Analyzing Gender Representation in Multilingual Models

TL;DR

This paper investigates how gender distinctions are internally represented in multilingual language models, revealing shared and language-specific components that explain transferability and challenges in gender removal across languages.

Contribution

It provides a detailed analysis of gender representation in multilingual models, highlighting the coexistence of shared and language-specific components and their impact on transferability.

Findings

Gender representations have shared and language-specific components.

Gender classification transfers well across languages.

Gender removal interventions do not transfer easily across languages.

Abstract

Multilingual language models were shown to allow for nontrivial transfer across scripts and languages. In this work, we study the structure of the internal representations that enable this transfer. We focus on the representation of gender distinctions as a practical case study, and examine the extent to which the gender concept is encoded in shared subspaces across different languages. Our analysis shows that gender representations consist of several prominent components that are shared across languages, alongside language-specific components. The existence of language-independent and language-specific components provides an explanation for an intriguing empirical observation we make: while gender classification transfers well across languages, interventions for gender removal, trained on a single language, do not transfer easily to others.