VN ChatGPT translation is great, but it always merges lines which makes it unreliable for patches... How to prevent that?

[Batenxx]

So I'm working on a ChatGPT translation of an eroge, however the big issue is that ChatGPT no matter what merges lines. I can fix that somehow by making the lines a numerical list, however then it's being treated as separate strings and translated without context and proper flow. Does anyone have a better solution?

I tried telling ChatGPT to leave the omitted line empty if it merges it. This way I could fix it manually in a spreadsheet, however it told me that it's incapable of doing this. Does anyone have any idea? I don't want to feed ChatGPT line by line either as the beauty of it is to have proper context and that is only possible by feeding it entire blocks of text.

 

[RedEye781]

You can feed it line by line and it will 100% understand the context based on the previous line, however as good as it is, it's still an idiot bot. First, you got the fact that no matter how you write your prompt it will still be mistranslated into a milder version.
The line could say "I groped her tits." and GPT will give you, "I gently kneaded her breasts." Second, god help you if you give it a metaphorical line or something that uses onomatopoeia as dialogue like 「ひゃ....ひぅぅぁぁぁ!　ビリビリが走ってぇぇぇ....あ,　ああ!」 GPT will give you "Hyaa... Hiuuuahh! Electricity is running through me... Ah, ahhh!" instead of a more correct "Hyaa... Hiuuuahh! My body's tingling all overrrr... Ah, ahhh!"

If all you're doing is feeding GPT lines and expecting good translation, then sorry, you'll never get that.

As for your main issue, maybe try to write a prompt with clear instructions, such as, for each 「」 start on a new line or something like that.

Personally, no matter how I tried to use it in full sentences it always failed me at some point.
I am, of course, talking about 3.5, 4.0 is a lot smarter and fixes a good 80% of 3.5's issues, so break out that wallet if you want the best result a bot can do. Alas, both still failed that line without clear instruction from me, even though I was feeding them the script line by line.


TL;DR ChatGPT is meh, not great.

 

[Batenxx]

Until now I used some free site which used GPT 3 and it even translated h-scenes accurately and rarely merged lines. But that page doesn't work anymore since few days ago. It was

You must be registered to see the links

(warning, it's full of ads right now, no idea what happened there) Now I'm done with 3.5 out of 5 routes of an entire eroge and I'm stuck. Can't use the official ChatGPT for the above reasons. Tried Bing which supposedly uses GPT 4 but it's not working the way I want either. Damn...

 

[RedEye781]

Why not use a site like this one?

You must be registered to see the links

They let you build prompts and use multiple different AIs. Spoiler GPT 4 is the best one.
3.5 is the second best.

Oh, fair warning the site has issues currently, so wait a day or two.
There are others like c.ai or Poe, but I've never used them.

I'm not sure what you mean by Bing, are you referring to the search help feature that uses 4? It's only based on GPT-4 and it's designed to search not translate lol and it has the same BS no porn rule so...

 

[xorxorrax]

So I'm working on a ChatGPT translation of an eroge, however the big issue is that ChatGPT no matter what merges lines. I can fix that somehow by making the lines a numerical list, however then it's being treated as separate strings and translated without context and proper flow. Does anyone have a better solution?

I tried telling ChatGPT to leave the omitted line empty if it merges it. This way I could fix it manually in a spreadsheet, however it told me that it's incapable of doing this. Does anyone have any idea? I don't want to feed ChatGPT line by line either as the beauty of it is to have proper context and that is only possible by feeding it entire blocks of text.

From my experience if its saying "I can't do that" its usually hitting the content filter, try removing some lines to coax it into translating.

Other things you might want to try to keep the lines separate:

• Format input/output as a json document.
  

  [ {
  "n": "0",
  "id": "C00000000",
  "msg": "――広瀬家の朝は、いつも包丁の音から始まる。"
  },
  {
  "n": "1",
  "id": "C00000001",
  "msg": "「……そろそろ、お父さん起きてくるかな」"
  },
  {
  "n": "2",
  "id": "C00000002",
  "msg": "僕は、素早く仕上げを済ませてしまうと、それぞれの料理を皿に盛り付け、テーブルに配膳していく。"
      },...

