Extractive-Abstractive Text Automatic Summary Based on the BERT Graph Model

The extracted text summary method is not completely divorced from the document itself, but it has poor accuracy and readability. The generative method is more flexible and can avoid redundancy, but there are lack of coherence and logic. For these problems, puts forward a kind of extraction-generative text summary automatic generation method, through the BERT model for word vector, cosine similarity method to calculate the similarity between sentences, with TextRank algorithm iteration calculated each sentence score, extract top key sentence, input to the pointer generation network model to get the final summary. The model was tested on the TTNews dataset, and the ROUGE index value is significantly improved.