Characterizing Stereotypical Bias from Privacy-preserving Pre-Training
Stefan Arnold, Rene Gröbner, Annika Schreiner
TL;DR
This work investigates whether differential privacy–driven text privatization reduces stereotypical bias in language models and whether such reductions are uniform across social domains. It trains BERT on privacy-modified WebText using word-level Madlib perturbations with GloVe embeddings and evaluates bias via StereoSet and CrowS-Pairs, reporting results with a bias-sensitive pseudo-likelihood metric. The findings show a general decline in stereotypical bias as privacy tightens, but effects are heterogeneous across attributes and tasks, with some biases amplified or unchanged in certain domains. The study highlights that privacy-preserving text perturbation does not guarantee uniform fairness improvements and calls for careful, multi-faceted bias diagnostics when deploying privatized LMs, especially across diverse social categories.
Abstract
Differential Privacy (DP) can be applied to raw text by exploiting the spatial arrangement of words in an embedding space. We investigate the implications of such text privatization on Language Models (LMs) and their tendency towards stereotypical associations. Since previous studies documented that linguistic proficiency correlates with stereotypical bias, one could assume that techniques for text privatization, which are known to degrade language modeling capabilities, would cancel out undesirable biases. By testing BERT models trained on texts containing biased statements primed with varying degrees of privacy, our study reveals that while stereotypical bias generally diminishes when privacy is tightened, text privatization does not uniformly equate to diminishing bias across all social domains. This highlights the need for careful diagnosis of bias in LMs that undergo text privatization.