isn't a ngram bayesian model sufficient here?

also, isn't the test data linear dependent from the training set here so creating skewed performance measurement?