Generative Encoder-Decoder Models for Task-Oriented Spoken Dialog Systems with Chatting Capability.