Cur nt Issues in Linguisti Theor o
-L
EDITED BY
Br> n'dj Rud ka- si n
JOH BEN) INS PUBLISHI C CO P NY

TOPICS IN COGNITIVE LINGUISTICS

AMSTERDAM STUDIES IN THE THEORY AND
HISTORY OF LINGUISTIC SCIENCE
General Editor
E.F. KONRAD KOERNER
(University of Ottawa)
Series IV - CURRENT ISSUES IN LINGUISTIC THEORY
Advisory Editorial Board
Henning Andersen (Buffalo, N.Y.); Raimo Anttila (Los Angeles)
Thomas V. Gamkrelidze (Tbilisi); Hans-Heinrich Lieb (Berlin)
J. Peter Maher (Chicago); Ernst Pulgram (Ann Arbor, Mich.)
E.Wyn Roberts (Vancouver, B.C.); Danny Steinberg (Tokyo)
Volume 50
Brygida Rudzka-Ostyn (ed.)
TOPICS IN COGNITIVE LINGUISTICS

TOPICS IN
COGNITIVE LINGUISTICS
Edited by
BRYGIDA RUDZKA-OSTYN
University of Leuven
JOHN BENJAMINS PUBLISHING COMPANY
AMSTERDAM/PHILADELPHIA
1988

Library of Congress Cataloging-in-Publication Data
Topics in cognitive linguistics / edited by Brygida Rudzka-Ostyn
p. cm. — (Amsterdam studies in the theory and history of linguistic science.
Series IV, Current issues in linguistic theory, ISSN 0304-0763; v. 50)
Bibliography: p.
Includes index.
1. Cognitive grammar. I. Rudzka-Ostyn, Brygida. II. Series.
P165.T65 1988
415--de 19 87-37495
ISBN 90 272 3544 9 (alk. paper) CIP
© Copyright 1988 - All rights reserved
No part of this book may be reproduced in any form, by print, photoprint, microfilm, or
any other means, without written permission of the copyright holders. Please direct all
enquiries to the publishers.

To the memory of my mother

Contents
Preface ix
I. Toward a Coherent and Comprehensive Linguistic Theory
An Overview of Cognitive Grammar
Ronald W. Langacker
A View of Linguistic Semantics
Ronald W. Langacker
The Nature of Grammatical Valence
Ronald W. Langacker
A Usage-Based Model
Ronald W. Langacker
II. Aspects of a Multifaceted Research Program
The Relation of Grammar to Cognition
Leonard Talmy
Where Does Prototypicality Come From?
Dirk Geeraerts
The Natural Category MEDIUM: An Alternative to Selection
Restrictions and Similar Constructs 231
Bruce W. Hawkins
Spatial Expressions and the Plasticity of Meaning 271
Annette Herskovits
Contrasting Prepositional Categories: English and Italian 299
John R. Taylor
The Mapping of Elements of Cognitive Space onto Grammatical
Relations: An Example from Russian Verbal Prefixation 327
Laura A. Janda
3
49
91
127
165
207

Vlll
CONTENTS
Conventionalization of Cora Locationals 345
Eugene H. Casad
The Conceptualisation of Vertical Space in English: The Case of Tall 379
Rene Dirven and John R. Taylor
Length, Width, and Potential Passing 403
Claude Vandeloise
On Bounding in Ik 429
Fritz Serzisko
A Discourse Perspective on Tense and Aspect in Standard Modern
Greek and English 447
Wolf Paprotte
Semantic Extensions into the Domain of Verbal Communication 507
Brygida Rudzka-Ostyn
Spatial Metaphor in German Causative Constructions 555
Robert Thomas King
Nahuatl Causative/Applicatives in Cognitive Grammar 587
David Tuggy
III. A Historical Perspective
Grammatical Categories and Human Conceptualization:
Aristotle and the Modistae 621
Pierre Swiggers
Cognitive Grammar and the History of Lexical Semantics 647
Dirk Geeraerts
References 679
Subject Index 695

Preface
Recent years have witnessed a growing interest in cognitive linguistics,
a framework aiming at an adequate account of the relationship between
language and cognition and as such involving human psychology,
interpersonal relations, culture, and a host of other domains. Our purpose has been
to present the theoretical premises of this framework and to explore its
descriptive and explanatory potential with respect to a wide range of
language phenomena. In pursuing this goal, we have frequently relied on
corpus analyses, intensive field work, and other means of empirical
verification. Crossing the boundaries of particular languages or language families
has lent additional support to the findings emerging from our research.
If one had to name a key notion of cognitive grammar, one would
certainly point to the dependence of linguistic structure on conceptualization
as well as the conceptualized perspective. Within this framework,
meanings are defined relative to conceptual domains, particular linguistic choices
are often found to hinge upon the vantage point from which a given
situation is viewed, and category boundaries are seen as fluctuating and
dependent on, among other things, the conceptualized experience or purpose.
This relativism extends to the very structure of the framework. The reader
will soon discover that not all contributors to the volume draw the same
distinctions.Neither do they use identical descriptive tools. Their choice and
nature vary with the purpose as well as perspective adopted.
This inherent flexibility renders the framework exceptionally receptive
to findings in other disciplines. Only some of the current interdisciplinary
crossovers could be signalled here; but they suffice to show the enormous
potential of cognitive grammar in capturing various facets of language. As
language is such a complex phenomenon, it can be adequately described
only when approached from different angles. By relativizing its own
methodology, cognitive linguistics can accommodate these different angles
readily and naturally.

X
PREFACE
To place cognitive grammar against a broader background, we have
explored some earlier linguistic theories. While adding another dimension
to our research, this exploration has unveiled interesting links between
present and past attempts at grasping the relation of language to cognition.
The project has materialized thanks to the help of several people. I am
indebted to Rene Dirven who came up with the idea and suggested to me
the role of editor. The conferences organized by him, first at the University
of Trier and now in Duisburg, have allowed many of us to come together
and discuss topics of common interest; and for this wonderful forum we
remain grateful.
As editor, I wish to thank the authors for their contributions, but also
for their good humor and the spirit of cooperation. Much of the present
volume is the fruit of an intensive exchange of ideas and materials, and also
numerous revisions. In this context, a special word of thanks is due to
Rene Dirven, Dirk Geeraerts, Bruce Hawkins and Pierre Swiggers, all of
whom offered to act as referees on several occasions. To Peter Kelly, I owe
a particular debt of gratitude for sharing with me his native-speaker
intuition.
Finally, I must record my great appreciation for my husband's
encouragement and help at all stages of the project.
Brygida Rudzka-Ostyn

PARTI
TOWARD A COHERENT AND
COMPREHENSIVE LINGUISTIC THEORY

An Overview of Cognitive Grammar
Ronald W. Langacker
University of California, San Diego
Orientation
Cognitive grammar (formerly "space grammar") is a theory of
linguistic structure that I have been developing and articulating since 1976.
Though neither finished nor formalized, it has achieved a substantial
measure of internal coherence and is being applied to an expanding array of
languages and grammatical phenomena (see Langacker 1981, 1982a , 1984,
1985, in press; Casad 1982a; Casad and Langacker 1985; Hawkins 1984;
Lindner 1981, 1982; Smith 1985; Tuggy 1981; Vandeloise 1984, 1985a).
These efforts have been prompted by the feeling tha* established theories
fail to come to grips in any sensible way with the real problems of language
structure, as they are based on interlocking sets of concepts, attitudes, and
assumptions that misconstrue the nature of linguistic phenomena and thus
actually hinder our understanding of them. It is therefore necessary to start
anew and erect a theory on very different conceptual foundations.
Cognitive grammar thus diverges quite radically from the mainstream
of contemporary linguistic theory, particularly as represented in the
generative tradition. The differences are not confined to matters of detail, but
reach to the level of philosophy and organizing assumptions. I will
succinctly sketch these differences as they pertain to the nature of linguistic
investigation, the nature of a linguistic system, the nature of grammatical
structure, and the nature of meaning. My presentation of the "orthodox"
position is admittedly a caricature; I state it without the necessary
qualifications for sake of brevity, and also to underscore the substantially different
spirit of the two approaches.
® Ronald W. Langacker

4
RONALD W. LANGACKER
With respect to the nature of linguistic investigation, orthodox theory
holds that language (or at least grammar) is describable as an algorithmic
system. Linguistics is thus a formal science akin to logic and certain
branches of mathematics (e.g. automata theory). Of paramount importance
is the construction of an all-embracing linguistic theory incorporating
explanatory principles; ongoing description is considered most valuable
when formulated in terms of current theory and directed towards the
testing and refinement of its predictions. Discrete categories and absolute
principles are sought, on the grounds that a theory should be maximally
restrictive and make the strongest possible claims. Moreover, economy is a prime
concern in formulating the grammar of a language: redundancy of
statement implies the loss of significant generalizations.
The cognitive grammar "heresy" sees biology as providing a better
metaphor for linguistic research than the formal sciences. While certain
aspects of language may be discrete and "algebraic", in general a language
is more accurately likened to a biological organism; our expectations
concerning the nature of revealing analysis and viable description must be
adjusted accordingly. For instance, absolute predictability is normally an
unrealistic expectation for natural language: much is a matter of degree,
and the role of convention is substantial. Considerations of economy must
cede priority to psychological accuracy; redundancy is plausibly expected in
the cognitive representation of linguistic structure, and does not in principle
conflict with the capturing of significant generalizations. Further, linguistic
theory should emerge organically from a solid descriptive foundation.
Preoccupation with theory may be deleterious if premature, for it stifles the
investigation of non-conforming phenomena and prevents them from being
understood in their own terms.
In the orthodox view, the grammar of a language consists of a number
of distinct "components". The grammar is conceived as a "generative"
device which provides a fully explicit enumeration of "all and only the
grammatical sentences" of the language. The linguistic system is
self-contained, and hence describable without essential reference to broader
cognitive concerns. Language may represent a separate "module" of
psychological structure.
Cognitive grammar views the linguistic system in a very different
fashion. It assumes that language evokes other cognitive systems and must be
described as an integral facet of overall psychological organization. The
grammar of a language is non-generative and non-constructive, for the

AN OVERVIEW OF COGNITIVE GRAMMAR
5
expressions of a language do not constitute a well-defined, algorithmically-
computable set. The grammar of a language simply provides its speakers
with an inventory of symbolic resources — using these resources to
construct and evaluate appropriate expressions is something that speakers do
(not grammars) by virtue of their general categorizing and problem-solving
abilities. Only semantic, phonological, and symbolic units are posited, and
the division of symbolic units into separate components is considered
arbitrary.
Orthodox theory treats grammar (and syntax in particular) as an
independent level or dimension of linguistic structure. Grammar (or at least
syntax) is considered distinct from both lexicon and semantics, and describe
able as an autonomous system. The independence of grammatical structure
is argued by claiming that grammatical categories are based on formal
rather than semantic properties. Speakers are capable of ignoring meaning
and making discrete well-formedness judgments based on grammatical
structure alone.
By contrast, cognitive grammar claims that grammar is intrinsically
symbolic, having no independent existence apart from semantic and
phonological structure. Grammar is describable by means of symbolic units
alone, with lexicon, morphology, and syntax forming a continuum of
symbolic structures. Basic grammatical categories (e.g. noun and verb) are
semantically definable, and the unpredictable membership of other classes
(those defined by occurrence in particular morphological or syntactic
constructions) does not itself establish the independence of grammatical
structure. Well-formedness judgments are often matters of degree, and reflect
the subtle interplay of semantic and contextual factors.
Finally, it is commonplace to reject a "conceptual" or "ideational"
theory of meaning as being untenable for the scientific investigation of
language. It is assumed instead that the meanings of linguistic expressions are
describable in terms of truth conditions, and that some type of formal logic
is appropriate for natural language. It is held that a principled distinction
can be made between semantics and pragmatics (or between linguistic and
extralinguistic knowledge), that semantic structure is fully compositional,
and that such phenomena as metaphor and semantic extension lie outside
the scope of linguistic description.
In the cognitive grammar heresy, meaning is equated with
conceptualization (interpreted quite broadly), to be explicated in terms of cognitive
processing. Formal logic is held to be inadequate for the description of

6
RONALD W. LANGACKER
semantic structure, which is subjective in nature and incorporates
conventional "imagery" — defined as alternate ways of construing or mentally
portraying a conceived situation. Linguistic semantics is properly considered
encyclopedic in scope: the distinction between semantics and pragmatics is
arbitrary. Semantic structure is only partially compositional, and
phenomena like metaphor and semantic extension are central to the proper
analysis of lexicon and grammar.
Meaning and Semantic Structure
An "objectivist" view of meaning has long been predominant in
semantic theory. Rigorous analysis, it is maintained, cannot be based on
anything so mysterious and inaccessible as "concepts" or "ideas"; instead,
the meaning of an expression is equated with the set of conditions under
which it is true, and some type of formal logic is deemed appropriate for the
description of natural language semantics. Without denying its
accomplishments, I believe the objectivist program to be inherently limited and
misguided in fundamental respects: standard objections to the ideational view
are spurious, and a formal semantics based on truth conditions is attainable
only by arbitrarily excluding from its domain numerous aspects of meaning
that are of critical linguistic significance (cf. Chafe 1970: 73-75; Langacker
in press: Part II; Hudson 1984).
Cognitive grammar explicitly equates meaning with
"conceptualization" (or "mental experience"), this term being interpreted quite broadly.
It is meant to include not just fixed concepts, but also novel conceptions
and experiences, even as they occur. It includes not just abstract,
"intellectual" conceptions, but also such phenomena as sensory, emotive, and
kinesthetic sensations. It further embraces a person's awareness of the
physical, social, and linguistic context of speech events. There is nothing
inherently mysterious about conceptualization: it is simply cognitive
processing (neurological activity). Entertaining a particular conceptualization, or
having a certain mental experience, resides in the occurrence of some
complex "cognitive event" (reducing ultimately to the coordinated firing of
neurons). An established concept is simply a "cognitive routine", i.e. a
cognitive event (or event type) sufficiently well "entrenched" to be elicited as
an integral whole.
Cognitive grammar embraces a "subjectivist" view of meaning. The
semantic value of an expression does not reside solely in the inherent prop-

AN OVERVIEW OF COGNITIVE GRAMMAR
7
erties of the entity or situation it describes, but crucially involves as well the
way we choose to think about this entity or situation and mentally portray
it. Expressions that are true under the same conditions, or which have the
same reference or extension, often contrast in meaning nonetheless by
virtue of representing alternate ways of mentally construing the same
objective circumstances. I would argue, for example, that each pair of sentences
in (1) embodies a semantic contrast that a viable linguistic analysis cannot
ignore:
(1) (a) This is a triangle.
(a') This is a three-sided polygon.
(b) The glass is half-empty.
(b>) The glass is half-full.
(c) This roof slopes upward,
(c') This roof slopes downward.
(d) Louise resembles Rebecca,
(d') Rebecca resembles Louise.
(e) Russia invaded Afghanistan.
(e') Afghanistan was invaded by Russia.
(f) I mailed a package to Bill.
(f) I mailed Bill a package.
I use the term "imagery" to indicate our ability to mentally construe a
conceived situation in alternate ways (hence the term does not refer
specifically or exclusively to sensory or visual imagery (cf. Kosslyn 1980; Block
1981)). A pivotal claim of cognitive grammar is that linguistic expressions
and grammatical constructions embody conventional imagery, which
constitutes an essential aspect of their semantic value. In choosing a particular
expression or construction, a speaker construes the conceived situation in a
certain way, i.e. he selects one particular image (from a range of
alternatives) to structure its conceptual content for expressive purposes. Despite
the objective equivalence of the sentence pairs in (1), the members of each
are semantically distinct because they impose contrasting images on the
conceived situation.
At this point, I will confine my discussion of imagery to a single
example, namely the semantic contrasts distinguishing the universal quantifiers
of English. It is intuitively obvious that the sentences in (2) have subtly
different meanings despite their truth-functional equivalence:

8
RONALD W. LANGACKER
(2) (a) All cats are playful.
(b) Any cat is playful.
(c) Every cat is playful.
(d) Each cat is playful.
These sentences share the conceptual content of a property (playfulness)
being attributed to all the members of a class (the set of cats). They
nevertheless employ distinct images with respect to how one "reaches" or
"mentally accesses" the class members for this purpose. An informal sketch
of the contrasts is attempted in Fig. 1, where circles indicate class members,
a box represents the class as a whole, and heavy lines enclose the entity
designated by the quantifier. All refers to the class collectively, as an
undifferentiated mass, and makes a blanket attribution of the property to its
members. The other three quantifiers each refer to a single, arbitrary class
member, but this member is conceived as being selected in such a fashion
that the property attributed to it is similarly attributed to all the other
members. The image conveyed by any is one of random selection: if one chooses
a member at random, it will invariably display the property in question.
Every and each are alike (and contrast with all) in attributing the property
to the full set of class members on an individual (rather than a collective)
basis. The difference between them is that each further suggests that the
members are examined sequentially — one at a time — for this purpose.
(a)
ALL
ANY
l-ooo-l
c) EVERY
|-o t
1 j
??'"
1 ♦
1
1
1
1
1
1
(d) EACH
|-o <
f J
1
1
1
* 1
1
1
??•■■
1 A
i
I
1
1
1
1
Figure 1
This analysis is at best preliminary and suggestive, but the putative
imagic contrasts account for a variety of otherwise puzzling phenomena.
First, they explicate the clear differences in meaning among the sentences
in (3) (first cited by George Lakoff):
(3) (a) Tonight you can see any star in the Milky Way.
(b) Tonight you can see every star in the Milky Way.
(c) Tonight you can see each star in the Milky Way.

AN OVERVIEW OF COGNITIVE GRAMMAR
9
Speakers readily agree that (3)(a) describes the ability to see a single star:
if a particular star is in the Milky Way, tonight you can see it, whichever
one you choose to look at. In (3)(b), the stars of the Milky Way are viewed
simultaneously, but nevertheless stand out as individuals. This
individuation is even stronger in (3)(c), which evokes the image of the viewer shifting
his gaze across the sky, looking first at one star, then at another, and so on.
The analysis further explains the grammatical behavior of the
quantifiers, for instance the ability of any, every, and each (but not all) to occur
with a singular count noun or the pro form one:
(4) (a) any star
(b) every star
(c) each star
(d) * all star
(5) (a) any one of those stars
(b) every one of those stars
(c) each one of those stars
(d) * all one of those stars
Moreover, because they specify individual attribution of the property to
multiple class members, every and each construe the class with a greater
degree of individuation than either all (which views the class collectively) or
any (which specifically invokes only a single, randomly selected member);
the former two quantifiers specifically highlight the status of the class as an
aggregate of distinct individuals. It is natural, then, that mass nouns
(including plurals) should occur with only all and any, since the hallmark of a
mass is internal homogeneity:
(6) (a) all water
(b) any water
(c) * every water
(d) *each water
(7) (a) all cats
(b) any cats
(c) * every cats
(d) * each cats
The analysis also explains the judgments in (8) and (9).
(8) (a) He examined each one in turn,
(b) He examined every one in turn.

10
RONALD W. LANGACKER
(c) ?He examined all in turn.
(d) *He examined any one in turn.
(9) (a) All cats have something in common.
(b) ?Every cat has something in common.
(c) *?Each cat has something in common.
(d) * Any cat has something in common.
In turn specifies tliat the examination described in (8) affects class members
individually and sequentially. This is fully compatible with each and every
(and reinforces something suggested by the former). Sentence (8)(c) is a bit
peculiar, since the collective construal imposed by all is at cross-purposes
with the individuating force of the adverbial, i.e. there is a certain amount
of tension between the images evoked by different components of the
expression. In (8)(d) this conflict in imagery is even more pronounced,
since any specifically picks out only a single member of the class for
attribution of the property, while in turn implies the participation of multiple
individuals. The semantic effect of have something in common runs counter to
that of in turn, for the notion of commonality requires the simultaneous
conception of class members. The well-formedness judgments in (9) are
thus essentially the inverse of those in (8). The collective character of all
renders (9)(a) unproblematic, while the individuation suggested by every
and each accounts for the relative infelicity of (9)(b) and (c), greater (as
expected) in the case of each. However the sentence with any is once again
the least acceptable, and for the same reason as before: only a single class
member is specifically selected (at random) for attribution of the property,
but have something in common (like in turn) implies access to multiple
members.
The notion of sequential examination may further account for the fact
that each is non-generic, i.e. a sentence like (2)(d) is only used with respect
to some restricted subset of class members identifiable from the context
(not the set of cats as a whole). The reason, perhaps, is that sequential
examination is not easily conceived as providing mental access to all
members of an open-ended, indefinitely expandable class.
This is one example of how the lexical and grammatical resources of a
language embody conventional imagery, which is an inherent and essential
aspect of their semantic value. Two expressions may be functionally
equivalent and serve as approximate paraphrases or translations of one another,
and yet be semantically distinct by virtue of the contrasting images they

AN OVERVIEW OF COGNITIVE GRAMMAR
11
incorporate. The imagery is conventional because the symbolic elements
available to speakers are language-specific and differ unpredictably —
languages simply say things in different ways, even when comparable thoughts
are expressed, as exemplified by the sentences in (10).
(10) (a) I'm cold.
(b) Pai froid.
(c) Mir ist kalt.
In this respect the cognitive grammar view of linguistic semantics is
Whorfian or relativistic. However it is not Whorfian in the sense that this imagery
is taken as imposing powerful or even significant constraints on how or
what we are able to think. I take its influence to be fairly superficial, a
matter of how we package our thoughts for expressive purposes. We shift from
image to image with great facility, even within the confines of a single
sentence, and freely create new ones when those suggested by linguistic
convention do not satisfy our needs.
The Nature of a Grammar
The grammar of a language is characterized as a "structured inventory
of conventional linguistic units". By "unit", I mean a thoroughly mastered
structure (i.e. a cognitive routine). A unit may be highly complex
internally, yet it is simple in the sense of constituting a prepackaged assembly
that speakers can employ in essentially automatic fashion, without
attending to the details of its composition. Examples of units include a lexical
item, an established concept, or the ability to articulate a particular sound
or sound sequence. The units comprising a grammar represent a speaker's
grasp of linguistic convention. I do not limit the term "conventional" to
structures that are arbitrary, unmotivated, or unpredictable: linguistic
structures form a gradation with respect to their degree of motivation, and
no coherence or special status attaches to the subclass of structures that are
fully motivated or to those that are fully arbitrary; speakers operate with
the entire spectrum of structures as an integrated system. Finally, this
inventory of conventional units is "structured" in the sense that some units
function as components (subroutines) of others.
Only three basic types of units are posited: semantic, phonological,
and symbolic. Symbolic units are "bipolar", consisting in the symbolic
relationship between a semantic unit (its "semantic pole") and a phonolog-

12
RONALD W. LANGACKER
ical unit (its "phonological pole"). I claim that lexicon, morphology, and
syntax form a continuum of symbolic structures divided only arbitrarily into
separate components of the grammar. Symbolic units are by no means a
homogeneous class — they vary greatly along certain parameters (e.g.
specificity, complexity, entrenchment, productivity, regularity) — but the
types grade into one another, and do not fall naturally into discrete, non-
overlapping blocks. Grammar (i.e. morphology and syntax) is
accommodated solely by means of symbolic units. The intrinsically symbolic
character of grammatical structure represents a fundamental claim whose import
and viability will be examined in the remainder of this paper.
Treating grammar as symbolic in nature (and not a separate level or
autonomous facet of linguistic structure) enables us to adopt the highly
restrictive "content requirement", which holds that no units are allowed in
the grammar of a language apart from (i) semantic, phonological, and
symbolic units that occur overtly; (ii) "schemas" for the structures in (i); and
(iii) "categorizing relationships" involving the structures in (i) and (ii). For
example, the phonological sequence [tip] occurs overtly, so it is permitted
by (i). The syllable canon [CVC] is "schematic" for [tip] and numerous
other phonological units (i.e. it is fully compatible with their specifications
but characterized with a lesser degree of specificity), so (ii) allows its
inclusion in the grammar. Permitted by (iii) is the categorizing relationship
between the phonological schema [CVC] and a specific sequence like [tip]
that "instantiates" or "elaborates" this schema. I will indicate this
relationship in the following way: [[CVC] —> [tip]], where a solid arrow stands for
the judgment that one structure is schematic for another, and square
brackets enclose a structure with the status of a unit. Hence the categorizing
relationship [[CVC] —> [tip]] is a complex unit containing the simpler units
[CVC] and [tip] as substructures. The schema [CVC] defines a phonological
category, and [[CVC] —> [tip]] characterizes [tip] as an instance of this
category.
The effect of the content requirement is to rule out many sorts of
arbitrary descriptive devices routinely employed in other frameworks. For
example, it rules out the use of "dummies" having neither semantic nor
phonological content, invoked solely to drive the machinery of formal
syntax. Also proscribed is any appeal to arbitrary diacritics or contentless
features (more about this later). It further precludes the derivation of an overt
structure from a hypothetical underlying structure of radically different
character (e.g. the derivation of a passive from an underlying active). There

AN OVERVIEW OF COGNITIVE GRAMMAR
13
is a real sense in which the content requirement is far more constraining
than the conditions, principles, and restrictions which continue to
proliferate (and often evaporate) in the theoretical literature. If valid, it will help
immeasurably to ensure the naturalness of linguistic descriptions.
In describing the grammar of a language as an "inventory'' of
conventional units, I refer to my conception of a grammar as being non-generative
and non-constructive. Specifically rejected is the notion that a grammar can
be regarded as an algorithmic device serving to generate (or give as output)
"all and only the grammatical sentences of a language", at least if the
description of a sentence is taken as including a full semantic
representation. The reason for this rejection is that the full set of expressions is
neither well-defined nor algorithmically computable unless one imposes
arbitrary restrictions on the scope of linguistic description and makes gratuitous
(and seemingly erroneous) assumptions about the nature of language.
Treating the grammar as a generative, constructive device requires (i) the
assumption that well-formedness is absolute, not a matter of degree; (ii) the
exclusion of contextual meaning from the scope of semantic description;
(iii) the omission of metaphor and figurative language from the coverage of
a grammar; (iv) the supposition that semantic structure is fully
compositional; and (v) the claim that a motivated distinction can be made between
semantics and pragmatics (or between linguistic and "extra-linguistic"
knowledge). These positions are commonly taken, but not because the facts
of language use cry out for their adoption. Rather, they are adopted
primarily for methodological reasons, there being no other way to make
language appear to be a self-contained, algorithmically-describable system.
I will simply comment that the convenience to the theorist of positing a self-
contained formal system does not constitute a valid argument for its factual
correctness. I opt for a cognitively and linguistically realistic conception of
language over one that achieves formal neatness at the expense of
drastically distorting and impoverishing its subject matter.
Hence the grammar of a language is not conceived as a generative or
algorithmic device, nor does it construct any expressions or give them as
outputs — it is simply an inventory of symbolic resources. It is up to the
language user to exploit these resources, and doing so is a matter of problem-
solving activity involving categorizing judgments. The results of these
operations draw upon the full body of a speaker's knowledge and cognitive
ability, and are thus not algorithmically computable by any limited,
self-contained system.

