ArticlePDF Available

Conceptual Spaces as a Framework for Knowledge Representation


Abstract and Figures

The dominating models of information processes have been based on symbolic representations of information and knowledge. During the last decades, a variety of non-symbolic models have been proposed as superior. The prime examples of models within the non-symbolic approach are neural networks. However, to a large extent they lack a higher-level theory of representation. In this paper, conceptual spaces are suggested as an appropriate framework for non- symbolic models. Conceptual spaces consist of a number of 'quality dimensions' that often are derived from perceptual mechanisms. It will be outlined how conceptual spaces can represent various kind of information and how they can be used to describe concept learning. The connections to prototype theory will also be presented.
Content may be subject to copyright.
2004 Mind and Matter Vol. 2(2), pp. 9–27
Conceptual Spaces as a Framework
for Knowledge Representation
Peter G¨ardenfors
Department of Cognitive Science
Lund University, Sweden
The dominating models of information processes have been based
on symbolic representations of information and knowledge. Dur-
ing the last decades, a variety of non-symbolic models have been
proposed as superior. The prime examples of models within the
non-symbolic approach are neural networks. However, to a large
extent they lack a higher-level theory of representation. In this pa-
per, conceptual spaces are suggested as an appropriate framework
for non-symbolic models. Conceptual spaces consist of a number of
“quality dimensions” that often are derived from perceptual mech-
anisms. It will be outlined how conceptual spaces can represent
various kind of information and how they can be used to describe
concept learning. The connections to prototype theory will also be
1. The Problem of Modeling Representations
Cognitive science has two overarching goals. One is explanatory:By
studying the cognitive activities of humans and other animals, one formu-
lates theories of different aspects of cognition. The theories are tested by
experiments or by computer simulations. The other goal is constructive:
By building artifacts like chess-playing programs, robots, animats, etc.,
one attempts to construct systems that can accomplish various cognitive
tasks. For both kinds of goals, a key problem is how the representations
used by the cognitive system are to be modeled in an appropriate way.
In cognitive science, there are currently two dominating approaches to
the problem of modeling representations. The symbolic approach starts
from the assumption that cognitive systems should be modeled by Turing
machines. On this view, cognition is seen as essentially involving symbol
manipulation. The second approach is associationism, where associations
between different kinds of information elements carry the main burden of
representation. Connectionism is a special case of associationism, which
models associations by artificial neuron networks. Both the symbolic and
the associationistic approaches have their advantages and disadvantages.
They are often presented as competing paradigms, but since they attack
10 ardenfors
cognitive problems on different levels, I shall argue later that they should
rather be seen as complementary methodologies.
However, there are aspects of cognitive phenomena for which neither
symbolic representation nor connectionism seem to offer appropriate mod-
eling tools. In this article, I will advocate a third form of representing
information that is based on using geometrical structures rather than
symbols or connections between neurons. Using these structures similar-
ity relations can be modeled in a natural way. The notion of similarity is
crucial for the understanding of many cognitive phenomena. I shall call
my way of representing information the conceptual form since I believe
that the essential aspects of concept formation are best described using
this kind of representations.
Again, conceptual representations should not be seen as competing
with symbolic or associationist (connectivist) representations. Rather,
the three kinds can be seen as three levels of representations of cognition
with different scales of resolution.
I shall outline a theory of conceptual spaces as a particular framework
for representing information on the conceptual level. A conceptual space
is built up from geometrical representations based on a number of qual-
ity dimensions. The emphasis of the theory will be on the constructive
side of cognitive science. However, I believe that it also can explain sev-
eral aspects of what is known about representations in various biological
2. Quality Dimensions
One notion that is severely downplayed in symbolic representations
is that of similarity. I submit that judgments of similarity are central
for a large number of cognitive processes. Judgments of similarity reveal
the dimensions of our perceptions and their structures. For many kinds
of dimensions it will be possible to talk about distances. The general
assumption is that the smaller the distances is between the representations
of two objects, the more similar they are. In this way, the similarity of
two objects can be defined via the distance between their representing
points in the space. Thus conceptual spaces provide us with a natural
way of representing similarities. In general, the epistemological role of
the conceptual spaces is to serve as a tool in sorting out various relations
between perceptions.
As introductory examples of quality dimensions one can mention tem-
perature, weight, brightness, pitch and the three ordinary spatial dimen-
sions height, width and depth. I have chosen these examples because
they are closely connected to what is produced by our sensory receptors
(Schiffman 1982). The spatial dimensions height, width and depth as well
Conceptual Spaces 11
as brightness are perceived by the visual sensory system, pitch by the
auditory system, temperature by thermal sensors and weight, finally, by
the kinesthetic sensors. There are additional quality dimensions that are
of an abstract, non-sensory character.
The primary function of the quality dimensions is to represent various
“qualities” of objects.1They correspond to the different ways stimuli are
judged to be similar or different. In most cases, judgments of similar-
ity and difference generate an ordering relation of stimuli (Clark 1993, p.
114). For example, one can judge tones by their pitch which will generate
and ordering of the perceptions. The dimensions form the “framework”
used to assign properties to objects and to specify relations between them.
The coordinates of a point within a conceptual space represent particu-
lar instances of each dimension, for example a particular temperature, a
particular weight, etc.
The quality dimensions are taken to be independent of symbolic rep-
resentations in the sense that we and other animals can represent the
qualities of objects, for example when planning an action, without pre-
suming an internal language or another symbolic system in which these
qualities are expressed. In other words, the dimensions are the building
blocks of representations on the conceptual level.
When the explanatory aim of cognitive science is in focus, the qual-
ity dimensions should be seen as theoretical entities used as a modeling
factor in describing cognitive activities of organisms. When construct-
ing artificial systems, the dimensions function as the framework for the
representations used by the systems.
The notion of a dimension should be understood literally. It is as-
sumed that each of the quality dimensions is endowed with certain geo-
metrical structures (in some cases they are topological or orderings). As
a first example to illustrate such a structure, Fig. 1 shows the dimension
of “weight” which is one-dimensional with a zero point, and thus isomor-
phic to the half-line of non-negative numbers. A basic constraint on this
dimension that is commonly made in science is that there are no negative
Figure 1: The weight dimension.
