it reminds me of the kind of tokenization that LLMs do (though that isn't strictly per word)