Looking at data gathering difficulty is a good way to look at it. Though I may disagree with his classification hard / very hard / very very hard: you can defi

Looking at data gathering difficulty is a good way to look at it. Though I may disagree with his classification hard / very hard / very very hard: you can definitely work on some ranking tasks with just behavioural data (i.e. no labels). And to build decent OCR, you need to label your own datasets, the problems is not that different from speech. Also, what is hard / not hard changes w/ state of the art. Translation is a good example: it used to be that you had translate 1:1 each text. But not anymore w/ recent advanced in NLP (transformers, etc.).