14
RONALD W. LANGACKER
The basic scheme is sketched in Fig. 2(a). The language user brings
many kinds of knowledge and abilities to bear on the task of constructing
and understanding a linguistic expression; these include the conventional
symbolic units provided by the grammar, general knowledge, knowledge of
the immediate context, communicative objectives, esthetic judgments, and
so on. Both the speaker and the addressee face the "coding" problem:
given the resources at their disposal, they must successfully accommodate a
"usage event". The semantic pole of the usage event is identified as the
detailed conceptualization that constitutes the expression's full contextual
value, i.e. how it is actually understood in context. The phonological pole
of the usage event is the actual vocalization employed to symbolize this
conception, in all its phonetic detail. Roughly, the speaker starts from the
conceptualization and must arrive at the proper vocalization, while the
addressee proceeds in the opposite direction (encoding vs. decoding). However
both of them, by employing the varied resources at hand, must deal with
the full, bipolar usage event for its occurrence to amount to a meaningful
symbolic experience. Doing so necessarily involves categorization and
problem-solving more generally.
(a)
symbolic units (G)
general knowledge
knowledge of context
communicative objectives
esthetic judgment
etc.
CODING
problem-solving
categorization
RESOURCES
USAGE EVENT
(o)
other
symbolic j_
units L^kh
(inactive)
SU
GRAMMAR
^r^
)~
(COMPOSITIONAL VALUE)
Figure 2
I USAGE \
' **1 EVENT I
ACTUAL
CONTEXTUAL
VALUE
One facet of this coding operation, namely the relation between the
grammar and the usage event, is diagrammed in Fig. 2(b). In arriving at the
usage event, and evaluating it with respect to linguistic convention, the

AN OVERVIEW OF COGNITIVE GRAMMAR
15
speaker or addressee must select a particular array of symbolic units (SU)
and activate them for this purpose; I take it that the cognitive reality of
linguistic units (and their activation in language use) is self-evident and uncon-
troversial. Moreover, since the usage event is identified as the utterance
itself and how it is actually understood in context, it is obviously real — by
definition the usage event does in fact occur. The only facet of Fig. 2(b)
whose status is uncertain is the intervening structure, representing the
"compositional value" of the expression, i.e. the value that could, in
principle, be obtained by algorithmic computation based solely on the
conventional values of the symbolic units employed. It is not unlikely that the
language user computes this compositional value as one step in the coding
process leading to the full usage event. Whether he does so consistently or only
on certain occasions is something we presently have no way of knowing.
What is, however, clear is that the usage event is not in general
equivalent to the compositional value. It is virtually always more specific than
anything strictly predictable from established symbolic units, and thus
represents an elaboration or "specialization" of the compositional value; or
else it conflicts in some way with the compositional value and thus
constitutes an "extension" from it. This departure of the usage event from the
hypothetical compositional value, whether by elaboration or by extension,
reflects the contribution to the coding process of extra-grammatical
resources deployed by the language user.
An example of extension is a novel metaphor, e.g. the use of cabbage
harvester to designate a guillotine. The compositional semantic value of this
expression is roughly 'something that harvests cabbage'; its actual,
contextual semantic value is 'guillotine'; the comparison of these two values,
serving to register their points of similarity and conflict (and thus responsible
for the expression being perceived as metaphoric), is a type of categorizing
judgment; and the symbolic units activated in the coding process include
the lexical items cabbage and harvester, as well as the schematic symbolic
unit representing the relevant pattern of compound formation.
As an example of specialization, suppose I show you a new gadget,
used to sharpen chalk, and refer to it as a chalk sharpener. The
compositional value of this expression is simply 'something that sharpens chalk', but
in context — where you actually see the gadget — your understanding of
this novel expression is far more detailed: you see that it is a mechanical
device, and not a person; you note its approximate size; you may observe
how it operates; and so on. All of this constitutes the expression's contex-

16
RONALD W. LANGACKER
tual semantic value, the semantic pole of the usage event. As is typically the
case, the compositional value substantially "underspecifies" this contextual
value, i.e. the former is schematic for the latter in the categorizing
relationship between the two.
Suppose, now, that the gadget I have shown you becomes widely used,
and that the term chalk sharpener establishes itself as the conventional term
for this type of object. The conventional semantic value of this lexical item
now includes many of those specifications that represented
non-compositional facets of its contextual value when the expression first occurred.
(This is precisely what happened with pencil sharpener, which means not
just 'something that sharpens pencils', but is further understood as
indicating a particular sort of mechanical device.) The implication of this type of
development ought to be apparent: since the non-compositional aspects of
an expression's meaning are part of its contextual value (i.e. how it is
actually understood) the very first time it occurs, and further become part of its
conventional value when it is established as a unit in the grammar, it is
pointless (indeed misguided) to arbitrarily exclude these facets of its
meaning from the domain of linguistic semantics. Moreover, conventionalization
is a matter of degree; there is no particular threshold at which
non-compositional specifications suddenly undergo a change in status from being extra-
linguistic to being conventional and hence linguistic. The process is gradual,
and is initiated when the expression is first employed.
This discrepancy between compositional and contextual value is not
limited to short expressions of the sort that conventionalize as lexical items;
it is characteristic of novel sentences as well. Far more is contributed to the
understanding of a sentence by context and general knowledge (and far less
by compositional principles) than is normally acknowledged. For instance,
(1 l)(a) would generally be interpreted in the fashion of Fig. 3(a), but given
the appropriate context it could easily be used and understood for any of
the situations in 3(b)-(g) and indefinitely many others.
(11) (a) The ball is under the table.
(b) He is barely keeping his head above the water.
While (ll)(b) evokes the conception of somebody struggling to keep afloat
while swimming, this conception cannot be derived by regular
compositional principles from the conventional meanings of its component lexical
items; in another context, the sentence could be understood in a radically
different way. (Imagine a race over the ocean by helicopter, where the con-

AN OVERVIEW OF COGNITIVE GRAMMAR
17
testants must transport a severed head, suspended by a rope from the
helicopter, from the starting line to the finish; a contestant is disqualified if
the head he is carrying ever dips below the water's surface.) An
expression's compositional value owes nothing to context, so in cases like these it
must be sufficiently abstract (schematic) to be compatible with all the
countless contextual interpretations the expression might receive — and
that is very abstract indeed. Thus, if semantics is restricted to what is
algorithmically computable from linguistic units, the resulting semantic
representations will be so limited and impoverished relative to how expressions
are actually understood that we would hardly recognize them as reasonable
approximations to their meanings. The domain of semantic analysis can
certainly be defined in this way, but one must then question whether its
content, so delimited, is independently coherent or worthy of serious interest.
Figure 3
In short, the dictum that linguistic semantics is fully compositional does
not rest on empirical observation, but is rather a matter of a priori
definition by theorists who wish to consider language as a self-contained formal
system. Semantic structure is rendered compositional simply by defining
non-compositional aspects of meaning as belonging to pragmatics or extra-
linguistic knowledge. The position of cognitive grammar is that this
distinction is arbitrary, and that semantics is only partially compositional. There
are indeed patterns of composition, and these are described at the semantic
pole of the conventional units representing grammatical constructions.

18
RONALD W. LANGACKER
However, the meaning of a complex expression (whatever its degree of
conventionalization) is recognized as a distinct entity not in general
algorithmically derivable from the meanings of its parts. Its compositional
value (assuming for sake of discussion the coherence of this notion) need
not always be separately computed, and if it is, this computation is only one
step in arriving at how the expression is actually understood.
Grammar as a Symbolic Phenomenon
The "autonomy" of grammar (and of syntax in particular) is a
fundamental tenet of contemporary theoretical orthodoxy (cf. Newmeyer 1983).
Grammar (or at least syntax) supposedly constitutes an independent level
or dimension of linguistic structure, describable as an autonomous
"component" within the linguistic system. The remainder of this paper addresses
the autonomy issue and argues for the viability of a radical alternative:
grammar is inherently symbolic, and therefore distinct from neither lexicon
nor semantics.
In view of the pivotal nature of this issue, it is surprising how little
attention has been devoted to clarifying precisely what it means for
grammatical structure to be autonomous, or to specifying what form a
conceivable alternative might take. For our purposes here, I will understand the
autonomy thesis as including the following points: (i) Grammatical
description requires the postulation of tree-like structures which represent an
independent facet of linguistic organization; commonly assumed are phrase
trees such as Fig. 4, which specify constituency, linear ordering, and
category membership (the latter through node labels), (ii) Lexical items are
"inserted" into such trees, which themselves have neither semantic nor
phonological content; rules of semantic and phonological interpretation
apply to the content provided by the inserted lexical items, and take
grammatical tree structure into account, but these trees per se are not semantic
or phonological objects, (iii) Being susceptible to neither semantic nor
phonological characterization, a distinct set of grammatical categories must
be posited; the class memberships of a lexical item are specified by syntactic
features (or some comparable diacritic device).

AN OVERVIEW OF COGNITIVE GRAMMAR
19
Figure 4
The proposed alternative does not employ phrase trees and rejects
points (i)-(iii). It claims that grammar (both morphology and syntax) is
describable using only symbolic elements, each of which has both a
semantic and a phonological pole. The symbolic units characterizing grammatical
structure form a continuum with lexicon: while they differ from typical
lexical items with respect to such factors as complexity and abstractness, the
differences are only a matter of degree, and lexical items themselves range
widely along these parameters. I will therefore begin my presentation of the
alternative view with a brief discussion of lexical units.
Lexical items vary greatly in their internal complexity. We are
concerned here in particular with "symbolic complexity", i.e. whether an item
decomposes into smaller symbolic units. Those which do not, and are
consequently minimal from the symbolic standpoint, are known as
"morphemes". Examples are given in Fig. 5(a), which illustrates the notations I
will use here for symbolic units: the semantic pole is shown at the top, its
content indicated by capital letters (thus SHARP abbreviates the meaning
of sharp); the phonological pole is shown at the bottom (orthographic
representations suffice for present purposes); the line between the two poles
stands for their symbolic relationship; and the box enclosing them indicates
that the symbolic structure as a whole has the status of a conventional unit.

20
RONALD W. LANGACKER
(a)
SHARP!
sharp]
IPENCII
pencil
J
en!
-en]
|er
|-er
ft)
SHARP-EN
sharp-en
/
SHARP
sharp
x
EN
-en
composite
structure
component
structures
(c)
"exploded" format
[[SHARP
Jsharp
-en)
"compacted"
format
(d)
| [SHARP
|| sharp
~en~|
-en|
er]
-erl
(e)
PENCIL^
pencil]
SHARP
[sharp
EN
-enl
ER
-er
Figure 5
The majority of lexical units are morphemically complex, as illustrated
for sharpen in Fig. 5(b). I will speak of the constituent morphemes,
[SHARP/sharp] and [EN/-en], as "component structures". The structure at
the top in 5(b), what I call the ''composite structure", represents the
semantic and phonological value of the expression as an integral whole. I do not
assume full compositionality for either the semantic or the phonological
pole, so the composite structure need not be strictly computable from its
components; possible discrepancies between the compositional value and
the actual composite structure are not our immediate concern, however. To
the extent that an expression does approach full compositionality, the
composite structure reflects the "integration", at each pole, of the contents of
the two component structures. Thus [SHARP-EN], the composite semantic
structure, reflects the integration of [SHARP] and [EN], while the
composite phonological structure [sharp-en] results from the integration of [sharp]
and [-en]. Fig. 5(b) depicts the complex expression in an "exploded"
format, showing the composite structure separately from its components. The
"compacted" representation of 5(c) is adopted for convenience as an
abbreviatory notational variant: it implies both component and composite
structures, even though this latter is not shown individually. There is of
course no intrinsic limit on the symbolic complexity of lexical items, since a
composite structure may function as one component of a larger expression
at a higher level of organization. Fig. 5(d) diagrams the integration of
sharpen and -er to form the derived noun sharpener. Sharpener in turn combines
with pencil to form the compound pencil sharpener, as seen in 5(e). All of
these expressions are conventional units of English.

AN OVERVIEW OF COGNITIVE GRAMMAR
21
Lexical units also vary greatly along the parameter of schematicity, i.e.
they are characterized at differing levels of precision and detail. This is
most obvious at the semantic pole. The semantic pole of bird, for instance,
is schematic for that of sparrow, the categorizing unit [[BIRD] -»
[SPARROW]] indicates that the specifications of [SPARROW] achieve a finer
grain (or higher resolution) than those of [BIRD], and represents the
conventional judgment that sparrows instantiate the bird category. There is
similarly a relationship of schematicity between two basic senses of cat, that
which indicates any feline, and that which contrasts with kitten and
indicates an adult feline in particular. The grammar of English thus includes the
categorizing unit shown in Fig. 6(a): the conventional symbolic structure
[FELINE/cat] is schematic with respect to the equally conventional
symbolic unit [ADULT FELINE/cat]. Note that the locus of elaboration is
confined to the semantic pole (the two symbolic units are identical at the
phonological pole).
(a)
FELINE
cat
ADULT
FELINE
cat
PL"
| -lar
" PL
-ler J
Figure 6
Variation along the parameter of schematicity is also observable for
lexical units at the phonological pole. Depending on vowel harmony, for
instance, the Turkish plural morpheme appears as either -lar or -ler, hence
[PL/-lar] and [PL/-ler] are symbolic units that occur overtly. Presumably
speakers capture the obvious generalization by extracting a schema to
represent the commonality of these two expressions; this schema can be given
as [PL/-lAr], where A is schematic for a and e (i.e. it is characterized simply
as a non-high, unrounded vowel). The grammar of Turkish thus includes
the units and categorizing relationships shown in Fig. 6(b). To
accommodate the special root structure of Arabic, we can posit phonologically
schematic symbolic units such as [WRITE/...k...t...b...], where '...' indi-

22
RONALD W. LANGACKER
cates the possible occurrence of a vowel (but does not specify its quality).
We see, then, that lexical units manifest varying degrees of schematic-
ity at either the semantic or the phonological pole. Some symbolic units are
schematic at both poles, e.g. Turkish [PLAlAr] ([PL] designates a set of
entities belonging to the same category, but this category is identified only
schematically). I suggest that schematicity, at one or both poles, is precisely
what characterizes those symbolic units responsible for grammatical
structure. Though a gradation can be observed, essential grammatical units are
highly schematic at both poles. Moreover, the greater the schematicity of a
symbolic unit, the more likely it is to fall within the traditionally-recognized
realm of grammar (as opposed to lexicon).
In Part II of Langacker in press, I try to demonstrate that a subjectivist
view of meaning which properly accommodates conventional imagery
makes possible a descriptively adequate and revealing semantic
characterization of basic grammatical classes (e.g. noun, verb, adjective); reasonably
precise definitions are offered that are grounded in plausible assumptions
about cognitive processing. Here I will simply assume that such descriptions
are possible in principle, and show where this assumption takes us.
If nouns, verbs, and other basic categories are susceptible to notional
definition, we may posit a schematic symbolic unit serving to characterize
each class: there will be a "noun schema" expressing the commonality of all
nouns, a "verb schema" defining the class of verbs, and so on. Such
schemas are maximally schematic at each pole. Phonologically, we can say
little more about nouns or verbs as a class than that they have "some
phonological content"; I use [X] and [Y] to indicate these highly schematic
phonological structures in Fig. 7(a)-(b). At the semantic pole, I claim that
every noun designates a "thing", while every verb designates a "process";
note that "thing" and "process" are technical terms whose import is highly
abstract (the class of "things", for instance, is not restricted to physical
objects). The membership of a lexical unit in one of these categories is
represented in the grammar not by any diacritic or syntactic feature, but rather
by a categorizing unit. The unit depicted in Fig. 7(c) specifies that pencil
instantiates the noun schema, and is therefore a member of the class that
this schema defines. Similarly, the categorizing unit in 7(d) indicates that
write is a verb by virtue of bearing an elaborative relationship to the verb
schema.

AN OVERVIEW OF COGNITIVE GRAMMAR
23
(a)
(b)
[THING]
I X 1
[PROCESS
1 Y
noun (c)
s chema
] verb
1 s chema
1 THING]
1 x 1
, * ,
PENCIL
[pencilj
1 (d)
[PROCESS
1 Y 1
f
IwriteI
|write)1
Figure 7
We see, then, that this framework accommodates basic grammatical
categories and category membership within the restrictive confines of the
content requirement: nothing has been invoked other than specific
symbolic units (e.g. pencil), schematic symbolic units (e.g. the noun schema),
and categorizing relationships between the two. What about grammatical
rules and constructions? These are not distinguished in cognitive grammar,
but are treated instead as alternate ways of regarding the same entities,
namely symbolic units that are both schematic and complex (in the sense of
having smaller symbolic units as components). I refer to these entities as
"constructional schemas".
Fig. 8(a) represents the constructional schema for deverbal nominali-
zations in -er {talker, swimmer, complainer, painter, mixer, opener, and so
on). Despite the abbreviatory notations, this structure is a complex
symbolic unit in which component structures are integrated to form a composite
structure at both the semantic and the phonological pole. One component
structure is the verb schema; semantically it designates a process of
unspecified character, while phonologically its content is maximally
schematic. The other component structure is the morpheme -er, which is
specific phonologically but semantically abstract (it designates a "thing"
identified only by the role it plays in a schematic process). The composite
structure (not separately shown in the compacted format of 8(a)) can be
abbreviated as follows: [PROCESS-ER/Y-er]. Semantically, it designates a
thing, and the process in which it plays a specified role is equated with that
designated by the schematic verb stem [PROCESS/Y]. Phonologically, it
specifies the suffixation of -er to the content provided by the stem.

24
RONALD W. LANGACKER
(a)
(b)
: (process)
T Y J
ER
|-er
llTALK
Htalk
er~|
-er]
constructional
schema
instantiation
of schema
[PROCESS
1 Y
1
ER j
-er) 1
i 4 &
Italk]
Italkl
ER 1 j
-er j 1
categorizing
relationship
(structural
description)
Figure 8
In short, the constructional schema is exactly parallel in formation and
internal structure to any of its instantiations, such as talker, diagrammed in
Fig. 8(b); the only difference is that the schematic verb stem in the former
is replaced in the latter by the more elaborate content of talk. The
constructional schema can therefore be regarded as a symbolically-complex
expression, albeit one that is too abstract semantically to be very useful for
communicative purposes, and too abstract phonologically to actually be
pronounced or perceived. Instead it serves a classificatory function, defining
and characterizing a morphological construction by representing the
commonality of its varied instantiations. The relationship between the schema
and a specific instantiating expression resides in a categorizing unit of the
sort depicted in Fig. 8(c). The global categorization, indicated by arrow T,
amounts to a relationship of schematicity: talker elaborates the abstract
specifications of the schema, but is fully compatible with them, and thus
instantiates the morphological pattern the schema describes. This global
categorization can be resolved into local categorizations between particular
substructures. Arrow 42' indicates that talk is categorized as a verb and
elaborates the schematic stem within the constructional schema. The
morpheme -er occurs in both the schema and the specific expressions that
instantiate it; the double-headed solid arrow, labeled '3', marks this
relationship of identity.
A symbolically-complex expression can of course be incorporated as
one component of an expression that is more complex still. A derived verb
like sharpen can therefore function as the stem in the morphological pattern
just described, resulting in sharpener, this noun in turn combines with
pencil in the compound pencil sharpener, whose organization is sketched at the
bottom in Fig. 9(a). In the same fashion, one constructional schema is
capable of being incorporated as a component of a larger schema describing

AN OVERVIEW OF COGNITIVE GRAMMAR
25
expressions of greater complexity. At the top in Fig. 9(a), we see the
higher-order constructional schema responsible for expressions like pencil
sharpener, mountain climber, lawn mower, taxi driver, flamethrower, etc.
One of its components is the constructional schema for nominalizations
with -er, as represented in Fig. 8(a); the other component is the noun
schema (cf. Fig. 7(a)). Phonologically, the higher-order schema specifies
the juxtaposition of the two stems to form a compound: [X Y-er]. Semanti-
cally, it specifies that the "thing" symbolized by [X] is equated with the
object of the process symbolized by [Y]. The global categorizing
relationship between this constructional schema and a specific compound like
pencil sharpener reflects local categorizing relationships at different levels of
organization. In Fig. 9(a), arrow '1' again indicates that the overall
relationship is elaborative (i.e. pencil sharpener conforms to the specifications of
the schema but is characterized in finer detail). At a lower hierarchical
level, arrows '2' and '3' show that pencil and sharpener qualify respectively
as instantiations of the noun schema and of the morphological pattern
previously considered. Finally, at the lowest hierarchical level, relationship '3'
is resolved into the local categorizations '4' (which classes sharpen as a
verb) and '5' (an identity relation).
(a)
(THING!
| X |
2
I \
(process
1 Y
~i~.
f
[PENCIL
1 [pencil
ER II
-er II
—i i
k \
I
T
JSHARI
1 JEN ]
(sharp [
-en]
r3
I—»
5
If
, f
"ER
-er|
(t)
THING
CHALK
chalk
PROCESS]
ER
-er|
flSHARP]
1 [sharp)
|EN]!
|-en|
Figure 9
Grammar, I claim, is nothing more than patterns for successively
combining symbolic expressions to form expressions of progressively greater
complexity. These patterns take the form of constructional schemas, some
of which incorporate others as components. Constructional schemas have
multiple functions in this model. First, they capture generalizations by
representing the commonality observed in the formation of specific
instantiating expressions. Second, they provide the basis for categorizing relation-

26
RONALD W. LANGACKER
ships which show the status of specific expressions with respect to structures
and patterns of greater generality. An expression's "structural description"
is simply the set of categorizing relationships in which it participates; for
instance, the structural description of pencil sharpener includes the
categorizing relationships depicted in Fig. 9(a) (together with others not
shown). Finally, a constructional schema serves as a template for the
computation of novel instantiating expressions. Consider a previous example,
namely the coinage of chalk sharpener to designate a previously unfamiliar
gadget. Symbolic resources available to the speaker include the units and
categorizations indicated in Fig. 9(b): the lexical units chalk and sharpen,
their respective categorizations as a noun and a verb, and the constructional
schema representing a compounding pattern. To compute an appropriate
compound, the speaker need only co-activate these structures, i.e. carry out
the elaborative operations relating chalk to the noun schema, and sharpen
to the verb schema, in the setting of the schema for the construction. Of
course, the composite semantic structure obtained in this way will only
approximate the expression's actual, contextually-determined value.
To conclude this section, let us once more consider phrase trees like
Fig. 4, which specify linear ordering, constituency, and constituent types. It
may now be apparent that all three sorts of information are also provided in
the proposed alternative, which posits only symbolic units for the
description of grammatical structure. Linear ordering (actually, temporal
ordering) is simply one dimension of phonological structure. A symbolic
structure — whether simple or complex, specific or schematic — specifies the
temporal sequencing of its phonological components as an inherent part of
the characterization of its phonological pole. Thus sharp, sharpen,
sharpener, and pencil sharpener all specify, as part of their internal phonological
structure, the temporal ordering of their segments, syllables, morphemes,
and stems; the same is true of the schemas which these expressions
instantiate, except that the phonological elements in question are partially
schematic (e.g. the constructional schema of Fig. 8(a) specifies, in its
composite structure, that -er follows the schematically-characterized verb
stem). Syntax and morphology are not sharply distinct in this regard. The
difference between them reduces to whether the construction involves
multiple words at the phonological pole, or parts of a single word (compounds
are thus a borderline case).
Constituency is not a separate facet of linguistic structure, but merely
reflects the order in which symbolic structures are successively combined in

AN OVERVIEW OF COGNITIVE GRAMMAR
27
the formation of a complex expression. We may speak of a "compositional
path" leading from individual morphemes, through intermediate-level
composite structures, to the highest-level composite structure representing the
value of a complex expression overall. The compositional path remains
implicit in "compacted" diagrams like Fig. 9(a), which fail to separately
depict composite structures, but is immediately apparent when these are
converted into the equivalent '"exploded" format, as illustrated in Fig. 10.
On the left in Fig. 10 is the compositional path of pencil sharpener;
intermediate-level composite structures show that sharpen and sharpener are
constituents (in addition to the individual morphemes and the expression as
a whole). On the right in Fig. 10 is the compositional path of the
constructional schema that pencil sharpener instantiates.
PENCIL-SHARP-EN-ER
pencil sharp-en-er
1
THING-PROCESS-ER1
X Y-er
PENCIL
1 pencil
\ 2
THING]
X
SHARP-EN-ER
sharp-en-er
1SHARP-EN
[sharp-en
^ k
PROCESS
Y ]
SHARP
sharp
enI
-en
Figure 10
These are not phrase trees, but simply ordered assemblies of symbolic
units. They differ from phrase trees in several ways. First, every node is a
symbolic structure incorporating both semantic and phonological content

28
RONALD W. LANGACKER
(and nothing else). Second, these structures are not linearly ordered: each
node specifies temporal ordering internally at the phonological pole, but
the nodes are not temporally ordered with respect to one another. A third
difference is the absence of node labels specifying the grammatical class of
constituents. Class membership is specified not by labels or features, but
rather by categorizing relationships, each reflecting the assessment that a
structure instantiates a particular schema. In the case of grammatical
structure, the categorizing schemas are symbolic, with actual semantic and
phonological content (though often this content is abstract, even to the
point of being essentially vacuous). Collectively, these categorizing
judgments constitute the expression's structural description.
Distribution and Predictability
Cognitive grammar posits only symbolic units for the description of
grammatical structure. Having examined the general character of such
analysis, we turn now to specific phenomena that are often taken as
demonstrating the autonomy of grammar.
One class of arguments pertains to the impossibility of predicting the
membership of "distributional classes", i.e. classes defined on the basis of
morphological or syntactic behavior. For example, there is no way to
predict, on either semantic or phonological grounds, precisely which English
verbs form their past tense by ablauting / to a (i.e. [1] to [ae]); this pattern is
possible only with a small set of verb stems (sit, swim, begin, ring, sing,
etc.) that are neither coherent semantically nor unique phonologically. Nor
is it possible, apparently, to predict the exact membership of the class of
verbs occurring in the so-called "dative shift" construction. While the
construction favors verbs of transfer that are monosyllabic (e.g. give, send,
mail, ship, and buy, but not transfer, communicate, purchase, or propose),
certain pairs of verbs that appear quite comparable both semantically and
phonologically exhibit contrasting behavior:
(12) (a) I told the same thing to Bill,
(a') I told Bill the same thing,
(b) I said the same thing to Bill.
(b') * I said Bill the same thing.
(13) (a) I peeled a banana for Bill,
(a') I peeled Bill a banana.

