Research On Key Technologies And Model Implementation Of Task-based Dialogue System

As two important research directions of Task-based dialogue system,the end-to-end task-based dialogue system and the task-based dialogue sys-tem based on reinforcement learning have attracted more and more schol-ars' attention.A typical task-based dialogue system consists of several in-dependent modules,which are complex in structure and difficult in coor-dination.The end-to-end task-based dialogue system uses the end-to-end neural network to integrate the task-based dialogue system.However,the current research still needs additional dialogue state tracker,which fails to form a complete training integrated end-to-end architecture,and relies too much on the order of recurrent neural network Structure,which leads to time dependence and sequence dependence,restricts the training speed of Task-based dialogue system.The task-based dialogue system based on re-inforcement learning adopts reinforcement learning to keep the optimiza-tion direction of the task-based dialogue system consistent with the opti-mization index of the user's needs.However,the current reinforcement learning task-based dialogue system has poor robustness and brings users poor experience.Moreover,the reinforcement learning training mode does not directly depend on natural language,but on natural language Therefore,a natural language understanding module with good semantic analysis abil-ity is needed.In view of the problems existing in the two research directions of the above-mentioned task-based dialogue system,this paper proposes a full parallel end-to-end task-based dialogue system based on convolution neu-ral network and attention mechanism to build a task-based dialogue system that can train integration and fast;and proposes a deep reinforcement learn-ing task-based dialogue system combined with natural language under-standing to build a task-based dialogue system that can understand natural language A more powerful and robust task-based dialogue system.The specific work includes the following three points:(1)A task-based dialogue system based on convolutional neural net-work and attention mechanism is proposed.This task-based Dialogue framework aims at the multi round ability of Task-based dialogue,designs a copy network based flow context modeling method to improve the multi round dialogue modeling ability,and designs a full parallel codec Dialogue framework based on convolutional neural network and attention mecha-nism to solve the sequence dependence and time dependence problems caused by recursive neural network of codec structure High model training speed,and based on beam search and policy gradient to optimize the dialog results.In this paper,the experimental demonstration is carried out on the open data set CamRest676 and KVRET of the task-based multi round dia-logue.The results show that the scheme in this paper achieves the optimal effect on multiple indicators of the task-based dialogue,especially on the training speed,compared with the existing scheme;(2)This paper proposes a task-based dialogue system of deep reinforce-ment learning combined with natural language understanding.This task-based dialogue system framework aims at the problem that the current nat-ural language understanding does not make full use of the interaction be-tween intention recognition and slot value extraction,and proposes a joint modeling natural language understanding model based on the two-way in-teraction of intention slot value to enhance natural language understanding.At the same time,the current reinforcement learning task-based dialogue system is not robust enough and the task completion rate is not high For high-level problems,a task-based dialogue model of double DQN deep re-inforcement learning is adopted.In this paper,experiments are carried out on the natural language understanding open data set ATIS and Snips and the public movie ticket booking data set movie ticket booking.The results show that the model in this paper achieves the best effect on the natural language understanding index,and has a better effect on the task comple-tion rate of booking movie tickets;(3)On the basis of the above two kinds of researches on Task-based dialogue system,this paper uses Django website construction technology to complete the construction of end-to-end task-based dialogue system and task-based dialogue system demonstration platform based on reinforce-ment learning.When a user enters the corresponding natural language di-alogue text through the front-end interface,the background automatically calls the corresponding dialogue module to train the model to get the phase It should output,return the end-to-end task-based dialogue results,natural language understanding results,task-based dialogue results based on rein-forcement learning to the user,and conduct multiple rounds of dialogue interaction.