Abstract. The wide-spread applications of document digitization have lead to the use of structured digital representation methods such as the XML language. Extraction methodologies for the formatting metadata can be used on such structured documents for enhancing their accessibility, including augmented audio representation of documents. To the best of our knowledge, an effort has yet to be made to produce an automatic extraction system of semantic information of the document formatting, solely from document layout, without the use of natural language processing. In this study a corpus of XML representations of several issues of a Greek newspaper is used in order to create and evaluate a semantic classifier of text formatting, based on Bayesian Networks.