International Journal of Science, Engineering and TechnologyAuthors: Sakshi Dinesh Pandit, V. D. Dabhade
Abstract: The rapid growth in online hostile behavior, in- cluding hate speech and cyberbullying, has created significant challenges for users and platform moderators. While early detection methods relied on traditional machine learning ap- proaches like SVM, Naïve Bayes, and Random Forests using hand-crafted features [3], [4], these techniques struggled with noisy, multilingual, and imbalanced social media data. Recent advances in deep learning, particularly transformer-based models such as BERT, RoBERTa, and architectures like CNN, LSTM, and BiLSTM, have substantially improved detection capabilities by automatically learning rich textual representations [5]–[8]. However, most existing systems generate deterministic, overcon- fident predictions even in ambiguous contexts, posing risks in high-stakes moderation scenarios. This review comprehensively examines hate speech and cyberbullying detection techniques with emphasis on uncertainty-aware learning approaches, in- cluding Monte Carlo dropout, Bayesian modeling, and ensemble methods. We identify critical research gaps related to dataset bias, annotation inconsistency, limited multilingual coverage, and insufficient robustness. Our findings highlight the necessity of integrating uncertainty quantification into detection systems to enhance reliability, interpretability, and real-world applicability, guiding future development of trustworthy and ethically respon- sible content moderation frameworks
DOI: http://doi.org/10.5281/zenodo.20856177
