Research On Multi-domain Dialogue System Based On Deep Learning

With the continuous development of artificial intelligence technology,intelligent robots have gradually penetrated into people’s daily life while changing the way of social production.In order to make it easier for users to use intelligent robots,the research of human-machine dialogue system is urgent.With the continuous development of deep learning,researchers have integrated deep learning technology into the study of humanmachine dialogue systems.Studies have shown that the human-machine dialogue system based on deep learning has achieved more excellent results.However,existing methods cannot accurately predict unknown slot values,and most models require a large amount of training data,which affects the accuracy of the dialogue system to a certain extent.Therefore,this paper studies the related problems of end-to-end task-oriented multi-domain humanmachine dialogue systems from two aspects: improving the accuracy of dialogue systems and simulation implementation.First,for the feature extraction problem of multi-domain dialogue,this paper proposes a dialogue state tracking model based on multi-encoder structure for multi-domain dialogue.In this model,two kinds of encoders are set up to extract the feature information in the conversation.In addition,the antagonism training method is adopted to support the shared encoder to generate the domain shared feature,and the private encoder is used to extract the feature information of each domain.In addition,this model uses a dynamic fusion network to extract the correlations between all domains to further utilize the fine-grained correlation information between domains,making it possible for domain expansion tasks.Secondly,aiming at the correlation between intent detection task and slot filling task in dialogue system,this paper proposes an end-to-end model based on Stack-Propagation framework.Stack-Propagation framework is used in decoder to add the results of intention detection to the input of slot filling task,that is,the results of intention detection were used to further guide the progress of slot filling task.The model can not only use the correlation information between the intent detection task and the slot filling task,but also achieve the effect of mutual promotion through joint learning.Finally,in order to further verify the proposed method,this paper simulates the process that the robot obtains user instructions through man-machine dialogue and helps users complete the task of picking up specified items.During the task demonstration,it mainly involves three aspects: the design and construction of the simulation environment,the speech recognition and synthesis,and the multi-point navigation.In the demonstration of this paper,the ROS platform is selected to build the simulated home environment used in the experiment,and the Turtlebot robot is equipped with speech recognition and synthesis functions for related implementation.