The invention discloses a listed-company announcement classification and abstract generation method based on deep learning. The method comprises the following steps: step 1, acquiring announcement original-text data, extracting text, picture and form information, and establishing structured documents. step 2, establishing a classification rule word library of different announcements on the basis of industry knowledge of announcement fields according to various company operation change event keyword differences, and carrying out statistical judgment on announcement classes; and step 3, for the announcements of the different classes, extracting announcement document contents, combining the rule word library of corresponding class keywords to train an announcement content classification model, and automatically generating document abstract contents, wherein content extraction, training set selection, keyword model optimization, model training, model testing, result analysis and content generation are included. The method can solve technical problems of automatically classifying the announcements for a large amount of announcement information generated each day, automatically extracting key and important information according to classification situations, generating the abstract contents and the like.