Assessing the Impact of Algospeak on Sentiment Analysis Models

The innovative use of linguistic strategies in written communication on social media platforms has emerged as a method to circumvent harmful content detection mechanisms. While current algorithms exhibit satisfactory performance in content moderation, their efficacy in identifying and processing coded speech remains underexplored. This study analyses social media posts that incorporate altered words and two commonly employed strategies: symbols and hyphens. Through a comparative analysis of sentiment scores for both coded and non-coded forms, our findings highlight the effectiveness of these linguistic modifications in altering model predictions. Furthermore, we provide a qualitative analysis of the posts that presented greater shifts in polarity across different sentiment analysers. Implications of digital communication innovations are discussed in relation to their interaction with detection mechanisms.