The major goal in transfer learning is that the knowledge learned in one environment will help new tasks in another or changing environment. In this paper, a novel transfer learning approach is presented and the transfer knowledge will be applied to text categorization. First, we will learn the transfer knowledge from different category data respectively, and then, different classifiers will be constructed, final, transfer knowledge will guide other categorization task. We compared with SVM, K-NN and Centroid methods. Experiments showed that transfer learning method was effective and got a better performance in text categorization, it can help new tasks in another new environment or changing environment.