Research On Diverse Response Generation For Opden-domain Dialog System

As one of the important problems in the field of artificial intelligence,dialogue system has important research significance and application value in today's era of deep learning.This field is not only an important part of human-machine interaction research,but also has rich applications in intelligent customer service,personal assistant,leisure and entertainment scenes.The research of this thesis is based on an important branch of dialogue system-open domain dialogue system,focusing on the improvement of generative model.In recent years,the research of generative models in chatty conversations has made a series of development and progress,but there are still some problems and challenges: Generating universal responses,lack of context consistency,lack of background knowledge,etc.Neural network-based conversation models tend to produce trivial or inconclusive responses,usually including high-frequency phrases like "I don't know" or "I'm fine";And the bot's responses are inconsistent,often contradicting what the user has already said,or conflicting descriptions of the system itself.This thesis focuses on the lack of semantic and topical diversity of reply sentences caused by the current situation of universal reply generated by dialog system,and carries out relevant research.This thesis proposes the solution of this problem from three angles.Firstly,from the perspective of diversity based on the state of the dialogue,the ductility of the communication topic is improved by identifying multiple candidate topics.Then consider how the introduction of role information can improve the diversity of responses.Then,from the modeling perspective of the generated model,the dialogue is modeled as a one-to-many relationship,which is different from the one-to-one mapping of the machine translation model.The main work of this thesis includes the following aspects:(1)Multiple response generation based on conversation state.The responses generated by the generative model usually do not involve the topic(that is,the status information in the chat conversation),and do not focus on a specific topic,which makes the responses lack the diversity of topic level.In this thesis,state recognition and reply generation is modeled as a two-stage task.Considering the category information of state,different levels of relationship extractors are set to use the category information in a global local way.Finally,the effectiveness of the method is verified by experiments.(2)Multiple response generation based on system roles.Existing generative models usually consider the history of the dialogue,but the representation of the robot's personality(role)will be inconsistent,and the introduction of the role can improve the diversity and richness of the dialogue.According to the character of character information encoding,memory network is introduced to store the character information,and pointer network is used to retrieve memory information.Finally,the validity of the method is verified by experiments.(3)Multiple response generation based on sentence pattern control.The generative model,which builds a one-to-one mapping of user messages to responses,is suitable for machine translation scenarios but needs to be improved in dialogue scenarios.Language interaction in a dialogue scenario is a one-to-many mapping.Referring to the concept of sentence pattern control,this thesis can control the sentence pattern of reply sentence in the dialogue process to bring diversity.Finally,the validity of the method is verified by experiments.