Iterative Multilingual Spectral Attribute Erasure

TL;DR

This paper introduces IMSAE, an iterative spectral method that effectively debiases multilingual models across multiple languages and demographic dimensions, including zero-shot scenarios, while preserving model utility.

Contribution

IMSAE is a novel iterative spectral approach that identifies and mitigates joint bias subspaces across multiple languages, improving multilingual debiasing effectiveness.

Findings

IMSAE outperforms monolingual and cross-lingual debiasing methods.

Effective in zero-shot settings with linguistically similar languages.

Maintains model utility across diverse language models.

Abstract

Multilingual representations embed words with similar meanings to share a common semantic space across languages, creating opportunities to transfer debiasing effects between languages. However, existing methods for debiasing are unable to exploit this opportunity because they operate on individual languages. We present Iterative Multilingual Spectral Attribute Erasure (IMSAE), which identifies and mitigates joint bias subspaces across multiple languages through iterative SVD-based truncation. Evaluating IMSAE across eight languages and five demographic dimensions, we demonstrate its effectiveness in both standard and zero-shot settings, where target language data is unavailable, but linguistically similar languages can be used for debiasing. Our comprehensive experiments across diverse language models (BERT, LLaMA, Mistral) show that IMSAE outperforms traditional monolingual and cross-lingual approaches while maintaining model utility.