Text classification-oriented feature selection method, training method and data enhancement method

The text classification-oriented feature selection method is realized through a method in the field of natural language processing. Text information and label information in the text classification data set are input, and text classification data are enhanced by designing an adversarial training model comprising a generator for generating word weights and a classifier for judging original sentence corpus classification categories; wherein the generator is used for generating word weights and word weight vectors, meanwhile, the word weight vectors are input into the classifier, the classifier firstly uses the word weight vectors to multiply word vectors of original input sentences, then classification is carried out through an RNN classifier, and the word vectors of the original input sentences are classified through the RNN classifier. And model parameters are updated by using a classification result of the classifier. According to the scheme, the sentence polarity label information in the text data is used,