As a final test we applied our best model to the test set given in [8] and got better results than they got with their own model. Our model gives 72.72% Bs correct, 4.27% NB incorrect, while their best gives 70% Bs correct, with 5% NB incorrect. To be fair we cannot claim that this improvement is due solely to our model as our model was trained on a significantly larger training set, but it does show that our model is not specific to our particular test set.