Have you tried approaching this with the kanji instead? That seems like free tokenisation.