Almost as good as human translators at getting the basic meaning right and putting words in the right order, for similar language pairs, translating wikipedia articles.
The skill of human translators isn't based on getting the information right, that's a given. The difference between a good and a bad translator is how well they capture the thoughts and intent of the speaker, whether they can make their audience feel the same way as the original did.
I say it every time one of these articles comes up, a machine capable of universal translation will be none other than a true AI.
Is there an intermediate representation or is it a big matrix of pairwise language translations? If it's the former we could just start speaking that
I believe we are already doing that, even in this thread;)
No intermediate representation, which is puzzling to me. There's so much shared structure between languages, especially things like closely related European languages. You get none of that with this approach.
Either way the intermediate representation would probably be a vector, so not very practical for human use.
> Google researcher Quoc Le says Google’s big translation upgrade could also lead to improved relations between people and machines.