Representational Formats in Cognitive Semantics

Dirk Geeraerts

Abstract
The paper reviews the representational fortnals that are currently in use in Cognitive Semantics for
the diagrammatic representation of the semasiological structure of lexical items. Three major types
of representation will be distinguished: the radial set model popularized by Lakoff (1987), the
schematic network model defined by Langacker (1987, 1991), and the overlapping sets model
introduced by Geeraerts (1989b). It will be argued that (given a number of straightforward adapta-
tions and additions), these three models are notational variants, in the sense that they exhibit the
same representational potentialities. This conclusion will be reached by examining whether the
three models can provide for the representation of four different types of data that arise from a
prototype-theoretical conception of semantic structure. These four types of data are; salience ef-
fects among readings, non-hierarchical semantics links among readings (like metaphor and meton-
ymy), hierarchical semantic links among readings, and discrepancies between intuitive and analyti-
cal definitions of polysemy.

1. Introductory Remarks
If Cognitive Semantics is defined as the type of semantic research conducted
within the tradition of Cognitive Linguistics as represented by Langacker (1987,
1991) and Lakoff (1987), then the introduction and elaboration of a prototype-
theoretical conception of semantic stnicture constitutes a major contribution of
Cognitive Semantics to the study of word meaning (cf. Taylor 1989); other con-
tributions would include Lakoff's theory of generalized metaphors (Lakoff &
Johnson 1980), Fillmore's frame semantics (1977, 1984), and Talmy's lexical-
semantic typological research (1983, 1985), The present paper will review the
most common representational formats currently used by Cognitive Semantics
researchers for describing prototype-oriented semantic structures. These fomiats
include the radial set model popularized by Lakoff (1987), the schematic network
model defined by Langacker (1987, 1991), and the overlapping sets model intro-
duced by Geeraerts (1989b). It will be argued that these three models are by and
large notationally equivalent. Specifically, it will be shown that the three repre-

FoliaUnguisticaXXIX/1-2 0165-4004/95/29-21 S 2.-

(C) Mouton de Gruyter, Berlin — Socieuis Liiif^iiisticd Euio/nwo
22

sentational formats can (in some cases, with slight adaptations) deal adequately
with various types of semantic data that crucially arise in the context of a proto-
type-theoretical conception of semantic structure.
There are, to avoid possible misunderstandings, a number of things that the
paper will not do. The discussion is restricted to representations of semasiological
structures — roughly, the relationship between the various readings of lexical
items. This implies that it is not primarily about onomasiological phenomena
(although a comparison with lexical field representations will be given below),
nor about representations of individual readings. The representational mecha-
nisms developed by Langacker for dealing with phenomena like figure/ground
distinctions, for instance, will not be treated here, as they pertain to the distribu-
tion of information within individual meanings rather than to the way in which
several such meanings are mutually related within the semantic structure of lexi-
cal items.
Also, the paper does not deal with representation as a mechanism for storing
and manipulating lexical information in formalized grammars, NLP programmes,
or AI systems. Deliberately, the title mentions formats rather than formalisms:
the formats meant here are primarily graphical ways of visually representing
lexical-semantic analyses for expository purposes, not formal descriptions that
can be automatically manipulated in the context of algorithmic rule systems. This
is not to say that such fomialisms are undesirable or unattainable. Promising
attempts have been made, for instance, to model lexical-semantic prototype ef-
fects in connectionist terms. The analysis of those representations is, however, a
topic of its own that will be ignored altogether in the present paper.
It would be wrong to think, incidentally, that graphical representational for-
mats are largely unrestricted in comparison with symbolic formalisms. The set of
basic elements (lines, boxes, or whatever) chosen for a graphical representation
carries a symbolic meaning: boxes may be stipulated to represent sets, different
types of lines may be chosen to represent various kinds of relations, and so on.
These choices determine how the representational elements may be combined,
and restrict the kind of things that can be represented at all. The general question
asked in this paper may therefore be formulated as follows: do the representa-
tional choices made in the radial set model, the schematic network model, and the
overlapping sets model restrict the representational possibilities of the three
models in such a way that certain phenomena that are important in a prototype-
theoretical conception of semantic structure could not be represented? Although
the three representational models as currently used highlight different phenomena
from among the set of relevant facts, it will be shown that this is a restriction of
practice rather than principle. All three models can be adapted and extended in
 23

such a way as to deal adequately with all the relevant phenomena that are here
taken as point of departure for the comparison.

2. Criteria for Evaluating Semasiological Representations

An investigation of the type intended here requires a set of criteria with which to
evaluate the competing models. Such a set of criteria may take the form of a
number of data types that the different formats should be able to represent. These
data types in turn will derive from a theoretical model of semasiological struc-
ture: what kind of phenomena can be expected to occur in the semantics of lexical
items? Within a Cognitive Semantics framework, the latter question primarily
evokes the definition of prototypicality: what kinds of phenomena fall under the
conceptual scope of the notion "prototypicality"?
In Geeraerts (1989a), it is argued that prototypicality effects fall into four, not
necessarily coinciding classes. The classification involves a distinction between
flexibility (or, defined negatively, the absence of rigidity) and difterences of
structural weight (or, negatively, the absence of equality) as structural character-
istics of prototype-based categories. These two features may be identified on an
intensional level (where definitional descriptions of a category are at stake) and
on an extensional level (where the members of a category are envisaged). In the
context of the present paper, however, we will be concerned primarily with the
intensional side of the fourfold classification that results from the cross-
classification of the two conceptual pairs just described. Such a restriction to
intensional phenomena is not uncommon in lexical semantics, given that defini-
tion is one of its major concerns. Moreover, the extensional aspect of the matter
will not be entirely neglected; it will become clear further on in the article, how-
ever, that it can be easily incorporated into a representation that starts off on an
intensional basis. To begin, then, let us have a closer look at the two crucial phe-
nomena as defined from an intensional perspective.

