Today I want to clarify the nomenclature a bit.
A dear reader of the list,Andreas Lopez, responded to yesterday’s email:
My wife is a linguist and fluent in 4 languages (Spanish, English, French, and Japanese). She is actually thoroughly disappointed in the LLM translations because too often they ignore cultural relevance and dialects and mix and match them which can lead to confusion for the people you are making it – the localized versions.
Especially in interactive fiction it’s not about actual translation but localizing it. We have many ‘odd’ phrases in Germany we exchange every day but a different phrase reveals your locale or homestate.
Andreas refers to the fact that translating is not localizing .
There is a difference between putting a text in another language and capturing the essence of that text to reproduce it with local forms and manners.
This is a hard problem to solve, and most game translations suffer from this issue at different levels, even when done by humans.
I will “ignore” this problem for now, because I think this is a problem that LLMs will solve eventually.
I may be wrong, because I’m not aware of the computational complexity of this problem, but I think it might get solved with future LLM generations.
I would keep my hands off of LLM just because the risk of pissing people off with bad writing is just too big for my personal pride and risk levels.
As an indie with $0 budget you might be better off asking for volunteers or just accept you might piss people off due to insensitive translations which you won’t be able to check unless you know the language you have the machine translate into.
Yes. But also maybe not?
You can’t release a game translated with LLMs just like that. That’s a given.
Getting volunteers is not always feasible, especially with games with high wordcount.
The readers of the list can guess my approach: playtesting.
You should never release anything without it being tested, and translations are no exception.
In the next emails, we will discuss some approaches for localization using LLMs (or not).
