I think Chinese is really not a pictorial language either; using your apt description, it's like if every character was a "shortcut". Harder to learn, but still fairly standard in conveying meaning as far as modern languages go. Some characters are still visually similar to what they represent, but at this point they're mostly a bit of a stretch.

A true pictorial language would convey most meaning through the symbols themselves, and I don't think any modern languages fulfill that definition. Maybe sign languages are the closest thing we have to pictorial language, in terms of the way some things are expressed symbolically?

Right so under the hood do you normalize to romajii? How do you deal with kunyomi vs onyomi - do they even matter? Even if you had a stream of kana, how do you tokenize when they're not delimited?

(don't know much about NLP beyond surface level software experience but have an amateur interest in Japanese)