It’s the consequence of translation engines being trained on a corpus that is mainly Spanish and English. Meaning, that most txt in other languages have been translated to either Spanish or english, and those translations are what is used to train contextual inference for translations. It needs a corpus of text of two languages to even attempt translation from A to B. If it doesn’t have enough, it will use bridge languages.
How many texts are there both in Ukrainian and Japanese? Effectively none, so it makes perfect sense to use a bridge language. The issue here is that the context wasn’t properly kept in the bridging, and they should theoretically be able to do that. A buddy I shared my master’s with was Iranian and he explained the semantic context issue to me, he was working on exactly this issue.
The middle language is not a spoken language. Googles AI created an intermediary language (like 10+ years ago) to translate between all languages. It’s not written/spoken, it’s done on weights. It’s referred to as an ‘interlingua’
Very old article, they are absolutely just using llms nowdays
Edit: this is the paper https://arxiv.org/pdf/1611.04558 the article is about
In my experience, with the amount of mistakes similar to the one in the OP, I strongly suspect that language is somehow unbelievably close (semantically) to English then.
I decided to quickly experiment with two languages I somewhat know, and got 3/3 incorrect translations consistent with English as intermediary:
- Russian ширинка (fly, meaning “zipper on your trousers”) -> Georgian ფრენა (to fly, as in “flight”) - completely incorrect translation, consistent with English as intermediary
- Russian парковаться (to park, as in “park your car”) -> Georgian პარკი (park, meaning “green space with trees”) - completely incorrect translation, consistent with English as intermediary
- Russia лежать (to lie, as in “lie on your back”) -> Georgian ტყუილი (lie, as meaning “untruth”) - completely incorrect translation, consistent with English as intermediary
It makes exactly the same mistakes if you translate from Russian to Spanish and French too.
I don’t know if there are many other languages where “zipper on your trousers” and “flight”, “park your car” and “green space with trees”, “lie on your back” and “untruth” would all be pairwise the exact same word. I strongly suspect not, or at least it would be a language really close to English.
you’d think they would use embeddings nowadays instead, with all the llm hype going on
If you say so.
They don’t. OP just bulk reposts other people’s complaints. Talk about mildlyinfuriating
I was just pointing out that I (and probably most lemmy users) have to take OPs word for there actually being a mistranslation in the screenshot.
And I’m just saying they’ll never confirm it either way because it’s not a human posting
You notice it when you look at international instructions on packets of food. The non-English texts have a distinct English vibe even when the product is, say, French or German.
Or Microsoft documentation pages which are all sloppily translated by ai without proofreading
For German Dutch is does seem to do it directly. But I think it mostly is because a person has put in the effort for it.
The way translation engines are trained is with a body of written text in two languages, which teaches the engine both how words translate and how they fit in equivalent semantic contexts. That way it can decide which synonym is the best one to use given the context, and lose the least amount of context in the bridging.
Of course a person as put in effort, so has in every language. But that’s irrelevant if there isn’t a large body of text in both languages to learn from. I’m sure German and Dutch have a looooot of shared texts.
Doesn’t German -> Dutch basically involve just translating half the words to English and adding a silly accent though?
I wish. Love, Dutch person learning German.
Yeah, in my experience it absolutely does translate everything via English. If you need to translate from Ukranian directly to other languages, ironically Yandex Translate is probably your best bet. It’s worse at Russian <-> English and Ukranian <-> English but much better at direct pairs between other languages.
Are they? I know that its been that way years ago, but now with ai technology, it should be able to translate text not just by the word definition, but also context and conceptual meaning. At lest it is supposed to work this way if they used ai for actually useful stuff and not another slop generator.
it should be able to translate text not just by the word definition, but also context and conceptual meaning.
It has been doing that at least since 2016, and it’s using LLM-adjacent tech (transformer + RNN) since 2020. The problem is that it still seems to be doing things via English, which is especially problematic for single words with no context. In this example, “зачиняти” means “to close”. Google translates it to English as simply “close”, which loses the context of which meaning or even part of speech it is. Google then translates it to Japanese, but since the context is lost it assumes that “close” means “near” rather than “block an opening” and translates it as “chikai”.
(actually, Yandex Translate fucks this up as well and translates “зачиняти” as “黙れ” (shut up), which is somehow worse)
how about deepl.com? it’s been my go-to for a while but no idea how it performs on these langs
It translated correctly to 閉める
I have generally found it very good for Japanese, but I’ve only done Japanese <-> English pairing
