Robustness issues in text mining