Lately, I have been using Google’s Gemini AI assistant to help me learn German. It does very well with prompts like, “I would like to talk with a German friend about community theater. Help me learn German vocabulary about this topic.” I’ve also gotten very helpful responses to prompts like, “Help me understand the difference between ‘gleich’ and ‘gerade’.”
Since the gradings for German books on Natively are still so sparse, I’ve tried prompts like, “Suggest some books to read in German at an A2 level.” I’ve even tried to use it as a Natively substitute with prompts like, "Which book is harder to read in German, “Die Kleine Hexe” or “Petronella Apfelmus”?
When I tried to get german book recs out of ai it kept making books up. I didn’t feel it was useful at all. It’s ok as a convo practice partner. Also frequently wrong when I asked it to break down grammar.
In jp, I sometimes use it to explain grammar questions I get wrong, but I always make sure to double check. It is relatively obvious when it’s wrong because there wil be no internal consistency in what it is saying. Honestly best thing it does is identify what grammar pattern is being tested in my question, once I know that I can find human made resources easily. It hallucinates around 20 percent of the time in this context in my experience. Still useful in cases where I can’t realistically ask someone. But if you can’t identify when it goes wrong, functionally useless.
I think people either act like ai is the best thing ever for language learning or not to be used at all but it’s, like most things, in-between. An useful tool that you need to know how to use.
oh definitely. I actually was thinking of a recent discussion I saw on asking simple grammar discussions to ai vs on open forums like reddit or specific discords. Unless you have a tutor/someone you know is high language, there is no reason to implicitly trust humans either (and tutors etc can make mistakes too ofc).
I experimented with using chatgpt (which I’m otherwise very averse to) as a conversation partner for like 10 min. I get the impression it’s good for practicing generating your own sentences, but it wasn’t a particularly good conversation partner (nor am I). I do wonder if it would be good for “how do I say this in タメ口/敬語/謙譲語” sorta questions
While I personally wouldn’t use it for grammar breakdowns, I don’t think it’s inherently doomed… but I doubt it’s a good idea to send anyone who isn’t decently far along to these things - since I’m not sure they’ll be able to tell when it’s hallucinating… Whereas at least with humans online, someone is likely to correct someone else giving wrong information. It is also possible that somebody will later realize, and correct themselves
I always get so happy/amused when I discover errors in books
Reddit especially is way too variable. Otoh sites like HelloNative or Chiebukuro are good sources.
You could always ask here, no? Should be pretty clear what people’s approximate knowledge level is, since we have book levels and such
I use chatgpt fairly often for simpler sentences when jpdb doesn’t have bilingual sentences, or sometimes even if they do. Occasionally
you’ll get jpdb sentences like pen: this is a pen, which you may as well not have a sentence at all. And the ones without English translations tend to be long and meandering passages without much context, so even when I do understand all of the words (which is maybe 30% of the time) it’s still not that helpful for me as an aid to recall a meaning.
Another place I’ve been using it with jpdb is with leech cards to help me come up with a better definition. A lot of time on my bad leeches it will have a meaning that I just struggle to remember or have such disparate meanings that it’s hard for me to really grasp what the word means. I started asking it for simple definitions of the words and that’s helped a lot on leech cards.
I also do a lot of writing to my friends on LINE, and after explaining who I am (age, gender, etc), my relationship with the people I’m messaging, and what kind of vibe I want, it does a good job of taking my sentences that just feel not quite right and both giving me better options plus telling me what it is that was off about what I had.
A long time ago, I had asked it for book recommendations but like @earwen got a lot of fake book hallucinations. Now that ChatGPT can search the internet I’d maybe be more inclined to give it a list of award winning works and ask it which ones are in my level or rank them in difficulty. Honestly, Natively is way better for that for me, but I do appreciate that outside of Japanese (and to some extent Korean) Natively just doesn’t have enough of a user base to be that helpful. (I know that this thread is about AI, but I do have a few A1/A2 German books/graded readers that I’d recommend if you’re interested @HerrTimmy).
I think my takeaway with AI for language learning is the same as my takeaway for AI more generally: it’s really most useful when used for things that you have a base level of knowledge with as either a second pair of “eyes” or at most in a peer mentor role. You really do need to be able to spot its bs, which I don’t think you’re able to until you’re into an intermediate level. That said, for something like a language partner I do trust it, since language generation is actually what LLMs are meant to be doing. Although I wouldn’t really want to pick up some of the linguistic quirks it had, it’s good at roleplay so you could just ask it to be the type of person you’d expect to encounter.
tl;dr I’m with @暁のルナ that it’s neither a life changing tool nor is it absolutely to be avoided. But that said I do use it for language learning almost daily I’d guess.
Yeah jpdb example sentences are wildly variable (and sometimes straight up wrong). I ignore them 80% of the time. Mostly I look at them for hints about how the word or grammar pattern works.
Interesting. I’m not connected to anyone on LINE (it’s fun tho ), so my messages aren’t realtime, and I just compose in Google docs, which gives me (usually reliable) grammar corrections
I generally compose in Google doc
I’m really curious if I can get it to talk to me in 1920s speech patterns, or ojou-sama type character lol. I assume it will not have the right grammar patterns or orthography for the former.
I usually don’t post in real time either, we use it more like email, which given time zones and Japanese phone messaging history makes sense.
The thing I like for gpt over the grammar correction is that there are times when what I’ve written is grammatically correct but it doesn’t feel quite natural. (What I’ve learned from both my friends and gpt is the the answer is to usually add some よねs or emoji to the end of my sentences ).
It will depend on what they ingested for Japanese sources. On the one hand, it’s free (although it’s not like a silly thing like copyright stopped them from using texts as training data ), on the other hand, how much did/do they care about non English language sources?
It’d be worth a try, honestly, but definitely be skeptical.
For my purposes, the best chatgpt prompt type has been “give me 3 different ways of expressing (some concept) in Japanese”, and then working with those as example sentences to construct what I actually want. Need to be able to understand what is dodgy though, because it can give non-obvious nonsense.
Also, LLM roleplay/storytelling is a rabbithole that, y’know - I don’t necessarily recommend that anyone goes down, but it’s pretty solid practice.
Sure but ai will give instant answers back to back I’m not completely sold on ai for language learning, but it’s something I can throw questions at without either needing to wait or/feeling like a bother at least.
(also I’d argue simply being able to read at a certain level does not mean equivalent grammar knowledge or ability to explain it well, but that’s another matter)
Hey, I don’t need convincing that ai hallucinates but I think I already wrote my thoguhts in prev comments and I’d just be repeating myself. I was actually surprised at how much people use ai in a recent book club.
3 Likes
As an Amazon Associate, Natively earns from qualifying purchases through any Amazon links on the site.
), so my messages aren’t realtime, and I just compose in Google docs, which gives me (usually reliable) grammar corrections
).
), on the other hand, how much did/do they care about non English language sources?
I’m not completely sold on ai for language learning, but it’s something I can throw questions at without either needing to wait or/feeling like a bother at least.
but I think I already wrote my thoguhts in prev comments and I’d just be repeating myself. I was actually surprised at how much people use ai in a recent book club.