I'm curious as to how well these AI applications actually do translate. I did once work in a tech sector that used machine translation followed by human "editing." The company I worked for updated it's manuals every quarter. We had to be very judicious and change as little as possible of the previously tranlated stuff so that it didn't have to be retranslated.
I've been watching some English language television shows with the closed captions on. The number of mistakes I've seen in them is quite astonishing, especially when there is an accent variant involved (say Scottish). They are often things that most English speakers would have heard correctly. Would AI be able to deal with the various accents in Japanese?
I've often wondered about the translations from Japanese into English. From what I've been able to gather, Japanese can be—very unspecific, I guess I'd call it—and there is often wide room for interpretation. And I've watched much subtitled Japanese material, some of which was translated almost word for word, which makes the it like reading word salad, especially if there are idioms. Would AI do better, I wonder? Or was it AI that made it so confusing? I once had to edit closed captions that were presumably typed by a human (but now I'm not so sure). I did so by listening to the material and comparing it to the audio because the message of the talk was considered to be very important. My biggest catch was finding that there was a proposed cyclotron in the middle of my city's downtown. It turns out it was really a proposed cyle path. At least whoever or whatever was typing didn't type "psychopath." But what I had heard sound nothing like either of those. A human ear/eye is always going to be necessary for a translation that makes sense for to the users. No human translators? Human editors or so-called proofreaders cost money, too.
LLMs can be extremely useful but they are not in any way capable of producing finalized prose, especially in the context of manga, where much of the nuance has to be derived from imagery. Japanese is a highly contextual language, where pronouns and even subjects are very often dropped in speech, and the listener/reader is expected to "use common sense" to infer meaning. This is the major stumbling block for LLMs and it's difficult to imagine them producing content that won't require large amounts of cleanup (which defeats the supposed time-saving in the end.) In my experience the translation component is not any more or less of a bottleneck than the many other steps in the process.
Yes, the idea that people are "into manga," the medium, rather than "into manga," the specific stories and characters. There's a reason many manga don't get picked up for distribution. Are these companies eager to stake their reputations on fetish porn? Because that's a big chunk of what doesn't get released.
Can’t blame them for ‘knowing’ the key market. Boobs sell. Manga Ent. knew that, ADV knew that, and so on. The problem is so much is tied up to old ‘playbooks’, both here in the US and in Japan. The mantra is, if I remember right, 5000 units for video, 10,000 units for book. If you can’t hit those numbers you don’t bother trying. But those numbers are based on 1980’s retail reality. The walk-in retail for anime and manga is effectively dead. In the old days all you had to do is solicit Suncoast and bang, there’s your 5000 videos right there. Everything sold to Tower Records, Best Buy, FYE, Circuit City was extra profit.
Manga, the same. Walden’s, Boarders, Barnes and Nobel, plus over 3000 comic book shops and all the independents, you could move 10,000 copies of anything.
All gone. B&N is still strong on manga but they don’t do ‘depth’ and ‘fill’ like they used to. FYI still exists but they’re clearly on life support. And Amazon is NOT filling the gap in sales lost to stores because so much shopping is IMPULSE. You see it, you buy it.
So for manga, publishers need to seriously recalibrate. And I could go on as you know too well, Matt. 😁
Is it possible to have a permalink for this article that does not include the k-word? As the current URL is being blocked/downrated/hidden by social media AI/algorithms due to its inclusion.
This actually illustrates the point of this essay, which is that AI / algorithms by definition lack the common sense to make editorial decisions. How telling is it that they would block a line from Shakespeare?
Thanks, Matt. This topic - and you and Hiroko - have been on my mind. Thanks for making the important distinction between translation and localization. I've been making a similar one with bilingual/bicultural students who show concern about what generative AI means to their career prospects. Short answer: don't worry. Longer answer - people who are culturally savvy in two or more geographic work cultures generally find work!
Generally, when the cost of doing something goes down, we do more of that thing. This was first pointed out as Jevon's Paradox - it was noted when the cost of coal extraction went down, we used more, not less, coal. So too with li-ion batteries. As the cost of those has come down (from $200 per kwh to $100 or below), the cost of EVs has come down. So generally, this leads to more affordable EVs.
So too with search. It does cost Google and others money to serve a response to search queries (and that cost has gone up with generative AI responses). But free search induced a lot of search demand.
Where I'm going with this - even before Orange or Mantra had come along, DeepL was there, as was Google Translate. I remember using Gengo - basically, semipro translators - for a project in like 2014. I would use DeepL for that now. But I still have to edit/smooth DeepL when its done (in my case, going from E to J or vice-versa). So AI has eaten the semipro, so-so translation service. But AI quality still needs some professional TLC to be enterprise-grade.
For localization, if the cost comes down, I would think more content, and maybe lower-quality, less-known content, would get localized. That *could* be a positive, from the perspective of helping more content to more people, more affordably.
So I wonder if the DSC market is a comparable for high-touch, localization professionals . Meaning, DSCs didn't go away when smartphones came along. Rather, DSCs moved upmarket, and kept serving professionals and serious hobbyists. And more people took more pictures since everyone has a pretty good but not great connected camera with their smartphone. The bottom of the pyramid got a lot wider.
