Detecting Hate Speech Utilizing Deep Convolutional Network and Transformer Models
Utkarsh Mittal
Abstract
Online social networks exhibit a significant prevalence of hate speeches, which poses a potential threat to the society and fosters targeted animosity towards specific communities and authorities. Although online platforms are available to automate few mechanisms of hate speeches but classification w.r.t. to different domains and their accuracy are the big issues, and are challenging the researchers, media, and the academic world. The present study addresses the identifying of Hate Speeches through the comparative analysis of the classification efficacy and model intricacy of four distinct Deep Neural Network models; namely CNN (baseline), bidirectional LSTM with attention, pretrained BERT, and fine-tuned RoBERTa transformer models, and utilizing a ternary classification system (hate, offensive, non-hate). The performance of the subject under consideration was assessed through the application of Accuracy, F1-score, and Matthew's correlation coefficient (MCC) metrics on the test set.