| The implementation of artificial intelligence and the promotion of smart devices are changing human beings' lives,and more and more users have obtained efficient and convenient services through human-machine conversation.The performance of natural language understanding is crucial for evaluating the intelligence of the dialogue system,and it will directly affect the subsequent dialogue state tracking and dialogue generation tasks of the dialogue system.At present,due to the lack of public datasets in specific tasks,the colloquialism of user expressions,the diversity of semantics,and the implicitness of intentions,Chinese semantic understanding has attracted much attention.Therefore,the research of semantic understanding in task-specific dialogue systems has great scientific and application value.In view of the current dilemma of Chinese semantic understanding,we build a natural language understanding data set for Chinese task-specific dialog systems.Based on the basic data,it is extended and labeled by entering sentences into search engines to find similar sentences and artificially making sentences.The expanded data set includes seven fields,each of which contains corresponding domain labels,intent labels and slots,the total amount of data is 53,327.In addition,we propose an end-to-end joint detection model based on three tasks: domain classification,intent detection and slot filling,the sentence accuracy of the joint model based on ERNIE is95.27%.Additionally,comparison experiments of the basic model method are performed to verify the feasibility and effectiveness of the proposed model.In order to apply our method to application,we use existing health service application,develop the system architecture of a smart ssistant,Xiaozhi Assistant.Xiaozhi Assistant's system architecture mainly includes input and output modules,intelligent question answering modules and service composition modules.For each module,a corresponding solution is proposed. |
PDF Full Text Request