Re: [Senserelate-users] acceptable values for contextScore and pairScore

[Ted P. <dul...@gm...>]

Hi Gil,

There are some subtleties to this question, and I'll answer at more length
in the next day or two. However, the first thing is to establish the bounds
of whatever measure
you are using. Some of the measures work on a range of 0 to 1 (e.g., lin,
wup, vector) whereas others work on less constrained ranges (e.g., lesk,
jcn, res). So, if you are working with lesk, the upper bound of that measure
is not really precisely defined, but it can easily be in the 100 or 200
range (sometimes more). So, with lesk a very small cutoff like .01 or .001
will probably have little or no effect, whereas with wup it might have a
very big impact. This would be true for both the parameters (pairScore and
contextScore).

We don't have precise recommendations for these scores, so there is going to
be some trial and error here I think. With lesk I would probably start with
pairScore around 10 and work my way up in increments of 10, and for the
contextScore I think it will probably depend a bit on your window size...but
I'll look at that a little and comment more in the coming days.

More soon, but just wanted to get started with the above. Hope this helps,
and let us know if you find or see anything interesting or problematic!

Thanks,
Ted

On Thu, Apr 10, 2008 at 11:00 AM, Gil Vidals <gv...@gm...> wrote:

> I would like to find words in my sentence where their is a minimum level
> of confidence that the disambiguation worked using
> WordNet::SenseRelate::AllWords. I see that there are two values listed in
> the wsd.pl doc -- contextScore and pairScore. These seem perfect for the
> job, but what values do I use? Should I use a minimum score of 0.001 or 0.1
> or 100 or 1000. What order of magnitude should I use here? Any help would be
> greatly appreciated.
>
> --Gil
>
> --contextScore=*REAL*
>
> If no sense of the target word achieves this minimum score, then no winner
> will be projected (e.g., it is assumed that there is no best sense or that
> none of the senses are sufficiently related to the surrounding context). The
> default is zero.
> --pairScore=*REAL*
>
> The minimum pairwise score between a sense of the target word and the best
> sense of a context word that will be used in computing the overall score for
> that sense of the target word. Setting this to be greater than zero (but not
> too large) will reduce noise. The default is zero.
>
> -------------------------------------------------------------------------
> This SF.net email is sponsored by the 2008 JavaOne(SM) Conference
> Don't miss this year's exciting event. There's still time to save $100.
> Use priority code J8TL2D2.
>
> http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone
> _______________________________________________
> senserelate-users mailing list
> sen...@li...
> https://lists.sourceforge.net/lists/listinfo/senserelate-users
>
>


-- 
Ted Pedersen
http://www.d.umn.edu/~tpederse