[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: a propos of artificial intelligence, linguistics, and geckos



On Thu, 20 Jan 2000 01:09:36 -0600 (CST), Jay Jacobs wrote:

>On Thu, 20 Jan 2000 scozens@pwj.co.jp wrote:
>> We're also seeing the communication aspect coming up, which is
>> an important issue. (Are we starting to model social patterns?)
>> What do *you* do when you need to find something out? You
>> either already know it (it's in your DBM, as it were) or you
>> can ask someone else, you can look it up on the web or in an
>> encyclopaedia. In some situations, you may think you've got the
>> answer, but it doesn't seem to fit the question somehow. A
>> human can tell pretty easily when he's got the wrong reference
>> source for a topic; the query has a series of mental
>> associations. When we're looking up quavers, we can usually
>> work out if the answer's supposed to relate to music or cheesy
>> potato chips. We're about to hit a really nasty knowledge
>> representation issue here.
>
>wow.
>
>  I'm only going to comment on the first one... I started thinking about
>information validity.  It sounds like the path is being paved to a much
>smarter AI, and information validity will be a key component.  Everyone
>who's run an infobot knows how much invalid data is scraped from IRC, so
>I see these main topics:
>
>  How and where to get and accept information.

Accept liberally, and consider the source.  Is it some stranger?  Is
it the producer of quality factoids?

What defines quality?  Karma?  The number of factoids that haven't
been erased or overwritten by someone else?  Peer review, perhaps
with a karma-like system?

>  How to process and reference, and relate information.

I have been cross-referencing ambiguous factoid keys or related
values on purl, the EFNet #perl infobot.  It's as simple as:

  intermezzo is (see: inter-mezzo)

It would be interesting if infobot could follow the links, pulling
out the "inter-mezzo" factoid even when someone asks about "intermezzo".
That brings up issues like: Should changes to intermezzo be propagated
to inter-mezzo, or should factoids be copy-on-write?  Probably the
former, but I'm not willing to say always.

  use Can::Worms;
  my $can = new Can::Worms;
  $can->open();

The problem becomes harder if a link accompanies other information;
following it should be the reader's job.  For example, the referenced
"poe slides" factoid may have only a tenuous connection to
"inter-mezzo":

  inter-mezzo is a distributed file system at http://www.inter-mezzo.org/
  (see: poe slides) or version 001 is 90% lethal

>  How to store new information for future refence.
>
>  How to detect and remove bogus information that's already stored.

The human brane seems to forget least-used information.  How safe
would it be for geckobot to quietly prune deadwood?  What defines
deadwood?  Age since last recollection?  Relative fetch frequency?
Accumulators that increase by different amounts depending on the
access method (i.e., directed fetches, like "geckobot: factoid"
count more than undirected ones like "factoid?") and decay gradually
over time?

For the paranoid, decayed factoids could appear in a report instead
of die off quietly.  They could also be presented in a form-based
factoid manager, which, if geckobot's current design works out, may
possibly be integrated into the bot as a personal web server.

>[... removed by troc@netrus.net]

Of course, I'm handwaving a whole helluvalotta details.  Grains of
salt available upon request.

-- Rocco Caputo / troc@netrus.net