Enhanced Convolutional Neural Network Model Using Bidirectional Encoder Representations from Transformers and Long Short-Term Memory for Hate Speech Detection

Abstract

Hate speech has become a prevalent and concerning issue in this digital age as a result of the adoption of social media as a means of communication, social interaction, and driving social discussion. This trend has opened a novel front in cybercrime, cyberbullying, privacy violations, and exposure of the people component of information security to violence, subsequently violating the governance and internal security of nations. The detection and mitigation of hate speech on online platforms have gained significant attention due to the growing influence of social media and the internet as communication channels and the adoption of the same as a means of propagating hate and hate crimes. This thesis offers a model for the more accurate detection of hate speech on an online platform using Bidirectional Encoder Representations from Transformers (BERT) to enhance the feature selection layer of Convolutional Neural Networks (CNN) and a hybrid of CNN and Long-Short-Term Memory (LSTM) to detect hate speech. The proposed BERT-CNN-LSTM model achieved an accuracy of 96.1%, which is an increase of over 0.45% from 95.63% in the benchmark paper, thereby contributing a novel and more accurate tool that, when adapted into the security policies of organisations and governments, will help safeguard the rights and privacy of users, thereby protecting them from both physical and psychological harm resulting from hate propagation over the internet.