Transformer-Based Hybrid Model for Robust Abusive Content Detection in Online Social Networks (OSNs)

Identifying abusive content in online social networks (OSNs) continues to be a considerable issue due to linguistic complexity, code-mixing, and the existence of implicit abuse. This article presents a transformer-based hybrid model that combines contextual embeddings from transformer models with convolutional neural networks (CNNs) for pattern recognition and subsequently employs an attention mechanism to improve contextual interpretation. The model proficiently captures both word-level and character-level features, overcoming the drawbacks of current techniques that frequently break with implicit and multilingual abuse. An F1-score of 96.7%, an accuracy of 95.73%, a precision of 94.5%, a recall of 96.8%, and detailed tests on benchmark datasets show that the proposed model works very well. A comparison with cutting-edge methodologies further proves the model’s efficacy in minimizing false positives and false negatives. The findings underscore the effectiveness of the suggested method in detecting abusive content across diverse contexts. This article enhances the safety of digital platforms by offering a dependable and scalable approach for the real-time detection of abusive content.