Framework for Detecting Toxic Speech Using BERT and Deep Learning

We have witnessed a big surge in the use of online social media platforms in the past few years with the widespread availability of mobile phones and cheaper data rates. This has also led to an increase in misuse of these platforms to spread hatred in the society. The stats show that most of the cyberbullying happens on social media platforms like Facebook, WhatsApp, etc. The Hatred speech was used for spreading hatred against specific persons, communities, etc. It is necessary to check for the hate speech on social media platforms before it can lead to damaging the peace situation. Automating the task of hate speech detection is a very important step in mitigating this issue as such contents can be quickly acted upon before it spreads hatred. In this paper, we are proposing an Automated Hate speech identification system. The datasets are obtained from HASOC which stands for Hate Speech and Offensive Content Identification in Indo-European Languages, for sub-task b to classify into one of the four categories, Hate, Profane, Offensive, or None for English and Hindi languages. The experiments are carried out to compare the performance of Bert-based architecture with conventional deep learning architectures including CNN (Convolution Neural Network) architecture for English subtask and for Hindi Subtask we fine-tuned the Bert model which gave better results than conventional solutions. We obtained 75.41% accuracy with BERT-based fine tuning better than baseline methods LSTM (Long Short-Term Memory) with 41% accuracy for Hindi dataset and for English dataset, we achieved an accuracy of 82.92% compared to 81.57% with baseline models.