Few-Shot Structured Policy Learning for Multi-Domain and Multi-Task Dialogues