<< . .

( 29)

. . >>

ing point, combining properties, and so forth, but that they do vary,
when encountered, with respect to size, weight, shape, ownership, mon-
etary value, place they were mined (for mined minerals), and so forth.
You know that there are properties that the chemical kinds have qua be-
ing those very chemicals again, and that there are other properties that
only samples of them have.This is because you understand the category
chemical compound to correspond not merely (if at all) to a substance
(what, if any, are the determinate properties that every chemical com-
pound has?) but to a substance template (Section 1.8). Similarly, you
probably take me to be right about Xavier, not because you know him
personally, but because you understand him to fall under the substance
template human being, and you have a good idea what determinables are
likely to be constant for substances falling in this category.
But if you also agreed with me about Felis domesticus, why was that?
Was it because you know that every species of animal is uniform, for
example, with respect to adult size? But snakes and alligators keep right
on growing. Or because you know that every species of animal is liable
to vary in color from individual to individual? But this is not true of
most species. Nor is it true of most species “ perhaps only of mammals
and some birds “ that their behavior patterns may vary significantly
from individual to individual. Animal is not as well focused a template
as either chemical compound or human being. Knowing just that something
is an animal, you will have lots of ideas about the kinds of questions that
can be asked about it, but for a significant proportion of these, you may
not know in advance whether they can be answered univocally for the
species as a whole. If you are given that the animal is a mammal, of
course, this may help quite a lot. Categories like animal and mammal
correspond, of course, to substances “ each has some univocal proper-
ties of its own “ but, more important, for the project of gathering
knowledge, they bring with them substance templates. The categories
animal species, mammal species, person, crab, pebble, bridge, road, musical compo-
sition, chemical element, and book, for example, are all substance templates
as well as corresponding to substances in their own right. As templates,
they take predicates like “have shapes,” “have colors,” “have metabolism
rates,” “have specific gravities,” “have spatial lengths,” “have temporal
lengths,” “have designers,” “are written in languages (French, German),”
and so forth.
I have mentioned that substances vary both in the number of induc-
tions they support and in the reliability of the inductions they support.
Here we have a third kind of variability in substance quality. Substances

vary in the availability and sharpness of focus of recognizable substance
templates covering them. Where good substance templates are available
and known, concepts of the substances falling under them are extremely
easy to develop, for it is known in advance what kinds of determinables
will be determinate for these substances, hence what kinds of induc-
tions they can support. The discovery of substance templates requires
something like meta-inductions, although there is some evidence that
certain meta-inductions may be bypassed by human infants. Some grasp
of certain templates, some grasp of the structure of certain substance
domains, may be wired in (compare Atran 1989; Boyer 1998; Carey
1985; Gallistel et al. 1993; Gelman and Coley 1991; Keil 1979, 1989;
Markman 1989; Marler 1993, Spelke 1989, 1993). For those disciplines
systematic enough to be clearly labeled as well-developed empirical sci-
ences, the substances studied typically fall under well-focused substance
templates, or under a hierarchy of such templates. Especially, well-
founded second-order inductions of this sort would seem to underlie all
of what Kuhn labeled “normal science.” The basic principles of good
scientific induction are never found in logic alone; all inductive reason-
ing rests on a posteriori projectability judgments (compare, for example,
Boyd 1991).
Many substances do not fall under well-focused substance templates,
however. Consider, for example, the substance chair. I have argued that
this is a historical substance, but what substance template do chairs fit
under? It is clear that one would not want to project a science of fur-
niture, for example, for although there may be one or two questions
pretty certain to have answers for each kind of furniture (what was it
designed to be used for?), there are not nearly enough to delimit in ad-
vance all or most of the determinables that are relatively reliably deter-
mined for most chairs.
Aristotle thought there was a hierarchial ordering among all sub-
stances. According to the doctrine of “real definition” or of natural or-
dering by genus and differentia, substances were supposed to form a
logical tree. I think this doctrine was seriously wrong. The structure of
the domain of substances is frankly a logical mess, a mare™s nest of over-
lappings and crisscrossings.There are multitudes of entwined substances,
very very many more, surely, than we have ideas of. The ones that are
picked up by thought and by language are only those that have proper-
ties of interest to us. But that they are interesting does nothing, of
course, to make their status as substances less than fully objective. Tree
structure is good for a general classification system to have (Section

