Research On Automatic Response Generation For Chat-Oriented Systems

With the rapid development of internet techniques and the boom of social network platforms,the communication between people has gradually shifted from purely face-toface interaction to social network platforms.Benefiting from this transformation,a large number of one-turn or multi-turn conversations were accumulated on these platforms.These dialogues not only provide corpora to the dialog research,but also make it pratical to build automatic dialog systems.More importantly,the regularity and pattern of real conversations would motivate the research progress of dialog techniques.Futhermore,there are many application scenarios of dialog systems,such as customer service and personal assistant.Dialog systems are able to reduce human resources and promote the efficiency of one's work and life.Therefore,in-depth studies of the dialog system would be of great importance for the researches of natural languange processing techniques and the development of internet related business.This disseration focuses on the essential problems of selecting or generating response in the dialog system.Our research investigates influences of the dialog procedure,and their functions in response generation task.Then,these influences are employed to address challenges of information retrieval based dialog system or generative dialog model.In detail,the major contents of this disseration include the following four parts.(1)As one of essential factors of dialog understanding,the background knowledge provides topic clues for dialogues.These clues are helpful for the deepning and extension of conversation topics and guarantee the topic convergence of conversations.Meanwhile,there exist topic divergency problems in information retrieval based dialog systems.This disseration presents a method incorporating the background knowledge to address this issue.After investigating the characteristics of background knowledge in dialog scenario,this disseration proposes an apporach to extract background knowledge from the dialog corpora.Then,this disseration incorporates the background knowledge into the response ranking procedure.In order to take fully use of background knowledge,this disseration presents a recall mechanism for selecting appropriate knowledge in response ranking procedure.The experimental results have indicated that the proposed model is able to improve the response ranking performance using clues provided by related background knowledge.(2)Users are the primary participants of the dialog and effect the process of dialogue with no doubt.But existing dialog systems rarely take account of users' individualized information in the response selection procedure.Even if one system involves personalized characteristics,only about interest label,gender and age,and these features just cover a part of personalized information.To fully mine personalized characteristics,this disseration proposes a user modeling approach to learn individualized information from the matching relationship between users and their generated content.On this basis,this disseration presents a response ranking method incorporating with personalized characteristics,so as to select appropriate reply for the corresponding user.The experimental results have indicated that the proposed model is able to improve improve the response ranking performance.(3)To solve the "safe-response" problem,this disseration introduces a feedback mechanism to evaluate the quality of generated responses.In detail,this disseration presents a response generation model based on generative adversarial net,in which the discriminator is able to measure the differece between the ground response and generated one.As feedbacks of the response generation model,such differences would guide the model to produce more realistic replies.Meanwhile,to solve the discrete sampling of the response generation procedure,this disseration proposes an approximate embedding layer to connect the discriminator and generator,so as to propagate feedbacks from the discriminator to the generator directly.Moreover,this disseration also improves the feedback measurement to fit the response generation task.(4)The dialog context benefits response generation models by providing topic clues directly since these clues are the bases of the dialog.Therefore,making fully use of the dialog context is important for generating responses.However,the context is exploited inadequately in existing response generation models because of limited capability of context modeling and far away dialog histories generally aren't taken into account.After investigating the characteristics of the dialog context,this disseration proposes a dynamic working memory for modeling context using a special memory updating mechanism to capture and maintain topic clues in the context.On this basis,a context enhanced response generation model is presented to produce context-aware replies.