| As one of the core technologies in the field of artificial intelligence,human-machine dialogue has attracted the attention of academia and industry with its broad application prospects and attractive commercial value.For decades,from the early Eliza and Parry to intelligent personal assistants like Siri,to today's chatbots like Xiaolce,the human-machine dialogue system has constantly changed people's、lives.Chatbots aim to meet the needs of communication and emotional companionship for human through natural,fluid and di-verse dialogues in human-computer interaction.Therefore,a good chatbot should be equipped with both IQ and EQ.However,the existing work on the dialogue system mainly focuses on improving the content quality of responses,while paying less attention to emotion.Therefore,this thesis focuses on emotional conversation generation,which aims to generate responses that are not only emotional but also semantically natural,so as to achieve more natural human-computer interaction.In recent years,with the popularity of deep learning technology,researchers have gradually adopted seq2seq generation model to implement a dialogue system,rather than relying on template-and retrieval-based methods.Therefore,this thesis combines seq2seq model with the topic model and variational autoencoder respectively to improve the relevance and diversity of responses,and based on this,the emotional factors are combined to enable the model to generate emotional responses.The main work of this thesis is as follows:1)In order to generate responses that contain emotions and are related to the content of input,this thesis combines emotional factors and topic information with seq2seq model to construct a topic-enhanced emotional dialogue generation model(TE-ECG)based on the attention mechanism.First,the input is encoded through an encoder based on a bidirectional Long Short-Term memory network,and then a Twitter LDA model is used to obtain the keywords of input as an additional input of the model.The integration of the topic information makes input and response share the same topic,ensuring content relevance between response and input.Finally,the model captures emotion-related information in the input text and additional topic words simultaneously through a dynamic emotional attention mechanism.Experiments are conducted on the emotional dialogue corpus released by NLPCC-ICCPOL 2017.The experimental results based on manual evaluation and automatic e valuation show that TE-ECG performs the best compared with other methods.2)The TE-ECG model integrates emotion and improves the quality of sentences,but the reply content still lacks diversity.In order to alleviate this problem,this thesis proposes an emotional conversation generation model based on variational autoencoder(VAE-ECG),which can better model the underlying semantics of text by taking advantage of the characteristics of VAE.The model is divided into three modules:a variational encoder is used to encode the input and output;a variational inference is used to model the approximate posterior distribution of the hidden variable;a variational decoder uses the contextual vector,the hidden variable,and the emotional state to generate a reply.This thesis conducts experiments on the emotional dialogue corpus released by NLPCC-ICCPOL 2017.Experimental results based on automatic evaluation show that the performance of VAE-ECG is improved compared with other benchmark models on the diversity of Unigram and Bigram,which verifies the effectiveness of this model. |
