Local Contrastive Editing of Gender Stereotypes
This addresses the issue of stereotypical bias in language models for safer AI applications, representing an incremental advance in parameter-efficient editing methods.
The paper tackles the problem of gender stereotypes in language models by introducing local contrastive editing to localize and edit a small subset of weights (< 0.5%) associated with bias, enabling precise control over these parameters.
Stereotypical bias encoded in language models (LMs) poses a threat to safe language technology, yet our understanding of how bias manifests in the parameters of LMs remains incomplete. We introduce local contrastive editing that enables the localization and editing of a subset of weights in a target model in relation to a reference model. We deploy this approach to identify and modify subsets of weights that are associated with gender stereotypes in LMs. Through a series of experiments, we demonstrate that local contrastive editing can precisely localize and control a small subset (< 0.5%) of weights that encode gender bias. Our work (i) advances our understanding of how stereotypical biases can manifest in the parameter space of LMs and (ii) opens up new avenues for developing parameter-efficient strategies for controlling model properties in a contrastive manner.