In order to solve the “labeling bottleneck” problem of short text categorization, a novel Semi-Supervised Expectation-Maximization short text categorization method based on Random Subspace (RS-EM) is used in this paper. RS-EM performs an iterative EM style training where multiple models are trained on subsets by using random subspace method. This combination of the stochastic discrimination theory and semi-supervised EM algorithm is used to compensate for the weaknesses of the standard EM algorithm to over-training. Experimental on real corpus show that the proposed method is more effectively exploit unlabeled data to enhance the learning performance, and is superior to standard semi-supervised EM algorithm in the learning efficiency and the classification generalization.