Research On Abstractive Text Summarization Method Based On Hybrid Neural Network

With the advent of the era of big data,the amount of text data is growing rapidly,and the demand for retrieval,understanding,and analysis of text information is also growing.In the military field,how to extract the main content from the massive military intelligence information to provide reference for the commander is an urgent problem to be solved.To solve people's needs under the condition of massive information,the field of automatic text summarization is rapidly developed.Automatic text summarization is to automatically understand and analyze one or more documents through machine learning or deep learning,by extracting sentences from document or automatically generating sentences of the main information of the document,and finally get a short length,informative,and readable summary,to help people understand and use a large amount of information in a short period of time,greatly saving time has been used to improve the efficiency of information utilization and to achieve accurate use of information in the field of military intelligence.The purpose of this paper is to improve the quality of the summary generated by the model,to mine the long-term information and global information of the document,and to improve the encoder or decoder based on the sequence-to-sequence structure,to construct a text summary model of traditional neural network mixed with memory network or non-local network,so as to improve the model's ability to capture long-term information or global information in documents and generate higher-quality summaries through the deep learning method.This paper focuses on abstractive automatic text summarization methods,mainly to complete the works as follows:(1).We propose a Memory-Enhanced Abstractive Summarization model(MEAS model)This paper presents a Memory-Enhanced Abstractive Summarization model consisting of a memory enhancement module and a sequence-to-sequence module.Firstly,the memory enhancement module is constructed by the idea of memory network,then the improved encoder is used to encode the document,extract the text features and potential information of the document,store it through the memory enhancement module,and then use the decoder to refer to the feature representation and long-term information of the document to decode and generate high quality,informative sentences form a summary.Our model improves the encoder in the sequence-to-sequence module through the memory enhancement module,strengthens the model for capturing and storing long-term information,overcomes the long-term information loss problem of the RNN structure,enriches the amount of information in the generated summary to improve the accuracy of the generated summary.(2).We propose a Hybrid Non-Local Network Abstractive Summarization model(HNNAS model)In this paper,the non-local network is innovated from the field of computer vision to the field of text summarization.The Hybrid Non-Local Network Abstractive Summarization model is proposed to improve the traditional sequence-to-sequence model.First,we encode the document through the encoder in the sequence-to-sequence model,extract the potential features of the document,and then the decoder is improved by using the non-local network module,and the global information of the deep layer of the document is captured by the non-local network module,and the document features and global information referenced by the decoder are enriched.On this basis,the decoder decodes and generate more informative statements constitutes a summary.We use non-local networks to capture potential connections between sentences or words in distant distances in the document,thereby obtaining richer global information and providing richer information for the decoder to generate sentences.In order to verify the effect of the abstractive summarization models proposed in this paper,we conduct experimental testing and analysis of all models on the CNN news datasets.The experimental results showed that:(1)Compared with the basic model,the MEAS model increased by 2.9%,1.6%,and 2.3% on the ROUGE-1,ROUGE-2,and ROUGE-L scores,respectively.Explain that our model is better able to capture and store long-term information about documents,resulting in a smoother,more informative summary.At the same time,the experimental results show that our model is better at generating longer summaries.(2)Compared with the basic model,the HNNAS model increased the scores of ROUGE-1,ROUGE-2 and ROUGE-L by 1.8%,0.5% and 0.3%,respectively.Explain that our model can capture more global information through non-local network,more effectively capture the deep relationship between sentences or words that are far apart in the document,and generate a summary containing more information.At the same time,the experimental results show that the HNNAS model performs better when generating longer summaries.