) so any deficiencies are all y'alls fault . No useful comments on what the "error rate" could possibly mean, so would appreciate actual insight into that figure beyond "it doesn't make sense, so it's wrong!".
https://www.forbes.com/sites/jackke...-job-losses-at-ibm-and-chegg/?sh=2ae21c8975a4Going to copy and elaborate on a question I posted in the Ancient Magus' Bride Ch. 96 thread, because I think a conversation about this (likely to expand) trend is worthy of having out in the open. It's worth noting beforehand that I've had a bit of time since I first asked the question to do more reading up on Mantra and their claim the engine is more than just supercharged MTL. The engine apparently actually breaks down the source image and identifies nontextual factors (gender of speakers is specifically mentioned) to better guide the engine in selecting the best output. In my mind this targets the current greatest advantage of a human translator--the possession of in situ and cultural knowledge that allows them to choose the better word/sentence in context. I doubt Mantra are quite there yet...but I find it likely they will be before we know it.
Back to the actual question:
I'd be interested in seeing an objective pro/con analysis of the Mantra-assisted translation. The claim on the Mantra press release with their partnership with Shueisha is that the error rate of translation in their model is ~1.6%. No source for that number, best I could see in the PR, but I would not expect much from a public-facing corpo press release. The real question is, does the alleged accuracy hold up on inspection?
-----
I've seen some comments that the translation suffers from odd phrasing or poor word choice. Fair criticism, although I see that as equally a fault of the editor/proofreader as of the AI. I should also note that Mantra itself claims that its LLM was trained on fan translations (hopefully under an appropriately structured NDA...but not hopeful on that end because piracygoodbad
Pixiv/Fanbox apparently has an autotranslation tool powered by Mantra? I don't have any experience with those sites, so...any stories about how well it works?
Also, a 2020 paper on the Mantra model, which I've only briefly scanned for the juicy gossip, indicates that the strength of their AI LLM over traditional MTL is that it is "context-aware" to better guide the model in selecting the correct words to ultimately use, by adding non-textual factors derived from OCR of the source image into the LLM. Not misgendering (wince) characters a la Ichigo Reader sounds pretty swell. I have no background in the statistical models they use to evaluate accuracy, so...I'll have to assume that the BLEU model is a widely accepted and appropriate one. Doesn't quite explain where that 1.6% figure comes from, alas. Expert input greatly appreciated.
-----
Back to the original point: if that alleged level of accuracy does hold up, even within a ballpark, that means that:
1) instead of paying a translator hundreds of dollars (or equivalent in JPY) for a translation that may take hours, they can pay pennies for one done with 98%+ accuracy in just seconds. Output/JPY jumps by several orders of magnitude (that's factors of ten/hundreds/thousands, for you who don't math). Or another way...imagine translating the entirety of Detective Conan at the same cost as a single human-translated chapter.
2) potentially more time spent on ED/PR once the raw AI translation is complete, but only a 1.6% (-ish) increase in the big scheme of things, so the savings from (1) should still vastly outweigh any increased ED/PR costs. Depending on how editing is currently structured and budgeted (and there should be ED/PR even if a human did the original TL) this may not increase costs at all. So...massive savings from (1) - negligible increased costs from (2) = some savings per chapter/title
3) Savings per title = more titles(?) (yeah...not gonna touch that discussion. Evil Corporation (R) has little relevance to whether AI translation is a good or bad thing)
Thus far all I've heard is complaints no better than an allergy to AI/MTL without any, you know, facts to back up that feeling. I'm no particular fan of AI, and I do worry about the loss of jobs coming from them (my own job category will likely disappear in 5-10 years at the current rate) but I hate it when a community decides things are bad based entirely on hearsay and ego. I want to see the bad, along with the good, and let it be judged accordingly.
TL;DR if there are objective analyses breakind down the actual performance of Mantra (or any competitor, honestly) for better or worse, go ahead and spill the tea please. Sources much appreciated. Practical experience in the translation and publication fields would be even better.
there's no way i'll accept this statement "manga has "continuously suffered" damages from online piracy and illegal English translations"Going to copy and elaborate on a question I posted in the Ancient Magus' Bride Ch. 96 thread, because I think a conversation about this (likely to expand) trend is worthy of having out in the open. It's worth noting beforehand that I've had a bit of time since I first asked the question to do more reading up on Mantra and their claim the engine is more than just supercharged MTL. The engine apparently actually breaks down the source image and identifies nontextual factors (gender of speakers is specifically mentioned) to better guide the engine in selecting the best output. In my mind this targets the current greatest advantage of a human translator--the possession of in situ and cultural knowledge that allows them to choose the better word/sentence in context. I doubt Mantra are quite there yet...but I find it likely they will be before we know it.
