Self-Attention for Cyberbullying Detection

In recent years, cyberbullying has grown out of proportion due to the increasing usage of social media platforms along with the benefit of user anonymization over the Internet. Affecting people across all demographics, the effect of cyberbullying has been more pronounced over adolescents and insecure individuals. Victims suffer from societal isolation, depression, degrading self-confidence and suicidal thoughts. Thus, prevention of cyberbullying becomes a necessity and requires timely detection. Recent advances in Deep learning and Natural Language Processing have provided suitable models to predict whether a text sample is an example of cyberbullying. In this context, we explore the adaptivity and efficiency of self-attention models in detecting cyberbullying. Though a few of the recent works in this context have employed models like deep neural networks, SVM, CNN, LSTM and other hybrid models, to the best of our knowledge, this is the first work exploring self-attention models which have achieved state-of-the-art accuracies in Machine Translation tasks since 2017. We experiment with the Wikipedia, Formspring and Twitter cyberbullying datasets and achieve more efficient results over existing cyberbullying detection models. We also propose new research directions within cyberbullying detection over recent forms of media like Internet memes which pose a variety of new and hybrid problems.