A method for generating title by loop neural network based on multi-layer attention mechanism, wherein the loop neural network comprises a filter layer, an encoder, an attention mechanism and a decoder. The method comprises the following steps: S100: generating a plurality of coarse digests after inputting source text into a plurality of digest generation methods of the filter layer; S200: inputting the plurality of coarseness digests into the encoder for encoding respectively, and outputting respective hidden layer representation sequences; S300: merging the hidden layer representation sequence to generate summary semantics representing important information of the entire document, and paying attention to the generated rough summary and the words in the encoder by using the attention context; S400: the attention context and the summary semantics are input to the decoder to generate a corresponding title. Compared with the existing methods, this method improves the accuracy and readability of the generated title.