Attention-based Gender-Stereotype Detection

Gender stereotypes are perceptions about the typical physical, emotional, and social characteristics of individuals. Unlike gender bias which can result in the systematic and unfair treatment of individuals based on their gender, gender stereotypes do not always perpetuate a negative impact. Hence, there is no solid definition that frames what is considered a gender-stereotype in text. In addition, there is also a lack of labelled gender-stereotype datasets. This has led to most of the work in literature being about gender bias and not gender stereotypes. Therefore, in our research, we frame a clear definition of what constitutes a gender-stereotype in text. And using fairness benchmark datasets that were not originally intended for gender-stereotype studies, we propose an end-to-end approach to automatically identify language features that influence a gender-stereotypical sentence or text. We do this using the learned attention weights of a model. Our evaluations on the features that the model was using to predict gender-stereotypes aligned with our definition of gender-stereotypes.https://arrow.tudublin.ie/cddpos/1015/thumbnail.jp