Research And Implementation Of Task-oriented Dialogue System In Low-resource Scenarios

The rapid development of artificial intelligence and the popularization of smart devices have brought broader landing scenarios and richer technical challenges to taskbased dialogue systems,so they have received widespread attention.More and more business fields have begun to adopt intelligent customer service to reduce labour costs on simple tasks.However,most fields lack high-quality corpus to support the development of task-based dialogue systems based on deep learning,that is,low resources scenarios with scarce data.In low-resource scenarios,the scarcity of intent and slot samples limits the accuracy of the natural language understanding model,and the lack of dialogue sequence annotation data will make the construction of policy models face the problem of cold start,confusing the prediction of dialogue actions.Aiming at the above problems faced by task-based dialogue systems in low-resource scenarios,this paper studies and implements a task-based dialogue system for lowresource scenarios.The main work of the paper includes the following three points:(1)A natural language understanding model based on prompt tuning is proposed.It adopts the learnable template in the form of natural language,extracts implicit knowledge from the pre-trained language model,and combines the tasks of intent recognition and slot extraction for multi-task learning,which solves the problem of low recognition accuracy in low-resource scenarios.The model achieves a slot accuracy rate of 72.3 and an intent accuracy rate of 70.4 under the few-shot setting of the SNIPS dataset,outperforming traditional methods based on fine-tuning.(2)A deep learning dialogue policy model based on dialogue flow bootstrapping is proposed.Use the state machine to design the initial dialogue flow and convert it into the dialogue sequence annotation data for the deep learning dialogue policy model to learn,to solve the cold start problem of the dialogue policy model,and combine the machine teaching module to make the model easier to maintain and extend.(3)Based on the above improvements,this paper implements a complete taskbased dialogue system display platform for low-resource scenarios and applies it to the intelligent customer service chatbot in the scientific and technological resource integration project.The platform includes natural language understanding data and model management,dialogue strategy data and model management,log management and other sub-services.Compared with the traditional task-based dialogue system,it has higher practical value.