One of important steps in hybrid statistical-structural recognition method for handwritten characters is to label primitives for classifier training and label structural position information for structural recognition. In this paper, we propose a sem