Copy Link
Add to Bookmark
Report
AIList Digest Volume 5 Issue 181
AIList Digest Wednesday, 15 Jul 1987 Volume 5 : Issue 181
Today's Topics:
Classification - Natural kinds & Fuzzy Categories,
Comment - Need for Harnad-Style Discussions
----------------------------------------------------------------------
Date: 10 Jul 87 1019 PDT
From: John McCarthy <JMC@SAIL.STANFORD.EDU>
Subject: Natural kinds
Recently philosophers, Hilary Putnam I think, introduced the concept
of natural kind which, in my opinion, is one of the few things they
have done that is useful for AI. Most nouns designate natural kinds,
uncontroversially "bird", and in my opinion, even "chair". (I don't
consider "natural kind" to be a linguistic term, because there may
be undiscovered natural kinds and never articulated natural kinds).
The clearest examples of natural kind are biological species -
say penguin. We don't have a definition of penguin; rather we
have learned to recognize penguins. Penguins have many properties
I don't know about; some unknown even to penguin specialists.
However, I can tell penguins from seagulls without a precise definition,
because there aren't any intermediates existing in nature.
Therefore, the criteria used by people or by the programs we build
can be quite rough, and we don't all need to use the same criteria,
because we will come out with the same answer in the cases that
actually arise.
In my view the same is true of chairs. With apologies to Don Norman,
I note that my 20 month old son Timothy recognizes chairs and tables.
So far as I know, he is always right about the whether the objects
in our house are chairs. He also recognizes toy chairs, but just
calls them "chair" and similarly treats pictures of chairs in books.
He doesn't yet say "real chair", "toy chair" and "picture of a chair",
but he doesn't try to sit on pictures of chairs. He is entirely
prepared to be corrected about what an object is. For example, he
called a tomato "apple" and accepted correction.
We should try to make AI systems as good as children in this respect.
When a an object is named, the system should generate a
gensym, e.g. G00137. To this symbol should be attached the name
and what the system is to remember about the instance. (Whether it
remembers a prototype or a criterion is independent of this discussion;
my prejudice is that it should do both if it can. The utility of
prototypes depends on how good we have made it in handling similarities.)
The system should presume (defeasibly) that there is more to the concept
than it has learned and that some of what it has learned may be wrong.
It should also presume (although will usually be built into the design
rather than be linguistically represented) that the new concept is
a useful way to distinguish features of the world, although some new
concepts will turn out to be mere social conventions.
Attaching if-and-only-if definitions to concepts will sometimes be
possible, and mathematical concepts often are introduced by definitions.
However, this is a rare case in common sense experience.
I'm not sure that philosophers will agree with treating chairs as
natural kinds, because it is easy to invent intermediates between
chairs and other furniture. However, I think it is psychologically
correct and advantageous for AI, because we and our robots exist
in a world in which doubtful cases are rare.
The mini-controversy about penguins can be treated from this point of
view. That penguins are birds and whales are mammals has been discovered
by science. Many of the properties that penguins have in common with
other birds have not even been discovered yet, but we are confident that
they exist. It is not a matter of definition. He who gets fanatical
about arbitrary definitions will make many mistakes - for example,
classifying penguins with seals will lead to not finding tasty penguin
eggs.
------------------------------
Date: Sat 11 Jul 87 21:45:36-PDT
From: Ken Laws <Laws@Stripe.SRI.Com>
Reply-to: AIList-Request@STRIPE.SRI.COM
Subject: Natural Kinds
I would not be so quick to thank recent philosophers for the concept
of natural kinds. While I am not familiar with their contributions,
the notion seems similar to "species" in biology and "cluster" in
engineering and statistics. Cluster and discriminant analysis go
back to at least the 1930s, and have always depended on the tendency
of objects under study to group into classes.
-- Ken
------------------------------
Date: 13 Jul 87 16:31:17 GMT
From: uwslh!lishka@rsch.wisc.edu (Christopher Lishka)
Subject: Re: The symbol grounding problem: "Fuzzy" categories?
In article <3930@sunybcs.UUCP> dmark@marvin.UUCP (David M. Mark) writes:
>In article <974@mind.UUCP> harnad@mind.UUCP (Stevan Harnad) writes:
>>
>>
>>In Article 185 of comp.cog-eng sher@rochester.arpa (David Sher) of U of
>>Rochester, CS Dept, Rochester, NY responded as follows to my claim that
>>"Most of our object categories are indeed all-or-none, not graded. A penguin
>>is not a bird as a matter of degree. It's a bird, period." --
>>
>>> Personally I have trouble imagining how to test such a claim...
