Factorio Forums

mmmPI wrote: Sun Jul 27, 2025 5:29 pm Unfortunatly, i cannot evaluate wether or not this is proper german langage due to my ignorance of the german langage. I can however tell that "Schiene" and "Gleis" are different words, and since it is the title of the article, it's not a good start. I suppose a richer glossary would have some value, and i suggest using the game's locale file, which you can find in the game's folder, with a bit of editing or maybe using another AI you can get a curated list fairly fast.

The translation and handling of markup is very good, only minimal editing afterwards required. Good except for "Gleis", which should be "Schiene", and as you already found out it's a thing to add to the glossary. That's what the glossary is for, and that's a regular task during translation and nothing special. In the past I used Deepl, and Deepl also has a glossary, but only 10 entries in the free version, so it's tedious to constantly swap in required entries and removing not required entries. Adding the whole game terminology to the AI glossary would be overkill, so I add stuff if I see there's some terminology not translated correctly.

mmmPI wrote: Sun Jul 27, 2025 5:29 pm However even when done "right" AI-only translation are currently far from the quality of human translation also done "right", which i believe is the case of the on the wiki, judging from french/english translations. It has a certain style of expression, that an AI could typically mimick, but it would end up doing a lot of non-sense that a human would need to curate. I feel this is the main point.

My point is that I as translator have a choice between two alternatives. Either spend a huge amount of time and provide a few selected articles in "human translated" quality and the vast majority of articles still untranslated. Or spend the same amount of time to get every article translated but with a slightly mechanical touch.

I went with the second option. I wanted to enable German people who cannot cope with English using the Factorio wiki. These people don't want perfectly worded articles, they want to understand the game! And I gave them this understanding. I got the whole German part of the wiki translated. A machine translator enabled me to translate 10 articles within an hour or two. This went on for about a year (I had too much time on my hands).
I cannot judge how good the French articles are and if they are machine translated or not, but German is translated very good and very natural by current machine translators.
Something not possible if I tried to do that manually. I started the whole thing by translating manually, but my own translations ended up with clumsy and really badly worded text that was clearly visible as "translated from English". Another hour and it would be good. One hour and more for just 1 article.

This would also mean burning out after a few articles. We see many wiki editors with just a few translated articles - they started with good intentions but burnt out fast. And I got the job done for 1.1. I'd like to do that for Space Age as well, but I don't have too much time on my hands any more.

And really, the machine translated text is so good, it's rarely required to edit significantly. It's not the robotic mess from 20 years ago any more. The most work goes into fixing up the wiki markup.

Tertius wrote: Sun Jul 27, 2025 6:51 pm A machine translator enabled me to translate 10 articles within an hour or two. This went on for about a year (I had too much time on my hands).
...
Something not possible if I tried to do that manually. I started the whole thing by translating manually, but my own translations ended up with clumsy and really badly worded text that was clearly visible as "translated from English". Another hour and it would be good. One hour and more for just 1 article.

It's real and it's true.

I was tempted to lead with a carpenter and a tablesaw. A dog is much, much cuter than a tablesaw. You won't find a carpenter without a tablesaw though.

My follow up was from a company that localizes things with AI professionally. It's been their job to localize things since 2017.

I have a concern because they indicated something slightly different than you have.

What you've said about this thing not learning, is there a specific hook or flag to turn that off? I know there is a local version, I presume it has a terminal command or a configuration script that prevents this? Online you lose that control, but presumably a corporation like google has something set up.

If there is no such thing... the wise thing is to find a friend that speaks English and German. That would be continuing with caution I think.

There's got to be someone who can in Germany or all the internet. And if you get them playing Factorio for a minute I think they'll get it. At worst, add someone to your friend list.

As for whether or not it ever got me anything meaningful. It got me a good grade and several won games of scrabble on my way to my Bachelors Degree. And it continues to give me some lovely images of... furniture. Yes. I'm looking at furniture. Beautiful drawers. You can say I've kept abreast of AI's developments and don't want it to go away. I know a carpenter. I can see how many fingers he has, know he's good.

coffee-factorio wrote: Mon Jul 28, 2025 2:26 am I was tempted to lead with a carpenter and a tablesaw. A dog is much, much cuter than a tablesaw. You won't find a carpenter without a tablesaw though.

Sounds like someone who walk on a rake, and take it in the face.

Tertius wrote: Sun Jul 27, 2025 6:51 pm

I've been off-grid a bit and was thinking a bit more, about the "dead angle", one thing you can consider as "a risk" is if you start considerng maintenance. The english wiki is containing the lenghier article from my survey on other langage. More pages, and more text on pages. This is also because there are more contributors i suppose. Now if you bravely task yourself of translating the whole thing in german, and you managed to do so all by yourself with the help of AI, that would surely be very impressive, but that would only be the "first step". Because then you are left with a situation where there is an "english wiki" and a "german wiki" with the same amount of things to update when changes are made but far few contributors to notice in german.

Whereas if the wiki's size is more accuratly representing the amount of time contributors can spend into it, it because it was only translated "slowly" , it increases the chances that "someone" , the person who've spent time doing the translation, "remembers" that some pages will need an update. There is less descrepency, i'm obviously exagerating when implying you'd be the only one doing german translation ^^

The extreme case where you push "automation" in translation "too good", then you don't even know what has been translated, you don't need to oversight the machine.

