Thanks! Earlier in the project I tried to find a way to generate English but with Chinese grammar - something like “you need to grab your shoelaces and tie them well” - as a way to help memorize grammatical structures. But I didn’t get very far with that. For one reason, I couldn’t coax an LLM to output such unnatural English (!)
So in the end I decided that I can ask GPT-4 to explain alternative structures based on the translation I get
Ideally, I would like to find a reliable way to directly transform my first attempt at an expression into a native-level expression, without necessary routing via English. It didn’t seem doable with the LLMs I tried so far, but there are others I can try, and maybe some training would help too - as you say
I use GPT-4 for this kind of exposition too. If you haven’t come across Pleco [1] yet, I can highly recommend it. It’s a high quality Chinese-English dictionary with a lot of helpful features for reading Chinese
One feature I use a lot is the clipboard reader, which lets you step through Chinese text character by character and shows you the meaning of each character. It can be faster than using an LLM, but you don’t get the whole-sentence meaning or the ability to ask follow-up questions like you can with an LLM. But it’s a great app all round (some features are paid)
I also use GPT-4 for explaining the meaning of sentences in more detail (as in JimDabell’s comment). Often my questions are like “how would a native speaker say this colloquially” - I’ve found it really valuable to be able to have a back-and-forth on why something works the way it does
Just for fun, I’m trying to force it to misinterpret what I’m saying by changing “dei tie” to “di tie” (as in subway). It really doesn’t want to interpret it differently! I think “shoelaces” must be such a strong signal of context for the LLM
After working on this project, I’ve started using MS translate as my go-to app if I want to translate an existing sentence. Previously I tended to use Google Translate by default. But I’ve genuinely been surprised by the differences between the apps
I have seen some variation in the interpretation - although it does seem to be pretty consistent. When I enter “服了u” into unscrambler I get the interpretation “I give up” every time I try it
So in the end I decided that I can ask GPT-4 to explain alternative structures based on the translation I get
Ideally, I would like to find a reliable way to directly transform my first attempt at an expression into a native-level expression, without necessary routing via English. It didn’t seem doable with the LLMs I tried so far, but there are others I can try, and maybe some training would help too - as you say