2.7 Intensional non-rigidity

Intensional non-rigidity involves the absence of classical definitions for a cate-
gory: if no definition in terms of necessary-and-sufficient attributes is available
for a category, then that category is defined less rigidly than the classical model
of definitions predicts. Instead of a single description consisting of individually
necessary and jointly sufficient features, the definition takes the fonn of a multi-
plicity of partial descriptions.
However, as argued in Geeraerts (1987), the so-called absence of classical
definitions as such does not suffice to establish the non-orthodo.x. prototype-based
nature of lexical categories. Even in the classical model, the absence of a single
24

definition in terms of necessary and sufficient attributes is a regular feature of

lexical categories — in those cases where they are polysemous, in fact: if a
polysemous category is conceived of as one that cannot be adequately described
by means of a single definition, then it necessarily fits the description of inten-
sional non-rigidity mentioned above. A solution for this conceptual problem re-
quires making a distinction between various operational definitions of polysemy
(for a more extensive treatment, see Geeraerts 1993). The absence of classical
definability, then, is only a non-orthodox feature of lexical categories if it relates
not to the (polysemous) category as a whole, but rather to the individual readings
within that polysemous set. This involves distinguishing between two ways of
determining what a distinct reading of a lexical item is.
On the one hand, a number of operational tests reveals what is, from an in-
tuitive point of view, a different meaning of a lexical item. One of these tests is
Quine's "p and not p" test (1960): taking into account the readings "harbour" and
"fortified sweet wine from Portugal" of port, these readings are revealed as truly
different meanings (or "senses") by the possibility of uttering, in a specific con-
text, a sentence like Sandeman is a port but not a port. If port were merely
vague rather than polysemous between the "harbour" and the "wine" reading,
such a sentence would turn out to be invariably contradictory.
On the other hand, a definitional criterion for polysemy (as informally stated
by Aristotle in the Posterior Analytics Il.xiii) says that an item has more than one
lexical meaning if there is no minimally specific definition covering the extension
of the item as a whole, and that it has no more lexical meanings than there are
maximally general definitions necessary to describe its extension. Definitions of
lexical items should be maximally general in the sense that they should cover as
large a subset of the extension of an item as possible. Thus, separate definitions
for "blended sweet fortified wine from Portugal" and "vintage sweet fortified
wine from Portugal" could not be considered definitions of lexical meanings,
because they can be brought together under the definition "sweet fortified wine
from Portugal" On the other hand, definitions should be minimally specific in the
sense that they should be sufficient to distinguish the item from other non-
synonymous items. A maximally general definition covering both port "harbour"
and port "kind of wine" under the definition "thing, entity" is excluded because it
does not capture the specificity ot port as distinct from other items.
Given these two approaches, it may now be shown that discrepancies be-
tween both (and hence, non-traditional forms of semantic structure) may exist.
On the one hand, word meanings that have to be recognized as single senses ac-
cording to an intuitive approach, appear to lack a classical definition (i.e. would
have to be recognized as polysemous according to the definitional criterion). This
 25

is the case for the "biological species" reading of bird, as analyzed in figure 1:
the reading "member of the biological species Aves" functions as a single mean-
ing (a penguin is a bird but not a bird is decidedly odd), but there is no classical
definition complying with the necessity-cum-sufficiency criterion. (The features
that are common to all birds as indicated in figure 1 are not jointly distinctive, as
they also apply to the duck-billed platypus.) On the other hand, there exist cases
where what is an intuitively distinct meaning is smaller than what would be rec-
ognized as a separate sense according to the definitional approach. Examples are
provided by autohyponymous words like dog, which has both the general reading
"member of the species Canis familiaris", and the restricted meaning "male
member of the species Canis familiaris" In the latter reading, dog is an antonym
of bitch; in the former, it contrasts with other animal terms (like cat). Now, the
set-theoretically more restricted meaning "male dog" does not have the generality
that is required according to the definitional criterion (because the "male dog"
cases can always be subsumed under the definition of the hyperonymous reading
"Canis familaris in general, regardless of sex"), but at the same time, Mirza is a
dog but not a dog makes perfect sense, assuming that Mirza is a bitch.

Figure I: A definitional analysis of "bird" in an overlapping sets

model

e.g. robin e.g. ostrich e.g.

