Perplexity in "ngram"

[Andreas Stolcke]

Mats Svenson wrote:
> Hi,
>  I have tried to use "ngram" to count perplexity of my
> LMs. However, I am not sure how does the srilm
> implementation treat OOVs in terms of counting
> perplexity. Is it that "log P(<unk>|history) != 0" or
> OOVs are just ignored? If a model with a higher number
>   
SRILM excludes words with zero probability from the perplexity 
computation and
reports their tally separately.  That includes OOV words when the LM 
doesn't contain
an unknown word (<unk>) token.

> of OOVs has a lower perplexity than another LM, does
> it mean that it is "better" in this -ppl
> implementation?
>   
Possibly.  You should not compare perplexities of LMs with different 
vocabularies.
> Second, in some discussions, I have heard about -ppl1
> option, but the current version does not seem to have
> it. In what -ppl1 differs from -ppl?
>   
There is no -ppl1 option.  -ppl reports a statistic labeled "ppl1", 
which is explained
in the ngram man page.
> Third, is there a way how to meaningfully compute
> perplexity for a hidden event LM? Or another way how
> to evaluate hidden event LM quality?
>   
Hidden event LMs are LMs, so you can compute a word-based perplexity just
like for any other LM.  If the goal of the HE-LM is to decode hidden events
(like sentence boundaries) then you can obviously evaluate that task as 
well.

Andreas