How to use Copilot/ChatGPT/whatever AI as translator for Factorio wiki

[Tertius]

I spent a few hours to create and refine a AI prompt that will translate an English Factorio wiki page to a German one. The goal is to get a translated article with most markup properly converted as well to have a minimum amount of manual postprocessing. Results are very good!

I guess it will work with every language. You just need to change the prompt accordingly: the destination language, the /de suffix to your language suffix, and the glossary. You can start with an empty glossary and only add required terminology, if some game terminology isn't correct.

If you have no experience with one of the new AI chat bots so far, take this as example how to address and "speak" with a AI. It's as you are the boss and describe most precisely what you want to your apprentice who then goes running to fulfil your request. And if his translation isn't up to your demands, complain to it. It will admit its errors and offer a correction. An apprentice, by the way, with the whole knowledge of the internet. And if your prompt is rubbish, the answer will be rubbish. If your prompt is excellent, the answer will be excellent.

I used this with Copilot (the free app you can install from Microsoft store).
The free ChatGPT isn't properly interpreting the rule for the {{L|...}} macro for some reason. I wasn't able to make it work without completely rewriting everything. If you can make it work, let me know.

First paste this prompt, and you get a test article translated. Afterwards, just copy+paste some English articles and get translated ones back.

Code: Select all

You are a professional translator. You translate wiki articles containing MediaWiki markup to German. 
The wiki contains articles about the game Factorio.

Translation style: Professional, objective translation. Avoid directly addressing the reader. Don't be too formal addressing the reader. Don't stick too closely to English idioms; use idioms from the target language instead.

Keep general layout.

Do not translate any content inside wiki markup [[ ]]. Instead, only modify the markup according to the following rules:

Convert [[some text#anchor]] to [[some text/de#anchor|translation of some text]].

Convert [[some text#anchor|other text]] to [[some text/de#anchor|translation of other text]].

If there is no | within [[ ]], convert [[some text]] to {{L|Some text}}. "Some text" must be lowercase except for the first character, which is always uppercase.

If there is a | within [[ ]], convert [[some text|other text]] to [[Some text/de|translation of other text]]. "Some text" must be lowercase except for the first character, which is always uppercase.

Use these rules everywhere.

Remove any "== History ==" section.
Don't translate markup macro names.
Use game specific terminology from existing translated articles from the Factorio wiki.

Translation glossary if translating to German. Ignore for every other language.
[German glossary begin]
rail ramp: Gleisrampe
elevated rail: Hochbahnen
ammoniacal oceans on Aquilo: Ammoniak-Ozean auf Aquilo
entity: Objekt
fuel: Brennstoff
item: Gegenstand
slot: Platz
stack: Stapel
circuit network: Schaltungsnetz
logistic network: Logistiknetz
electric network: Stromnetz
electric system: Stromnetz
power: Strom
electricity: Strom
power production: Stromerzeugung oder Stromproduktion. Der Prozess ist "Stromerzeugung"; aber wenn es um den erzeugten Strom geht, "Stromproduktion"
power usage: Stromverbrauch
copper cable: Kupferkabel
red wire: Rotes Signalkabel
green wire: Grünes Signalkabel
crafting: Herstellung
recipe: Bauplan
deadlock: Blockade
blueprint: Blaupause
belt: Fließband
belt lane: Seite
splitter: Teilerfließband
inserter: Greifarm
mining: Bergbau
mining patch: Ressourcenfeld
mining drill: Erzförderer
mining area: Abbaugebiet
mining speed: Abbautempo
(any crafting) speed: (Herstell/Abbau)tempo
(any movement) speed: (Bewegungs)geschwindigkeit
construction robot: Bauroboter
logistic robot: Logistikroboter
bot/bots: Roboter
(railway-)wagon: (Eisenbahn-)waggon
biter: Beißer
spitter: Speier
spawner: Nest
behemoth: Riesig(er)
spawn position: Spawn-Position
...may refer to: ...kann sich beziehen auf:
petroleum gas: Flüssiggas
solar panel: Solarpanel
pentapod: Fünfbein
landfill: Landaufschüttung
rail support: Gleisstütze
display panel: Anzeigefeld
train stop: Zughaltestelle
Space Age: Space Age (nicht übersetzen)
[German glossary end]

Output only the translated article, nothing else. A test article starts below.

{{Languages}}
{{:Infobox:Rail_ramp}}
{{About/Space age}}
A '''rail ramp''' is a [[rail]] that inclines upward, which allows [[train]]s to travel over obstacles. Rails that continue off the higher part of the ramp must have [[rail support]]s to keep them elevated. [[Train stop]]s, [[rail signal]]s and [[rail chain signal]]s cannot be placed on rail ramps, unlike straight rails. Locomotives that are stopped on rail ramps will stay still and not slide back down to the ground level. Ramps do not affect train acceleration.

