categorization and allusion (238)
Willard McCarty (MCCARTY@VM.EPAS.UTORONTO.CA)
Wed, 22 Mar 89 20:08:29 EST
Humanist Mailing List, Vol. 2, No. 755. Wednesday, 22 Mar 1989.
Date: Wed, 22 Mar 89 14:58:29 EST
From: "James H. Coombs" <JAZBO@BROWNVM>
Subject: Categorization; allusion
Peter Junger (20 Mar 89) asks about "A List Taken from Borges"
Peter quotes that
the following message was posted in PHILRELSOC ISSUE #33
My suggestion of what there is (following Borges): (a) belonging to
the emperor, (b) embalmed, (c) tame, (d) sucking pigs, (e) sirens,
(f) fabulous, (g) stray dogs, (h) included in the present
classification, (i) frenzied, (j) innumerable, (k) drawn with a very
fine camelhair brush, (l) et cetera, (m) having just broken the water
pitcher, (n) that from a long way off look like flies.'
The point is this: it's one thing to enumerate a list of categories,
it's quite another to show how they derive and how they cohere. Take
Aristotle: he doesn't simply give a list in arbitrary fashion, but
suggests a principle according to which his categories are to be
derived. He also has an elaborate theory of how the various senses
of being are to be related.
But see Eleanor Rosch. 1978. "Principles of Categorization." In Rosch,
Eleanor, and Barbara B. Lloyd. Cognition and Categorization. Hillsdale, NJ:
Lawrence Erlbaum. 27-48.
Rosch disagrees somewhat on the quotation, so I give her version (which has a
more literary style and is more satisfying to read). Of central importance,
Rosch uses the passage to illustrate a very different point. According to
Rosch, it is not that we need justification for this categorization; it is
that this categorization will not occur in non-literary information processing.
The following is a taxonomy of the animal kingdom. It has been
attributed to an ancient Chinese encyclopedia entitled the Celestial
Emporium of Benevolent Knowledge:
On those remote pages it is written that animals are divided into (a)
those that belong to the Emperor, (b) embalmed ones, (c) those that
are trained, (d) suckling pigs, (e) mermaids, (f) fabulous ones, (g)
stray dogs, (h) those that are included in this classification, (i)
those that tremble as if they were mad, (j) innumerable ones, (k)
those drawn with a very fine camel's hair brush, (l) others, (m)
those that have just broken a flower vase, (n) those that resemble
flies from a distance (Borges, 1966, p. 108).
Certain types of categorizations may appear in the imagination of
poets, but they are never found in the practical or linguistic
classes of organisms or of man-made objects used by any of the
cultures of the world. For some years, I have argued that human
categorization should not be considered the arbitrary product of
historical accident or of whimsy but rather the result of
psychological principles of categorization, which are subject to
investigation. This chapter is a summary and discussion of those
principles.
I suspect that this passage will seem sadly naive to many readers. Rosch's
work is among the most sophisticated that I have seen. She knows Wittgenstein
and has a paper (with Mervis) on family resemblances in human categorization
(fuzzy sets). Basically, she introduced fuzzy set theory into (cognitive)
psychology. (Zadeh 1966 was in information theory.) No one is more aware of
the influences of language and culture on cognition and categorization. Her
work is supported by experiments. The experiments of the last 15 years have
refined her work. (This quick, awkward summary seeks to help those of us with
contemporary literary theoretical training---to help keep us from dismissing
Rosch's work without investigating it for ourselves.) I would guess that the
quickest way to get an overview of categorization theory would be to read
George Lakoff's *Women, Fire, and Other Dangerous Things*, which Lakoff
represents as a summary of the work of the last 15 years, beginning with about
60 pages on the Roschian work of the mid-70s. I don't vouch for it because I
haven't been able to get my hands on it yet.
I doubt very much that it is worth attempting to discuss these theories here.
Harnad already tried in psychnet, and things quickly blew up---not because
Harnad and Lakoff could not discuss the theory with equanimity but because
Harnad and most of the potential participants simply did not know the relevant
theory. (Harnad recently published a book on categorical perception, but the
line of inquiry is not quite the same.) In the end, the issue was there as it
is everywhere: should one spend one's time in open discussion of topics about
which one knows little, or should one become expert in a field and restrict
discussion to informal discussion with experts and indirect (via publications)
discussion with others. And how much of our attention should we direct toward
metadiscussion? Harnad says "I have time to talk"; Lakoff says "I have work
to do; let me know when you have done your homework." Harnad says "I don't
insist that you read *my* book." (These are paraphrases.)
In part, those few issues of psychnet (starting around 28?) will be interesting
to people who think about electronic communications. In part, let's not
repeat that discussion here.
Back to Peter's note:
Since any computer program that can deal with "legal reasoning"-- or
with the recognition of allusions, to mention a topic discussed on
this list--will have to deal with problems of categorization, I
suspect that it raises, in a backhanded way, one of the critical
problems in humanities computing: How can .... or, better: Can a
computer recognize when a human being would say that a certain
instance belongs to a certain category?
Our discussion of allusion began with a mention of fuzzy-set theory.
According to my understanding of allusion, an utterance is allusive to degree
0 or 1, with no in betweens. I have read nothing in categorization theory
to shake my theory of allusion.
