Research on abstractive automatic summarization technology based on deep learning

Automatic summarization technology is a method to obtain important information from documents, which can alleviate people's time and energy problems in the era of information explosion. This paper mainly studied the abstractive automatic summarization technology based on deep learning. Abstractive automatic summarization is consistent with the human habit of writing abstract, and has the characteristics of simplicity, flexibility and diversity. The experimental results based on English automatic summary data set (CNN/Daily Mail) and Chinese short text summary data set (LCSTS) showed that after the Attention Mechanism, Pointer Networks and Coverage Mechanism were added to the Seq2Seq model, the automatic summary Rouge evaluation index had an apparent improvement. In addition, comparative experiments were carried out from the neural network types (LSTM, GRU, SRU, etc.), the impact of Pointer Networks and Coverage Mechanism and the role of position features and Beam Search. After adding Batch Normalization and location features to the Point-Generator Network, there was a significant improvement in the Rouge1 and Rouge2 evaluation index.