I am a translator for Japanese. I don’t think this will help much - in fact it might possibly increase my workload if someone approved those AI translations without correcting them properly.
We have already translated pretty much everything that can be easily correctly translated by AI, and the untranslated ones are all very tricky (and AI won’t be able to translate them well, because it requires knowledges about how iNat works).
Sounds good. I see that MT is already prefilled for Hungarian; I’ll gather my feedback and circle back. @kueda - can you remind me when you’re approving the unapproved translations? I want to make sure I approve HU before that.
I generally approve on the first US work day of the month, though I’ll be out of town at the end of April so I may not get to it until May 2 or May 5.
There is at least one exception on this rule, the lady bird case where I removed the translation of the common names and returned to the English names.
https://help.inaturalist.org/en/support/solutions/articles/151000194901-how-do-identifications-work-
https://help.inaturalist.org/nl/support/solutions/articles/151000194901-hoe-werken-determinaties-
Ok, it’s been a month. How are people feeling about pre-translate? I’m about to go through my usual review and approve everything that doesn’t seem to have technical problems, so some of those machine translations are about to get integrated into the code and deployed (probably next week). Does anyone who’s been translating in Hungarian, Danish, Russian, French, Afrikaans, Japanese, or Korean want me to turn off pre-translate for those languages?
@kueda - my experience so far:
Additionally, I looked into the CrowdIn help center and found that using MT with custom glossary would be possible (https://support.crowdin.com/machine-translation/#google-cloud-automl-custom-glossary-configuration), if this is an option for the iNat team.
Finally, I was positively surprised that CrowdIn has pre-translation accuracy reports, see https://support.crowdin.com/project-reports/#pre-translation-accuracy
Thanks, I really appreciate this detailed feedback. Would you like me to turn off pre-translate for Hungarian? This goes for all the proofreaders: if you don’t like it, please let me know and I’ll turn it off.
Additionally, I looked into the CrowdIn help center and found that using MT with custom glossary would be possible (https://support.crowdin.com/machine-translation/#google-cloud-automl-custom-glossary-configuration), if this is an option for the iNat team.
Looks like this is only for the Google integration, which we could try, but frankly if we try a different automation strategy I’d be more inclined toward an LLM-based approach. My guess is that the Google-based Machine Translation will suffer from the same problems you’re seeing with the Microsoft one, but I’d guess an LLM might be significantly different.
I don’t translate - but would be wary of automated Afrikaans translation.
On FB it is often 
nonsense. I get more sense from using my school Afrikaans.
Would you like me to turn off pre-translate for Hungarian?
I’d like to keep it on for Hungarian if quasi auto-approving them can be “turned off” :)
If the latter not an option, I’m leaning towards turning pre-translate off, as it doesn’t combine well with auto-approvals.
but frankly if we try a different automation strategy I’d be more inclined toward an LLM-based approach
A few Gemini-versions ago I actually tried translating iNat strings in AI Studio the manual way and I was impressed by the quality of the translations. It properly respected prompts like “the style should be informal”, “keep the markup”, “make it more compact” etc - and that was without feeding it TM and previous translations.
I tend to agree.
German turns out to be ‘rotten’ language to automatically translate to/from, and the culprit is English – because automated translations tend to revolve around English, e.g. Google does that => Italian gets first translated into English and from there to, say, French, and while that kinda works (the similar sentence structure helps a bit), forget any combination with German (I suspect Dutch and Swedish or Danish can be tricky too, but I can’t judge that as I do not speak those languages).
Assuming that a bad translation is always better than none at all may be delusional. Trying to save myself time I passed on a German webpage about bees in forests to an Italian friend with translation turned on – the result was unintelligible even to me who had read the original and knew what to expect. In two instances, in what looked like grammatically correct sentences, it had turned the meaning upside down, stating the exact opposite of the original. I ended up translating it myself.
In the past few months I noticed that the quality of translations went further south. AI doesn’t seem to improve things, actually, quite the opposite.
Which translation service did you use? If you didn’t pay for it, it’s probably not LLM-based.
My guess is it was Google for that webpage. My friend is not paying for translation services.
Since “auto-approval” is basically manual, I don’t trust myself to remember that I shouldn’t approve things in Hungarian every month, so I’ll just turn pre-translate off for Hungarian. I’m happy to do the same for other languages if you’re seeing unintelligible translations on iNat in your language (though an even better solution would be to add better translations on Crowdin).