AN OVERVIEW OF COGNITIVE GRAMMAR
29
(b) I cored an apple for Bill,
(b') *I cored Bill an apple.
Much more can be said about this issue, but let us assume the worst,
namely that the class of dative-shift verbs is unpredictable and must
somehow be listed in the grammar.
In providing such information, linguists generally resort to some type
of diacritic or grammatical feature. For example, sit might be marked by a
diacritic indicating its membership in the class of verbs (e.g. "Class 2B")
that ablaut / to a to mark the past tense. Or, in addition to semantic and
phonological features, give might be attributed a syntactic rule feature such
as [4- Dative Shift], which specifies its ability to undergo the dative shift
rule. Whatever device is used, the marking is assumed to have neither
semantic nor phonological content, since the class it identifies is not
uniquely predictable on the basis of either form or meaning. The autonomy
of grammar is commonly assumed to follow as an inescapable consequence:
since grammatical behavior forces one to posit a distinct set of specifically
"grammatical" classes and descriptive constructs, grammar must constitute
an independent domain of linguistic organization.
This argument is fallacious, for it confuses two issues that can in
principle be distinguished: (i) what kinds of structures there are; and (ii) the
predictability of their behavior. I call this the "type/predictability fallacy". It is
not logically incoherent to maintain that only symbolic units are required
for the description of grammatical structure, even though one cannot
always predict, in absolute terms, precisely which symbolic units occur in a
given construction. Some type of marking or listing is required to provide
this information, but there is no a priori reason to believe that a contentless
feature or diacritic is the proper device for this purpose. Indeed, little
cognitive plausibility attaches to the claim that speakers possess any direct
analog of empty markers like [Class 2B] or [4- Dative Shift] in their mental
representation of linguistic structure.
It is possible to furnish the requisite distributional information without
resorting to anything other than symbolic units. To say that a particular
verb, e.g. sit, marks its past tense by ablauting / to a is equivalent to saying
that sat is a conventional unit of English. Similarly, to say that a particular
verb, e.g. give, occurs in the dative shift construction is equivalent to saying
that the schema describing this construction is instantiated by subschema
having give as its verbal element. I have argued elsewhere (Langacker
1982a, in press) for a "usage-based" model of linguistic structure, wherein

30
RONALD W. LANGACKER
both schemas and their instantiations are included in the grammar of a
language, provided that they have the status of conventional units. Schemas
capture generalizations by representing patterns observable across
expressions. Unit instantiations of constructional schemas (both specific
expressions and subschemas at varying levels of abstraction) describe the actual
implementation of these generalizations by specifying their conventional
range of application.
The specification that sit marks its past tense by ablauting / to a is thus
accommodated by including in the grammar of English the symbolic
structures sketched in Fig. 11(a). At the top in 11(a) is the constructional schema
describing the ablaut pattern in general terms. There are two component
structures. The first is the verb stem, characterized semantically only as
designating a process; its phonological pole is also schematic, specifying
only the inclusion of /. The other component structure is the appropriate
past-tense allomorph; its semantic pole is given as [PAST], and its
phonological pole as [i —> a] (i.e. past tense is symbolized by an ablauting
operation applied to the stem). The compacted format is used, so the
resulting composite structure, namely [PROCESS-PAST/...a...], is not
separately shown. Instantiating this constructional schema is the structure
at the bottom in 11(a), which is nothing other than the lexical unit sat. This
complex expression is naturally parallel in organization to the schema it
instantiates. The two component structures are [SIT/sit], which elaborates
the schematic verb stem, and the past-tense allomorph [PAST/i --> a], which
also occurs in the constructional schema. Though not shown separately, the
composite structure [SIT-PAST/sat] instantiates that of the schema. Other
past-tense verbs that follow this pattern are also included in the grammar as
conventional units, each of them combining with the constructional schema
in a categorizing unit analogous to 11(a). Both the morphological pattern
itself and its lexical extension are thereby specified using only the apparatus
permitted by the content requirement.
:| (process
1 Y
1
THING
X
' ♦ ;
[GIVE
[give
- ■■ |
"THING 1 j I
z l||
I' ■ j
THING
X
L- -J
J. .-.
THING111
Z |
Figure 11
l| PROCESS
...i...
1
PAST it
i--»*a|| 1
* ;
1 x" r
[SIT
sit
f
PAST |||
1—a||

AN OVERVIEW OF COGNITIVE GRAMMAR
31
Fig. 11(b) sketches the categorizing unit identifying give as a verb that
occurs in the dative shift construction. At the top in 11(b) is a simplified
representation of the constructional schema. There are three component
structures, the verb schema and two instances of the noun schema (one
would actually have to indicate that these are full noun phrases). At the
semantic pole, the composite structure (not separately shown) specifies that
the "things" designated by the noun phrases participate in a possessive
relationship that results from the verbal process; its phonological pole
specifies that the noun phrases are directly contiguous and follow the verb.
At the bottom in 11(b) is a subschema representing the generalization that
give appears in this construction. It is precisely parallel to the schema,
except that the specific content of give replaces that of the verb schema at
both the component- and composite-structure levels. Similar categorizing
units are posited for other dative-shift verbs.
There is much more that needs to be said about distribution,
predictability, and the related issue of economy in linguistic descriptions, but this is
not the place to pursue the matter. Our present concern is merely to
observe that inability to predict the exact membership of distributional
classes does not, in principle, establish the autonomy of grammar (as earlier
defined). I have outlined a strategy for providing the necessary information
using only symbolic units. Moreover, only constructs and descriptive
devices needed for independent reasons arc employed.
Arbitrary Grammatical Markings
The autonomy of grammar is commonly asserted on the grounds that
languages impose arbitrary requirements on the form of permissible
expressions. The validity of such an argument depends on how the pivotal notions
are understood, in particular the notions "form" and "autonomy". My own
definition of what it means for grammar to be autonomous was presented
earlier; the alternative I propose relics solely on symbolic units for the
description of grammatical structure. From this perspective, the only
sensible definition of "form" equates it with the phonological pole of linguistic
units and expressions — specifically rejected are phrase trees like Fig. 4,
which supposedly represent an independent level or dimension of
"grammatical form". Linear ordering is simply one dimension of phonological
space, and we have seen how the symbolic alternative accommodates both
constituency and category membership.

32
RONALD W. LANGACKER
Obviously, languages do impose requirements on the form of
expressions. Most of these requirements have to be learned, as they are not
strictly predictable from any independent factor; in this sense they are
arbitrary (or at least conventional). That itself does not distinguish
grammar from lexicon — after all, a lexical item such as [APPLE/apple] can be
regarded as imposing the arbitrary formal requirement that the notion
[APPLE] be symbolized by [apple]. Nor does it demonstrate that
grammatical structure involves anything other than symbolic units. Consider the
formal restriction that a simple adjective in English must precede rather than
follow the noun it modifies, as in ripe apple. Although this formal
relationship is essentially arbitrary, we must certainly recognize its symbolic
import: the fact that [ripe] directly precedes [apple] along the temporal axis
is precisely what symbolizes the fact that the property designated by [RIPE]
is attributed to the "thing" designated by [APPLE]. The regularity of this
relationship is captured by positing a constructional schema having a
schematic adjective and a schematic noun for its component structures; the
schema's phonological pole specifies the temporal contiguity and ordering
of the adjective and noun, while its semantic pole specifies the fact and
nature of their semantic integration. The formal restriction is therefore
simply the phonological pole of a particular type of symbolic unit.
Discussions of the autonomy issue generally cite the existence of
"empty" markers, which appear to serve purely grammatical purposes
without making any semantic contribution. Frequent examples are the auxiliary
verbs be and do; certain subordinators; prepositions that are governed by a
particular verb or construction (especially when no alternative choice is
permitted); markings induced by agreement; and also the markers for case,
gender, and similar categories. We cannot examine all these matters in
detail (previous works on cognitive grammar describe the semantic value of
numerous "grammatical" elements). I will nevertheless argue that they are
not intrinsically problematic for the symbolic conception of grammatical
structure.
We may certainly concede the existence of markings that do not
substantially affect the information content of the expressions in which they
appear. However, there is much more to meaning than information content
(or truth conditions): equally important is imagery, i.e. how that content is
structured or construed for communicative purposes. Given a linguistically
appropriate view of meaning, one that accommodates conventional
imagery, grammatical markers like those noted above can in fact be considered

AN OVERVIEW OF COGNITIVE GRAMMAR
33
meaningful. The nature and degree of their semantic contribution varies —
as a limiting case, it may be essentially vacuous; but this extreme should be
regarded as the endpoint of a continuous scale, and not as criterial or
prototypical for grammatical elements.
Let us begin by considering certain factors that are sometimes taken as
indicating that a grammatical marker is meaningless. These include: (i)
being redundant; (ii) being obligatory; (iii) serving a specifiable
grammatical function; and (iv) being language-specific. Granted various concepts
and assumptions of the present framework, none of these factors implies
the meaninglessness of a grammatical element.
A redundant element may be meaningless from the standpoint of
information theory, but that is only a minor concern of linguistic semantics. For
natural language, it is important that we distinguish clearly between two
notions that are certainly not equivalent, namely "meaningfulness" and
"non-overlapping meaning". There is always some semantic overlap
between two component structures that combine to form a composite
expression; it is precisely because the component structures overlap in some
fashion that they can be integrated to form a coherent composite
conceptualization. In ripe apple, for instance, ripe specifies the maturity of fruit
and its readiness for consumption, and hence introduces the notion of fruit
schematically, while apple subsumes and elaborates this schematic notion
by characterizing a particular type of fruit. Since all complex expressions
involve semantic redundancy of this sort, the meaning of one component
cannot be determined just by subtracting the meaning of its companion
from that of the expression as a whole — the meanings of the component
structures are not strictly complementary, so each must be described in its
own terms.
As a limiting case, one component structure is properly considered
meaningful even if its semantic content is fully subsumed by that of the
other, and therefore contributes nothing to the composite semantic
structure that would not be available from the second element alone. This can be
exemplified by a compound like oak tree or puppy dog. The content of tree
is fully subsumed by that of oak, and the content of dog by puppy, but no
one would argue that tree and dog are semantically empty morphemes.
Semantic overlap is inherent to grammatical constructions, cases like oak
tree and puppy dog being special only in that the region of overlap happens
to exhaust the content of one component. Expressions like oak tree
illustrate what I call the "schematic transparency principle": when combined, a

34
RONALD W. LANGACKER
schema and its instantiation merge to form a composite conceptualization
equivalent to the instantiation. Since [TREE] is schematic for [OAK], the
content of the former being fully subsumed by that of the latter, the
composite notion obtained by merging their specifications is identical to the
instantiation [OAK]. But while tree adds nothing to the information content
of oak tree, it does make a semantic contribution from the standpoint of
imagery: explicit mention of the superordinate category tree makes the
membership of oaks in this category somewhat more salient in the
expression oak tree than it is in the simple expression oak.
With this in mind, consider the auxiliary verb do and the contrast
between pairs of sentences like (14):
(14) (a) My cousin smokes.
(b) My cousin does smoke.
I analyze the auxiliary do as designating a maximally schematic process, i.e.
it is a true verb that conforms to the specifications of the verb schema (Fig.
7(b)) and in fact is identical to that schema at the semantic pole. Semanti-
cally, do is therefore schematic for smoke, as it is for any other verb. From
the schematic transparency principle, it follows that the composite structure
of do smoke is equivalent to that of smoke alone, i.e. do contributes
nothing to the information content of the complex expression. One can
understand why many linguists have concluded that do is semantically
empty, but I have tried to show that this conclusion does not really follow.
The redundancy of do in composition with a specific verb does not preclude
its analysis as a symbolic unit with actual (though schematic) semantic
content. Moreover, this content helps explain the semantic contrast of sentence
pairs such as (14). Like that between oak and oak tree, the contrast between
smoke and do smoke pertains to imagery rather than information content,
and resides in the added salience accorded a superordinate category (the
[PROCESS] category in this case) by virtue of explicit symbolization. In
conjunction with emphatic stress, the occurrence of do + V (rather than V
alone) underscores the notion of a process and thus lends a nuance of
"emphasis1' or "assertion" to the verbal expression.
Hence the semantic redundancy of a grammatical marker (from the
standpoint of information content) is compatible in principle with its being
both meaningful and describable as a symbolic unit. We can also coherently
assert the meaningfulness of a grammatical marker that is obligatory, is
chosen to the exclusion of other alternatives, and serves a specifiable gram-

AN OVERVIEW OF COGNITIVE GRAMMAR
35
matical function. The auxiliary verb do is once again a convenient example.
If (14)(a) is negated, questioned, used as a tag, or rendered elliptic, do
occurs in addition to smoke or in lieu of it:
(15) (a) My cousin does not smoke.
(b) Does my cousin smoke?
(c) My cousin smokes, doesn't she?
(d) She does.
Do is generally attributed a specific grammatical function in these
constructions: it carries the markings of tense and subject-agreement, which would
otherwise be "stranded". However, nothing inherently prevents a
meaningful element from being conventionally adopted to play a particular
grammatical role, and elements with highly schematic content would seem to be
ideally suited for this purpose. (By the same token, subordinators like
since, because, while, although, etc. indicate the beginning of a subordinate
clause and thus have a specific grammatical function, but they are clearly
meaningful.) The schematicity of do with respect to the class of verbs
explains its semantic "transparency" in the constructions of (15)(a) and (b),
and also its ability to function as a verbal "pro form" in those of (c) and (d).
What about the fact that do is obligatory in these constructions and
fully predictable (i.e. no other verb can take its place)? Only granted a
strict information-theoretic definition of meaning does this imply that do is
meaningless; there is no such implication when meaning is related instead
to conceptualization and conventional imagery, and semantic overlap is
recognized as an inherent feature of grammatical constructions. A
constructional schema indicates the conventionality (well-formedness) of
expressions conforming to its specifications. Should such a schema mention a
particular symbolic element (e.g. do) as one of its component structures, the
expressions it sanctions are limited to those that incorporate the element in
question. The absence of sanctioned alternatives does not prevent this
element from having a semantic pole with actual content, regardless of how
schematic this content might be, or how redundant it might be with respect
to the other component. One facet of conventional imagery resides in the
compositional path through which the composite structure of a complex
expression is progressively assembled, and even redundant elements
contribute in this regard, as we have seen.
Finally, consider the fact that grammatical markings vary from
language to language — not just their form, but also their inventory and dis-

36
RONALD W. LANGACKER
tribution. This fails to establish that such markings are meaningless or non-
symbolic. It merely indicates, in the realm of grammatical constructions,
what is already evident from the study of lexicon: languages make different
inventories of symbolic resources available to their speakers, who
consequently say comparable things in different ways. To convey roughly the
same content, the speakers of two languages may be forced, by the
constructional schemas at their disposal, to employ expressions that differ in
such factors as how precisely they specify some parameter (e.g. definite-
ness, number, gender), the amount of redundancy they incorporate, or the
nature of the compositional path through which they arrive at the
composite semantic structure. In short, semantic structure is not universal when
imagery is properly taken into account, so the non-identity of grammatical
markings across languages does not itself demonstrate their semantic
irrelevance.
I conclude, then, that the factors considered above do not establish the
meaninglessness of grammatical markers, as they are commonly assumed to
do. With this background, we can now examine some specific phenomena
that are almost invariably cited in support of the autonomy thesis:
agreement, number, case, and gender. I must limit myself to sketching the
general strategy for dealing with them in the context of cognitive grammar.
Let us start with subject-verb agreement in English, as illustrated in
(16):
(16) (a) The kitten is playful.
(b) *The kitten are playful.
(c) *The kittens is playful.
(d) The kittens are playful.
Standard analyses attribute no semantic value to the person and number
distinctions marked on the verb. Instead, some mechanism is invoked (e.g.
a transformational rule) that copies the person and number features of the
subject noun phrase onto the verb, where they are eventually spelled out
morphologically. Being mechanically induced by morphosyntactic
processes, verb agreement does nothing more than duplicate information
provided elsewhere in the sentence, and is therefore semantically empty even
if the person and number categories are considered semantic.
We will confine our attention to number, and assume for the moment
that the number category, as marked on nominal expressions, is indeed
semantic. There are cogent reasons for questioning the validity of the fea-

AN OVERVIEW OF COGNITIVE GRAMMAR
37
ture-copying analysis; though it can doubtless be made to work, its
implementation is far less straightforward than is generally recognized (cf.
Morgan 1972), and it offers no real insight into the phenomenon. Let us
consider just one type of problem, namely instances where the same subject
nominal occurs with either a singular or a plural verb:
(17) (a) The faculty agrees,
(b) The faculty agree.
(18) (a) Drinking and singing is fun.
(b) Drinking and singing are fun.
(19) (a) The team is playing hard,
(b) The team are playing hard.
In the case of (17), one might argue that faculty can be valued either ' + ' or
' — ' for the feature [PLURAL]. Some corroboration is seemingly found in
the fact that only the singular demonstratives (this, that) can substitute for
the definite article in (17)(a), and only the plural demonstratives (these,
those) in (17)(b). For (18), one might similarly propose that nominals
formed by conjunction with and are sometimes free to take either feature
value (it would not, however, be easy to specify just when that is possible).
This type of analysis will not work for (19) in any event, since team is clearly
singular in both examples (note this team, *these team, teams).
More, then, appears to be going on than is suggested by a mechanical
feature-copying analysis. How can these phenomena be dealt with in
cognitive grammar, which posits only symbolic units for the description of
grammatical structure? The central claim is that number-marking on the verb is
independently meaningful. A plural verb contrasts semantically with the
corresponding singular verb by portraying its subject as exhibiting a greater
measure of individuation among its subparts. Often this specification is
redundant, as in (16)(d); because a plural noun like kittens designates a
mass consisting of indefinitely many distinct instances of the type of entity
designated by the stem, substantial individuation is implied by the
morphology of the subject nominal itself. However, since "meaningfulness" is not
the same as "non-overlapping meaning", such redundancy does not affect
the viability of the analysis. Moreover, the deviance of sentences like
(16)(b) and (c) is attributable to the semantic incompatibility of separate
specifications concerning the same conceived entity: either the verb
attributes substantial individuation to this entity while the subject nominal does
not, or conversely.

38
RONALD W. LANGACKER
Number marking on the verb is neither redundant nor contradictory
with respect to the subject nominal when the entity designated by the latter
lends itself to alternate construals implying different degrees of
individuation. In (17), it is the number of the verb that indicates whether the faculty
is construed holistically as a unified body, or whether its status as a
collection of individuals is given greater weight; accordingly, (17)(a) conveys the
notion that the faculty, as a unitary formal entity, agrees with some
externally-generated proposal, while (17)(b) portrays the faculty agreeing as
individuals, perhaps with one another. Similarly, the choice between is and
are in (18) signals whether drinking and singing are being construed as two
facets of a single complex activity, or as two essentially separate activities,
each of which is fun individually. As for (19), it is clear that team is
inherently ambivalent, implying distinct individuals who nevertheless
subordinate themselves to cooperative endeavor toward a common goal. No
particular importance attaches to the fact that the choice between (19)(a) and
(b) is in large measure dialectally determined (American vs. British). For
one thing, individual speakers may well be familiar with both patterns, and
use them to make a semantic distinction (this is coming to be true of my
own speech). But what if speakers have no option, so that one pattern or
the other must be employed (strictly according to dialect), even though
either construal is conceptually quite natural? This merely reflects the
conventionality of the imagery embodied by the symbolic resources of a
language: out of all the ways of construing a given type of situation, certain
possibilities become conventionally established (i.e. represented in the
grammar by symbolic units) to the exclusion of others. Like languages,
dialects often diverge in this regard. I would argue, then, that the dialectal
difference between (19)(a) and (b) is also a subtle difference in meaning —
speakers of the two dialects conventionally employ slightly different images
to construe the situation for expressive purposes.
An argument sometimes advanced for the autonomy of grammar
maintains that "semantic number" and "grammatical number" must be
distinguished, with grammatical rules sensitive only to the latter (cf. Hudson
1976: 6). For instance, nouns like binoculars, scissors, pants, glasses, tongs,
tweezers, pliers, shorts, and trousers are said to be semantically singular, but
they show the agreement pattern of plurals:
(20) (a) These binoculars are powerful.
(b) *This binoculars is powerful.

AN OVERVIEW OF COGNITIVE GRAMMAR
39
The problem with this argument is that it fails to acknowledge either the
phenomenon of imagery or its conventional character. Why are binoculars,
scissors, pants, etc. considered semantically singular? Presumably because
each designates a discrete physical object, not a mass consisting of separate
discrete objects all of the same type. Yet it is surely no accident that these
nouns in particular act as plurals; a prominent feature of each designated
object is that it consists of two more-or-less identical parts that have no
separate function or identity — they occur only as complementary halves of the
object in question. Consequently, our conception of these objects embraces
both unity and duality, either of which might be highlighted by the
conventional units of a given language. Anything which emphasizes one facet of a
complex notion at the expense of another counts as imagery. Since
alternatives are conceivable, the imagery is conventional, i.e. speakers must learn
the proper way of expressing these notions. Our inability to predict what
pattern a language might use does not entail that the choice has no semantic
basis.
To summarize, the patterns illustrated in (16) (for "regular" nouns like
kitten) and (20) (for nouns like binoculars) are fixed and obligatory, involve
a certain amount of redundancy, and could not be predicted if one did not
happen to know them. In the present framework, we can nevertheless
coherently maintain that the number indications of the subject nominal and
the verb represent separate semantic specifications and are independently
meaningful; their contributions are sometimes consistent, sometimes
incompatible, and sometimes complementary. Plural marking on the verb
attributes a certain measure of individuation to the subparts of the entity
designated by the subject nominal; the nature and the degree of
individuation implied require further study — most likely they vary depending on the
inherent properties of the subject. The plural marking that occurs on nouns
is also variable in its precise semantic import, and is not necessarily
identical in value to that which co-occurs on verbs. Prototypically, plural
inflection indicates a mass consisting of indefinitely many instances of the type of
discrete object designated by the noun stem. From this basic sense of the
plural morpheme, extended senses develop to accommodate the semantic
peculiarities of special types of nouns. I thus identify the final -s of nouns
like binoculars as the plural morpheme, even though the remainder docs
not always occur individually (this is unproblematic in cognitive grammar
— see ch. 12 of Langacker in press); presumably it indicates the inherent
duality of the object designated by the overall expression.

40
RONALD W. LANGACKER
Like number, the categories of case and gender are often subject to
"agreement", i.e. the case or gender of a particular entity is often marked
in more than one place. It should now be evident that the redundancy,
obligatoriness, and language-specific character of such markings do not per
se demonstrate the autonomy of grammar. The pivotal question is whether
a marker for case or gender can be attributed some kind of semantic value
and treated as a symbolic unit. If so, the phenomena can be analyzed along
the general lines suggested above for number agreement. Each occurrence
of such a marker constitutes a separate semantic specification. Whether for
functional reasons or out of sheer perversity, languages sometimes demand
multiple specifications pertaining to the same entity, despite their full
semantic overlap. Conventional agreement patterns are represented in the
grammar of a language by constructional schemas which incorporate only
symbolic units. These schemas embody conventional imagery. They subtly
influence the meaning of the complex expressions they sanction, both by
specifying the inclusion of the case or gender markings (however tenuous or
abstract their semantic content may be), and also by dictating a particular
compositional path for arriving at the ultimate composite structures.
We must now address the issue of whether the categories of case and
gender are in fact to be considered meaningful. My comments on case will
be very brief and general, as this topic has only begun to be investigated in
the context of cognitive grammar (cf. Smith 1985). We may distinguish
between nominative and accusative case on the one hand, and "oblique"
cases on the other. Basically, the question of whether markings for
nominative and accusative case are meaningful reduces to the question of whether
the grammatical relations "subject" and "direct object" have a semantic
basis. I claim that they do — the characterization pertains to the relative
prominence of relational participants (prominence being an important
dimension of imagery) — but discussion and justification of this claim are
beyond the scope of this paper (see Part II of Langacker in press). The
assertion that oblique cases like dative, genitive, instrumental, ablative,
etc. make a semantic contribution is perhaps less controversial. I suggest
that they always have a semantic value of some sort, though it may be
redundant or even vacuous on occasion. It is important that the basic
question of their meaningfulness be distinguished from two subsidiary issues
that do not bear on it directly. First is the issue of whether a given case has
a single meaning in all its uses; I neither claim nor expect that it does.
Lexical items in common use almost invariably have a family of distinct but

AN OVERVIEW OF COGNITIVE GRAMMAR
41
interrelated senses, and it is reasonable to assume (as I did earlier for plural
markings) that the same is generally true of the more abstract symbolic
units regarded as grammatical markers. Having a multiplicity of
interrelated meanings is not equivalent to being meaningless. The second issue
concerns the grammatical function of case markings, and the fact that a
verb, preposition, or construction often governs a particular case. I have
already argued that such considerations do not establish that the element in
question is semantically empty. Meaningful elements can perfectly well
serve specifiable grammatical functions, even when no alternatives are
permitted.
As with number, a distinction is commonly drawn between
"grammatical gender'1 and "semantic" or "natural gender". It is recognized that
gender-like categories tend to be semantically motivated for a substantial range
of vocabulary (e.g. le filslla fille, le chat/la chatte, etc.), but even within that
range there may be exceptions (la sentinelle), and outside that range the
category membership of a noun is basically arbitrary (le fromagella viande,
le crayon/la plume). While the arbitrariness of gender-like classes is
sometimes overstated, the basic facts arc clear enough. It would be foolish to
maintain that every French noun taking le or la does so because it is
semantically 'masculine' or 'feminine' (using these terms in their normal sense).
Nor can one plausibly argue that the gender categorizations of nouns are
strictly predictable from their meanings in all cases. Many scholars have
taken such facts as demonstrating the autonomy of grammar and
grammatical classes.
Nevertheless, the data is readily accommodated in a model that posits
only symbolic units for the description of grammatical structure. Recall, to
begin with, that a meaningful element need not have precisely the same
meaning in all its uses — both lexical and grammatical morphemes typically
display a network of distinct but interrelated senses (cf. Lindner 1981, 1982;
Hawkins 1984; Casad and Langacker 1985). We can therefore attribute to a
gender morpheme such highly specific semantic content as 'masculine' or
'feminine' in many of its occurrences, notably with the vast majority of
animate nouns, without thereby committing ourselves to the claim that it
retains this content in all its occurrences. Indeed, with inanimates its
semantic value may well be highly schematic and contribute very little to
the meaning of the composite expression. But schematic conceptions are
meanings nonetheless, and markers with schematic content are properly
analyzed as symbolic units.

