Approximate Distribution Matching for Sequence-to-Sequence Learning