O) chicken
as
[^
e.g. penguin ^

a. 'Being able to fly* a. "Not domesticated"

b. "Having feathers' f. "Being born from eggs"
c. 'Being S-shaped" g. 'Having a beak or bili'
d. 'Having wings'

In short, what is at stake with regard to intensional non-rigidity as one of the four
types of prototypicality effects is not the absence of classical definability as such,
but rather a discrepancy between intuitive and definitional conceptions of
polysemy.
26

2.2 Intensional non-equality

Intensional non-equality takes the form of differences in the structural importance
of the various subsets that may be definitionally distinguished within the range of
application of a lexical item. There is a deliberate vagueness in this formulation,
to the extent that the "subsets" in question may either constitute different mean-
ings in the traditional sense or not. If they do, the intensional differences of
structural weight involve differences of salience among the various meanings of a
word. Let us observe, for instance, that bird might be used metaphorically to in-
dicate an airplane (as in a gigantic silver bird approached from the west). That
metaphorical extension from the "biological species" reading would certainly be
less salient than the latter as such. This is true from a logical point of view (to the
extent that the metaphorical reading is a semantic extension of the former), from
a psychological point of (to the extent that the metaphorical reading is less likely
to be permanently stored in the mental lexicon of the language user), and from a
statistical point of view (to the extent that the metaphorical reading is less com-
mon than the literal one). Admittedly, it need not always be the case that the vari-
ous indices of salience coincide as neatly as in the example, but at least it will be
clear what criteria may be taken into account to establish salience.
If the differently weighted subsets do not constitute word meanings in the
traditional sense, they involve subsets of the kind found in figure 1. The defini-
tional analysis of bird divides the range of application of that word (in its literal
reading) into a number of subsets: we may, for instance, distinguish the central
subset comprising robins and other usual birds from a peripheral subset such as
the one including ostriches. The centrality relationship between the sets as such
indicates the differences in salience: the central set includes typical birds. It may
be noted that there is a second way of defining subsets with regard to figure 1.
Instead of talking about the minimal subsets distinguished in the figure (in the
sense in which ostriches belong to a different subset than robins), maximal sub-
sets may be envisaged (in the sense in which ostriches, robins, and chickens be-
long to the subset defined by the feature "having feathers"). Minimal subsets in
the figure are defined in terms of multiple descriptive features, whereas maximal
subsets are based on single features. It is a traditional part of prototype-
theoretical semantics, then, that the most salient minimal subset tends to coincide
with that area where a maximal number of maximal subsets overlap. Or, in other
words, the prototypical instances of a category are those in which a maximal
number of structurally relevant characteristics coincide.
Regardless of whether the intensional salience effects at stake here are situ-
ated within or among different meanings, they invariably involve cases of struc-
tured polysemy. The term polysemy is obviously somewhat confusing here, since
 27

it normally only relates to situations of the first type described above, i.e. those
involving different meanings in the traditional sense. A more neutral term, en-
compassing both situations, could be multiple semantic applicability, but as that
expression introduces an unacceptable cumbersomeness, we may stick to
polysemy for the present purposes. The polysemy is a structured one, because the
differently weighted subsets are related among each other by means of specific
structural links. In the bird example, the subsets within the literal reading
("member of the biological species Aves") are related by similarity: the periph-
eral subsets are related by partial similarity to the central subset that includes
robins and the other regular birds. Further, the literal reading is related by a
metaphorical link to the peripheral "airplane" reading. (It may be remarked that
metaphor also involves relations of similarity, but the terminological point would
then be to distinguish literal from figurative similarity. In what follows, we will
use similarity to refer to literal similarity, and metaphor for figurative similarity.)
Similarity and metaphor do not exhaust the set of possible relations in polyse-
mous clusters as intended here. In general, all relations that have been tradition-
ally identified in diachronic semantics may recur in the synchronic structure of
lexical items; metonymical relations, for instance, can be expected to be very
frequent.
One specific kind of semantic relationship deserves to be mentioned sepa-
rately, as it plays a crucial role in one of the representational formats that we will
be dealing with in paragraph 3. In traditional diachronic semantics,
"specialization" and "generalization" refer to hierarchical semantic relations, i.e.
relations that can be situated along the vertical dimension of a taxonomical tree.
The specialization/generalization relationship holds, for instance, between the
"male Canis familiaris" reading of dog and its "Canis familiaris in general"
reading. In extensional terms, the more specialized reading is always a proper
subset of the more general reading. Other terms beside "specialization" and
"generalization" are often used; "abstraction" and "schematization" in particular
may be met with as synonyms of "generalization" The importance of hierarchi-
cal semantic relations in the classical definability debate will be obvious from the
discussion in section 2.1.: the ideal of classical definability implies that for any
intuitively distinct reading of an item, a distinctive definition can be found that
generalizes over the specific instances of use that fall within the range of that
reading.
28

3. The Current Representational Formats

3.1 The overlapping sets model, the radial set model, and the schematic net-
work model
The three representational models that are currently used in Cognitive Semantics
are the overlapping sets model, the radial set model, and the schematic network
model. The overlapping sets model is illustrated by figure 1; applications may be
found in studies such as Geeraerts (1990), Cuyckens (1991), Schmid (1993),
Geeraerts, Grondelaers & Bakema (1994). The basic elements in this representa-
tional format are the members of a category (such as the types of birds in figure
1), or, in some cases, instances of use of the category as found in a text corpus.
These basic elements are grouped together on the basis of the features that they
share or the senses that they exemplify. Each grouping is typographically repre-
sented by means of a Venn-diagram. The different groupings may overlap; the
area in the figure where the sets overlap maximally constitutes the prototypical
center of the category.