42
RONALD W. LANG ACKER
Consider the noun endings -o and -a of Spanish. With animate nouns,
they generally indicate 'masculine' and 'feminine' respectively; hence perro
'male dog' vs. perra 'female dog', muchacho 4boy' vs. muchacha 'girl', etc.
The complex symbolic units representing the two forms for 'dog' are
sketched in Fig. 12(a)-(b); note that each component morpheme
contributes something substantial to the composite semantic value. How, then,
do we analyze such nouns as palo 'stick' and mesa 'table', which have no
natural gender but function grammatically as masculine and feminine
parallel to perro and perra?
1 1 DOG
i perr
masculine]]
-o 1 j
Fdog"
11 perr
FEMINII^Ej]
-a I]
(stick
1 pal
ihingI j
-o 11
1] TABLE
1 mes
THINGM
-a
Figure 12
A possible analysis is given in Fig. 12(c)-(d). Let us assume that
speakers segment -o and -a as gender markers. What semantic value do they
attribute to these seemingly empty morphemes? Possibly their content,
necessarily schematic with inanimate stems, is simply to be identified as
[THING], the semantic pole of the noun schema (just as the content of the
auxiliary verb do is the same as [PROCESS], the semantic pole of the verb
schema). The content of the ending is then fully subsumed by the stem,
since [THING] is schematic for both [STICK] and [TABLE] (and any other
noun). By the schematic transparency principle, the composite structures
[STICK-THING] and [TABLE-THING] are equivalent to the respective
component structures [STICK] and [TABLE]; this would account for why
the ending appears to be meaningless. Still, only symbolic units are
employed in the description. Presumably a speaker of Spanish masters all
the conventional units depicted in Fig. 12. Those in 12(a) and (b) indicate
that the stem for 'dog' occurs with either ending to specify its "natural"
gender. Those in (c) and (d) specify that 'stick' and 'table' take -o and -a
respectively, though all the essential content is provided by the stem. The
categorization of these nouns as grammatically "masculine" or "feminine"

AN OVERVIEW OF COGNITIVE GRAMMAR
43
is represented by these units themselves — separate diacritics or
grammatical features are not required.
Grammatically Judgments
Since the earliest days of generative grammar, linguists have supported
the autonomy thesis by asserting that speakers are capable of judging the
grammatical well-formedness of sentences out of context and without
regard for their meaning (cf. Chomsky 1957). Thus (21)(a) is accepted as
grammatical despite its semantic anomaly, while (21)(b) is judged ill-
formed.
(21) (a) Colorless green ideas sleep furiously,
(b) * Green furiously ideas colorless sleep.
The same point is often made with Jabberwocky, where sentences are
recognized as grammatical even when nonsense forms are used for the stems,
so that only the "grammatical morphemes" are familiar.
Let me state at the outset that the treatment of so-called "grammatical-
ity judgments" in the generative tradition has been little short of
scandalous, with respect to conceptual confusion in discussing the issue, the
absence of any firm empirical basis for assessing grammaticality, and the
dependence of sweeping theoretical claims on judgments that are in fact
quite fragile and variable. We have no reason whatever to believe that
natural-language well-formedness judgments resemble the prototype
suggested by automata theory, where a mechanical procedure yields a
discrete decision of "grammatical" or "ungrammatical" simply by scanning a
string of vocabulary symbols devoid of content. The natural-language
judgments reflected in the generative literature by asterisks and comparable
notations are often not discrete (hence the presence vs. the absence of an
asterisk is not sufficiently delicate to represent them), and almost invariably
take semantic considerations into account in one way or another (regardless
of whether they are recognized as such). These assessments are also quite
sensitive to contextual factors. We should seriously consider Dinsmore's
suggestion (1979) that sentences are always construed relative to some
context, and that a "normal" context is implicitly invoked for supposedly aeon-
textual judgments.
Still, there is a legitimate difference between (21)(a) and (b) that has to
be accounted for. What do such examples actually demonstrate? First, they

44
RONALD W. LANGACKER
show that languages impose restrictions on the form of permissible
expressions, restrictions that refer in some manner to grammatical classes (e.g. N,
V, ADJ, ADV, NP). Second, they show that speakers are capable of
assigning words to such categories even when their meanings, integrated
according to the usual compositional patterns, fail to yield a coherent
composite semantic structure. These observations do not, however, establish
the autonomy of grammar (as earlier defined). We saw previously that the
symbolic alternative accommodates both formal restrictions and
membership in grammatical classes. At least in principle, a description employing
only symbolic units is sufficient to handle the phenomenon.
We have been assuming that basic grammatical categories are
susceptible to notional characterization, and are represented in the grammar by
highly schematic symbolic units (cf. Fig. 7). Constructions, i.e.
conventional patterns for assembling symbolically complex expressions, are
similarly described by schemas for such expressions. A constructional schema
incorporates symbolic units as component structures and specifies how they
integrate to form a composite structure (Figs. 8-11): at the semantic pole, it
represents a pattern of semantic composition; at the phonological pole, it
determines (among other things) the temporal sequencing of the
component elements. The judgments in (21) thus reflect the availability in English
of constructional schemas that permit the assembly of certain [ADJ ADJ N
V ADV] sequences, and the unavailability of schemas that would sanction
the construction of comparable sequences with the temporal ordering [ADJ
ADV N ADJ V]. The former requires at least three constructional
schemas: one permitting nominals of the form [ADJ ADJ N]; a second for
the modifying relationship [V ADV]; and a third allowing the integration of
the resulting composite structures at a higher level of organization. (This
third schema actually allows the integration of a subject and a predicate
regardless of their internal composition. For sake of discussion, let us
assume — quite plausibly — the existence of a subschema that takes as
components a subject nominal and a predicate with [ADJ ADJ N] and [V
ADV] in particular as their respective internal structures.)
The requisite constructional schemas (in compacted format) are
sketched at the top in Fig. 13(a). The schema describing the relevant type
of nominal expressions has for component structures two instances of the
adjective schema and one of the noun schema. The second constructional
schema takes the verb schema and the adverb schema as component
structures. The third constructional schema incorporates the first two as its com-

AN OVERVIEW OF COGNITIVE GRAMMAR
45
ponents, and specifies their integration at a higher level of constituency.
This entire complex structure (which contains only symbolic units) allows
the computation of an open-ended class of instantiating expressions that are
coherent at both the semantic and the phonological pole (e.g. Tired young
boys sleep soundly). Of course, nothing guarantees that a speaker (or
linguist) will always employ the symbolic resources at his disposal with
semantic coherence as his ultimate objective. It is quite possible to choose lexical
items which conform to local specifications of the overall schema on an
individual basis, slot by slot, but which cannot be successfully integrated in
the manner that the schema requires.
(a)
1 ADJECTIVAL
RELATION
1 x
1
ADJECTIVAL
RELATION
Y
'
ICOLORIESS]
[colorless]
1
thing]
Z 1
'
GREEK
green
<
PROCESS
1 W
f
IDEAS
ideas]
ADVERBIAL11
RELATION
u HI
f
1 sleep!
[sleep]
f
[FURIOUSLY 1
[furiously)
(b)
colorless green ideas sleep furiously
COLORLESS
colorless
GREEN,
green
IDEAS]
ideas]
? 1
sleep furiously]
/
SLEEP
sleep
/
\
FURIOUSLY'
furiously
Figure 13
Suppose, then, that the lexical units colorless, green, ideas, sleep, and
furiously are selected to instantiate the lowest-level schemas, as indicated in
Fig. 13(a). What happens when these units are integrated, level by level, in
accordance with the specifications of the constructional schemas? The

46
RONALD W. LANGACKER
result is sketched (using the exploded format) in Fig. 13(b). At the
phonological pole, this computation yields precisely the sequence Colorless
green ideas sleep furiously, henee our judgment that (21)(a) is formally
correct. The outcome is quite different at the semantic pole, however. When
one attempts to integrate the component semantic structures in the fashion
specified by the constructional schemas, a coherent composite structure
fails to emerge at any level of organization (this is because the lexical units
selected attribute conflicting properties to entities that the schemas
stipulate as being identical). An analysis employing only symbolic units is
therefore capable of accounting for the fact that (21)(a) is semantically
anomalous despite its formal correctness. It also accounts for the difference
between (21)(a) and (21)(b) — in addition to its semantic anomaly, the
latter violates the formal specifications of the constructional schemas with
respect to temporal sequencing, and could not be computed from any other
available schemas.
The issue, then, is not the existence of ''grammatical patterns" and
''formal restrictions11, nor the possibility of an expression being recognized
as conforming to the latter despite its semantic incoherence. Rather the
issue is whether the representation of those patterns and restrictions
requires anything other than symbolic units. Proponents of the autonomy
thesis answer in the affirmative: in addition to meaning and phonological
shape, they posit a separate dimension of "grammatical form1', whose
primary ingredients are linear ordering, constituency, and labeling for
membership in grammatical categories (cf. Fig. 4). The symbolic alternative does
not deny the validity of these factors, but interprets them in a very different
way. Linear ordering is simply temporal ordering, one dimension of
phonological space. Constituency is a very general notion, amounting to
nothing more than the order in which simpler structures are progressively
assembled into larger ones; grammatical constituency thus reduces to the
path of composition followed in the assembly of complex symbolic
structures. Grammatical classes are defined by schematic symbolic units, and
class membership is specified by categorizing relationships holding between
schemas and their instantiations. The various elements of "grammatical
form11 are recognized, therefore, but they do not constitute a separate, self-
contained, or individually coherent structural domain — they are viewed
instead as integral facets of symbolic structures and their interconnections.
Languages provide for the symbolization of ideas by observable
sequences of sounds. Wc must therefore acknowledge in some fashion the

AN OVERVIEW OF COGNITIVE GRAMMAR
47
reality of meanings, sounds, and the symbolic associations between them.
The status of a distinct and autonomous domain of "grammar" is on the
face of it less secure, for it is hard to envisage a realm of "grammatical
content" analogous to semantic and phonological content. Undeniably, there
are conventional patterns for the assembly of complex symbolic expressions
out of simpler ones, and "grammar" is the obvious term for referring to
them. But the reality of "grammar", so defined, does not itself establish the
character of the elements that specify these patterns. In particular, it leaves
open the possibility that these elements are themselves symbolic in nature,
schematic units which embody the commonality of well-formed symbolic
expressions and serve as templates for the computation of novel
instantiations. What could be more natural and straightforward? If workable, this
approach affords a coherent and integrated view of the various facets of
linguistic organization, and permits the adoption of the highly restrictive
content requirement. I consider it a curious anomaly in the history of our
discipline that this natural conception must be explained and defended, while
the autonomy thesis — with all its unnaturalness and conceptual obfusca-
tion — commands widespread and often unquestioning allegiance.
Conclusion
Certain ways of viewing linguistic structure have become so standard
and deeply ingrained in recent years that scholars tend to accept their
validity as self-evident and to discount the potential viability of radical
alternatives. My basic objective has been to outline a very different conception of
linguistic organization, and to show what things look like when viewed in
this perspective. I do not pretend to have demonstrated the incorrectness of
currently predominant theories, or to have proved the non-autonomy of
grammatical structure. I have nevertheless considered various phenomena
commonly cited as supporting the autonomy thesis, and mapped out
strategies for attacking them within the type of model proposed. I have also
suggested the coherence, the naturalness, and even the restrictiveness of
the descriptive framework. To be sure, this overview resembles grammar
itself in being neither autonomous nor self-contained. It is schematic in its
description of critical notions, such as imagery, categorization,
constructions, the semantic definition of basic grammatical classes, and the usage-
based approach. For elaboration and justification, I must refer the reader
to other works on cognitive grammar.

48
RONALD W. LANGACKER
I have emphasized two fundamental issues that underlie virtually every
question of linguistic theory and determine how we talk and think about
language structure: the nature of meaning, and the putative autonomy of
grammatical structure. I have tried to indicate that these two issues are
inextricably linked. Only with an impoverished, objectivist conception of
meaning and linguistic semantics do the standard arguments for the
autonomy thesis have any force. Given a more inclusive and linguistically
appropriate conception of semantic structure, one which treats meaning as
a subjective phenomenon and accommodates conventional imagery, these
arguments are seen as invalid. For example, a subjectivist semantics makes
possible a notional characterization of basic grammatical categories; the
supposed impossibility of notional definitions is a major pillar supporting
the edifice of autonomous grammar.
Another major theme has been the inappropriateness of relying on
absolute predictability as criterial for determining the nature of linguistic
structure and organization. I have noted the prevalence of the
type/predictability fallacy, which confuses the two distinct issues of (i) what kinds of
structures there are, and (ii) the predictability of their behavior. A usage-
based model, which accommodates the full complexity of linguistic
knowledge by incorporating in a grammar both schematic structures and unit
instantiations at various levels of abstractness, accounts for distribution
without resorting to structures that violate the content requirement (which
allows only semantic, phonological, and symbolic units). It is commonly
assumed that the autonomy of grammar is established unless one can
demonstrate that every peculiarity of grammatical structure and behavior is
strictly predictable from independent factors (cf. Newmeyer 1983: ch. 1);
under this assumption, grammar is non-autonomous only if we can "get it
for free" as automatic consequences of something else. A severely
reductionist program of this sort is doomed to failure; the grammatical patterns
of a language are in large measure conventional and have to be specifically
learned (regardless of their degree of naturalness and functional
motivation). However, the conventionality of grammatical structure does not per
se prove that it is autonomous in the sense of constituting a separate level
or dimension of linguistic organization demanding special, purely
"grammatical" classes and elements. I have shown that conventional patterns,
including arbitrary restrictions on the form of expressions, are in principle
describable using only symbolic units for the representation of grammatical
structure.

A View of Linguistic Semantics
Ronald W. Langacker
University of California, San Diego
The Nature of Meaning
A language enables its speakers to effect an open-ended set of
symbolic correspondences between meanings and phonological sequences. The
set is open-ended because speakers learn not only a vast (though limited)
inventory of conventional symbolic expressions, but also an array of
patterns for successively combining simpler expressions into expressions of
ever greater complexity. These patterns constitute the grammatical
structure of the language. The theory of "cognitive grammar" (Langacker in
press) is founded on the claim that grammatical structure itself is inherently
symbolic, i.e. it is fully describable by means of symbolic elements, each
with both semantic and phonological content (though this content may be
quite abstract). On this view, grammatical analysis and semantic analysis
are indissociable: a description of grammatical structure that makes no
reference to meaning is ultimately no more revealing than a dictionary
providing only a list of undefined forms. As a basis for grammatical analysis,
consequently, this theory must offer at least a programmatic account of
meaning and semantic structure.
Our initial discussion will focus on five basic claims about the nature of
linguistic meaning. These are summarized in (1).
(1) Thesis A: Meaning reduces to conceptualization (mental
experience).
Thesis B: A frequently-used expression typically displays a
network of interrelated senses.
® Ronald W. Langacker

50
RONALD W. LANGACKER
Thesis C: Semantic structures are characterized relative to
"cognitive domains".
Thesis D: A semantic structure derives its value through the
imposition of a "profile" (designatum) on a "base".
Thesis E: Semantic structures incorporate conventional
"imagery", i.e. they construe a situation in a particular
fashion.
We will consider these claims in the order listed, the later ones requiring
more extensive discussion. While all of them are in some measure
controversial, I am more concerned with articulating a coherent view than with
arguing against alternatives or reviewing the history of pertinent issues.
With Thesis A, I embrace a form of the "conceptual" or "ideational"
theory of meaning that semantic theorists so commonly feel obliged to
challenge (e.g. Lyons 1977: 1/113; Kempson 1977: 15-20; Palmer 1981: 24-29).
Conceptualization is ultimately to be explicated in terms of cognitive
processing; there is nothing inherently mysterious about its basic character, nor
is cognition beyond the realm of scientific inquiry. I speak of
"conceptualization" in order to emphasize the "subjective" nature of linguistic meaning.
I understand the term in a maximally inclusive way with respect to the
domain of mental experience: it subsumes both established concepts and
novel conceptions; includes sensory, emotive, and kinesthetic sensations;
and extends to our awareness of the physical, social, and linguistic context.
Semantic structure is therefore treated as a special case of conceptual
structure (as it is in Jackendoff 1983, which otherwise shows virtually no
resemblance to the present framework in its view of meaning and
grammar). Linguistic expressions (including the symbolic units of lexicon and
grammar) are "bipolar", each having a "semantic pole" and a
"phonological pole"; a "semantic structure" is thus definable as a conceptualization
conforming to linguistic convention that functions as the semantic pole of
an expression. I will also refer to semantic structures as "predications", and
to the semantic pole of a morpheme (regardless of type) as a "predicate".
Hence all predicates qualify as predications, but the predications
corresponding to symbolically-complex expressions are not predicates. An
important feature of cognitive-grammar semantics is that predications of all sizes
are characterized by means of a single array of descriptive constructs.
Thesis B simply recognizes that polysemy is the norm for lexical units,
and must therefore be accommodated by linguistic theory as a natural,

A VIEW OF LINGUISTIC SEMANTICS
51
unproblematic phenomenon. I propose a network model, where each node
in the network represents one established sense of the lexical item, and
each arc connecting two nodes indicates the nature of their association.
Nodes are associated via "categorizing relationships", of which there are
two basic types: relationships of "schematicity" (indicated by solid arrows)
and of "extension" (marked by dashed arrows). The notation [[A] —> [B]]
says that [A] is "schematic" relative to [B], which constitutes an
"elaboration" or "instantiation" of [A]; [B] is consistent with [A], but [BJ's
specifications are more detailed and achieve a finer "grain" than [A]'s (hence the
relationship is one of semantic "specialization", or its inverse,
"abstraction"). By contrast, semantic extension implies some conflict in value; in
the relationship [[A] -~> [B]], certain specifications of the basic sense [A]
must be suspended or modified to arrive at the extended sense [B].
These notions are illustrated in Fig. 1, which shows a fragment of the
network describing the conventional meanings of the English noun ring.
The concept 'circular entity' is schematic relative to 'circular mark' and
'circular object'; the latter is in turn schematic relative to 'circular piece of
jewelry', which is further elaborated by conceptions representing jewelry
worn on different parts of the body. The term ring is also applied to an
arena used in boxing, wrestling, or bullfighting; the sense 'arena'
constitutes an extension relative to 'circular object' or 'circular entity' because it
does not incorporate the specification of circularity (boxing and wrestling
rings are typically rectangular). It is certainly not assumed that all these
nodes and categorizing relationships are equal in status. The nodes vary
greatly in their degree of entrenchment or cognitive salience, some being
far more readily activated than others. In the case of ring, for instance, the
sense 'circular piece of jewelry worn around finger' is presumably the
category "prototype", and is likely to be activated in preference to others in a
neutral context. Moreover, categorizing relationships vary not only in their
cognitive salience, but also in their "distance" (i.e. the extent to which [A]
must be extended or elaborated to yield [B]).

52 RONALD W. LANGACKER
[CIRCULAR ENTITY \- ~_ _ ___
~ ^ "" ""M ARENA I
CIRCULAR MARK] [CIRCULAR OBJECTK
CIRCULAR PIECE OF JEWELRYl
CIRCULAR PIECE OF JEWELRY
WORN AROUND FINGER
CIRCULAR PIECE OF JEWELRY
WORN THRU NOSE
Figure 1
This network, when fully articulated, represents the conventional
usage of the lexical item, i.e. the range of conceptions for which it is
normally employed. The precise form of the network may vary across
speakers, depending on their experience and the categorizing judgments they
happen to have made, but this will hardly be a problem for communication
provided that enough nodes are shared. Also, we have no immediate way
of knowing the degree of "delicacy" that speakers achieve in their mastery
of conventional usage (i.e. how far "downward" they articulate the network
into progressively more specialized applications), nor the degree of
schematization (how far "upward" they build the network through
successive levels of abstraction). But the specific details are less important than
the appreciation that some type of network must in fact be posited: any
attempt to reduce the meaning of a lexical item to a single node will in
general prove unworkable. The obvious candidates for such reduction are the
category prototype and the highest-level schema, corresponding
respectively to the primary sense and the most abstract sense. However, we
cannot presume that every lexical category has a clearly identifiable prototype,
nor that speakers invariably extract a "superschema" having all other nodes
in the network as elaborations (none is shown in Fig. 1); often the necessary
abstractness of an all-subsuming superschema would leave it essentially
void of content. Furthermore, even if a prototype or superschema were
always available, it would afford no basis for predicting precisely which
array of extensions and elaborations — out of all those that are conceivable
and semantically plausible — happen to be conventionally established in
the language. The conventional usage of a lexical item must be learned, as

A VIEW OF LINGUISTIC SEMANTICS
53
it is not strictly predictable.
A speaker's knowledge of the conventional meaning of a lexical unit
therefore embraces the entire network, and must be characterized as such.
Each node in such a network represents a distinct predication, and together
with the phonological pole it defines a distinct "semantic variant" of the
lexical item. Nevertheless, since polysemy is not our immediate concern, the
remainder of our discussion will focus on individual nodes. What is
necessary to characterize a lexical item (or some novel expression) when it is
understood in one particular way?
Cognitive Domains
According to Thesis C, linguistic predications are characterized
relative to "cognitive domains". The basic observation is that certain
conceptions presuppose and thus incorporate others, which provide the necessary
basis for their emergence and characterization. For instance, the notion
hypotenuse presupposes the conception of a right triangle, being incoherent
without it; it thus incorporates this conception as a necessary part of its own
characterization, i.e. as a cognitive domain. Similarly, the notion corner
kick presupposes, as one of its domains, a considerable body of knowledge
pertaining to the rules and objectives of football.
We can therefore reasonably posit hierarchies of conceptual
complexity, where a concept at one hierarchical level derives through various
cognitive operations performed on concepts at lower levels, as sketched in Fig. 2.
These operations, indicated by double arrows, include the coordination of
lower-level concepts (i.e. the integration of their contents), and also
adjustments pertaining to profiling and other dimensions of imagery, as discussed
in later sections. For example, by coordinating multiple instances of the
notions line segment and angle, we obtain the conception triangle. That,
together with the concept of perpendicularity, yields the notion right
triangle, which functions as a cognitive domain for hypotenuse. This latter
notion is in turn a potential domain for such higher-order conceptions as
midpoint of the hypotenuse, and so on indefinitely.

54
RONALD W. LANGACKER
w
A
B
y
C
increasing complexity-
Figure 2
The central claim is that a conceptualization occupying any level in
such a hierarchy can function as a cognitive domain for the characterization
of a linguistic predication. In Fig. 2, for instance, [A] is depicted as a
complex notion serving as the domain for predication [B], obtained by
performing an operation on [A] and coordinating it with additional content.
However, since a predication is itself a conceptualization, [B] can itself be
invoked as the cognitive domain for predication [C], derived by further
operations. The import of this claim is that the starting point for the
semantic analysis of a given expression is an integrated conceptualization which
may have any degree of internal complexity. Obvious though it may seem,
this type of account differs significantly from certain others that are
commonly assumed. First, it contrasts with the view that the meaning of an
expression is directly describable in terms of a fixed vocabulary of semantic
primitives; a cognitive domain need not be primitive, nor is there any fixed,
limited set of them. Second, it rejects the idea that an expression's meaning
is represented as a bundle of semantic markers or features; a cognitive
domain is an integrated conceptualization in its own right, not a feature
bundle.
What occupies the lowest level in hierarchies of conceptual
complexity? I do not necessarily posit conceptual primitives, being essentially
neutral as to their existence. It is however necessary to assume some inborn
capacity for mental experience, i.e. a set of cognitively irreducible
representational spaces or fields of conceptual potential; I refer to these as
"basic domains". Among these basic domains are the experience of time
and the ability to conceptualize configurations in 2- and 3-dimensional
space. There are basic domains associated with the various senses: color
space (i.e. a range of possible color sensations), the ability to perceive a

A VIEW OF LINGUISTIC SEMANTICS
55
particular range of pitches, domains defining possible sensations of taste
and smell, and so on. Moreover, special domains are doubtless required for
emotive and kinesthetic sensations. I offer no specific inventory of basic
domains — the important point is simply that certain types of mental
experience are cognitively irreducible. Presumably we begin constructing
our mental universe out of experience registered in basic domains, arriving
at ever higher levels of conceptual organization by means of innately
specified cognitive operations.
Some linguistic predications may be characterized directly with respect
to basic domains; possible examples are red (color space), hot
(temperature), line (2-dimensional space), and before (time). However most
predications presuppose domains representing higher levels of conceptual
organization. For instance, the conception of a right triangle, the domain
for hypotenuse, is non-basic (or "abstract") because it is reducible to more
fundamental notions; the fact that a right triangle occupies 2-dimensional
space (and thus incorporates this basic domain) does not obviate its
conceptual (hence non-basic) status. Similarly, the noun tip presupposes the
conception of an elongated (generally thin) object, an elbow is characterized in
part by its position within the overall configuration of the human arm, April
invokes for its definition the complex notion of a cycle of months
exhaustively dividing a year, the verb castle presupposes substantial knowledge of
the rules and strategies of chess, and so forth. There is obviously no fixed or
universal set of cognitive domains. Any facet of our mental experience or
conceptual world can be appropriated to function in this capacity.
In principle, then, to describe a predication exhaustively one would
have to furnish complete descriptions of the cognitive domains it
presupposes. This would in turn require characterizations of the more
fundamental conceptions that each domain itself presupposes, and so on, until finally
one reached the level of basic domains. One would also need to
characterize the various operations and abilities by means of which these
hierarchies of conceptual complexity are constructed during the course of
cognitive development. In short, a complete and definitive account of semantic
structure would require an essentially full description of developmental
cognition and the conceptual hierarchies it produces!
This is obviously not a practical short-term goal. In practice we must
do the best we can with partial and provisional accounts, describing
relevant domains to the best of our ability in whatever degree of detail is
minimally necessary for particular purposes. It is important, though, to realize

56
RONALD W. LANGACKER
what a cognitively accurate description of semantic structure would actually
require were it attainable. There is no point in deceiving ourselves by
maintaining the gratuitous belief that meaning can be understood without a
comparable understanding of cognitive processing and the conceptual
hierarchies and knowledge systems to which it gives rise. Semantic structure is not
a separate or autonomous "module" of psychological organization that can
be pulled out and studied in isolation from the full, rich fabric of our mental
experience. Viewing semantic structure as a self-contained, algorithmically
describable component of the linguistic system is theoretically convenient,
but it necessarily impoverishes and distorts the subject matter in drastic
ways.
Another consideration leading to a similar conclusion pertains to the
number of cognitive domains presupposed by a given predication. The full
set of domains required for a predications's description will be called its
"matrix". The matrix is said to be "complex" if it consists of more than one
domain, as it typically does. Even with basic domains, one easily finds
examples where a single domain is not sufficient. For instance, a beep is a
sound of short duration that to some extent approximates a pure tone; a
characterization of this notion therefore presupposes the basic domains of
both time and pitch. A flash is a brief, intense, and possibly diffuse light
sensation; the description of this noun consequently refers to time, color
space (particularly the brightness dimension), and the extensionality of the
visual field.
For predications invoking non-basic domains, the matrix is often quite
large, and sometimes open-ended. The complex matrix of knife is sketched
in Fig. 3. One specification of this predicate is a characterization of the
typical shape of a knife (or a family of typical shapes), presupposing the basic
domain of space. Another salient specification is the role a knife plays as an
instrument of cutting; the conception of a canonical episode of cutting
therefore functions as a second, non-basic domain. Other specifications,
each invoking some non-basic domain, pertain to the normal size of a knife;
its weight; the material from which it is made; the inclusion of a knife
(together with a fork and a spoon) in a standard place-setting; the games
played with knives; their cultural associations; the existence of
knife-throwing acts in circuses; the role of knives in the conquest of the American
frontier; and so on indefinitely. What ties these diverse conceptions together is
the participation in each of them of the entity that knife "designates" (see
below).

