INTRODUCTION TO
COGNITIVE LINGUISTICS 8
TRUONG VAN ANH
SAI GON UNIVERSITY

Categorisation and idealised cognitive
models

8.1 Categorisation and cognitive semantics
In the 1970s the term prototype theory
emerged from the research of Eleanor Rosch
and her colleagues.
Humans categorise not by means of the
necessary
and
sufficient
conditions
assumed by the classical theory, but with
reference to a prototype: a relatively abstract
mental representation that assembles the key
attributes or features.

The most recent theories of categorisation
assert that a key aspect of knowledge
representation is the dynamic ability to form
simulations, an idea that was introduced in
the previous chapter.

Firstly, an investigation of prototype theory
provides a picture of the historical context
against which cognitive linguistics emerged as a
discipline. The development of prototype theory
in the 1970s resonated in important ways with
linguists whose research would eventually
contribute to defining the field of cognitive
semantics.
In this way, 'Prototype Theory' inspired some of
the early research in cognitive semantics.

Secondly, and perhaps more importantly,
although it now seems that prototype theory
cannot be straightforwardly interpreted as a
theory of knowledge representation. Thirdly, as
we mentioned above, Lakoff's (1987) book set
the scene for the development of three
important strands of research within cognitive
linguistics: (1) Conceptual Metaphor Theory
(Chapter 9); (2) cognitive lexical semantics
(Chapter 10); and (3) a cognitive approach to
grammar

8.1.1 The classical theory
Consider once more the familiar lexical concept
BACHELOR. For an entity to belong to this category, it
must adhere to the following conditions: 'is not
married'; 'is male'; 'is an adult'. Each of these
conditions is necessary for defining the category, but
none of them is individually sufficient because 'is not
married' could equally hold for SPINSTER, while 'is
male' could equally hold for HUSBAND, and so on.
The English noun chair names a category that can be
decomposed into the set of semantic features or
markers shown in Table 8.1.

8.1.2 The definitional problem
While the classical theory holds that categories have
definitional structure, in practice it is remarkably difficult
to identify a precise set of conditions that are necessary
and sufficient to define a category.
Let's discuss the category GAME. While some games
are characterised by AMUSEMENT, others are
characterised by LUCK, like dice games, still others by
SKILL or by COMPETITION, like chess.
Consider the category CAT. We might define this
category as follows: 'is a mammal'; 'has four legs'; 'is
furry'; 'has a long tail'; 'has pointy ears'. What happens if
your cat gets into a fight and loses an ear?

8.1.3 The problem of conceptual fuzziness
A second problem with the classical view is that
definitional structure entails that categories
have definite and distinct boundaries. In other
words, an entity either will or will not possess
the 'right' properties for category membership.
Indeed, this appears to be the case for many
categories. Consider the category ODD
NUMBER. As we learn at school, members of
this category are all those numbers that cannot
be divided by 2 without leaving a remainder: 1,
3, 5, 7, 9 and so on. This category has clearly
defined boundaries, because number is either

8.1.4 The problem of prototypicality
The third problem with the definitional view of
categories is related to the problem of concepual
fuzziness, but while the problem of conceptual
fuzziness concerns what happens at the boundaries of
a category, the problem of prototypicality conerns what
happens at the centre of a category. Findings from
experimental cognitive psychology reveal that
categories give rise to prototype or typicality effects.
For example, while people judge TABLE or CHAIR as
'good examples' or 'typical examples' of the category
FURNITURE, CARPET is judged as a less good
example. These asymmetries between category
members are called typicality effects.

8.1.5 Further problems
The problem of psychological reality relates to the fact
that there is no evidence for definitional structure in
psychological experiments. For example, we might
expect words with a relatively 'simple' definitional
structure or small set of features (like, say, man) to be
recognised
more
rapidly
in
word-recognition
experiments than words with a more 'complex'
definitional structure or greater number of features (like,
say, cousin). Possessing a concept is not dependent
upon knowing its definition. For example, it is possible
to have the concept WHALE while mistakenly believing
that it belongs to the category FISH rather than the
category MAMMAL.

