Interpolating with -lambda 1.0

[Tanel Alumäe]

Hello,

The -bayes 0 switch didn't help, although it did change the calculated
perplexity values for lambda < 1.0

However, I discovered that I had the following in my classdefs:
<s> 1 <s>
</s> 1 </s>

After removing those lines, the interpolated perplexity with -lambda 1.0
is equal to the perplexity of the pure word trigram, as expected.

Regards,

Tanel A.


On Wed, 2005-02-23 at 10:23 -0800, Andreas Stolcke wrote:
> Try using -bayes 0 when running the interpolated model.
> Without it, ngram will construct a merged ngram model in memory,
> which does not work well when combining word and class-based models.
> 
> --Andreas
> 
> In message <1109146260.28101.1.camel at markov>you wrote:
> > Hello,
> > 
> > I'm a bit confused with interpolation.
> > I want to calculate test text's perplexity using different interpolation
> > weights (lambdas). Everything is OK until I set lambda to 1.0. Shouldn't
> > I then get the same perplexity as using only the base language model?
> > This doesn't seem to be the case:
> > 
> > $ ngram -lm trigram.arpa -ppl <testtxt> 
> > file <testtxt>: 2394 sentences, 29475 words, 1224 OOVs
> > 0 zeroprobs, logprob= -86274.9 ppl= 653.583 ppl1= 1132.06
> > 
> > $ ngram -lm trigram.arpa -ppl <testtxt> -classes <classdefs>  -mix-lm
> > class-trigram.arpa -lambda 1.0
> > file <testtxt>: 2394 sentences, 29475 words, 1224 OOVs
> > 0 zeroprobs, logprob= -85554.4 ppl= 619.144 ppl1= 1067.5
> > 
> > As shown, the perplexity is 653.539 when using standalone trigram, and
> > 619.144 when interpolating the trigram with the class-trigam, using
> > lambda 1.0. Why are they not equal?
> > 
> > Both word trigram and class trigram are close-vocabulary LMs, if it
> > matters.
> > 
> > Regards,
> > 
> > Tanel A.
> > 
> > 
> > 
>