End--to--End Task--oriented Dialog System Based On Knowledge Fusion

Conversational systems help people in many ways by communicating with them in natural language.In particular,task-oriented conversation systems aim to accomplish a user's goal(for example,restaurant or airline reservations)in the fewest conversation sessions.The earliest systems were designed with lots of expert hand-crafted rules and templates,which were expensive and limited.As a result,data-driven statistical conversa-tional systems,including powerful neurosystem-based ones,have received considerable attention in the past few decades to reduce costs and provide robustness.One of the main challenges in constructing neural task-oriented conversation systems is to model long conversation contexts and external knowledge information.Some neu-ral conversational systems are modular.Although they are considered stable and easy to interpret,they often require expensive manual tags for each component and have unnec-essary module dependencies.The end-to-end approach,on the other hand,automatically learns the hidden conversation representation and retrieves/generates the system response directly.They require less human involvement,especially in data set construction.How-ever,most existing models incorporate too much information in an end-to-end learning framework.In this paper,we focus on task-oriented dialogue system based on deep learn-ing model,which is an important research direction in the field of natural language pro-cessing.Firstly,a multi-turn memory model based on dynamic attention is proposed in this paper.The model is divided into two parts.The dynamic memory dialogue module is used to encode the context dialogue.Knowledge memory model is responsible for cap-turing knowledge information.The independent modeling of the two modules and the information interaction between rounds can better understand the changing intentions of users and give reasonable answers.Secondly,for the problem Of Out Of Vocabulary(OOV),this paper proposes an end-to-end dialogue model based on graph attention network.The model uses the graph structure to model the external knowledge and preserves the structural information in the external knowledge so as to obtain a better entity representation.The potential structural relations in the knowledge base are combined with the semantics of the historical dialogue,and better auxiliary reasoning is used to get high-quality answers.Finally,we propose a multi-modal interactive Transformer model based on Trans-former model for multi-modal task-type dialogue with picture information.In this model,different modes are modeled separately,and special mask schemes are used to make the information between different modes interact,and finally generate high-quality responses.