1In traditional philosophy, following Locke, a distinction between “primary” and
“secondary” qualities is often made. This distinction corresponds roughly to the dis-
tinction between “scientific” and “phenomenal” dimensions to be presented in the
following section.
2However, it is interesting to note (cf. Kuhn 1970) that during a period of phlogiston
chemistry, scientists were considering negative weights in order to evade some of the
anomalies of the theory.
12 ardenfors
In previous writings on conceptual spaces, I have used the example
of the perceptual color space to illustrate a more structured set of qual-
ity dimensions (G¨ardenfors 1990, 1991, 2000). However, we can also find
related spatial structures for other sensory qualities. For example, con-
sider the quality dimension of pitch, which is basically a continuous one-
dimensional structure going from low to high tones. This representation
is directly connected to the neurophysiology of pitch perception.
Apart from the basic frequency dimension of tones, it is possible to
identify some further structure in the mental representation of tones. Nat-
ural tones are not simple sinusoidal tones of only one frequency, but are
constituted of a number of higher harmonics. The timbre of a tone, which
is a phenomenal dimension, is determined by the relative strength of the
higher harmonics of the fundamental frequency of the tone. An interesting
perceptual phenomenon is “the case of the missing fundamental”. If the
fundamental frequency is removed by artificial methods from a complex
physical tone, the phenomenal pitch of the tone is still perceived as that
corresponding to the removed fundamental.3Apparently, the fundamen-
tal frequency is not indispensable for pitch perception, but the perceived
pitch is determined by a combination of the lower harmonics.
Thus, the harmonics of a tone are essential for how it is perceived. This
entails that tones which share a number of harmonics will be perceived
to be similar. The tone that shares the most harmonics with a given
tone is its octave, the second most similar is the fifth, the third most
similar is the fourth and so on. This additional “geometrical” structure
on the pitch dimension, which can be derived from the wave structure of
tones, provides the foundational explanation for the perception of musical
For another example of sensory space representations let me only men-
tion that the human perception of taste appears to be generated from
four distinct types of receptors: saline, sour, sweet, and bitter. Thus the
quality space representing tastes could be described as a 4-dimensional
space. One such model was put forward by Henning (1961), who sug-
gested that phenomenal gustatory space could be described as a tetrahe-
dron (see Fig. 2). Actually, Henning speculated that any taste could be
described as a mixture of only three primaries. This means that any taste
can be represented as a point on one of the planes of the tetrahedron, so
that no taste is mapped onto the interior.
However, there are other models which propose more than four fun-
damental tastes.5The best model of the phenomenal gustatory space
remains to be established. This will involve sophisticated psychophysical
3See e.g. Gabrielsson (1981), pp. 20–21.
4For some further discussion of the structure of musical space see G¨ardenfors (1988),
Sections 7–9.
5See Schiffman (1982), Chap. 9, for an exposition of some such theories.
Conceptual Spaces 13
Figure 2. Henning’s taste tetrahedron.
measurement techniques. Suffice it to say that the gustatory space quite
clearly has some non-trivial geometrical structure. For instance, we can
meaningfully claim that the taste of a walnut is closer to the taste of a
hazelnut than to the taste of popcorn in the same way as we can say that
the color orange is closer to yellow than to blue.
It should be noted that some quality “dimensions” have only a discrete
structure, that is, they merely divide objects into disjoint classes. Two
examples are classifications of biological species and kinship relations in
a human society. One example of a phylogenetic tree of the kind found in
biology is shown in Fig. 3. Here the nodes represents different species in
the evolution of, for example, a family of organisms, where nodes higher
up in the tree represent evolutionarily older (extinct) species.
Figure 3. Phylogenetic tree.
The distance between two nodes can be measured by the length of
the path that connects them. This means that even for discrete dimen-
sions one can distinguish a rudimentary geometrical structure. For exam-
ple, in the phylogenetic classification of animals that mirrors evolutionary
branchings it is meaningful to say that rats and whales are more closely
related than whales and fish.
14 ardenfors
3. Phenomenal and Scientific Interpretations
of Dimensions
In order to separate different uses of quality dimensions it is impor-
tant to introduce a distinction between a phenomenal (or psychological)
and a scientific interpretation. The phenomenal interpretation concerns
the cognitive structures (perceptions, memories, etc.) of humans or other
organisms. The scientific interpretation, on the other hand, treats dimen-
sions as a part of a scientific theory.
The distinction is relevant in relation to the two goals of cognitive
science presented above. When the dimensions are seen as cognitive en-
tities, that is, when the goal is to explain natural cognitive processes,
their geometrical structure should not be determined by scientific theo-
ries which attempt at giving a “realistic” description of the world, but by
psychophysical measurements which determine the structure of how our
perceptions are represented. Furthermore, when it comes to providing a
semantics for a natural language, the phenomenal interpretations of the
quality dimensions are in focus.
On the other hand, when we are constructing an artificial system, the
function of sensors, effectors and various control devices are in general
described in terms of scientifically modeled dimensions. For example, the
input variables of a robot may be a small number of physically measured
magnitudes, like brightness, delay of a radar echo, or the pressure from
a mechanical grip. With the aid of the programmed goals of the robot,
these variables can then be transformed into a number of physical output
magnitudes as, for example, the voltages of the motors controlling the left
and the right wheels.
To give an example of the distinction, consider color. The distinction
introduced here is supported by Gallistel (1990, p. 518–519) who writes:
The facts about color vision suggest how deeply the nervous system
may be committed to representing stimuli as points in descriptive
spaces of modest dimensionality. It does this even for spectral com-
positions, which does not lend itself to such a representation. The
resulting lack of correspondence between the psychological repre-
sentation of spectral composition and spectral composition itself is
a source of confusion and misunderstanding in scientific discussions
of color. Scientists persist in referring to the physical characteristics
of the stimulus and to the tuning characteristics of the transducers
(the cones) as if psychological color terms like red,green,andblue
had some straightforward translation into physical reality, when in
fact they do not.
