I understand training sentence piece model in monolingual case. But in multilingual case, its not clear enough. It's because dataset sizes across languages varies greatly. I think this leads to biased shared vocabulary.
Is it using sampling technique while training sentence piece as well?
If yes, how many times is sampling performed?
Isn't it better to go through all the text in dataset to create sub-words vocab instead of just the samples?
I understand training sentence piece model in monolingual case. But in multilingual case, its not clear enough. It's because dataset sizes across languages varies greatly. I think this leads to biased shared vocabulary.