Hierarchical Reinforcement Learning and Hidden Markov Models for Task-Oriented Natural Language Generation