| In the past few years,deep neural network has been developed rapidly and the model based on recurrent neural network has achieved remarkable results in the field of machine translation.Because the generation of dialogue can be regarded as a kind of translation from input to response,it is also a very promising way to apply it to dialogue system.A lot of open source dialogue data sets have emerged on the network due to the development of the Internet,which makes it possible to train generative dialogue system in a datadriven and end-to-end way.This thesis mainly studies the open domain generative dialogue system.Different from the domain specific dialogue system,its main purpose is to chat with users and relax.The special feature of generative dialogue system is that it can generate sentences that do not appear in the idiom database,so it can generate more interesting content than the dialogue system based on retrieval and template.However,there are some problems in the traditional sequence-to-sequence model:(1)because greedy decoding is used to select the word with the highest probability at each time,the model tends to generate very conservative and general response,which greatly reduces the diversity of response;(2)because the current pre-trained word embedding and training dataset do not contain more emotional information,Therefore,the generated sentences often lack affection.Recently,Conditional Variational Auto Encoder model has made a breakthrough in the diversity of generated responses,so this thesis bases on the Conditional Variational Auto Encoder to handle problems aboved.Our contributions are:(1)This thesis proposes an open domain dialogue generation model Affect-CVAE,which combines emotion dictionary with Conditional Variational Auto Encoder.In this model,the emotion dictionary labeled by human is used as the affect embedding of words,so the prior knowledge of emotion is introduced into the training data.On the other hand,the model reconstructs the word emotion in the response to make the hidden variables capture the emotion distribution of the reply,and generate sentences with more natural emotion and more diverse semantics.In addition,the model establishes reranking rules from the syntax,semantics and affection of the generated responses,which keeps the diversity of the generated response while keeping the content and emotion appropriate.(2)This thesis proposes an open domain multi-round dialogue generation model RLVHRED,which combines word level emotion and sentence level emotion with VHRED model and uses reinforcement learning technology.In order to guide the generation of response,reward is designed for the emotional diversity,semantic diversity and contextual coherence of reply.In addition,the effectiveness of the reranking mechanism in the Affect-CVAE model is verified in multi-round dialogue generation.Generally speaking,this thesis mainly studies the methods to enhance the affection and diversity of response in single-round and multi-round open domain generative dialogue systems.The experimental results on real datasets show that the Affect-CVAE model and RL-VHRED model are effective in improving the affection and diversity of generated responses. |
