People often get harassed and bullied online, and both of these things can hurt their mental health. To keep online spaces safe, it is important to find and recognize patterns of abusive behavior. Most of the time, though, the present detection approaches are either rule-based algorithms or simple machine learning models. These approaches cannot handle the complexity of user-generated content when it comes to context and sequence features. We propose a hybrid deep learning framework dubbed the Bidirectional Long Short-Term Memory-Convolutional Neural Network (BiLSTM-CNN). This framework is meant to fix the problems that have been found. The CNN layer’s job is to find important local features so that classification can be better. On the other hand, the BiLSTM component performs well to diagnose the flow of words in both directions. Using the BiLSTM-CNN model, we were able to more accurately and sensitively detect risky content through our analysis of the texts that originated from social media posts. Analysis of the experimental evaluation data indicates that the BiLSTM-CNN model outperformed traditional models by detecting cyberbullying patterns more quickly and accurately. The model used a 97.4% recall, 97.8 accuracy, 98.5% F1-score, and 96.2% precision. These hybrid requirements helped improve recognition accuracy and strengthened the model’s ability to moderate content on online platforms.
Alsadh et al. (Thu,) studied this question.