Does Context Help Mitigate Gender Bias in Neural Machine Translation?

TL;DR

Does Context Help Mitigate Gender Bias in Neural Machine Translation? investigates whether context can reduce gender bias in NMT by examining two phenomena: English→German/French translation of stereotypical professions and Basque→Spanish translation with non-informative context. The study compares sentence-level baselines with 2to1 context-aware Transformer models across datasets like MT-GenEval and COH-TGT:GENDER, finding that context can improve feminine-form accuracy but may preserve or amplify masculine bias. Non-informative context can further exacerbate bias in Basque→Spanish, especially in domains with strong masculine bias, underscoring the need for finer-grained, domain-aware bias-mitigation strategies. Overall, context improves certain gendered translations but does not reliably mitigate bias, highlighting the complexity of bias in NMT and the need for targeted evaluation and mitigation approaches.

Abstract

Neural Machine Translation models tend to perpetuate gender bias present in their training data distribution. Context-aware models have been previously suggested as a means to mitigate this type of bias. In this work, we examine this claim by analysing in detail the translation of stereotypical professions in English to German, and translation with non-informative context in Basque to Spanish. Our results show that, although context-aware models can significantly enhance translation accuracy for feminine terms, they can still maintain or even amplify gender bias. These results highlight the need for more fine-grained approaches to bias mitigation in Neural Machine Translation.

Figures (1)

• Figure 1: Distribution of sentence-level predictions in stereotipically feminine professions.