Research And Implementation Of Task-Oriented Dialogue System For Government Affairs

With the development of Internet and artificial intelligence related technologies,the scenes and needs of government services are gradually diversified.The number of traditional artificial windows is limited,so the development of Task-oriented dialogue system in the field of government affairs by using natural language processing and other related technologies can effectively solve this problem and effectively improve the efficiency of government services and the efficiency of the masses.However,at this stage,the Task-oriented dialogue system is less applied in the field of government affairs.At this stage,the QA system in the field of government affairs mainly depends on KBQA one question and one answer and so on.Therefore,this thesis mainly studies the construction of Task-oriented dialogue system in the field of government affairs.The main work includes:(1)Using the common characteristics of dialogue management and reinforcement learning,deep reinforcement learning is introduced into the construction of dialogue management model,and a dialogue management model based on A3 C algorithm is proposed to complete the learning task of dialogue strategy.Based on actor critic architecture,the model can dynamically improve the learning of dialogue strategies and improve the prediction accuracy of dialogue management model through internal evaluation mechanism.At the same time,the model is based on asynchronous architecture training,and can execute multiple sub networks at the same time,which can effectively improve the training efficiency of the model.(2)A joint model of intention recognition and word slot filling based on Bert and self attention mechanism is proposed.The model uses the ability of Bert to encode texts,and extract the feature information of intention and word slot through training,and then use Softmax layer to probabilistic intention to complete intention recognition,Then,the characteristics of two-way long-term and short-term memory network and self attention mechanism are used to identify the word slot of the text and complete the word slot filling task.Experimental results show that the recognition accuracy of the joint model in the two tasks of intention and word slot recognition has been improved.(3)A task-based dialogue system for government affairs is designed and implemented.Firstly,analyze the needs of users,design the system architecture according to the needs,and apply the two models proposed in this thesis.Then it defines the intention and word slot of the field involved in the system,defines the template generated by natural language,and completes the implementation of the system.Finally,the function and performance of the system are tested.