I promised someone that I would elaborate on my maxim of non-repetition:
"Avoid repetition (of your own or of anyone's discourse or any features
thereof)." As I recall it, the point of confusion was focused on the idea of
avoidance. Allusions rely on repetition, so how would a maxim against
repetition explain allusions? Briefly and loosely, if I repeat recognizable
features of an utterance and you recognize that repetition, your attention is
going to be diverted to some extent away from my utterance and toward that
other utterance. But, you think to yourself, why would he want me to think
about that other utterance? He knows that I know that utterance. And he
knows that I will recognize the repetition. By all appearances, he hasn't
stopped making sense; I want to continue this conversation; he is usually
cooperative and seems to be attempting to cooperate with me now in an
interchange of information. So that other utterance must be somehow
relevant to the current utterance.
ALLUSION: Aha! I see. Wordsworth means to convey that his poem (The Prelude)
is in some way a continuation, a going-beyond, of Milton's (Paradise Lost).
*That* is why he said "the earth is all before me." Also, like Adam and
Eve,....
NOT AN ALLUSION: Blast Wordsworth! Everytime I read this passage, I think of
Samson Agonistes, but I can't find any reason for Wordsworth to direct my
attention to SA here. I have been over it and over it. Maybe he just likes
the words. Maybe I am missing something. Whatever it is, I am tired of being
distracted by this repetition, and it never has been anything more than a
distraction.
That's the basic idea. If I repeat recognizable features, I have been
uncooperative unless I intend that repetition to convey further information.
(Or unless our mutual contextual beliefs are such that you will not be
tempted to take the repetition as another of those instances of conveying
information by repeating.)
So, to allude, violate the maxim; otherwise, don't violate it. Just as to
convey something ironically, violate the maxim of quality (say only that which
you believe to be true); otherwise, don't violate it.
Back to Peter's question. I don't know whether or not a computer could be
programmed to recognize allusions. I do know that it would not be worth the
effort. The problem is not important enough to solve all of the sub-problems.
I could not easily enumerate the repeatable features of an utterance. There
are many levels to an utterance, and all of them have repeatable features.
Beyond repetition, we have a problem of recognition. I regularly say, "And
that's true too." I have yet to hear someone respond, "That's from King
Lear." For the Wordsworth allusion, we would have to have not only both texts
on line and some pattern matcher but also something to filter out the
"unrecognizable" repetitions or, more properly, the similarities that are not
repetitions at all. Also, some module has to make judgments of intentionality
and of implication. And we have to be able to recognize repetitions of
rhythms and situations and themes and character types and combinations of
character types ad indefinitum.
When we solve non-allusive natural language processing, we will have most of
what we need for allusion. Such processing will have to cover mutual
contextual beliefs (MCBs). It will probably have to have some minimal ability
to process conversational implicature. Then we can worry about repetition and
the like.
Unfortunately?, MCBs can be very specific. There are MCB sets that I share
with individuals. All of us have participated in English conversations that
were completely opaque to acquaintances that could hear every word that was
exchanged. There would not be much reason to make these MCBs available to a
computer, and few of us would want to give up our rights to privacy. Every
group has its MCBs, and new comers learn them gradually, are only gradually
allowed to help shape them.
We are going to have to restrict our processing of MCBs and, consequently, our
processing of natural language to relatively public utterances. (I mean
utterances intended to be understood by large and "important" groups.) The
practice will be similar to the practice of selecting words for general
dictionaries: it goes in iff it occurs in a range of publications over at least
3-5 years (details determined by editor).
Can a computer recognize when a human being would say that a certain
instance belongs to a certain category?
Which human being? Which day of the week? What sort of category? Is it a
relatively typical instance? Is it typical for that category-context pair?
A variety of experiments have found that the most typical bird is the robin.
Turkey is much lower on the scale of typicality. Unless it's Thanksgiving.
Then "cooking up a bird" makes people think of turkeys.
We have a lot of evidence for categorization into superordinate, basic level,
and subordinate. Basic level shifts with context and with expertise. "I saw
John in a sedan the other day." More likely that I saw him in a "car." Most
of us would wonder what was so special about John's sedanity. What, is he a
convertible nut or something? If so, and I know it and all of that, then it
would be strange to say that I saw him in a "car." Is this just a matter of
discourse strategies? Did we also categorize at the car level? My guess
is that we did not. Researchers are trying to distinguish representation from
retrieval. They are finding a lot of evidence against "static" semantics. I
might well create a representation at the sedan level only. I certainly could
later infer that he was in a car, but I have insufficient evidence at this
point for believing that I activate a node in a semantic network and that I
can, therefore, be said to have categorized at a number of levels at the time
of perception and representation. (Not just semantic networks: any static
structure.)
Computers certainly can perform some kinds of categorization very accurately.
Parsers can categorize words into classes. Most of the time, a person who
disagrees with the program will be wrong. (Most people don't know a verb
from a noun anyway---well, they know, but they can't attach those labels.)
Lotus Agenda has a category editor and a rules system. I can gradually
construct rules so that it will not only know that "Meeting with Norm" should
be categorized as: event_type=meeting, people=norm, priority=high, etc. We
believe that a system of this sort would not be of much value for a large,
shared database, however, and have elected not to pursue that direction in our
work on Intermedia. According to the reports, most people want something that
is more structured than Agenda anyway. People just don't have time to train
a program to recognize instances of categories by enumerating them. The rules
that do a little more work---such as categorizing a meeting as high priority
if...---such formalisms are more complicated than most people want to work
with.
I hope that I have given some sense of the range of possibilities. Have to
go. --Jim
Dr. James H. Coombs
Software Engineer, Research
Institute for Research in Information and Scholarship (IRIS)
Brown University
jazbo@brownvm.bitnet