Detecting White Supremacist Hate Speech Using Domain Specific Word Embedding with Deep Learning and BERT

Abstract

White supremacist hate speech is one of the most recently observed harmful content on social media. The critical influence of these radical groups is no longer limited to social media and can negatively affect society by promoting racial hatred and violence. Traditional channels of reporting hate speech have proved inadequate due to the tremendous explosion of information and the implicit nature of hate speech. Therefore, it is necessary to detect such speech automatically and in a timely manner. This research investigates the feasibility of automatically detecting white supremacist hate speech on Twitter using deep learning and natural language processing techniques. Two deep learning models are investigated in this research. The first approach utilizes a bidirectional Long Short-Term Memory (BiLSTM) model along with domain-specific word embeddings extracted from white supremacist corpus to capture the semantic of white supremacist slangs and coded words. The second approach utilizes one of the most recent language models, which is Bidirectional Encoder Representations from Transformers (BERT). The BiLSTM model achieved 0.75 F1-score and BERT reached a 0.80 F1-score. Both models are tested on a balanced dataset combined from Twitter and a Stormfront dataset compiled from white supremacist forum.