Application of a Structural Support Vector Machine method to N-gram based text classification in Serbian

Abstract

The paper presents classification results of a hierarchically organized document corpus in Serbian, by using Support Vector Machine method (SVM). Two techniques have been applied derived from the SVM with structural output: multiclass flat and hierarchical classification. Common representation model of a document and a class or a hierarchy of classes the document belongs to, specific for this form of the SVM method, is based on different length byte n-grams. Four tf-idf statistics have been used that define significance of an n-gram for a specific document. The techniques and statistics described have been tested on a hierarchically structured subset of the Ebart corpus of newspaper texts. The results obtained for both types of classifiers are similar for the corpus as a whole, while hierarchical classifier performs better for most specific classes with small number of texts.