Figure 2: A definitional analysis of "bird" in a radial

set model

The radial set model is described in Lakoff (1987); examples may be found in
the work of Brugman (1989), Janda (1990), Nikifoddou (1991), Goldberg
(1992), and others. The basic elements in a radial set representation are the
meanings or senses of a category; these are connected in pairwise fashion by
means of relational links that indicate how one reading is an extension of an an-
other (for instance, on the basis of similarity). The typographical distribution of
 29

the various readings on the page illustrates the prototypical structure of the cate-
gory: the prototypical sense is situated roughly in the middle of the figure, while
the extensions that emanate from this central sense are grouped radially around it.
In more complex figures, a sense from which many others emanate is likely to be
highly salient within the category. Figure 2, which is adapted from Cuyckens
(1991), shows how the overlapping sets representation of figure 1 can be trans-
formed into a radial set representation. (Note that the features 1-7 in figure 2
correspond to features a-g in figure 1.) Figure 2 is, however, not entirely canoni-
cal as far as radial set representations go: the visual metaphor indicating the cen-
tral position of the 1234567-case is not included (but this could, of course, be
remedied by shifting the 1234567-case to the middle of the square formed by the
other readings). Also, figure 2 differs from the usual radial set representations to
the extent that the elements of the radial set (types of birds) are referential sub-
sets (or, if one wishes, members) of the category "bird", rather than meanings or
senses in the usual sense.

Figure 3: A definitional analysis of "bird" in a

schematic network model

bird

/ \ N

bird' ^ Chicken

robin sparrow blackbird

The schematic network model is described in detail by Langacker (i 987, 1991);

it is illustrated by the work of Rudzka-Ostyn (1985, 1989), Tuggy (1987, 1993),
Taylor (1992), Casad (1992), Schuize (1993), and others. The basic elements in
the schematic network model may be meanings or members of a category. As in
the radial set model, these elements are connected by means of relational links,
but a systematic distinction is maintained between two kinds of links: links of
schematization and links of extension. Schematicity involves the relationship
between a subordinate node and a superordinate node in a taxonomical hierarch>.
30

The category "bird", for instance, is schematic with regard to "robin",

"sparrow", "ostrich", and other types of birds. Extension, on the other hand, in-
volves partial schematicity: assuming that the subset comprising robins, spar-
rows, blackbirds (and others) constitutes the prototypical center of the category
"bird", the subset comprising chickens is an extension from that prototype.
Chickens do not fall within the prototypical subset, but the concept "chicken" can
be seen as an extension (based on a relationship of similarity) of the prototypical
sense. (And the same holds, obviously, for "kiwi", "ostrich", and "penguin".)
Precisely because the example involves similarity, the relation is one of partial
schematicity. Typographically, the schematization links are indicated by solid
arrows along the vertical dimension of the figure, whereas the extension links are
represented by broken arrows along the horizontal dimension of the representa-
tion. Prototypicality may be indicated by using thicker lines for drawing the
boxes in the figures. Figure 3 represents part of the "bird"-category in the form of
a schematic network. (The concept BIRD' corresponds with the prototypical case
of figures 1 and 2; BIRD represents the category as a whole. The introduction of a
BiRD'-node is necessary if a separate representation of the prototype of the cate-
gory is required.)

3.2 Onomasiological parallels

An instructive parallel may be drawn between, on the one hand, the distinction
between radial set and schematic network representations versus an overlapping
sets representation, and on the other hand, alternative representations pertaining
to the structure of lexical fields. In the former case (the case that primarily inter-
est us here), the representations involve the semasiological structure of individual
lexical items — roughly, the relationship between the various readings of a lexi-
cal item. In the latter case, the representations involve the onomasiological
structure of lexical fields — the semantic associations between various words
that are somehow related in meaning. (Note, incidentally, that the term onomasi-
ology is used here in a broad sense, where it includes any association of lexical
items on the basis of semantic relatedness. There also exists a more restricted
interpretation of the term, in which it refers exclusively to alternative lexicaliza-
tions of specific senses or referents.)
In lexical field research, then, three major types of representational formal-
isms may be distinguished:
First, the traditional representation as found in, for instance, Lehrer (1974)
takes the field metaphor more or less literally, by positioning the lexical items in
the field in a two-dimensional space. The distinction between the items may then
be indicated by dividing the field on the basis of .semantic dimensions (that each
occur with specific values). In the upper part of figure 4, an abstract example is
 31

presented with four items distinguished by two semantic dimensions (numbered 1

and 2). Each dimension is a binary one, so that the dimensional values may be
indicated by means of plus and minus signs.
The second representational format is the componential one popularized
(though not invented) by Katz & Fodor (1963). A componential analysis turns the
two-dimensional analysis inside out: the distinctive dimensions that structure the
field are now placed "within" the lexical items, as part of their definition. By
attributing a separate definition to each item in isolation, the associative links
among the items remain implicit: they have to be derived from the presence or
absence of specific features in the componential definitions.
Finally (as exemplified in the lower part of figure 4), a relational representa-
tion joins the traditional type in that it does not look inside the lexical items but
presents them as entities without internal structure linked by external relations.
The distinction between the first type of representation and the present one re-
sides in the nature of the links grouping together lexical items. The first approach
distinguishes sets of items on the basis of shared features like the dimensional
values -1-1 or -2; the sets overlap in the sense in which, for instance, item A be-
longs both to the set defined by -i-l and to that defined by -\-2. In the relational
representation, the groupings consist of individual links between pairs of lexical
items; following the definition of the relational approach in Lyons (1963), the
relations usually envisaged are taken from a restricted set including at least