8.2 Prototype theory
8.2.1 Principles of categorization
Prototype theory posits that there are two basic
principles that guide the formation of categories in the
human mind: (1) the principle of cognitive economy,
and (2) the principle of perceived world structure.
These principles together give rise to the human
categorisation system.
Principle of cognitive economy
This principle states that an organism, like a human
being, attempts to gain as much information as possible
about its environment while minimising cognitive effort
and resources. This cost-benefit balance drives
category formation. In other words, rather than storing
separate information about every individual stimulus

experienced, humans can group similar stimuli into
categories, which maintains economy in cognitive
representation.
Principle of perceived world structure
The world around us has correlational structure. For
instance, it is a fact about the world that wings most
frequently co-occur with feathers and the ability to fly
(as in birds), rather than with fur or the ability to
breathe underwater. This principle states that humans
rely upon correlational structure of this kind in order to
form and organise categories.

8.2.2 The categorisation system
These two principles give rise to the human
categorisation system. While the principle of cognitive
economy has implications for the level of detail or level
of inclusiveness with which categories are formed, the
principle of correlational structure has implications for
the representativeness or prototype structure of the
categories formed. Rosch suggests that this gives rise
to a categorisation system that has two dimensions: a
horizontal and a vertical dimension. This idea is
represented in Figure 8.1.

The vertical dimension relates to the level of inclusiveness
of a particular category: the higher up the vertical axis a
particular category is, the more inclusive it is. Consider the
category DOG in Figure 8.1. Relative to this category, the
category MAMMAL is higher up the vertical axis and
includes more members than the category DOG. The
category MAMMAL is therefore more inclusive than the
category DOG. The category COLLIE, however, is lower on
the vertical axis and has fewer members; this category is
less inclusive than the category DOG. In contrast, the
horizontal dimension relates to the category distinctions at
the same level of inclusiveness. Hence, while DOG and
CAR are distinct categories, they operate at the same level
of detail. In the next two subsections, we look in more detail
at the evidence for these two dimensions of categorisation.

8.2.3 The vertical dimension
The vertical dimension derives from the discovery by
Rosch and her colleagues (Rosch et al. 1976) that
categories can be distinguished according to level of
inclusiveness. Inclusiveness relates to what is
subsumed within a particular category. As we have
seen, the category FURNITURE is more inclusive than
the category CHAIR because it includes entities like
DESK and TABLE in addition to CHAIR. In turn,
CHAIR is more inclusive than ROCKING CHAIR
because it includes other types of chairs in addition to
rocking chairs. The category ROCKING CHAIR only
includes rocking chairs, and therefore represents the
least inclusive level of this category.

There is a level of inclusiveness that is optimal for human beings
in terms of providing optimum cognitive economy. This level of
inclusiveness was found to be at the mid-level of detail, between
the most inclusive and least inclusive levels: the level associated
with categories like CAR, DOG and CHAIR. This level of
inclusiveness is called the basic level, and categories at this level
are called basic-level categories. Categories higher up the
vertical axis, which provide less detail, are called superordinate
categories. Those lower down the vertical axis, which provide
more detail, are called subordinate categories. This is illustrated
in Table 8.3. In a remarkable series of experiments, basic-level
categories provided the most inclusive level of detail at which
members of a particular category share features in common. In
other words, while the superordinate level (e.g. MAMMAL) is the
most inclusive level, members of categories at this level of
inclusiveness share relatively little in common when compared to
members of categories located at the basic level of inclusiveness
(e.g. DOG).

Attributes
Rosch et al. (1976) found that the basic level is the level
at which humans are best able to list a cluster of
common attributes for a category. To investigate this,
Rosch and her colleagues gave subjects 90 seconds to
list all the attributes they could think of for each of the
individual items listed in a particular taxonomy. Six of the
taxonomies used by Rosch et al. are presented in Table
8.4. (It is worth pointing out to British English readers
that because Rosch's experiments were carried out in
the United States, some of the American English
expressions may be unfamiliar.)