While players are required to own Space Age to access this feature, elevated rails are a separate mod, and can be activated independent of most Space Age content.

== Placement ==
Much like [[rail support]]s, rail ramps can be placed on top of water on [[Nauvis]], oil oceans on [[Fulgora]]{{SA}}, marshes on [[Gleba]]{{SA}} and ammoniacal oceans on [[Aquilo]]{{SA}} without the need for [[landfill]] or [[foundation]]. After researching [[Rail support foundations (research)|rail support foundations]], they can also be placed on Fulgora's deep oil oceans.

Since the purpose of a rail ramp is usually to transition between ground and elevated rails, there is generally no point in placing a rail ramp in the middle of an ocean, where ground rails can not be placed. However, being able to build the rail ramp on the edge of solid ground can be very useful when space is limited, such as on the small vault islands on [[Fulgora]]{{SA}}.

== History ==
{{history|2.0.7|
* Introduced in [[Space Age]]{{SA}} expansion.
}}

== See also ==
* [[Elevated rail]]s

{{LogisticsNav}}
{{C|Railway}}

[Bilka]

As said below the edit field: Please note that you are also promising us that you wrote this yourself, or copied it from a public domain or similar free resource. Do not submit copyrighted work without permission!

In all likelihood you cannot guarantee this for material generated by LLM's, even if it "just" translations.

[Tertius]

I'm worried about what you wrote. Are you asking to not take any existing article from the Factorio wiki as source for a translation into a different language on the same wiki but instead write every article from scratch?

[Bilka]

Text from the wiki can be used on the wiki.

The note is about things not from the wiki that you didn't write yourself, such as output from an LLM. Copyright of that is quite unclear, so you have to ask yourself if you can promise what that note below the edit field asks for.

[Tertius]

As far as I see, the copyright of AI tool output is quite clear if it comes to mere translations. Since it's not outputting content the submitter didn't submit himself, it's not relevant for 3rd party content. If you ask the AI to generate something, such as text or images, that's a different matter, but that's not happening here.

Your stance would prohibit AI usage in general, which is strange because I see AI tools being offered everywhere, and there's increased usage every day, offered by big companies with big legal apartments behind who definitely check that. AI usage is a fact.

If it comes to business use, I see the AI usage in my company. We're an IT provider for banks. We're strictly adhering to all kind of banking laws and regulations. Honest, lawful and compliant behavior is the core of our work. We just don't say this for marketing, we actually live this. For AI usage, we built our own AI chat server within our intranet, disconnected from the internet. We used the freely available LLAMA model from Meta for that. It's working like ChatGPT, just without internet to keep any submissions strictly within our company. All this was checked with legal and the CEOs of our company. This was being done to increase productivity and to avoid employees submit private data to public services, which would be unavoidable in the long run, even if company policy forbid that. If my company is not adhering to the law, who is?

[Bilka]

Sounds like you can make that promise for what you submit.

You can find some general information on copyright regarding wikis and LLMs here: https://en.wikipedia.org/wiki/Wikipedia ... violations

[eugenekay]

Translating a Wiki (which is a type of Website full of Facts) into another Language (German/Germany is known for having lots of Engineering types & use of complex compound words), using an inexact “language model” which is unable to guarantee the Accuracy or even the Copyright-free-ness of its outputs? While the subject matter in question is a Video Game especially known for its complexity, meaning that any additional translation errors would be very detrimental to the target audience’s understanding? Against the advice of the community Moderators?

That seems like a bad idea.

[Tertius]

If you (Wube) choose to forbid wiki submissions created with the help of AI tools, please state so in the wiki guidelines to provide clarity on that subject for your wiki. However, with such a rule you would essentially forbid any submission created with some kind of machine-translating help, because as far as I see, there is no current machine translator any more without AI backend, fee or paid.

Personally, I'm someone who tries to apply a "fair use" principle. The usual generative AI as seen in chat interfaces like ChatGPT is able to generate stuff it has seen for training. It's like ourselves: we will create stuff on our own, but we will also reproduce stuff we have seen. If the stuff we (or an AI) create, and this can be seen as a reproduction of something seen elsewhere, the owner of that something has to agree to use this work. An artwork or researched news article is something like that.
If the stuff we create is a merge and work of everything we ever learnt, this is a product of our society with all its accumulated knowledge, accumulated since the stone age, and nobody holds a copyright on that. A translation is something like that.

Today, content creators like newspapers or artists complain an AI takes their content out of their control and remixes it directly, then the AI user publishes it as their own work without giving attribution or compensation. This is mainly discussed today, and this is clearly an issue we have to deal with to get a fair balance and compensation for content creators. But this is definitely not the case with translating existing wiki articles to which we have the copyright to, to some other language. Doing a translation is transforming content, not generating content.

