← 返回日报
精读 预计 2 分钟

23 languages, one I can check

摘要

独立开发者 Jake 分享了他用 AI 将应用 Popsicle Boat 的约 1500 条界面字符串翻译成 23 种语言的过程。他本人只精通英语和德语,通过仔细审校德语翻译,发现 AI 翻译质量很高,不仅语法正确,还符合德语母语者的习惯表达,因此对不熟悉语言的翻译也建立了信心。他建议其他开发者先检查自己熟悉的语言的翻译质量,再决定是否信任其他语言,并做好接受用户反馈修正的准备。文章还提到,没有母语用户反馈的语言翻译很难验证,但用户反馈链接是快速修正的途径。

荐读理由

文章给出了可复用的 AI 翻译工作流:先接入 gettext 保证所有字符串走翻译函数,再用自己精通的语言(如德语)仔细审校 AI 输出,以此作为质量锚点,最后通过用户反馈修正其他语言。这直接解决了独立开发者多语言本地化的成本与验证难题,且是真实的一手实践而非炒作。

原文

23 languages, one I can check

Projects

Popsicle Boat is available in 23 languages besides English now. I speak two of them: English and German.

I wired gettext in early, so every visible string in the app has been going through a translation function from the beginning. That part was just discipline. What I never had was the translations themselves. There are about 1,500 strings in there. Getting those into Arabic, Thai, Korean, Ukrainian, and twenty more is not something I was ever going to do myself, and paying 23 translators for a hobby project was never going to happen either. So it just sat there, one language, waiting.

The AI did it. All of them, in an afternoon.

It lands on the right word. Usually.

The part that bothered me

You can’t check the work. That’s the whole problem. I can look at the Thai file and tell you it’s Thai, and that’s the end of what I know. If it’s stiff, or wrong, or accidentally rude, I have no way to find out until a person who speaks it tells me. Shipping a thing you can’t read to people who can read it is a little scary. I want to get things right, but seriously. How would I know the Thai translation tracks for a native Thai speaker?

Why I shipped it anyway

German. I read it, write it, and speak it fluently. I lived in Germany for 3 years. German happens to be one of the 23, and it’s the only one aside from English I can verify.

So I read the German. I didn’t skim, I read. I scoured. And it’s good. Not “technically correct” good, actually good. The tone matches, the informal tonality is consistent, the UI strings are the ones a German app would use instead of the literal translation of the English. A few things I’d have phrased differently, which is also true of things I write in English. So in general, it’s been a good exercise in reviewing copy that faces users. Probably (almost certainly) still rough edges especially in languages I don’t know, but also the two I do. But AI brought me a loooooooong way to supporting languages I simply don’t know.

What I’d suggest you do if you’re curious

This has been an experiment, mostly. Of course I’d love a global audience, but this was not an anticipatory effort. It was more like “Can AI do this?”.

One thing I didn’t expect: shipping a language nobody reads yet teaches you nothing, because nobody is there to tell you it’s wrong.

If you’re interested in trying this for your own site, here’s the general flow I followed: check the language translations you know -> decide if it’s good enough to have confidence in the ones you don’t -> commit, but brace yourself for corrections from users

If I didn’t know German, I don’t think I’d have pushed this. But because I know my own fluency, I could spot check the German translation and it was good. It was like, really good. So that gave me confidence on all the languages I do not know.

Way cool, friends. This technology is not just a toy. It’s wildly powerful.

There’s a feedback link in the sidebar and it goes to me. If something reads wrong in your language, that’s the fastest way to fix it, and honestly I’d rather hear it from you than not know.

Cheers Jake

Lobsters · 1 赞 · 1 评 讨论 → 阅读原文 →

这条对你有帮助吗?