Gallistel’s warning against confusion and misunderstanding of the two
types of representation should be taken seriously. It is very easy to con-
found what science says about the characteristics of reality and our per-
Conceptual Spaces 15
ception of it. In this article, it is the phenomenal representation that will
be in focus.
Aconceptual space can now be defined as a collection of one or more
quality dimensions. However, the dimensions of a conceptual space should
not be seen as totally independent entities, but they are correlated in
various ways since the properties of the objects modeled in the space
covary. For example, the ripeness and the color dimensions covary in the
space of fruits.
4. On the Origins of Quality Dimensions
In the previous sections, I have given some examples of quality dimen-
sions from different kinds of domains. There seem to be different types of
dimensions, so a warranted question is: Where do the dimensions come
from? I do not believe there is a unique answer to this question. In this
section, I will try to trace the origins of the different kinds of quality
Firstly, some of the quality dimensions seem to be innate or devel-
oped very early in life. They are to some extent hardwired in our nervous
system, as for example the sensory dimensions presented above. This
probably also applies to our representations of ordinary space. Since do-
mains of this kind are obviously extremely important for basic activities
like finding food, avoiding danger, and getting around in the environment
there is evolutionary justification for the innateness assumption. Humans
and other animals who did not have a sufficiently adequate representa-
tion of the spatial structure of the external world were disadvantaged by
natural selection.
The brain of humans and animals contains topographic areas mapping
different kinds of sense modalities onto spatial areas. The structuring
principles of these mappings are basically innate, even if the fine tuning
is established during the development of the human or animal. The same
principles seem to govern most of the animal kingdom. Gallistel (1990,
p. 105) argues:
[...] the intuitive belief that the cognitive maps of “lower” ani-
mals are weaker than our own is not well-founded. They may be
impoverished relative to our own (have less on them) but they are
not weaker in their formal characteristics. There is experimental
evidence that even insect maps are metric maps.
Quine notes that something like innate quality dimensions is needed
to make learning possible (Quine 1969, p. 123):
Without some such prior spacing of qualities, we could never ac-
quire a habit; all stimuli would be equally alike and equally differ-
16 ardenfors
ent. These spacings of qualities, on the part of men and other ani-
mals, can be explored and mapped in the laboratory by experiments
in conditioning and extinction. Needed as they are for all learning,
these distinctive spacings cannot themselves all be learned; some
must be innate.
However, once the process has started, new dimensions can be added
by the learning process.6One kind of examples comes from studies of chil-
dren’s cognitive development. Two-year-olds can represent whole objects,
but they cannot reason about the dimensions of the objects.
Learning new concepts is, consequently, often connected with expand-
ing one’s conceptual space with new quality dimensions. For example,
consider the (phenomenal) dimension of volume. The experiments con-
cerning “conservation” performed by Piaget and his followers indicate that
small children have no separate mental dimension of volume; they confuse
the volume of a liquid with the height of the liquid in its container. It
is only at about an age of five years that they learn to represent the two
dimensions separately. Similarly, three- and four-year-olds confuse high
with tall,big with bright, etc (Carey 1978). Smith (1989, p. 146–147)
argues that
working out a system of perceptual dimension, a system of kinds
of similarities, may be one of the major intellectual achievements
of early childhood. [. . . ] The basic developmental notion is one of
differentiation, from global syncretic classes of perceptual resem-
blance and magnitude to dimensionally specific kinds of sameness
and magnitude.
Still other dimensions may be culturally dependent.7Take “time” as
an example: In some cultures time is conceived to be circular – the world
keeps returning to the same point in time and the same events occur
over and over again; and in other cultures it is hardly meaningful at all
to speak of time as a dimension. A sophisticated time dimension, with
the full metric structure, is needed for advanced forms of planning and
coordination with other individuals, but is not necessary for the most
basic activities of an organism. As a matter of fact, the standard Western
conception of time is a comparatively recent phenomenon (see Toulmin
and Goodfield 1965).
The examples given here indicate that many of the quality dimensions
of human conceptual spaces are not directly generated from sensory in-
puts. This is even clearer when we use concepts based on the functions
6It must be noted that it is impossible to draw a sharp distinction between innate
and learned quality dimensions, since many sensory dimensions are structurally pre-
pared in the neural tissue at birth, but require exposure to sensory experiences to lay
out the exact geometrical structure of the mapping.
7I do not claim that my typology of the origins of quality dimensions is exclusive,
since, in a sense, all culturally dependent dimensions are also learned.
Conceptual Spaces 17
of artifacts or the social roles of people in a society. Even if we do not
know much about the geometrical structures of these dimensions, it is
quite obvious that there is some non-trivial such structure. This has been
argued by Marr and Vaina (1982) and Vaina (1983), who give an analysis
of functional representation where functions of an object are determined
in terms of the actions it allows.
Culture, in the form of interaction between people, may in itself gen-
erate constraints on conceptual spaces. For example, Freyd (1983) puts
forward the intriguing proposal that conceptual spaces may evolve as a
representational form in a community just because people have to share
knowledge (Freyd 1983, pp. 193–194):
There have been a number of different approaches towards analyz-
ing the structures in semantic domains, but what these approaches
have in common is the goal of discovering constraints on knowledge
representation. I argue that the structures the different semantic
analyses uncover may stem from shareability constraints on knowl-
edge representation. [...] So, if a set of terms can be shown to
behave as if they are represented in a three-dimensional space, one
inference that is often made is that there is both some psychological
reality to the spatial reality (or some formally equivalent formula-
tion) and some innate necessity to it. But it might be that the
structural properties of the knowledge domain came about because
such structural properties provide for the most efficient sharing of
concepts. That is, we cannot be sure that the regularities tell us
anything about how the brain can represent things, or even “prefer”
to, if it didn’t have to share concepts with other brains.
Here Freyd hints at an econo mic explanation of why we have conceptual
spaces: they facilitate the sharing of knowledge.
