I use the corpora of webtext to train n-gram model in NLTK.And i use katz back-off smoothing to solve 0 count-problem.But one problem is:while the smoothing used in based-on characters language model,r* (computed good-turning) >r(original count),as a result the counting coefficient alapa<0.what's the problem? And that problem doesn't come up in based in word language model. i need your help ,thanks

  • 4
    I'm voting to close this question as off-topic because this question focuses on programming details. It is better asked on stackoverflow.com or in Data Science. Commented Apr 11, 2017 at 10:32
  • I disagree. I think this is definitely an NLP question.
    – Adam_G
    Commented Apr 11, 2017 at 22:30
  • Is really r* > r? I thought that because some weight is backed off the r* counts are actually smaller than the real counts?
    – Claude
    Commented Apr 14, 2017 at 20:00
  • yes,i has turned the method of Witten-Bell backoff.
    – Tina_seven
    Commented Apr 30, 2017 at 13:19


Browse other questions tagged or ask your own question.