy gains are modest, significant F1-score improvements for underrepresented classes demonstrate augmentation’s efficacy in addressing imbalance. This study contributes to the growing field of NLP for low-resource languages by providing insight into augmentation techniques for text classification. Additionally, we identified certain limitations in existing Bangla data augmentation techniques, which may affect the quality and diversity of generated samples.
