Hi,

you are correct that for POS LMs the lower order n-gram counts
are very different and smoothing is less relevant.

You could train a 7-gram LM with Good Turing smoothing for the lower
order n-grams and Kneser-Ney for the higher order n-grams.
I have done this occasionally.

-phi


On Tue, Nov 24, 2009 at 6:59 PM, Ben Gottesman <[email protected]> wrote:
> Hi,
>
> The Moses manual recommends using the following switches when building
> a language model with SRILM:
>
>  -interpolate -kndiscount
>
> I assume this recommendation applies specifically to surface-string
> language models.  For a part-of-speech language model, KN-discounting
> is inappropriate because it is based on counts-of-counts, and the
> counts-of-counts for POSes are odd in that there are very few POSes
> that occur only once or twice in a given corpus.
>
> Are there particular switches that are recommended for building a POS
> language model with SRILM?
>
> Regards,
> Ben Gottesman
> _______________________________________________
> Moses-support mailing list
> [email protected]
> http://mailman.mit.edu/mailman/listinfo/moses-support
>

_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to