1. Explore the performance of your classifier for different values of n in n-grams. Does it change significantly?
2. The charts in the previous section suggest that some stemming might be needed, as some of the top features are quite similar in meaning. Apply stemming and look at the shape of the ROC curve.
3. Experiment with other classifiers, for instance, random forests or support vector machines, and different kinds of n-grams. Can you improve the performance?