Macro Features Based Text Categorization

Text Categorization (TC) is one of the key techniques in web information processing. A lot of approaches have been proposed to do TC; most of them are based on the text representation using the distributions and relationships of terms, few of them take the document level relationships into account. In this paper, the document level distributions and relationships are used as a novel type features for TC. We called them macro features to differentiate from term based features. Two methods are proposed for macro features extraction. The first one is semi-supervised method based on document clustering technique. The second one constructs the macro feature vector of a text using the centroid of each text category. Experiments conducted on standard corpora Reuters-21578 and 20-newsgroup, show that the proposed methods can bring great performance improvement by simply combining macro features with classical term based features.