Explainable Cyberbullying Detection for Multilingual Online Platforms Using NLP and Augmentation-Enhanced Deep Learning

The growing prevalence of online abuse has increased the need for automatic cyberbullying and hate speech detection, particularly in under-resourced and morphologically complex languages with limited annotated data and language resources. This study presents a multilingual English–Lithuanian framework combining machine learning, deep learning, transformer models, three data augmentation techniques, and explainable AI to address class imbalance, limited data, and model transparency. Overall, transformer-based approaches outperformed traditional methods. For the English dataset, averaged results over five runs with different random seeds showed model-dependent performance. Syn-Aug achieved the best F1-scores for LR (0.8791 ± 0.0029), SVM (0.8810 ± 0.0028), and RoBERTa (0.9287 ± 0.0023), while contextual BERT augmentation achieved the best F1-scores for CNN (0.9124 ± 0.0020) and BERT (0.9274 ± 0.0021). The highest English performance was obtained with the soft-voting BERT-RoBERTa ensemble, with an F1-score of 0.9356 ± 0.0035. For Lithuanian hate speech classification, hold-out validation and five-fold cross-validation were employed and evaluated on the same independent test set. For the five-fold cross-validation, XLM-RoBERTa and Syn-Aug achieved the best accuracy (0.9458) and F1-score (0.9363), followed by XLM-RoBERTa with XLM-Aug (accuracy 0.9072; F1-score 0.9336). Under the hold-out protocol, XLM-RoBERTa with XLM-Aug achieved the best accuracy (0.824) and F1-score (0.8109). These results show that Lithuanian classification performance varied with augmentation and evaluation protocol. Cross-language transfer using semantically aligned binary labels further highlighted the importance of language-specific data. Overall, the proposed approach supports efficient, interpretable, and reliable multilingual cyberbullying detection.