Research Of Generating Text Via Generative Adversarial Nets

The birth of Generative Adversarial Nets has led to many breakthroughs in the research of neural networks and machine learning.However,language model of discrete space output form,limits the ability of GANs' abilityin natural language processing.WGAN,as a variant of the generative confrontation network,succeeded in extending the GANs' application data space from continuous data space to discrete data space.This research provides a possibility for the application research of Generative Adversarial Nets in text sequence generation and provides a new idea for solving the problems of neural network optimization.Based on WGAN,this paper analyzes and summarizes the problems existing in the task of text generation by neural networks.Based on the nature and advantages of the generated antagonism network,this paper proposes a new unsupervised Label GANs method.This method uses character-level text construction methods and has the characteristics of being widely applicable to a variety of sequence processing tasks.At the same time,the deep convolutional neural network is used as a generator and discriminator in the process of generating confrontation.The formalized definition of deep neural networks for generating networks and discriminant networks,the structural design and optimization strategy of neural networks are given respectively.During the construction of the generator,the noise data space of the generated data is labeled,and the specific tag structure and generating strategy are given.The label noise is constrained in the discriminator process,which is reflected in the structure of the discriminator,the optimization function and the entire network structure,so called Label GANs.In the experiment,based on tensorflow platform and GPU hardware accelerated computing structure,English news and Chinese poetry were selected as training data respectively.These two kinds of experimental data have the characteristics of great difference in character space and language habit.In addition,in order to better measure the effect of generated data,this paper also presents a loss measurement method based on generating confrontation process,which provides quantitative reference for performance analysis of generated networks without increasing the amount of computation.Finally,using WGAN as a comparative experimental method,we conduct a comprehensive experiment and analysis of the proposed methods and assumptions,including the selection of experimental parameters and the discussion of experimental results.The experimental results show that using Label GANs method can effectively intervene and constrain the generated data space,and can generate the results that are closer to the human language habit when selecting the appropriate parameters and training data,and can be expressed on different data sets Out of this feature.Similarly,during the experiment,the numerical results given by the quantitative loss calculation method are also in line with human intuition.Based on the WGAN,Label GANs is an innovative application of generating confrontation networks to generate discrete text sequences.Combining with the spatial distribution of noise data and constraint methods,Label GANs provides a reference for textual research on Generative Adversarial Nets.