CYBERBULLYING DETECTION ON SOCIAL MEDIA TEXTUAL CONTENTS USING DEEP LEARNING

Abstract

Cyberbullying has emerged as a significant concern across social media platforms, given its profound psychological and societal impact. Traditional moderation strategies—such as reporting or blocking users—are largely manual and often fall short in effectively mitigating harmful content. To address this gap, we present an automated cyberbullying detection system powered by a hybrid deep learning model that integrates XLNet and RoBERTa architectures. This combination harnesses the contextual understanding capabilities of Transformer-based language models, enabling more accurate and reliable identification of abusive language. The model is trained on a curated, balanced dataset that includes various forms of cyberbullying, such as those targeting gender or sexual orientation, ethnicity or race, and religion, along with neutral (non-cyberbullying) content.
Our hybrid approach significantly outperforms traditional machine learning techniques and standalone deep learning models, reaching a validation accuracy of approximately 96.7%. Unlike earlier methods that require extensive manual feature extraction and data preprocessing, our solution benefits from transfer learning to simplify the pipeline while boosting performance. For ease of use, the model is deployed through a Gradio-based interface, enabling real-time user interaction and content evaluation. The results underscore the potential of hybrid Transformer architectures in building scalable, efficient, and accessible tools for tackling online abuse.