Research And Implementation Of Task-oriented Dialogue System Integrating Attention Mechanism

With the development of science and technology,task-oriented dialogue system,as one of the important applications of natural language processing technology,has attracted the attention of many industries.Task-oriented dialogue system aims to complete specific service requirements,obtain user information in multiple rounds of interaction,and ultimately help users to achieve specific tasks.The implementation methods of task-oriented dialogue system are divided into pipeline method and end-to-end method.The pipeline method is composed of spoken language understanding module,dialogue management module and natural language generation module.Each module can be implemented by different methods so that the module method is more relevant to the module function,and the data requirement of the module is also lower than the end-to-end method.Therefore,pipeline method is the mainstream implementation method of task-oriented dialogue system.There are some problems in the application of open academic research in the process of system development.First,in the natural language understanding task,the capture of task related information by model needs to be optimized,and the accuracy of intention recognition and semantic slot filling still needs to be improved.Second,in the dialogue management task,due to the lack of data sets,the development of the module is faced with the problem of cold start.For the natural language generation module,the method based on neural network relies on data-driven,and the results are random,which will bring uncontrollable risks to the dialogue process.The rule-based method can solve the generation problem simply and effectively,and the manually defined rules are easy to understand and improve,so it is a better choice to apply to the system implementation.Therefore,this thesis mainly researches the problems of the first two modules,proposes new models or methods and verifies them.This thesis first studies the joint model of intent recognition and slot filling of spoken language understanding task,and proposes CIASC-BiLSTM model which integrates attention mechanism.This model uses self attention mechanism to capture more semantic information in the process of encoding and decoding,and proposes cross mechanism to enable two tasks to share context information.Experimental results show that the model has excellent performance in intent recognition and slot filling tasks.In addition,this thesis also proposes a dialogue management method based on slot framework and state machine.In the case of cold start,a flexible and extensible dialogue model can be customized according to the business which can maximize the use of input information from user and optimize the dialogue experience.Finally,the proposed research method is applied to the spoken language understanding module and the dialogue management module of the task-oriented dialogue system based on the pipeline method.A dialogue system of the doctor appointment business is developed and tested which verify the feasibility of the research method in the implementation of the task-oriented dialogue system.