Research On Dialogue Generation Based On Emotion And Content Features

The human-machine dialogue system has many advantages such as convenience,speed,and efficiency,and is regarded as one of the important forms of human-computer interaction in the future.Its development trend has also received widespread attention in the academic and industrial circles.As an important part of the human-machine dialogue system,dialogue generation is aimed at generating replies based on the user's dialogue messages.The quality of the dialogue directly affects the user's experience.With the development of deep learning technology,it provides new impetus for the development of dialogue generation research.However,in the research of dialogue generation in the open field,the basic dialogue generation model has the problems of tending to generate a kind of lack of content response such as "don't know”and the generated response is of low interest and sustainability。The expression of language needs the support of emotion and content,which can make language more vivid,natural,substantial and appealing.Considering the emotion and content in the context in the dialogue generation,not only can provide the dialogue with emotional embellishment and content filling,alleviate the problem that the system is prone to generate a lack of content reply,and at the same time enhance the diversity of emotions and content of the reply,can increase the dialogue interesting and sustainable.This thesis will study the generation of dialogues based on emotion and content features.That is,in the process of dialogue generation,according to the emotion and content features contained in user messages,psychological "emotional empathy"mechanism and the transformation of emotion and content features are adopted to guide the deep learning model to generate responses that meet the emotional needs of users and are related to content.The main research work of this article is as follows:(1)A dialogue generation model(CECF-DG)combining emotion and content features is proposed.In order to obtain the emotion and content feature contained in user messages,the model uses circular neural network and convolutional neural network to obtain the potential emotion and key content feature in user messages.In order to promote the fusion of emotion and content,the model constructs a decoder of emotion and content characteristics based on attention mechanism.In the decoder,an independent feature decoding unit is designed for each feature to learn the expression of different features.In the process of decoding,the attention mechanism is used to dynamically acquire the encoding of the user's dialog message according to the decoding state of the feature.Through the design of comparative experiments,the results show that the model is improved in each index of automatic evaluation and manual evaluation.(2)A dialogue generation model(ECFC-DG)based on the conversion of emotion and content characteristics is proposed.In real conversation messages,users not only need to respond with consistent emotions and contents,but also need to respond with more diversified emotions and contents to enhance the interest and sustainability of the conversation.Considering that there is a certain correlation between emotion and content between real messages and replies,this correlation has the effect of enhancing the diversity of emotion and content.Based on this,this model is based on the Seq2seq model,which introduces the feature transfer matrix and learning function to learn the transformation relationship between the emotion and content characteristics between the source statement and the target statement,and then integrates the transformed emotion and content characteristics into the model decoding process,and finally generates the response.The experimental results show that the responses generated by the model are improved in the evaluation indexes of emotion and content diversity.