Document-Driven Task-Oriented Dialogue System

The task-oriented dialogue system aims to help user complete one or more specific tasks by interacting with user in several turns in natural language.The task-oriented dialogue systems are widely used in different types of smart terminals and applications for improving the interactive experience,such as Xiao Mi XiaoAi,Baidu DuMi,Ali XiaoMi,etc.The mainstream solution of the task-oriented dialogue systems is the slot-filling,which collects information from users to fill the slots by interacting with users in several turns in natural language.It needs to build the dialogue ontology by the experts of business firstly,collect and label the dataset based on the dialogue ontology,finally train the model of dialogue and build the dialogue system.Although the slot-filling method are widely used in different task-oriented dialogue systems,there are three problems of this method:1.The process of designing dialogue ontology depends on the knowledge of expert,which makes the development of dialogue system cost more human resources and time.2.The transfer and reusable capability of dialogue system is week.The dialogue ontology and labeled dataset in specific domain cannot directly used in another domain.3.The relationship among the slots is flatten,which limits the ability of dialogue system to accomplish complex tasks.This paper proposes a document-driven task-oriented dialogue system model,which completes the tasks of users by interacting with users in several turns,based on the document describing the business logits.Specifically,this model first builds the dialogue ontology by parsing the structure of the document of business,and then complects the dialogue by modeling the dialogue history,making the dialogue strategy every turn and natural language generation.It decreases the costs of time and human resources,because it does not depend on the dialogue ontology designed human expert.It improves the capability to transfer dialogue system from different domains by replacing the document of business instead of designing again.It improves the ability of dialogue system to accomplish complex tasks,using the structure of document modeled by two relationenhanced graph attention networks designed in this paper for building dialogue ontology.In this paper,the validation and feasibility of this model is verified by a series of experiments and analysis.Based on the above works,this paper designs a task-oriented dialogue system in political domain.This system is able to help users complect the question by multi-turn dialogue.