This fairly anecdotal, but LLMs aren’t that good at translating Japanese. From what I’ve seen they are pretty good with many other languages, but they really struggle with Japanese → English translations, possibly because the languages are so very different. I have yet to see a machine translation of more than a couple paragraphs of native content that didn’t make huge mistakes on the nuances, and a lot of times there isn’t even a straightforward translation from Japanese to English. The first example that comes to mind is the “Watashi… Watakushi? Boku? Ore?” from Your Name, which, using a direct translation, just means “I… I? I? I?”.
Comment on Crunchyroll Accidentally Reveals They've Been Using ChatGPT for Sub Translations
KairuByte@lemmy.dbzer0.com 4 days ago
While this is some bs from a worker perspective, this is also one of the main places LLMs thrive: translation.
If it weren’t for the fact that they are sidestepping the cost of a translator (and subsequently costing someone(s) a job), I’d actually be not appalled by this.
Grumpy@ani.social 3 days ago
Hugin@lemmy.world 3 days ago
It can’t be worse than some of the netflix translations.
I just watched Apothecary Diaries with the wife. The subtitles are basicly fanfiction. I know characters reply wakarimasu a lot but you don’t need to make up dialog. The new stuff often conflicts with the actual story.
glilimith@lemmy.blahaj.zone 4 days ago
LLMs should be good for translation, but the thing is that they aren’t, for two main reasons:
First, the avoidable problem: CR is running them on the audio, not the script. The AI is mishearing stuff all the time, leading to inconsistent character naming and nonsensical sentences. There are a few CR anime this season that are extremely hard to follow at times because the AI subtitles can’t even begin to understand what’s happening.
Second, a more permanent problem: translation of fiction is much more of an art than a science. Translators don’t just translate words, they translate connotation and tone, as well. A translator interprets the meaning of the lines - what information is being conveyed, spoken or unspoken - and tries to put that same meaning into a different sentence. This is a task that LLMs are going to struggle with a lot. For example, how is the AI going to convey that there’s been a pun, or a reference to a well-known turn of phrase, or an innuendo using words with multiple meanings?
Unboxious@ani.social 3 days ago
There’s also a third problem: visual context. Someone translating a manga or anime would normally be able to see what’s happening, which can be essential when translating certain otherwise-ambiguous lines.