Emotional Dialogue Generation Based On Conditional Variational Autoencoders

Human-machine dialogue is one of the most challenging tasks in the field of natural language processing,and it is also the foundation for the realization of a human-machine inclusive society in the future.In recent years,thanks to the development of deep learning technology and the massive amount of data accumulated in the era of big data,the neural network-based dialogue generation method has attracted more and more attention from academia and industry.At present,most of the generative dialogue models based on deep learning use maximum likelihood estimation as the training goal.This method is easy to generate generic responses with single content and no meaningful information.The perception and expression of emotion play an important role in human-to-human communication.However,the existing researches mainly focus on the semantic understanding of discourse,with relatively few considerations of emotional factors.Aiming at the shortcomings of the existing methods,this paper studies the emotional dialogue generation algorithm based on the conditional variational autoencoders technology,which aims to improve the diversity of reply information and embed emotional factors in the generated reply,as well as to improve the user experience in the human-machine dialogue.The specific research work of this article is as follows:1.Emotional dialogue generation model based on conditional variational autoencoders:We fuse the Seq2 Seq model with the conditional variational autoencoders,and use the latent variables of the conditional variational autoencoders to learn the latent semantics in the text sequence to improve the diversity of the reply information.In addition,in order to improve the model's ability to perceive and express emotion,a dual emotion framework is designed to capture the emotional response in the conversation and control the apparent emotion of the response.In order to identify the emotion information implicit in the utterance sequence,an emotion classification model based on Bi LSTM-AT is proposed.Experiment shows that the model proposed in this paper can not only generate replies with diverse content information,but also generate controllable emotional responses.2.Transformer-based emotional dialogue generation model: The Transformer model is used to extract the semantic features of the text sequence to improve the utilization of the semantic information of the text sequence.In order to increase the diversity of the response information,the Transformer and the conditional variational autoencoders are fused,and latent variables are introduced into the decoder,so that the model can use the latent semantic information in the decoding process.In addition,in order to enhance the empathy ability of the model,an emotion perception encoder is used to encode user emotion information,and a pre-trained emotion classification model based on BERT is proposed to detect the emotion information implicit in the utterance.Experiment shows that the model proposed in this paper has stronger generation ability,more diverse information generated in response,and stronger empathy ability.