Abstract

Neural machine translation has significantly pushed forward the quality ofthe field. However, there are remaining big issues with the translations andone of them is fairness. Neural models are trained on large text corpora whichcontains biases and stereotypes. As a consequence, models inherit these socialbiases. Recent methods have shown results in reducing gender bias in othernatural language processing applications such as word embeddings. We takeadvantage of the fact that word embeddings are used in neural machinetranslation to propose the first debiased machine translation system.Specifically, we propose, experiment and analyze the integration of twodebiasing techniques over GloVe embeddings in the Transformer translationarchitecture. We evaluate our proposed system on a generic English-Spanishtask, showing gains up to one BLEU point. As for the gender bias evaluation, wegenerate a test set of occupations and we show that our proposed system learnsto equalize existing biases from the baseline system.

 

Introduction (beta)

None

 

Conclusion (beta)

None