FairAdapt: An Exploration of Fairness-promoting Adapters

Anonymous

TL;DR: Exploring the intersection between fairness-promoting algorithms and adapters

Abstract: Large language models work better for some than others, and lightweight mitigation of performance disparities across social groups could help bridge inequality gaps. Here, we explore fairness-promoting adapters as a potential mitigation technique. We find that generally adapters lead to as good or better performance than full fine-tuning, with mixed effects on group disparity. Combining fairness-promoting adapters does not lead to smaller group disparity, and while Adapter Fusion is superior to model stipulation, such systems fail to outperform non-fairness promoting adapters. Combinations of fairness-promoting adapters seem to positively effect group fairness under temporal concept drift, although, as expected, we observe a generalized performance drop. From the perspective of group fairness, our results are somewhat negative, and we discuss the potential bottlenecks for current approaches to mitigating group disparity.

Paper Type: short

Research Area: Ethics, Bias, and Fairness

Contribution Types: Model analysis & interpretability

Languages Studied: Chinese, German