Commercial MT fails gender-neutral pronouns, ACL study finds
TL;DR
- Three commercial machine translation systems produce grammatical and semantic errors when translating gender-neutral pronouns, an ACL 2023 paper reports.
- The study covered English paired with Danish, Farsi, French, German, and Italian, plus Danish to English, and found gender neutrality 'often not preserved.'
- Authors Lauscher, Nozza, Miltersen, Crowley, and Hovy surveyed affected native speakers to draft recommendations and warn wrong pronoun translations can discriminate against non-binary individuals.
Three commercial machine translation systems routinely produce grammatical and semantic errors when the input uses gender-neutral or neopronouns, and often strip gender neutrality from the output. That is the finding of an ACL 2023 paper by Anne Lauscher, Debora Nozza, Ehm Miltersen, Archie Crowley, and Dirk Hovy, framed by the authors as a "reality check" on where commercial MT stands on identity-inclusive language.
The team compared translations of gendered and gender-neutral pronouns between English and five other languages (Danish, Farsi, French, German, Italian), plus Danish to English. Their error analysis reports that "the presence of a gender-neutral pronoun often leads to grammatical and semantic translation errors" and that "gender neutrality is often not preserved." To shape recommendations for future MT research, they surveyed "affected native speakers from diverse languages."
The paper does not name the three vendors it tested. It states the stakes in its own words: "Wrong pronoun translations can discriminate against marginalized groups, e.g., non-binary individuals." The work was presented at the 61st Annual Meeting of the Association for Computational Linguistics in Toronto, pages 377 to 392. Two researchers in our tracker resurfaced the link this week.
Shared on Bluesky by 2 AI experts
-
#MemoryMonday #NLProc Lauscher et al.'s (2023) study, 'How Commercial Machine Translation Fails to Handle (Neo-)Pronouns', unveils gender-biased errors in pronoun translations, amplifying the need for gender identity-in…
View on Bluesky →
Originally reported by aclanthology.org
Read the original article →Original headline: What about “em”? How Commercial Machine Translation Fails to Handle (Neo-)Pronouns