Research And Application Of Deep Learning In Emotional Dialogue Generation

In recent years,in the background of increasing computer computing power,deep learning techniques have made continuous progress.Research in human-machine interaction and dialogue systems has also developed considerably.Emotions play a huge role in promoting understanding between people,and users prefer a conversational system to a dull,mechanical software.How to integrate emotion into human-machine interaction so that machines can understand human emotion becomes an important research direction.The task of generating emotional dialogue is divided into two aspects: emotion analysis and dialogue generation.First,the contextual emotion of the user is analyzed,and then the emotion category is used as the parameter to guide the dialogue generation.By rewriting the dialogue raw elements from the model,sir into semantically smooth response,then rewriting according to the emotional category,producing multiple rounds of dialogue with emotional tendencies,and designing an emotional analysis dialogue system to test the model.The research work of this paper mainly includes the following aspects.(1)Construct a model that integrates attention mechanisms to categorize emotions in the context of conversation.The model includes encoder and decoder,which integrates the semantic information of historical sentences in the dialogue into the decoder,and the emotional tag information of decoded sentences into the decoder.The attention mechanism is used to mine context-related historical information of dialogue and to incorporate semantic representation in sentences.At the same time,we combine the semantic information of historical sentences in the encoder and the emotional tag information of historical sentences in the decoder to help with the emotional classification of dialogue.(2)A two-stage conversational generation model based on emotional variables is proposed.Inspired by the secondary rewriting,we considered dividing the generation of emotional dialogue into two stages: generating responses in semantic messages and adapting responses to emotional variables,which not only ensures the fluency of the generated statement,but also enriches the emotion of the generated statement.In real life,the expression of emotion does not always point to a particular type of emotion,mixed emotions are common.Therefore,this paper uses Transformer encoder to realize emotion modeling in context,and divides emotion variables into primary emotion variables and mixed emotion variables.The main emotion variable is determined by the emotion classifier,and the main emotion variable is obtained according to the context.Mixing emotional variables,we get them by sampling and then combine the two types of emotional variables into global emotional variables.Finally,the global emotional variables are combined with the response semantic information generated in phase 1 to form a character level input to the rewrite model in phase 2,resulting in multiple emotional responses.(3)On the basis of the above two models,an open-area conversational system is designed and tested that integrates the mood of the dialogue.The models presented in this paper are not only rich in conversational expressions,improve the quality of sentence generation,etc.,but also provide a practical basis for further research.