Research On Abstractive Text Summarization Based On Deep Learning

With the development of Internet technology,textual information has grown exponentially,and people spend a lot of time and effort on textual information.Automatic summarization technology is a method that can get important information from documents.It can help people reduce the waste of time and energy brought to people in an era of information explosion.In the term of sentences of summarization,Automatic summarization technology is broadly divided into extractive summarization and abstractive summarization.Extractive methods select sentences in the original text by evaluating their importance.The abstractive methods use natural language processing techniques to generate more concise sentences.Compared with extractive methods,summaries created by abstractive methods are closer to what a human might express.ive summarization has an advantages of simplicity,flexibility,and diversity.In recent years,deep learning has promoted the rapid development of abstractive summarization technique.Recently,abstractive summarization mainly uses the Seq2 Seq models.In the Seq2 Seq model,document can be represented by a vector and then model decode from the vector into many output.With the Seq2 Seq model,by encoding many inputs into one vector,and decoding from one vector into summary.In the paper,we research the technology of abstractive summarization,the main content consists of the following three aspects:(1)Single-document summarization based on deep learning.In this paper,we use Attention Based Summarization(ABS)model and compare the effects of different neural units and forward and reverse input on the model.Moreover,we use batch normalization to speed up the training of the model.With the Pointer Networks,coverage mechanism and the position encoding,which is combined with word embedding,the performance of the abstractive summarization model improves.Finally we use beam search for decoding.(2)Multi-document summarization based on deep learning.In order to solve the problem of lack of multi-document summary dataset,we analyzes the relationship between single-document and multi-document abstracts,and uses the transfer learning technology for multi-document summarization.From keywords level and key sentences level,we use single document as the source domain and mapping single document and multiple documents into the same feature,which alleviates the problem of lack of multi-document summary dataset.(3)Automatic summary system design and implementation.In this paper,we introduce the framework and design of the website system,which is combined with the automatic summarization models.We show the document summary with a website,and prove the practicality and validity of the automatic summarization.