Not one bit sexist: a competitive game for counterfactuals against sexist language

We will challenge users to a competitive game where two or more parties try to turn sexist online messages non-sexist. This will achieve three distinct goals: (i) create an exceptionally rich dataset of counterfactual examples to improve automated sexism detection, (ii) gain unprecedented insights on how cultures and demographics play into people’s perceptions of sexism, and (iii) facilitate the discourse around the issue of sexist language online, fostering resilience in civil society. The game requires encoding rules, tracking the different phases of a match, and supporting the interaction of distinct player roles such as participants and jury: we will leverage Mephisto to manage this complexity. By making players compete to create alternatives to amend sexism, we will inductively unearth what impacts perceptions of sexism, especially in cases that lead to disagreements between individuals. This procedure will serve as a blueprint for a gamut of subjective NLP tasks beyond sexism detection.