Cyberbullying Detection using Pre-Trained BERT Model

Cyberbullying is spread across various social media platforms. It is a wrong deed in which the victim is harassed by receiving the derogatory / provocative / sensitive images or text messages by the bully. Detection of such message/post in such large platforms is very difficult and may sometimes lead to false detection. Recently, deep neural network based models have shown significant improvement over traditional models in detecting cyberbullying. Also, new and more complex deep learning architectures are being developed which are proving to be useful in various NLP tasks. Google researchers has recently developed a language learning model called BERT, which is capable of generating contextual embeddings and is also able to produce task specific embeddings for classification. A new approach is proposed to cyberbullying detection in social media platforms by using the novel pre-trained BERT model with a single linear neural network layer on top as a classifier, which improves over the existing results. The model is trained and evaluated on two social media datasets of which one dataset is small size and the second dataset is relatively larger size.