>>
>>Try sampling concrete nouns in a dictionary.
>
>Well, a dictionary may not always be a good authority for this sort of
>thing.
I don't want to start a huge discussion on a related topic, but I guess I'll
throw in my two-cents worth.
Mr. Harnad states that one should try sampling concrete nouns in a
dictionary. It seems to me that a short while ago there was some
discussion around the country as to what a dictionary's purpose
actually is, to which a prominent authority on the subject replied
that a dictionary is *only* a description of what people are commonly
using certain words for. Now, one upshot of this seems to be that a
dictionary, in the end, is NOT a final authority on many words (if not
all of them included). It can only provide a current description of
what the public in general is using the word for.
In the case of some words, many people will use them for many
different things. This may be one reason for the problems with the
word 'map.' In the case of a penguin, scientifically it is considered
a bird. I consider it a bird, although a penguin certainly does not
fly in the air. However, if every English-speaking person except a
few, say myself and Mr. Harnad, suddenly decided to think of a penguin
as something other than a bird, than a dictionary's description would
need to be changed, for myself and Mr. Harnad would be far outweighed.
I suspect that the dictionary would have some entry as to the
historical meaning of 'penguin' (i.e. a penguin used to be considered
a bird, but now it is something else). However, since a dictionary is
supposed to be descriptive of a language in its current usage, the
entry for penguin would have to be modified.
Which brings me to my point. Given that a dictionary is a descriptive
tool that seeks to give a good view of a language as it is currently being
used, can it really be used as a final authority? My feeling is no;
just look at all the different uses of a certain word among your
friends, not to mention the entire state you live in, not to mention
your continent, not to mention the entire English-speaking population
of the world. Holy cow! You've suddenly got a lot of little
differences in meaning for a certain word. Not to mention slang and
local terms (e.g. has anyone ever heard of the word 'bubbler?' It
means a 'Water Fountain' here in Wisconsin, but you'd be surprised how
many people don't know this term). In this case you can only look at
words as a 'graded' term, not an all-or-none term if you are using a
dictionary as the basis for a definition. Sure, if you want to use a
scientific definition for penguin, go ahead...since science seems to
seek to be unambiguos (unlike general spoken language), then you will
have a better all-or-none description. But I don't think you can go
about using a dictionary, which is a descriptive tool, as an
all-or-none decisive authority on what a word means. If I remember
back to a Linguistics course I took, this is the same difference as
denotation vs. connotation.
A couple notes: if you notice above (and right here), I use the word
'you' (as a technical writer would use the word 'one') to refer to a
person in general (i.e. the reader). This is not generally accepted
as proper English by the people who seek to define proper English, but
it is the term that is used by most people that I have known (here in
Wisconsin). It seems to me that this is further evidence of my
argument above, because I do not think twice in using this term 'you;'
it is how I was raised.
Also, please don't start a discussion on language in this group unless
it pertains to A.I. (and in some case it does); I just felt that
someone ought to speak up on the ambiguity of words, and how to
different people there might be problems with using a dictionary as a
basis for judgement. If you want to continue this discussion, please
e-mail me, and I will respond in a decent amount of time (after I cool
off in the case of flames ;-)
--
Chris Lishka /lishka@uwslh.uucp
Wisconsin State Lab of Hygiene <-lishka%uwslh.uucp@rsch.wisc.edu
\{seismo, harvard,topaz,...}!uwvax!uwslh!lishka
------------------------------
Date: 10 Jul 87 16:47:54 GMT
From: trwrb!aero!venera.isi.edu!smoliar@ucbvax.Berkeley.EDU (Stephen
Smoliar)
Subject: Re: The symbol grounding problem: "Fuzzy" categories?
In article <3930@sunybcs.UUCP> dmark@marvin.UUCP (David M. Mark) writes:
> we conducted
>a number of experiments and found many ambiguous stimuli near the boundary
>of the concept "map". Air photos and satellite images are an excellent
>example: they fit the dictionary definition, and some people feel very
>strongly that they *are* maps, others sharply reject that claim, etc.
>Museum floor plans, topographic cross-profiles, digital cartographic
>data files on tape, verbal driving directions for navigation, etc., are
>just some examples of the ambiguous ("fuzzy"?) boundary of the concept
>to which the English word "map" correctly applies. I strongly suspect
>that "map" is not unique in this regard!
