Thread by @WellsLucasSanto

Last talk for this #FAccT21 session is "Towards Cross-Lingual Generalization of Translation Gender Bias" with Won Ik Cho, Jiwon Kim, Jaeyoung Yang, Nam Soo Kim.

Remember the Google translate case study that added sexist gender pronouns when translating? This is about that.

Languages like Turkish, Korean, Japanese, etc. use gender-neutral pronouns, but when translating to languages like English, often use gender-specific pronouns. But also, languages like Spanish and French, have gendered *expressions* as well to keep in mind.

This matters because existing translation systems could contain biases that could generate translated results that are offensive and stereotypical, and not always accurate.

Note that not all languages have colloquially used gender neutral pronouns (like the English "they").

Interesting... part of what they measured was disparate impact: the differences of the impact on female and male pronoun translation accuracy. But one thing they noted was that they *did not* see gender-neutral pronoun translations happen (like using the English "they").

For them, they defined a translation from a gender-neutral language to English where the gender-neutral "they" was used would be an *accurate* translation. Makes me wonder why this isn't just the norm, for us to *always* do this type of translation.

Anyways, their results and take-away suggest that "inductive bias as a social stereotype is a major factor causing the errors and augmenting training corpora may not be a solution". Essentially, it's a human problem!! We can't just fix the datasets to fix the results.