This paper deals with automatic classification of Arabic web documents. Such a classification is very useful for affording directory search functionality, which has been used by many web portals and search engines to cope with an ever-increasing number of documents on the web. In this paper, Naive Bayes (NB) which is a statistical machine learning algorithm, is used to classify non-vocalized Arabic web documents (after their words have been transformed to the corresponding canonical form, i.e., roots) to one of five pre-defined categories. Cross validation experiments are used to evaluate the NB categorizer. The data set used during these experiments consists of 300 web documents per category. The results of cross validation in the leave-one-out experiment show that, using 2,000 terms/roots, the categorization accuracy varies from one category to another with an average accuracy over all categories of 68.78 %. Furthermore, the best categorization performance by category during cross validation experiments goes up to 92.8%. Further tests carried out on a manually collected evaluation set which consists of 10 documents from each of the 5 categories, show that the overall classification accuracy achieved over all categories is 62%, and that the best result by category reaches 90%.

No abstract

No abstract