Research On Response Generation With Dialogue Constraints

Recently,human-machine dialogue systems have received more and more attention due to their great academic and commercial value.With the rapid development of social media and deep learning technology,chatbots play a wide range of roles in life,which can not only meet people’s communication needs,but also save the service cost of enterprises.In order to be more in line with practical applications,people often study multi-round dialogue generation systems,which can use more dialogue history information,and also poses new challenges on how to better capture features that fit the dialogue scene.In order to better model the context information,in the multi-round dialogue generation,it is necessary to model the dialogue constraints,capture the attribute information in the dialogue,improve the attribute control ability of the model,and increase the diversity of generated responses.The main work of this paper is as follows:Firstly,for the topic and personality constraints in dialogue,this paper proposes a research method for response generation based on topic and personality constraints.In everyday conversations,the interlocutors usually revolve around a specific topic and each sentence has a clear emotion and intent.Emotion and intent describe the personality of the speaker.The method combines the identification of the topic,emotion and intention of the dialogue,and realizes the constraints on the topic,emotion and intention of the generated response by sharing parameters,so as to generate the response with reasonable emotion and intention and related to the topic of the dialogue.Secondly,for the problem of dialogue context constraints,this paper proposes a research method for response generation based on context constraints.The importance of modeling contextual clause associations is often overlooked in existing research on multiround dialogue generation.Therefore,this method aims to model contextual associations to improve the quality of the generated responses.In this method,the main consideration is to model the associations between the same speaker clauses in the context and the associations between the same round of dialogue clauses.Modeling the associations between clauses of the same speaker can learn the roles and dialogue motivations assumed by the speakers in the dialogue;modeling the associations between clauses in the same round of dialogue can learn the logic interaction information between speakers.The method uses graph convolutional networks to model contextual associations,capturing role and logical information in the context.Finally,for the problem of combining context and personality constraints in dialogue,this paper proposes a research method for response generation based on pre-trained models and constraints.The correlation between emotion and intention in discourse has often been overlooked in previous studies.Therefore,this method aims to improve the quality of personality constraints by using the correlation between emotion and intention on the basis of context constraints to assist in response generation.In this method,the conditional probability during decoding in generative method is mainly considered to model the relationship between emotion and intention,so as to improve the accuracy of emotion and intention recognition.Then,the model parameters,that is,the learned personality information,are transferred to the dialogue generation to constrain the generated responses.