4 ms·

I think the sweet spot is machine-assisted translation. If you can get the output to be 90-95% correct, you can then display the raw and the machine output sid

by cbhl 6y ago

I think the sweet spot is machine-assisted translation. If you can get the output to be 90-95% correct, you can then display the raw and the machine output side-by-side, and have a human make corrections inline. Instead of a team of four working around the clock for a day or two, maybe you could have a translation as fast as it takes to proofread it three or four times end-to-end. Rev is the same idea in the speech transcription space -- they have humans listen to the audio and fix up a machine-generated transcription.

cinntaile 6y ago
I think you underestimate the amount of work that's required when "proofreading", you need another translator that could have done the work themselves. And it's not like you can just exchange a few words here and there and consider it done. If a human has to make corrections they'll likely rewrite the entire sentence and then the advantage of machine-assisted translation is pretty much gone. Overall you'll save time but the amount of time saved is probably less than what we think.
- numpad0 6y ago
  I think I’ve come across a PSA about this scheme. The story was that unaware translators puts discount on proofreading, so some clients send in dummy machine translation to qualify as “proofreading” task to get full translation at proofreading pricing.
Hamuko 6y ago
I've seen some people run visual novels through automated translations in aggregate like this: https://static.wikia.nocookie.net/muvluv/images/8/89/Be_a_good_boy_and_you_might_get_a_MiG-27_for_Christmas.jpg https://static.wikia.nocookie.net/muvluv/images/8/89/Be_a_go... You can definitely get some of the gist in there, but some of the automated translations are just way off. And pretty much none of them result in good prose. In addition, none of the translations get the names fully correct, so you definitely need someone to go and fix it.
numpad0 6y ago
You can’t get output to be “90-95% correct” from machine translation. People think of language translation as some sort of same dimension transformation but it’s more like re-projection that involve rotation in upper dimension. Simple warping goes only so far, neural networks give some uncanny slurries, human artists add a lot of their own brush strokes and it’s a lossy process both ways. Speech transcription is much more straightforward because speakers are supposed to have corresponding single literal expressions for each segments of voice.
kevingadd 6y ago
Machine-assisted translation is deceptively bad, especially for fiction content like manga. The machine TL might mix up the order of a sentence like 'bob verbed alice', fundamentally changing the meaning, alter context, or omit implied information. All of these errors will produce english sentences that appear sensical and valid and will get past most editors and quality checkers unless they're familiar with the source language or very familiar with the work. This sort of error even creeps into official human translations of anime/manga in some cases when the translators are working quickly without a good editor (i.e. simulcasts where they have tight deadlines and low budgets) In practice if you look at the fan translation community for manga machine-assisted translation is not given much more respect than machine translation - they both produce bad results and in many cases the people who normally welcome even a clumsy translation will reject machine TL and attempt to have it removed, because it often causes people to fundamentally misunderstand the work. The worst cases of machine-assisted TL in manga become infamous to the point of becoming shared memes - try googling "abaj" or "duwang" sometime. For a very simple pervasive example: Japanese frequently uses gender-neutral pronouns and when translating to English you'll need to appropriately select the right gendered pronoun (or proper name) for each one, if you can. This is something a human can do pretty accurately if they have enough context and knowledge of the material, but it is nearly impossible for a computer to do it accurately without a ton of assistance. In a novel this would be an easier problem because all the necessary context is in the text instead of in the art and panel layouts. You'll note this arxiv paper intentionally cheats on the gender problem.

by cbhl 6y ago

cinntaile 6y ago

I think you underestimate the amount of work that's required when "proofreading", you need another translator that could have done the work themselves. And it's not like you can just exchange a few words here and there and consider it done. If a human has to make corrections they'll likely rewrite the entire sentence and then the advantage of machine-assisted translation is pretty much gone. Overall you'll save time but the amount of time saved is probably less than what we think.

numpad0 6y ago
I think I’ve come across a PSA about this scheme. The story was that unaware translators puts discount on proofreading, so some clients send in dummy machine translation to qualify as “proofreading” task to get full translation at proofreading pricing.

Hamuko 6y ago

I've seen some people run visual novels through automated translations in aggregate like this: https://static.wikia.nocookie.net/muvluv/images/8/89/Be_a_good_boy_and_you_might_get_a_MiG-27_for_Christmas.jpg https://static.wikia.nocookie.net/muvluv/images/8/89/Be_a_go... You can definitely get some of the gist in there, but some of the automated translations are just way off. And pretty much none of them result in good prose. In addition, none of the translations get the names fully correct, so you definitely need someone to go and fix it.

numpad0 6y ago

You can’t get output to be “90-95% correct” from machine translation. People think of language translation as some sort of same dimension transformation but it’s more like re-projection that involve rotation in upper dimension. Simple warping goes only so far, neural networks give some uncanny slurries, human artists add a lot of their own brush strokes and it’s a lossy process both ways. Speech transcription is much more straightforward because speakers are supposed to have corresponding single literal expressions for each segments of voice.

kevingadd 6y ago

Machine-assisted translation is deceptively bad, especially for fiction content like manga. The machine TL might mix up the order of a sentence like 'bob verbed alice', fundamentally changing the meaning, alter context, or omit implied information. All of these errors will produce english sentences that appear sensical and valid and will get past most editors and quality checkers unless they're familiar with the source language or very familiar with the work. This sort of error even creeps into official human translations of anime/manga in some cases when the translators are working quickly without a good editor (i.e. simulcasts where they have tight deadlines and low budgets) In practice if you look at the fan translation community for manga machine-assisted translation is not given much more respect than machine translation - they both produce bad results and in many cases the people who normally welcome even a clumsy translation will reject machine TL and attempt to have it removed, because it often causes people to fundamentally misunderstand the work. The worst cases of machine-assisted TL in manga become infamous to the point of becoming shared memes - try googling "abaj" or "duwang" sometime. For a very simple pervasive example: Japanese frequently uses gender-neutral pronouns and when translating to English you'll need to appropriately select the right gendered pronoun (or proper name) for each one, if you can. This is something a human can do pretty accurately if they have enough context and knowledge of the material, but it is nearly impossible for a computer to do it accurately without a ton of assistance. In a novel this would be an easier problem because all the necessary context is in the text instead of in the art and panel layouts. You'll note this arxiv paper intentionally cheats on the gender problem.