A smoothing technique for n-gram language models that improves probability estimates for rare word sequences.