Abstract: In order to simulate the Chinese character acquisition process, this paper set up a multilayer selforganizing maps (SOM) network model based on improved Kohonen network. The model's output maps, which adapt modified algorithm and expand neuron's neighborhood, were connected via associative links updated by Hebbian learning. After training the model could learn Chinese character architecture successfully and also do well in Chinese character component recognition. The simulation results demonstrated that the feasibility of further research in Chinese character acquisition and even Chinese language learning with this model was possible.