Table 8.5 lists common attributes found for three of these
taxonomies. In the table, lower levels are assumed to have all
the attributes listed for higher levels and are therefore not
repeated. Table 8.5 illustrates the fact that subjects were only
able to provide a minimal number of shared attributes for
superordinate categories. In contrast, a large number of
attributes were listed as being shared by basic-level categories,
while just one or two more specific attributes were added for
subordinate categories. Hence, while subordinate categories
have slightly more attributes, the basic level is the most inclusive
level at which there is a cluster of shared attributes.

Motor movements
The experiment of Rosch and her colleagues revealed
that basic level categories were the most inclusive level
at which members of categories share motor
movements (physical interaction). While there are few
motor movements common to members of a
superordinate category, there are several specific motor
movements listed for entities at the basic level, while
entities at the subordinate level make use of essentially
the same motor movements. This provides further
evidence that the basic level is the most inclusive level,
this time with respect to common interactional
experiences. Rosch et al. sought to establish the most
inclusive level of categorisation at which shapes of
objects in a given category are most similar.

In order to investigate this, the researchers collected
around 100 images from sources like magazines and
books representing each object at each level in the
taxonomies listed in Table 8.4. The shapes were scaled
to the same size and then superimposed upon one
another. Areas of overlap ratios were then measured,
which allowed the experimenters to determine the
degree of similarity in shape. While objects at the
superordinate level are not very similar in terms of
shape, and while objects at the subordinate level are
extremely similar, the basic level was shown to the most
inclusive level at which object shapes are similar. In
other words, the basic level includes a much greater
number of instances of a category than the
superordinate level (for example, DOG versus COLLIE)
that can be identified on the basis of shape similarity.

Identification based on averaged shapes
Subjects were shown the shapes and provided with
superordinate, basic-level and subordinate terms to
which they were asked to match the shapes. Again,
although there is a greater degree of similarity at the
subordinate level, the basic level is more inclusive. The
absence of shape similarity at the superordinate level
compared to the evident shape similarity at the basic
level goes some way towards explaining why the basic
level is the optimum categorisation level for the human
categorization system.

Cognitive economy versus level of detail
The basic level of categorisation is the most important
level for human categorisation because it is the most
inclusive and thus most informative level. The
subordinate level is at least as informative as the basic
level, if not more so, given that it provides more detailed
information in addition to the information represented at
the basic level. When asked to list attributes of CAR and
SPORTS CAR, subjects typically listed more attributes
for SPORTS CAR than for CAR. This is because the
subordinate category SPORTS CAR is likely to be
identified with the same attributes as CAR, plus some
extra attributes specific to SPORTS CAR.

The reason why the basic level is the most salient
level of categorisation relates to the tension between
similarity of members of a category and the principle of
cognitive economy. While entities at the subordinate
level are most alike (rocking chairs have most in
common with other rocking chairs), different categories
at the subordinate level are also very similar (rocking
chairs are pretty similar to kitchen chairs). At the basic
level, on the other hand, while there are also
similarities within a particular category (all chairs are
pretty similar to one another), there are far fewer
between-category similarities (a chair is not that similar
to a table).

Crucially, for a category to achieve cognitive
economy (to provide the greatest amount of
information at the lowest processing cost), it
must share as many common within-category
attributes as possible. In intuitive terms, it is
easier to spot the differences between a chair
and a lamp than between a desk lamp and a
floor lamp. This demonstrates why the basic
level of categorisation is 'special': it is the level
which best reconciles the conflicting demands
of cognitive economy. Therefore the basic level
is the most informative level of categorisation.

This notion of cognitive economy has been described
in terms of cue validity. A particular cue – or attribute
– becomes more valid or relevant to a given category
the more frequently it is associated with members of
that category. Conversely, a particular attribute
becomes less valid or relevant to a category the more
frequently it is associated with members of other
categories. Thus 'is used for sitting on' has 'high cue
validity' for the category CHAIR, but 'is found in the
home' has low cue validity for the category CHAIR
because many other different categories of object can
be found in the home in addition to chairs.

