Computational linguistics

A trigram model has little evidence for a particular two-token history. Which strategy can use lower-order estimates while still using the trigram estimate when supported?