Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!seismo!ut-sally!berleant
From: berleant@ut-sally.UUCP (Dan Berleant)
Newsgroups: comp.cog-eng,comp.ai
Subject: Re: The symbol grounding problem: 3 routes to grounding needed?
Message-ID: <8404@ut-sally.UUCP>
Date: Fri, 3-Jul-87 18:15:43 EDT
Article-I.D.: ut-sally.8404
Posted: Fri Jul  3 18:15:43 1987
Date-Received: Sat, 4-Jul-87 14:33:29 EDT
References: <.... <6174@diamond.BBN.COM> <917@mind.UUCP> <14269@teknowledge-vaxc.ARPA> <955@mind.UUCP>
Reply-To: berleant@ut-sally.UUCP (Dan Berleant)
Organization: U. Texas CS Dept., Austin, Texas
Lines: 84
Xref: mnetor comp.cog-eng:180 comp.ai:611

In article <955@mind.UUCP> harnad@mind.UUCP (Stevan Harnad) writes:
>Another red herring in Wittegenstein's "family resemblance" metaphor was
>the issue of negative and disjunctive features. Not-F is a perfectly good
>feature. So is Not-F & Not-G. Which quite naturally yields the
>disjunctive feature F-or-G 
[sic] [...]
>There's absolutely no reason to restrict "features" to monadic,
>conjunctive features that subjects can report by introspection.

To my mind, this reconciles the 'classical' view of concept
representation with the 'probabilistic' one. There may not be much
difference between a classical view augmented to allow *arbitrary*
boolean expressions of features (instead of just a conjunction), and a
probabilistic view which proposes a list of possibly-present features
each associated with a number describing (e.g.) the probability of
category X given feature F. There is some difference, I think, which I
will ignore for the purposes of this posting.

>The
>problem in principle is whether there are any logical (and nonmagical)
>alternatives to a feature-set sufficient to sort the confusable
>alternatives correctly. I would argue that -- apart from contrived,
>gerrymandered cases that no one would want to argue formed the real
>basis of our ability to categorize -- there are none.

You may be dismissing the typicality and reaction time results that have
been interpreted as supporting probabilistic and exemplar-based category
representations too quickly. We need to be able to explain both these
results, and, the observed correctness of categorization even for atypical
instances. Yes, we can classify even a platypus as a mammal, but the fact
that it takes longer is not irrelevant.
   The answer may lie in the hypothesis that there are 2 representations
for categories: a 'core' of defining features, and a heuristic
categorizer that uses other features that are handy and useful rather
than defining (or maybe the heuristic categorizer uses exemplars instead
of features). For example, my concept of thunder contains defining
features like 'caused by lightning', and 'is an atmospheric phenomenon'.
However, in actual practice I am more likely to identify thunder using
heuristic features like 'rumbling noise' and 'associated with
rainstorms'. The heuristic categorizer then accounts for things like
typicality and reaction time results in psychological experiments. The
core is used as a slower-but-surer checker, which ensures correctness
(e.g.  informing me that a certain loud boom is thunder, while a loud
rumbling truck passing in the rain is not producing thunder, even though
the faster heuristic categorizer might have disagreed).
   Thus, we have 2 pathways along which categories are grounded. What is
the third?
   Anders Weinstein writes that the semantic meaning of terms is not
dealt with by such arguments, but is nevertheless an important part of a
category. He illustrates this with the fact that thunder may mean (to
some!) 'angry gods nearby'. How is this aspect of a category to be
grounded? First of all, the terms in the definition presumably are
grounded via the 2 routes discussed above (even 'gods', I suppose).  But
what basis is there for claiming the definition exists in the first
place? One basis is, you ask people and they tell you. But it would be
preferable to have an objective basis for grounding since people are not
always the most reliable measuring instruments. 
   I pointed out in a previous article that perhaps logic can help here.
Consider a sentence with 2 variables, e.g. FISH SWIM, where FISH and
SWIM are variables. Thus the sentence we are considering is equivalent
to the sentence A B. Obviously, many bindings would satisfy the sentence.
For example, FISH=fish and SWIM=swim, or FISH=mountains and SWIM=erode
(because fish do swim, and mountains do erode). 
   By adding many more true sentences, the possible bindings of the
variables become much more constrained. To illustrate, consider now the
sentence FISH LIVE, where FISH and LIVE are variables. Clearly the
variable FISH can be potentially bound to fewer meanings than before,
since whatever FISH is bound to must both SWIM and LIVE.  As we add more
and more sentences, eventually perhaps there will no longer be a set of
variable bindings for which all the sentences (millions of them, maybe,
just go to a library and start counting!)  are true and in which FISH
means 'mountain'. On the other hand, this is only a hypothesis: Maybe a
Martian attempting to deduce the meanings of English words could figure
out a way to do it consistently with FISH=mountain. Indeed maybe your
neighbor has already figured out a way to do that, but as long as you
both agree on the truthfulness of all the sentences you are mutually
aware of, there is no way to tell! Shades of the Turing test...
   My question is, would this method of 'grounding' the semantics
of categories be sufficient to do the job? Only in theory? Potentially
in practice? ...

Dan Berleant
UUCP: {gatech,ucbvax,ihnp4,seismo...& etc.}!ut-sally!berleant
ARPA: ai.berleant@r20.utexas.edu