A VIEW OF LINGUISTIC SEMANTICS
57
SPACE
DOMAIN 1
_j_—
r^n
CUTTING
DOMAIN 2
"-•••...
0 i w
\
H1
SILVERWARE
*••
DOMAIN 3
COMPLEX MATRIX |
KNIFE
Figure 3
A fundamental question of semantic theory arises at this juncture.
Which of the specifications in a complex matrix are part of the conventional
meaning of a linguistic expression, and which ones represent extra-linguistic
knowledge? In fact, can such a distinction be made at all? The standard
view is that it can. Theorists generally assume that the meaning of a word
like knife is not to be equated with a speaker's vast and open-ended
knowledge of the objects in question, but rather with some restricted portion of
this overall body of knowledge; the linguistic description of a term's
conventional value more closely resembles an entry in a dictionary than an
entry in an encyclopedia. A definite division is therefore presumed possible
between linguistic and extra-linguistic knowledge (or between "semantics"
and "pragmatics"). The problem with this view is that the facts of language
offer little basis for accepting its validity. Haiman (1980) has surveyed the
obvious rationales that might be invoked for drawing the line between the
linguistic and extra-linguistic knowledge associated with a term, and found
them all to be inadequate. In truth, the existence of a clear-cut boundary
has been assumed on methodological (not factual) grounds: only under this
assumption can semantic structure (and more generally, linguistic
structure) be described as a self-contained system amenable to algorithmic
formal treatment. Theoretical convenience does not, however, constitute
empirical evidence. I see no a priori reason to accept the reality of the
semantics/pragmatics dichotomy.
Instead, I adopt an "encyclopedic" conception of linguistic semantics.
I posit no specific boundary between our linguistic and non-linguistic
knowledge of the entity designated by a term, such that all those
specifications on one side of the boundary clearly fall within the purview of seman-

58
RONALD W. LANGACKER
tics, while all those on the other side are safely relegated to pragmatics. Far
more realistic, I believe, is to posit a gradation of "centrality" in the
specifications constituting our encyclopedic knowledge of an entity: some
domains and specifications are obviously more salient and linguistically
important than others, and in practice we will focus our attention primarily
on these, but the imposition of any precise or rigid boundary is considered
arbitrary. A number of factors contribute to a specification's degree of
centrality, among them whether the specified property is inherent or
contingent, whether it is generic or specific, whether it is peculiar to the entity
in question or shared by many others, and so on (for fuller discussion, see
Langacker in press: ch. 4). At the level of cognitive processing, centrality
can be explicated as the likelihood that a particular specification will be
activated on a given occasion when a term is used.
The prevalence of the "dictionary" view of linguistic semantics is
attributable in part to the ubiquitous "container metaphor", which
permeates our thought about language and is reflected in how we talk about it
(cf. Reddy 1979). This metaphor portrays lexical items as "containers" for
an abstract substance called "meaning"; speakers exchange meanings by
sending strings of such containers back and forth. Any actual container has
a definite and limited volume, so the metaphor encourages us to think of a
lexical item as having some restricted set of semantic specifications — we in
fact have some initial difficulty in conceptualizing the meaning of a term as
extending indefinitely. We must not allow the formulation of linguistic
theory to be dictated by a metaphor of dubious validity. As an alternative
to the container metaphor, we can profitably regard a lexical item as
providing "access" to knowledge systems of open-ended, encyclopedic
proportions. Speakers do not send meanings to one another (all they transmit, in
fact, is sound waves). Rather, their detection of a lexical item in the speech
stream enables them to activate selected portions of the knowledge base
they already possess; communication is possible to the extent that the
knowledge systems of different speakers are comparable, but it is never
perfect, since these systems are never fully identical.
Profile and Base
Thesis D holds that the imposition of a "profile" on a "base" is crucial
to the value of every linguistic predication. The base for a predication is
nothing more than its matrix, i.e. the set of operative domains (or more

A VIEW OF LINGUISTIC SEMANTICS
59
precisely, those portions of such domains which the predication actually
invokes and requires). Some facet of the base is invariably raised to a
distinctive level of prominence, and serves (intuitively speaking) as its focal
point; this substructure is the predication's profile. I do not know whether
profiling reduces to any independently attested psychological phenomenon.
Factors that suggest themselves as possibly being relevant include figure/
ground organization, focus of attention, and level of activation, but the
simple identification of profiling with any such notion is problematic. Whatever
its cognitive basis turns out to be, profiling is essential and fundamental to
linguistic semantics.
I will further describe the profile as comprising those portions of the
base which the predication "designates". (Designation, then, is not to be
understood as the relation between a linguistic expression and the "world",
but rather one that holds between an overall conceptualization and certain
of its subparts.) For instance, the conception of a right triangle functions as
the base for hypotenuse, whose profile (or "designatum") is one of the
constituent line segments, as illustrated in Fig. 4(a). When predications are
diagrammed, the profiled entity is depicted with heavy lines to indicate the
special prominence that distinguishes it from the remainder of the base.
(a) (b) (c)
HYPOTENUSE
Figure 4
Profiling is one of the cognitive operations responsible for the growth
of hierarchies of conceptual complexity (cf. Fig. 2): a predication
presupposes and incorporates its base, from which it derives a distinct, higher-
order conception through the imposition of a particular profile. An
expression's meaning does not, then, reside in either the base or the profile alone.
Both facets of a predication are crucial to its value, which reflects the
elevation of some particular part of the base to a special level of prominence and

60
RONALD W. LANGACKER
functional significance. For instance, if we suppress the profiling of
hypotenuse, as in Fig. 4(b), what results is no longer the conception of a
hypotenuse, but simply that of a right triangle; if we suppress the unprofiled
portions of the base, as in Fig. 4(c), there is no basis for identifying the
remaining line segment as being a hypotenuse, which exists only in the
context of a right triangle. The base of a predication can thus be thought of as
the "frame" needed to establish the character and identity of the intended
designatum: a person qualifies as a cousin only when linked to another
individual through a series of kinship connections; a span of idle time
constitutes an intermission only if it interrupts some type of performance; and
only the extremity of an elongated object can be recognized as a tip.
All expressions, regardless of complexity, are characterized semanti-
cally by the imposition of a profile on a base. A pattern of grammatical
composition must specify, at its semantic pole, precisely which entity (out
of those contributed by the component expressions) is selected as the
profile of the composite expressions it derives. What is traditionally known as
the "head" of a given construction is simply that component from which the
composite expression inherits its profile (and hence its grammatical
categorization). For example, consider the nominal expressions in (2),
assuming that each of them invokes for its base the conceived situation
sketched in Fig. 5.
(2) (a) the lamp above the table
(b) the table below the lamp
(c) the leg of the table below the lamp
(d) the light from the lamp above the table
Figs. 5(a)-(d) represent the composite values of (2)(a)-(d) respectively.
These expressions contrast semantically, but not by virtue of any
differences in their conceptual content — we have assumed a common base,
incorporating the conception of a lamp (hence light), a table (hence legs),
and the locative relationship between the two. The semantic contrasts are
attributable primarily to the imposition of alternate profiles on this base;
starting from the same basic conception, each expression selects a different
substructure as the one it designates. It is readily observed that the entity
designated by the expression as a whole is in each case the same one that is
profiled by the head noun. The compositional pattern permitting a noun to
be modified by a prepositional phrase describes the integration of their
contents to form a coherent composite conception, and further specifies that

A VIEW OF LINGUISTIC SEMANTICS
61
the composite-structure profile is the same entity that is designated by the
noun.
(c)
r
W
s
Figure 5
The expressions just considered have all been nominal in character,
but the notion of profiling (designation) is also considered applicable to
relational predications (including verbs, adverbs, adjectives, prepositions,
and others). A relational predication profiles the "interconnections" among
conceived entities, where an "entity" can be either a "thing" or some other
relation; interconnections can be thought of as cognitive operations
assessing the relative position of entities within a domain. For exemplification,
let us compare the meanings of go, away, and gone, as reflected in (3):
(3) (a) You've been here long enough — please go now.
(b) California is very far away.
(c) By the time I arrived, she was already gone.
All these words are polysemous, but we will in each case confine our
attention to the single sense illustrated.
As a verb, go designates a "process", i.e. a series of relational
configurations followed sequentially through conceived time. Time is therefore
one domain in its matrix, as indicated by the arrow in Fig. 6(a), and space
is another. The diagram shows explicitly only four component "states" —
including the initial and final states — out of the continuous series that the
process comprises. There are two major participants, represented diagram-
matically by circles. I refer to one of them as the "trajector" (tr), and to the
other as the "landmark" (/m). The dotted lines signify "correspondence";
they indicate that the trajector is the same from one state to the next, as is
the landmark. The heavy, dashed lines connecting the trajector and
landmark within each state represent the profiled interconnections responsible
for the relational character of the predication, i.e. the cognitive operations
that register the position of the trajector within the domain relative to that

62
RONALD W. LANGACKER
of the landmark. The initial state finds the trajector in the landmark's
"neighborhood", given as an ellipse. The trajector's position changes from
state to state with the passage of time, and in the final state it lies outside
the neighborhood of the landmark.
(a) GO Cb) AWAY (c) GONE
Figure 6
Away is not a verb, and time is not an active domain for the semantic
variant under examination. In (3)(b), away profiles a single spatial
configuration, diagrammed in Fig. 6(b). Observe that this configuration is the
same one that constitutes the final state in the profile of go: it situates the
trajector outside the vicinity of the landmark. Hence the process designated
by go results in a locative relationship appropriately described by away.
However, we see from examples like (3)(b) that this relationship need not
be construed as coming about by any motion of the trajector (presumably
California is stationary for the time period under consideration).
What about gone? An examination of Fig. 6(c) reveals that it differs
from both go and away, but resembles each of them in a certain respect.
Gone matches away in profiling a single locative relationship in which the
trajector is outside the neighborhood of the landmark. These two
predications differ, however, in their base: for the relevant sense of away, the base
is simply some portion of the spatial domain, but the base for gone is
precisely the process profiled by go. That is, something is properly described as
being gone only if its position results from an instance of the process of
going, hence the conception of this process constitutes its base. Go and
gone are therefore identical in their basic conceptual content — they differ
in the profile imposed on this content. This difference in profiling is the
semantic contribution of the past participial predication (which itself has
several semantic variants — see Langacker 1982a). The participial
morpheme takes a schematically-characterized process as its base, and profiles

A VIEW OF LINGUISTIC SEMANTICS
63
only the final state of this process. The participial construction identifies the
schematic base of the participial morpheme with the specific process
profiled by the verb stem, but specifies that the profile of the former prevails at
the composite-structure level (i.e. the participial morpheme is the "head" in
this construction). In short, the effect of the participial inflection, at the
semantic pole, is to convert a process like go into a non-processual relation
that profiles only the final state of that process.
Imagery
The remainder of this paper deals with Thesis E. It claims that
semantic structures incorporate conventional "imagery", where that term is
understood in a special technical sense. By imagery, I do not refer to visual
imagery or sensory imagery more generally, as studied by Kosslyn (1980),
Shepard (1978), and others (though I believe this phenomenon to be both
cognitively and semantically significant). I refer instead to our amazing
mental ability to "structure" or "construe" a conceived situation in many
alternate ways. In view of its obvious significance, it is surprising that this
ability has received so little systematic attention. I would argue, in fact, that
imagery is crucial to a revealing account of either semantic or grammatical
structure, and provides the necessary foundation for a subjectivist theory of
meaning.
For the description of lexical and grammatical structure, cognitive
grammar posits only symbolic units, each attributed some measure of
conceptual "content". This content, moreover, is structured in a particular
fashion, i.e. the unit incorporates conventional imagery, which contributes
to its semantic value. Because of imagery, two expressions that have the
same content, or describe the same objective situation, may nevertheless
have different meanings; recall, for example, that go and gone have the
same content (cf. Fig. 6), but differ semantically due to profiling. The
lexical and grammatical resources of a language are therefore not semantically
neutral — inherent to their nature is the structuring of conceptual content
for symbolic purposes. I should emphasize that the distinction between
content and imagery is drawn in part for expository convenience, and should
not be thought of as a strict dichotomy. These two facets of meaning grade
into one another and may well be indissociable: every conception reflects
some particular construal of its content, if only by the selection of default-
case options.

64
RONALD W. LANGACKER
Linguistic semantics must therefore accommodate conventional
imagery in a natural and integral manner. Imagery cannot be treated as an
afterthought — its various dimensions are intrinsic to the conceptualization
process, and a linguistic predication, by its very formation, assumes particular
values along these parameters. Some important dimensions of imagery are
catalogued in (4); they will be discussed in the order listed:
(4) (a) level of specificity
(b) background assumptions and expectations
(c) secondary activation
(d) scale and scope of predication
(e) relative salience of substructures
(f) perspective
Some of these dimensions are further divisible into factors requiring
separate consideration. No substantive claim is intended by either the groupings
or the sequencing.
The first dimension of imagery pertains to the degree of precision and
detail with which an entity is characterized. Here the indissociability of
imagery from content is quite apparent: in an obvious sense, a schematic
notion has less content than one spelled out with finer "grain" or greater
"resolution". Level of specificity has already been noted as one parameter
of lexical polysemy; certain conventional senses of a lexical item are
commonly schematic (or conversely, specialized) relative to others (cf. Fig. 1).
The same type of relationship often holds between distinct lexical items, in
which case it is common to speak of "hyponymy" or "taxonomic
hierarchies". We thus find ordered series of terms like those in (5)(a)-(b), where
the solid arrows indicate relationships of schematicity.
(5) (a) move -> locomote -» run —» sprint
(b) animal —» reptile —» snake —> rattlesnake —» sidewinder
(c) long —> over 6 feet long —> about 6-1/2 feet long —>
precisely 6 feet 7 inches long
I also consider each expression in (5)(c) to be schematic for the one
that follows: a schema introduces certain domains (a scale of length in this
case), and specifies a relatively broad range of possible values within those
domains; its instantiations confine the permitted values to narrower ranges
(and may themselves introduce further domains). We see from this example
that level of specificity, like all the other dimensions of imagery, is
applicable to novel expressions of any degree of complexity, not just lexical items.

A VIEW OF LINGUISTIC SEMANTICS
65
Whenever a speaker constructs a sentence, he necessarily opts to portray
the conceived situation with a certain degree of specificity. For example, he
might use any of the sentences in (6) to describe the same event:
(6) (a) I saw an animal and moved on.
(b) I saw a long snake and ran away.
(c) I saw a rattlesnake about 6-1/2 feet long and sprinted to
safety.
While the last sentence is certainly the most dramatic, a speaker more
concerned with preserving a reputation for fearlessness might well prefer the
greater schematicity of (6)(a).
Relationships of schematicity also play a significant role in grammatical
constructions. A construction is a pattern for integrating two or more
"component" structures to form a "composite" structure. Patterns of this sort are
described in cognitive grammar by means of "constructional schemas":
complex symbolic structures that are fully parallel in formation to
instantiating expressions, but characterized at a level of schematicity that
neutralizes their differences. As the embodiment of a structural regularity, a
constructional schema can be regarded as a template for the computation of
novel instantiating expressions, which are obtained by elaborating its
schematic components. For example, the structure on the left in Fig. 7
sketches the semantic pole of the constructional schema describing the
integration of a modifying adjective with a head noun (in a full description,
relations between the component and composite structures would similarly
be given'for the phonological pole — in the case of English, they specify
that the adjective immediately precedes the noun). A specific expression
such as long snake is judged well-formed by virtue of instantiating the
schema, and the categorizing relationship between the two constitutes the
expression's "structural description".

66
RONALD W. LANGACKER
CONSTRUCTIONAL INSTANTIATING
SCHEMA EXPRESSION
ADJ-N
u-~©
_y\
lm tr
; ADJ
—^
O
N
LONG-SNAKE
1 I
/\
LONG
—^-
fj]t
SNAKE 1
Figure 7
In a typical construction, schematicity also figures in the syntagmatic
relationship between the two component structures. Consider the
constructional schema of Fig. 7. The nominal predication profiles a "thing",
indicated by a circle. The adjectival predication is relational, profiling the
interconnections between its trajector, also a "thing", and some kind of
landmark. Every construction hinges on correspondences established between
subparts of the component structures; in this ADJ + N construction, the
pivotal correspondence holds between the trajector of the adjective and the
thing profiled by the noun. The dotted correspondence line thus identifies
the adjectival trajector with the nominal profile, but the adjectival
predication generally characterizes this entity only schematically, while the noun is
far more specific about its properties; this additional information is labeled
'X' in the diagram. There is consequently a relationship of schematicity
between the trajector of ADJ and the profile of N, as indicated by the solid
arrow. The composite structure, shown at the top, retains the profiling of
the noun, but to its characterization of the profiled entity is added — as
part of the base — the participation of this entity in the relation designated
by the adjective.
In the particular case of long snake, the adjectival predication has for
its domain the conception of a length scale. The relational interconnections

A VIEW OF LINGUISTIC SEMANTICS
67
specify that the trajector, when aligned with this scale, extends into the
region that lies beyond the neighborhood of the norm (n) for its category; this
region constitutes the relational landmark. Long characterizes its trajector
quite schematically. This schematic trajector is placed in correspondence
with the specific "thing" designated by snake, which provides much fuller
information about the correspondent, i.e. the noun elaborates the
schematic trajector of the adjective. The resulting composite structure profiles the
snake, but incorporates in its base the information furnished by the
adjective concerning its length. Observe that the instantiating expression is
precisely parallel to the constructional schema in the structures it incorporates
and the relationships among them. In particular, the component and
composite structures of long snake elaborate their respective schematic
counterparts, and the integration of long and snake follows the pattern represented
in the schema.
The second dimension of imagery, construal relative to different
background assumptions and expectations, will not be discussed in any
detail. It subsumes a variety of phenomena, including the familiar contrast
between (7)(a) and (b):
(7) (a) The glass is half-empty,
(b) The glass is half-full.
These sentences are true under the same objective circumstances, but
certainly they differ in meaning (cf. Tuggy 1980). Though I will not consider it
here, presupposition certainly falls under the rubric of background
assumptions. The same is true of discourse phenomena, such as the topic-comment
asymmetry, the given-new distinction, and the occurrence in English of
contrastive (or unreduced) stress on those elements considered informative
or otherwise noteworthy:
(8) (a) He likes LIVER.
(b) He LIKES liver.
(c) HE likes liver.
The base of a predication can also be regarded as the background for
viewing its profile. Recall that away and gone profile the same relational
configuration, but construe it against different bases.
One additional example is the semantic contrast between few and a
few. The sentences in (9) could perfectly well be used to describe the same
situation (let us suppose that precisely three friends are involved in each
case), but they clearly differ semantically.

68
RONALD W. LANGACKER
(9) (a) He has few friends in high places,
(b) He has a few friends in high places.
On intuitive grounds, one is inclined to say that few is somehow negative,
while a few is basically positive. This is corroborated by the data in (10);
only few provides the negative context required by any:
(10) (a) Few people have any friends in high places,
(b) * A few people have any friends in high places.
I suggest that few specifies the negative departure of a quantity from an
implicit norm or expectation; for instance, (9)(a) indicates that the number
of friends is less than a well-connected person ought to have. By contrast, a
few carries no special assumptions, and simply specifies that the quantity
departs somewhat from the baseline of zero.
The next dimension of imagery, secondary activation, pertains to the
network model of complex categories, as illustrated in Fig. 1. The network
model has general applicability, and is employed for such disparate classes
as the alternate senses of a lexical item, the allophones of a phoneme, and
the family of complex structures — including schemas, subschemas, and
instantiating expressions — representing the conventional range of a
grammatical construction (cf. Langacker in press: ch. 10). The network
describing a complex category consists of a set of structures (nodes) connected by
categorizing relationships. The structures may have any degree of internal
complexity (e.g. an entire constructional schema might function as a single
node). In the case of lexical polysemy, our prime concern here, the nodes
comprising a network are symbolic units, which are distinct at the semantic
pole but identical at the phonological pole (note that only the semantic pole
is depicted in Fig. 1). Each node thus constitutes a semantic variant of the
lexical item.
For the most part, only one semantic variant is directly relevant when
a speaker employs a particular lexical item in a specific context. Suppose a
term is needed to describe the arena for a boxing match. In this context, the
symbolic unit [ARENA/ring] is far more relevant than [CIRCULAR
MARK/ring], [CIRCULAR OBJECT/ring], or even the prototype
[CIRCULAR PIECE OF JEWELRY WORN AROUND FINGER/ring],
despite its peripheral status in the ring category. The variant [ARENA/
ring] is the one that the speaker activates when he decides to use the term:
he matches its semantic specifications against his conception of the entity to
be described, and accepts the term as appropriate if this categorizing judg-

A VIEW OF LINGUISTIC SEMANTICS
69
ment reveals these notions to be sufficiently compatible. The node invoked
in this fashion for primary categorization of the target conception will be
called the network's "active node". Clearly, different nodes serve this
function on different occasions.
We can reasonably assume that activation tends to spread across
associated cognitive structures (cf. Collins and Loftus 1975). Hence the
primary activation of a network's active node tends to elicit the secondary
activation of other nodes linked to it by categorizing relationships. While
the active node determines the basic applicability of a term, the additional
nodes activated secondarily contribute their own nuances and thereby
enrich the semantic value that the term acquires on a particular occasion of
its use. Let us assume, for example, that [ARENA/ring] is categorized as a
semantic extension from the more basic variant [CIRCULAR OBJECT/
ring]. In this case, the primary activation of the extended variant in
describing a boxing arena is likely to evoke the secondary activation of the basic
variant. The subsidiary activation of [CIRCULAR OBJECT/ring]
reinforces and thus renders more salient the notion of an enclosure completely
surrounding the locus of activity. There is no such effect when the term
arena is used instead, since [ARENA/arena] is not directly linked by any
categorizing relationship to [CIRCULAR OBJECT/ring] and does not
elicit its secondary activation.
Metaphorical expressions are simply more extreme instances of
semantic extension. For instance, the conventional usage of pig to designate a
glutton implies the semantic variant [GLUTTON/pig], which is categorized
as an extension from the basic variant [PIG/pig] and evokes its secondary
activation. The same phenomenon is observable with morphemically
complex expressions, e.g. established idioms like spill the beans or take the bull
by the horns. When these expressions are used idiomatically, the extended
or "figurative" sense functions as the active node — it represents the actual
notion to be conveyed — while the basic or "literal" sense is activated
secondarily. This two-level semantic representation accounts for the special
quality of metaphorical expressions, but the enhancement of meaning by
the activation of subsidiary nodes is present to some degree in virtually
every expression. The metaphorical use of novel expressions has the same
fundamental character, the only difference being that the extended variant
and categorizing judgment have not yet achieved the status of conventional
units.