3.2), but it is not the structure of the logical space of substances nor of
most of its subspaces.
Consider butter on the one hand and human beings on the other.
Clearly there is no way to hang these on the same logical tree.They are
neither beside one another (horizontal) under some higher substance,
nor is one included in the other (vertical), nor is there some more in-
clusive substance covering them both. (Aristotle might have said they
are both subsumed under substance and under Being, but substance is not
a substance and neither is Being.) When we look within domains rather
than across them, matters are no tidier. Susan is a 1990s American
mother and a professor and a diabetic. Each of these is a rough sub-
stance category, but there is no logical tree on which they all hang.
Heated modern debates among biologists about principles of classifica-
tion (phenetics, cladistics, evolutionary classification) reflect exactly this:
There is no natural way to organize the substances that are of interest
to the zoologist or botanist into a single hierarchy.The demand that bi-
ological taxonomy should settle on a single hierarchy is of course quite
rational. A good classification system is needed for information storage
and retrieval among the various biologists. The actual systems of classi-
fication used by biologists are compromises between good classification
and respect for natural substance boundaries (compare Mayr 1981). In
the natural domain of substances there is a confusing crisscrossing, every
which way. On the other hand, wherever there exist substances that are
also substance templates, a degree of hierarchy and order is naturally im-
posed on the domain of substances.
For every substance, one can ask how many inductions, if I knew to
venture them, would yield reliable results.We also can ask how many of
these inductions I could know to venture in advance through grasp of
a good template for the substance. The latter question is the more in-
teresting to the epistemologist. The interesting question of inductive
potential concerns how many determinables you know you can find
stable values for, not how many stable properties the substance actually
has. The best substances are the ones for which there are rich, known,
substance templates, for example, the chemical elements and com-
pounds, the various living species, and also individual members of these
species, and most more ordinary individual physical objects. These are
things we know how to learn many things about without wasting time
on dozens of observations verifying the stability of each trait.
A question that has sometimes been asked by psychologists interested
in categorization concerns which level of substance categories are in-

ductively the most “fertile” to have a grasp of (see Komatsu 1998). The
question assumes, of course, some degree of hierarchial structure within
the domain to be considered. Now if one were to recognize only the
lowest level substances, say, only the individual animals or only the
species, although it is true that these have the greatest number of prop-
erties, learning about these properties would be a hopelessly inefficient
process. One would have to start all over with each individual object or
species, exploring its individual features, with no contribution from
prior knowledge of higher substances carrying substance templates, ei-
ther about its properties or its relevant determinables. It seems that there
is no particular level at which greatest “fertility” lies. It results, rather,
from an interaction between levels.

Classifying, Identifying, and the
Function of Substance Concepts


Substances, as I have described them, are whatever one can learn from
given only one or a few encounters, various skills or information that
will apply on other encounters. Further, this possibility must be
grounded in some kind of natural necessity.The function of a substance
concept is to make possible this sort of learning and use of knowledge
for a specific substance. For this, the cognizing organism must be able to
recognize the specific substance under a variety of different conditions,
as many as possible. It needs to do this, first, to grasp that the substance
it is learning about over various encounters is one and the same so that
knowledge of it can accumulate and, second, so that the accumulated
knowledge can be applied. For substance concepts to be employed in
the service of theoretical knowing “ employed for knowing that rather
than knowing how “ the substance must be represented in thought in a
univocal way, the same substance always represented as being the same.
This makes possible a stable, unequivocal, and nonredundant inner rep-
resentational system.
The ability to recognize what is objectively the same substance again
as the same despite wide variations in the faces it shows to the senses is
necessarily fallible. Although you surely have many ways of identifying
each member of your immediate family “ similarly for water and for
cats “ there will always be possible conditions under which you would
misidentify them, mistaking them for someone or something else. If a
concept is genuinely a substance concept, if its extension is really a sub-
stance, this extension is not determined by one™s fallible dispositions to

recognize it. These dispositions, given any of numerous adverse condi-
tions, would break up the extension of the substance or mix it with
other things.The extension of a substance concept is determined not by
one™s dispositions (rightly or wrongly) to recognize it but, first, by the
real extent of the substance in nature.
The purpose of this chapter is to clear the ground of certain rubble
left by classical and contemporary theories according to which disposi-
tions to identify do determine the extensions of substance concepts.1
Chapter 4 will then begin the foundation for another explanation of
how a substance concept hooks onto its extension “ onto a substance,
whole “ and how it is determined onto which substance it is hooked.2
It will begin to establish a version of what is sometimes called “direct
reference” theory.
To understand how the extension of a substance concept is deter-
mined, first we must understand more exactly what the functions of
substance concepts are, hence to what sort of things in nature they need
to be hitched in order to serve these functions. The standard view has
been that terms for kinds and stuffs correspond to capacities to classify
instances falling under these terms. I claim instead that most such terms
correspond in the first instance to abilities to identify substances, and
that they are only secondarily used as classifiers. The result of the stan-
dard view has been a thorough confusion between two quite different
kinds of functions, the functions of identifying and the functions of


