Research And Application Of Dialogue Management Model Based On Deep Reinforcement Learning

Goal-driven dialogue systems are designed to provide the man-machine interactive way based on natural language,these systems have been widely used in smart customer services,personal assistant software,and many other tasks.A goal-driven dialogue system usually has three components:natural language understanding,dialogue management and natural language generation.Dialogue management is the core part in a goal-driven dialogue system,which is responsible for controlling the state and flow of the conversation.Because these three components are independently built in a pipelined way,there are some important limitations exist in the traditional dialogue systems,for example,the accumulative errors problem.In order to solve these problems,many researchers have proposed models to deal with these components jointly.In this thesis,the author is focusing on the joint model for natural language understanding and dialogue management.First,this thesis proposed a joint model for natural language understanding and dialogue management based on deep reinforcement learning.There are three networks in the model.At the bottom of the model,there are two recurrent networks which encode sentences into the current dialogue embedding.At the top of the model,it is an action scorer based on the deep neural network.By using these cascaded networks,the model maps text inputs to dialogue actions in an end-to-end method and it is trained by deep reinforcement learning method.Second,this thesis implements a man-machine dialogue system for meeting-room booking task,and the dialogue system uses the proposed model to replace traditional natural language understanding and dialogue management modules.Experimental results show that the model proposed in this thesis outperforms both traditional Markov decision process model and other joint models.Visualization of dialogue embeddings illustrates that the model can learn the representation of the dialogue states.