Finally, some quality dimensions are introduced by science. Witness,
for example, Newton’s distinction between weight and mass, which is of
crucial importance for the development of his mechanics, but which hardly
has any correspondence in human perception. To the extent we have men-
tal representations of the masses of objects in distinction to their weights,
these are not given by the senses but have to be learned by adopting the
conceptual space of Newtonian mechanics in our representations.
The most drastic changes in science occur when the underlying con-
ceptual space is changed. I believe that most of the “paradigm shifts” dis-
cussed by Kuhn (1970) can be understood as shifts of conceptual spaces.
I do not see any principal difference between this kind of change and
the change involved in the development of a child’s conceptual space.
Introducing the distinction between “height” and “volume” is the same
kind of phenomenon as when Newton introduced the distinction between
“weight” and “mass”. That distinction is nowadays ubiquitous in physics,
even though there is only scant sensory support for it.
18 ardenfors
The conceptual space of Newtonian mechanics is, of course, based on
scientific (theoretical) quality dimensions and not on phenomenal (psy-
chological) dimensions. The quality dimensions of this theory are ordi-
nary space (3-D Euclidean), time (isomorphic to the real numbers), mass
(isomorphic to the non-negative real numbers), and force (3-D Euclidean
space). Once a particle has been assigned a value for these eight dimen-
sions, it is fully described as far as Newtonian mechanics is concerned. In
this theory, an object is thus represented as a point in an 8-dimensional
5. Concept Formation Described with the Aid
of Conceptual Spaces
In more abstract terms, a conceptual space Sis established by a class
of quality dimensions D1, ..., Dn. A point in Sis represented by a vector
1, ..., dn>with one index for each dimension. Each of the
dimensions is endowed with a certain topological or metrical structure.
The purpose of this section is to show how conceptual spaces can be used
to model concepts.
A first rough idea is to describe a concept as a region of a conceptual
space S, where “region” should be understood as a spatial notion deter-
mined by the topology and metric of S. For example, the point in the
time dimension representing “now” divides this dimension, and thus the
space of vectors, into two regions corresponding to the concepts “past”
and “future”. But the proposal suffers from a lack of precision as regards
the notion of a “region”. A more precise and powerful idea is the following
criterion where the topological characteristics of the quality dimensions
are utilized to introduce a spatial structure on concepts:
Criterion P:
Anatural concept is a convex region of a conceptual space.
Aconvex region is characterized by the criterion that for every pair
of points v1and v2in the region all points in between v1and v2are also
in the region. The motivation for the criterion is that if some objects
which are located at v1and v2in relation to some quality dimension (or
several dimensions) both are examples of a concept C, then any object
that is located between v1and v2on the quality dimension(s) will also be
an example of C. I shall argue later that this criterion is psychologically
realistic. It presumes that the notion of betweennes s is meaningful for
the relevant quality dimensions. This is a rather weak assumption which
demands very little of the underlying topological structure.
Most concepts expressed by simple words in natural languages are
natural concepts in the sense specified here. For instance, I conjecture
that all color terms in natural languages express natural concepts with
Conceptual Spaces 19
respect to the psychological representation of the three color dimensions.
In other words, the conjecture predicts that if some object o1is described
by the color term Cin a given language and another object o2is also said
to have color C, then any object o3with a color that lies between the color
of o1and that of o2will also be described by the color term C. It is well-
known that different languages carve up the color circle in different ways,
but all carvings seems to be done in terms of convex sets. Strong support
for this conjecture can be found in Berlin and Kay (1969), although they
do not treat color terms in general but concentrate on basic color terms.
On the other hand, the reference of an artificial color term like “grue”
(Goodman 1955) will not be a convex region in the ordinary conceptual
space and thus it is not a natural concept according to Criterion P.8
Another illustration of how the convexity of regions determines con-
cepts and categorizations is the phonetic identification of vowels in various
languages. According to phonetic theory, what determines a vowel are the
relations between the basic frequency of the sound and its formants (higher
frequencies that are present at the same time). In general, the first two
formants F1and F2are sufficient to identify a vowel. This means that the
coordinates of two-dimensional space spanned by F1and F2(in relation
to a fixed basic pitch F0) can be used as a fairly accurate description of a
vowel. Fairbanks and Grubb (1961) investigated how people produce and
recognize vowels in “General American” speech.
Figure 4 summarizes some of their findings. The scale of the abscissa
and ordinate are the logarithm of the frequencies of F1and F2(the basic
frequency of the vowels was 130 Hz). As can be seen from the diagram,
the preferred, identified and self-approved examples of different vowels
form convex subregions of the space determined by F1and F2with the
given scales.9As in the case of color terms, different languages carve up
the phonetic space in different ways (the number of vowels identified in
different languages varies considerably), but I conjecture again that each
vowel in a language will correspond to a convex region of the formant
An important thing to note in this example is that identifying F1and
F2as the relevant dimensions for vowel formation is a phonetic discovery.
We had the concepts of vowels already before this discovery, but the spatial
analysis makes it possible for us to understand several features of the
classifications of vowels in different languages.
Criterion Pprovides an account of concepts that is independent of
8For an extended analysis of this example see G¨ardenfors (1989).
9Aself-approved vowel is one that was produced by the speaker and later approved
of as an example of the intended kind. An identified sample of a vowel is one that
was correctly identified by 75% of the observers. The preferred samples of a vowel
are those which are “the most representative samples from among the most readily
identified samples” (Fairbanks and Grubb 1961, p. 210).
20 ardenfors
500200 250 500 IK
Self - approved
F = F
Figure 4: Frequency ranges of different vowels in the two-dimensional
space generated by the first two formants (from Fairbanks and Grubb
both possible worlds and individuals and it satisfies Stalnaker’s desidera-
tum that a concept “... must be not just a rule for grouping individuals,
but a feature of individuals in virtue of which they may be grouped”
(Stalnaker 1981, p. 347). However it should be emphasized that I only
view the criterion as a necess ary but perhaps not sufficient condition on
a natural concept. The criterion delimits the class of concepts that are
useful for cognitive purposes, but it may not be sufficiently restrictive.
