Abstract: Social media produces a huge amount of dirty data. This allows harmful posts and comments to spread easily and create social violence. It is hard to automatically detect and stop these harmful messages. Therefore, we propose a model for identifying expressions that harm the health of mental youth. We explore the potential of understanding and identifying multimodal posts on X (formally Twitter) depicting the encouragement of violence in tweets. Moreover, the proposed model is used to refine the parameters of the Convolutional Neural Network (CNN) to maximize its strength. We have evaluated the performance of multiple models like CNN, BERT, Multilayer Perceptron, Random Forest, Logistic regression, BERT+CNN, BERT+MLP and BERT+LSTM. Based on our evaluation, the accuracy of models is as follows: Logistic regression at 94.08%, Random forest at 93.84%, Decision tree at 91.70%, Naïve Bayes at 43.53%, BERT+CNN at 61.67%, BERT+MLP at 91% and LSTM+MLP at 95.44%. The results reveal that the BERT+LSTM model performs very well in comparison to other models showing excellent results as 95.44% accuracy and 95.57% F1-Score. LSTM gives better results than all other models in the detection of the abusive data.

Key words: social media, abusive data, mental health, machine learning, deep learning