Men Also Like Shopping: Reducing Gender Bias Amplication Using Corpus-Level Constraints
Language is increasingly being used to define rich visual recognition problems with supporting image collections sourced from the web. Structured prediction models, which are used in these tasks to take advantage of correlations between co-occurring labels and visual input, risk inadvertently encoding and reinforcing social biases found in web corpora. We studied data and models associated with multi-label object classification and visual semantic role labeling, and found that (a) datasets for these tasks contain significant gender bias and (b) models trained on these datasets amplify those bias.
For example, the activity <em>cooking</em> is over 33% more likely to involve females than males in a training set, and a trained model further increases the disparity to 68% at test time. We propose adding corpus-level constraints that calibrate existing structured prediction models and designing an algorithm based on Lagrangian relaxation for collective inference. Our method results in almost no performance loss for the underlying recognition task but decreases the magnitude of bias amplification by 47.5% and 40.5% for multi-label classification and visual semantic role labeling, respectively.