6. Relations to Prototype Theory
Describing concepts as convex regions of conceptual spaces fits very
well with the so called prototype theory of categorization developed by
Rosch and her collaborators (Rosch 1975, 1978, Mervis and Rosch 1981,
Lakoff 1987). The main idea of prototype theory is that within a category
of objects, like those instantiating a concept, certain members are judged
to be more representative of the category than others. For example, robins
are judged to be more representative of the category “bird” than are
ravens, penguins and emus; and desk chairs are more typical instances of
the category “chair” than rocking chairs, deck-chairs, and beanbag chairs.
The most representative members of a category are called prototypical
Conceptual Spaces 21
members. It is well-known that some concepts, like “red” and “bald”
have no sharp boundaries and for these it is perhaps not surprising that
one finds prototypical effects. However, these effects have been found for
most concepts including those with comparatively clear boundaries like
“bird” and “chair”.
In traditional philosophical analyses of concepts, based on truth-funct-
ions or possible worlds it is very difficult to explain such prototype effects
(see G¨ardenfors 1991). Either an object is a member of the class assigned
to a concept (relative to a given possible world) or it is not and all members
of the class have equal status as category members. Rosch’s research
has been aimed at showing asymmetries among category members and
asymmetric structures within categories. Since the traditional definition
of a concept neither predicts nor explains such asymmetries, something
else must be going on.
In contrast, if concepts are described as convex regions of a conceptual
space, prototype effects are indeed to be expected. In a convex region one
can describe positions as being more or less centra l. For example, if color
concepts are identified with convex subsets of the color space, the central
points of these regions would be the most prototypical examples of the
color. In a series of experiments, Rosch has been able to demonstrate
the psychological reality of such “focal” colors. For another illustration
we can return to the categorization of vowels presented in the previous
section. Here the structure of the subjects’ different kinds of responses
shows clear prototype effects.
For more complex categories like “bird” it is perhaps more difficult
to describe the underlying conceptual space. However, if something like
the analysis of shapes by Marr and Nishihara (1978) is adopted, we can
begin to see how such a space would appear.10 Their scheme for describing
biological forms uses hierarchies of cylinder-like modeling primitives. Each
cylinder is described by two coordinates (length and width). Cylinders
are combined by determining the angle between the dominating cylinder
and the added one (two polar coordinates) and the position of the added
cylinder in relation to the dominating one (two coordinates). The details
of the representation are not important in the present context, but it is
worth noting that on each level of the hierarchy an object is described by
a comparatively small number of coordinates based on lengths and angles.
Thus the object can be identified as a hierarchically structured vector in a
(higher order) conceptual space. Figure 5 provides an illustration of this
hierarchical structure.
It should be noted that even if different members of a category are
judged to be more or less prototypical, it does not follow that some of the
10This analysis is expanded in Marr (1982), Chap. 5. A related model, together with
some psychological grounding, is presented by Biederman (1987).
22 ardenfors
Figure 5: Representing shapes by cylinders (from Marr and Nishihara 1978).
existing objects must represent “the prototype”. If a concept is viewed
as a convex region of a conceptual space this is easily explained, since
the central member of the region (if unique) is a possible individual in
the sense discussed above (if all its dimensions are specified) but need
not be among the existing members of the category. Such a prototype
point in the region need not be completely described as an individual, but
is normally represented as a partial vector, where only the values of the
dimensions that are relevant to the concept have been determined. For
example, the general shape of the prototypical bird would be included in
the vector, but its color or age presumably would not.
It is possible to argue in the converse direction, too, and show that,
if prototype theory is adopted, then the representation of concepts as
convex regions is to be expected. Assume that some quality dimensions
of a conceptual space are given, for example the dimensions of color space,
Conceptual Spaces 23
and that we want to partition it into a number of categories, for example
color categories. If we start from a set of prototypes p1, ..., pnof the
categories, for example the focal colors, then these should be the central
points in the categories they represent. One way of using this information
is to assume that for every point pin the space one can measure the
distance from pto each of the pi’s. If we now stipulate that pbelongs to
the same category as the closest prototype pi, it can be shown that this
rule will generate a partitioning of the space that consists of convex areas
(convexity is here defined in terms of an assumed distance measure). This
is the so-called Voronoi tessellation, a two-dimensional example of which
is illustrated in Figure 6.
p1 p2 p3
p4 p5
Figure 6: Voronoi tessellation of the plane into convex sets.
Thus, assuming that a metric is defined on the subspace that is sub-
ject to categorization, a set of prototypes will by this method generate a
unique partitioning of the subspace into convex regions. Hence there is an
intimate link between prototype theory and the description of concepts
as convex regions in a conceptual space.
As a concrete instance of this technique, Petitot (1989) applied Voronoi
categorization to explain some aspects related to the categorical percep-
tion of phonemes. In particular, he analyzed the relations between the so
called stop consonants /b/, /d/, /g/, /p/, /t/, /k/. The relations between
these consonants are expressed with the aid of two dimensions: one is the
voiced-unvoiced dimension, the other is the labial-dental-velar dimension
which relates to the place of articulation of the consonant. Both these di-
mension can be treated as continuous. Figure 7 shows how he represents
the boundaries between the six consonants.
As an example of the information contained in this model, Petitot
points out (Petitot 1989, p. 68):
The geometry of the system of boundaries can provide precious
information about the hierarchical relations that stop consonants
24 ardenfors
labial dental velar
/p/ /k/
/b/ /g/
Figure 7: A Voronoi model of the boundaries of stop consonants (from
Petitot (1989), p. 69).
maintain with each other. The fact that in the model of Massaro
and Oden, the domains of /p/ and /d/ are adjacent, whereas those
of /b/ and /t/ are separated, indicates that the contrast between
/b/ and /t/ is much greater than that between /p/ and /d/.
