Characterizing Stereotypical Bias from Privacy-preserving Pre-Training
This work addresses bias mitigation in privacy-preserving language models for AI fairness, but it is incremental as it builds on known links between privacy and bias.
The study investigated how differential privacy in text pre-training affects stereotypical bias in language models, finding that while bias generally decreases with tighter privacy, the reduction is not uniform across all social domains.
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.