The works applying meta-heuristic approaches for Arabic text is limited due to the complex of Arabic parsing and derivation rules, its rich morphology and complex grammatical rules. In this research we propose a new architecture for Arabic documents that applies the Grey Wolf Optimization used as a supervised learning approach by simulating the behavior of Gray Wolves in searching, encircling, and chasing their prey. The proposed method is divided into four phases: The term weighting phase, based on TF-IDF, and the term relevance frequency in documents and classes. The feature selection phase uses the Gray Wolf Algorithm to reduce features and identify relevant features for building the model. The next phase uses the Firey Algorithm to classify documents into several classes using some objective function. In the last phase, the researchers compare the performances of these approaches. The researchers experimented on two collections of reference documents and presented a comparative study on different systems of terms weighting. The experimental results allow us to deduce that our proposed approach based on Grey Wolf with Firefly Algorithm allows us to obtain better results than the other statistical algorithms.