When a trigram backoff language model is created from a large body of text, trigrams and bigrams that occur few times in the training text are often excluded from the model in order to decrease the model size. Generally, the elimination of n-grams with very low counts is believed to not...