Understanding the use of emojis in hate speech

Abstract

The latest advances in NLP (natural language processing) has led to the launch of the much needed machine-driven hate speech detection. Nevertheless, people continuously find new forms of hateful expressions that are easily identified by humans, but not by machines. One such expression is the mix of text and emojis, a type of visual hate speech that is increasingly used to evade algorithmic moderation. This research analyzes chat conversations from the popular streaming platform Twitch to understand the varied types of visual hate speech. Emotes were used sometimes to replace a letter, seek attention, or for emotional expression. We created a labeled dataset that contains 29,721 cases of emotes replacing letters. Based on the dataset, we built a neural network classifier and identify visual hate speech that would otherwise be undetected through traditional methods and caught an additional 1.3% examples of hate speech out of 15 million chat utterances.