From: Anders Sandberg (asa@nada.kth.se)
Date: Thu May 03 2001 - 08:15:15 MDT
torsdagen den 3 maj 2001 14:59 Ben Goertzel wrote:
> > Yes, it definitely is. The only problem is implementing it - that is far
> > trickier, nearly AI-complete.
>
> Using our Webmind Classification System product, we could make an
> extropy-detector that would agree with humans as often as they agree with
> each other. You'd need to give it training data -- a few hundred positive
> and negative example of extropy-relevant messages, according to a few
> different raters.
>
> One thing that you find when doing such work is that different raters don't
> agree nearly as well as you'd think about what belongs in a given category
> and what doesn't...
This is the usual problem with hand-labelling training data for learning
systems. Raters can be extremely unreliable. I recall a lecture at the
Swedish neural network society some years ago where one system for crater
detection and classification in radar images turned out to be only limited by
humans disagreeing on what constitutes a crater...
It would be interesting to see what webmind would be able to do with mailing
list posts.
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 08:07:27 MST