As an aside, there seems to have been a dramatic change in Google Translate’s underlying approach several years ago: before it, the fr>en and zh>en translations

As an aside, there seems to have been a dramatic change in Google Translate’s underlying approach several years ago: before it, the fr>en and zh>en translations were reasonable (if stilted and occasionally ungrammatical) while both de>en and ja>en absolutely sucked (just made no sense whatsoever over groups of more than three or so words); I suspect the difficulty being the “global” transformations needed to translate SOV to SVO word order. It would be very interesting to know what they did (there are old-style statistical approaches that involve learning pairs of corresponding syntax trees, but I didn’t get the impression they were practical? might have been an artifact of the original DoD research being focused on zh and ru, which are both SVO—though ar is VSO and fa is SOV...).