Gender-Neutral English to Portuguese Machine Translator: Promoting Inclusive Language

• Ricardo Trainotti Rabonato Unifesp
• Evangelos Milios Dalhousie University
• Lilian Berton Unifesp

Resumo

Machine translation (MT) plays a crucial role in globalization, making access to information more inclusive, although challenges persist for less popular languages, like Portuguese. One of the most complex challenges in the automatic translation into languages such as Portuguese is the precise preservation of masculine and feminine grammatical gender. There are still situations where translation does not adequately reflect gender equality, often reinforcing societal stereotypes. We aim to explore approaches to ensure fairness in English to Portuguese MT through post-processing techniques, which aim to apply some transformation to the model’s output. To this end, we used the MarianMT model as our foundation, then we fine-tuned it using a dataset of English-Portuguese sentences that was generated and carefully crafted to mitigate gender bias within the sentences. The results on gender disparities metrics, based on the WinoMT test set for MT such as ΔG, ΔS, and the overall accuracy (preserving the gender of the entity from the original) significantly improved with some drop in BLEU (Bilingual Evaluation Understudy) score. Our study focuses on addressing gender bias in the Portuguese language. However, it can also be adapted to other languages, since it is crucial to ensure truly fair and inclusive global communication.