7. Conclusions
The main purpose of this article is to present the core of the theory of
conceptual spaces. In this connection an important question is: what kind
of theory is the theory of conceptual spaces? Is it an empirical, normative,
computational, psychological, neuroscientific, or linguistic theory?
As was stated in Sect. 1, cognitive science has two predominant goals:
to explain cognitive phenomena and to construct artificial systems that
can solve various cognitive tasks. The theory of conceptual spaces is
presented as a framework for representing information. It should be seen
as a theory that complements the symbolic and the connectionist models
and forms a bridge between these forms of representation.
The primary aim is to use the theory of conceptual spaces in con-
structive tasks. In previous work, I have shown how it can be used in
computational models of concept formation (G¨ardenfors 1992) and induc-
tion (G¨ardenfors 1990, 1993) and that it is useful for representing the
meanings of different kinds of linguistic expressions in a computational
approach to semantics.
The borderline between constructive and explanatory uses of concep-
tual spaces is not sharp. When, for example, constructing the represen-
tational world of a robot, it is often worthwhile to take lessons from how
Conceptual Spaces 25
biology has solved the problems in the brains of humans and other animals.
Conversely, the construction of an artificial system that can successfully
solve a particular cognitive problem may provide clues to how an em-
pirical investigation of biological systems should proceed. Consequently,
there is a spiraling interaction between constructive and explanatory uses
of conceptual spaces.
This article has been asking questions about the geometry of thought.
With the aid of the notion of conceptual spaces I have provided an analysis
of concepts. A key notion is that of a natural concept which is defined
in terms of well-behaved regions of conceptual spaces – a definition that
crucially involves the geometrical structure of the various domains.
In my opinion, a conceptual level of representation should play a cen-
tral role in the cognitives sciences. After having been dominant for many
years, the symbolic approach was challenged by connectionism (which is
nowadays broadened to a wider study of dynamical systems). However,
for many purposes the symbolic level of representation is too coarse, and
the connectionist too fine-grained. In relation to the two goals of cogni-
tive science, I submit that the conceptual level will add significantly to
our explanatory capacities when it comes to understanding cognitive pro-
cesses, in particular those connected with concept formation and language
Where do we go from here? The main factor preventing a rapid de-
velopment of different applications of conceptual spaces is the lack of
knowledge about the relevant quality dimensions. It is almost only for
perceptual dimensions that psychophysical research has succeeded in iden-
tifying the underlying geometrical and topological structures (and, in rare
cases, the psychological metric). For example, we only have a very sketchy
understanding of how we perceive and conceptualize things according to
their shapes.
When the structure of the dimensions of a particular domain is discov-
ered, this often leads to fruitful research. For example, the development
of the vowel space that was presented in section 5 led to a wealth of new
results in phonetics and a deeper understanding of the speech process.
Thus, those who want to contribute to the research program should
start hunting for the hidden conceptual spaces. Even if results may not
be easily forthcoming, they are sure to have repercussions in other areas
of cognitive science as well.
Berlin B. and Kay P. (1969): Basic Color Terms: Their Universality and
Evolution, University of California Press, Berkeley.
Biederman I. (1987): Recognition-by-components: a theory of human image
understanding. Psychological Review 94, 115–147.
26 ardenfors
Clark A. (1993): Sensory Qualities, Clarendon Press, Oxford.
Fairbanks G. and Grubb P. (1961): A psychophysical investigation of vowel
formants. Jou rnal of Speech and H earing Resea rch 4, 203–219.
Freyd J. (1983): Shareability: the social psychology of epistemology. Cognitive
Science 7, 191–210.
Gabrielsson A. (1981): Music psychology – a survey of problems and current
research activities. In Basic Musical Functions and Musical Ability, Publications
of the Royal Swedish Academy of Music, No. 32, pp. 7–80.
Gallistel C.R. (1990): The Organization of Learning, MIT Press, Cambridge,
ardenfors P. (1988): Semantics, conceptual spaces and music. In Essays on the
Philosophy of Music (Acta Philosophica Fennica, vol. 43), ed. by V. Rantala,
L. Rowell and E. Tarasti, The Philosophical Society of Finland, Helsinki, pp. 9–
ardenfors P. (1990): Induction, conceptual spaces and AI. Philosophy of Sci-
ence 57, 78–95.
ardenfors P. (1991): Frameworks for properties: possible worlds vs. conceptual
spaces. In Language, Knowledge and Intentionality (Acta Philosophica Fennica,
vol. 49), ed. by L. Haaparanta, M. Kusch, and I. Niiniluoto, The Philosophical
Society of Finland, Helsinki, pp. 383–407.
ardenfors P. (1993): Induction and the evolution of conceptual spaces. In
Charles S. Peirce and the Philosophy of Science, ed. by E.C. Moore, University
of Alabama Press, Tuscaloosa, pp. 72–88.
ardenfors P. (2000): Conceptual Spaces: On the Geometry of Thought,MIT
Press, Cambridge, MA.
Goodman N. (1955): Fact, Fiction and Forecast, Harvard University Press,
Cambridge, MA.
Henning H. (1916): Die Qualit¨atenreihe des Geschmacks. Zeitschrift f¨ur Psy-
chologie und Physiologie der Sinnesorgane 74, 203–219.
Kuhn T. (1970): The Structure of Scientific Revolutions, University of Chicago
Press, Chicago.
Lakoff G. (1987): Women, Fire, and Dangerous Things, University of Chicago
Press, Chicago.
Marr D. (1982): Vision, Freeman, San Francisco.
Marr D. and Nishihara H.K. (1978): Representation and recognition of the spa-
tial organization of three-dimensional shapes. Proceedings o f the Royal Soc iety
London B 200, 269–294.
Marr D. and Vaina L. (1982): Representation and recognition of the movements
of shapes. Proceedi ngs of the Royal S ociety Lond on B 214, 501–524.
Mervis C. and Rosch E. (1981): Categorization of natural objects. Annual
Review of Psychology 32, 89–115.
Petitot J. (1989): Morphodynamics and the categorical perception of phonolog-
ical units. Theoretical Linguistics 15, 25–71.
