Translation is nothing like self driving cars when it comes to AI. Translation starts with a perfectly known input so there is no reason a LLM wouldn’t be able to perform just as well as a translator. Self driving cars have problems with pattern matching on the input they get. Like seeing the difference between a plastic bag and a solid object on the road. Our vision pattern matching is something we still don’t fully understand besides the funny ways to trick it with 2d images and that makes it difficult to train AI to do the same…
This isn’t actually true if you translate things like poetry, philosophy, or even law and science, where words get used (and misused) in novel ways. Language is actually very imprecise, metaphorical, idiomatic, vague, context-specific, etc. There are probably 100 different “definitions” for many words (but they aren’t actually definitions, as definitions convey false precision, see Wittgenstein on the meaning of “game”). However, with a large enough library machine translation goes very far. I’m certainly no expert, but I think a fair amount has been written regarding Canadian French-English translation of all official docs and what they do these days.
2 Likes
As a Canadian I would legit be making double what I am now if I could speak French. Awhile back we had a Belgian as a temp worker and the Quebecious complained so much about him he got fired despite being good at his job because he didn’t speak French "properly ".
There is incredible demand at positions for bilingual people and it will stay that way for the foreseeable future.
For 2x, I’d be pounding duolingo every night and anything else that might help.
1 Like
If a LLM can’t translate something correctly then neither can a translator. Happy to agree poetry can’t be translated but that has nothing to do with the capabilities of a LLM compared to a human.
French-Canadians are a special breed.
If one is serious becoming fluent in a foreign language, the last thing one should be doing is wasting time on Duolingo.
Yeah, agree. I have been saying that all along. I just used it since it’s come up a lot in this thread and lots of people seem to like it. I personally don’t think it is that helpful.
Do you really not pay the “gringo tax” if you speak Spanish? Don’t take this the wrong way, but I don’t care how fluent you may become, you ain’t fooling anyone about your gringo status.
Maybe I’m not understanding the term. How I’ve heard it used is that a Spanish speaking gringo is still a gringo.
What they do at the UN is a great comparison.
Let’s break it down.
Firstly. The UN has lots of speeches, where people talk in long blocks. This is much more amenable to allowing a lag as the translator follows a few seconds behind.
Secondly. The subject and the context is known. You aren’t going off on tangents and different subjects, so the translator has a pretty good sense of the subject matter faster
Third. It’s a controlled environment with a single talker. Good audio. Etc
Fourth… These people are elite translators. With subject matter expertise in the subject and the top of their game at translation.
In short. This id actually a much easier problem for a translator to solve AND it requires elite expertise. And chatGPT is always just a mediocre layperson in any field. Miles away from real expertise.
I don’t think AI will get anywhere close to doing this for a while. And that’s still an easier problem than many real world translation problems.
1 Like
As argued by others above. This isn’t even true if we are using text.
If we are using audio, it gets worse. People.mumble. they use the wrong word. They start a sentence, switch to something else, then switch back. They respond to things that are happening visually or audibly. They respond to facial expressions from the other person, and you need to know what that expression was to understand what is said next
Revisit any word for word transcriptions you have read. They are not whole sentences.
Most language isn’t like this, but enough is to make it very hard to get to 100%
It’s both.
Two arguments being made
A LLM can’t translate as well as an expert, and won’t be able to for a while. But the cases it can translate will get better and better
And
A translation is never the same as the original, you always lose something, a god translation is always a smart judgement call, and is often creating something new that is similar but different to the original.
Thered a reason why ambassadors etc are selected for knowing languages and then put a lot of time and effort into learning new languages. They have high quality translation available for everything important they do, but they still want to learn the language to work in it and understand the culture they are dealing with.
1 Like
Here’s the thing, though. For a human to be able to understand all of those things that you mentioned perfectly it is extremely hard and almost no self-taught duolingo person is getting to that level. To get there you really need to be immersed in the language to the point where you are thinking in the language. It’s a difficult feat for someone who makes a conscious decision to go learn a language. You normally have to be exposed as a kid and then just soak it up.
So it doesn’t really matter if AI can’t do that, because humans, in general, can’t do that either without years of dedication. Close is good enough for most things and AI is pretty close to close already.
I think the vast majority of American ambassadors are very different from what you have in mind. A good number are assigned based on who donated the most to the President’s campaign.
Now the important ones are often pros with native language skills, but many aren’t that at all.
Agree with this.
But not this. It’s certainly possible to achieve near-native fluency in a foreign language as an adult–to the point where one can understand the nuances Rugby described. But as you indicated above, it requires immense effort and deep immersion in said language (and culture).
Agree, that why I said “normally”. I think the vast majority of highly fluent speakers of anything learned it as a kid. It’s been a while since I’ve read about this, but IIRC research in language acquisition shows that kids brains are just wired to accept new languages more readily.
It’s definitely possible. It’s just really, really hard. And if you’re a random who suddenly wants to learn Arabic to converse with Arabic speaking people better and your options are dedicate years to learning it or use some sort of AI that’s like 80%, most people are going to be very happy with the AI earpiece.
How is it not true for text translations. There is zero chance LLM is not going to beat humans in text translations in the near future. The issues with speech translations is mostly about preferring speed over accuracy so the automated translations right now are not waiting as long as humans do before translating. The other issues you mention cause human translators to make mistakes as well.
I didn’t say American ambassadors. Lol.
Well, if you don’t specify and it’s mostly Americans here, it’s hard to assume you’re excluding them. I guess in Australia it is actually mostly skill-based.