The invention relates to a Chinese zero anaphora resolution method based on LSTM and aims at solving the problem that according to an existing method, a Chinese zero anaphora resolution task is low in accuracy and the accuracy of understanding semantic information is low. The method comprises the steps of 1, processing each word in existing text data, and training each word in the processed text data by employing a word2vec tool, thereby obtaining a word vector dictionary; 2, selecting an antecedent candidate set of zero anaphora; 3, if candidate phrases in the current antecedent candidate set of the zero anaphora is true antecedents of the zero anaphora, determining that the training samples are positive example samples, otherwise determining that the training samples are negative example samples; and 4, connecting a Dropout layer with a logistic regression layer, representing probability value that model input samples are judged as the positive example samples, and outputting the value as a model. The method is applicable to the field of natural language processing.