Conceptual Spaces 27
Quine W.V.O. (1969): Natural kinds. In Ontological Relativity and Other Es-
says, Columbia University Press, New York, pp. 114–138.
Rosch E. (1975): Cognitive representations of semantic categories. Journal of
Experimental Psychology: General 104, 192–233.
Rosch E. (1978): Prototype classification and logical classification: the two
systems. In New Trends in Cognitive Representation: Challenges to Piaget’s
Theory, ed. by E. Scholnik, Lawrence Erlbaum, Hillsdale, pp. 73–86.
Schiffman H.R. (1982): Sensation and Perception, Wiley, New York.
Smith L.B. (1989): From global similarities to kinds of similarities – the con-
struction of dimensions in development. In Similarity and Analogical Reasoning,
ed. by S. Vosniadou and A. Ortony, Cambridge University Press, Cambridge,
pp. 146–178.
Stalnaker R. (1981): Antiessentialism. Midwest Studies of Philosophy 4, 343–
Toulmin S. and Goodfield J. (1965): The Discovery of Time, Penguin, Har-
Vaina L. (1983): From shapes and movements to objects and actions. Synthese
54, 3–36.
... It is utilized here as a new foundation for the application of data and information fusion. Cognitive models have two overarching goals as stated by Gärdenfors [2]; The first goal is explanatory: by studying the cognitive activities of humans and other animals one can formulate theories about different aspects of cognition. The second goal is constructive: by building artifacts one attempts to construct systems that can accomplish various cognitive tasks. ...
... The second kind of model for cognitive thinking is the associative model, where associations between different types of information elements are the means of representation [2]. This form of cognitive model works on the principle of association. ...
... Take for example the dimension of brightness within the color domain; Figure 3a shows an example of what this dimension would look like. Two or more dimensions are called integral (meaning 'integrated') when a value can be assigned to any one dimension only by simultaneously assigning values to the other dimensions [2,9,10]. Consider the color domain. ...
Conference Paper
Full-text available
This paper describes an AFOSR-supported basic research program that focuses on developing a new framework for combining hard with soft data in order to improve space situational awareness. The goal is to provide, in an automatic and near real-time fashion, a ranking of possible threats to blue assets (assets trying to be protected) from red assets (assets with hostile intentions). The approach is based on Conceptual Spaces models, which combine features from traditional associative and symbolic cognitive models. While Conceptual Spaces are revolutionary , they lack an underlying mathematical framework. Several such frameworks have attempted to represent Conceptual Spaces, but by far the most robust is the model developed by Holen-der. His model utilizes integer linear programming in order to obtain an overall similarity value between observations and concepts that support the formation of hypotheses. This paper will describe a method for building Conceptual Spaces models for threats that utilizes ontologies as a means to provide a clear semantic foundation for this inferencing process; in particular threat ontologies and space domain ontologies are developed and employed in this approach. A space situational awareness use-case is presented involving a kinetic kill scenario and results are shown to assess the performance of this fusion-based inferencing framework.
... The study of concept combination may be seen as the touchstone of concept theories. Indeed, while elementary concepts can be suitably represented through different models (see [2], [5], [17], [29], [22], [34], [37] or [27] for a general overview), the extension to compound concepts of such representations requires a robustness of the model that may be lacking in the original theory. Except for some particular cases, the basic notions linked with concept theories are not compositional. ...
Full-text available
We extend to a certain class of compound concepts the binary model classically used for the representation of simple concepts. This class consists of concepts that are determined or modified by a single feature. The treatment of such a mixed composition shows the need to differentiate between exceptional and non exceptional modifiers. In the first case, typicality is easily retrieved from the components of the composition, while, in the second case, it is necessary to isolate the characteristic features of the initial concept that contradict the modifier. The distinction between exceptional and non exceptional modifiers plays a key role in the evaluation of resemblance with mixed compositions.
... However, such data provides a grounding to neural processes and behavior that does not exist with statistical approaches. This difference becomes evident when the CAR semantic model is compared to approaches such as Conceptual Spaces (Gardenfors, 2004;Bechberger and Kühnberger, 2019;CS), and distributional semantic models (Landauer and Dumais, 1997;Burgess, 1998;Mitchell and Lapata, 2010;Lapata, 2012, 2014;Anderson et al., 2013;Mikolov et al., 2013;Bruni et al., 2014;DSMs). Both, CAR theory and CS characterize concepts with a list of features or dimensions as the building blocks. ...
Full-text available
How are words connected to the thoughts they help to express? Recent brain imaging studies suggest that word representations are embodied in different neural systems through which the words are experienced. Building on this idea, embodied approaches such as the Concept Attribute Representations (CAR) theory represents concepts as a set of semantic features (attributes) mapped to different brain systems. An intriguing challenge to this theory is that people weigh concept attributes differently based on context, i.e., they construct meaning dynamically according to the combination of concepts that occur in the sentence. This research addresses this challenge through the Context-dEpendent meaning REpresentations in the BRAin (CEREBRA) neural network model. Based on changes in the brain images, CEREBRA quantifies the effect of sentence context on word meanings. Computational experiments demonstrated that words in different contexts have different representations, the changes observed in the concept attributes reveal unique conceptual combinations, and that the new representations are more similar to the other words in the sentence than to the original representations. Behavioral analysis further confirmed that the changes produced by CEREBRA are actionable knowledge that can be used to predict human responses. These experiments constitute a comprehensive evaluation of CEREBRA's context-based representations, showing that CARs can be dynamic and change based on context. Thus, CEREBRA is a useful tool for understanding how word meanings are represented in the brain, providing a framework for future interdisciplinary research on the mental lexicon.
... The conceptual schemas of meanings are often based on geometric or spatial construction. The conceptual space has different quality dimensions which can be either sensory related dimensions (for instance, color, pitch, temperature, weight) or abstract, non-sensory character (for example, time) (see Gärdenfors 2004Gärdenfors :11, 1999Gärdenfors :22, 1996b. These quality dimensions primarily serve to represent various qualities of objects in different domains (Gärdenfors 2001:168). ...