Another thing is if the translation is not "top notch", in one case it means the machine "failed", but if it's human-made it's easier to understand the imprecisions, and it makes it easier to participate or interact than when the "sum of content" is ginormous because machine-generated and all human changes feels insignificant touches that will get lost in the next machine-generated update.

Thanks again for sharing the script though =) I think it's a very good example of where you can use AI, the formatting, understanding the glossary, creating a draft , all your points are convincing i find, it's very functionnal, i'm sorry for the ramble but it feels like the necessary task. And if there was a need for very unrelated metaphor, it's like if you explain mechanic to people so they can fix their car, and you also need to tell them they can use this knowledge to create dangerously fast engines if they unthrottle their scooter or motorbike and they shouldn't use the repair knowledge for unsafe modification to their thing =)

Been using "internet" or "google" or "deepL" or dictionnaries for translations, sometimes i even send a text to an AI and ask if it's properly written, or to tell me if there are grammar mistakes and things like that, also ask what the text says, and if it's clear, and wether "this or that' is well explained, including traps with topics that are not mentionned to see if the AI is accurate x). Also asking Ai "what do you think a person will understand from the text ? " , and a follow up with "to me it means XXXX, do you think it's the same thing as what you just said ? what are the nuances" ? I will gladly use similar approach as the one you described in the future as a new one in the list

I'm excited how much AI-assisted translation improved these days. I vastly improved my prompt (and my ability to create AI prompts - see updated prompt in OP), and it's actually useful now. It understands context in combination with the glossary and it is still able to transform links to the {{L|...}} macro which is de facto obligatory in all translations and vastly different to the English pages. This was my original motivation for all this. Even huge articles like the Enemies article require just minimal rework afterwards. Previously, manually correcting all the links was the biggest time consumer (an issue with AI translation in general) and the most discouraging task due to its tediousness.

About AI: I learnt I need to "discuss" errors with the AI and point out errors. It will apologize and correct. It will reveal part of its "thinking" process. Previously, I assumed if I create the perfect prompt with the perfect ruleset, the translation will be perfect. I was furious the AI never got it 100% right on the 1st try in a fresh session. It always violated some rule. I discussed with the AI why it didn't adhere to rule X, and it explained, and I realized language is ambiguous and an AI is just lazy sometimes (sic!), almost similar to a human. A 100% perfect ruleset with every edge case is a program without flexibility, and translation requires flexibility.
And I learnt how to improve the prompt by requesting the AI to write the ruleset (including discussion results) as new prompt. So I see how the AI understands the prompt and I was able to improve inaccurate rules.
An incredible tool.

Time per article from taking the English original to posting the final translated and manually reworked article:
manual:
trivial: 10 min, regular: 1-2 hours, complex: 1-2 days (and you don't do complex articles more than 1 per month due to exhaustion)

AI assisted:
trivial: 2-3 min, regular: 5 min, complex: 20 min

And that's what I've learnt from this and will take with me: how to give instructions to an AI so that it does what I want.

Tertius wrote: Mon Apr 27, 2026 11:12 am About AI: I learnt I need to "discuss" errors with the AI and point out errors. It will apologize and correct. It will reveal part of its "thinking" process. Previously, I assumed if I create the perfect prompt with the perfect ruleset, the translation will be perfect. I was furious the AI never got it 100% right on the 1st try in a fresh session. It always violated some rule. I discussed with the AI why it didn't adhere to rule X, and it explained, and I realized language is ambiguous and an AI is just lazy sometimes (sic!), almost similar to a human. A 100% perfect ruleset with every edge case is a program without flexibility, and translation requires flexibility.
And I learnt how to improve the prompt by requesting the AI to write the ruleset (including discussion results) as new prompt. So I see how the AI understands the prompt and I was able to improve inaccurate rules.
An incredible tool.

This is nicely put ! It makes it so obvious when reading it this way !

It made me think though that sometime the AI apologize and correct their output to please the user, they tell what you want to hear, even if it's not correct, depending on how stubborn you are in the discussion

I was nitpicking, and in its thinking process the AI commented to itself along the lines "...this person is quite difficult...". I was complaining about rule X being violated, and I remarked the AI is a machine and machines are able to produce 100% exact results and demanded a reason why rule X was still not followed. It made me laugh, because I started to discuss something with a soulless machine like it was a human.

But this is a strength of the tool, because in the thinking process it became obvious what things it did in addition that I didn't mention in the rules myself. This improved the result. You have to start with a good prompt, but you also need to let it correct the things it didn't do right in the first try.
Everyone who uses AI in their daily work will probably know this already, however for me it's just occasional use, and my typical AI queries are just simple questions that I used to ask Google search. This will probably use just 1% of what an AI can do for you.

You can shape the general answering style. If you find answers submissive and don't want estimates and hallucinations, you can configure your global prompt-prefix to specify how answers should be in general. I requested exact answers, and if the AI doesn't know an answer it should tell about it and not start inventing things and hallucinating. This works reasonably well. You have to still verify answers but it's better this way. In older copilot/chatgpt versions I also requested to stop being enthusiastic, but this ceased on its own to an acceptable level.

These prompt prefixes actually have quite an impact. One day I asked the AI to generate an outline of a small adventure for my D&D pen&paper roleplaying group and forgot to remove my standard technical prefix. The resulting outline was quite strange and shallow. Only after I told it should forget the standard orders and instead switch to creative mode and encouraged inventing things it produced a quite nice new adventure story.