Research And Application Of Mulit-Task Dialogue State Tracking

The goal of task-oriented human-machine dialogue system is to complete specific tasks through multiple rounds of dialogue,such as booking a hotel,ordering a meal,etc.Multi-task dialogue means completing multiple tasks in one dialogue.Dialogue state tracking is a key module in the task-oriented dialogue system,and its goal is to obtain the dialogue state at the current moment,that is,the completion of slot filling.Multi-task dialogue state tracking faces an important challenge:the number of slots in multi-task increases linearly with the number of tasks.If a separate module is designed for each slot for tracking,the model size will increase linearly with the number of slots.Aiming at this problem,this paper conducts research on the basis of fully investigating the existing related work.The main work includes the following two aspects:First,a multi-task dialogue state tracking model based on task slot graph is proposed.The model uses the combination encoding,adds the encoding of tasks and slot types in the encoding module,and interacts with the dialogue content to obtain task vectors and slot type vectors,and combine the two to obtain slot vectors;the task-slot relationship graph is introduced.The masked multi-head attention layer realizes the interaction between tasks and tasks,tasks and slot types,slot types and slot types based on Graph-structured;Adopts the method of extracting slot value span based on dot product,and uses slot vectors to extract slot value spans;all slots are Sharing the same slot gate module,reference classification module and slot value span extraction module realizes parameter sharing between slots,so that the model size does not increase linearly with the increase of the number of slots.Experimental results on the public MultiWOZ 2.1 multitask dialogue dataset show that the model achieves higher joint accuracy than other baseline models.Second,the multi-task dialogue state tracking model based on task slot graph proposed in this paper is used to build a task-oriented humanmachine dialogue system.The test results show that the task-oriented human-machine dialogue system can meet the expected requirements.