As a practical matter, I would suspect that the presence of DeepL or Orange or Mantra might prevent "surge pricing" in localization. This is what has happened in hospitality. US hotels lost their ability to surge-price since there's more non-professional supply (rooms) in the market now.
There is a huge demand for "mediocre" translations (perhaps better described as "for-information" translations) that aren't intended to be consumed by anyone other than the requestor, for their own interest/knowledge. LLMs excel at this kind of thing: newspaper articles, reports, etc. (As an aside, a good 30% of my time as a translator at the US Patent and Trademark office was spent doing this, reading off patent specs to give examiners more of a hint as to if they needed full translations. I suspect LLMs have now replaced the humans in this role.) The issue with these startups is that they're using that mediocre content as the springboard to create prose for mass consumption. Nobody wants to be on the side of the horse and buggy when the automobile is around the corner, which is why so many companies are pouring money into this space right now (DeepL just landed 300mil in investment for expanding the capabilities of its translation model.) But everyone knows the content isn't ready for prime time, and won't be soon, so this feels less like streamlining the process than making the worklives of the remaining staff (editors, etc.) hell.
Ha! Is that Roger from Toybox DX? I was going to bring up my fav example from that vid! There is a scene where the ‘gun fan’ is reeling off knowledge and he produces a shotgun. The subs call it a ‘ Rallyot’ as if it were a brand, but it should have been ‘riot’ , as in a police riot shotgun. Simple thing easily caught if one has even a smattering of gun knowledge, or bothering to ask someone.
Ever since I entered this floating world I’ve seen wars over translating. I think it goes to the old stigma of adaption, but getting all knotted up about localization doesn’t distinguish between 7-Zark-7 and calling rice balls chocolate cake.
I finally finished Dunbine (and yes I have comments on the recent Pure Tokyoscope 😄) and man what a mess that ended up. It was said long ago that the ‘ruling class’ characters spoke an older, more formal version of Japanese but it comes across as something like google translate filtered thru a not-native English speaker. And if I read one more time “YOU CANNOT DO AS YOU LIKE
I'm curious as to how well these AI applications actually do translate. I did once work in a tech sector that used machine translation followed by human "editing." The company I worked for updated it's manuals every quarter. We had to be very judicious and change as little as possible of the previously tranlated stuff so that it didn't have to be retranslated.
I've been watching some English language television shows with the closed captions on. The number of mistakes I've seen in them is quite astonishing, especially when there is an accent variant involved (say Scottish). They are often things that most English speakers would have heard correctly. Would AI be able to deal with the various accents in Japanese?
I've often wondered about the translations from Japanese into English. From what I've been able to gather, Japanese can be—very unspecific, I guess I'd call it—and there is often wide room for interpretation. And I've watched much subtitled Japanese material, some of which was translated almost word for word, which makes the it like reading word salad, especially if there are idioms. Would AI do better, I wonder? Or was it AI that made it so confusing? I once had to edit closed captions that were presumably typed by a human (but now I'm not so sure). I did so by listening to the material and comparing it to the audio because the message of the talk was considered to be very important. My biggest catch was finding that there was a proposed cyclotron in the middle of my city's downtown. It turns out it was really a proposed cyle path. At least whoever or whatever was typing didn't type "psychopath." But what I had heard sound nothing like either of those. A human ear/eye is always going to be necessary for a translation that makes sense for to the users. No human translators? Human editors or so-called proofreaders cost money, too.
LLMs can be extremely useful but they are not in any way capable of producing finalized prose, especially in the context of manga, where much of the nuance has to be derived from imagery. Japanese is a highly contextual language, where pronouns and even subjects are very often dropped in speech, and the listener/reader is expected to "use common sense" to infer meaning. This is the major stumbling block for LLMs and it's difficult to imagine them producing content that won't require large amounts of cleanup (which defeats the supposed time-saving in the end.) In my experience the translation component is not any more or less of a bottleneck than the many other steps in the process.
“It is difficult to get a man to understand something when his salary depends on his not understanding it.” —Upton Sinclair
I would love to see a demonstration of how poorly these AI translation engines would handle just 10 minutes of Otaku No Video.
Exactly.
"Westerners currently enjoy the cream of the crop; will flooding the market with unsold leftovers satisfy, or nauseate?"
Wasn't this the TokyoPop method of curation? We all know how that ended up.
Yes, the idea that people are "into manga," the medium, rather than "into manga," the specific stories and characters. There's a reason many manga don't get picked up for distribution. Are these companies eager to stake their reputations on fetish porn? Because that's a big chunk of what doesn't get released.
Can’t blame them for ‘knowing’ the key market. Boobs sell. Manga Ent. knew that, ADV knew that, and so on. The problem is so much is tied up to old ‘playbooks’, both here in the US and in Japan. The mantra is, if I remember right, 5000 units for video, 10,000 units for book. If you can’t hit those numbers you don’t bother trying. But those numbers are based on 1980’s retail reality. The walk-in retail for anime and manga is effectively dead. In the old days all you had to do is solicit Suncoast and bang, there’s your 5000 videos right there. Everything sold to Tower Records, Best Buy, FYE, Circuit City was extra profit.
