Several probabilistic methods used for Part of speech (POS) tagging are based on Hidden Markov Models (HMM), these methods have difficulties especially in estimating transition probabilities accurately from limited amounts of training data. Consequently, a new method appeared to avoid problems that HMM face. However, the transition probabilities are estimated using a decision tree. Based on this method a language independent POS tagger (called TreeTagger) has been implemented. The main purpose of this work is to create the language model to adapt TreeTagger for Arabic POS tagging and lemmatization. Furthermore, different configurations have been done, namely, collecting lexical resources, as well as the annotated training corpora. In addition, we used the proposed universal tagset that consists of common POS categories of 22 different languages including Arabic. We highlight the use of this tagger via various experiments on vowelled and unvowelled text from both Modern Standard Arabic and Classical Arabic. In fact, the obtained accuracies rates are 99.4%, 92.6% and 81.9% for respectively the Quranic vowelled corpus "Al-Mus'haf", the unvowelled "Al-Mus'haf1" corpus and for the NEMLAR corpus.