AI and language learning - What have you tried?

[JLS]

I think I know what's going on here, and where people are concerned.

There are real issues with AI, which some feel strongly about. I think Le Baron is concerned that AI could be used to create an illusion of learning, when in fact AI has simply done the learning for you; and even some "drudgery" can be productive in its own right. I could go buy a pre-made flash card deck and get through more words; but I may learn fewer words much better if I handwrite them.

Others are concerned with unproductive drudgery. For example, how much do you really gain by looking up a word in a physical dictionary instead of looking it up on a professional dictionary website?

I use AI to transcribe Mandarin subtitles. I can't read the subtitles, and the learning curve to use either a physical or online Chinese dictionary are steep. In this case, AI captures all the hanzi (so far) perfectly, with pinyin, instantly. It'd be great to have the skill to use either a physical or online dictionary without AI's help. However, it's not useful for me to know Chinese at that level, so I'm foregoing that skill.

Technologies are all give-and-take. Autopilot is great, but could jeopardize manual flying skill (true story). Writing is great, but we use our memory less because of it. Application of tech is a matter of judgment. We gain some things, and inevitably lose others. I'm most interested in how to move forward with AI, with proper checks in place.

[leosmith]

What have others here tried?

This was 10 months ago:

I recently did a 5 month, 500 hr spurt in German from zero to approximately B1, and I used ChatGPT a lot. Here are the main ways:
1) To quickly give me the plural of a noun or past participle of a verb. I know this info is available all over the place, but I find it quicker on ChatGPT. 5 stars.
2) Compare two items. For example, two verbs with the same root but different prefixes, two completely different verbs, two nouns with overlapping meanings, two sentences that are equivalent or similar, etc. It is great for this. I can google reddit discussions and such for these, but again ChatGPT is faster. 5 stars.
3) Generate sets of 10 simple sentences to practice my word endings. It’s impossible to have correct word endings if I don’t even know the words. So I gave a list of nouns and adjectives to ChatGPT and asked it to use them to generate sentences to practice my word endings. I made it include singular, plural and all cases. Doing this was far easier/more effective/less frustrating than having to use random resources to practice word endings. That being said, ChatGPT really misbehaved; I think I was really taxing it. For example, it often used words that were not on my list. After a lot of experimenting, my work-around was to ask for the English sentences and quickly edit them to eliminate unknow words, then ask it to translate to German (which I wouldn’t look at yet). Then I would translate from English to German and compare my answer to its answer. This method worked quite well. Only 3 stars, but still worth it imo.

Italian grammar is much easier, so I’m only using 2) from that list. But as of this 4 month old thread I’ve been using AI a lot for generating or improving subtitles for videos. I prefer it when a video has accurate soft subs, because I can use them in my reading tool directly. If they are youtube auto-generated though, there are usually lots of errors and little or no punctuation. I can fix that by putting them in ChatGPT and telling it to add punctuation without changing any words. If there are no subs, or just hard subs which is often the case, I can strip the audio and feed it to a site like turboscribe or riverside which use AI to generate subs for free. Youtube doesn’t auto generate Swahili, so I’ve used these sites a lot while working on my new channel. But one thing about these AI subs – they aren’t very accurate for less popular languages. They’ll give you something to read, but if you want a near perfect match with the actual audio, someone is going to have to edit them.

It's about wise time management under severe time constraints which would increase your meaningful engagement with the language.

Nicely said. AI gave me a huge advantage over non-AI methods for 3) above. I wouldn’t use it for languages similar to my L1, like Italian, but I could see it being helpful for something like Latin.

The fixation with trying to be 'efficient' through AI is ridiculous.

I don’t know. That’s probably been said about computers, and even about book learnin’ at some point in time. I especially look forward to it making conversation more available. I’m still waiting for my hot robotress to talk to me in 13 languages while doing a striptease!

[Cainntear]

I think I know what's going on here, and where people are concerned.

Well (to use my favourite word) a lot of people are focusing on superficial details, which makes it harder to respond about deeper ideas without overgeneralising..

There are real issues with AI, which some feel strongly about. I think Le Baron is concerned that AI could be used to create an illusion of learning, when in fact AI has simply done the learning for you; and even some "drudgery" can be productive in its own right.

Absolutely, and I've been trying to push the discussion into the deeper stuff.

I could go buy a pre-made flash card deck and get through more words; but I may learn fewer words much better if I handwrite them.
...
Technologies are all give-and-take. Autopilot is great, but could jeopardize manual flying skill (true story). Writing is great, but we use our memory less because of it. Application of tech is a matter of judgment. We gain some things, and inevitably lose others. I'm most interested in how to move forward with AI, with proper checks in place.

I've brought those two paragraphs together because I want to make an important point: in many respects, and in many of its uses, AI is just a particular class of technology, and technology only matters in its influence on techniques.

Do we draw a line between buying a premade deck of physical flashcards and buying an Anki deck? If so, why do we even see a difference? Is it just that physical flashcards take a longer time to make than entering text (and possibly images ripped from the internet) into Anki? Because if so, that begs the question of how much "drugery" can be useful... and I think that varies from person to person, because the unanswerable question is what your brain does while carrying out tasks that don't really occupy the brain.

