A reinforcement learning approach that separately optimizes local component-level judgments and global alignment scoring.