The article presents the results of the experiments performed on selected sub-set of Wikipedia which we categorized automaticly. We analyze two methods of text representation: based on references and word content. Using them we introduced joint representation that has been used to build groups of similar articles based on Kohonen Self-Organizing Maps. To fulfill efficiency of the data processing, we performed dimensionality reduction of raw data using Principal Component Analysis performed on similarity matrix. Changing the granularity of SOM network allows to build hierarchical categories and find significant relations between articles in documents repository.