Back to the actual question:
I'd be interested in seeing an objective pro/con analysis of the Mantra-assisted translation. The claim on the Mantra press release with their partnership with Shueisha is that the error rate of translation in their model is ~1.6%. No source for that number, best I could see in the PR, but I would not expect much from a public-facing corpo press release. The real question is, does the alleged accuracy hold up on inspection?
-----
I've seen some comments that the translation suffers from odd phrasing or poor word choice. Fair criticism, although I see that as equally a fault of the editor/proofreader as of the AI. I should also note that Mantra itself claims that its LLM was trained on fan translations (hopefully under an appropriately structured NDA...but not hopeful on that end because piracygoodbad
Pixiv/Fanbox apparently has an autotranslation tool powered by Mantra? I don't have any experience with those sites, so...any stories about how well it works?
Also, a 2020 paper on the Mantra model, which I've only briefly scanned for the juicy gossip, indicates that the strength of their AI LLM over traditional MTL is that it is "context-aware" to better guide the model in selecting the correct words to ultimately use, by adding non-textual factors derived from OCR of the source image into the LLM. Not misgendering (wince) characters a la Ichigo Reader sounds pretty swell. I have no background in the statistical models they use to evaluate accuracy, so...I'll have to assume that the BLEU model is a widely accepted and appropriate one. Doesn't quite explain where that 1.6% figure comes from, alas. Expert input greatly appreciated.
-----
Back to the original point: if that alleged level of accuracy does hold up, even within a ballpark, that means that:
1) instead of paying a translator hundreds of dollars (or equivalent in JPY) for a translation that may take hours, they can pay pennies for one done with 98%+ accuracy in just seconds. Output/JPY jumps by several orders of magnitude (that's factors of ten/hundreds/thousands, for you who don't math). Or another way...imagine translating the entirety of Detective Conan at the same cost as a single human-translated chapter.
2) potentially more time spent on ED/PR once the raw AI translation is complete, but only a 1.6% (-ish) increase in the big scheme of things, so the savings from (1) should still vastly outweigh any increased ED/PR costs. Depending on how editing is currently structured and budgeted (and there should be ED/PR even if a human did the original TL) this may not increase costs at all. So...massive savings from (1) - negligible increased costs from (2) = some savings per chapter/title
3) Savings per title = more titles(?) (yeah...not gonna touch that discussion. Evil Corporation (R) has little relevance to whether AI translation is a good or bad thing)
Thus far all I've heard is complaints no better than an allergy to AI/MTL without any, you know, facts to back up that feeling. I'm no particular fan of AI, and I do worry about the loss of jobs coming from them (my own job category will likely disappear in 5-10 years at the current rate) but I hate it when a community decides things are bad based entirely on hearsay and ego. I want to see the bad, along with the good, and let it be judged accordingly.
TL;DR if there are objective analyses breakind down the actual performance of Mantra (or any competitor, honestly) for better or worse, go ahead and spill the tea please. Sources much appreciated. Practical experience in the translation and publication fields would be even better.
Here's an objective analysis from someone who's read the shit they drop for $2/mo in their app: It's garbage lmao. It's not necessarily a matter of "accuracy," (it's obviously hard to tell if a translation is accurate without knowing and reading the source language) but rather it has zero ability to "voice" characters, so every single bit of dialog ends up sounding like an out of character summary of what they would be saying rather than feeling like an actual distinct character talking. The end product is essentially completely sucked dry of flavor except from what scraps you can glean from the art alone. Anything that's subtext will be lost. More than half of any given character's personality will be lost. Most jokes will fall flat, you might not even notice they were trying to be funny. It'll pass the time, you'll technically know the plot, but at that point you might as well just read a fandom wiki summary of the manga instead.
Sora the Troll (a JP TLer, voice actor, our aho) showcased how to use ChatGPT4 to have good JP -> EN translations. I wouldn't put it past Mantra for not having actual human personnel to do Editing and PR when the AI itself can do it.
Editing and PR are still be handled at the publisher's level, I think, rather than Mantra's. It's a fair point, though, that with current technology there is a lot of financial pressure to
Subjective hyperbole aside, do you have some concrete examples to back up your conclusions? I have seen comments in the Magus thread noting a decrease in the effective reading grade level of the AI output (vs a fan TL), which is the kind of analysis I would like to see. Not trying to knock your position here, just genuinely curious to see details beyond broad generalizations.
