Arabic is one of the Semitic languages in antiquity and one of the six official languages of the UN. Also, Arabic classification plays a significant and essential role in modern applications. There is a big difference between handling English text and Arabic text classification; preprocessing is also challenging for Arabic text. This paper presents the implementation of a Naïve Bayes classifier for Arabic text with and without stemmer. A set of four categories and 800 documents were used from the Text Retrieval Conference (TREC) 2001 dataset. The results showed that Naïve Bayes with light stemmer achieves better results than Naïve Bayes without stemmer. The findings of the classifier accuracy by employing stemmer and without stemmer are as preprocessing. It reveals that the accuracy resulted from the light stemmer was better than the classifier without stemmer detection, which Naïve Bayes Classification with light stemmer got 35.0745 higher than the Naïve Bayes Classification 33.831% without stemmer. After contrasting them, the stemmer got better accuracy than the classifier.