Experiment: German translation 🇩🇪🇨🇭🇦🇹

[caugner]

tl;dr

We're launching an experimental German version of MDN, which is targeted to those that prefer to read MDN content in German rather than English. The content is automatically translated from English using OpenAI's latest GPT-4o model, and is now available on (developer.mozilla.org). The sources can be found in the mdn/translated-content-de repository.

Note: This is an experiment, and work in progress. Translations may contain inaccuracies.

Details

MDN was available in German (de) until August 2022, when we had to remove the locale, because it was incomplete, outdated, and no longer maintained. Although we hypothesized that there was still demand for the German locale, we saw no possibility to bring it back at the time: We had evaluated automatic translations as not good enough, and considered manual translation as impossible due to the volume of MDN content, and the pace at which it is updated.

In April 2024, we re-evaluated our options, and identified LLM-based machine translation as a potential solution to revive the German locale, and to maintain it in a semi-automated way going forward.

In May 2024, we ran initial tests with LLM-based machine translation from English to German, which confirmed that the approach yields sufficient results even with a simple prompt.

In June 2024, we ran a survey (see results below) that confirmed interest in the German locale, and informed us about translation criteria, and user concerns.

In September 2024, we launched the new experimental Remember language feature, which we had identified as a blocker for the German locale in the survey. The feature effectively allows users to opt-out of the German locale.

In October 2024, we launched the public preview of the German language in an isolated test environment.

Experiment

In November 2024, we're now launching the German language as an experiment on production (without exposure to search engines). This is the first fully-translated language covering all MDN Web Docs content with over 12'500 pages, translated from the English language using OpenAI's latest gpt-4o-2024-08-06 model. (Note that most of the user interface, MDN Plus, Curriculum, and Blog are not translated.)

In December 2024, we started exposing the German translation to search engines.

To provide feedback, please rate good and bad translations using the thumbs (👍👎) in the "MDN-Feedback-Box" below the articles. If you notice any issues, please use the "Report a problem with this translation" link, write us on Discord, comment below, or send us an email (mdn-german[at]mozilla[dot]com). Please respect the Mozilla Community Participation Guidelines to ensure a healthy and constructive collaboration.

Translation process

As of December 4, 2024, the translation process includes a daily automated workflow that includes the following steps:

1. The English Markdown file is pre-processed, in particular by replacing code blocks with placeholders, converting selected macros to links, converting the YAML frontmatter to Markdown, and chunking longer documents.
2. The resulting input is passed to GPT-4o with a system prompt.
3. The resulting output is post-processed, in particular by cleaning up frontmatter, normalizing links, fixing well-known recurring syntax errors, re-inserting code blocks, and replacing link anchors.
4. The resulting German Markdown files are committed with automatic markdownlint and prettier fixes to the mdn/translated-content-de repository.

Next steps

We believe that the machine-translated German locale is helpful in its current state, even if it may contain inaccuracies.

For now, we gather user feedback, and aim to iterate on a weekly basis to address issues, and improve the quality of the translation, and the reliability of the translation process.

Our vision is for German to become a regular translated locale, and the first locale to be maintained using a combination of machine translation, and community reviews.

Appendix: Survey

In June 2024, we ran a survey on MDN for 5 days that was shown to all users whose browser preference indicated German as the preferred language.

Goal

The goal of the survey was to test our hypothesis that there is demand for the German language, understand our users' language proficiency and preference, translation priorities, and concerns.

Participation

The survey was seen by over 60'000 users, and over 1'900 users completed it (3%).

Language proficiency and preference

We asked about language proficiency:

• 90% of participants indicated German as their native language.
• Of these, 33% rated their English as basic or advanced, not fluent.

(Note: German language preference correlated inversely with English language proficiency.)

We asked what language they would use if MDN was available in German:

• 62% participants indicated a preference for English (54% would use mostly English, 8% more English),
• 35% participants indicated a preference for German (24% mostly German, 8% more German), and
• 3% participants indicated no preference.

Translation priorities

We asked what would be important in a German version of MDN:

1. Technical correctness (89% Very important)
2. Consistency of technical terms (73% Very important)
3. Up-to-date ness (68% Very important)
4. Completeness (56% Very important)
5. Comprehensibility (50% Very important)
6. Linguistic correctness (31% Very important)
7. User interface in German (14% Very important)
8. Graphics in German (8.3% Very important)
9. Code examples in German (7.5% Very important)

Concerns

We asked an open-ended question, and 13% of participants answered it, providing context for their responses, or sharing concerns.

