Inference-Time Selective Debiasing to Enhance Fairness in Text Classification Models

TL;DR

This paper introduces a novel inference-time selective debiasing method called LEACE, which improves fairness in text classification models by identifying and reducing bias in predictions without retraining the model.

Contribution

The paper presents a new inference-time debiasing approach that uses KL divergence for bias detection and applies LEACE to enhance fairness in text classification models.

Findings

Selective debiasing reduces bias more effectively than standard uncertainty methods.

The approach narrows the performance gap between post-processing and at-training debiasing methods.

Experiments show improved fairness and prediction quality on text datasets.

Abstract

We propose selective debiasing -- an inference-time safety mechanism designed to enhance the overall model quality in terms of prediction performance and fairness, especially in scenarios where retraining the model is impractical. The method draws inspiration from selective classification, where at inference time, predictions with low quality, as indicated by their uncertainty scores, are discarded. In our approach, we identify the potentially biased model predictions and, instead of discarding them, we remove bias from these predictions using LEACE -- a post-processing debiasing method. To select problematic predictions, we propose a bias quantification approach based on KL divergence, which achieves better results than standard uncertainty quantification methods. Experiments on text classification datasets with encoder-based classification models demonstrate that selective debiasing helps to reduce the performance gap between post-processing methods and debiasing techniques from the at-training and pre-processing categories.