Figure 4: Alternative representations of semantic relations in a lexical field

TRADITIONAL FIELD REPRESENTATION:

DIMENSIONAL DIMENSIONAL
VALUE-1-1 VALUE-1
DIMENSIONAL
Item A Item B
VALUE + 2
DIMENSIONAL
ItemC itemD
VALUE-2

COMPONENTIAL ANALYSIS:

Item A Item B Item C Item

+] -1 +1 -1
+2 +2 -2 -2

RELATIONAL NETWORK REPRESENTATION:

• r
32

hyponymy, antonymy, and synonymy (but extensions of the set are not unusual,
cf. Cruse 1986). In the example of figure 4, the opposite values of the items with
regard to the dimensions 1 and 2 are interpreted as an antonymous relationship
(in the sense in which antonyms like alive and dead might be componentially
represented as H-LIVE versus -LIVE).
The semasiological representations compared in the present paper may now
be aligned with the onomasiological ones just reviewed. On the one hand, the
radial set representation and the schematic network representation are both rela-
tional formats to the extent that they primarily link individual readings of lexical
items in pairwise fashion. They are, in other words, network representations. The
overlapping sets representation, on the other hand, is of the same "set-theoretical"
kind as the first type distinguished in the case of lexical fields. The relations be-
tween the approaches presented so far are schematically represented in figure 5.

Figure 5: The relationship between the semasiological and the

onomasiological representational formats

EXTERNALC ROUPING OF INTERNAL

RELEVAKlr ENTITIES ANALYSIS OF
RELEVANT
ENTITIES
BINARY OVERLAPPING
RELATIONS SETS

SEMASIO- -radial sets overlapping componential

LOGY -schematic sets model analysis
networks

ONOMASIO- relational traditional componential

LOGY approach field approach analysis

An important remark with regard to the schematic overview in figure 5 concerns

the presence of the componential approach on the semasiological level. First of
all, it should be noted that the componential approach in its Katzian form is al-
ready both a semasiological and an onomasiological representational format: the
structural analysis of the lexical field automatically yields a definition of the
items in the field. Moreover, the componential approach need not take its starting-
point in a lexical field, but may also start from the polysemous set of readings
attached to a single lexical item. One may compare, for instance, Pottier's field-
based application of componential analysis (1964) with Katz & Fodor's word-
based approach. Within the domain of Cognitive Semantics, radial set represen-
 33

tations have sometimes been enriched with componential descriptions of the

nodes in the network; see, for instance, Brugman's analysis of over (1989), and
the example presented in figure 2, as drawn from Cuyckens (1991). As an overall
representation of semasiological data, however. Cognitive Semantics has tended
to avoid componential analysis (compare Fillmore 1975). The basic reason seems
to be that a componential analysis, describing various readings in isolation, tends
to obscure the structural relations among those word meanings. Specifically, the
prototype-based differences of salience (or structural weight) that are crucial to
Cognitive Semantics are not automatically incorporated into a componential rep-
resentation.

4. Current Formats as Notational Variants

The representational formats of paragraph 3 may now be compared among each
other on the basis of the requirements formulated in paragraph 2. We will indicate
how each of the three representations in their usual form deals with the various
types of information discussed in section 2. It will be shown that they are, by and
large, notational variants.
Differences of salience among subsets in the range of application of an item
are represented in roughly the same ways in the three models. On the one hand,
salience effects receive an indirect, structural representation. In the overlapping
sets model, for instance, salient subsets are those that are constituted by multiple
overlapping, i.e. by the coincidence of a large number of relevant descriptive
features. Similarly, in the radial set model and the schematic network model, sali-
ent readings would be those that constitute a semantic center, in the sense that
they are the basis for numerous semantic extensions. Conceptually, salient read-
ings contribute significantly to the semantic coherence of the category; typo-
graphically, they are the basis from which multiple extensions to other readings
depart.
On the other hand, salience may be represented more directly with typo-
graphical means, like shading (cf. figure 1) or thicker lines (cf. figure 3). This
way of representing salience is more generally applicable than the former, pre-
cisely because the different indices of salience need not coincide: what is statisti-
cally prominent need not be the logical, semantic center of the category. This
often happens, for instance, when the center of a category shifts from its etymo-
logical origin to one of the extensions arising from that origin. Through semantic
reorganisations, the statistically dominant extension would then eventually be-
come the new semantic center of things, but at least at one stage of the word's
semantic history, the most frequent reading need not be the center of semantic
coherence. In this sense, then, a neutral, direct indication of salience will be use-
34