70
RONALD W. LANGACKER
Scale and Scope of Predication
The "scale" of a predication is roughly analogous to the scale of a map.
Many predications presuppose for this parameter a value that falls within a
certain range. For example, a body of land completely surrounded by water
is called an island, but this term would not normally be used for a handful
of mud in the middle of a puddle — it is only applicable to situations on a
larger scale. In similar fashion, the term continent is applied to land masses
on the scale of Australia, Europe, and North America, but hardly to
Ireland, Tahiti, or Luxembourg. The contrast between bay and cove is also
one of scale: whereas either might refer to precisely the same land/water
configuration, cove is limited to small-scale configurations, for which bay is
marginal. Scale of predication is important even for relational expressions
like near and close, which may assume essentially any value along this
parameter. An expression of the form A is near B or A is close to B is
equally appropriate for galaxies, nations, objects in a room, and atoms in a
molecule, indicating in each instance that A lies in the neighborhood of B.
But what constitutes a neighborhood? The absolute distances involved vary
enormously from one example to the next. Obviously, the neighborhood
conception must be characterized in terms of relative distances assessed
with respect to relationships on a particular scale.
The "scope" of a predication is its coverage in relevant domains, i.e.
how much is included in the conception that functions as its base. Every
predication has some implicit scope, though it need not be precisely
delimited or cognitively salient. Consider the land/water configuration depicted
in Fig. 8, where the outer box and the dashed lines indicate alternate scopes
of predication. Scope (a) is sufficient to identify the land mass as an island,
but (c) and (d) are not; to qualify as an island, a land mass must be
completely surrounded by water, and neither (c) nor (d) is inclusive enough to
provide this information. What about scope (b), which contains only the
land together with a narrow strip of water all around its periphery?
Interestingly, we could not properly identify the land as an island with only this
much coverage; it is possible, for instance, that the strip of water is a moat
protecting a castle, but it is not conventional to speak of the land inside a
moat as an island. The water surrounding an island must have a certain
expanse — it must extend outward from the island a substantial distance (at
least on most sides). Though the scope of predication is not sharply
delimited, it must at least be large enough to reveal this expanse. As another

A VIEW OF LINGUISTIC SEMANTICS
71
example, consider the lexeme peninsula. In this case we find that scopes (a)
and (c) suffice, while (b) and (d) are inadequate for different reasons. The
insufficiency of scope (b) is similar to that observed for island: a certain
minimal expanse is required for the body of water into which a peninsula
projects. Scope (d) fails because the finger of land constituting a peninsula
must be fairly small relative to the overall land mass of which it is a part;
scopes (a) and (c) are sufficient to establish the relative sizes, but (d) is not.
Figure 8
The scope construct has substantial grammatical significance (cf. Casad
and Langacker 1985). For one thing, the scope of a predication can be
referred to explicitly in the characterization of a grammatical construction.
Consider the "nested locative" construction illustrated in (11).
(11) (a) The rake is in the yard by the back fence near the gate,
(b) The camera is upstairs in the bedroom in the large closet
on the top shelf underneath the quilt.
Each sentence contains a series of locative expressions, which collectively
describe the location of the entity designated by the subject. Under the
relevant interpretation, each successive locative in the series narrows the
"search" for this entity, confining it to a region that is properly included in
the region singled out by the locative that immediately precedes. Following
Hawkins (1984), I will use the term "search domain" to indicate the region

72
RONALD W. LANGACKER
to which a locative predication confines its trajector (subject).
Fig. 9 diagrams the syntagmatic relations connecting the three locatives
of (ll)(a). Each locative has a scope of predication, and within that scope
it specifies the position of the trajector relative to a landmark. It is
mandatory in this construction that the trajector be the same for all the locatives in
the sequence; the identity of the trajector from one predication to the next
is indicated by dotted correspondence lines. Cross-hatching is used to mark
the search domain: each locative tells us that the trajector is somewhere
within this area. The first one, in the yard, confines the trajector to the
interior of its landmark (namely, the yard). The other two landmarks, by
the back fence and near the gate, are "neighborhood" predications; each
situates the trajector in the vicinity of the landmark identified by its
prepositional object.
, s<
locative 1
0997777791
:ope 1
1m
locative 2
V/////4W//'*''
scope 2
locative 3
wht^im '
scope ^§W
Figure 9
The "nested" character of these locative specifications derives from a
second set of correspondences, marked by the two lower dotted lines. They
specify that the search domain of one locative constitutes the scope of
predication for the next. Hence the interior of the yard, the search domain of
the first locative, functions as the scope of predication for the second, i.e.
only the interior of the yard is considered for purposes of identifying the
back fence in question and assessing the proximity relationship. The
neighborhood of the back fence is the search domain of the second locative,
and consequently the scope of predication for the third: the gate to which
the trajector is proximate must be part of the back fence in particular, not
some other fence outside this scope (observe that the gate may well be con-
textually unique — as signaled by the definite article — only with respect to
scope 3). The composite structure that results from integrating these
locatives in the specified manner finds the trajector located simultaneously in
all three nested search domains, each calculated with respect to a different
landmark. We see from this example that highly abstract and cognitively

A VIEW OF LINGUISTIC SEMANTICS
73
non-salient constructs, notably scope of predication and search domain, are
capable of participating in correspondences that are pivotal to a
grammatical construction.
When scopes of predication are nested in this fashion, and an entity of
concern is simultaneously located in all of them, I will refer to the
innermost scope (i.e. the most restricted one) as the "immediate" scope of
predication. With respect to (ll)(a) and Fig. 9, the locative relationship near the
gate has scope 3 as its immediate scope of predication; scope 3 figures
directly in the computation of this proximity relation. By contrast, scopes 2
and 1 figure only indirectly in the characterization of near the gate: they
provide the initial basis for identifying scope 3, hence scope 3 mediates their
role in the predication. This nesting of scopes (with special status accorded
the immediate scope) not only arises through syntagmatic combination, but
is observable as well in the relationships among distinct expressions. This
phenomenon is nicely illustrated by English body-part terms.
It is important not to confuse the hierarchical part/whole relationships
of a complex entity with the conceptual hierarchy that corresponds to it —
the former pertains to the structure of the entity itself, and the latter to the
structure of our conceptualizations. Thus the fact that arms, legs, head, and
torso are parts of the human body does not entail that the concepts arm,
leg, head, and torso are conceptual "parts" out of which the higher-order
concept body is assembled. Indeed, it is more plausible to assume precisely
the opposite: a crucial specification of the concept arm is the position of the
designated entity within the overall configuration of the human body, hence
the notion arm presupposes and incorporates the notion body, rather than
the converse. In the relevant hierarchy of conceptual complexity, body is
arguably more fundamental than arm, leg, etc., and functions as a
conceptual component of these higher-order notions. I therefore assume that a
rough gestalt representing the overall shape of the human body constitutes
one domain in the matrix for such predicates as arm, leg, head, and torso,
which profile particular substructures within this base. To be sure, the
matrices for these terms contain many other domains, pertaining to substance,
function, and so on, but these can be ignored for present purposes.
The notion body is thus analyzed as the base and scope of predication
for the higher-order concepts arm, leg, etc. Continuing along the
conceptual hierarchy, each of these concepts functions in turn as the scope of
predication for the characterization of additional terms (defined on a smaller
scale) designating their subparts; the conception of an arm, for instance, is

74
RONALD W. LANGACKER
the base for describing such higher-order notions as elbow, forearm, and
hand. The conception of a hand then provides the scope of predication for
palm, thumb, and finger, with finger giving rise to such small-scale notions
as knuckle, fingertip, and fingernail. The conceptual hierarchy is therefore
defined by nested scopes of predication. For knuckle, the sequence is body
> arm > hand > finger, with finger the innermost layer, i.e. the immediate
scope. It is intuitively obvious that knuckle is characterized directly with
reference to the conception of a finger, and progressively more indirectly
with respect to notions farther away along the hierarchy. It is not that these
more distant notions are totally irrelevant — after all, finger presupposes
hand, which presupposes arm, and so forth; all these notions are within the
scope of predication, broadly construed. Clearly, though, finger is the most
prominent within knuckle, and provides the immediate basis for its
characterization.
These constructs have linguistic consequences, one of which is
illustrated in (12):
(12) (a) A finger has 3 knuckles and 1 nail.
(b) ?? An arm has 14 knuckles and 5 nails.
(c) ??? A body has 56 knuckles and 20 nails.
In sentences with have that describe part/whole relationships, there are
differences in felicity that reflect the putative conceptual hierarchy (cf. Bever
and Rosenbaum 1970; Cruse 1979). Specifically, such a sentence is most
felicitous when the subject provides the immediate scope of predication for
the object, and declines in felicity (other things being equal) as the distance
between them along the hierarchy increases. Compounds designating body
parts are subject to a similar restriction:
(13) (a) fingertip; fingernail; toenail; eyelash; eyelid
(b) *bodytip; *armnail; *footnail; *headlash; *facelid
The terms in (13)(a) are the conventional English expressions for the body
parts in question. Note that the first member of the compound is in each
instance the immediate scope of predication for the second (in the case of
eyelash and eyelid, eye must be interpreted as designating the entire eye
region, not the eyeball in particular). The compounds in (13)(b) are to be
construed as designating the same body parts as their respective
counterparts in (13)(a). In (13)(b), however, the first member of the compound is
not the immediate scope of predication for the second, but is farther
removed along the appropriate conceptual hierarchy. Not only are these

A VIEW OF LINGUISTIC SEMANTICS
75
compounds non-conventional, but they seem strikingly unnatural — if new
terms were needed for these body parts, the ones in (13)(b) would hardly
suggest themselves as suitable candidates. While it is conceivable that
languages might occasionally conventionalize such expressions, I would expect
them to be relatively infrequent.
Relative Salience of Substructures
Unfortunately, the claim that a particular element is "salient" or
"prominent" relative to others is much easier to make than to verify. We
have no way at present to measure objectively the relative salience of the
various facets of a conceptualization. However, detection by direct and
objective means is hardly a realistic criterion for the adoption of linguistic
constructs, and theorists seldom hold themselves to this standard (note the
abstract, indirect basis for positing such entities as traces and zero pro
forms in government-binding theory). In contrast to many theoretical
constructs, the idea that varying levels of prominence distinguish the subparts
of linguistic structures is both cognitively plausible and intuitively natural.
Still, the notion of salience is overly vague if left unexplicated, being
applicable to numerous linguistic phenomena of diverse character. To render the
notion somewhat more precise and substantive, we can attempt to
distinguish particular types of linguistic salience, and to examine as best we can
their basis and structural significance.
Just three types of prominence will be considered, the first of which —
profiling — has already been discussed. As the "focal point" within a
predication, the profile (i.e. the substructure that is designated) is prominent
relative to the remainder of the base. A predication's profile determines its
basic grammatical category: a nominal expression profiles a "thing"
(defined abstractly as a "region in some domain"), while such classes as verbs,
adjectives, and adverbs profile different sorts of relations. In a complex
expression, that component which contributes its profile to the expression
as a whole can be identified as the "head". Snake is thus the head in the
nominal expression long snake, since the composite structure designates the
reptile, not the relationship of length (cf. Fig. 7).
A second type of salience pertains to relational predications. With few
if any exceptions, relational predications manifest a subtle asymmetry in
how they portray their participants: one participant is always singled out for
special prominence — this is the one that I have labeled the "trajector" in

76
RONALD W. LANGACKER
previous diagrams (Figs. 6 and 7). Though I will not pursue the issue here
in any detail, there is some basis for supposing that the special status of the
trajector is a matter of figure/ground organization (cf. Langacker in press:
ch. 6); if so, a trajector can be characterized as uthe figure within a
relational profile". I use the term "landmark" for salient participants other than
the trajector. If there are multiple landmarks, the most prominent among
them is called the "primary landmark". The terms trajector and landmark
are inspired by prototypical predications of motion and physical activity
(e.g. go, run, hit), where the trajector is the mover and the landmark serves
as a point of reference for specifying its path. Observe, however, that the
definition of these constructs refers only to relative prominence.
The trajector/landmark asymmetry underlies the subject/object
distinction, but they are not equivalent. The terms "subject" and "object" are
generally reserved for overt nominals (i.e. noun phrases) that have a
particular role in clause-level organization. The trajector/landmark asymmetry
differs in several respects. First, it pertains to a relational predication's
internal structure; substructures are identified as trajector and landmark
regardless of whether these participants are spelled out overtly by
accompanying expressions. In Fig. 7, for example, the adjectival predicate long
has both a trajector and a landmark — respectively, a thing and a region
along the length scale — even though only the former is capable of being
elaborated by an accompanying noun. Second, the trajector/landmark
asymmetry is observable at any level of organization, not just the clause
level. With minor exceptions (e.g. prepositional objects), only verbs are
traditionally regarded as having subjects and objects; yet verbs, adjectives,
adverbs, and prepositions are all attributed trajectors and landmarks,
regardless of whether they function as clausal heads. Third, only nominal
expressions are normally considered to be subjects or objects, but the
trajector or the landmark of a relational predication can itself be a relation.
In fact, having a relational trajector is what distinguishes adverbs from
adjectives.
We may summarize as follows. Relational predications profile
interconnections among conceived entities, some of these entities being
recognized as central "participants". One aspect of conventional imagery is the
level of prominence accorded the various participants in the construal of a
relationship: one participant, termed the trajector, achieves the greatest
prominence by virtue of being selected as the figure within the relational
profile; to some extent, the remaining participants (landmarks) can them-

A VIEW OF LINGUISTIC SEMANTICS
77
selves be ranked for salience. These participants, with their differing
degrees of prominence, characterize the internal structure of a relational
predication, regardless of type or level of grammatical organization. Often
a relational predication occurs with another expression in a grammatical
construction serving to render one of its participants overt; specifically, the
relational participant is placed in correspondence with the profile of the
accompanying expression, and is typically elaborated by this expression
(e.g. the schematic trajector of long in Fig. 7 corresponds to the profile of
snake, which elaborates this trajector). If the relational predication
happens to be a clausal head, and the elaborating expression a nominal, the
latter is traditionally recognized as a subject or direct object. A subject is thus
characterized in this framework as a nominal expression that corresponds to
the trajector of a clausal head, and a direct object as one that corresponds
to its (primary) landmark.
What could be the conceptual basis for the trajector/landmark
asymmetry? If we seek a schematic definition of trajectors, i.e. one that is
applicable to all relational predications at all levels of organization, we obviously
cannot focus on such notions as "agent", "controller", and "topic", which
are commonly proposed for the characterization of prototypical clausal
subjects. A universally valid definition must be sufficiently abstract and flexible
to accommodate sentence pairs like (14)-(16):
(14) (a) The painting is above the light switch,
(b) The light switch is below the painting.
(15) (a) A trumpet resembles a cornet,
(b) A cornet resembles a trumpet.
(16) (a) Iraq invaded Iran.
(b) Iran was invaded by Iraq.
All three pairs display a semantic contrast that I attribute to the choice of
trajector within the clausal head (this trajector being elaborated by the
subject nominal), yet there is no action in (14) or (15) (hence no agent or
controller), while in (16) the trajector is the agent in one instance and the
patient in the other. In fact, we observe the trajector/landmark asymmetry
even with verbs, like resemble, that designate a symmetrical relationship. It
appears, then, that the trajector/landmark asymmetry is essentially
independent of a predication's content, i.e. the choice of trajector is not strictly
predictable on that basis. Nor does the trajector/landmark asymmetry
reduce to discourse notions such as "topic", for it is observable even in the

78
RONALD W. LANGACKER
semantic contrast between pairs of words such as above and below,
considered in isolation and out of context.
For several reasons, defining a trajector as the figure within a
relational profile is reasonable as a working hypothesis. First, the importance
of figure/ground organization as a cognitive phenomenon is well established
(see Talmy 1978b and Wallace 1982 for discussion of its linguistic
applicability). Also, the variability and independence from content observable in
the choice of relational trajector is mirrored by our ability to make
alternate selections of figure within a given scene. Moreover, figure/ground
organization affords a plausible analysis of problematic examples like (14)-
(16). It nicely explicates the semantic contrast observable in (14), if we
assimilate the notion "landmark" or "reference point" to the more inclusive
notion "ground". Specifically, in (14)(a) the light switch is construed as a
point of reference (landmark) for locating the painting, while the converse
is true in (14)(b). Though the sentences in (15) pertain to more abstract
relationships, the semantic contrast is precisely parallel: (15)(a) takes a
cornet as the standard of comparison (abstract reference point), and evaluates
a trumpet with respect to this standard; (15)(b) reverses these roles. For
situations involving motion or other activity, the mover or the actor is most
naturally selected as the figure; linguistically, the same entity represents the
unmarked choice for trajector and subject, as seen in (16)(a). Marked
constructions allowing some other participant as trajector and subject, such as
the passive in (16)(b), can then be explicated as instances of figure/ground
reversal.
Finally, there is a notable cross-linguistic tendency for the participant
most readily selected as figure on perceptual grounds to represent as well
the unmarked choice of trajector in a relational predication. This is true not
only for action sentences, but also for expressions like above and below that
describe static configurations. Intuitively, it is evident that above is in some
sense the unmarked (or "positive") member of the pair; in similar fashion,
over and in front of are unmarked relative to under and in back of,
respectively. In each instance, the unmarked member is the one whose trajector
would normally be the most salient perceptually: something is more easily
perceived and selected as figure within the scene when it is above, over, or
in front of something else than when it is below, under, or in back of it.

A VIEW OF LINGUISTIC SEMANTICS
79
(a) (b)
i
V
E
K
T
1 O"
i
i
i
o
[ HORIZ
ABOVE BELOW
Figure 10
Such pairs also demonstrate that linguistic semantics cannot be based
on content alone. Everyone agrees, for instance, that above and below are
semantically distinct, yet they appear to have the same content — even the
same profiling — as sketched in Fig. 10. The domain is oriented space, i.e.
space organized into the horizontal and vertical dimensions. Each
predication profiles the interconnections between two entities in this domain,
which are differentially located along the vertical axis. Where, then, does
the semantic contrast lie? It can only lie with the choice of trajector,
identified as the upper entity in the case of above, and the lower one for below.
Apart from the figure/ground organization imposed on the relational
profile, the two predications are basically identical.
The last type of salience to be considered is a concomitant of "analyza-
bility". A complex expression is "analyzable" to the extent that speakers
are cognizant of the contributions of component morphemes to the value of
the composite whole. Analyzability need not take the form of conscious
awareness (though it often does); it is sufficient that the component
structures are activated at some level of processing when the composite structure
is constructed or manipulated. Novel expressions are necessarily analyza-
ble, but once a complex expression has the status of a familiar unit, it is
conceivable that a speaker might activate its composite structure
independently from its components. Intuitively, fixed expressions show a gradation
of analyzability. Speakers definitely recognize the contribution of complain
to complainer, but the presence of compute in computer lies at the margins
of awareness, and propel may not be perceived at all in propeller.
1
i V
E
K
T
Otr
1
1
1
0lm i
[ HORIZ

80
RONALD W. LANGACKER
The analyzability of an expression affects its semantic value. I would
argue, for example, that pork contrasts semantically with pig meat, father
with male parent, gravel with pebbles, and oak with oak tree, even if we
ignore any differences in connotation or information content that might
distinguish the members of each pair. Specifically, I suggest that analyzability
augments the salience of the semantic elements profiled by the component
structures. If we assume for sake of discussion that pork and pig meat have
precisely the same composite structure, they nonetheless differ in meaning
because the latter expression provides individual symbolization to the
conceptual components pig and meat, thereby rendering these notions more
salient within the composite whole than they would otherwise be. As a
derived plural, pebbles highlights the fact that the designated mass consists
of multiple instances of the category pebble, whereas the monomorphemic
gravel conveys comparable information about its constituency but fails to
emphasize it in any way. The semantic effect of analyzability becomes more
pronounced as the paraphrases become more elaborate. Thus triangle is
quite different in meaning from three-sided polygon despite their identical
extensions: the latter expression calls explicit attention to a factor
remaining latent in the former, namely the membership of the designated entity in
a larger class of geometrical figures. And few would want to claim that sink
has precisely the same semantic value as passively descend through a
medium under the force of gravity (even assuming the accuracy of the
paraphrase).
If the semantic import of analyzability is acknowledged, how is it to be
accommodated? We need only recognize that the meaning of a complex
expression does not reside in the composite structure alone, despite its
privileged status; also included, as a subsidiary aspect of an expression's
meaning, is the compositional path through which the composite structure
is progressively assembled. Suppose, for example, that an expression has
the composite semantic structure [XYZ], where X, Y, and Z are distinct
(though possibly overlapping) facets of this conception. There are many
potential compositional paths for arriving at [XYZ], some of which are
diagrammed in Fig. 11; each path represents a slightly different meaning,
despite the identity of content at the composite-structure level. The path in
11(a) corresponds to a monomorphemic expression, in which the entire
structure [XYZ] is symbolized as an unanalyzed whole by the phonological
sequence [w]. The paths in (b) and (c) correspond to bimorphemic
expressions, where [X] is individually symbolized in one instance, and [Z] in the

A VIEW OF LINGUISTIC SEMANTICS
81
other. All three facets of the composite structure are separately symbolized
in (d), which represents a trimorphemic expression with internal
constituency. Because they have the same composite structure, the four
expressions schematized in Fig. 11 are paraphrases in the sense of conveying the
same content and designating the same entity. Yet they contrast
semantically, their overall semantic values incorporating the following sets of
structures: (a) [XYZ]; (b) [X], [YZ], [XYZ]; (c) [XY], [Z], [XYZ]; (d) [X], [Y],
[Z],[YZ],[XYZ].
(a)
XYZ1
00
(c)
XYZ
wz
XY] [Z]
~w"j |zj
Figure 11
(d)
"XYZj
xyz
|xj
[yzI
I
rYjfzi
LElEj
An expression's composite structure is thus construed against the
background provided by its path of assembly; its meaning embraces the
entire, hierarchically-organized set of component structures through which
the speaker arrives at the composite conception. Although this is quite an
elaborate definition of meaning, there are specific advantages to adopting
it. For one thing, it resolves a classic problem of truth-conditional semantics
that arises in regard to semantic anomaly. Because an anomalous sentence
has no truth conditions, all such sentences are semantically equivalent. The
sentences in (17) are thus treated as synonymous:
(17) (a) * Existence spanked a yellow aroma,
(b) * Colorless green ideas sleep furiously.
Barring special interpretations of the component lexemes, such sentences
are certainly incoherent, but speakers do not readily accept either their
meaninglessness or their semantic identity. Furthermore, sentences like
(18) are semantically well-formed (and certainly non-equivalent), despite
their incorporation of an anomalous constituent:

82
RONALD W. LANGACKER
(18) (a) It is nonsensical to speak of a yellow aroma.
(b) It is nonsensical to speak of colorless green ideas.
This is problematic for traditional compositional semantics, since the
anomaly of a constituent prevents the emergence of well-formed semantic
representations at higher levels of composition.
These problems essentially disappear with a conceptualist view of
meaning acknowledging the semantic contribution of compositional paths.
A phrase like yellow aroma is indeed anomalous; when the two words are
combined in accordance with the proper constructional schema, conflicting
properties are attributed to corresponding entities that the schema specifies
as being the same (cf. Fig. 7), hence a coherent composite
conceptualization fails to emerge. However, an inconsistent conceptualization is still a
conceptualization, and the expression does have a compositional path
(incorporating the meanings of the individual words) even if it terminates in a
composite structure that is empty or incoherent. Moreover, yellow aroma
and colorless green ideas contrast semantically because their compositional
paths incorporate semantically distinct components. The coherence of
sentences like (18) can also be explained when meaning is identified with
conceptualization. Some conceptions refer to other conceptions as
substructures, and nothing prevents a defective conception from being referred to
in this way. If an anomalous conception occurs in conjunction with a
predication like nonsensical, which specifically comments on such anomaly, the
overall conception achieves a certain coherence. Sentences that predicate
the anomaly of a specific notion such as yellow aroma need not themselves
be any more anomalous than words, like nonsensical, which refer
schematically to an anomalous conception as part of their internal structure.
Defining an expression's meaning as including its entire compositional
path at the semantic pole has the further advantage of accounting for the
otherwise inexplicable semantic contrast between certain pairs of
expressions. Consider the words unrecover and reuncover, which feel a bit
awkward but nonetheless conform to established derivational patterns.
Speakers agree that they are semantically distinct: roughly, unrecover means
'undo the act of recovering', while reuncover means 'repeat the act of
uncovering'. (Note that recover is to be interpreted as 'cover again', not 'get
back'.) Yet the two expressions, at the composite structure level, have
precisely the same base, and profile the same substructure within this base.
The semantic distinction cannot be explained unless their compositional
paths are included as part of their meanings.

A VIEW OF LINGUISTIC SEMANTICS
83
>
(I)
> 1
(II)
1 >
(III)
(iv) ;
?R0FII£D EVENT
BASE
Figure 12
The base for each word comprises a sequence of four events, labeled I-
IV in Fig. 12. The first event consists in the placement of a cover on some
object, and the second, in its removal; events III and IV repeat the cycle.
Events I-IV must all be included in the base because they provide the
minimal context required for the proper identification of the profile (event IV).
To be in a position to unrecover something, i.e. to undo the act of
recovering it, one must already have recovered it; this implies an original act of
covering (I), a second act of covering (III), and some intervening event (II)
permitting the covering to be repeated. Unrecover thus evokes the entire
sequence of events in Fig. 12, though it profiles only the last. What about
reuncover? To reuncover something, one must first uncover it (II) and then
repeat this act (IV), and under normal circumstances each act of
uncovering requires a previous act of covering to undo (I and III). Hence reuncover
also presupposes the entire sequence of events, and designates only the
last.
Consequently, the semantic contrast between unrecover and reuncover
cannot be attributed to either their base or their profile at the composite
structure level. But obviously, they arrive at this complex conception
through different compositional paths. Each path begins with the root
cover, whose content is limited to the first event, and terminates with the
composite structure, whose content includes the full sequence; let us
therefore represent these structures as [I] and [I-II-III-IV], respectively. The
difference between the two paths lies in the intermediate-level structure
resulting from the initial derivational process. In the case of unrecover, this
intermediate structure corresponds to the derived stem recover, with the
content [I-II-III]. For reuncover, on the other hand, the intermediate
structure corresponds to the stem uncover, whose content is simply [I-II].
Distinct sets of structures thus constitute the compositional paths of the two

84
RONALD W. LANGACKER
expressions. Leaving aside the meanings of the individual derivational
morphemes, the path consists of [I], [I-II-III], [I-II-III-IV] for unrecover, and
[I], [III], [I-II-III-IV] for reuncover.
Perspective
The final dimension of imagery, perspective, subsumes a number of
more specific factors: orientation, vantage point, directionality, and how
subjectively or objectively an entity is construed. All these factors imply a
viewer, or more abstractly, a conceptualizer whose construal of the
conceived situation can be thought of as roughly analogous to a perceptual
relationship (cf. Langacker 1985). As a default-case option, the viewer (or
conceptualizer) is identified with the speaker (and secondarily, the
addressee).
The notion of orientation is largely self-evident, so I will say little
about it. Certain predications, most obviously those involving left and right,
crucially presuppose a particular orientation on the scene being described.
Thus, if V is the viewer in Fig. 13, and the arrow indicates the direction in
which he is looking, he would describe X as being to the left in the context
of 13(a), but to the right in 13(b). Many predications rely on a canonical
orientation that is independent of the viewer's immediate circumstances.
For instance, the domain of oriented space — space organized into the
horizontal and vertical dimensions — reflects the typical orientation of human
observers in relation to the surface of the earth. This domain is invoked for
predications like above and below (cf. Fig. 10), whereby one entity can be
characterized consistently as being above or below another whether the
viewer observes the scene while upright, standing on his head, or lying on his
side.
(c)
i
i
1
0 !
Figure 13
The notion of vantage point is similarly straightforward. Clearly, the
use of left and right in Fig. 13 is dependent on the viewer's vantage point as

A VIEW OF LINGUISTIC SEMANTICS
85
well as his orientation. Noteworthy here is the phenomenon of "mental
transfer" (Vandeloise 1984), i.e. a speaker's ability to mentally assume a
vantage point other than his actual one for purposes of calculating a
particular relationship. For example, from his actual vantage point the viewer
in Fig. 13(c) would describe X as being to the left of Y, but he could also,
by assuming Y's vantage point and orientation, say that X is to the right of
Y. The same phenomenon is well known from sentence pairs like (19),
either of which might be uttered by a speaker in San Diego talking on the
phone to somebody in Chicago:
(19) (a) I will go to Chicago tomorrow,
(b) I will come to Chicago tomorrow.
In (19)(a), the speaker describes the envisaged motion from his own
vantage point, whereas in (19)(b) he adopts for this purpose the vantage point
of the addressee.
In some languages, the notion of vantage point is of greater systematic
importance than it appears to be in English. A case in point is Cora, a Uto-
Aztecan language of Mexico (cf. Casad 1982a; Casad and Langacker 1985).
Cora has an elaborate system of locative particles and verb prefixes, some
combination of which occurs in virtually every sentence, and vantage point
is often essential to their characterization. A particularly striking example is
provided by the sentences in (20), both of which refer to a dog:
(20) (a) u-h-k*-tya-pu'u
inside-vertical-short-middle-planted
'Its tail is cut short.'
(b) a-h-ki-tya-pu'u
outside-vertical-short-middle-planted
'Its tail is cut short.'
Despite the identical English translations, these sentences are semantically
distinct, and reflect the vantage point of the speaker: the first would be
spoken by someone viewing the dog directly from behind, and the second when
the dog is viewed from the side. The way in which the speaker's position is
indicated proves both interesting and instructive.
The locative elements in (20) are the contrasting prefix combinations
u-h- and a-h-. They belong to a larger set of combinations describing
topographic relationships, which are often extended to characterize relationships
in non-topographic domains. In its basic topographic sense, -h- locates

