In this paper, we propose a Bayesian network framework for explicitly modeling strokes and their relationships of characters. A character is modeled as a composition of stroke models, and a stroke as a composition of point models. A point is modeled with 2-D Gaussian distribution for its X-Y position. Relationships between points and strokes are modeled as their positional dependencies. All the models and relationships are represented probabilistically in Bayesian networks. The recognition experiment with on-line handwritten digits showed promising results; the recognition errors of the proposed system were greatly reduced by dependency modeling, and its recognition rates were higher than those of previous methods. (C) 2003 Pattern Recognition Society. Published by Elsevier Ltd. All rights reserved.