BiasEdit: Debiasing Stereotyped Language Models via Model Editing
This addresses the issue of bias in AI language models for users and developers, offering a more efficient alternative to existing debiasing methods, though it is incremental in improving upon prior techniques.
The paper tackles the problem of stereotyped biases in language models by proposing BiasEdit, an efficient model editing method that uses lightweight networks to generate parameter updates, achieving effective debiasing with minimal impact on language modeling abilities as demonstrated on StereoSet and Crows-Pairs benchmarks.
Previous studies have established that language models manifest stereotyped biases. Existing debiasing strategies, such as retraining a model with counterfactual data, representation projection, and prompting often fail to efficiently eliminate bias or directly alter the models' biased internal representations. To address these issues, we propose BiasEdit, an efficient model editing method to remove stereotypical bias from language models through lightweight networks that act as editors to generate parameter updates. BiasEdit employs a debiasing loss guiding editor networks to conduct local edits on partial parameters of a language model for debiasing while preserving the language modeling abilities during editing through a retention loss. Experiments on StereoSet and Crows-Pairs demonstrate the effectiveness, efficiency, and robustness of BiasEdit in eliminating bias compared to tangental debiasing baselines and little to no impact on the language models' general capabilities. In addition, we conduct bias tracing to probe bias in various modules and explore bias editing impacts on different components of language models.