Research On Dialogue State Tracking In Task-Based Dialogue Systems Based On System Action

Task-oriented human-machine dialogue system is usually oriented to vertical domains,with clear user requirements,and aim to complete specific tasks.In the actual humanmachine dialogue,the dialogue system cannot get all the information needed to complete the user’s task from a single round of dialogue.Therefore,task-oriented human-machine dialogue system needs to rely on the dialogue state tracking module to record and update the user’s dialogue state.Accurately tracking the dialogue state can help the system understand the current task process,and provide a basis for the system to formulate correct and reasonable dialogue policies,so as to better help user complete the task.In recent years,with the rapid development of artificial intelligence technology,the dialogue state tracking model based on deep learning can directly obtain the user’s dialogue state according to historical context information and the current dialogue,and has achieved good results.These methods usually track the dialogue state according to the response generated by the natural language generation module at the previous moment and the current user dialogue,without using the system action information at the previous moment.However,system action often contains explicit types of system action request by system(e.g.,request,acknowledge,etc.)and the slot information needed to complete the task,which is helpful to better identify the slot value pairs that need to be updated in the dialogue state.Based on this,this paper explores the use of the previous system action and the user’s current response to track the user’s dialogue state.The main work of this paper is as follows:(1)proposes a dialogue state tracking model based on system action.Different from most dialogue state tracking methods,this paper directly takes the previous system action and the current user dialogue as input.The model contains a slot operation classifier and a slot value generator.The slot operation classifier uses the BERT model as the encoder,and integrates the information of system action into the slot representation through the selfattention mechanism of BERT to obtain a richer slot representation,so as to identify the slots that need to be updated in the dialogue state.The slot value generator uses a Pointer Generation Network to predict the slot value corresponding to the slot that needs to be updated.In the comparative experiments,the effect of system action on dialogue state tracking is analyzed,and it is proved that the proposed model can effectively use the system action information and improve the effect of dialogue state tracking.At the same time,the experimental results also show that the performance of the slot operation classifier has an important impact on the accuracy of dialogue state tracking.(2)Considering the importance of slot operation classifier for dialogue state tracking,this paper further improves the performance of slot operation classifier,and proposes a dialogue state tracking model based on system action gate mechanism.The proposed model directly incorporates system action into the slot operation classifier to help the slot operation classifier identify slots that need to be updated.Since the influence of system action on each slot in the dialogue state is different,this paper vectorizes the system action and designs the gate mechanism based on the system action.The gate mechanism can automatically learn the correlation weight between the system action and each slot in the dialogue state,and integrate the weighted system action vector into the slot representation to obtain a better slot representation,so as to improve the accuracy of the slot operation classifier.The experimental results show that the proposed model improves the accuracy of slot operation classifier and further improves the accuracy of dialogue state tracking.