ful over and above the indirect indication of salience that derives from the se-
mantic structure of the category.
Hierarchical semantic relations receive a less uniform treatment in the three
models. In the radial set model, schematicity may be represented as one particular
type of semantic relation next to many others; graphically, this might take the
form of a binary link labeled "schematization" or "generalization" In the sche-
matic network model, of course, such links receive a special treatment in the
sense that they will invariably be situated along the vertical dimension of the
figure, whereas the other links (like metaphor and metonymy) will be placed
along the horizontal dimension. In the overlapping sets model, the schematization
relation surfaces in the form of sets encompassing others: any box in the diagram
that properly includes an other box, generalizes over the latter.
One specific case along the hierarchical dimension needs to be mentioned
separately. Individual members of a category are situated at the bottom of a tax-
onomy; "individual members" as meant here are primarily conceived of as indi-
viduals in a philosophical sense: separate entities rather than groups or classes.
Thus, the individual members of bird would be birds (like Pete, your auntie's
parrot) rather than specific classes of birds. In a set-theoretical sense, individuals
constitute singletons. Now, once a representational format contains a mechanism
for depicting hierarchical relations, it can in principle also deal with individual
members of a category.
However, considering that these individual members belong to the exten-
sional level rather than to the intensional level of the category, they may also be
treated somewhat differently than the other levels of a taxonomy. Consider the
overlapping sets model. Representing individual members as bundles of features
(i.e. overlapping sets) meets with the difficulty that the number of features identi-
fying an individual could be unlimited, or at least very large. Rather, individual
members of a category are often represented by means of "point-like" represen-
tations — by listing them with an individual name, so to speak. In the overlapping
sets model, this can be achieved by placing individual category members
(represented graphically as points, for instance) within the appropriate areas
constituted by the featural sets. In network models, a specific link (like maybe a
"membership link") might be used with the same purpose of distinguishing indi-
viduals.
Two further nuances are necessary here. First, it would seem that the bottom
level of a taxonomy may sometimes be shifted somewhat. In figure 1, for in-
stance, one could say that "ostrich" and "chicken" etc. are treated as the relevant
members of the category "bird" (rather than auntie's parrot). This also implies
that the distinction between the intensional and the extensional level may be sub-
 35

ject to contextual shifts. In those cases in which they constitute the bottom of the
hierarchy, types of birds like "ostrich", "chicken" etc. are treated as the actual
members of the category "bird", i.e. they are taken to belong to the extensional
level. If, however, auntie's parrot is added to the set, the extensional level be-
comes more specific. Pete the parrot is an extensional token of the intensional
type "parrot", and of the intensional type "bird"; but in representations like figure
1, the category "parrot" is treated as a token of the category "bird", i.e. is treated
as an entity rather than a class.
Second, "individual members" may also be instances of a category as found
in, for instance, corpus-based analyses. If the specific instances of use of the
various senses of a word are considered members of the extension of the word, a
reference to illustrative quotations from the corpus can then be incorporated into
the representation in the same way that Pete the parrot would be listed as a mem-
ber of the category.
Non-hierarchical semantic relations are treated as binary links in both the
radial set and the schematic network model; in the latter, they obviously occur
along the horizontal axis, as the vertical axis is reserved for relations of schema-
ticity. Depending on the nature of the link, labels like "metaphor" or "metonymy"
or "similarity" (or others) may be attached to the links. In the overlapping sets
model, on the other hand, a distinction has to be maintained between similarity
relations and relations like metaphor and metonymy. Clearly, similarity relations

Figure 6: The representational scope of the

overlapping sets model
36

are represented in the form of overlapping sets defined in terms of the character-
istics featuring in the similarity relation. But it is less obvious to represent meta-
phor and metonymy in the same way, i.e. on the basis of shared features.
For metaphorical relations, this is not entirely impossible; the "airplane"
reading of bird undoubtedly shares the feature of "flying" with (many) birds. The
differences between birds and airplanes could then be indicated by adding fea-
tures that both readings do not share, such as the fact that the literal reading in-
volves living beings while the metaphorical one involves machines. However, if
the distinction between literal and metaphorical similarity is to be representa-
tionally marked, it is advisable to indicate metaphorical relations explicitly by
means of labelled links.
Such labelled links would seem to be indispensable for representing me-
tonymical links at any rate: it is difficult to imagine how metonymical relations in
general can be reduced to shared features. The description of metonymy basically
involves relational predicates linking the two readings that are associated through
metonymy: the extended reading refers to, for instance, something that is part of
(or caused by, or contained in, or characterized by, etc.) the referent of the pri-
mary reading. This relational nature of metonymy can be most easily represented
with relational means, i.e. by means of labelled links.
In order to further establish the equivalence between the three models, an

Figure 7: The representational scope of the

schematic network model

(5) (1) (2) (3) (7) (6)

 37

abstract example may again be useful. Let us assume we have a lexical item
whose literal meaning is a cluster of three subsets, constituted by the overlapping
features {a} and {b}; also, all instances of the literal reading share feature {c).
Three subsets then have to be distinguished: {c,a}, {c,b}, and {c,a,b}. Further,
reading {c,b} is metonymically extended towards a reading defined by {d}. Each
of the readings distinguished so far is represented in a corpus by a number of
examples (in a text corpus, these could be sentences in which the word is used in
the reading in question): examples (1), (2), and (3) exemplify {c,b}; examples
(4), (5), and (6) illustrate {c,a,b}, {c,a}, and {d} respectively. There is no exam-
ple illustrating {c} in isolation, but there does occur an example (7) exemplifying
a metaphorical extension {e} of {c}. Given the frequency of the examples, {c,b}
is clearly more salient than the other subsets. Figures 6-8 (on pages 35-37) illus-
trate how the same information may be presented in each of the three representa-
tional approaches.
The representation of the discrepancy between intuitive and analytic
polysemy requires further additions to the diagrams exemplified by figures 6-8.
In the overlapping sets representation, any area of the entire cluster that is intui-
tively felt to be a distinct reading may be singled out by typographical means. In
the /7/«f-example of figure 1, for instance, the dotted line that indicates that the
features 6 and 7, when taken together, do not suffice to define the category in a
distinctive fashion, could also be used to indicate that the cluster of sets that falls
within the range of the dotted line usually functions as a single reading. The net-

