Research On The Generation Of The Open Domain Emotional Conversation Based On Deep Learning

Open domain emotional dialogue refers to endowing machine-generated response sentences with corresponding emotions in non-task-oriented small talk,thus making human-computer interaction more natural,cordial and vivid.With the rapid development of social networks,massive conversational corpus provides rich data resources for the dialogue system,while the development of computer hardware and deep learning provides strong technical support for the generation of dialogues.In recent years,more and more attention has been paid to dialogue system in academia and industry.In the development of human-machine dialogue,it is an important aspect to endow the dialogue system with emotional perception and interaction.First of all,large-scale conversational corpus with emotional labeling is required in terms of data.Secondly,how to make the quality of dialogue generation smooth and context-related,and how to make the generated dialogue contain the corresponding emotion.Finally,there is no unique answer statement for open domain dialogues,and how to evaluate machine-generated dialogues is another challenge.In this paper,the data set of NLPCC 2017 emotional dialogue generation task is adopted to solve the challenges in the generation of emotional dialogue by using the deep learning model,and a variety of evaluation indicators and methods are adopted and tried.Specifically,in the large-scale conversation corpus with emotion labels,the deep learning technology is used to train the sequence-to-sequence model to generate the dialogue statements.On this basis,the pre-trained emotion supervision model is used to assign corresponding emotions to the response statements generated from the sequence-to-sequence model.The methods of automatic evaluation and manual evaluation are adopted.The specific work of this paper is as follows:This paper proposes an open domain dialogue generation model with emotional supervision based on deep learning.In this paper,the word embedding representation for the dialogue post and response is first adopted by combining words and characters.The bidirectional LSTM is used to encode the word embedding of the input statement to get the hidden vector,and then LSTM is used to decode the hidden vector to predict the response.In addition,the pre-trained emotion classifier can classify the emotion of the dialogue text.Based on this,two kinds of affective dialogue tasks are studied in this paper,one is the generation of specified emotion response and the other is the generation of unspecified emotion response.In the dialogue generation of specified emotion,response can be endowed with five types of specific emotions,namely happiness,sadness,angry,disgust and like.With the help of emotion classifier,the emotion vector of generating response is obtained in the emotion space,and the cosine similarity between it and the emotion vector corresponding to the specified emotion is calculated,and it is made to approach the specified emotion vector continuously in the emotion vector space,so as to generate the response with the specified emotion.The unspecified emotion generation is to generate a response that is emotionally close or distant to the current post.With the help of emotion classifier,the vector of the current input statement and the generated response statement in the emotion space are obtained respectively,and the cosine similarity of them is calculated.If the similarity is minimized,the dialogues with similar emotions will be generated,otherwise,the dialogues with different emotional distances will be generated.