Sharpening the distinction between the terms “identify” and “classify”
somewhat for expository purposes, the difference between identifying
and classifying lies both in purpose and in the psychological structure of
these acts.The purposes of classification and identification are hopelessly
entangled, for example, in the following contemporary descriptions by
psychologists of the functions of what the authors call “categorization”
and “concepts”:

1 Appendix B also concerns this matter. It concerns information theories of mental content
according to which to have a concept of Xes, one must be able to “discriminate” between
Xes and all other things.
2 The explanation will not be completely finished, however, until the full implications of the
theory of abilities to be introduced in Chapter 4 have been drawn out in Chapters 13 and

Categorization . . . is a means of simplifying the environment, of reducing the
load on memory, and of helping us to store and retrieve information efficiently.
(Markman 1989, p. 11)

Without concepts, mental life would be chaotic. If we perceived each entity as
unique, we would be overwhelmed by the sheer diversity of what we experi-
ence and unable to remember more than a minute fraction of what we en-
counter. And if each individual entity needed a distinct name, our language
would be staggeringly complex and communication virtually impossible.”
(Smith and Medin 1981, p. 1)

. . . concepts are used to classify . . . if you know nothing about a novel object
but are told it is an instance of X, you can infer that the object has all or many
of X™s properties. . . . (Smith and Medin 1981, p. 8)

A good classification system aids efficient information storage, re-
trieval, and transfer of information, or efficient storage and retrieval of
the objects classified, or, in a different but related way, efficient com-
munication. It aids the efficient storage and retrieval of what we al-
ready know when we use dictionaries, encyclopedias, telephone
books, guide books, filing systems, classification systems in libraries,
and so forth. It aids in putting objects away where they can be found
again in grocery stores, hardware stores, museums, home workshops,
and again, libraries. It aids in communication in the following way.
Shared classification systems allow one person efficiently to convey
enough information about a thing for another to retrieve it, either lit-
erally or from memory. For example, saying it is “the red book on my
table in my study” that I want you to bring will get me what I want
only if we share a way of classifying things into those that are studies
and those that are not, those that are mine and those that are not, and
into books and nonbooks, red things and nonred things, tables and
nontables. Then I can swiftly convey enough for you to retrieve the
object I intend.
These being its principle uses, an ideal general classification system,
designed to cover a general domain, will draw sharp lines around the
classes it contains, so that each member in its domain falls determinately
either in or out of each class. General domain classification systems are
used, for example, by libraries and grocery stores. These lines need not
cut between items in all possible worlds, however, but only in the actual
domain where the classification system is to be used. Classification for
purposes of communication, on the other hand, does not correspond to

a single general domain. The domains that are involved in ordinary in-
formal communication typically are severely restricted by context, vary-
ing radically from one speaker-hearer pair and from one occasion to an-
other. For this reason, words whose natural extensions have very vague
boundaries can still be used in specific communicative contexts to clas-
sify objects precisely. I refer to what I want simply as “red,” but given
the books on my table it is clear enough what object I want, even
though the entire domain of red things shades off gradually into pink
things, purple things, orange things, and so forth. If both my hearer and
I know my intended local domain and I choose my words well with
reference to it, I can often use very vague words to effect accurate
hearer classifications in context. It remains true, however, that clear
boundaries between classes relative to the actual members in the do-
main of its use is helpful for all of the functions of classification.
An ideal general classification system also has a tree or a grid struc-
ture, so that each item can be located within it by answering a determi-
nate set of questions in order. This assures that each member of the rel-
evant domain has one and only one location within the classification
system, hence that it can be efficiently put away and retrieved. This is
true also for classification used for communication. In asking for “the red
book on the table in my study,” each of my descriptive words, used here
as classifiers, partitions my intended domain into two classes in such a
way that, taken together, they separate off just the intended (unit) class.
The initial data for a classification task must include a specification of
each property of each object to be classified that is relevant to its clas-
sification. It must be possible to answer each of the questions determin-
ing its classification. A librarian would not try to classify a book, for ex-
ample, without carefully examining its contents, and to classify an object
as a red book on the table in my study you must know it is red, a book,
on a table, in a study, and that the study is mine “ every one of these.
On the other hand, most of the myriad properties of any object will not
be relevant to a given classification task. Librarians don™t need to note
the colors, numbers of pages, numbers of illustrations, publishers, type
fonts or, usually, the shapes and sizes of the books they classify. But the
properties that define the classes in the classification system being used
obviously do need to be determined, whether by observation or by in-
ference, as either present or absent for each instance of a domain prior
to classifying it. Consonant with this, in classical “categorization” exper-
iments in psychology, since these are usually taken to be studies of clas-

