Research And Application Of Dialogue Management In Task-based Dialogue System

With the development of artificial intelligence technology,the interaction between human and intelligent devices is becoming more and more intelligent.People can interact with intelligent robots through multi-turn dialogue for completing a task or a service.Dialogue management,as an important part of human-machine dialogue system,includes two parts: dialogue state tracking and dialogue policy.Dialogue state tracking module tracks and updates the status of the dialogue in real time.Dialogue policy module analyses the response action of the system based on the status tracking of the dialogue.The performance of dialog management directly affects the performance of dialog system.Therefore,the research and application of dialog management in Task-based dialog system is of great significance.Dialogue state tracking task faces the problems of sparse data and user target slot carry-over.The main problems faced by the dialogue policy module are the lack of cold start of training data and how to build a dialogue strategy model considering the search results of knowledge base.In this paper,a series of work has been carried out to solve the above problems,including:Aiming at the problem of sparse data and user target carry-over in dialogue state tracking,this paper constructs a joint model of natural language understanding and dialogue state tracking.The model uses a coder based on bidirectional recurrent network to encode system actions,user utterance and candidate slot-value pairs.The experimental results on open English datasets and Chinese datasets show that the proposed model achieves better performance than the existing models.Compared with the comparative model,the accuracy of joint target is improved by 17.5%,4.6%,3.9%,4.1% and 0.2%,respectively.The model effectively alleviates the problem of data sparsity,and can capture the change of slot value when the user's target is carried over.In view of the cold start problem of dialog policy task,this paper proposes a dialog policy algorithm which integrates Markov Decision Process and attribute information entropy,and constructs a dialog policy model considering knowledge base search results.Experiments on multi-turn dialogue data in the field of music search show that the fusion algorithm can improve the task completion rate of the first three turns of dialogue,and effectively shorten the number of turns of dialogue of music search task,which is 2.24,0.84 and 0.03 shorter than random strategy,rule-based strategy and information entropy-based strategy respectively.Finally,the dialogue state tracking model and the dialogue policy algorithm proposed in this paper are applied to the music search system based on multi-turn dialogue,and the dialogue management part of the dialogue system is realized.By demonstrating the functions of the system,the feasibility of dialogue state tracking based on hierarchical structure coding and the strategy of integrating Markov Decision Process and attribute information entropy in task-based dialogue system is verified.