Kneser-Ney context counts

Andreas Stolcke stolcke at speech.sri.com
Sat Jan 3 20:14:05 PST 2009


Alexy Khrabrov wrote:
> Greetings Andreas -- I'd like to access the number of contexts for any 
> given ngram, used in Kneser-Ney computation (those with the fat dot).  
> What's a good way to get at them via the C++ API?
You create a ModKneserNey object (Discount.h). Be sure to leave the 
countsAreModified parameter at the default value (false).

Then invoke ModKneserNey:: estimate() on your counts.  As  a side 
effect, the lower-order counts will be modified to reflect the context 
type counts.  Note that the counts of ngrams starting with <s> are 
unchanged since there are no preceding words for them.

Andreas






More information about the SRILM-User mailing list