Towards Fair Retrieval: Controlling Bias through a Backpack-Inspired Architecture

Women are still underrepresented pretty much everywhere power hangs out—about 27% of MPs worldwide (Aug 2025) and ~27.5% of managerial roles (2022). Not exactly a rounding error. See IPU Parline and UN Women.

tl;dr: Ever searched “developer” images and gotten wall-to-wall dudes—or “nurse” and seen mostly women? That’s data bias doing cosplay as relevance. This paper uses Backpack language models (J. Hewitt et al., 2023) (which split each token into multiple, interpretable “sense” vectors) to identify which senses carry gender signal and then turns those down at inference—no retraining, just fewer stereotype vibes. On MS MARCO and a gender-bias IR benchmark, it cuts RaB/ARaB (less skew) (N. Rekabsaz et al., 2020) with only a tiny dent in NDCG/MRR, and the Backpack ranker still edges a similarly sized GPT-2 baseline.

Download the preprint from here.