Indeed, it almost seems as if "What is a map?" is not really the appropriate
question. The better question might be "What can be used as a map?" or
perhaps "How can I use a FOO as a map?" Furthermore, I agree that "map"
is probably not unique. There are probably any number of bindings for
BAR for which "What is a BAR?" runs into similar difficulty and for which
"How can I use a FOO as a BAR?" is the more useful question.
One candidate I might propose to discuss along these lines is the concept
of "algorithm." There are any number of entities which might be regarded
as being used as algorithms, ranging from Julia Child's recipies to
chromosomes. It would seem that any desire to classify such entities
as algorithms is only valuable to the extent that we are interested in
the algorithmic properties such entities possess. For example, we might
be interested in the nature of recipes which incorporate "while loops"
because we are concerned with how such loops terminate.
In an earlier posting, Harnad gave the example of how we classify works of
art according to particular styles. Such classifications may also be
susceptible to this intermediate level of interpretation. Thus, you
may or may not choose to view a particular tapestry as an allegory.
You may or may not choose to view it as a pastoral. Such decisions
influence the way you see it and "parse" it as part of your artistic
appreciation, regardless of whether or not your particular view coincides
with that of the creator!
I suspect there is a considerable amount of such relativity in the way we
detect categories. That relativity is guided not by what the categories
are or what their features are but by how we intend to put those
categories to use. (In other words, the issue isn't "What features
are present?" but "What features do we want to be present?")
------------------------------
Date: 14 Jul 87 15:37:00 GMT
From: apollo!laporta@beaver.cs.washington.edu (John X. Laporta)
Subject: Re: The symbol grounding problem: "Fuzzy" categories?
In article <245@uwslh.UUCP> lishka@uwslh.UUCP (Christopher Lishka) writes:
>Given that a dictionary is a descriptive
>tool that seeks to give a good view of a language as it is currently being
>used, can it really be used as a final authority? My feeling is no;
SUMMARY
(1) You are absolutely right. There is no 'final authority' because language
changes even as one tries to pin it down, with a dictionary, for example.
(2) AI programs designed to 'understand' natural language must include
an encyclopedic as well as a lexicological (dictionary) competence.
(3) The nonexistence to date of perfect artificial understanders of natural
language should not be surprising, given the enormity of the task of
constructing an artificial encyclopedic competence.
(4) The encyclopedia in this instance must grow with the language, preserving
past states, simulating present states, and predicting future states.
ELABORATION
Tackling (2) first:
While dictionary definitions are helpful guides in some respects, the nature of
linguistic competence is encyclopedic rather than lexicological. For instance,
you might hear someone say:
Because I was going to give a cocktail party, I went to the mall
to buy whiskey, peanuts, and motor oil.
A lexicological competence would deem this sentence grammatical and
unremarkably consistent, since 'mall' includes the availability of all the
items mentioned. An encyclopedic competence, on the other hand, would
mark this sentence as strange, since 'motor oil' is not a part of 'cocktail
party,' unless, I suppose, you were willing to assume that some of the guests
needed mechanical, not social, lubrication. Even this conjecture is unlikely,
however, because 'cocktail party' includes humans consuming alcoholic
beverages. A case of Billy Beer at the local Exxon is not a cocktail party.
Car mechanics do not come to work in little black dresses. An encyclopedic
competence is able (a) to isolate the assumptions an utterance requires for
coherence, (b) to rank their probability, and (c) thus to evaluate the
coherence of the utterance as a whole.
Further, 'encyclopedic' in this context includes more than is found in the
_Brittanica_. A humorist might write (in the character of a droll garage
mechanic) about a parley to negotiate sale of a gas station. He decides to
provide a little festive atmosphere by bringing along some beer. But even
this hypothesis doesn't eliminate all strangeness: why is the mechanic
buying motor oil at the supermarket? Certainly he could get a better price
from his distributor.
This sentence is a mine of linguistics lessons, but the above should be
enough to suggest my point. Encyclopedic competence, however provided,
(scripts or semantically marked graphs of words, to give two examples
which are not mutually exclusive) is crucial to understanding even the
topic of an utterance.
The wider question evolves from (1) ... :
Language is an elaboration of symbols which refer to other symbols. The
'last stop' (the boundary of semiotic analysis, not the the boundary of the
linguistic process itself in actual beings or machines) is the connection of
certain signs to 'cultural units.' These pieces of memory are what ground
symbol nets to whatever they are grounded upon. (I prefer Harnad's
formulation, but that is not crucial for this discussion.) When Og the Caveman
remembers one morning the shape of the stone that he used as a scraper
yesterday, a cultural unit exists, and stones of that shape are the first signs
dependent upon it. To oversimplify, the process continues infinitely as signs
are connected to other signs, new cultural units are formed, signs modify
other signs, etc.
