Stemming is one of the most effective techniques, which has been adopted in many applications, such as machine learning, machine translation, document classification (DC), information retrieval, and natural language processing. The stemming technique is meant to be applied during the classification of documents to reduce the high dimensionality of the feature space, which, in turn, raises the functioning of the classification system, particularly with extreme modulated language, for instance, Arabic language. This paper aims to study the impact of stemming techniques, namely Information Science Research Institute (ISRI), Tashaphyne, and ARLStem on Arabic DC. The classification algorithms, namely Naïve Bayesian (NB), support vector machine (SVM), and $K$ -nearest neighbors (KNN), are used in this paper. In addition, the chi-square feature selection is used to select the most relevant features. Experiments are conducted on CNN Arabic corpus, which is collected from Arabic websites to assess the performance of the classification system. In order to evaluate the classifiers, the $K$ -fold cross-validation method and Micro-F1 are used. Findings of this paper indicate that the ARLStem outperforms the ISRI and Tashaphyne stemmers. The outcomes clearly showed the effectiveness of the SVM over the KNN and NB classifiers, which achieved 94.64% Micro-F1 value when using the ARLStem stemmer.