translators and gender-neutral words
2021-08-30
Algorithms built on large datasets have been showing a curious trait lately: sometimes they turn out racist, or sexist.
In Uzbek, when we talk about a third person, there is no gender. We always use the same personal pronoun, "u". In English and Russian the person's gender matters: he/she/his/her, он/она/его/её. In modern English, when the gender isn't known, "they" (or sometimes "she") is considered the relatively gender-neutral choice.
And yet it turns out today's Google Translate gets somewhat sexist when it translates that genderless "u" from gender-neutral languages into English. A thread on Twitter showed a text translated from Hungarian into English:
Google Translate decides the woman is beautiful and the man is clever; he reads a book, she washes the dishes; he teaches, she cooks.
Google Translate picks the pronoun for "u" based on the profession
or the activity.
Out of curiosity, I ran the same text from Uzbek into English. Apart from one or two cases, the translation came out the same. Yandex Translate goes with "he" in most cases.
When you build on existing data, it matters which segment that data came from. A speech recognizer trained mostly on American English, for instance, will do noticeably worse on other accents.