Reading

vs +/"Kneser-Ney" $NOTES/ws/nlp-natural-language-processing/reading/eisenstein-nlp-notes.txt

Kneser-Ney smoothing

Based on absolute discounting, but it redistributes the resulting probability mass in a different way from Katz backoff.

Empirical evidence points to Kneser-Ney smoothing as the state-of-art for n-gram LMing.