In this paper, we address the problems of Arabic Text Classification and root extraction using transducers and rational kernels. We introduce a new root extraction approach on the basis of the use of Arabic patterns (Pattern Based Stemmer). Transducers are used to model these patterns and root extraction is done without relying on any dictionary. Using transducers for extracting roots, documents are transformed into finite state transducers. This document representation allows us to use and explore rational kernels as a framework for Arabic Text Classification. Root extraction experiments are conducted on three word collections and yield 75.6% of accuracy. Classification experiments are done on the Saudi Press Agency dataset and N-gram kernels are tested with different values of N. Accuracy and F1 report 90.79% and 62.93% respectively. These results show that our approach, when compared with other approaches, is promising specially in terms of accuracy and F1.