Search Results

Automated translation (AT) plays a pivotal role in breaking down language barriers and fostering cross-cultural communication. However, gender bias in AT remains a pressing concern, particularly for language pairs where the source language generally does not have grammatical gender and the target language does. In this context, the present research aims to assess gender bias in English-Portuguese AT outputs. Two machine translation (MT) systems and two LLMs are evaluated in this study, examining the correlation between translated gender and societal stereotypes, and how the different models handle nouns whose gender is unknown. The outputs of two commercial MT systems (Google Translate and DeepL Translator) and two general-purpose LLMs (ChatGPT-4o and DeepSeek-V3), translating from English into European Portuguese, were analyzed. We used a subset of the WinoMT challenge set to assess gender rendering accuracy across three conditions: gender-defined nouns, gender-undefined nouns, and anaphoric pronoun resolution. Our main goal is therefore to evaluate the accuracy of gender information transmission and to understand the extent to which gender bias is present in AT outputs. Our results indicate that all four models struggle to faithfully convey gender information from source to target, with male-centric outputs predominating across conditions and MT systems showing a more pronounced bias. This work contributes a preliminary cross-system comparison for an underexplored language pair and lays the groundwork for larger-scale evaluations of gender-inclusive AT.