[SRILM User List] linear interpolation of LM

[Andreas Stolcke]

On 10/18/2013 8:38 AM, MATHUR, Prashant wrote:
> Hi,
>
> I wanted to know how do I do linear interpolation of several models 
> given their weights?
> Also, can I interpolated more than 10 models at once?
>
> I tried several hit/trials so far. Nothing seems to work for me.
>
> $ngram -lm small.lm.1 -lambda 0.7 -mix-lm big.lm.1 -unk
> It doesn't throw any output or error.
>
> when I try -write option
> $ngram -lm small.lm.1 -lambda 0.7 -mix-lm big.lm.1 -unk  -write-lm 
> mixed.lm.1
> write() method not implemented
> error writing mixed.lm.1
The above command should work, unless you are not giving the complete 
command line in your example (if you add the -bayes option then you will 
see the "not implemented"  error).


>
> but when I try
> $ngram -lm small.lm.1 -lambda 0.7 -mix-lm2 big.lm.1 -mix-lambda2 0.3 
> -unk  -write-lm mixed.lm.1
> Then the mixed.lm.1 file is the same as small.lm.1
>
> My SRILM version is 1.5.3.
> I read that there are many ways of interpolation such as count-based 
> and log-linear interpolation.
> I tried the options -count-lm (throws a format error), -loglinear-mix 
> (didn't do anything)
>
> I am out of options. Please help!

You are using a very old version of SRILM.   Please get the latest 
stable version (1.7.0).

If you want to try the current beta version (1.7.1) you will find a new 
option (ngram -read-mix-lms)  that allows you to specify the mixture 
component LMs in a separate file, and also allows an arbitrary number of 
components.

Andreas


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.speech.sri.com/pipermail/srilm-user/attachments/20131018/9cd84e4f/attachment.html>