Manga, the same. Walden’s, Boarders, Barnes and Nobel, plus over 3000 comic book shops and all the independents, you could move 10,000 copies of anything.
All gone. B&N is still strong on manga but they don’t do ‘depth’ and ‘fill’ like they used to. FYI still exists but they’re clearly on life support. And Amazon is NOT filling the gap in sales lost to stores because so much shopping is IMPULSE. You see it, you buy it.
So for manga, publishers need to seriously recalibrate. And I could go on as you know too well, Matt. 😁
Is it possible to have a permalink for this article that does not include the k-word? As the current URL is being blocked/downrated/hidden by social media AI/algorithms due to its inclusion.
Yes! Here you go: https://substack.com/home/post/p-144860963
This actually illustrates the point of this essay, which is that AI / algorithms by definition lack the common sense to make editorial decisions. How telling is it that they would block a line from Shakespeare?
Thanks, Matt. This topic - and you and Hiroko - have been on my mind. Thanks for making the important distinction between translation and localization. I've been making a similar one with bilingual/bicultural students who show concern about what generative AI means to their career prospects. Short answer: don't worry. Longer answer - people who are culturally savvy in two or more geographic work cultures generally find work!
Generally, when the cost of doing something goes down, we do more of that thing. This was first pointed out as Jevon's Paradox - it was noted when the cost of coal extraction went down, we used more, not less, coal. So too with li-ion batteries. As the cost of those has come down (from $200 per kwh to $100 or below), the cost of EVs has come down. So generally, this leads to more affordable EVs.
So too with search. It does cost Google and others money to serve a response to search queries (and that cost has gone up with generative AI responses). But free search induced a lot of search demand.
Where I'm going with this - even before Orange or Mantra had come along, DeepL was there, as was Google Translate. I remember using Gengo - basically, semipro translators - for a project in like 2014. I would use DeepL for that now. But I still have to edit/smooth DeepL when its done (in my case, going from E to J or vice-versa). So AI has eaten the semipro, so-so translation service. But AI quality still needs some professional TLC to be enterprise-grade.
For localization, if the cost comes down, I would think more content, and maybe lower-quality, less-known content, would get localized. That *could* be a positive, from the perspective of helping more content to more people, more affordably.
So I wonder if the DSC market is a comparable for high-touch, localization professionals . Meaning, DSCs didn't go away when smartphones came along. Rather, DSCs moved upmarket, and kept serving professionals and serious hobbyists. And more people took more pictures since everyone has a pretty good but not great connected camera with their smartphone. The bottom of the pyramid got a lot wider.
As a practical matter, I would suspect that the presence of DeepL or Orange or Mantra might prevent "surge pricing" in localization. This is what has happened in hospitality. US hotels lost their ability to surge-price since there's more non-professional supply (rooms) in the market now.
Thanks again for sharing.
There is a huge demand for "mediocre" translations (perhaps better described as "for-information" translations) that aren't intended to be consumed by anyone other than the requestor, for their own interest/knowledge. LLMs excel at this kind of thing: newspaper articles, reports, etc. (As an aside, a good 30% of my time as a translator at the US Patent and Trademark office was spent doing this, reading off patent specs to give examiners more of a hint as to if they needed full translations. I suspect LLMs have now replaced the humans in this role.) The issue with these startups is that they're using that mediocre content as the springboard to create prose for mass consumption. Nobody wants to be on the side of the horse and buggy when the automobile is around the corner, which is why so many companies are pouring money into this space right now (DeepL just landed 300mil in investment for expanding the capabilities of its translation model.) But everyone knows the content isn't ready for prime time, and won't be soon, so this feels less like streamlining the process than making the worklives of the remaining staff (editors, etc.) hell.
Ahem. Hit post. Can’t edit. Hurm. Baka. 😄
Ha! Is that Roger from Toybox DX? I was going to bring up my fav example from that vid! There is a scene where the ‘gun fan’ is reeling off knowledge and he produces a shotgun. The subs call it a ‘ Rallyot’ as if it were a brand, but it should have been ‘riot’ , as in a police riot shotgun. Simple thing easily caught if one has even a smattering of gun knowledge, or bothering to ask someone.
Ever since I entered this floating world I’ve seen wars over translating. I think it goes to the old stigma of adaption, but getting all knotted up about localization doesn’t distinguish between 7-Zark-7 and calling rice balls chocolate cake.
I finally finished Dunbine (and yes I have comments on the recent Pure Tokyoscope 😄) and man what a mess that ended up. It was said long ago that the ‘ruling class’ characters spoke an older, more formal version of Japanese but it comes across as something like google translate filtered thru a not-native English speaker. And if I read one more time “YOU CANNOT DO AS YOU LIKE
“video game development, which interestingly doesn’t seem to be anywhere nearly as enthusiastic about AI translation”
Maybe because they are already a lot closer to AI in their work and know how much human editing it needs.
Just so!
I wish you or someone in this damn world would finish translating "the five star stories" manga