Research On Emotional Dialogue Generation Model Based On Deep Learning

As one of the core technologies in the field of artificial intelligence,human-machine dialogue has always attracted the attention of academia and industry.In the early days,the construction of the human-machine dialogue systems were mainly based on rules,templates or retrieval methods.In recent years,the development of deep learning technology has made breakthroughs in existing human-machine dialogue technology.With the expansion of the application scenario of the human-machine dialogue system,people have put forward higher demands.They hope that the human-machine dialogue system has both IQ and EQ,and can communicate with people flexibly,whether it is task-oriented or open-domain.The machine is expected to understand and express emotions in the conversation,rather than giving a cold response.However,most of the existing dialogue generation researches are devoted to improving the relevance of the response content,and there are few studies on the emotional dialogue generation task.This thesis is based on deep learning to study the emotional dialogue generation,which according to the user messages in dialogue and the specified emotion category,that include {other,like,sad,disgusted,angry,happy} six emotions,to generate content-and emotional-related responses.The main work of this thesis is as follows.This thesis proposes an emotional dialogue generation model based on joint decoding of content and emotion(EDG-JDCE).Related researches indicate that adding emotional factors to the dialogue generation model leads to a decrease in the grammatical correctness and content relevance of the generated responses.In order to alleviate the above problem,the model designs a joint attention mechanism and joint decoder based on content and emotion.The independent content decoding units and emotion decoding units are designed in the decoder to learn the expression ability of the content and emotion in the dialogue respectively.In the decoding process,a joint attention mechanism based on content and emotion is introduced.The attention weight is obtained according to the expression state of content and emotion,and then the user message in dialogue is dynamically encoded.At the same time,this thesis explores the impact of two common emotion category representations of One-hot and Embedding on this model.This thesis conducts comparative experiments on the emotional dialogue corpus published by NLPCC2017.The experimental results show that this model has a certain degree of improvement in each metrics compared with other models.The EDG-JDEC model uses the maximum likelihood function as the objective function,which leads to the same problem that is easy to generate"safe responses"under this method,such as"I don't know"'In order to improve the diversity of generated responses,this thesis proposes an emotional dialogue generation model based on sequence generative adversarial network(EDG-SeqGAN),which mainly includes generator,content discriminator and emotion discriminator.The generator uses the emotional dialogue generation model based on content and emotionjoint decoding(EDG-JDEC),and both the content discriminator and emotion discriminator are based on the Bi-directional Long Short-Term Memory Network(Bi-LSTM).In the adversarial training process,the content discriminator is used to ensure that the generator generates responses related to the content of the user messages in dialogue,and the emotion discriminator is used to guide the generator to generate responses containing the specified emotion category.This process ultimately allows the generator to generate responses that are closer to the real dialogue,and increases the diversity of the generated responses.The experimental results show that the responses generated by this model are not only related to content and emotion,but also richer and more diverse.