Figure 8: The representational scope of the radial set model

metaphor

M) {1){2)(3) (6) (7)

38

work representations (i.e. the schematic network and the radial set representation)
require the definition of a new type of node, or alternatively, a new type of la-
belled link. In fact, in situations like that exemplified by bird, the semantic link
between the analytically distinguishable subsets and the intuitive reading could
be described as one of alleged generalization: the language user so to speak pre-
tends that there is a classically definable, unitary concept generalizing over the
semantically more specific applications of the category. Switching from the la-
belled links to the nodes, the nodes created by the process of "alleged generaliza-
tion" take the form of a disjunction: various definitionally distinct subsets are
lumped together as if they constituted a single classically definable concept.
Figure 9 illustrates how a radial set representation may be enriched to incor-
porate "alleged generalization" as an example of the way in which intuitive and
analytical conceptions of polysemy need not coincide. Obviously, the radial set
representation is decidedly more cumbersome in these cases than the overlapping
sets representation: the graphical changes that have to be applied to the figure to
achieve the required distinction are more far-reaching than in the other represen-
tational format. In principle, however, the representational possibilities are basi-
cally the same.

Figure 9: Enriched radial set representation of "bird"

123A5B7
or 234567
or 23467
or 4567
or 3567
 39

5. Concluding Remarks
In various respects, the story told so far is not yet complete. The exploration, that
is, of adequate representational formats for prototype-oriented lexical semantics
should not stop at the observation that the three formats currently used are no-
tional variants (or at least, can be turned into notational variants). Even disre-
garding the whole domain of onomasiological research, the following observa-
tions have to be made.
First, the representations developed so far do not explicitly provide a place
for syntagmatic data. The readings incorporated into the representations may
each be subject to specific syntagmatic restrictions. In the case of spatial rela-
tional concepts, for instance, a particular relation may occur only with particular
pairs of landmarks and trajectors; furthermore, this set of syntagmatic contexts is
likely to be subject to typicality effects just like the paradigmatic readings them-
selves. The diagrammatic representations, then, will have to be elaborated with a
further layer on which the syntagmatics of each of the paradigmatic readings is
specified.
Second, the notational equivalence of the three models does not imply that no
choices have to be made when working with these formats. There are, for one
thing, empirical matters to be clarified: what the relevant prototypical centers or
the psychologically real schematic meanings within a concept are, is not some-
thing that can be derived from the models as such: it is an empirical matter that
has to be settled before representations are suggested. (In this respect, see Taylor
1990 for an insightful discussion of the empirical aspects of choosing between
prototypicality and schematization; compare also Winters 1992.) For another
thing, the choice of one form of representation rather than another may be a mat-
ter of pragmatic focus: when the research interest lies mainly with definitional
problems (involving the differences between intuitive and analytical definitions of
polysemy), the overlapping sets model may be more useful than the radial set
model. The latter, conversely, may be more appropriate when the focus is on
mechanisms of semantic extension, as in diachronic semantics.
Finally, the limits of the representational formats have to be taken into ac-
count: the larger the set of relevant semantic dimensions, the more difficult it
becomes to devise a graphical representation that is both elegant and complete.
Splitting up the representation in various levels is one way of extending the repre-
sentational possibilities: each of the components of a simplified, skeletal repre-
sentation at level 1 may then be treated separately and in more detail at level 2.
for instance. In general, the representations should not be given absolute value:
they merely serve to present a linguistic analysis of word meaning in as clear and
systematic a way as possible. The analysis itself, then, obviously comes first.
40

both logically and chronologically; the representation should never become an

end in itself.

Address of the author:

Dirk Geeraerts
Departement Linguistiek
University of Leuven
Blijde-Inkomststraat 21
B-3000 Leuven (Belgium)
e-mail: [email protected]

