Deep learning approaches for Bengali cyberbullying cetection on social media: a comparative study of BiLSTM, BiGRU and BERT models

Abstract

As technology becomes more accessible, it is now much easier than ever to abuse someone by misusing it. Usually, people use slang or absurd language with the goal of bullying, harassing, and harming someone by using social media. Moreover, these types of cyberbullying activities are more widespread among teenagers and young people despite knowing the fact that these may break someone down emotionally and may lead them towards suicidal activities. Hence, our goal is to detect cyberbullying happening on social media in the Bengali language with the help of state-of-theart deep learning and Natural Language Processing (NLP) techniques. We have examined with 3 different algorithms such as Bi-LSTM, Bi-GRU and BERT for both multiclass and binary classification. For both binary and multiclass classifications, BERT outperformed the other two models in terms of performance with the f1 score of 0.89 for binary and 0.85 for multiclass classification. Our proposed state-of-theart transformer model BERT will detect whether a message or comment is sent to harass someone or not and could help to take immediate action against them. Therefore, our research might have a positive impact on changing the social media environment by detecting hate speeches and bullying messages.