Blank Slate

[flatcap]

Writing the first of anything is hard.
You have to decide what to say, how to say it.
I imagine it's the same for translations.

I have a crazy idea that might just help to encourage new translators to get started.

Recently, I found "translate-shell" -- a script which taps into Google's translate API (it can use other services too). These free translators use standard texts and AI in order to produce meaningful translations (allegedly :-)

• https://github.com/soimort/translate-shell

Would it be worth using them for all our unsupported languages?
We could fill-in approximate strings for all the untranslated and fuzzy entries.

It would be a lot of work.
Does anyone think it would be worth it?
Would anyone like to try?

[l2dy]

I think machine translated strings with no review must be set fuzzy because incorrect translation can turn out to be worse than no translation.

Lack of context could make things worse. e.g. msgid "created: " is used to show the time a message was signed, but it was incorrectly translated in zh_CN to "created: something" rather than "created: timestamp". Without the necessary context, even humans can't figure this out.

 #: ncrypt/crypt_gpgme.c:1482 ncrypt/crypt_gpgme.c:1487
 msgid "created: "
-msgstr "已建立："
+msgstr "创建于：" 

FWIW, I found https://github.com/OzzyCzech/potrans, but it will be billed per https://cloud.google.com/translate/pricing ($20 per 1,000,000 characters). Also, some web-based translation platforms like Crowdin support machine translation and site-wide translation memory.

[flatcap]

strings with no review must be set fuzzy

Absolutely

incorrectly translated ... "created: something" rather than "created: timestamp"

Please feel free to add 'L10N' notes in the source for these -- they're sure to mislead other translators.

billed ... ($20 per 1,000,000 characters)

50K chars * 30 languages = 1.5M chars => $30
I hadn't considered paying for it, but that's not bad!

some web-based translation platforms like Crowdin support machine translation and site-wide translation memory

Bored one evening, I downloaded the entire source of Fedora and extracted all the po files (~8GB). I wondered if we could use that as a translation memory. (Alternatively, I might just drown in too much data).

[l2dy]

use that as a translation memory

Zanata supports site-wide TM, but 8 GB... I'm not sure if it can process that much data efficiently.

[flatcap]

Interesting.

If anyone wants to take a look, it's 444MB compressed.
https://flatcap.org/mutt/fedora-po.tar.xz

(It's a new month, so I've got bandwidth to burn :-)