Not for anything involving art. Real humans flub shit often enough, I don’t trust an llm to convey any important context.
Comment on Crunchyroll & Now Netflix Both Confirm Anime Sub Watchers Are the Minority on Their Platforms
brown567@sh.itjust.works 3 weeks agoIsn’t translation one of the few actually effective uses for llms? Or am I remembering wrong?
Duamerthrax@lemmy.world 3 weeks ago
brown567@sh.itjust.works 3 weeks ago
It seems like my monolingual perspective led me to underestimate the complexity of translating! I have a lot more respect now for the people who do it
jacksilver@lemmy.world 3 weeks ago
As with most AI/LLMs it’s about how much error you can tolerate.
Translating basic info, sure the user can infer any slight mess ups. Translating jokes, idoms, tone, cultural differences, etc. is hard assuming you even got the original text right.
Unboxious@ani.social 3 weeks ago
Maybe that works if you’re translating large blocks of text, but when it’s small bits of isolated text where the viewer is expected to have visual context which the LLM won’t have it’s a hot mess. I tried experimenting with some translation software on a page of manga and it got a lot of stuff wrong.
brown567@sh.itjust.works 3 weeks ago
Makes sense!
GreenKnight23@lemmy.world 3 weeks ago
just watched Nagatoro call Senpai a bitch.
it’s lore now, but was totally NOT what she said.
goferking0@lemmy.sdf.org 3 weeks ago
Aren’t they only doing that so they don’t need to pay for the sub licenses?
Only reason for why they refuse to have subs on their dubs
nekusoul@lemmy.nekusoul.de 3 weeks ago
If you’re using it just to translate a few paragraphs of text on a website here or there, then yes, it’s much better than what we had before.
For anything complex however it can’t even begin to compare with a professionally done translation/localization. Japanese is one of the more difficult languages to localize due to a bunch of linguistic concepts that don’t translate well to other languages and need creative solutions that carry over the same intent.
More important however is consistency: Even if an AI translates some of the language ticks of the characters instead of completely glossing over them, it needs to do so consistently and apply the same translation across the whole script.
The same goes for any named items. If there’s a “Soul Stone” for example, you need to make sure to call it “Soul Stone” every single time and not “Spirit Rock”.
Khrux@ttrpg.network 3 weeks ago
I 100% agree on consistency, but I actually suspect LLMs would be pretty decent at translating linguistic concepts. Their grasp of contextual language is always surprisingly good, as seen in how well they rewrite tone to very specific parameters.
I don’t use ChatGPT loads, but I do use it a little, often alongside a method of project that can’t hallucinate. The other week I was talking in one chat about the vocal tone of the novel Blood Meridian by Cormac McCarthy, that has this laconic, worn down dialogue with old testament feeling prose. In another chat, I was prompting ‘sage advice’ for a matcha scone recipe I wanted to make where I was marbling strawberry directly into the dough. I got which chat I was responding in wrong and accidentally had it give me a the scone recipe in Cormac McCarthy’s writing style.
Now this is actually super easy to critique as a recipe, because it fucking sucks. In adaptation, it’s lost a lot of critical instruction. This is effectively a translation that must adhere to narrow and specific tone to achieve a meaning that cannot be translated without grasping a meaning to language that exists beyond the words.
Obviously this is English to English, but a big difference is that there is far more Japanese out there than Cormac McCarthy.
That being said, nothing cements what you’re saying about consistency more than how badly butchered the underlying instructions to this recipe are.
brown567@sh.itjust.works 3 weeks ago
True! Consistency is definitely something they aren’t great at