The Research On Text Abstract Generation Based On Deep Learning

With the continuous development of Internet technology,the amount of text information has shown an explosive growth.Reading so much text information requires a lot of time and effort.How to quickly capture effective information from a large amount of information and apply it appropriately has become an urgent problem to be solved today.In this situation,automatic text summarization technology has emerged as the times require.This technology can help people quickly obtain the key content of a large amount of text information,reduce the burden of information processing and reading,and is an important technology in the field of natural language processing.However,there are still some problems,such as the inability to effectively extract key information,the high proportion of words outside the text set in the output summary,and problems such as information loss and inconsistent sentences when processing long text summaries.Therefore,in response to these issues,existing text summarization techniques are studied and improved.The main research contents are as follows:(1)An improved Transformer model is proposed.This model builds an Aggregation module between the encoder and decoder,which makes the generated summary more relevant to the original content.At the decoding end of the model,pointer network is introduced to complete the decoding of text summary.The function of pointer network allowing direct copying of unlisted words in the glossary from the source text is used to generate text summary,which improves the flexibility of summary generation,Use the News2016 zh dataset to test the performance of the model.The experimental results show that the improved Transformer model achieves better results than the benchmark model in terms of ROUGE evaluation indicators.(2)A end-to-end automatic text summarization model based on BERT model and sequence to sequence model is proposed.Firstly,the BERT model is used to extract text features as part of the encoder.Secondly,utilizing long and short memory networks to generate semantic features of text as another major part of the encoder.Combining the two parts to generate an intermediate vector can more accurately capture the key content and Semantic information of text information.Finally,the intermediate negotiation with attention mechanism is input into the decoder to generate a text summary summary.The experimental results on the CSL and LCSTS datasets show that the improved sequence to sequence model achieves significant improvements compared to the baseline model.The research in this thesis focuses on how to optimize the automatic text summarization model and improve the accuracy of text summarization by improving the traditional model structure.And a series of experiments were conducted to verify that the improved model in this article achieved better results on the dataset News2016 zh,CSL,and LCSTS.