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

TL;DR

This paper explores automatic detection of white supremacist hate speech on Twitter using domain-specific embeddings with LSTM and BERT, achieving high F1-scores and demonstrating the effectiveness of deep learning in identifying harmful content.

Contribution

It introduces a novel approach combining domain-specific embeddings and BERT for detecting white supremacist hate speech, with experimental validation showing improved accuracy over traditional methods.

Findings

LSTM approach achieved 0.74890 F1-score.

BERT approach achieved 0.79605 F1-score.

Both methods effectively detect hate speech in social media texts.

Abstract

White supremacists embrace a radical ideology that considers white people superior to people of other races. The critical influence of these groups is no longer limited to social media; they also have a significant effect on society in many ways by promoting racial hatred and violence. White supremacist hate speech is one of the most recently observed harmful content on social media.Traditional channels of reporting hate speech have proved inadequate due to the tremendous explosion of information, and therefore, it is necessary to find an automatic way to detect such speech in a timely manner. This research investigates the viability of automatically detecting white supremacist hate speech on Twitter by using deep learning and natural language processing techniques. Through our experiments, we used two approaches, the first approach is by using domain-specific embeddings which are extracted from white supremacist corpus in order to catch the meaning of this white supremacist slang with bidirectional Long Short-Term Memory (LSTM) deep learning model, this approach reached a 0.74890 F1-score. The second approach is by using the one of the most recent language model which is BERT, BERT model provides the state of the art of most NLP tasks. It reached to a 0.79605 F1-score. Both approaches are tested on a balanced dataset given that our experiments were based on textual data only. The dataset was combined from dataset created from Twitter and a Stormfront dataset compiled from that white supremacist forum.