Cue validity is maximised at the basic level, because
basic level categories share the largest number of
attributes possible while minimising the extent to which
these features are shared by other categories. This
means that basic-level categories simultaneously
maximise their inclusiveness (the vertical dimension)
and their distinctiveness (the horizontal dimension)
which results in optimal cognitive economy by
providing a maximally efficient way of representing
information about frequently encountered objects.

Perceptual salience
The basic level appears to be the most abstract
(that is, the most inclusive and thus the least
specific) level at which it is possible to form a
mental image. After all, we are unable to form an
image of the category FURNITURE without
imagining a specific item like a chair or a table: a
basic-level object. This is consistent with the
finding that averaged shapes cannot be identified
at the superordinate level as there are
insufficient similarities between entities at this
very high level of inclusiveness.

Based on a picture verification task, Rosch et al. (1976)
also found that objects are perceived as members of
basic-level categories more rapidly than as members of
superordinate or subordinate categories. In this
experiment, subjects heard a word like chair.
Immediately afterwards, they were presented with a
visual image. This enabled experimenters to measure
the reaction times of the subjects. It emerged that
subjects were consistently faster at identifying whether
an object matched or failed to match a basic level word
than they were when verifying images against a
superordinate or subordinate level word. This suggests
that in terms of perceptual verification, objects are
recognised more rapidly as members of basic-level
categories than other sorts of categories.

Language acquisition
Basic-level terms are among the first concrete
nouns to emerge in child language. This
investigation was based on a case study of a
single child, consisting of weekly two-hour
recordings dating from the initial period of
language production. All relevant utterances
were independently rated by two assessors in
order to determine whether they were
superordinate, basic or subordinate level terms.
The study revealed that the individual noun-like
utterances were overwhelmingly situated at the
basic level. This finding provided further support
for the primacy of the basic level.

Basic-level terms in language
The language system itself also reveals the primacy of the basic
level in a number of ways. Firstly, basic-level terms are typically
monolexemic: comprised of a single word-like unit. This contrasts
with terms for subordinate level categories which are often
comprised of two or more lexemes – compare chair (basic-level
object) with rocking chair (subordinate-level object). Secondly,
basic-level terms appear to occur more frequently in language use
than superordinate or subordinate level expressions. The
superordinate level (for example, VEHICLE) highlights the
functional attributes of the category (vehicles are for moving
people around), while also performing a collecting function
(grouping together categories that are closely linked in our
knowledge representation system). Subordinate categories, on
the other hand, fulfil a specificity function.

8.2.4 The horizontal dimension
The horizontal dimension of the categorisation system
(recall Figure 8.1) relates in particular to the principle of
perceived world structure. This principle states that the
world is not unstructured, but possesses correlational
structure. As Rosch points out, 'wings correlate with
feathers more than fur' (Rosch 1978: 253). The world
itself has structure, which provides constraints on the
kinds of categories that humans represent within the
cognitive system. One consequence of the existence of
correlational structure in the world is that cognitive
categories themselves reflect this structure: the category
prototype reflects the greater number of correlational
features.

Members of a category that are judged as highly
prototypical (most representative of that
category) can be described as category
prototypes. This feature of category structure was
investigated in a series of experiments reported
in Rosch (1975), which established that
prototypical members of a category were found
to exhibit a large number of attributes common
to many members in the category, while less
prototypical members were found to exhibit
fewer attributes common to other members of
the category. In other words, not only do
categories exhibit typicality effects (having more
or less prototypical members), category members
also exhibit family resemblance relations.

Family resemblance
Prototype structure, as exhibited by goodness-ofexample ratings, serves to maximise shared
information
contained
within
a
category.
'Prototypes appear to be those members of a
category that most reflect the redundancy
structure of the category as a whole'. In other
words, the more frequent a particular attribute is
among members of a particular category, the
more representative it is. The prototype structure
of the category reflects this 'redundancy' in terms
of repeated attributes across distinct members,
or exemplars. This entails that another way of
assessing prototype structure is by establishing
the set of attributes that a particular entity has.

