This paper evaluates the performance of Maximum
Entropy (MaxEnt), Support Vector Machine (SVM) and Naıve
Bayes (NB) techniques for Indonesian text classification. Performance
of MaxEnt and SVM techniques are compared against
baseline NB technique. We also investigate the effect of language
dependent tools such as Indonesian stemming and stop words
removal can have on these techniques for text classification performances.
Up to now, there is no experimental report about the
effect of Indonesian stemmer on the text classification accuracy.
From our experiments, we conclude that maximum entropy
performs better than other classifiers in general. Language
dependent tools such as stemming and stop words removal have
only little effect on the accuracy of text classification. However
stemmed approach scored highest average accuracy and due to
the dimension reduction of feature vectors used in classification,
makes this approach is viable step in pre-processing stage.
To appear in SITIA 2010.