Research And Implementation Of Automatic Text Summarization Technology Based On Deep Learnin

With the rapid development of the Internet,text data surge,how to accurately and quickly extract useful information from massive text has become an urgent problem in the era of big data.Automatic text summarization is one of the most popular research fields in natural language processing,which uses computer technology to compress text information and produce concise content.This process can help Users better browse and absorb massive information on the Internet,reduce the information load of users and improve the efficiency of user knowledge extraction.At first,Most text summarization methods are based on the extraction,but with the deepening of the study,the researchers found that the generated summary in text coherence are often not very desirable,poor readability,thus gradually the focus of research shifted from extraction summarization to abstractive summarization.With the rapid development of Deep Learning,abstractive summarization is becoming more and more mature,especially sequence to sequence model based on encoder and decoder framework is widely used in summary generation.This thesis mainly studies abstractive summarization based on Deep Learning,and the main work includes the following parts:(1)Proposed a abstractive summarization model RCTP based on improved Transformer and pointer generation network.It aims to solve the problem that standard Transformer model’s position encoding is not rich enough and Out Of Vocabulary words and repeated words are generated in summary generation.By introducing RC-Encoder based on bidirectional GRU,the model’s ability to capture sequence information and local information is improved.On this basis,a hybrid model is built by introducing pointer generation network,which can effectively alleviate the problem of Out Of Vocabulary words by deciding whether to copy a word from the original text or to generate a new word from the vocabulary during sequential decoding.In order to verify the effectiveness of RCTP model and various improvement strategies,experiments were carried out on CNN/Daily Mail data sets and LCSTS data sets respectively.The experimental results show that RCTP model can improve three Rouge indicators and obtain better summary results compared with the benchmark model.(2)Proposed a segmented text summarization method based on BERT pre-training model and key information fusion.Its purpose is to solve two problems.One is that the model is not enough to mine the semantic features of long-distance text,and the other is that the generated summary often misses the key information of text.By combining the advantages of extractive summarization and abstractive summarization,we extract key sentences from long texts and compress them into short texts,and then generate summaries on the basis of short texts.In addition,in order to enhance the key information extraction capabilities of the model,this thesis extracts keywords from key sentences,integrates keyword information into the model as external information,and enforces higher weight of keywords to guide the generation of summary by introducing keyword attention mechanism.A comparative experiment on CNN/Daily Mail and NLPCC 2018 dataset(text length>300)shows that the segmented abstractive summarization model achieves good results,and verifies the effectiveness of key sentences and keyword extraction tasks for summary generation.(3)In order to solve the problem that the traditional unsupervised extraction algorithm lacks the ability to extract deep semantic features,this thesis transfers the BERT pre-training model to the key sentence extraction task and proposes the BTM unsupervised key sentence extraction algorithm.For the problem of low accuracy of single extraction algorithm,a keyword extraction algorithm based on multi-feature fusion is proposed to improve the information extraction capability.