The invention provides a semantic-based text retrieval method, and solves the problems that polysemy or synonymy cannot be distinguished and retrieval results are mistakenly selected or missed in the retrieval process. According to the method, concepts of words replace words for searching and retrieval, and the retrieved files are sorted; the method specifically comprises steps as follows: S1, a concept tree is established according to the concepts of words, and a word similarity matrix is calculated; S2, the concept of a target file is extracted in reference of a preset body, the target file is subjected to indexing processing according to the concept, and an index file is generated; S3, word segmentation is performed on initial query of a user, similar items whose similarity to the query words is larger than the threshold value M are found out from the word similarity matrix, and the similar items are added to the user query in an OR manner; S4, a search engine searches the target file according to the user query; S5, according to the similarity of words, the similarity of the files is evaluated, and the files are sorted; and S6, file data are read, and a sorting result is output.