I’m building something similar for Chinese[0]. Every time a new language model was released, I tested it with pinyin conversion (the standard Mandarin phonetic romanization), and they all failed miserably - until ChatGPT, which so far has performed flawlessly.
Sentence generation isn’t perfect though (at least with my rudimentary prompting) - it will regularly generate “Chinglish” phrases (like word-for-word transliterations from English which are not native Chinese) and occasionally just regurgitate nonsense. It’s not something that I would freely unleash on the end user but there is definitely a place for it somewhere in the pipeline in future.
If you wanted pinyin output, you could use a library like pypinyin? (we use it, no complaints). I assumed GPT models performed best in English, you could machine translate the input/output. MT is pretty fantastic these days, although, you introduce fidelity issues, and an API adds to cost and latency.
I didn't actually need the pinyin conversion (I implement that myself), it was just a test to see how well LLMs* were progressing at "out-of-domain" tasks. I always used recent newspaper headlines, so the possibility of finding matching character/pinyin pairs somewhere in the huge training set was negligible.
For this specific task, Chat-GPT is the first model I've come across that actually shows some ability to chain together concrete steps to form a task (i.e. look up character, fetch pinyin), rather than the simplistic "predict the next word in the sequence" in pure language models.
* Large Language Model - I know ChatGPT isn't a pure language model (which probably explains why it performs so much better), but I don't know what we're calling these models now so I'm sticking with LLM for now.
Even if you just had a set of lessons to run through, that'd be a lot of fun. Prompts that make it the counter party in text book dialogues. Character, pinyin, and translation shown.
The email sign up didn't seem to work on your page.
That's basically the idea! Thanks for the note on the email signup, I'll have a look - I threw it together very quickly just before I went on a plane a few weeks ago so it's very fragile.
Thanks! I agree, I'll stay away from characters, at least at an early/beginner level (although I am including them mostly for aesthetic purposes). It's a huge hurdle that lots of people don't necessarily need/want to overcome.
Sentence generation isn’t perfect though (at least with my rudimentary prompting) - it will regularly generate “Chinglish” phrases (like word-for-word transliterations from English which are not native Chinese) and occasionally just regurgitate nonsense. It’s not something that I would freely unleash on the end user but there is definitely a place for it somewhere in the pipeline in future.
[0] https://polyvox.app