General Phrase Debiaser: Debiasing Masked Language Models at a Multi-Token Level

TL;DR

This paper tackles the problem of biases in pretrained masked language models, focusing on biases at the phrase level which are under-addressed by word-level debiasing methods. It introduces General Phrase Debiaser, a two-stage pipeline: a Phrase Filter Stage that extracts stereotypical phrases from Wikipedia using topic seeds, and a Model Debias Stage that discovers multi-token prompts and fine-tunes the MLM to reduce bias. The method leverages a multi-token cloze-probing framework with distributions computed via $P^{(n)}$ and measures prompt-bias discrepancies with Jensen-Shannon Divergence ($JSD$), optimized through beam search and subsequent full-model fine-tuning using the selected prompts; the approach centers the loss on stereotypical phrases rather than the entire vocabulary, enabling targeted debiasing. Empirical results on SEAT and GLUE across BERT, ALBERT, and DistilBERT show state-of-the-art debiasing performance with only modest trade-offs in language understanding, demonstrating effective cross-domain bias reduction while preserving capabilities. These findings provide a practical pathway for phrase-aware debiasing in encoder-based models and suggest extendability to cross-modal settings, with code and models released for reproducibility.

Abstract

The social biases and unwelcome stereotypes revealed by pretrained language models are becoming obstacles to their application. Compared to numerous debiasing methods targeting word level, there has been relatively less attention on biases present at phrase level, limiting the performance of debiasing in discipline domains. In this paper, we propose an automatic multi-token debiasing pipeline called \textbf{General Phrase Debiaser}, which is capable of mitigating phrase-level biases in masked language models. Specifically, our method consists of a \textit{phrase filter stage} that generates stereotypical phrases from Wikipedia pages as well as a \textit{model debias stage} that can debias models at the multi-token level to tackle bias challenges on phrases. The latter searches for prompts that trigger model's bias, and then uses them for debiasing. State-of-the-art results on standard datasets and metrics show that our approach can significantly reduce gender biases on both career and multiple disciplines, across models with varying parameter sizes.

Figures (2)

• Figure 1: The proposed General phrase Debiaser pipeline has two stages. In phrase filter stage, we filter out stereotypical phrases $S_{weighted}$ and $S_{unweighted}$ from WikiPedia pages with MLM and stereotypical seeds. In model debias stage, we search biased prompts at multi-token granularity, and fine-tuning MLM with them.
• Figure 2: Computation of multi-token JSD loss.