[SRILM User List] Smoothing Error

[Andreas Stolcke]

KN smoothing modifies the lower-order counts before estimating
probabilities.
Therefore, the unigram counts in the two cases you ran are not the same.
(You could use the -write option to dump out the modified counts.)
For details, see the Chen & Goodman paper.

BTW, please join the srilm-user group if you plan to use SRILM in a
serious way, and address questions like these to the group.

Andreas

On 3/8/2010 11:25 PM, 罗华清 wrote:
> Dear sir:
> I am new to SRILM. When I tried to build a language model from the
> count file "nettalk.count", I got the following error:
> but when input other command, the error was gone:
> Why did I get different values of n(1, 2, 3, 4) when both commands
> above were using modkn smoothing for 1-grams?
> Thanks.
> Huaqing Luo
> Department of EE, TsingHua University, China.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.speech.sri.com/pipermail/srilm-user/attachments/20100309/4a4230d2/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: image/bmp
Size: 465258 bytes
Desc: not available
URL: <http://www.speech.sri.com/pipermail/srilm-user/attachments/20100309/4a4230d2/attachment.bin>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: image/bmp
Size: 560706 bytes
Desc: not available
URL: <http://www.speech.sri.com/pipermail/srilm-user/attachments/20100309/4a4230d2/attachment-0001.bin>