Research On Deep Learning Based Emotional Dialogue Generation Methods

The research on emotion-based dialogue generation models is of great significance for enhancing the user's experience in man-machine interaction.At present,existing research mainly relies on the basic framework by injecting emotion vectors into the sequence-tosequence generative model.A variety of variant models using different emotion injection methods have derived from this framework,such as introducing emotion embedding in the input layer and output layer of the model or directly leveraging an external emotion dictionary to encourage more attention to emotion words in the decoding process.The focus of the above research is to give a certain type of emotion,and let the model generate a response with such emotion,which essentially solves the problem of emotion-controlled dialogue generation.However,the obvious shortcomings of the above methods are that these models can not automatically make emotional choices according to the other's emotion,and do not have the ability to perceive and predict emotions.Therefore,they need to rely on human judgment to specify emotions in practical applications,which increases the difficulty of automated online deployment.In view of the problems of these models,this thesis focuses on how to better tackle the single-turn open-domain emotional dialogue generation problem.Therefore,the main contributions of this thesis are as follows:1.This thesis proposes a self-attention-based emotional dialogue generation model(EACM),which can automatically perceive emotions,manage emotions,and express emotions.Specifically,the model first uses a self-attention mechanism,emotional and semantic word embedding,and fusion network to process user input sentences and extracts effective semantic and emotional information,then uses the prediction network to estimate the emotional distribution,and finally uses emotion-biased attention mechanism to generate responses with the predicted emotions.2.This thesis proposes another dialogue generation model using prior/posterior emotional distribution estimation(TG-EACM).This model introduces prior/posterior networks to acquire empirical information in the historical data,and uses sufficient emotional interacting information from the posterior distribution to guide the prior network.KL divergence loss is designed as an auxiliary objective function for multi-task learning.These methods improve the accuracy of emotion prediction,and achieve more accurate modeling of the emotion interaction in dialogues.3.Through baseline comparison experiments and module ablation experiments,this thesis proves that the proposed models can surpass existing models on automatic metrics such as BLEU-n and Distinct-n and human-evaluation metrics from both emotional and semantic aspects.At the same time,the mechanisms proposed in this paper are very effective in increasing the accuracy of emotion prediction,reducing the difference in the prior network distribution,and increasing the convergence speed of models.Through the case study of the emotional responses generated by the model,we can intuitively see that the proposed model can effectively perceive user's emotions and generate responses with appropriate emotions,surpassing the existing dialogue models in terms of user experience and comfort.