86
RONALD W. LANGACKER
some entity in the face of a slope (as opposed to the foot or the head of a
slope); by extension, it locates an entity on any vertical surface. The
prefixes u- and a- are ubiquitous, and form a contrastive pair with a
considerable array of specific senses. The glosses 'inside' and 'outside' reflect their
prototypical values. In topographic expressions, these glosses are still
accurate, but must be construed relative to a very special domain. This domain
involves the conception of a viewer (generally the speaker) standing at the
foot of a slope and looking straight up the face of the slope toward the top.
With respect to this abstract domain, we can define an entity describable as
the "line-of-sight region": this region includes a relatively narrow corridor
that runs along the viewer's line of sight up to the crest of the slope (and
also subsumes the viewer's immediate surroundings on all sides). It is this
subjectively-determined line-of-sight region that is pivotal to the
topographic value of u- vs. a-, which are relational predications taking this region
as their landmark. Specifically, u- indicates that the trajector falls 'inside'
the line-of-sight region, and a-, that it is located 'outside' this region. The
basic topographic values of u-h- and a-h- are consequently 'inside the line of
sight in the face of a slope' and 'outside the line of sight (i.e. off to the side)
in the face of a slope', respectively.
Figure 14

A VIEW OF LINGUISTIC SEMANTICS
87
By extension, these prefix combinations apply to non-topographic
domains, where they pertain to any sort of vertical surface. In (20), the
vertical surface in question is some facet of a dog's torso, which is represented
as a block in Fig. 14. Fig. 14(a) corresponds to (20)(a), in which the dog is
viewed from behind. The trajector of the prefix combination is a relation,
namely that of the tail being short; the use of u- indicates that this trajector
falls within the viewer's line of sight extending upwards along a vertical
surface. Because a Cora knows where a dog's tail is attached, he can therefore
determine that the viewer is examining the dog from the rear. In (20)(b), on
the other hand, the viewer is examining the dog from the side, as sketched
in Fig. 14(b). The occurrence of a- indicates that the trajector (i.e. the tail
configuration) lies outside the line-of-sight region on a vertical surface,
hence the viewer cannot be examining the dog end-on. He must instead be
looking at it from the side.
Let us turn now to directionality, the penultimate factor listed under
the heading of perspective. The main observation I wish to make is that the
relevance of this notion is not limited, as one might expect, to sentences
describing physical or even abstract motion. Directionality is common even
in sentences describing static configurations of indefinite duration, as we
see from (21):
(21) (a) The roof slopes upward at a steep angle,
(a') The roof slopes downward at a steep angle.
(b) The road widens just outside the town,
(b') The road narrows just outside the town.
(c) The hill gently rises from the bank of the river,
(c') The hill gently falls to the bank of the river.
(d) The line from point A to point B is straight,
(d') The line from point B to point A is straight.
The sentences in each pair are semantically distinct despite their
applicability to precisely the same objective situations. Moreover, it is intuitively
clear that the meaning contrast hinges on directionality. But if nothing is
conceived as moving, how does this sense of directionality arise?
Evidently, it reflects the manner in which the conceptualizer (notably
the speaker) construes the conceived situation: at some level of processing,
the conceptualizer mentally "scans through" the static scene in a particular
direction. Hence the feeling of directionality does not reflect the contents

88
RONALD W. LANGACKER
of the conceptualization, but rather how it is mentally accessed; ultimately,
it stems from the sequencing of cognitive events. As it stands, of course,
this proposal is quite vague. A substantive analysis will have to provide a
broadly-grounded body of descriptive constructs allowing the origin of the
intuitively-perceived directionality to be made explicit in a motivated way
(some preliminary proposals are made in chs. 3 and 4 of Langacker in
press). Still, directionality must be recognized as an important facet of
imagery, however dimly we understand the underlying mechanisms.
The last factor subsumed under perspective pertains to how
subjectively or objectively an entity is construed. In this context, the terms
"subjective", "objective", and their derivatives must be understood in a
particular sense (I am not presently referring to the general claim that meaning is
a subjective phenomenon). The construal of an entity is said to be purely
"subjective" when it functions solely as the conceptualizer, and not at all as
an object of conceptualization. Conversely, the construal of an entity is
purely "objective" when it functions solely as the object of
conceptualization and is fully distinct from the conceptualizer. The only pure case of a
subjectively-construed entity is consequently the conceptualizer himself
when he is totally absorbed in the conception of some other entity, to the
extent of losing all awareness of himself and even of his own role in the
conceptualization process. Whether the construal of any entity is ever purely
subjective or purely objective is immaterial — subjectivity is a matter of
degree, and the pure cases are primarily invoked as reference points on
what is probably a continuous scale.
These notions are subtle and perhaps elusive, but their ultimate
importance for semantic and grammatical structure is hard to overstate (see
Langacker 1985 for extensive discussion). By way of brief illustration, let us
consider their application to the problem of directionality. Sentence (22)(a)
is a typical description of physical motion: the entity designated by the
subject moves along the spatial path specified by the locative phrases.
(22) (a) A boy walked across the field, through the woods, and
over the hill,
(b) There was a fire last night across the river, through the
canyon, and over the mountain.
Though a comparable series of locatives appears in (22)(b), there need be
no motion involved; the speaker may simply be indicating where the fire
occurred, with no thought of anybody going there. The noteworthy feature

A VIEW OF LINGUISTIC SEMANTICS
89
of (22)(b) is that the locative sequence is identical in form to one that
describes an extended spatial path, yet its function is evidently to identify a
single, point-like location.
I analyze the locative series in both sentences as describing an
extended spatial path — what differs is not the path, but the nature of the
movement that follows it. The movement in (22)(a) is concrete and
objectively construed: the mover is an entity (the boy) fully distinct from the con-
ceptualizer (the speaker), as is the spatial path itself; the motion is concrete
in that the mover physically occupies the successive points that constitute
the path. By contrast, the movement in (22)(b) is abstract and subjectively
construed. The path itself is only partially objective, for it extends from the
speaker to the location of the fire, and is thus not fully distinct from the
speaker. It is the speaker who moves along this path, but this motion is
conceptual rather than physical; it consists in the speaker successively
activating his conception of the ordered points or segments that the path
comprises. Starting from his own position, the speaker mentally constructs a
progressively more elaborate path conception whose endpoint, upon
completion, is identified with the location of the fire. Why is this motion
subjective? Because the speaker is not concerned with his own movement along
the path — indeed, he does not think of himself as moving at all. Only from
the external perspective of the analyst is the speaker construed objectively
as following a path. From his own perspective, the speaker is simply
conceiving and expressing the location of the previous night's fire. The motion,
in other words, is not itself the object of conception, but rather an inherent
facet of the conceptualization process. It lies much closer to the subjective
than to the objective end of the scale, since it represents activity by the con-
ceptualizer in his role as such.
A similar analysis is proposed for the sentences in (21). Each of them
describes a complex spatial configuration, and we can reasonably suppose
that a speaker builds up to this complex conception by successively and
cumulatively activating various facets of it. The order in which he does so
(at some stage or level within the overall processing activity) determines the
perceived directionality of the corresponding expression. In the case of
(21)(d), for instance, the speaker starts with point A, and progressively
builds up the conception of a line extending away from it, until the
conception so arrived at includes point B as its terminus; in (21)(d'), he proceeds
in the opposite direction. Once more, the speaker can be characterized as
moving mentally along the directed path, but this motion is both abstract

90
RONALD W. LANGACKER
and subjectively construed: what the sentences actually describe and
designate is the locative configuration, not the speaker's construal of it.
Conclusion
The ideational view of meaning is looked on with little favor by
mainstream semantic theory. I contend, however, that the heavy emphasis
in current theory on truth conditions and rigorous formalization stems from
a priori methodological considerations, not from any evidence that a self-
contained, objectivist semantics is optimal or even appropriate for natural
language. I have outlined a subjectivist approach to semantics that, while
admittedly programmatic, nevertheless goes considerably beyond the
simple, essentially vacuous assertion that the meaning of a word is a concept
(of undefined character). For an ideational account to be substantive, one
must actually describe the internal structure of the concepts at issue in an
explicit, descriptively adequate manner. I have attempted to catalog some
of the factors that such an account must accommodate. In particular,
linguistic semantics cannot confine itself to the information content of
expressions; our capacity for imagery — the ability to construe a conceived
situation in alternate ways — is a fundamental aspect of meaning that is indissoc-
iable from conceptual content. The view of meaning that emerges from
these considerations is complex and elaborate. The phenomena we have
examined will severely test the adequacy of any comprehensive model of
cognitive processing.

The Nature of Grammatical Valence
Ronald W. Langacker
University of California, San Diego
Preliminary Notions
Cognitive grammar seeks to characterize a speaker's grasp of
established linguistic convention. This knowledge is assumed to take the form of
a "structured inventory of conventional linguistic units", where a "unit" is a
structural complex that the speaker has fully mastered. Only three basic
types of units are posited: "semantic", "phonological", and "symbolic".
Symbolic units are "bipolar", with a semantic unit at one pole in symbolic
correspondence to a phonological unit at the other. The inventory of
conventional units is "structured" in that some units function as components of
others. The vowel [e], for instance, is one component of the phonological
unit [[b]-[£]-[djl, which is in turn a component of the symbolic unit [[BED]/
[[b]-[e]-[d]]],andsoon.
Three facets of the model must be explicated in slightly greater detail:
schematicity, grammatical structure, and semantic structure. "Schematic-
ity" is the relation between superordinate and subordinate nodes in a
taxonomic hierarchy. For example, the concept [TREE] is schematic
relative to more highly elaborated concepts such as [OAK] and [PINE];
superordinate to phonological units like [a] and [i] we can posit the
schematic phonological unit [VOWEL]. A schema is said to be "elaborated" by the
structures subordinate to it, called its "instantiations", and this elaborative
relation is indicated by a solid arrow, e.g. [TREE] —> [OAK]. Relatively
speaking, a schema specifies a notion only in gross terms, while its
instantiations specify it in finer detail — it is like the difference between a graph of
the stock market plotted on a coarse grid, showing only general trends from
Ronald W. Langacker

92
RONALD W. LANGACKER
month to month, and a graph plotted on a fine grid, showing day-to-day
fluctuations with greater precision. The conventional units of a grammar
include both schemas and their instantiations. Schemas have a categorizing
function and also a sanctioning function in the creation of novel
expressions. Relations of schematicity also play a role in grammatical valence,
which will be our basic concern here.
Cognitive grammar does not posit special morphological or syntactic
units. Grammatical structure, both morphology and syntax, is claimed to be
symbolic in nature, forming a continuum with lexicon. It is therefore
accommodated by symbolic units, each consisting of a semantic and a
phonological pole. Grammatical patterns, or "rules", are represented in the
form of schematic symbolic units, parallel in structure to the specific
expressions that instantiate these patterns. These schematic units capture
generalizations by representing the commonality of instantiating
expressions, and also function as templates that a speaker can employ in
constructing and evaluating novel expressions.
For a concrete example, consider the plural noun beds. What must we
say to fully describe this complex structure? First we must characterize the
two component morphemes, each a symbolic unit with a semantic and a
phonological pole. The second task is to specify precisely how the two
morphemes combine — this is the problem of grammatical valence. For now I
simply observe that the combination is bipolar and symbolic. It is bipolar in
the sense that we must state not only how the phonological units [bed] and
[-z] integrate to form the composite phonological structure [bed-z], but also
how the semantic units [BED] and [PL] integrate to form the composite
semantic unit [BED-PL]. Moreover, the phonological integration
symbolizes the semantic integration: the fact that the plurality designated by
[-z] pertains to [BED] in particular (rather than some other nominal
concept) is conveyed specifically by the fact that [-z] suffixes to [bed] rather
than some other stem. In short, the semantic relation between the meaning
components [BED] and [PL] is symbolized by a specific relation (one of
suffixation) between their respective phonological representations.

THE NATURE OF GRAMMATICAL VALENCE
93
SEMANTIC POLE ! PHONOLOGICAL POLE
Figure 1
Fig. 1 summarizes these relations of symbolization and integration. It
shows an integrative relation between the semantic units [BED] and [PL],
to yield the composite semantic structure [BED-PL], and also between the
phonological units [bed] and [-z] to yield the composite phonological
structure [bed-z]. It shows symbolic relations between [BED] and [bed],
between [PL] and [-z], and between the modes of integration of [BED] and
[PL] at the semantic pole and of [bed] and [-z] at the phonological pole. The
composite structures [BED-PL] and [bed-z] stand in a symbolic relationship
derivative of the ones linking their components and the integration of these
components.
This may strike one as an overly complicated way to describe a simple
case of suffixation, but in actuality it is not. I have only made explicit
reference to all the facets of structure and organization that go into a stem-affix
combination, since these facets must ultimately be described in full detail
rather than being taken for granted, as they normally are. An account of
each of the structures and relations in Fig. 1 is the minimum required for an
adequate description of a simple instance of grammatical valence between
two morphemes. Fig. 1 of course represents a specific instance of the
general pattern of plural formation with [-z]. The pattern itself — the cognitive
grammar equivalent of a morphological or derivational rule — is simply a
schematic version of Fig. 1, where the specifications of bed are replaced by
those of the count-noun schema. The schema for the morphological
construction coexists, in a speaker's cognitive representation of the linguistic
system, with those of its instantiations that have been mastered as estab-

94
RONALD W. LANGACKER
lished units.
Semantic structure is viewed in cognitive grammar as
"conventionalized conceptual structure", i.e. the form our conceptualizations must
assume for ready linguistic expression given the symbolic conventions and
resources of the language. A semantic structure is a conceptualization that
functions as the semantic pole of a linguistic expression, either fixed or
novel. I refer to any such structure as a "predication", and to the semantic
pole of a single morpheme as a "predicate".
A predication is characterized relative to one or more cognitive
"domains", collectively referred to as its "matrix". Some are "basic domains",
i.e. primitive fields of representation not reducible to other, more
fundamental conceptions; examples of basic domains include space, time, color,
taste, emotive domains, etc. Most domains are reducible and hence non-
basic: they are related to basic domains through conceptualization or some
conceptual hierarchy. Random examples include our conception of the
human face; knowing how to skin a rabbit; the calendrical cycle; the notion
that objects or events can be ranked on a scale with respect to some
property; and so on indefinitely. Any integrated conception or knowledge
system can function as a domain for the characterization of a linguistic
predication.
Non-basic domains are structured hierarchically, in the sense that
simpler conceptions can be coordinated or transformed to create more
complex, higher-order conceptions, which can in turn combine to form still
more complex conceptions, and so on. A given predication must be
characterized at an appropriate level in such hierarchies of complexity. It would
be pointless, for example, to try to define the geometrical notion [ARC]
directly in the basic domain of two-dimensional space, since an arc exists
only by virtue of its status as a segment of a circle. [ARC] is thus superordi-
nate to [CIRCLE] in the hierarchy of conceptual complexity, and we must
approach its characterization in two steps. First, [CIRCLE] can be defined
as a configuration in the basic spatial domain. Then, since [CIRCLE] is a
coherent concept, it can serve as the cognitive domain for the
characterization of derivative notions like [ARC].
It is claimed that a predication always derives its value through the
imposition of a "profile" on a "base". The base for a predication is its
domain (or matrix), or some relevant subportion thereof. The profile is that
portion of the base which the expression "designates". It is thus the focal

THE NATURE OF GRAMMATICAL VALENCE
95
point within the predication, and the special prominence it receives makes
it a kind of figure with respect to the ground provided by the remainder of
the base. Neither the base nor the profile is sufficient in itself to define a
predication, which consists precisely in the relation between the two.
Without the base, the profile cannot be identified; without the profile, the base
makes no designation.
These notions are illustrated for [CIRCLE] and [ARC] in Fig. 2.
Heavy lines are used for profiled entities. The base for [CIRCLE], in 2(a),
is the basic domain of two-dimensional space. The profile for this predicate,
what the predicate designates, is a configuration in this domain. As a
conceptualization, [CIRCLE] can then serve as base for the characterization of
expressions like [ARC], sketched in 2(b). Observe that the conception of a
circle is part of the predicate [ARC]; this configuration constitutes the base
of [ARC] but is not profiled, though precisely the same configuration
functions as the profile in 2(a).
CIRCLE
ARC
Figure 2
Predicates fall into several basic types determined with respect to the
nature of their profile. A "nominal" predication designates a "thing",
defined as a region in some domain; this is a bounded region in the case of
count nouns. Both circle and arc are count nouns, since they designate
lines, a special type of bounded region in space. Yellow (as a noun, the
name of a color) designates a bounded region in the color domain, defined
primarily by its location along the hue dimension. More abstractly, a
paragraph s a bounded region in a written passage, Tuesday profiles a bounded
region in the cycle of days constituting a week, and octave designates a
bounded region on a musical scale. Physical objects, as bounded entities in
three-dimensional space, are prototypical members of the "thing" category,

96
RONALD W. LANGACKER
but nevertheless represent a special case in the spectrum of possibilities
defined by its schematic characterization.
A "relational" predication profiles the "interconnections" among two
or more conceived entities; we can think of interconnections as cognitive
operations assessing the relative position of these entities within a domain.
Typically a relational predication profiles the interconnections between two
things, as illustrated by [IN] in Fig. 3(a). (Observe the notational device of
using circles to indicate things.) However, one or more of the relational
participants may themselves be relational in character. [FAST], for
example, locates a process in a particular region along a scale of rapidity. The
conception of such a scale constitutes a non-basic domain, and [FAST]
situates a process in that region of the scale which lies beyond the
neighborhood of the norm (n) in a positive direction. In every relational predication,
one of the profiled participants is singled out as the "trajector" (tr),
characterized as the figure within the relational profile; other salient participants
■(c)
LEAVE
Figure 3
FAST
i
w
T '
(®)
tr
TIME
tr
P
^^in

THE NATURE OF GRAMMATICAL VALENCE 97
are referred to as "landmarks" (Im). As the terms suggest, the landmark
entity can be thought of as a point of reference for specifying the location of
the trajector.
The relations in 3(a) and (b) are "stative", i.e. each is conceptualized
as a static configuration in its domain, scanned as a simultaneously
available whole. A "process" is more complex and involves the tracking of a
conceived situation through time. It can be regarded as a continuous series
of states occupying a continuous series of points in time and scanned
sequentially; a process is thus seen as unfolding through time, and the span
of time through which it is tracked is referred to as its "temporal profile".
Essential facets of the processual predicate [LEAVE] are sketched in Fig.
3(c). The matrix for this predicate coordinates the basic domains of time
and space; the heavy-line portion of the time arrow indicates the temporal
profile of the process.
The Problem
We will further explore the cognitive grammar framework by
examining its treatment of valence and grammatical constructions. Exemplification
will be drawn from various American Indian languages of the Uto-Aztecan
family (see Langacker 1977 for a survey). A common feature of these
languages, and of languages generally, is the formation of complex verbs
through the layering of "verb-like" derivational affixes. The Luiseno data in
(1) is not untypical (the tense suffix -q indicates present or recent past).
(1) (a) noo rjee-q
I leave-TNS
T am leaving.'
(b) noo rjee-vicu-q
I leave-want-TNS
T want to leave.'
(c) noo poy rjee-ni-q
I him leave-make-TNS
T made him leave/
(d) noo poy rjee-vicu-ni-q
I him leave-want-make-TNS
T made him want to leave.'

98
RONALD W. LANGACKER
(e) noo poy rjee-ni-vicu-q
I him leave-make-want-TNS
'I want to make him leave.'
(f) noo poy rjee-vicu-ni-vicu-q
I him leave-want-make-want-TNS
'I want to make him want to leave.'
Perhaps the most intriguing aspect of such constructions is the tendency for
morphological layering to correlate with "semantic scope", so that an inner
layer of structure is semantically "in the scope of" the affix constituting the
layer immediately external to it.
One approach to this phenomenon was afforded by the theory of
generative semantics, as illustrated in Fig. 4(a). Scope was explicated in
terms of subordination in underlying structure; the surface complex verb
was derived by repeated application of rules such as Predicate Raising and
Equi-NP Deletion (cf. Langacker 1973). In retrospect, several objections
can be raised to this type of analysis. One pertains to the adequacy of the
posited underlying structures as representations of meaning. Another is the
absence of theory-independent motivation for deriving the root and
derivational affixes from clauses embedded to NP. Finally, the whole notion of
deriving the surface form from a radically different underlying structure is
rather suspect.
(a)
UNDERLYING
STRUCTURE
s
LEAVE
SURFACE
STRUCTURE
(RULES)
M
HIM V TNS
/\
V MAKE
V WANT
I
LEAVE
Figure 4
TNS
MAKE
/ \
I WANT
\
LEAVE
HE

THE NATURE OF GRAMMATICAL VALENCE 99
Another alternative, which avoids at least one of the objections cited
above, is to posit dependency trees like that in Fig. 4(b). Such trees
accommodate predicate-argument relationships without the arbitrary appeal to
syntactic embedding under NP-nodes. However, they do not themselves
address the matter of morphological layering; special provisions are
required to provide for the grouping of predicates into words. I will argue,
moreover, that predicate-argument representations of this kind are
crucially inexplicit even with respect to pivotal semantic relationships.
In the remainder of this paper, I approach the correlation between
semantic scope and morphological layering from a broader perspective on
the nature of grammatical valence. I provide a preliminary but reasonably
explicit account of how morphemes combine, and of the relation their
forms and meanings bear to those of the composite whole. It is suggested
that predicate-argument structures like Fig. 4(b) represent only a special
(though prototypical) case in the spectrum of possible grammatical and
semantic relationships. The correlation between semantic scope and
morphological layering receives a natural explication in a theory that treats
grammatical structure as an inherently symbolic phenomenon.
Grammatical Valence: Canonical Instances
Grammar is the combination of symbolic structures to form
progressively more complex symbolic expressions. This combination is bipolar,
with the integration of semantic structures standing in a symbolic relation to
the integration of their corresponding phonological structures. Our main
concern at present is the nature of this integration at the semantic pole.
What precisely is the character of the semantic integration that forms a
composite semantic structure from two or more component predications?
I emphasize at the start that I am not concerned with trying to predict
the valence of a morpheme on the basis of its internal semantic structure. In
fact, I do not think it is possible to predict valence in absolute terms. There
is no reason to suppose, lor instance, that the semantic pole of Luiseno rjee
is substantially different in any crucial respect from that of English leave,
yet rjee is consistently intransitive, while leave optionally bears a valence
relation to a nominal complement (this very optionality establishes the
point):

100
RONALD W. LANGACKER
(2) (a) The man left.
(b) The man left the building.
Instead of absolute predictability, we must settle for predictability of a
weaker sort, one more generally appropriate for language: the semantic
structure of a morpheme defines its valence potential and determines how
readily it lends itself to certain kinds of valence relations exploiting that
potential, but whether and how this potential will actually be realized is in
some measure a function of linguistic convention.
Our concern here, though, is with the nature of the valence ties (at the
semantic pole) between two structures that do in fact combine in
accordance with the conventions of the language. We will start with prototypical
valence relations, turning subsequently to cases which deviate from the
prototype in various ways.
Consider the following Hopi sentence:
(3) taaqa moosa-t tiwa
man cat-ACC see/find
The man found the cat.'
(I simplify matters by ignoring the accusative inflection on 'cat' as well as
definiteness and tense/aspect.) In standard predicate-argument terms, the
Hopi verb tiwa is a canonical two-place predicate, for it designates a
relation between two salient entities, a searcher/perceiver and the object
sought/perceived. Fig. 5 is a typical predicate-argument dependency tree
representation for the lexical morphemes of (3), showing the valence
relations between [FIND] and its arguments [MAN] and [CAT].
FIND
/ \
MAN CAT
Figure 5
Though Fig. 5 is perfectly acceptable as a first approximation, it is
undeniably inexplicit on many crucial points. For one thing, nothing of
substance is indicated about the internal structure of any of the three
predicates. Second, nothing explicitly shows that [MAN] and [CAT] have
different roles with respect to [FIND], or what these roles are. Third, how
[MAN] and [CAT] connect to [FIND], and what permits this combination

THE NATURE OF GRAMMATICAL VALENCE 101
in the first place, are left unspecified. Finally, there is no direct
characterization of the composite semantic structure that results from integrating the
three components.
The cognitive grammar account of grammatical valence can be
regarded as an attempt to be explicit on all these points, i.e. to explicate the
valid intuitions (pertaining to prototypical valence relations) that lie behind
the widespread acceptance of some version of predicate-argument
structure. We can begin with a characterization of the internal structure of the
predicates, most crucially [FIND]. The matrix for this predicate is complex,
involving not only the domain of physical space but also non-basic domains
pertaining to perception and cognition, including the knowledge that
perceiving individuals, at a given time, have perceptual access to a limited area
and make perceptual contact only with objects located within this field. I
have conflated these different domains in Fig. 6, which oversimplifies
matters but is sufficient for present purposes.
FIND
Figure 6
[FIND] traces through time the evolution of a situation involving
several entities. Two of these entities, both physical objects, are included in
the relational profile as central participants: the trajector, corresponding to
the searcher, functions as the figure within this profile; the object sought
and found functions as the primary landmark. Less salient in this
conception is a third entity, namely the perceptual field of the trajector, which I
have given as an ellipse. The base of [FIND] includes a search process of
indefinite duration. Only the final stages of that process are actually desig-

102
RONALD W. LANGACKER
nated by the predicate and hence profiled, namely the transition into the
situation where the landmark is located in the trajector's visual field. The
dotted "correspondence" lines in Fig. 6 indicate that the trajector is
identical for all the component states, as is the landmark.
I treat [MAN] and [CAT] in much lesser detail, using only a mnemonic
sketch representing their shape specifications in diagrams below. For our
purposes, the relevant observation is that they designate prototypical
things, bounded objects in physical space. Their full description also
includes characterizations with respect to other domains (e.g. specifications
of size, color, canonical activities, and so on).
Let us now consider the valence relations in Fig. 5. A valence relation
between two predications is possible just in case these predications overlap,
in the sense that some substructure within one corresponds to a
substructure within the other and is construed as identical to it. The dotted
correspondence lines in Fig. 7 therefore specify that the trajector of [FIND] is
construed as being identical to the object profiled by [MAN], and that the
landmark of [FIND] is construed as being identical to the profile of [CAT].
Establishing these correspondences is what permits [FIND] to combine
with [MAN] and [CAT] in a grammatical valence relation. The lines of
correspondence can also be thought of as the integration of the component
predications, i.e. as instructions for fitting these predications together to
form a coherent composite structure. In Fig. 7, they specify that the
specifications of [MAN] are to be superimposed on those of the trajector of
[FIND], and the specifications of [CAT] on those of the landmark of
MAN
4JCAI
FIND (simplified)
Figure 7

