A Rationale-Centric Framework for Human-in-the-loop Machine Learning