Research On Sentiment Dialogue Based On Generative Adversarial Networks

Human-machine dialogue generation is a major research direction in the field of natural language processing.It is a great challenge to generate high-quality,diverse,fluent,and emotional dialogue.With the rapid development of artificial intelligence and deep learning technology,the end-to-end neural network model provides an extensible dialogue generation framework,which makes it possible for machines to understand semantically and generate responses automatically.Neural network model also brings some new problems and challenges,the basic dialogue model framework tends to produce universal,meaningless and relatively “safe” responses.Based on the generative adversarial networks,we have some explores and researches on emotional dialogue generation task in this paper.The main work of this paper is as follows:(1)It is difficult for the basic dialogue generation model to get the emotional features of sentences from the dialogue text.In order to build a dialogue generation model which can generate the dialogue text with specified sentiment,the task of emotional dialogue generation is split.Several different models are trained to generate dialogue texts with different sentiments.Each model focuses on generating one kind of emotional dialogue text,in this way excludes the interference and influence of other sentiment in the process of generating the specified sentiment dialogue text so that improving the accuracy of generating dialogue texts with specified sentiment.(2)We proposed a new emotional dialogue generation framework SMC-GAN based on generative adversarial network to complete the emotional dialogue generation task.The proposed model includes a generative model and multiple discriminative models.The generative model was constructed based on the basic Seq2Seq(sequence to sequence)dialogue generation model,and the discriminative model includes the basic discriminative model,the sentiment discriminative model and the fluency discriminative model.The basic discriminative model can distinguish the fake text of generative model generated and the real text from the dataset.The sentiment discriminative model can distinguish whether the sentiment of generated sentence and specified sentiment category is the same,and guide the generative model to generate a dialogue text with specified sentiment.The fluency discriminative model can give the score for the input text sequence,and guide the generative model to generate smoother and more fluent sentences.The experimental results show that the SMC-GAN model can generate coherent,smooth and fluent dialogue texts with specified sentiment.Compared with the baseline mode,our model shows better performance on sentiment accuracy,coherence,and fluency.