The invention provides an abstract generation method and device, and the method comprises the steps of scoring each sentence in a received text content through employing a pre-trained deep neural network model, and obtaining a first score of each sentence; scoring each sentence of the text content by adopting a text sorting algorithm to obtain a second score of each sentence; processing the firstscore and the second score of each sentence by adopting a pre-trained first logistic regression model to obtain a third score of each sentence; selecting at least one target sentence from all sentences of the text content according to at least one submodule function, a greedy algorithm and a third score of each sentence; and generating an abstract of the text content according to at least one target sentence. Therefore, the importance and the richness (redundancy removal) of abstract contents are maximized, and the problem that the abstract contents are not smooth is solved to a larger extent.