Mining knowledge from structured data has been extensively addressed in the few past years. However, most proposed approaches are interested in flat structures. With the growing popularity of the Web, the number of semi-structured documents available is rapidly increasing. Structure of these objects is irregular and it is judicious to assume that a query on documents structure is almost as important as a query on data. Moreover, manipulated data is not static since it is constantly being updated. The problem of maintaining such sub-structures then becomes as much of a priority as researching them because, every time data is updated, found sub-structures could become invalid. In this paper we propose a system, called A.U.S.M.S. (Automatic Update Schema Mining System), which enables us to retrieve data, identify frequent sub-structures and keep up-to-date extracted knowledge after sources evolutions.