Why do we Hate Migrants? A Double Machine Learning-based Approach

AI-based NLP literature has explored antipathy toward the marginalized section of society, such as migrants, and their social acceptance. Broadly, extant literature has conceptualized this as an online hate speech detection task and employed predictive ML models. However, a crucial omission in this literature is the genesis (or causality) of online hate, i.e., why do we hate migrants? Drawing insights from social science literature, we have identified three antecedents of online hate: Cultural, Economic, and Security concerns. Subsequently, we probe -which of these concerns triggers higher toxicity on online platforms? Initially, we consider OLS-based regression analysis and SHAP framework to identify the predictors of toxicity, and subsequently, we use Double Machine Learning (DML)-based casual analysis to investigate whether good predictors of toxicity are also causally significant. We find that the causal effect of Cultural concerns on toxicity is higher than Security and Economic concerns.