I first noticed when going back to reread Golden Kamuy, a favorite series of mine that I first read with the excellent fan translation by Everyday Heroes here on mangadex:
I mean I could theoretically use Google Translate to translate manga chapters and then clean up the output so that it sounds grammatically correct in English without knowing an ounce of Japanese and most people would be none the wiser
yup 100%
Going to copy and elaborate on a question I posted in the Ancient Magus' Bride Ch. 96 thread, because I think a conversation about this (likely to expand) trend is worthy of having out in the open. It's worth noting beforehand that I've had a bit of time since I first asked the question to do more reading up on Mantra and their claim the engine is more than just supercharged MTL. The engine apparently actually breaks down the source image and identifies nontextual factors (gender of speakers is specifically mentioned) to better guide the engine in selecting the best output. In my mind this targets the current greatest advantage of a human translator--the possession of in situ and cultural knowledge that allows them to choose the better word/sentence in context. I doubt Mantra are quite there yet...but I find it likely they will be before we know it.
Back to the actual question:
I'd be interested in seeing an objective pro/con analysis of the Mantra-assisted translation. The claim on the Mantra press release with their partnership with Shueisha is that the error rate of translation in their model is ~1.6%. No source for that number, best I could see in the PR, but I would not expect much from a public-facing corpo press release. The real question is, does the alleged accuracy hold up on inspection?
-----
I've seen some comments that the translation suffers from odd phrasing or poor word choice. Fair criticism, although I see that as equally a fault of the editor/proofreader as of the AI. I should also note that Mantra itself claims that its LLM was trained on fan translations (hopefully under an appropriately structured NDA...but not hopeful on that end because piracygoodbad
Pixiv/Fanbox apparently has an autotranslation tool powered by Mantra? I don't have any experience with those sites, so...any stories about how well it works?
Also, a 2020 paper on the Mantra model, which I've only briefly scanned for the juicy gossip, indicates that the strength of their AI LLM over traditional MTL is that it is "context-aware" to better guide the model in selecting the correct words to ultimately use, by adding non-textual factors derived from OCR of the source image into the LLM. Not misgendering (wince) characters a la Ichigo Reader sounds pretty swell. I have no background in the statistical models they use to evaluate accuracy, so...I'll have to assume that the BLEU model is a widely accepted and appropriate one. Doesn't quite explain where that 1.6% figure comes from, alas. Expert input greatly appreciated.
-----
Back to the original point: if that alleged level of accuracy does hold up, even within a ballpark, that means that:
1) instead of paying a translator hundreds of dollars (or equivalent in JPY) for a translation that may take hours, they can pay pennies for one done with 98%+ accuracy in just seconds. Output/JPY jumps by several orders of magnitude (that's factors of ten/hundreds/thousands, for you who don't math). Or another way...imagine translating the entirety of Detective Conan at the same cost as a single human-translated chapter.
2) potentially more time spent on ED/PR once the raw AI translation is complete, but only a 1.6% (-ish) increase in the big scheme of things, so the savings from (1) should still vastly outweigh any increased ED/PR costs. Depending on how editing is currently structured and budgeted (and there should be ED/PR even if a human did the original TL) this may not increase costs at all. So...massive savings from (1) - negligible increased costs from (2) = some savings per chapter/title
3) Savings per title = more titles(?) (yeah...not gonna touch that discussion. Evil Corporation (R) has little relevance to whether AI translation is a good or bad thing)
Thus far all I've heard is complaints no better than an allergy to AI/MTL without any, you know, facts to back up that feeling. I'm no particular fan of AI, and I do worry about the loss of jobs coming from them (my own job category will likely disappear in 5-10 years at the current rate) but I hate it when a community decides things are bad based entirely on hearsay and ego. I want to see the bad, along with the good, and let it be judged accordingly.
TL;DR if there are objective analyses breakind down the actual performance of Mantra (or any competitor, honestly) for better or worse, go ahead and spill the tea please. Sources much appreciated. Practical experience in the translation and publication fields would be even better.
It exposed Japan as a den of pedophiles, preying on 16 yo and less and is seen much repeated drafts of Usagi Drop, Uncle - middle school niece grooming, etc. In addition to their obsession to high school students in mini skirts. Not to mention their masterpiece: Boku no Pico.
Well not really, it wouldn't be there if there were no anime translated.
not true.
