Cyberbullying Detection using Neural Network Models

Abstract

Cyberbullying is the spreading of hate among different religions and people through internet. With the worldwide growth in the twitter users, it is now more prone to bullying. There are groups, organizations and individuals who are involved in online bullying. Such people and individuals can be a severe threat to our society. It is the need of hour to detect bullying and then classify its types to check the severity of such bullying comments. Work has been done in order to detect the online abuse on social media platforms. ML models were used earlier which proved to be successful in order to detect bullying but they were not very accurate in case on large datasets. Later DL models using the techniques of NLP were introduced which have the ability to detect the bullying content accurately. Previous work on text classification has mostly handled binary class problems and very little work has been done for multi class classification of cyberbullying. In this study our purpose is to detect online abuse and classify different classes of cyberbullying on a large dataset ‘Malignant Comment Classification’ which is a publically available dataset on kaggle. We have trained our machine learning and then deep learning models and adopted supervised learning approach. For evaluating performance on traditional models SVM, KNN, Decision Tree and boosting algorithms were applied on the dataset. For deep learning we have used LSTM, BERT and its variant, distillation BERT. After training phase we have made a comparison between these two different approaches i.e. ML and DL approach. The experimental results shows that DL models are more effective in text classification problems as our BERT model obtained an accuracy better than other state of the art models.