Among users who indicated no preference or a preference for German there were no significant patterns, apart from praise and gratitude (n = 18; thank you!).

Users who indicated a preference for English …

• … explained their preference with programming languages and technical terms being in English (n = 59), and having better experiences with English documentation, being more complete and up-to-date (n = 44).
• … rejected automatic or AI translations, criticizing them as inaccurate and incomprehensible (n = 34), and sharing negative experiences with machine-translated technical documentation (n = 19).
• … emphasized that they don't want to be redirected to the German language (n = 21), even if their browser indicates German as the preferred language, and mentioned the need to permanently switch to the English version (n = 10).
• … suggested having translations reviewed by the community to improve their quality (n = 12).
• … acknowledged that some users, especially beginners, could benefit from a German translation to better understand technical concepts (n = 11).

Limitations

The main limitation of the survey is its appearance in the context of non-German content, so potential users who only speak German may be under-represented. On the other hand, users who are interested in the German locale may be more likely to respond to the survey, and therefore over-represented.

[yatil]

Love how this page in German encourages people to "make use of diarrhea" ("Ausnutzen des Durchfalls"). "AI" translations are really innovative stuff. /s

Hire translators.

[AliveDevil]

Microsoft can’t get their automatic translations up to quality, who own OpenAI, how’s MDN going to get this right? You’ll be constantly fighting with the model.

Piece if Evidence A: „DLL Surrogate“ is not „DLL-Ersatzzeichen“.

[caugner]

Thank you for trying out the experimental German translation, and for taking the time to provide this feedback. 🙌

Love how this page in German encourages people to "make use of diarrhea" ("Ausnutzen des Durchfalls").

First of all, great catch! 🙂

Looking into the translation history of this page, 3 out of 4 translations used the English term "fall-through", while only the most recent one replaced it with the German words "Durchfall" and "Durchfallen".

While this translation is not necessarily incorrect, it highlights a trade-off: leaving a term untranslated (e.g., "fall-through") vs. using a German equivalent that, while technically accurate, might feel unconventional or humorous to some readers. I have re-translated the page with our latest prompt, which now uses "fall-through" again: mdn/translated-content-de@b58ce94

@yatil Could you confirm whether "fall-through" should remain untranslated? If a German translation is preferred, do you have a suggestion for an alternative that might work better? Based on your input, we can update our translation glossary to ensure consistent handling of this term moving forward.

how’s MDN going to get this right? You’ll be constantly fighting with the model.

@AliveDevil MDN is committed to transparency in its approach to machine translation, and we encourage the Community to participate in refining the German translation. We believe that collaboration and feedback from both beginners and experienced developers will significantly improve translation quality over time. To support this, we’re actively seeking input through multiple channels, such as the "MDN-Feedback-Box" at the bottom of every translated page, discussions like this one, and the #german channel on Discord.

Please continue to share any terminology issues you spot, along with your suggestions for improvements. 🙏 Your feedback is invaluable to us.

PS: We kindly ask everyone to follow the Mozilla Community Participation Guidelines to ensure a constructive and respectful collaboration. Let’s stay direct, but also positive and solution-oriented. 😊

[yatil]

I will not help support machine-translated resources, so I cannot reply to your question, even if I knew the answer.

[kaiserkiwi]

While this translation is not necessarily incorrect

It is. "Ausnutzen des Durchfalls" is not the correct translation. "Taking advantage of fall-through" is translated with "Taking advantage of diarrhea" in this post.

"Ausnutzen des Durchfallens" would be technically correct. Even if it's a completely weird way to write.

You want to know what shouldn't be translated automatically from your AI? Everything. Hire translators or keep it in English. Please this does more harm for everyone involved that it helps anyone.

[caugner]

The re-translation of the switch page is now live, see: Ausnutzen des Fall-through-Verhaltens.

[tomayac]

Could you confirm whether "fall-through" should remain untranslated?

Yes, this should remain as is. "Ausnutzen des Fall-through"-Verhaltens" is how I would translate it.

[caugner]

Thank you. I double-checked, and this is the only page that uses this term, and the current translation now consistently uses the term "Fall-through".

[philo1989]

Pls just dont..... It's teribble. Make it an Option for people who need it but not the default pls....

Another part to consider... Do you really want to particapte in the llm hype and contribute an not insignificant energy need on a pet project like a german translation... . They need to build new power plants for this bs and you Support that?

[caugner]

@philo1989 Thank you for your feedback.

I invite you to try out the experimental Remember language feature that allows you to set your preferred language, and ensures that you will always be redirected there if the current page exists in that language.