SMRT Chatbots: Improving Non-Task-Oriented Dialog with Simulated Multiple Reference Training