Very much agree. Character bigrams will solve this perfectly with tiny amounts of training data (like a couple KB).