Research On The Model Of Text Summarization Based On Deep Learning

The abstract of an academic paper is a high-level summary and refinement of an academic paper,and it is also an indispensable part of the formal publication of the paper.Through the abstract,on the one hand,readers can have a preliminary understanding of the author's research content,research methods,and research results,so that other researchers can decide whether to read the entire paper,and improve the efficiency of paper screening.On the other hand,it has a certain ability to attract researchers in related fields,attracting interested readers to read the full text and obtain the latest scientific research content.Based on the above background,the main work of this paper is summarized as follows:First,the abstract generation of academic papers is a subtask of text generation.Among them,the adversarial neural network(GAN)is often used in the field of text generation.However,ordinary adversarial neural networks are less effective in text generation tasks,and they may not be able to be trained.Based on these situations,this paper proposes the EAD?GAN model.The EAD?GAN model adds a Monte Carlo Tree Search(MCTS)method to the generator to solve the gradient propagation problem.Before the formal training,the auto-encoding model is used to extract the feature distribution of the real text to improve the text Expressive ability.The discriminator uses a common CNN model,and adds cross-entropy and a reward mechanism in reinforcement learning to improve the innovation of the text.Experiments show that the EAD?GAN model improves the quality and effect of generating abstracts.Secondly,the abstract of the paper often contains a lot of information such as proper nouns,data explanations and experimental results,and the full text of the abstract is elaborated around the topic of the paper,and the overall structure is strong.However,the common neural network model cannot solve the above problems well.Based on the above situation,this paper proposes a GAT-Bi LSTM model based on graph attention model and two-way LSTM,which combines the abstract and topic content for training.First,use the SCIIE model to convert text data into highly aggregated knowledge graph data.The graph data contains node information and relationship information,which can solve the problem of reference and logical relationship in academic abstracts.Second,use the graph attention model to extract the features of the summary map.Third,use the two-way LSTM to extract the features of the title.Fourth,the two features are fused during training.Experiments show that the abstract generated by the GAT-Bi LSTM model conforms to the writing habits of academic papers,and the generated abstract is more innovative and closer to the abstract content written by humans.Finally,the emergence of new vocabulary is developing with the development of society.Using computer to generate abstracts provides writing ideas and inspiration for academic researchers to write abstracts,and helps to reveal the essence of language and literature research.At the same time,for computer-generated abstracts,automatic evaluation and manual evaluation are used to comprehensively evaluate the quality of generated abstracts from both subjective and objective aspects,which can accurately measure the effect of abstracts and provide certain ideas and methods for related research work.