Research And Application Of Self-dialogue In Dialogue Systems Based On Reinforcement Learning

Goal-driven Dialogue Systems are designed to enable machines to help the user complete the specific task through dialogue.It has been widely used in intelligent customer services of enterprises.There are many ways to train a Goal-driven Dialogue System,reinforcement learning is the most popular way at this stage.Although reinforcement learning has a lot of advantages over other methods,there are still some problems.It is necessary to build a user simulator when using reinforcement learning.At present,ways to build user simulators and train dialogue systems often have some problems,such as poor diversity and too much time-consuming.To solve these problems,this thesis has carried out the research and application of self-dialogue in dialogue systems based on reinforcement learning after extensive research.The specific work content is as follows:An asynchronous cooperative reinforcement learning is proposed for training Goal-driven Dialogue Systems.The algorithm has two advantages.First,the algorithm sends parameters from child processes to the main child for asynchronously updating the dialogue model.Dialogue models in different child processes cooperate to update the dialogue model in the main process and,user simulators in different child processes cooperate to update the user simulator in the main process.Second,the dialogue model interacts with the user simulator in the child processes.Both are trained by reinforcement learning so that the user simulator is also trained while training the dialogue system.The experimental results show that the algorithm proposed in this thesis has achieved significant performance improvements in the three indicators of time-consuming,dialogue success rate and dialogue diversity.A dialogue system is designed and implemented.The system includes a conference room booking system trained by the algorithm proposed above,and can also chat with real users.