Towards Usable Neural Comment Generation via Code-Comment Linkage Interpretation: Method and Empirical Study