sification, all properties of each “stimulus” and each “test item” consid-
ered to be relevant to the classification to be learned are clearly exhib-
ited to the learner.
Now consider Smith and Medin™s claim quoted above that “. . . con-
cepts are used to classify . . . if you know nothing about a novel object
but are told it is an instance of X, you can infer that the object has all
or many of X™s properties. . . .” Suppose that the extension of “X” is, as
suggested, merely a class and not also a substance. In that case, you will
not have to do any inferring in order to know that the object “has all or
most of X™s properties.” For in that case, the properties of the object
must include all of those properties used to determine that it falls in
class X, and no other properties will be inferable in a grounded way.
This is because grounded inferences from properties of some members
of a class to other members are possible only if the extension of the
class is a substance. Being a substance is the same thing as being some-
thing that grounds inferences of this kind; if no substance, then no
grounded inference. Concepts that merely classify, and do not also iden-
tify substances, contain exactly as much information as is analytically
put into them, no more and no less. If without concepts of this kind we
would be “unable to remember more than a minute fraction of what
we encounter,” we would be just as unable with them. Use of words
that are shorthand for strings of memorized properties may make clas-
sification and communication more efficient, but it cannot make mem-
ory or thought more efficient.
Similarly, when a classification system is used for storing away infor-
mation in encyclopedias or libraries, it is not the classification system
that contains the information. The information is in the encyclopedia
entries and in the books, not in the classification system. In order to re-
trieve the information, one must first find the right entries or retrieve
the right books. And it is exactly the same when I use words to classify
what I am talking about. By classifying my subject matter as a “red book
on the table in my study,” I hand you tools with which to circumscribe
that object, but if you are to retrieve any more information about it
than is already contained in my classification, you will either have to go
and find the extension itself and inspect it, or find it in your memory
so as to retrieve things you already happen to know about it. (On the
other hand, exactly because the word “book” is not merely a classifier
but corresponds to a rough substance, you also may retrieve from mem-
ory what you know about books generally and apply it to this case.)


Contrast now the functions of reidentifying. Reidentifying is required
not primarily for information storage, retrieval, and transfer, but for in-
formation acquisition and information use. Rather than knowing its rel-
evant properties in advance, as when one classifies something, one iden-
tifies a substance in order to come to know its properties. Identifying is
necessary in order to collect together over time knowledge of a thing™s
properties, hence in order to know its properties on particular occa-
sions, since many of these properties are manifest on encounter only
some of the time. Identifying is also necessary in order to apply one™s
knowledge of things. One applies one™s knowledge by managing to rec-
ognize a substance on the basis of whatever properties do happen to be
currently manifest, and then applying one™s prior knowledge of others of
its properties, properties not currently observed.
Notice first that these functions do not require the substance recog-
nized to have sharp boundaries. Grasp of substances very often affords
knowledge that is not invariant over the substance™s entire extension. If
the edges of the substance are vague, the variance is more marked to-
ward these edges. Well-fashioned substance concepts, well-fashioned
abilities to exploit substances as objects of knowledge, will include the
ability to recognize a difference between more central and less central
areas in the extent of the substance, and to portion out degrees of re-
liance on accumulated knowledge of the substance accordingly. When,
as is often the case, the boundaries of substances really are vague, obvi-
ously there is nothing, other than taking this into account, that the or-
ganism can do about it. Artificially imposing precision will not help.
By contrast, ideal general classification systems, I have said, are more
efficient if precise relative to the entities actually in their domains.
Where substance boundaries are vague in nature, the purposes of clas-
sification are sometimes served by drawing artificial boundaries around
the extensions of these substances. For certain classificatory purposes,
for example, what counts as war and who counts as a member of
the working class or as a full-time student may be quite sharply but
artificially defined.
Notice, second, that substances are not generally organized into tree
or grid structures. This was argued in Section 2.6.
Notice, third, that unlike a task of classifying, the task of identifying
a substance doesn™t require that any one particular set of the substance™s

<< . .

( 29)

. . >>

Copyright Design by: Sunlight webdesign