Intersectionality Bias Analysis: Annotating People in Images with Demographic and Contextual Attributes

Following [60], we annotate people in images via crowdsourcing. Details on the annotation process are provided in Section 3.3. For each person, we use Amazon Mechanical Turk[ https://www.mturk.com/] (AMT) to get four demographic and two contextual attributes. With up to six attributes per person, our goal is to analyze bias from an intersectional perspective.

Demographic attributes: We denote demographic attributes as characteristics of people that are intrinsic to their being and cannot be easily changed. We annotate four attributes with the following categorization: [Demographic categorization systems cannot represent all the different identities, and thus, it should only be seen as a rough and non-inclusive approximation to different social groups in order to analyze disparities.]