Cyber Bullying Detection

This research presents a comprehensive study on cyberbullying detection leveraging hybrid multimodal approaches by integrating textual, sentiment, emotion, and embedding-based features with advanced machine learning classifiers. Using multilingual datasets derived from real-world online conflict discourses and processed into structured CSV formats, we evaluate the performance of Logistic Regression and Linear SVC models under original and SMOTE-balanced conditions. The results highlight that contextual embeddings, particularly BERT and RoBERTa, significantly outperform traditional textual and sentiment features, while the hybrid multimodal model achieves the highest accuracy and F1-scores. Comparative experiments demonstrate that although the hybrid approach incurs higher computational costs, it provides superior robustness in detecting nuanced and context-dependent bullying patterns, thereby ensuring scalability and real-time relevance. This work contributes to advancing automated detection systems with potential deployment in social media monitoring and online conflict moderation.