PAN at CLEF 2021

While models built and tested on a specific dataset and for a specific task often achieve very good performance, they then fail to generalise when they are applied to new, unseen data. In this talk I will discuss the importance and challenges of achieving generalisable performance in social media research with a particular focus on fact verification and hate speech detection. I will present some of our recent work in this direction, as well as discuss open challenges to further the capacity of generalisation especially in hate speech detection.

Language has the power to reinforce stereotypes and project social biases onto others, either through overt hate or subtle biases. Accounting for this toxicity and social bias in language is crucial for natural language processing (NLP) systems to be safely and ethically deployed in the world. In this talk, I will first analyze a failure case of automatic hate speech detection, in which we find that models tend to flag speech by African Americans as toxic more often than by others. We trace the origins of the biases back to the annotated datasets, and show that we can reduce these biases, by making a tweet's dialect more explicit during the annotation process. Then, as an alternative to binary hate speech detection, I will present Social Bias Frames, a new structured formalism for distilling biased implications of language. Using a new corpus of 150k structured annotations, we show that models can learn to reason about high-level offensiveness of statements, but struggle to explain why a statement might be harmful. Finally, I will introduce PowerTransformer, a new unsupervised model for controllable debiasing of text through the lens of connotation frames of power and agency. With this model, we show that subtle gender biases in how characters are portrayed in stories and movies can be mitigated through automatic rewriting. I will conclude with future directions for better reasoning about toxicity and social biases in language.