This paper was accepted on the fifth Workshop on Gender Bias in Pure Language Processing 2024.
Machine translation (MT) techniques typically translate phrases with ambiguous gender (e.g., English time period “the nurse”) into the gendered type that’s most prevalent within the techniques’ coaching knowledge (e.g., “enfermera”, the Spanish time period for a feminine nurse). This typically displays and perpetuates dangerous stereotypes current in society. With MT consumer interfaces in thoughts that permit for resolving gender ambiguity in a frictionless method, we research the issue of producing all grammatically appropriate gendered translation alternate options. We open supply practice and take a look at datasets for 5 language pairs and set up benchmarks for this activity. Our key technical contribution is a novel semi-supervised answer for producing alternate options that integrates seamlessly with commonplace MT fashions and maintains excessive efficiency with out requiring extra elements or rising inference overhead.