Societal biases resonate in the retrieved contents of information retrieval (IR) systems, resulting in reinforcing existing stereotypes. Approaching this issue requires established measures of fairness regarding the representation of various social groups in retrieved contents, as well as methods to mitigate such biases, particularly in the light of the advances in deep ranking models. In this work, we first provide a novel framework to measure the fairness in the retrieved text contents of ranking models. Introducing a ranker-agnostic measurement, the framework also enables the disentanglement of the effect on fairness of collection from that of rankers. Second, we propose an adversarial bias mitigation approach applied to the stateof-the-art Bert rankers, which jointly learns to predict relevance and remove protected attributes. We conduct experiments on two passage retrieval collections (MS MARCO Passage Re-ranking and TREC Deep Learning 2019 Passage Re-ranking), which we extend by fairness annotations of a selected subset of queries regarding gender attributes. Our results on the MS MARCO benchmark show that, while the fairness of all ranking models is lower than the ones of ranker-agnostic baselines, the fairness in retrieved contents significantly improves when applying the proposed adversarial training. Lastly, we investigate the trade-off between fairness and utility, showing that through applying a combinatorial model selection method, we can maintain the significant improvements in fairness without any significant loss in utility.
CCS CONCEPTS• Information systems → Learning to rank; Test collections.