A note on document classification with small training data