An Empirical Survey of the Effectiveness of Debiasing Techniques for Pre-trained Language Models

TL;DR

This work provides a comprehensive empirical survey of five debiasing techniques (CDA, Dropout, INLP, Self-Debias, SentenceDebias) applied to multiple pre-trained language models across gender, racial, and religious biases, using SEAT, StereoSet, and CrowS-Pairs to quantify bias. It further assesses the impact of debiasing on language modeling (WikiText-2) and downstream NLU tasks (GLUE), revealing a general tradeoff: debiasing often increases perplexity and reduces LM scores while having limited adverse effects on downstream performance. Self-Debias consistently yields the strongest bias reduction across benchmarks, but variability in benchmark reliability and cross-model effects suggest caution in interpreting these improvements. The paper highlights the need for robust, multi-faceted evaluation and cross-cultural bias benchmarks to accurately measure debiasing efficacy and its practical implications for real-world deployments.

Abstract

Recent work has shown pre-trained language models capture social biases from the large amounts of text they are trained on. This has attracted attention to developing techniques that mitigate such biases. In this work, we perform an empirical survey of five recently proposed bias mitigation techniques: Counterfactual Data Augmentation (CDA), Dropout, Iterative Nullspace Projection, Self-Debias, and SentenceDebias. We quantify the effectiveness of each technique using three intrinsic bias benchmarks while also measuring the impact of these techniques on a model's language modeling ability, as well as its performance on downstream NLU tasks. We experimentally find that: (1) Self-Debias is the strongest debiasing technique, obtaining improved scores on all bias benchmarks; (2) Current debiasing techniques perform less consistently when mitigating non-gender biases; And (3) improvements on bias benchmarks such as StereoSet and CrowS-Pairs by using debiasing strategies are often accompanied by a decrease in language modeling ability, making it difficult to determine whether the bias mitigation was effective.