Text categorization is an important task in text mining process that consists in assigning a set of texts to a set of predefined categories based on learning algorithms. There exist two kinds of text categorization: monolingual and multilingual text categorization. The main problematic of this manuscript is how to exploit concepts and algorithms of machine learning in contextual categorization of multilingual texts. Our study on this subject allowed us to propose many solutions and provide many contributions, notably: (1) a simple, fast and effective algorithm to identify the language of a text in multilingual corpus. (2) An improved algorithm for Arabic stemming based on a statistical approach. Its main objective is to reduce the size of term vocabulary and thus increase the quality of the obtained categorization in TC and the effectiveness of search in IR. (3) A new multilingual stemmer which is general and completely independent of any language. (4) Application of new panoply of pseudo-distances to categorize texts of a big corpus such as Reuters21578 collection. All these solutions were the subject of many academic papers published in international conferences and journals.