> and I don't understand why some users are boycotting the translation due to the AI used in the translation work, but that's beside the point
Segagaga has a ton of obscure, referential, meta humor that isn't easily translated to English. The "cleaned up machine translation" approach means that a lot of this is lost. Looking at some screenshots of the game, the script seems stiff and overly formal, much like how direct machine translation of Japanese text reads. https://bsky.app/profile/did:plc:2jjromh55tf7pp7s4hsvurf4/po... Obviously it's better than nothing, but people are pissed off because the "edited machine translation" workflow leads to poor results, not because of some reflexive anti-AI bias or whatever.
Based on what I read in the article, MT was used as an experiment and proof of concept, and dedicated translation was done for the full release. Was that not the case?
Was the workflow MT -> human translation?
If so thats honestly such a lazy followthru given the technical hurdles overcome wrt the font tech
From the Github repo (https://github.com/ExxistanceDC/Segagaga-English-Translation), the translation went through a process called MTPE (Machine translation, post-editing). This works just like it sounds. The initial translation is done with machine translation, then human translators review and edit the resulting translation to try to correct any mistakes.
> What I call the “playtesting translation” — a base translation that allowed the artists and playtesters to get started early and understand what they were working on — was developed using a combination of DeepL and ChatGPT 4o/4.5. That translation then went through a substantial, months-long human translator review. I don't think that the end product feels “machine-translated,” but that’s ultimately for you, the player, to judge.
> MTPE (Machine translation, post-editing). This works just like it sounds. The initial translation is done with machine translation, then human translators review and edit the resulting translation to try to correct any mistakes.
And the consensus among professional translators is that MTPE only saves time if you're willing to accept a half-assed result. For them to edit MT up to the standard of manual translation takes just as much expertise and effort as translating it manually in the first place.
> And the consensus among professional translators is that MTPE only saves time if you're willing to accept a half-assed result.
I have no particular interest in translation, but clearly when the person saying X is bad depends financially on you not buying X, you must take their word with a grain of salt.
> consensus among professional translators
[citation needed]
https://blog.gts-translation.com/2025/04/07/the-state-of-mac...
> 12.08% [of translators] say MTPE produces high-quality output.
> A significant portion (around 50%) of respondents do not offer discounts for MTPE work, arguing that post-editing can take as much time as traditional translation.
> Among those who do offer discounts, the most common range is between 10-30%.
Oh boo hoo.
Just boooooo.
Disappointed. Gif
As a side bar, I found it funny how in a Stanford lecture teaching various routes to training llms on the Machine Translation benchmark, the sample used was a French to English translation of 'the teddy bear is blue' or something similar.
After the lecture I reviewed the current production grade google translate...it butchered the translation
I wouldnt trust machine translation.
i don't think the translators had any first hand experience with akiba culture of that time period, nor did they know how to speak japanese to actually research it (and most of those ezweb.ne.jp / imode sites are long gone)
Is it possible for AIs in the future catch up those gags and meta humor in the language? Could the translation team have done a better job prompting the AI with the current technology?
Afaik japanese media (manga) abuse alot the use of puns on its language due to the amount of homonyms on it.
ai translations of japanese still miss the point and screw up / hallucinate articles and subjects between passages (switching genders, for instance, 3 times in one paragraph)
Avante-garde will always be lost on an inferential technology because to 'get it' it needs to be trained on sufficient volumes.
67 67 67 67 67 67 67 67 67 67
Edit: I may not exactly be spitting pearls, but passive downvoting for a genuine take on the matter based on professional experience in AI is a canary gasping for oxygen.