... and concludes with (3) and (4):
Meaning is 'slippery' because language changes as it is used. A historically
amnesiac encyclopedic competence for 1980 would mark as improbable
sentences used daily at American slave auctions of the 1840's.
SOURCE NOTE: Nearly everything I have said here has been elaborated by
Umberto Eco in his book 'A Theory of Semiotics' and subsequent writings.
------------------------------
Date: 14 Jul 87 20:25:01 GMT
From: ritcv!cci632!dwp@cs.rochester.edu (Dana Paxson)
Subject: Re: Thanks. (was Re: Results of Symbol Grounding Poll)
In article <1010@mind.UUCP> ghn@mind.UUCP (Gregory Nelson) writes:
>In article <993@mind.UUCP> harnad@mind.UUCP (Stevan Harnad) writes:
>>[]
>>[make the] Net the reliable and respectable medium of scholarly communication
>>that I and (I trust) others are hoping it will evolve into.
>> ...
>>(4) I continue to be extremely enthusiastic about and committed to
>>developing the remarkable potential of electronic networks for scholarly
>>communication and the evolution of ideas. I take the present votes to
>>indicate that the current Usenet Newsgroups may not be the place to attempt
>>to start this.
>
> ... Perhaps you should take some time off to
>look at some of the other newsgroups. The comp.xxxx discussions are naturally
>oriented to computer people, but things like rec.xxx and sci.xxx are much
>more "broadminded" (if you will.) If you want a real surprise, try tuning
>in to the Deja Vu discussion on misc.psi or something like that.
>
I realize that this is belated input.
As one who followed along with an occasional understanding of
the discussion on symbol grounding, I have been attracted both
to the discussion and to the way in which Stevan Harnad
conducted it. I admire the discipline and rigor evident in his
postings, and see his work as an example of how a newsgroup
functioning often as a bulletin board with limited scope can
be enriched by some really difficult exploration. Some of the
other contributors to the discussion appeared to work well at a
level near Mr. Harnad's. It has been an exciting series of
exchanges.
I regret the loss of the discussion from the newsgroup. Any
reader of the most potent material on computer science will find
that the authors reach out to many fields to gain inspiration,
illustration, and, yes, even forms of grounding(!) for their
work. Especially grounding.
Like any other science area with meaning, computer science does
not begin in words (or bytes) and end in bytes (or words). It
ends in application, or at least applicability, to our lives.
In the AI realm, that applicability is becoming an intimate
metamorphism, a mapping/transformation, of how we work rather
than a translation of what we do. If I can characterize an
aspect of the symbol grounding discussion, it is a knife-sharp
exploration of the type of problem dismissed by so many as
having a self-evident solution. This class of problem is
precisely the type which is most difficult even to see, let
alone solve. Witness the depth and detail of the exchanges we
have seen. If others become impatient with the material, they
don't have to read it; but this topic area appears to be poorly
understood by anybody, and desperately needs close dialogue.
Personally, I feel strongly the need to extend my cognitive
framework with such powerful and challenging material.
Perhaps the outcomes from discussions like this one have too
much potential for making a lot of funded thesis work and
product development irrelevant... but then some outcomes can
unfold whole new realms of exploration and advancement. Unless
I am mistaken, these newsgroups can play an active role in this
unfoldment. I don't want to see anything this good be relegated
to an obscure electronic cranny, or lumped with a lot of diffuse
and irrelevant outpourings. Computer scientists have a lot to
learn from the symbol-grounding exchanges right here.
I sense that there are many quiet readers out there who have
powerful ideas relating to this subject, but who have kept
silent on seeing contemptuous and abusive complaints of
others about the length and content of the postings. For
complaints, it seems reasonable to address the complaints to
authors privately, or to the moderator if there is one; but
open criticism on the net discourages its use by those whose
insight and sensitivity exceed their boldness. Making one's
views public is an intimidating process in itself, so why should
we raise the level of intimidation?
For my part, I would like to ask for a citation for Mr. Harnad's
original article on the subject of symbol grounding; I want to
read it to find out what started the interchange I have seen. I
tuned in late in the process.
Thanks to all of the participants in this probing discussion.
The views expressed here are my own.
Dana Paxson
Systems Engineering
Computer Consoles, Incorporated
Rochester, New York
716 482-5000
CIS User ID: 76327,65
------------------------------
End of AIList Digest
********************
