And this is very likely the reason for the mistake. See, Google Translate seems to take a shortcut to translating between language A and B, by translating from A to English and English to B.
For instance, for the last five years or so I've been seeing GT translate the Greek word for "swallow", the bird, to the French word for "to swallow", the verb [1]. In Greek and French, the two words have nothing in common (Greek is helidoni/katapino, French is hirondele/avaler). But in English, they are homonyms (swallow/swallow). It seems that GT sees a Greek word, translates it in English, then finds a set of homonyms and simply chooses the translation that's most "likely", meaning the one that's used to most often- which of course means it gets one meaning wrong all the time.
Last year Google made a big todo about its neural nets inventing an "interlingua", an intermediary language to which and from which all other languages it knows can be translated. In practice, I think this "interlingua" ends up being just plain old English.
_________________
[1] Recently I started seeing the translation "machaon", a kind of butterfly, which Google translates in English as "swallowtail". "Avaler" is still available as an alternative.
> And this is very likely the reason for the mistake. See, Google Translate seems to take a shortcut to translating between language A and B, by translating from A to English and English to B.
Although in this case, if you translate "halb zwei" from German→English it correctly translates it to "half past one".
It's plausible that French-English dataset is an order of magnitude larger, mostly due to Canada's bilinguality causing a large trove of parallel data.
For instance, for the last five years or so I've been seeing GT translate the Greek word for "swallow", the bird, to the French word for "to swallow", the verb [1]. In Greek and French, the two words have nothing in common (Greek is helidoni/katapino, French is hirondele/avaler). But in English, they are homonyms (swallow/swallow). It seems that GT sees a Greek word, translates it in English, then finds a set of homonyms and simply chooses the translation that's most "likely", meaning the one that's used to most often- which of course means it gets one meaning wrong all the time.
Last year Google made a big todo about its neural nets inventing an "interlingua", an intermediary language to which and from which all other languages it knows can be translated. In practice, I think this "interlingua" ends up being just plain old English.
_________________
[1] Recently I started seeing the translation "machaon", a kind of butterfly, which Google translates in English as "swallowtail". "Avaler" is still available as an alternative.