Arabic text classification methods have emerged as a natural result of the existence of a massive amount of varied textual information (written in Arabic language) on the web. Document classification is a process that requires an extensive background work on the content of the selected document(s) where most documents do not have identifiers such as keywords or metadata. Furthermore, matching document text based on the orthographic form may not be sufficient for classification purposes. In Arabic documents specifically, document classification still poses a great challenge due the nature of Arabic language. Extracting roots or stems from the breakdown of multiple Arabic words and phrases are an important task that should be completed before applying text classification. Moreover, feature extraction in text classification is a crucial task since it highly affects accuracy of classification. Generally, two types of features could be extracted: Statistical-based features and semantic- or concept-based features. The main interest of this thesis is divided between finding the best statistical features to be utilized and specifying the most effective semantic-based features on Arabic text classification process. In statistical-based features, the thesis contributed in hybridizing several techniques including Bag of Words (tokenization), root extraction and lexical stemmers (finding original form of word) and part of speech (syntactical analysis). The best techniques in Arabic documents’ statistical feature extraction, according to the literature, were used and compared against each other. Techniques include Khoja Root Stemmer, Light Stemmer, Root Extractor and a novel technique for returning words to its roots using Arabic WordNet is devised. The experimental results are conducted on six publicly available datasets. Namely, BBC, CNN, CCA, Watan, Jazeera and Akhbar Al-Khaleej. Results indicated minor improvements and revealed that the dataset characteristics govern the definition of the best technique. On the other hand, The text classification process is conducted using two semantic-based features based on semantic relations of Arabic WordNet (AWN) ontology. The first feature is a List of Pertinent Synsets (LoPS), a list of concepts that have a direct predefined relation with the original terms in the tokenized document, and the second feature set is a List of Pertinent Words (LoPW), a list of concepts that have an extended predefined relation with the original terms in the tokenized document. Seventeen different relations (defined in AWN ontology) such as “hyponyms” and “related-to” were used for feature extraction with both proposed techniques. The Naïve Bayes classifier is used to perform the classification process. For testing the semantic-based features two datasets were utilized namely, BBC and CNN. Results ‎revealed that using LoPS feature set improves the accuracy of Arabic text ‎classification compared to the well-known Bag-of-Word feature and the ‎recent Bag-of-Concept feature by 13.7% and 6.6% for BBC dataset, respectively, and by 6.11% and 6.09% for CNN dataset respectively. Moreover, it was found that LoPW (especially with “related-to” relation) improves the classification accuracy compared to LoPS, Bag-of-Word, and Bag-of-Concept by 1.27%, 7.95% and 15.15% respectively. The experimental results are evaluated using precision, recall, and F1-Measure. Data were tested using percentage-split and cross validation approaches to ensure that the best method is implemented.