Characterizing Stereotypical Bias from Privacy-preserving Pre-Training

TL;DR

This paper examines how applying differential privacy to text affects stereotypical biases in language models, revealing that bias reduction is inconsistent across social domains and emphasizing the importance of bias diagnosis post-privatization.

Contribution

It provides an empirical analysis of the impact of privacy-preserving text privatization on stereotypical bias in language models, highlighting domain-specific effects.

Findings

Bias generally decreases with increased privacy

Bias reduction is inconsistent across social domains

Text privatization does not uniformly eliminate stereotypes

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.