Nowadays, with the Internet infrastructure and nearly global access, the amount and diversity of data are increasing rapidly. Many tasks require information retrieval and data collection for machine learn- ing, research, and survey reports in various fields such as meteorology, science, geography, literature, and more. However, manual data collection and classification can be time-consuming and prone to errors. Addition- ally, AI assistants used for drafting or writing can sometimes be corrected regarding writing style and inappropriate language for the given con- text. Faced with these needs, In this article, Vietnamese documents are classified using the TF-IDF method, TF-IDF combined with SVD, and FastText at three levels: word level, n-gram level, and character level. For this approach, 15 categories were gathered from various online news sources. The dataset was preprocessed and trained using machine learn- ing models such as SVM, Naive Bayes, Neural Network, and Random Forest to find the most effective method. The Random Forest combined with the FastText method was highly evaluated, achieving a success rate of 82% when measured against essential evaluation criteria of accuracy, precision, and F1 score.