Dialogue systems are attracting more and more attention recently. Dialogue systems can be categorized into open-domain dialogue systems and task-oriented dialogue systems. Task-oriented dialogue systems are designed to help user finish a specific task, and there are four modules, namely the spoken language understanding module, the dialogue state tracking module, the dialogue policy module and the natural language generation module. One of the most important modules is the dialogue policy module, which aims to choose the best reply according to the dialogue context. In this thesis, we focus on the dialogue policy of task-oriented dialogue systems.

Reinforcement learning is usually used in the dialogue policy. However, traditional reinforcement learning algorithms rely heavily on...[ Read more ]

Dialogue systems are attracting more and more attention recently. Dialogue systems can be categorized into open-domain dialogue systems and task-oriented dialogue systems. Task-oriented dialogue systems are designed to help user finish a specific task, and there are four modules, namely the spoken language understanding module, the dialogue state tracking module, the dialogue policy module and the natural language generation module. One of the most important modules is the dialogue policy module, which aims to choose the best reply according to the dialogue context. In this thesis, we focus on the dialogue policy of task-oriented dialogue systems.

Reinforcement learning is usually used in the dialogue policy. However, traditional reinforcement learning algorithms rely heavily on a large number of training data and accurate reward signals. Transfer learning can leverage knowledge from a source domain and improve the performance of a model in the target domain with little target domain data. However, traditional transfer learning methods focus on supervised learning setting, and they cannot handle knowledge transfer in reinforcement setting since they do not consider the states. Transfer reinforcement learning (TRL) aims to transfer dialogue policy knowledge across different domains. In the target domain, the state and action can be aligned to the source domain state and action, so the dialogue policy can be transferred from the source domain to the target domain.

The key to transfer reinforcement learning is learning to build the mapping between the source and the target domains, and transfer only domain independent common knowledge while minimizing the negative transfer caused by the domain-dependent knowledge. In this thesis, we propose a unified framework for transfer reinforcement learning problems in task-oriented dialogue systems, including 1) How to transfer dialogue policies across different users with different preferences in personalized task-oriented dialogue system? 2) How to transfer fine-granularity common knowledge when the common knowledge is mixed with the domain-dependent knowledge? 3) How to transfer dialogue policies across dialogue systems built with different sets of speech-acts and slots? We will use both large-scale simulations and large-scale real-world datasets to valid this research. The thesis will also discuss the latest progress in the field and point out some future directions for future investigation.