Enhancing the Rationale-Input Alignment for Self-explaining Rationalization