...and of course which tasks occupy the brain. My experience is that doing repetitive tasks on the computer tends to take more thought than doing them on paper, because there's a lot of bery abstract "actions" to express a semantic meaning to get the computer to do something, and on paper, you're just doing... well... actions -- it's all less indirect, so the brain's less busy.

Others are concerned with unproductive drudgery. For example, how much do you really gain by looking up a word in a physical dictionary instead of looking it up on a professional dictionary website?

That's a very very tricky thing to think about, so let's go back to flashcards for a moment...

Making physical flashcards may be less of a waste of time because you're not as absorbed in the task and your mind can wander more, but making computer flashcards takes less time overall so may be less of of a waste of time. There is no simple formula, and it probably differs between people, but it's all a matter of balancing things out.

Now the thing I see with dictionary websites is that when I open a new tab, I'm kind of changing context. Have you ever walked into a room and immediately forgotten why you did? It's extremely common, and I remember reading that psychologists had determined that walking through a door changes your context so radically that it feels like starting something new. I've found myself opening a new tab, typing in the address for a dictionary, and just completely forgetting what word I was intending to look up. To me, it feels like I've internalised switching between tabs effectively the same as walking between rooms. Yes, I have abnormal difficulties with attention, but I see that as simply making me more aware of underlying problems -- something that breaks my attention will be straining yours.

I use AI to transcribe Mandarin subtitles. I can't read the subtitles, and the learning curve to use either a physical or online Chinese dictionary are steep. In this case, AI captures all the hanzi (so far) perfectly, with pinyin, instantly. It'd be great to have the skill to use either a physical or online dictionary without AI's help. However, it's not useful for me to know Chinese at that level, so I'm foregoing that skill.

Now is that truly AI, or is it just a computer? My understanding is that hanzi->pinyin is pretty trivial, and it's only pinyin->hanzi that is computationally tricky. I don't know Chinese, but my understanding is that the biggest difficulty in hanzi->pinyin is identifying when a radical is a standalone word, and when it is part of a compound.

As such, if there is an advantage to AI over algorithmic approaches, it's a relatively minor improvement in accuracy.
That means this isn't a question of "Is AI good?" but "Is AI better than other computer-based approaches?" and that isn't addressing the topic of the thread particularly directly.

Well yes... JLS started the thread with the aim of looking at specific techniques that people have tried but let's consider an analogy. Imagine recorded audio is new, and imagine the question is "recorded audio and language learning -- what have you tried?" This leads someone who has used Pimsleur to say audio-based stuff is great, but doesn't really talk about the fact that he's doing it in between going to weekly night-classes. Someone then complains that Pimsleur is absolutely not complete, and the first respondent kicks back with "do you think that's all that I do? I didn't say it was!" Then someone else cuts in complaining about how language labs don't work for him, so the Pimsleur guy is wrong, and someone else cuts in to say audio is rubbish because Michel Thomas teaches Spanish with a very strong learner accent.

When you put a name on something, it becomes a single conceptual "thing", but that "thing" is different to everyone. In the case of audio-based learning, it might sound odd the idea that the breadth of techniques might ever be bundled into one thing, but it really wouldn't have been in the early days of the gramophone. Hell, even towards the end of last century, the simple act of putting more CDs with a language learning book was often resisted because everyone associated audio with language labs, and schools were ripping language labs out and doubling down on the importance of paper. Audio in their minds mean "language lab", language labs didn't work, therefore they felt audio didn't work.

[Le Baron]

I’m still waiting for my hot robotress to talk to me in 13 languages while doing a striptease!

Ah...When you put it that way.

[Cainntear]

I’m not sure how much I’d trust it for Finnish. It is useful for cleaning up formatting. When I copy autogenerated transcripts of YouTube videos the formatting is pretty bad. Maybe I’ll make it list vocabulary words and grammar concepts based on frequency and complexity. I’m not counting on it being the most accurate, but I can look stuff up.

That's kind of the main issue: trust.

If it fails so badly that you can notice it in a language that you've not even put on your language list yet, then it's absolutely not trustworthy in that language.

But how can you trust it in any language? All it has to do is be good enough that you can't see the mistakes, and you end up happily learning its mistakes.

The quality in any language is clearly related to how much of that particular language it's seen, but it's not a linear relationship -- AI will always make errors. Until and unless we know what the errors are, will it ever be safe to use AI as a self-directed learner?

The key problem here is a decades-old one:
AI is seduced by the "sexy" notion of a computer that does everything for the non-expert, so this is where research goes. But it still needs some kind of expert supervision. The power of LLMs in language education is that they can be used by teachers to generate exercises for their students much quicker than doing them by hand -- when I was teaching English, I was never happy with the supplied worksheets or the free ones on the internet, and I knew what I would have preferred, but I couldn't justify dedicating more time to writing worksheets than I was spending in the classroom, because they weren't paying me enough. But if I was teaching now, I could imagine myself hacking around with ChatGPT to get the exact set of exercises I wanted for each class, but that still relies on me acting as the editor and correcting mistakes. And note that I say "hacking around with ChatGPT", because ChatGPT has an interface set for non-experts, thus making it harder to ask the system to do the exact things that it's best at and best for... because that's not "sexy" stuff.