Model Interpretability and Rationale Extraction by Input Mask Optimization