CaesarDX
Newbie
- Jul 17, 2020
- 27
- 4
The fact it's like Math means AI is actually quite well suited for transaltion.
LIke you said either you're right or you're wrong.
What chat AI is at the end of the day is essentially advanced pattern recognition and using that to predict what response is "correct".
This results is made up info when you ask it for new information. But asking it to simply recognize patterns in language and translate accurately is exactly the king of thing ChatGPT should be good at.
However to do it truely effectively you'd need to optomize the model for that and restrict it to not get too creative and only offer a translation.
And no one's made that yet as far as I know.
It should absolutely be possible but if anyone's working on it they haven't released it yet.
CaesarDX
Newbie
- Jul 17, 2020
- 27
- 4
Maybe ur file was corrupted, u can try download once again. But if u still got same issue try uninstal the japanse local and instal again for All users
Ferghus
Engaged Member
- Aug 25, 2017
- 3,059
- 4,706
Language isn't literally analogous to math. It was the closest thing I could compare it that to could be universally applied since anyone with up to high school education encounters algebra. Like if you've taken any foreign language class, teachers will always tell you that the shit you're learning isn't actually how anyone in that native language speaks.
But if you examine English itself, you'll find that it's very inconsistent in what's considered correct and incorrect, like how [its] is possessive even though you generally need to use ['s] to denote possessiveness in other subjects, or how 11 isn't firsteen or 12 isn't secondteen, but these are conventions that we learn. There's a reason why foreigners commonly complain about how hard it is to learn English. The issue translating something to English is that anything not-exact matches go both ways.
AI wouldn't be able to, for example, draw an appropriate analogy to abstract concepts like wordplay, or recognize anything that requires common knowledge of the country or origin (or at least an extensive Google search), or even ironically, mimic foreigner-speak. You take all that away and what you get is guesswork sprinkled between things that actually make sense, which could change entire contexts. You'd be hiding a tree in a forest, so to speak. AI can't produce results comparable to abstract knowledge with just pattern recognition alone. The sheer amount of variability would make it impractical to try to input every possible instance. To go back to the math analogy, it would be like trying to map every result for every calculation, which would bloat the shit out of your computer. But again, translations aren't 1:1. Even within the same language, A+B may not necessarily equal B+A. "I hit the ball hard" is not the same as "I hardly hit the ball."
dikdik48
Well-Known Member
- Apr 5, 2020
- 1,312
- 1,170
Actually you'll find human speach is ... fairly predictable with enough data. EVen with wordplay someone's already probably done that wordplay before so an AI with enough data on that could handle ... a good bit of it.
I mean you bring up the "it's" issue but that's litterally just a rule of English we tend to find confusing but is a predictable pattern.
Same with 11 ... YOU find that confusing but an AI fed the proper data would just go "Oh sure this is that". Again we're not talking making new data with no guide, just a translation which REMOVES the guesswork rather than adding more.
Same with requiring common knowledge. You could litterally feed all data on a culture to an AI and make it more knowledgable on the subject than an actual Japanese person.
You're also over selling variability. Yes some words can mean multiple things in other languages. But determining the right meaning comes down to context of the overall text and not pure guesswork. If it is down to guesswork generally multiple variations are equally acceptable.
Even then if you feed the AI with examples of "this is Japanese ... and this is how we would translate that exact text to English" you could teach the AI the guesses we prefer to make.
This is litterally how these AI models work with weighted probabilities to use the arrangement of words to predict what response is most correct.
Now I want to be clear that a True, quality AI translator needs a personalized model with data focused on language, culture and translation examples.
BUt give it all that and it should do AT LEAST an okay job at it. Sure the translation might lack some creative flair, it might mess up translating some idioms and clultural phrases that were maybe phrased creatively so the AI missed the connection, and it might overall not be AAA perfect.
But if the only mistakes are some missed idioms then someone could just proof read and fix that. As long as most of it is translated well without mistakes in grammar or phrasing it's still a massive improvement over current MTL.
Got a fix for the Aya scenes (Fellatio2, Missionary2 and Missionary3) and the Shiho Video Voyeur menu, unzip and replace the Patch.xp3 file in the game directory. If there are any other crashes like this, let me know which scenes they are as it's an easy fix.
You don't have permission to view the spoiler content.
Log in or register now.
Last edited:
Ferghus
Engaged Member
- Aug 25, 2017
- 3,059
- 4,706
It's predictable because humans have critical thinking skills. We don't just mindlessly accept information, we draw conclusions from the information and apply our experiences and knowledge to rationalize why it can or can't be something else, which is beyond the capabilities of current models of pattern recognition software. You can technically map all the possible combinations, but that would as impractical as indexing every grain of sand to count it. You notice how your computer slows down the more files it has? What do you think would happen if a language program had a bunch of redundant and unique entries?
You ... REALLY don't understand how AI models work. First of all they work with data values in the billions easily. If not more. That's why these models run on servers and not really local, or you need strong as hell computers to run even lower end models.
You're also trying a bit too hard to be philisophical.
We don't need some bullshit about containing the entierty of human rationality. The fuck do you even think this conversation is about? It's translation dude don't be going crazy. You don't need to contain the entirety of human conversation to make these models work. Again the way they actually work is ... well complicated in detail but the short version is each word is given a token and the combination of these tokens is weighted in such a way the AI predicts what response should be given next.
In ths case of generating conversation from scratch it's already able to immitate people fairly well. It does have some flaws and some repetition that show it's weak points. But considering it's able to to do that much just from given input is impressive.
But with Translation the model wouldn't need to predict from scratch and would mainly just need data on what words translate into, use context that it can analyze to help when words have multiple meanings, and use it's grasp on gramar structure to handle the issue of sentences being structured differently in other languages.
You're talking like you're expecting the AI to be able to actually percieve the meaning of conversation ... but that's absurd and not really needed.
You don't need to be able to read author's intent for most translation unless it's seriously 50/50 on how to translate a phrase. But again most of the time that means either assumption works.
And FYI someone's already tried using ChatGPT as a translator with ... mixed results but better than Google Translate in terms of readability. Keep in mind this was a chat model being forced to work as a translator a Model and program designed for translatioin would be far more efficient and accurate.
Of course I do want to make it clear, there is a good chance even with a good translation focused AI the translations will not be perfect. Idioms or phrases that the author gets creative with may not translate well without real critical thinking skills and some words that the AI sees as likely to translate one way are meant to be another way. But with a good AI to accurately read most phrases this won't be much of a problem most of the time.
and again proof reading can always help catch those mistakes. It's not like fan translations are always great either. I've seen plenty of times they choose what feels like clearly the wrong translation. BUt somehow they didn't catch it.
Ferghus
Engaged Member
- Aug 25, 2017
- 3,059
- 4,706
At this point, I think we have very different ideas of what a reliable AI translation is supposed to look like and I don't think we're going to come to an agreement.
Thanks, I've updated the fix.
See this post for fix: https://f95zone.to/threads/home-guard-2-final-beel-ze-bub.170974/post-11513728