The invention relates to a subtopic discovery method for a semi-structure short text set based on a mutual constraint topic model. The main technical characteristics are as follows: applying data cleaning to a short text set containing topic tags; extracting a short text containing specified seed topic tags for a certain topic based on seed topic tags; generating an input file for the cleaned data; inputting the input file to a mutual constraint topic model for model training; obtaining a semantic vector representation of the topic tags in the set, an average semantic vector representation of the text where the topic tags are and a lexical vector representation of the text where the topic tags are; orderly connecting the three vector representations as a complete semantic representation for one topic tag; and clustering by use of a Kmeans clustering method and outputting the centroid of a class obtained by clustering as a subtopic. According to the invention, the design is rational; mutual constraint potential topic modelling is adopted to solve the problems of high sparsity and great noise existing in the traditional semi-structure short text topic semantic modelling technology.