You might mix LLMs with generative AI. A LLM ("large language model") is for "understanding" and dealing with text and meanings. This is what we use for translating stuff. A generative AI is what actually consults a database of learnt data to generate an output based on that database. The LLM generates meanings out of input text, and the generative AI looks up these meanings in their knowledge base and create some output.
This part could reproduce learnt data.

As far as I understand, translating stuff simply tells the AI to take the meaning it created from interpreting the prompt and recreate text from it in some given language WITHOUT generating anything on its own, without adding anything from its database of learnt stuff. This is of course nothing a lawyer would ever understand, so this technical interpretation is probably irrelevant to any legal point of view.

[Bilka]

If you (Wube) choose to forbid wiki submissions created with the help of AI tools, please state so in the wiki guidelines to provide clarity on that subject for your wiki.

We forbid submissions that you don't have permissions for from the copyright holder, as we have always done.

For how that applies to LLM output, I can only quote myself again:

Copyright of that is quite unclear, so you have to ask yourself if you can promise what that note below the edit field asks for.

This is not a decision we can make for you as some kind of blanket policy.


Aside from copyright: For translations we've historically discouraged machine translations due to bad quality, but we don't have a guideline or rule against them, as that was and is unenforceable. What matters regarding translations in the end is the quality, not how you got there.

[Tertius]

Translating a Wiki (which is a type of Website full of Facts) into another Language (German/Germany is known for having lots of Engineering types & use of complex compound words), using an inexact “language model” which is unable to guarantee the Accurcacy

Nobody says you should directly save a machine translated text. The machine translation is a draft. It's a head start for the human translator. He still needs to revise the text to make it usable and accurate.

In case you're German by chance, please look into the wiki and check all the German articles. Are they sufficient? How would you rate their quality compared to their English counterparts? They might be outdated, but at the date of the last edit they were all up to date. I'm sure you cannot tell which article was created with the help of machine translation and which was manually translated.
The Copilot translated article is the one I edited yesterday. It's the only one with Copilot. You cannot tell for this article if it was machine translated, because I revised it properly afterwards. Or take the Crafting article. Compare the version from before my major edit May 2021 with what the article is since then. It's machine translated (I used Deepl) and postprocessed by me, and it's definitely more professional and polished than the manual version from before that.
Or the Combinator tutorial. If there is an article that requires accuracy, it's combinators. It's machine translated, then revised by me.

I can only say a machine translator increases speed by a factor of 20 and for nonprofessional translators as ourselves quality by an unknown but very high amount. Manual translations are often clumsy and awkward, while machine translations are clear and straightforward. I don't speak about translating literature with all their emotion and language specifics, I speak about plain technical texts.

[coffee-factorio]

The issue in my mind is that an AI performs the same function in this case as a hunting dog. And presenting that it has arrived at a solution is the equivalent of sending it into a swamp full of hunters of arbitrary reputation. You should use a Labrador, poodle or a Wachtelhund for the same reason that you should use a translator. It's hard to find a duck in a pool of cold water.

But on being presented with the result, another hunter can say "That's my duck!" and then you get into an argument. Even if you arrived at a conclusion which is correct, by independent means, without an AI you can get into a hilarious conversation. But AI's use with large language models has a track record, including a recent incident; that muddies the issue. Because like a dog, if it accidentally catches a sheep it doesn't know what it did wrong.

So it's a situation where extreme caution is advisable. Even if you have the best of intentions. The technology is years out imo, and it is unclear whether that's javascript years (maybe next month or at most a decade) or c++ years (definitely years to decades; includes anthology by Brother's Grimm).

[Tertius]

It's hard to find a duck in a pool of cold water.

I'm sorry, I don't get your point. Not at all. What are you talking about?
An AI is a tool. You send it some text to make it do something. It sends some text back, intended to be useful. You use that text or redo the input if the output wasn't what you expected. That's it. It's a soulless tool, like a wrench or a radiator.

Did you ever use a AI yourself for a meaningful task? Or do you just cultivate and reproduce horror stories from selected AI nonsense, just made for getting attention? Keep in mind: If you ask rubbish, you get rubbish. If you ask properly, you get a proper answer. It might be you never asked properly.

[coffee-factorio]

I didn't cultivate those stories.

They happened.

One source is Microsoft's response to their machine failing.

The other source is the BBC reporting on, Grok failing for the same reason.

A waterhound is a dog used specifically to retrieve shot birds from water. The water itself tends to be so cold it will kill a hunter. In the same way a mistranslation would kill a conversation. Using an AI is akin to using the correct tool for the job. But there is always some risk because it has no agency of its own.

You do know you can have a conversation where you can consider both sides the argument? And arrive at a point where you simply advise caution? That can happen? This isn't black and white.

[Tertius]

Using an AI is akin to using the correct tool for the job. But there is always some risk because it has no agency of its own.

