Us vs. them: a dataset of political othering in Brazilian WhatsApp discourse

Political polarization is increasingly expressed through discursive strategies that construct moral and symbolic boundaries between “usž and “themž. Among these strategies, othering is particularly important because it portrays sociopolitical groups as fundamentally different, inferior, or threatening. However, despite extensive research on hate speech and misinformation, othering remains less studied, partly due to the lack of annotated resources that capture this phenomenon. In this paper, we introduce a large-scale labeled dataset for political othering from Brazilian public WhatsApp groups. Starting from a manually annotated gold set, we develop an LLM-assisted annotation pipeline with systematic human validation, resulting in 8.5k labeled messages. Our analysis shows that othering is not limited to explicit insults or traditional hate-speech targets. Instead, it frequently targets ideological, partisan, and institutional opponents through negative affect and the discursive construction of group boundaries. These findings highlight political othering as a distinct dimension of harmful discourse in online environments.