Exploiting structural information for semi-structured document categorization