NetNews Usenet Archive 1992 #26/NN_1992_26.iso/spool/bit/listserv/csgl/1383

Text File | 1992-11-04 | 6.4 KB | 127 lines

Comments: Gated by [e-mail]
Path: sparky!uunet!news.encore.com!csar!foxtail!sdd.hp.com!zaphod.mps.ohio-state.edu!darwin.sura.net!europa.asd.contel.com!paladin.american.edu!auvm!PARC.XEROX.COM!SIBUN
X-Delivery-Notice:  SMTP MAIL FROM does not correspond to sender.
X-else-reply-to: [e-mail]
Fake-Sender: [e-mail]
Message-ID: <[e-mail]>
Newsgroups: bit.listserv.csg-l
Date:         Wed, 4 Nov 1992 16:22:53 PST
Sender:       "Control Systems Group Network (CSGnet)" <[e-mail]>
From:         Penni Sibun <[e-mail]>
Subject:      Re: description
In-Reply-To:  "(Avery Andrews)"'s message of Wed,
              4 Nov 1992 11:06:17 -0800 <[e-mail]>
Lines: 111

(penni sibun 921104.1400)

   [Avery.Andrews (921104.1349)]
     (Rick Marken (921103.1500))

    >The important part of the process in the PCT model of language is that
    >it is closed loop -- the words that represent the meaning (ws) of what
    >is being said are generated by an error signal -- not by the intended
    >meaning itself, rm1. ...

   In a sense, something like this is already built into Salix.  The
   entities and relations in the KB that is to be expressed have a
   bit which marks `already said', and the generation process is organized
   so as to go on unitl everything is so-marked.

this actually isn't quite true.  salix (my generation system) doesn't
``say everything.''  when it's told to talk about something, say a
house or a family, it's told about a ``completeness criterion''--a
specification of what it must make sure to mention.  as soon as it's
mention the critical things, it can stop.  in a house, the
completeness criterion is all the rooms (possibly excluding the
bathroom; this criterion for house/apartment descriptions is
well-documented); in a family, it's something like all the family
members still living, or here at thanksgiving.  salix starts with a
representation of all the objects/properties in a house or family or
whatever it's talking about and the relations between them.
(actually, salix keeps all its knowledge of everything together.)
whenever salix includes an object or a relation in the text it
produces, it marks it as already-said.  it also can make simple
inferences and mark as already-said things that are not explicitly
said.  for example, if salix says ``penni's father is john,''
both (father-of john penni) and (daughter-of penni john) are marked.

when salix is done w/ a description, it may or may not have marked all
the objects as already said.  it *never* has marked all the
relations it knows about.  in principle, it could happen, but it just
never does.  for instance, in a house, all the rooms bear lots of
relations to each other such as proximity (or not), direction (eg,
north-of, east-of), and connected via a door (or not).  salix's job is
to talk about the rooms and how they are related, but it is sufficient
to express one or two of those relations.  (in fact, in a house
description, the direction relations are never directly expressed at
all.  instead, they are used to dynamically compute whether a room is,
eg, ``left of'' or ``right of'' another room.  such relations are
never marked.)

   The problem is that the reference level (the whole KB)

the reference level must be a function of the kb, mediated by the
completeness criterion, right?

   & the controlled
   perceptions (the portion of the KB marked `already said' are not
   continuous-valued vectors ...  -- people just don't know how to make
   connectionist systems do the sorts of things that speakers do.

exactly.  i absolutely agree that my symbolic representations are
extremely crude and almost certainly completely wrong.  on the other
hand, if i were to try to use continuous values, i'd still be trying
to figure out how.

   [Avery Andrews 921104.1654]

   Here's an aspect of the PCT-ification of Salix that I've been
   overlooking for a while.  Salix as it stands simply marks information
   as said when it thinks it's said it, but what's really going on
   has to be a bit different.  In the first place, for adult conversation,
   the things `to-be-said' should really be taken as things `to-be-known-
   by-the-hearer'; so that there reference `signal' is what the hearer is
   wanted to know, the perception what the hearer is `perceived' as
   knowing, where a person's being in the presence of X being said is normally
   taken as a sufficient basis for concluding that the person knows X (there
   must be a better way to say this, but it eludes me for the present).

well, remember that there's a little bit of this done by the inference
mech., as described above.  i agree that this sort of thing needs to
be captured somehow.  i avoided that sort of thing like the plague
when i drew up my thesis proposal, because in ai/comp ling, this sort
of thing falls into a black hole called ``user/hearer modeling.''  the
basic idea of this is that i build a model in my head of what you have
in your head, and we pretend infinite regress doesn't happen.  anyway,
i didn't want to get into that.  now i'm not constrained by passing a
thesis, i need to look into this in a more intelligent way.

   The point of all this: suppose two people, say a mother and father,
   are recounting some incident (what their kid did) to a third person.
   They can do this without both of them just reeling off everything they
   know about the incident, since one person's presenting an aspect of the
   incident cancels that contributor to the error signal for both of them
   (I think that what is going on is that one person takes the role of
   primary narrator, the other filling in with such bits as the first leaves
   out).  Having the loop closed throught the environment should make it
   easier to explain how cooperation works.

right.

   With children, I'm not so sure about all this.  My impressions are that
   (a) they just want to say what they want to say (and definitely don't
   want anyone else to say it) (b) aren't really focussed on conveying
   new information to a hearer--savoring common knowledge seems to be more
   like what they're on about.

i do think that some adult communication is like this.  as bill
suggested and as i hypothesize in my thesis, description (or story
telling) is like this, at least in the small.  in general, i think
lots of language is like this in the small.  but interspersed with
this, all sorts of interesting other stuff is going on.

cheers.

                --penni