Us vs. Them: A dataset of populist attitudes, news bias and emotions

Abstract

Computational modelling of political discourse tasks has become an increasingly important area of research in natural language processing. Populist rhetoric has risen across the political sphere in recent years; however, computational approaches to it have been scarce due to its complex nature. In this paper, we present the new Us vs. Them dataset, consisting of 6861 Reddit comments annotated for populist attitudes and the first large-scale computational models of this phenomenon. We investigate the relationship between populist mindsets and social groups, as well as a range of emotions typically associated with these. We set a baseline for two tasks related to populist attitudes and present a set of multi-task learning models that leverage and demonstrate the importance of emotion and group identification as auxiliary tasks.