Does Representational Fairness Imply Empirical Fairness?

Abstract

NLP technologies can cause unintended harmsif learned representations encode sensitive attributes of the author, or predictions systematically vary in quality across groups. Populardebiasing approaches, like adversarial training,remove sensitive information from representations in order to reduce disparate performance,however the relation between representationalfairness and empirical (performance) fairnesshas not been systematically studied. This paperfills this gap, and proposes a novel debiasingmethod building on contrastive learning to encourage a latent space that separates instancesbased on target label, while mixing instancesthat share protected attributes. Our results showthe effecti..