PEARC '22: Practice and Experience in Advanced Research Computing Proceedings

Abstract

The growing prevalence of online hate speech is concerning, given the massive growth of online platforms. Hate speech is defined as language that attacks, humiliates, or incites violence against specific groups. According to research, there is a link between online hate speech and real-world crimes, as well as victims' deteriorating mental health. To combat the online prevalence of abusive speech, hate speech detection models based on machine learning and natural language processing are being developed to automatically detect the toxicity of online content. However, current models tend to mislabel African American English (AAE) text as hate speech at a significantly higher rate than texts written in Standard American English (SAE). To confirm the existence of systematic racism within these models, I evaluate a logical regression model and a BERT model. Then, I determine the efficacy of the bias reduction method for the BERT model and the correlation between model performance and reduced bias.