• Use some verbose separator like ---------
  

  "――広瀬家の朝は、いつも包丁の音から始まる。"
  -------
  "「……そろそろ、お父さん起きてくるかな」"
  -------
  "僕は、素早く仕上げを済ませてしまうと、それぞれの料理を皿に盛り付け、テーブルに配膳していく。"
  -------
  ...

I also have the merged lines issue even with these so trying again with different/more or less lines helps.
Depending on your prompt I recommend sticking to 15-30 lines, with more lines merges start being really common.
Good luck, don't think there is any way to fix this issue 100% with only prompt engineering.

 

[Batenxx]

I'm not sure what you mean by Bing, are you referring to the search help feature that uses 4? It's only based on GPT-4 and it's designed to search not translate lol and it has the same BS no porn rule so...

Surprisingly the Bing solution reads the best. It is less over localized than GPT3.5 but omits characters talking about themselves in third person if it's clear they talk about themselves through context. It also doesn't merge lines 90% of the time, yet maintains a proper conversation flow.

For h-scenes, for now I'm using:

You must be registered to see the links

It's based on a Chinese open source chat AI, however the output is rather accurate and eloquent.

 

[Batenxx]

Why not use a site like this one?

You must be registered to see the links

They let you build prompts and use multiple different AIs. Spoiler GPT 4 is the best one.
3.5 is the second best.

Great page. I had an idea of making a prompt that numbers the lines. Whenever the line number is lower (so if lines are being merged) or higher than the expected number I would simply Regenerate. However after the second block of 50 lines to translate it forgets the number counting. And worse yet, it doesn't respond to any prompt afterward. Dumb AI.

 

[RedEye781]

Great page. I had an idea of making a prompt that numbers the lines. Whenever the line number is lower (so if lines are being merged) or higher than the expected number I would simply Regenerate. However after the second block of 50 lines to translate it forgets the number counting. And worse yet, it doesn't respond to any prompt afterward. Dumb AI.

It's not the AI's fault that you're asking it for something it wasn't designed for.
Remember it wasn't designed as a translation AI. It can only do this as a by-product of its ability to chat in different languages.
GPT-4 which was released after 3.5 based on the data 3.5 had interacting with users, is inherently a lot better. However, it's still not a capable translator and it certainly can't do bulk translation. Sugoi is actually better in the aspect of bulk translation, alas that thing is also dumb.

My conclusion remains the same. If you want the best result then buy GPT-4 and feed it line by line. I'm of course assuming you have a basic level of Japanese. If you don't, then the bots might as well talk in moon speak for you.

 

[Batenxx]

Sugoi is actually better in the aspect of bulk translation, alas that thing is also dumb.

But Sugoi doesn't translate the lines in context like Chat GPT does, no? From my understanding Chat GPT does it more similar to how a human would translate, it basically understands the story and content, which makes it more accurate to translate fiction.

 

[RedEye781]

But Sugoi doesn't translate the lines in context like Chat GPT does, no? From my understanding Chat GPT does it more similar to how a human would translate, it basically understands the story and content, which makes it more accurate to translate fiction.

Sugoi can definitely keep up based on context it just has lots of problems with pronouns and being too literal.
While chatGPT is better at making the thing sound natural, it does not mean it's the correct way to phrase the lines.
I gave you a perfect example of what chatGPT would do, to a metaphorical line.
There's also, and I can not stress this enough, it's fucking polite as fuck. Which given that we're translating porn here, is a very big problem.

Last time I checked all Japanese porn has a dash of rape in it, stemming from the fact that women aren't supposed to show enjoyment in having sex.

Good god, just try and do one of those games that make the woman into a cum-dumpster or an NTR that's non-consensual at first. You'll see my point then.

 

[RedEye781]

To give you a current example.

「すごく....すごくおまんこ擦られてぇぇぇ....Hee,ひぁぁぁぁぅぅ!」

GPT: "So intense... being rubbed so much... Hee, Hiaaaaaah!"

Correct: "So intense... my pussy is being rubbed soooooo much... Hee, Hiaaaaaah!"

It literally deleted the word おまんこ from the line.