Robins
are
judged
to
be
highly
prototypical: they possess a large number
of attributes found across other members
of the BIRD category. Conversely, ostriches,
which are judged not to be very good
examples of the category BIRD, are found to
have considerably fewer of the common
attributes found among members of the
category. Therefore, while OSTRICH and ROBIN
are representative to different degrees,
they nonetheless share a number of
attributes and thus exhibit a degree of
family resemblance.

8.2.5 Problems with prototype theory
It has been argued that prototype theory is
inadequate
as
a
theory
of
knowledge
representation. The study found that even a
'classical category' of this nature exhibits
typicality effects.
Prototype theory also suffers from the problem of
ignorance and error: it fails to explain how we can
possess a concept while not knowing or being mistaken
about its properties. The basis of this criticism is that a
concept with prototype structure might incorrectly
include an instance that is not in fact a member of that
category.

The third criticism that Laurence and Margolis
discuss is called the missing prototypes
problem: the fact that it is not possible to
describe a prototype for some categories. These
categories
include
'unsubstantiated'
(nonexistent) categories like US MONARCH and
heterogeneous categories like OBJECTS THAT WEIGH MORE
THAN A GRAM.
Lakoff (1987) therefore attempts to develop a
theory of cognitive models that might plausibly
explain the typicality effects uncovered by Rosch
and her colleagues. Lakoff's theory of cognitive
models avoids the problems that we summarised
above which follow from assuming Prototype
Theory as a model of knowledge representation.

8.3 The theory of idealised cognitive models
Lakoff argued that categories relate to idealised
cognitive models (ICMs). These are relatively
stable mental representations that represent
theories about the world. While ICMs are rich in
detail, they are 'idealised' because they abstract
across a range of experiences rather than
representing specific instances of a given
experience. ICMs guide cognitive processes like
categorization and reasoning. Categories of this
kind, which are constructed 'online' for local
reasoning, are constructed on the basis of preexisting ICMs.

8.3.1 Sources of typicality effects
Lakoff argues that typicality effects can arise in a range of
ways from a number of different sources. To illustrate,
consider the ICM to which the concept BACHELOR relates. An
individual's status as a bachelor is an 'all or nothing'
affair, because this notion is understood with respect to
the legal institution of MARRIAGE: the moment the marriage
vows have been taken, a bachelor ceases to be a
bachelor. The concept POPE, on the other hand, is primarily
understood with respect to the ICM of the CATHOLIC CHURCH
whose clergy are unable to marry. Clearly, there is a
mismatch between these two cognitive models: in the ICM
against which BACHELOR is understood, the Pope is 'strictly
speaking' a bachelor because he is unmarried. However,
the Pope is not a prototypical bachelor precisely because
the Pope is understood with respect to a CATHOLIC CHURCH ICM
in which marriage of Catholic clergy is prohibited.

Typicality effects due to cluster models
According to Lakoff, there is a second way in
which typicality effects can arise. This relates to
cluster models, which are models consisting of
a number of converging ICMs. The converging
models collectively give rise to a complex
cluster, which 'is psychologically more complex
than the models taken individually' (Lakoff 1987:
74). Lakoff illustrates this type of cognitive
model with the example of the category MOTHER,
which he suggests is structured by a cluster
model consisting of a number of different MOTHER
subcategories. These are listed below.

1. THE BIRTH MODEL: a mother is the person
who gives birth to the child.
2. THE GENETIC MODEL: a mother is the person
who provides the genetic material for the
child.
3. THE NURTURANCE MODEL: a mother is the
person who brings up and looks after the
child.
4. THE MARITAL MODEL: a mother is married to
the child's father.
5. THE GENEALOGICAL MODEL: a mother is a
particular female ancestor.