Research On Fairy Tale Text Generation Based On Improved GPT-2 Model

With the implementation of the second child policy,the number of newborns in our country is increasing year by year,and children's education is also increasingly valued.And now the fairy tales in our market are too old-fashioned and lack of customization,and the quantity is not enough to meet the growing demand now.On the other hand,with the rapid development of the field of deep learning,there are huge breakthroughs in many areas included in it,such as computer vision,natural language processing,and real-time decision-making.However,with the continuous development of models,how to combine these domains are also very important issues for us to solve real problems and reflect value.Therefore,this article attempts to combine the latest deep learning models to generate fairy tales from cartoon pictures and provide more corpora for preschool children.This paper proposes a combination of image description model and language model,and uses beam search algorithm to improve it.Constructed an end-to-end system that can generate fairy tales from cartoon pictures.The system is mainly divided into three modules: image caption module,connection module and text generation module.In the image caption module,a new encoder-decoder architecture is used to generate a short text description of the image from a picture;in the connection module,the cluster search algorithm is mainly used to perform the output part of module one and the input end of module three is improved,Integrate more image information into the text;in the text generation module,the latest language model GPT-2 is improved for generating long texts(fairy tales).In terms of model evaluation,the image description module was first evaluated with the BLEU machine translation index.The architecture of this paper is superior to the traditional architecture under the same training time.Second,the long text is evaluated to indicate that the generated text is good.Finally,a new index for evaluating such problems is proposed,which integrates factors such as text relevance,long-distance dependence,and sentence fluency,and proves that it can effectively evaluate the problem of generating long text from images.The innovations of this paper are the construction practical deep learning system;the use of DenseNet convolutional neural networks and the improvement character-level recurrent neural networks.Beam search is used to achieve this target.The algorithm organically fuses the image caption model and the language model;and also constructs an index to effectively evaluate such problems.And finally,it also fills the gap in the evaluation of long text image description problems.