References
Brugman, Claudia M. 1989. The Story of Over. Polysemy, Semantics, and the Structure of the
Lexicon. New York: Garland.
Casad, Eugene H. 1992. "Cognition, History, and Cora yee." Cognitive Linguistics 3, 151-186.
Cruse, D. Alan. 1986. Lexical Semantics. Cambridge: Cambridge University Press.
Cuyckens, Hubert. 1991. The Semantics of Spatial Prepositions in Dutch. A Cognitive-Linguistic
Exercise. PhD dissertation, Universltaire Instelling Antwerpen.
Fillmore, Charles 1975. "An Alternative to Checklist Theories of Meaning.'' Proceeding of the
Annual Meeting of the Berkeley Linguistics Society I, 123-131.
Fillmore, Charles. 1977. "Scenes-and-frames Semantics" In: Antonio Zampolli (ed.). Linguistic
Structures Processing. Amsterdam: North Holland, 55-81.
Fillmore, Charles. 1984. "Frames and the Semantics of Understanding". Quaderni di Semantica 6,
222-254.
Geeraerts, Dirk. 1987. "On Necessary and Sufficient Conditions'. Journal of Semantics 5, 275-291.
Geeraerts, Dirk. 1989a. "Prospects and Problems of Prototype Theory". Linguistics 27, 587-612.
Geeraerts, Dirk. 1989b. Wat er in een woord zit. Facetten van de lexicale semantiek. Leuven:
Peeters.
Geeraerts, Dirk. 1990. "The Lexicographical Treatment of Prototypical Polysemy" In: Savas L.
Tsohatzidis (ed.). Meanings and Prototypes. Studies in Linguistic Categorization. London -
New York: Routledge, 195-210.
Geeraerts, Dirk. 1993. "Vagueness's puzzles, polysemy's vagaries'' Cognitive Linguistics 4,
223-272.
Geeraerts, Dirk, Stef Grondelaers & Peter Bakema. 1994. The Structure of Lexical Variation.
Meaning, Naming, and Context. Berlin: Mouton de Gruyter.
Goldberg, Adele. 1992. "The Inherent Semantics of Argument Structure: the Case of the English
Ditransitive Construction''. Cognitive Linguistics 3, 37-74.
Janda, Laura A. 1990. 'The Radial Network of a Grammatical Category — its Genesis and Dy-
namic Structure". Cognitive Linguistics 1, 269-288.
Katz, Jerrold J. & Jerry A. Fodor. 1963. 'The Structure of a Semantic Theory" Language 39,
170-210.
Lakoff, George. 1987. Women, Fire, and Dangerous Things. What Categories Reveal about the
Mind. Chicago - London: The University of Chicago Press.
Lakoff, George & Mark Johnson. 1980. Metaphors We Live By. Chicago - London: The Univer-
sity of Chicago Press.
Langacker, Ronald W. 1987. Foundations of Cognitive Grammar I. Theoretical Prerequisites. Stan-
ford: Stanford University Press.
Langacker, Ronald W. 1991. Foundations of Cognitive Grammar II. Descriptive Application. Stan-
ford: Stanford University Press.
Lehrer, Adrienne. 1974. Semantic Fields and Lexical Structure. Amsterdam: North Holland.
 41

Lyons, John 1963. Structural Semantics. Oxford: Blackwell.

Nikiforidou, Kiki. 1991. "The Meanings of the Genitive: a Case Study in Semantic Structure and
Semantic Change". Cognitive Linguistics 2, 149-205.
Pottier, Bernard. 1964. "Vers une semantique moderne" Travaux de Linguistique et de Littera-
ture2, 107-137.
Quine, Willard V.O. 1960. Word and Object. Cambridge, Mass.: MIT Press.
Rudzka-Ostyn, Brygida. 1985. "Metaphoric Processes in Word-formation" In: Wolf Paprotte &
Ren6 Dirven (eds.). The Ubiquity of Metaphor. Amsterdam: Benjamins, 209-241.
Rudzka-Ostyn, Brygida 1989. "Prototypes, Schemas, and Cross-category Correspondences: the
Case of ask". Linguistics 27, 613-662.
Schmid, Hans-Jorg. 1993. Cottage und Co., idea, start vs. begin. Die Kategorisierung als Grund-
prinzip einer differenzierten Bedeutungsbeschreibung. Tubingen: Niemeyer.
Schulze, Rainer. 1993. "The Meaning of (a)round: a Study of an English Preposition". In: Richard
A. Geiger & Brygida Rudzka-Ostyn (eds.). Conceptualizations and Mental Processing in Lan-
guage. Berlin: Mouton de Gruyter, 399-431.
Talmy, Leonard. 1983. "How Language Structures Space" In: Herbert Pick & Linda Acredolo
(eds.). Spatial Orientation: Theory, Research, and Application. New York: Plenum Press, 225-
282.
Talmy, Leonard. 1985. "Lexicalization Patterns: Semantic Structure in Lexical Forms". In: Timothy
Shopen (ed.). Language Typology and Syntactic Description 3. Grammatical Categories and
the Lexicon. Cambridge: Cambridge University Press, 57-149.
Taylor, John. 1990. "Schemas, Prototypes, and Models: in Search of the Unity of the Sign" In:
Savas L. Tsohatzidis (ed.). Meanings & Prototypes. London - New York: Routledge, 521-534.
Taylor, John R. 1989. Linguistic Categorization. Prototypes in Linguistic Theory. Oxford: Claren-
don Press.
Taylor, John R. 1992. "Old Problems: Adjectives in Cognitive Grammar". Cognitive Linguistics 3,
1-36.
Tuggy, David. 1987. "Scarecrow Nouns, Generalizations, and Cognitive Grammar". Proceedings of
the Annual Meeting of the Pacific Linguistics Conference 3, 307-320.
Tuggy, David. 1993. "Ambiguity, Polysemy, and Vagueness". Cognitive Linguistics 4, 273-290.
Winters, Margaret E. 1992. "Schema's and Prototypes: Remarks on Syntax Change" In: Gunter
Kellermann & Michael D. Morrissey (eds.). Diachrony witliin Synchrony: Language History
and Cognition. Frankfurt am Main: Lang, 265-280.