Design And Implementation Of Task-oriented Dialogue System Based On Attention Mechanism

Human-Machine Conversation is the direction of several generations of computer and artificial intelligence researchers and practitioners to work and strive for,the current dialogue system is usually divided into chit-chat dialogue system,FAQ dialogue system and task-oriented dialogue system,this paper mainly focuses on the taskoriented dialogue system.There are two types of task-oriented dialogue systems: end-to-end and pipeline.The pipeline dialogue system usually consists of three modules,namely the dialogue state tracking module,the policy learning module and the reply generation module.When the existing task-oriented dialogue system communicates with users in multiple turns,it is often difficult to identify intents and slots,and cannot accurately convey information to users.At the same time,because the data acquisition cost of task-based dialogue systems is high,how to make dialogue systems still have certain effects in new fields with lack of data or even no data is a research hotspot in recent years.In this paper,the research is based on adding the description information of the dialogue task schema to the task.Through the model based on the attention mechanism and the method of using the runtime computing task structure and the information interaction of the dialogue content,the model can still use the organizational information of the task schema to identify the user’s intention and extract the slot value in the case of insufficient data or even no data.At the same time,the ADDST model proposed in this paper can better solve the problem of intent recognition and slot filling in the case of multiple turns of dialogue,especially in the classification type slot filling,which not only improves the model effect,but also reduces the amount of calculation.This paper also proposes a method to implement the policy learning module and the dialogue generation module in the dialogue system,the models used in these modules all have good results in the new field of zero-shot training.Using the implementation method of the dialogue state tracking module,policy learning module and reply generation module proposed in this paper,could build a dialogue system in pipeline structure which could talk multiple turns of to users smoothly.Finally,this paper designs a task-oriented dialogue system in the cabin scene.The system consists of a dialogue system terminal deployed in the cabin and a dialogue system server deployed in the cloud.The two systems work together to complete the entire dialogue system function,the function of the dialogue system deployed in the cloud is realized by the research results of the multi-turn dialogue system based on the attention mechanism.