Cyberbullying Detection from Bangla Text Using Cascaded Deep Hybrid Network

The use of social media is increasing tremendously due to the ease of the internet. As a result, cyberbullying or online harassment of people across social media platforms is increasing progressively. The detection and prevention of harassment texts and comments on social media are crucial for several major languages. Bengali is the sixth most widely used language worldwide, and more people are using social media. So, finding an efficient detection technique to manage and prevent cyberbullying is necessary. A significant number of studies on detecting cyberbullying use machine learning and have been conducted in English, Chinese, and Arabic. There are very few publications regarding Bengali languages. In this study, we simply present a hybrid deep learning model for Bengali text that can distinguish cyberbullying by analyzing and experimenting with several methods to find a feasible way of classifying such comments. We have used 44001 user comments from Facebook, divided into the following five categories: religious, sexual, threat, and not-bully. Our proposed hybrid deep neural network model performed better than baseline models and compared all these models with our proposed network. To analyze textual patterns in Bangla text, we conducted comprehensive experiments using baseline models of long-short-term memory (LSTM), BiLSTM, and CNN-BiGRU models. We proposed a CNN-BiLSTM model which cascaded by CNN and BiLSTM methods with various execution epochs, model layers, and tuning hyperparameters. To give a comparative comparison, different criteria were used to assess the effectiveness and performance of the models. The performance of our proposed CNN-BiLSTM architecture performed the most effective and accurate prediction, with validation accuracy of 88.5% and 80%, respectively, over binary classification and multi-class classification. The CNN-BiGRU performed with the second-largest accuracy of 86% over binary classification..