This paper aims at investigating metaphorical conceptions of LIFE and DEATH in Amharic from the perspective of cognitive semantics, specifically conceptual metaphor theory. In this theory, language is viewed as providing data that can show human beings conceive concepts (Lakoff & Johnson 2003: 116). The analysis shows that the Amharic speech community can metaphorically conceptualize LIFE in terms of space, specifically a journey and a bounded region in space, a precious possession, a taste, and a physical load. Also, lifetime can be viewed as the length of a physical entity. Moreover, among Amharic speakers, the abstract concept of DEATH can be understood in terms of space, more specifically a physically moving entity, falling, a unidirectional movement, and a final destination. It is also found out that the source domains of sleep, an adversary, the earth, a departure, a rest, and the loss of a precious possession can be mapped onto the target domain of death. Keywords: life, death, Amharic, conceptual metaphor, semantics
... In terms of their organization, concepts form conceptual spaces, which include relations between and operations with concepts as elements of these spaces [2][3][4]. The goal of this work is to study relations in conceptual spaces. ...
This paper discusses the theory of knowledge based on the idea of dynamical space. The goal of this effort is to comprehend the knowledge that remains beyond the human domain, e.g., of the artificial cognitive systems. This theory occurs in two versions, weak and strong. The weak version is limited to knowledge in which retention and articulation are performed through the discourse. The strong version is general and is not limited in any way. In the weak version, knowledge is represented by the trajectories of discourses in time, in a dynamical space called the discursive space, which has an arbitrary number of dimensions. Given space is used to represent a given part of knowledge. A manifold is introduced to represent knowledge with a wider scope (all knowledge). The strong version is an extrapolation of the weak version to cover all forms of knowledge, not necessarily human or manifesting in language. The use of dynamical space construction allows one to formalize knowledge as such. Such an effort requires us to initially consider knowledge as mainly a social and linguistic phenomenon, which also could be presented as a result of the evolution of the understanding of knowledge that took place in the 20th century.
Full-text available
Space is both conceptual and physical. My paper elucidates the ways in which conceptual space and physical space converge in order to serve the purposes of state apparatuses in the modernity that we inhabit, with close reference to literary and historical texts.
Manufacturers migrate their processes to Industry 4.0, which includes new technologies for improving productivity and efficiency of operations. One of the issues is capturing, recreating, and documenting the tacit knowledge of the aging workers. However, there are no systematic procedures to incorporate this knowledge into Enterprise Resource Planning systems and maintain a competitive advantage. This paper describes a solution proposal for a tacit knowledge elicitation process for capturing operational best practices of experienced workers in industrial domains based on a mix of algorithmic techniques and a cooperative game. We use domain ontologies for Industry 4.0 and reasoning techniques to discover and integrate new facts from textual sources into an Operational Knowledge Graph. We describe a concepts formation iterative process in a role game played by human and virtual agents through socialization and externalization for knowledge graph refinement. Ethical and societal concerns are discussed as well.
Full-text available
Conceptual spaces are a frequently applied framework for representing concepts. One of its central aims is to find criteria for what makes a concept natural. A prominent demand is that natural concepts cover convex regions in conceptual spaces. The first aim of this paper is to analyse the convexity thesis and the arguments that have been advanced in its favour or against it. Based on this, I argue that most supporting arguments focus on single-domain concepts (e.g., colours, smells, shapes). Unfortunately, these concepts are not the primary examples of natural concepts. Building on this observation, the second aim of the paper is to develop criteria for natural multi-domain concepts. The representation of such concepts has two main aspects: features that are associated with the concept and the probabilistic correlation pattern which the concept captures. Conceptual spaces, together with probabilistic considerations, provide a helpful framework to approach these aspects. With respect to feature representation, the existence of characteristic features (i.e., that apples have a specific taste) is essential. Moreover, natural concepts capture peaks of a probabilistic distribution over complex spaces. They carve up nature at its joints, that is, at areas with no or low probabilistic density. This last aspect is shown to be closely related to the convexity demand.
Full-text available
Ad hoc concepts (like “clothes to wear in the snow”, for instance) are highly-context dependent representations humans construct to deal with novel or uncommon situations and to interpret linguistic stimuli in communication. In the last decades, such concepts have been investigated both in experimental cognitive psychology and within pragmatics by proponents of so-called relevance theory. These two research lines have however proceeded in parallel, proposing two unconnected strategies to account for the construction and use of ad hoc concepts. The present work explores the relations between these two approaches and the possibility of merging them into a unique account of the internal structure of ad hoc representations and of the key processes involved in their constructions. To this purpose, we first present an integrated two-level account of the construction of ad hoc representations from lexical concepts; then, we show how our account can be embedded in a conceptual space framework that allows for a natural, geometrical interpretation of the main steps in such a construction process. After discussing in detail two main examples of the construction of ad hoc concepts within conceptual spaces, we conclude with some remarks on possible extensions of our approach.
Scitation is the online home of leading journals and conference proceedings from AIP Publishing and AIP Member Societies
Analyzing phonetic phenomena of categorical perception enables one to understand how the perceptual treatment of linguistic sounds segment and pattern the continuous audio-acoustic flux, and how this continuum can thus be transformed into strings of discrete elements from which a code can be built. These phonetic phenomena of categorical perception are the mediating processes between phonetic substance and phonological form, and have been so far assessed in terms of two conflicting conceptualizations: on the one hand, the reductionist and ‘substance-based’ approach of psycho-physical phonetics, and, on the other, the morphological and ‘form-based’ approach of structural phonology. This article attempts to show that phonetic phenomena of categorical perception are typical perceptual cases of critical phenomena (analogous to the phenomena of phase transitions that are observed in thermodynamics), and that their bifacial reality is congruent with the general models of critical phenomena propounded by Catastrophe Theory. Therefore it makes sense to apply the latter in the analysis of the former. The ultimate goal of this article, which summarizes in part a forthcoming book, is to lay the first bases for a catastrophe theoretic phonology.