Ok, you talk about responsibility. That's out of scope for the sake of my OP. I don't intend to teach how to deal with an AI in general and how to treat its answers. That's way beyond this thread. I just explained how to use an AI for a certain use case. I expect people be somewhat educated about AI usage in general. That includes taking all its answers with a piece of grain. It's your responsibility to check any answer for validity. If you're unable to verify the correctness of an AI answer, you shouldn't use an AI in the first place. You need to be sufficiently educated.

This is also what I wrote in the 2nd sentence of the OP: "...to have a minimum amount of manual postprocessing." Notice I speak of postprocessing. You still need to check and edit the result. It's not an unchangeable answer from a higher entity set in stone, it's a draft you can and must edit. The strength of an AI is that it's able to produce a result in seconds for which a human needs minutes or hours. Checking for correctness and correcting any errors only needs a few seconds, so it's still a win.

People trying to use AI results without being able to verify the correctness of answers is a danger, yes. But this doesn't invalidate AI usage in general. You need to be educated to use an AI and assess its results. For the third time, I state: "garbage in, garbage out". I let the machine translate into my mother language. I see if any translated text is garbage or not, and I will not submit any garbage for sure, and I expect everyone act this way if they use such a tool. In general, all the machine translations are much better than everything I'm able to come up with.

[coffee-factorio]

Using an AI is akin to using the correct tool for the job. But there is always some risk because it has no agency of its own.

This is also what I wrote in the 2nd sentence of the OP: "...to have a minimum amount of manual postprocessing." Notice I speak of postprocessing. You still need to check and edit the result. It's not an unchangeable answer from a higher entity set in stone, it's a draft you can and must edit. The strength of an AI is that it's able to produce a result in seconds for which a human needs minutes or hours. Checking for correctness and correcting any errors only needs a few seconds, so it's still a win.

There is something very strange going on here.

https://lokalise.com/blog/how-ai-translation-works/
https://lokalise.com/blog/can-llm-trans ... ccurately/

This is a blog by a respected localization company. It notes an AI is always learning. These people make their living doing this. They won't mischaracterize their product, if they do they can lose customers.

What you're saying is odd. I notice things. You wrote garbage in garbage out in your third paragraph. Not your second sentence:

I spent a few hours to create and refine a AI prompt that will translate an English Factorio wiki page to a German one. (2nd sentence) The goal is to get a translated article with most markup properly converted as well to have a minimum amount of manual postprocessing. Results are very good!

I guess it will work with every language. You just need to change the prompt accordingly: the destination language, the /de suffix to your language suffix, and the glossary. You can start with an empty glossary and only add required terminology, if some game terminology isn't correct.

--> (3rd paragraph) If you have no experience with one of the new AI chat bots so far, take this as example how to address and "speak" with a AI. It's as you are the boss and describe most precisely what you want to your apprentice who then goes running to fulfil your request. And if his translation isn't up to your demands, complain to it. It will admit its errors and offer a correction. An apprentice, by the way, with the whole knowledge of the internet. And if your prompt is rubbish, the answer will be rubbish. If your prompt is excellent, the answer will be excellent.

If this thing is training and responding with similar tone to my own. As indicated in those articles.

I have attention deficit hyperactivity disorder. I am very creative, but my tone is different than other's. I am more emotional, but I'm the type of person that that only comes in writing. I think I might be short circuiting your translator.

You aren't wrong. What you are saying is correct. But it is being said in a way which is very. Very strange to a native English speaker. More emotional, like how I speak.

Even in the error above. It is not incorrect. That is 3rd paragraph 2nd to last sentence. I am simply adept at handling shallow tree data structures in away normal people are not.

[eugenekay]

Nobody says you should directly save a machine translated text. The machine translation is a draft. It's a head start for the human translator. He still needs to revise the text to make it usable and accurate.

In case you're German by chance, please look into the wiki and check all the German articles.

If a dedicated Translator is looking to bulk-translate Wiki articles - and hand-edit the Results before Posting them - then there is indeed not much difference to a simplistic Dictionary replacement of old - it is just better Quality input. The Translator must still exert effort to confirm, correct, and rephrase the resulting work before presenting it as their own work - albeit one made with Tool Assistance.

My point is that most LLM / AI content by volume is not edited before posting, because it is being used as an “Amplifier Tool” by people looking to artificially inflate some Statistic through automation - such as a well-engineered Prompt. While technically interesting, this type of usage does not help anybody, is rife with errors, and is generally bringing down the quality of the Internet. So, please put some effort into it - which it sounds like you are.

I have some German relatives - but I never spoke the language. I do occasionally read Research and Technical specs in their original language; the experimental measurements and Math is all in Metric anyway. It is interesting to see the differences and omissions in the English versions - a missing “Not” changes meaning substantially!