Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Is this a decent fit for LLM?

“Talk to me in <language> and point out my grammar errors in English”

I imagine it’s risky, learning bad habits. But it seems like it might be very convenient. I believe the biggest issue for me is actually using a language regularly. But I’m way too socially afraid to do one of those “speak to a random person live” things.

Or even some sort of, “translate all my emails to <language>, but show English when I mouse over.”

I bolstered my French by setting almost all my video games to French in university. It helped me a ton, and was accessible because I understood the context.

Translation tech has come a long way. Might not even need LLMs.



The one and only actually useful use-case I've found for ChatGPT in my life (since it can't handle assisting my extremely basic coding work) has been "break this Japanese sentence down word-by-word and explain the grammar." On the surface, it seems more helpful for understanding and learning than simply putting the words into a JP/EN dictionary (which doesn't explain grammar at all) or putting the entire phrase into Google/Bing Translate (which makes it too easy to mentally ignore the grammar points I need to learn).

Reading the other couple of replies, though, maybe I should rethink doing even that.


Yes, as always it’s risky to use a LLM for something you’re not already familiar with. I guess for English or Spanish it’s good because it has a large corpus, but for a smaller language like Italian it’s quite bad.


Tried having it generate German puzzles (normal sentence with a missing word like "der" or "dem" or so) after someone blogged about that it would be worth like 90% of a language teacher for 1% of the price. I'm not very good at German but most things it proposed seemed wrong to me. The whole point is that I don't have to talk to a native speaker but I decided to show the conversation to one who then said something like "yeah no, you're correct half the time and the computer is wrong even more times"

Maybe I should feed it bits from Wikipedia and have it censor word classes for me (or is part-of-speech identification by human-made algorithms reliable?), but that's a lot more involved to code up than prompting it "hey just do this task". I'm sure I'm just holding it wrong and it can be a useful language teacher in some way, e.g. I have had good results with 1:1 translations, but don't expect it just does what you ask it when you can't verify the result




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: