Research On Automatic Text Generation Method Based On Generative Adversarial Nets

In recent years,artificial intelligence technology has developed rapidly,and the application fields of artificial intelligence technology are also expanding,such as smart city,automatic driving,content distribution and face recognition.In the field of natural language processing,there are also many scenes where artificial intelligence technology can be applied.One of them is the automatic generation of texts.As a basic work,the text generation model studied in this paper can be used as a module for more tasks such as the input method,the feature extraction of text,the automatic error correction of texts and the like.At present,the academic research on text generation tasks mainly focuses on the deep learning model,but most of the actual products still use template-based text generation methods.The main reason for the difference between the research and application is that the text generation model based on deep learning is not perfect enough,and the simulation of the target text is not meticulous enough.However,the construction of text generation template requires a lot of manual assistance and prior knowledge,and only for specific scenarios,and the scalability is not strong.Therefore,to study the text generation technology based on deep learning has more realistic significance.In order to break through the technical bottleneck of text generation,a more efficient text generation model is proposed in this paper,and its effect is tested on the text data set.Specifically,the research work of this paper can be divided into three parts:Extensive research has been carried out on the two areas of text generation and image generation.Analyze and compare the advantages and disadvantages of a single network and generative adversarial network.The basic text generation model is constructed by combining the generator network and reinforcement learning technology in the adversarial network,and the problem that the original generation of the adversarial network model can not deal with the discrete data is solved.According to the related work in machine translation,we design the "similarity score" to measure the quality of text generation.By using this index we compare the basic text generation model and the traditional recurrent neural network based text generation model on the ultrasonic inspection report data set and the novel data set.Deeply analyse of the advantages and disadvantages of the basic text generation model.According to the high variance of the actual reward function leading by Monte Carlo sampling search method and the slow convergence speed of the model leading by a lot of sampling operation during completing the text,we use convolutional neural network to improve the discriminator network.At the same time,due to the lack of feedback signal,the direction ofreinforcement learning training is not clear enough,and the higher order feature of the discriminator network is returned to the generator network.The improved text generation model is compared with the basic text generation model,the recurrent neural network based text generation model and the Rank GAN text generation model which is in the same field on the ultrasound report data set and the novel data set.Both in the two data set,the improved model achieves better results in the similarity score.