Cyberbullying Detection using LSTM-CNN architecture and its applications
Mihir Gada, Kaustubh Damania, Smita Sankhe
Abstract
In the present information age, there's been an increase in the use of social media, as a result, there have been multiple cyberbullying instances. To prevent or reduce cyberbullying, many existing approaches in the literature incorporate normal Machine Learning and Natural Language Processing text classification models without considering the sentence semantics. In this paper, we aim to target that issue. We have used word2vec to train our custom word embeddings upon which we built our LSTM - CNN architecture and finally it was trained on it. We made a comparative study of the above approaches by testing our model on Twitter posts and comments. The eminent performance achieved by our method has been observed in this study. We also created a web application that used the model to classify tweets as cyberbullying or not based on the toxicity score along with various features. The model was also implemented on Telegram Bot which is used to check and prevent cyberbullying. Two Chrome Extensions were built to redact and retract Not Safe For Work (NSFW) content and prevent cyberbullying on WhatsApp Web. We have been able to achieve excellent performance in the form of a 97% ROC AUC score for our model.