Research On Chinese Text Generation Based On Generative Adversarial Networks

Deep learning technology has developed rapidly in recent decades,and this technology has been applied in more and more fields.There are also many research directions in natural language processing that can use deep learning,and text generation is one of the important directions.Text generation is a basic research that can be applied to many real scenes,such as text abstract extraction,text style conversion,and automatic text error correction.Generative adversarial network is a framework model that has received much attention in deep learning.Existing text sequence models all have the characteristics of data discrete.When GAN is directly applied to text generation,it will face the problem that the parameter update cannot be completed through backpropagation during the training process.In addition,the training method of GAN is to map a noise distribution to the prior real text distribution.However,the current text generation task is generally character-level generation,and this method is prone to excessively high repetitiveness of the generated text and even mode collapse.Finally,due to the limitation of the generator in the original GAN,the ability to extract text features during the training process is limited,and the quality of the generated long text is generally low.In response to the above problems,this paper proposes a text generation model based on GAN.The main work content is as follows:(1)Based on the idea of seq GAN and self-attention,this paper proposes an unsupervised text generation algorithm.The generative network incorporates selfattention,and adds local modeling using Gaussian deviation to the original Transformer model.This method improves the problem that the original model cannot be parallelized,and at the same time improves the ability to capture long-distance text features.At the same time,the objective function uses the mechanism of minimizing the penalty and introduces a new way to measure the Wasserstein distance.This ensures that the gradient will not disappear,and the mode collapse problem is also improved.In this paper,the new model is tested on the product description data set to judge the effect of text generation in long text.The experimental results prove that compared with the comparative model,the generated text quality of the generated model proposed in this paper is higher,and the text diversity is also richer.(2)The text generation model proposed in this paper is applied to the field of poetry generation to prove the universality of the model.Aiming at the unique fixed word count and rhythm rules of poetry,based on the above work,this paper introduces the pinyin Chinese character comparison table,and adds the rhyming judgment part of the tail word to the algorithm.This method makes it possible not only to ensure that the number of words in the generated poetry conforms to the rules,but also that the metrical tones can also conform to the rules.This paper conducts comparative experiments on the Tang Dynasty quatrain poetry data set.Compared with the baseline model,the algorithm model proposed in this paper has achieved higher scores in the comprehensive evaluation,especially the rhyme part.The experimental data fully proves the value of the model proposed in this study in the field of text generation,and also provides a reference for the research direction of future work.