Quick example of machine translation bias. (The Finnish sentences don't have gendered pronouns)
https://twitter.com/annabrchisky/status/1368989109381980164
What baffles me is that English has a direct equivalent for this, called singular they. It's not like it even needs to take biased probabilistic data as a fallback when the algorithm isn't confident about what to do.
As soon as they added Finnish<->English translation, or any other combination of gendered and non-gendenred pronouns, they KNEW this would occur. It's a very basic knowledge for this type of product, so there's no way this is unnoticed. And still nothing has been done about it.
So
1. they use some neural model that is biased (which is expected because the world sucks hence the data will suck)
and
2. they don't debias the data nor combine with other translation methods since this specific problem could be fixed with a hardcoded rule
1. they use some neural model that is biased (which is expected because the world sucks hence the data will suck)
and
2. they don't debias the data nor combine with other translation methods since this specific problem could be fixed with a hardcoded rule
I'm not saying removing sexism from ML models is easy. On the contrary, this is super hard and there's a ton of research being done in this area. But this is a known problem since the beginning of offering this product, and they don't even offer low-tech fallbacks when possible.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
