Research On Method Of Emotional Text Generation In Human-machine Dialogue

As an important issue in the field of Artificial Intelligence,Human-Computer Dialogue has attracted much research attention.Open domain dialogue generation is one of the important issues in Human-Computer Dialogue.Its purpose is to ensure the generated dialogue responses much natural,smooth and diverse.In recent years,the rapid advancement of deep learning technology has greatly promoted the development of dialogue-related research,which makes dialogue generation no longer relying solely on template matching and retrieval.Currently,most deep learning based dialogue generation methods adopt the Seq2 Seq model following Encoder-Decoder framework.These methods use a large amount of data to learn the feature representation and response generation strategy.However,these methods tend to generate simple even meaningless responses.Aiming at the shortcomings of the existing methods,this thesis investigates the dialogue generation method based on Variational Auto-Encoder(VAE)to improve the diversity of model responses.Based on this,the emotional factors are embedded to support emotional response generation for further improving the user experience in human-machine dialogue.The main work in this work including:Target to the low quality dialogue generation problem of existing Seq2 Seq model based on RNN(Recurrent Neural Network),we investigate the dialogue generation model based on Variational Auto-Encoder.It introduces the VAE process into the decoding stage of Seq2 Seq model,and uses latent variables to model the potential distribution of text semantics.Meanwhile,this model incorporates the attention mechanism to enable the model "notice" the different parts of context during each time step in the decoding stage.The experimental results on the NLPCC 2017 Shared Task 4 dataset show that our model increase the content generation diversity measures based on Unigram and Bigram for 0.7% and 5.0%,respectively,compared to the traditional Seq2 Seq models.Target to problem that the existing Seq2 Seq model cannot generate emotional responses,we further investigate the response generation model,which embeds emotion factor.Firstly,referring to the enlightenment of convolutional operation,we proposed a sentiment classification model based on Convolution-based Memory Network(Conv-Mem).Conv-Mem models the sentence sequence features of “multi-words” level.Then,we design different emotion embedding method corresponding to different emotion labels.The model introduces the emotion embedding information into latent variables in the VAE process,so that the latent variables contain the corresponding emotional information.The experimental results on NLPCC 2017 Shared Task 4 dataset show that the proposed model may generate emotional responses corresponding to the assigned emotion label.This model effectively improves the naturalness of the generated text and improves the user experience of the human-machine dialogue.Meanwhile,the experimental results on NLPCC 2013 and NLPCC 2014 sentiment classification datasets show that Conv-Mem model outperforms the state-of-the-art baselines.