Text abstract automatic extraction method based on double-flow attention and position residual connection

The invention relates to a text abstract automatic extraction method based on double-flow attention and position residual connection. The text abstract automatic extraction method comprises the following steps: selecting a public data set and labeling data in the data set; constructing an RBPSum model, wherein the RBPSum model comprises a sentence encoder, a context encoder and an output layer; performing statement feature extraction on all statements in the training set by adopting a sentence encoder to obtain sentence feature representation of each statement; taking the sentence feature representation as the input of a context encoder, sequentially passing through L sentence enhancement layers, and outputting to obtain corresponding context position encoding information; and finally, obtaining a probability value that each statement is selected as an abstract through an output layer, and selecting statements corresponding to the first K probability values as a predicted abstract of the text. The performance