Using AI to translate Simutrans into languages ​​for which we don't have transla

[victor_18993]

Hello everyone, these past few days I've been experimenting with a specific AI lab for contextual translation of conversations and technical documentation based on Gemini. I became curious about whether it would be beneficial to translate from English, as the canonical language, with context for languages where we don't have translators.

This should allow us to reach more players and have the Steam version indicate that it's in their language, which is usually quite positive since there's no language barrier.

Since this involves AI and I enjoy discussing these topics with the community, I'd like to know your thoughts on translating for languages where we don't have translators.

It might help us attract translators, or perhaps it could detect a spelling error and have it reported.

As you know, I also have the NLS project underway, which is similar but focuses on the code area, allowing you to write code and have the text output adapt to languages that require different treatment.

Thank you very much.

[danivenk]

AI translations not always look right tho. Since translations do heavily depend on whether it makes sense or not (see how people always bash Google Translate when using sentences) it is always important to have it be checked by either natives or people who have a good enough grip on the language. Sure an AI could give you that initial "draft" but it would be good to still have it be checked. A kind of code review but for translations.

[victor_18993]

Yes, I agree with you that this never replaces a native speaker's review—there's no debate about that.

Unfortunately, Google doesn't use its translation models for its own tool, and it also lacks context, which is the crucial part for doing translations.

They could be drafts, but if we don't currently have translators for these languages, I find it hard to see anyone reviewing this specific section, though to me it's also valid to just leave the translation files so they can be reviewed if a native speaker comes along.

Specifically, I mean languages where we don't currently have players, but which could benefit from initial experimental coverage in their native language.

[danivenk]

Yes, I agree with you that this never replaces a native speaker's review—there's no debate about that.

Unfortunately, Google doesn't use its translation models for its own tool, and it also lacks context, which is the crucial part for doing translations.

They could be drafts, but if we don't currently have translators for these languages, I find it hard to see anyone reviewing this specific section, though to me it's also valid to just leave the translation files so they can be reviewed if a native speaker comes along.

Specifically, I mean languages where we don't currently have players, but which could benefit from initial experimental coverage in their native language.

I guess it should still show up in the system that it is a draft translation, which can then be approved by a native/speaker of the language later down the line. Just to be sure that things stay as correct as possible

[victor_18993]

If the notice you mention about it being preliminary seems important to me because it conveys the reality to the player, I don't think it would be difficult to include it right where we choose the game language in Simutrans.

[makie]

Hello everyone, these past few days I've been experimenting with a specific AI lab for contextual translation of conversations and technical documentation based on Gemini. I became curious about whether it would be beneficial to translate from English, as the canonical language, with context for languages where we don't have translators.

For pak128.german is German the source language. The English versions are often just cobbled together using Google Translate, and some English translations are outdated or poor.

That said, an AI translation is better than nothing and can serve as a foundation for a native speaker to revise, as they won't have to start from scratch.

I think automatic translations have become quite good these days.

[victor_18993]

For pak128.german is German the source language. The English versions are often just cobbled together using Google Translate, and some English translations are outdated or poor.
That said, an AI translation is better than nothing and can serve as a foundation for a native speaker to revise, as they won't have to start from scratch.
I think automatic translations have become quite good these days.

When the canonical language isn't English, I usually use a technique where I take it to the language with the best closeness percentage according to Google metrics and use the API. In those cases, what I do is a double translation: first I translate it to English, and then to the rest of the languages.

I also think a preliminary translation is better than nothing, since it helps break the language barrier, which in a game like Simutrans is important because of its complex mechanics.

[Isaac Eiland-Hall]

I would imagine that an LLM would generally be about the best practical tool (outside of volunteer or professional human translation) possible. Google Translate is alright, but I believe it focuses more on individual words, whereas an LLM can take wide context into account.

I think identifying translations done this way somewhere is not a bad idea, but even more than that, I think the idea of trying to provide translations to enhance the possible audience for the game is a worthy cause, and has the potential to bring joy to more people.

[victor_18993]

I've uploaded the Albanian translation as a test. Looking into it, I see on Steam that players from Brazil point out that it's not in their variant of Portuguese. The same happens with Spanish speakers living in Latin America—they speak another variant of Spanish that changes key game terms and how they are understood. It would be interesting to consider supporting variants whenever justified.

Which languages do you think should have priority for a provisional translation?

[prissi]

Preferably a language where we have players who can comment. The translation have many very short sentences or even single words. Especially the latter are not straightforward to translate without context of where they appear.

I think we have some players from Indonesia although none of them is active in the Forum.

[Isaac Eiland-Hall]

But it may still be interesting even if we don't have such players, if only because LLMs do know the difference.

In a single short paragraph, can you explain to me the differences between US and UK english as far as different words we used to call parts of a car or truck?

Where a US driver pops the hood to check the engine, a UK driver lifts the bonnet, and while Americans toss groceries in the trunk, Brits stow them in the boot. Roadside emergencies also require different gear, with Americans grabbing a spare tire and flashlight from the trunk while Brits retrieve a spare wheel and torch from the boot. Mechanics across the pond swap out fenders for wings, windshields for windscreens, and bumpers for bumpers—though Brits might casually refer to the whole machine as a motor rather than a car, and commercial haulers as lorries instead of trucks.

But if we do have some way of identifying AI translations, it might sitll be a Good Thing™.

[victor_18993]

I am filling in the missing translations for languages with less coverage; currently, I am not aware of any mechanism within Simutrans to identify them, but I can program an AI agent to monitor comments on Steam and platforms like the Play Store and notify me if anyone points out a poor translation.

[victor_18993]

But it may still be interesting even if we don't have such players, if only because LLMs do know the difference.

But if we do have some way of identifying AI translations, it might sitll be a Good Thing™.

In English, there are probably fewer variations in words. In Spanish, for example, in Spain they say Autobus and in Latin America they say Camioneta, Gugua, etc. In fact, there is a variant of Spanish called the neutral variant, which is usually based on the Spanish spoken in Mexico.