Abstract: A method that integrates language information and statistical information from the training corpus is put forward. The weight of these characters is computed from three parameters: word frequency, centralized degree, decentralized degree. After training, we get the vector space model of the text categorization. The classification of the input text is decided by K-nearest-neighbor.The result shows that the method improves the accuracy of the categorization.

Key words: natural language understanding, vector space model, K-nearest-neighbor, automatic text categorization