Research On Key Technology And Application Of Task-oriented Dialogue System

Task-oriented dialogue system is an important branch of natural language processing.It is usually oriented to specific fields and obtains domain-related information through interaction with humans to assist people to complete specific tasks such as restaurant reservation,schedule setting and weather inquiry.Task-oriented dialog system has two implementation schemes based on pipeline method and end-to-end method.Among them,the implementation scheme of task-oriented dialog system based on pipeline method decomposes the system into three modules: language understanding,dialog management and natural language generation.Because of its clear structure and stable effect,the scheme has been widely used in industry.Based on the pipeline method,this paper studies the multi-domain language understanding task and dialogue management task in the task-oriented dialogue system,designs and implements the task-oriented dialogue system for the Winter Olympic Games on the basis of the research results.Firstly,in the multi-domain language understanding task,we analyze the prediction results of some baseline models and conclude the slot-domain-error problem,that is,the prediction slot of the model does not belong to the set of candidate slots under the corresponding intention of the text.To solve this problem,we propose a joint model based on intent information enhancement,which explicitly introduces intent information in the form of vectors into slot prediction process.In the process of constructing intent information vector,we have tried three schemes: bi-directional long short-term memory neural network(BiLSTM),attention-based bi-directional long short-term memory neural network(Attention-based BiLSTM)and convolutional neural network(CNN).Finally,the experimental results on Snips,a multi-domain language understanding data set,show that the joint models based on the intent information enhancement are more effective than the baseline models;the three intent information vector construction schemes,the CNN scheme improves the experimental results the most,and the Attention-Based BiLSTM scheme takes the second place.The BiLSTM solution is the lowest.Secondly,in the dialogue management task,in order to alleviate the rule-based methods‘s problems,such as poor flexibility,time-consuming and expensive expert rule making,we introduce reinforcement learning algorithm,which is expected to automatically learn dialogue policies through interaction with the environment.We have tried two deep reinforcement learning algorithms,Deep Q-Network(DQN)and Dueling Deep QNetwork(Dueling-DQN).Finally,on our own data set,we prove the superiority of dialogue management based on reinforcement learning,and observe that Dueling-DQN algorithm is superior to DQN algorithm in both convergence speed and convergence results.Finally,we put our research results into practice,design and implement a taskoriented dialogue system for the winter Olympics.In the process of system design,we analyze the user's needs and complete the definition of key elements(domain,intent and slot)of the dialog system.In the process of system implementation,we propose a semiautomatic construction scheme of annotation data for language understanding to alleviate the lack of annotated data in Task-oriented dialog systems in new fields;we design a highly scalable framework based on reflection mechanism to reduce the difficulty of subsequent expansion and maintenance of the system;We implement the natural language generation module based on rules to ensure the fluency,readability and knowledge of the system reply text.