Abstract

Recent advances in deep learning based approaches enabled exciting new research frontiers for goal-oriented conversation systems. In this talk I will present an end-to-end dialog system, with components for language understanding, dialogue state tracking, policy, and language generation can be independently built and jointly optimized for dialogue quality and efficient task completion. I will focus on two components: language understanding and dialogue policy learning. I will present a joint, multi-domain recurrent neural network approach for language understanding and how we exploit context via end-to-end memory networks in multi-domain language understanding. Then in the second part, I will discuss policy learning and integration of immediate feedback via interactive reinforcement learning.