Re: Re: Error (signal 11) during auto-train

Graham Toal wrote:
> I almost despair of ever creating a completely clean test suite
> of any significant size.  Note that for a good testing regimen
> you need more than one corpus, so that you don't find yourself
> 'training the corpus'.  I.e. you may use one set for
> training, but you use a different set to measure the effectiveness
> of the training.  Anyone who has ever written a compiler will
> be intimately familiar with the problem here.

Yes I agree with that.  I've gone through mailboxes manually and had to 
make many passes to get them all right.  I have a test corpus that I 
think is pretty clean now but it's not based on recent email anymore so 
it's not perfect either.  One or two misclassifications in a test corpus 
can throw off the accuracy percentages fairly significantly.

I have been using a variation on the SA public corpus lately to get away 
from basing all of my tuning decisions on my own email.  I haven't 
manually classified those so I'm assuming they are all properly classified.

All the best,
++Brian


-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language
that extends applications into web and mobile media. Attend the live webcast
and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=110944&bid=241720&dat=121642