Hi Ko, very interesting article. Did you apply the “romanji” step to the original sentences with mecab ? If so how can this step be improved ? To my understanding kanji can be read as kunyomi or onyomi, and mecab seems to be a rule based tool, and I didn’t find any context based tool. I wonder if it is possible to create an accurate model that learn how to apply this “romanji” step based on context. Cheers