THE NATURE OF GRAMMATICAL VALENCE 103
[FIND]. [MAN] and [CAT] thus bear different relations to [FIND], since
their profiles correspond to different substructures within it.
I claim that all valence relations are based on correspondences
between subparts of the component structures. This is in fact the only
constant factor in valence relations. However, there is considerably more that
needs to be said about canonical relations like those in Fig. 5 if we want to
be fully explicit about their nature.
For example, what is the nature of the asymmetry between [FIND] on
the one hand and [MAN] and [CAT] on the other that leads us to put
[FIND] on top in the dependency tree and the others on the bottom? What
leads us to say, in predicate-argument terms, that [FIND] is the predicate,
and [MAN] and [CAT] the arguments, rather than the converse? It would
appear that this asymmetry is connected with the relational character of
[FIND]. [FIND] introduces and organizes a scene in which salient
participants interact in a specified way. [MAN] and [CAT] designate individual
objects, and while it is part of our knowledge of these objects that they
participate in relations with other objects, these external relationships are
neither salient nor profiled within these predicates. [FIND] thus makes salient
reference to two objects as part of its own internal structure — these
objects function as the trajector and landmark within its relational profile
— while neither [MAN] nor [CAT] profiles an external relation.
I speak of [FIND] as being "conceptually dependent", while [MAN]
and [CAT] are "conceptually autonomous", with respect to the
correspondences in Fig. 7. [FIND] is conceptually dependent because it presupposes,
as an inherent part of its own internal structure, the two things participating
in the correspondences; [MAN] and [CAT] are conceptually autonomous
because they do not similarly presuppose a salient external relationship.
One cannot conceptualize the [FIND] relationship without conceptualizing
the two things functioning as trajector and landmark of that relation (even
if they are conceived only in the vaguest terms, say as blobs), but it is
perfectly possible to conceptualize a man or a cat without mentally setting it in
a relation with some external object. I would emphasize that conceptual
autonomy and dependence are ultimately matters of degree, but in
canonical instances of grammatical valence there is a fairly clear asymmetry
between the dependent and autonomous predications along these lines.
The dependent structure can be equated with the predicate, in predicate-
argument terms, and the autonomous structures with its arguments. (Recall
that in the terminology of cognitive grammar, either the dependent or the

104
RONALD W. LANGACKER
autonomous structure can be a predicate — a predicate is defined as the
semantic pole of a single morpheme.)
Though [FIND] introduces two relational participants that correspond
to the profiles of the autonomous predicates, there is obviously a difference
in the degree to which the corresponding entities are specified by the
relational and nominal predications. [FIND] characterizes its trajector and
landmark only in schematic terms — the former only as a thing capable of
searching and perceiving, the latter only as a thing capable of being found
— while [MAN] and [CAT] specify the corresponding objects in far greater
detail. Hence there is a relation of schematicity between each profiled
participant of [FIND] and the autonomous predicate whose profile
corresponds to it. We can say that the dependent predicate organizes the scene,
setting up a relation between schematically specified objects, and that the
autonomous predicates fit into this scene and elaborate particular
substructures within it. These substructures can be called "elaboration sites" (e-
sites); they are cross-hatched in Fig. 8 for ease of identification. The arrow
leading from an e-site in the dependent predicate to the corresponding
autonomous predicate thus stands for a relationship of schematicity, as
before.
MAN
4L|cai
FIND (simplified)
Figure 8
Two further aspects of canonical valence relations must be considered.
The first is "constituency", which is treated here not as a separate
dimension of linguistic organization, but merely as reflecting the order in which
simpler symbolic structures successively combine to form progressively
more complex ones. We may speak of a "compositional path" leading from

THE NATURE OF GRAMMATICAL VALENCE 105
individual morphemes, through intermediate composite structures at
various levels of complexity, to the final composite structure representing the
overall semantic and phonological value of the entire expression. Typically
constituency is binary: at any particular level of organization, two
component structures integrate to form a composite structure, which may in turn
function as one of two component structures at the next higher level of
organization, and so on. It is important to observe that constituency per se
is not employed in this framework for the characterization of basic
grammatical relations. For example, the subject and direct object of a clause are
not defined with respect to a particular constituent-structure configuration
— instead the subject is defined as a nominal whose profile corresponds to
the trajector of the clausal head, and the direct object as one whose profile
corresponds to its primary landmark (cf. Fig. 8), regardless of the order in
which these components are assembled. Different compositional paths
(orders of assembly) often lead to the same overall composite structure, with
no effect on our ability to characterize grammatical relations. Inherent to
this account of grammatical valence, consequently, is the potential for
constituency to be fluid and variable. The alternate constituency groupings
suggested by phonological considerations (e.g. intonation breaks, different
phonological phrasings) generally prove workable for the semantic pole as
well.
Fig. 9 shows the most likely constituency grouping for the example
under discussion. At the lower level of constituency, [FIND] and [CAT] are
integrated to form the composite structure (FIND-CAT). The schematicity
arrow indicates that [FIND] is the dependent structure, and [CAT] the
autonomous one; more specifically, [CAT] elaborates the landmark of
[FIND]. (Observe that the structure which I identify as conceptually
dependent, in the sense described previously, is often not the same structure that
is considered dependent in the sense of dependency-grammar
representations.) At the higher level of constituency, [MAN] is integrated with the
composite structure (FIND-CAT) to form the higher-order composite
structure (FIND-CAT-MAN). At this second level, (FIND-CAT) is
dependent, and the autonomous structure [MAN] elaborates its trajector. Note
that rectangles and square brackets are employed to indicate that the
enclosed structure presumably constitutes a unit (i.e. a fully mastered
configuration) for a typical speaker; closed curves and parentheses are
employed for novel structures. Thus, while all the individual predicates in
Fig. 9 are obviously established units, the intermediate-level composite

106
RONALD W. LANGACKER
structure (FIND-CAT) may well be novel, as is the overall composite
structure (FIND-CAT-MAN).
(FIND-GAT-MAN)
(FIND-CAT > -[MAN]
[find} *{cat]
Figure 9
The underscores in Fig. 9 relate to the final aspect of canonical valence
relations. Consider the lower level of constituency. [FIND], a process
predicate, integrates with [CAT], which profiles a thing. What, then, will be the
profile of the composite structure (FIND-CAT)? Will it designate a thing or
a process? Clearly moosat tiwa 'found the cat' designates a process, not a
thing; it is the core of a clause, and a clause by definition is processual in
nature. However there is no inherent reason why the composite structure
would have to inherit the profile of the conceptually dependent component
rather than the conceptually autonomous one — the choice must be
specified as part of each grammatical construction. In a canonical valence
relation, therefore, one of the component structures must be singled out as
the "profile determinant", which means that its profile prevails in
determining the character of the composite structure. The underscores in Fig. 9
mark the profile determinants. (FIND-CAT) is thus a process, for it
inherits the processual profile of [FIND]. Similarly, at the second level of
constituency (FIND-CAT) is marked as the profile determinant, so the
composite structure for the entire expression is also processual, and (3)
taaqa moosat tiwa The man found the cat' qualifies as a clause.
Fig. 9 of course abbreviates the semantic structures at each level. Fig.
10 is a somewhat more adequate representation, depicting the internal
organization of these structures within the limits of the notations introduced
(obviously it is still quite informal). By now Fig. 10 should be largely self-
explanatory. At the lower level of constituency, [CAT] elaborates the
landmark of [FIND], which serves as elaboration site and corresponds to the
profile of [CAT]. [FIND] is the profile determinant — indicated by using
heavy lines for the box surrounding it. The novel composite structure
(FIND-CAT) is formed by superimposing the specifications of [CAT] on

THE NATURE OF GRAMMATICAL VALENCE
107
FIND-CAT-MAN
lm
Tit.
4—4 *
TIME
w a 4
TIME
Figure 10

108
RONALD W. LANGACKER
the schematic landmark within [FIND], while retaining the processual
profile of the latter. (FIND-CAT) is thus a process with unspecified trajector.
This trajector is elaborated at the second level of constituency, where it is
put in correspondence with the profile of [MAN]. Since (FIND-CAT) is the
profile determinant, the novel composite structure (FIND-CAT-MAN) is a
process whose trajector receives the specifications of [MAN].
It should be apparent that the same composite structure would be
obtained if the component structures were amalgamated in another order,
i.e. if the constituency grouping were different. For instance, if [MAN] first
elaborated the trajector of [FIND], the intermediate-level composite
structure (FIND-MAN) would constitute a processual predication with a specific
trajector and a schematic landmark. This landmark could then be
elaborated by [CAT] at the second level of constituency, with the resulting
overall composite structure being no different from the one shown in Fig. 10.
Regardless of the order of composition, taaqa 'man' would be identified as
the clausal subject, since it elaborates the trajector of tiwa 'find' (whose
processual profile is inherited by the clause as a whole); and moosat would
be identified as the direct object, for it elaborates the primary landmark. It
is reasonable to suppose that speakers do in fact employ this alternate
constituency in sentences displaying object topicalization, as illustrated in (4).
(4) moosat-t taaqa tiwa
cat-ACC man see/find
The cat the man found.'
Non-Canonical Valence Relations
The essential aspects of a canonical valence relation are summarized in
Fig. 11. It is a binary relation between two predications, one of which is
conceptually autonomous and the other dependent. The dependent
structure is relational and includes within its profile an entity, specifically a
thing, which corresponds to the profile of the autonomous structure. This
entity, only schematically specified within the dependent structure itself,
functions as an elaboration site; this e-site bears a relation of schematicity
to the autonomous structure, which serves to specify it in finer detail.
Finally, the dependent structure is the profile determinant and hence
imposes its relational profile on the composite structure.

THE NATURE OF GRAMMATICAL VALENCE
109
DEPENDENT AUTONOMOUS
STRUCTURE STRUCTURE
o
ash
e-site
Figure 11
This is the basic type of valence relation generally assumed in
predicate-argument accounts of semantics, but among the cluster of properties
defining it, only the existence of a correspondence between substructures of
the components appears to be a universal property of grammatical valence
relations. In this section we explore some of the ways in which such
relations commonly deviate from the prototype.
When there is a clear asymmetry between two predications along the
lines of conceptual autonomy/dependence, it is natural for the dependent
structure to function as profile determinant, as it canonically does. This is
quite consonant with its function of organizing a scene, of establishing
relations among schematically specified entities; the autonomous structure
simply fits into this scene and elaborates one of these entities. Choosing the
autonomous structure as profile determinant in such a situation amounts to
a kind of skewing, where one component is intrinsically suited to play a
scene-structuring role, but where the perspective of the other component is
adopted by the composite structure.
This skewing is the basis for the cognitive-grammar characterization of
the head-modifier relation. We speak of such a relation when there is a
clear asymmetry between a conceptually autonomous and a conceptually
dependent predication, and where the autonomous structure functions as
profile determinant: the autonomous component is then the "head", while
the dependent component is the "modifier". In the canonical alignment of
Fig. 11, by contrast, where the dependent component functions as profile
determinant, we speak instead of a "head-complement" relation.
Consider the Hopi expressions in (5).
o

110
RONALD W. LANGACKER
(5) (a) taaqa wiipa
man tall
'The man is tall.'
(b) wipa-taqa
tall-man
'tall man'
Sentence (5)(a) is clausal, hence relational in character, and here we would
normally speak of a subject-predicate relation between taaqa and wiipa,
referring to taaqa as a complement of wiipa. On the other hand, (5)(b) is
nominal in character, and in this case it is customary to speak of a head-
modifier relation. Yet in both instances [MAN] is autonomous and
elaborates the trajector of the dependent [TALL], as seen in Fig. 12(a). [TALL]
profiles a relation between an object and an abstract scale of comparison.
This scale pertains to the vertical extension of elongated objects when they
assume their canonical vertical orientation. [TALL] specifies that the upper
extreme of the trajector, with respect to this orientation, falls in the
landmark region of the scale's positive end, i.e. that portion beyond the
neighborhood of the norm (n).
TALL
MAN (IS) TALL
TALL MAN
Figure 12

THE NATURE OF GRAMMATICAL VALENCE
111
The relationship depicted in 12(a) is valid for both (5)(a) and (5)(b),
yet these expressions are quite different in meaning. They differ precisely in
that (5)(a) is relational, while (5)(b) profiles a thing and is therefore
nominal. Hence the composite semantic structure in (5)(a) inherits the relational
profile of [TALL], and that of (5)(b) the nominal profile of [MAN]. These
respective composite structures are sketched in 12(b) and (c). Observe that
in each instance a predication of tallness is made with respect to the man —
the only difference lies in whether that relation is profiled (designated) by
the overall expression, or whether it is merely part of the base.
The difference between an adjectival and an adverbial modifier is
simply that the latter has a relation rather than a thing for its trajector and e-
site. Consider the expression run fast. [FAST] was diagrammed in Fig.
3(b). Its trajector is a process rather than a thing. This schematic process is
elaborated by [RUN] in a way precisely parallel to the way in which [MAN]
elaborates the trajector of [TALL] in tall man. [RUN] is clearly the profile
determinant in run fast, since this expression designates a type of running,
not a type of rapidity — the composite structure is quite analogous to Fig.
12(c). Observe that the notion "adverb" receives a very simple
characterization in this framework: an adverb is a modifier whose head profiles a
relation (as opposed to a thing).
We have now seen two ways in which valence relations can depart from
the prototype: the autonomous structure may function as profile
determinant, and the e-site within the dependent structure may be a relation rather
than a thing. Additional kinds of departure from the prototype can be
illustrated by a series of composite locative particles in Cora, based on data and
analysis provided by Eugene Casad (1982a; Casad and Langacker 1985).
The particular locative particles in question are specialized for
topographical relations in mountainous terrain. Representative examples are
given in (6)(a)-(b).
(6) (a) yuu 'right here at the foot of the slope'
(b) mah 'away up there to the side in the face of the slope'
(c) y 'proximal'; m 'medial'; 0 'distal'
(d) u inside'; a 'outside'
(e) : 'foot of slope'; h 'face of slope'; n 'top of slope'
These particles prove to be essentially regular combinations of three
morphemes each. The alternatives for each morpheme position are given in
(6)(c)-(e).

112
RONALD W. LANGACKER
The semantic structure of mah 'away up there to the side in the face of
the slope' is sketched in Fig. 13. The first component, m 'medial', marks
distance from the speaker (S). The proximal range, indicated by a small,
light circle, essentially encompasses the area within the speaker's physical
reach. The medial range, functioning here as a salient landmark and shown
in heavy outline as a major relational participant, basically includes the
area within the speaker's visual field. The medial predication is thus a sta-
tive relation which locates the trajector within the medial range but outside
the proximal area.
Figure 13
The general contrast in Cora between u 'inside' and a 'outside' assumes
specialized values in particular contexts. The specific versions of u and a
relevant here are defined relative to a domain which finds the speaker

THE NATURE OF GRAMMATICAL VALENCE 113
standing at the foot of a slope and looking straight up the face of this slope.
The slope (including the foot, face, and top) is represented by a rectangle in
the middle diagram at the bottom in Fig. 13. The landmark area for u and
a is defined as the area along the speaker's line of sight as he looks up the
slope from this canonical position. Note that this landmark region extends
along the face of the slope only to the skyline, since the speaker's line of
sight from the foot of the slope cannot curve to take in the region on top.
The morphemes u and a designate stative locative relations, situating the
trajector either inside or outside the landmark region along the speaker's
canonical line of sight; a 'outside' is of course shown in Fig. 13.
Finally, h indicates that the trajector is located in the face of the slope,
as opposed to the foot or the top. Its representation in Fig. 13 should be
self-explanatory, as should be the lines of integration connecting the three
component predicates. It is an obligatory specification of this construction
that the trajectors of the three predicates correspond. The speaker is
obviously the same in the first two components, and the slope is the same in the
second two. The three predicates are therefore tightly integrated by
correspondences connecting shared substructures.
The composite structure is obtained simply by superimposing
corresponding entities. The result is a complex locative relationship in which the
trajector is simultaneously located with respect to three parameters and
thtee landmarks, one contributed by each component predicate. Taken
together these add up to a fairly precise specification that can be glossed as
in (6)(b).
This construction departs from prototypical valence relations in three
ways. First, the construction is not binary. There is no apparent reason to
break the three-morpheme sequence down into two levels of constituency;
the three specifications are essentially coordinate. Second, there is no
obvious sense in which any of the components is conceptually dependent
relative to the others, hence there is no e-site. Finally, there is no basis for
singling out any of the three component predicates as profile determinant. The
composite expression mah does not designate any of the three component
locative relations in particular, but rather the complex locative relation
defined by coordinating the locative specifications along the three
parameters. Because there is no profile determinant, nor any asymmetry in terms of
conceptual autonomy/dependence, none of the component structures can
be identified as a head, complement, or modifier with respect to the others.

114
RONALD W. LANGACKER
Further Departures from Canonical Valence Relations
For subsequent examples, I will adopt the abbreviatory notations given
in Fig. 14. A line between two entities indicates that they participate in a
stative relation, which may be characterized in either schematic or specific
terms. Only one component state is shown explicitly in the abbreviation for
a process, but it should be understood that there is one such state for every
point in its temporal profile, represented by the heavy-line portion of the
time arrow. The trajector of a process appears always to be a thing, as the
abbreviations suggest.
G □
P
rj
?
thing entity specified schematic specified schematic
(thing or stative stative process process
relation) relation relation
Figure 14
Additional types of departures from canonical valence relations can be
illustrated by the Cahuilla data in (7) (from Seiler 1977: 300f).
(7) (a) ne-'as kiyul
my-pet fish
'my fish'
(a') ne-'as tamawet
my-pet mockingbird
'my mockingbird'
(b) ne-wes-'a navet
my-plant-NR cactus
'my cactus'
(b') ne-wes-'a sandiya
my-plant-NR watermelon
'my watermelon'
This is a type of noun-classifier construction frequently found in Uto-Azte-
can possessive expressions. Instead of going directly on the possessed noun,

THE NATURE OF GRAMMATICAL VALENCE 115
the possessor prefix attaches to a more schematic noun, or classifier, to
which the possessed noun stands in a sort of appositive relation. Cahuilla
has a whole series of classifiers, only two of which are exemplified here.
The classifier 'as is used for pets and other domesticated animals, and wes-
'a, a nominalization of the verb stem wes 'plant', is used for crop plants
sown in a row.
The semantic structure of (7)(a) is sketched in Fig. 15. At the lower
level of constituency, [MY] functions as a modifier of [PET]. I assume that
[MY] profiles a schematically-characterized stative relation, whose
landmark is identified with the speaker; it simply indicates that the trajector
interacts with the speaker in some unspecified way. [PET] is a relational
noun: it profiles a thing, identified schematically as some kind of animal
(AN), but also makes salient internal reference to another individual, who
participates in an unprofiled relationship of ownership with the designated
animal, and thus functions as a type of landmark (reference point) within
this nominal predication. The valence relation between [MY] and [PET]
involves two correspondences. First, the schematic trajector of [MY] is put
in correspondence with the profile of [PET], and is elaborated by the latter
predication. A second correspondence associates the landmarks of the two
component predications — more precisely, the schematic relationship
profiled by [MY] is equated with the ownership relation in the base of [FETj.
Since [PET] is the profile determinant (head) in this construction, the
composite structure [MY-PET] is nominal rather than relational: it profiles an
animal whose characterization is schematic apart from the specification that
it is owned by the speaker.
Consider now the second level of constituency, where ne-'as 'my pet'
combines with kiyul 'fish'. The profile of [FISH] corresponds to the
schematically-characterized profile of [MY-PET], which is substantially
elaborated by the former predication. The profile of [MY-PET] therefore
functions as an e-site, and [MY-PET] can be regarded as conceptually
dependent with respect to [FISH]. The composite structure (MY-PET-
FISH), which designates a fish owned by the speaker, is surrounded by a
closed curve rather than a box to indicate its presumed status as a novel
expression (rather than a fixed unit).
This latter valence relation is noteworthy in two respects. For one
thing, both component predications are nominal: [MY-PET] is dependent
in the sense of being elaborated by [FISH] rather than conversely (cf. Fig.
11), but neither component structure is a relational predication. The lack of

116
RONALD W. LANGACKER
MY-PET-FISH
Figure 15
a relational element is problematic in standard predicate-argument
accounts of semantic structure (where predicates are considered n-place
relations), but it poses no problems whatever in the present framework,
where valence relations depend crucially only on the existence of
correspondences between substructures of the component predications. The
second noteworthy feature of this construction is that neither component
structure can be singled out as the profile determinant. We saw this situation
previously in the Cora example, but here the reason is different: since the
two component structures profile corresponding entities, the same
composite structure would result regardless of which one we identified as the
profile determinant. Because there is no clear-cut profile determinant, we lack
the basis for positing either a head-modifier or a head-complement relation;
the construction is basically appositional.
As a final example, Fig. 16 sketches the semantic structure of ne-wes-'a
'my plant', the classifier portion of (7)(b). At the lower level of
constituency, the nominalizer -'a combines with the verb stem wes 'plant'.

THE NATURE OF GRAMMATICAL VALENCE
117
[PLANT] is a process predicate, represented with the abbreviatory notation
introduced in Fig. 14. The nominalizing predication [NR] characterizes a
thing by virtue of the role it plays in a process. The base of [NR] is thus
analyzed as a schematic process, and its profile is identified as the landmark
of this base process. In the valence relation between [PLANT] and [NR],
the entire schematic process constituting the base of [NR] functions as e-
site; the specific process profiled by [PLANT] is put in correspondence with
the base process of [NR], so that [PLANT] spells out in specific terms the
process with respect to which the profile of the nominalization is charac-
MY-PLANT-NR
Figure 16

118
RONALD W. LANGACKER
terized. [NR] is both dependent (being elaborated by the other component)
and the profile determinant, so we must speak of a complement (rather
than a head-modifier) relation. The composite structure [PLANT-NR]
designates a thing identified as the landmark of the process of planting.
The integration of [PLANT] and [NR] illustrates two further
departures from the canonical valence relation schematized in Fig. 11. For the
first time in our examples the e-site is neither the profile nor a subpart of
the profile — rather it is a process in the base. Second, the e-site is not a
proper substructure within the dependent predicate, but is in fact
exhaustive of this predicate (apart from its profile specification). It is of course not
unexpected, as a limiting case, that the substructure functioning as e-site
might coincide with the whole.
At the second level of constituency, the trajector of [MY] corresponds
to the profile of [PLANT-NR], being parallel to Fig. 15 in this regard. A
second correspondence associates the landmark of [MY] (the speaker) with
the trajector of the process constituting the base of [PLANT-NR]; the
interaction which [MY] predicates between its trajector and landmark is
thus equated with the process of planting, and the speaker is identified with
the trajector of this unprofiled process. [PLANT-NR] is conceptually
autonomous relative to [MY] and also the profile determinant, so the two
participate in a head-modifier relationship. The composite structure [MY-
PLANT-NR] is a nominal predication which designates the landmark of the
process of planting, whose trajector is the speaker. The resulting expression
ne-wes-'a 'my plant/what I planted' is then capable of combining with an
appositional noun like navet 'cactus' at a higher level of organization, as in
our previous example.
We may now summarize the various ways in which a grammatical
valence relation can depart from the prototype. A valence relation need not be
binary, and it is not necessary that there be a clear asymmetry between an
autonomous and a dependent structure. If there is such an asymmetry, the
dependent structure need not be relational, and its e-site does not have to
be a thing included in the profile: it can be a relation rather than a thing,
and it can be an unprofiled facet of the base (or even subsume the base).
Either the autonomous or the dependent structure can function as profile
determinant, and in some instances the components contribute equally to
the profile of the composite structure. Finally, a valence relation often
involves multiple correspondences. The existence of at least one
correspondence is perhaps the only invariant feature of valence relations.

THE NATURE OF GRAMMATICAL VALENCE
119
Scope and Morphological Layering
We can now return to the problem posed earlier, namely the tendency
for morphological layering to correlate with semantic scope, as illustrated
by the Luiseno sentences in (1). The present conception of grammatical
valence allows a straightforward account of the correlation and the expressions
which display it. Let us focus on (l)(d), repeated here as (8).
(8) noo poy rjee-vicu-ni-q
I him leave-want-make-TNS
T made him want to leave.'
Limiting our attention for the moment to the verb, its constituency tree is
given in Fig. 17.
(LEAVE-WANT-MAKE-TNS)
(LEAVE-WANT-MAKE )-* -[TNS]
[ LEAVE-WANT}-- -{MAKE]
[ LEAVE}-* -[WANT]
Figure 17
Two factors combine to account for the correlation between
morphological layering and semantic scope. First, at each level in the
constituency hierarchy at the semantic pole, the conceptually dependent structure
is also the profile determinant (hence we are dealing with head-complement
rather than head-modifier relations). This is the canonical alignment, and it
constitutes what was recognized as "semantic scope" in the generative
semantic framework. At the lowest level, for instance, [LEAVE]
elaborates the landmark of [WANT], hence [WANT] is dependent; [WANT] is
also the profile determinant, since qee-vidu 'want to leave' is a kind of
wanting, not a kind of leaving. Thus [WANT] imposes its profile on the
composite structure, overriding the profile of [LEAVE], which serves to elaborate
a substructure of [WANT]. This is the type of relationship people have in
mind when they say that [LEAVE] is "in the semantic scope of" [WANT].
The second factor pertains to integration at the phonological pole,

120
RONALD W. LANGACKER
which we have largely ignored until now. The notions of autonomy and
dependence are equally important at the phonological pole in valence
relations as at the semantic pole. In the case of word structure, they amount to
the distinction between "root" or "stem" on the one hand, and "affix1' on
the other. An affix is morphologically dependent in the sense that it is
characterized in part by its position relative to a root or stem, and thus
makes inherent reference to a schematically-specified root or stem as part
of its own internal structure. This schematic stem within each affix serves as
e-site in a valence relation and is elaborated by a specified stem. A root or
stem is autonomous in the sense that it makes no salient internal reference
to another phonological entity relative to which it is positioned.
f rjeevicuniq j
Figure 18
The phonological pole of tjee-vicu-ni-q 'made want to leave' is sketched
in Fig. 18. At each successive level in the hierarchy, an affix (dependent)
combines with a root or stem (autonomous) to form a higher-order stem or
word (also autonomous). Observe, moreover, that at each level the
autonomous semantic structure is symbolized by the autonomous phonological
structure, and the dependent semantic structure by the dependent
phonological structure. There is consequently a kind of "harmony"
between the alignment of autonomous and dependent structures at the two
poles, a perfectly natural phenomenon reminiscent of numerous other types
of "harmonization" observable between different facets of linguistic
organization (note, for instance, the tendency for linear ordering to correlate with
the sequencing of events). It is hardly surprising that this parallel alignment
of autonomy/dependence at the two poles represents a strong universal
tendency (though exceptions can be found). When it also happens that the

THE NATURE OF GRAMMATICAL VALENCE
121
dependent semantic structure functions as the profile determinant (this too
being the canonical situation), we obtain the usual correlation between
morphological layering and semantic scope. As we work from the root
outwards, each successive morphological increment symbolizes the
introduction of a semantic predication which has the previously-constructed
predication "in its scope" (i.e. they participate in a head-complement
relationship).
Let us conclude by examining in specific detail how the complex verb
rjee-vicu-ni-q 'made want to leave' is assembled at the semantic pole. We
must first consider the semantic poles of the component morphemes.
[LEAVE] was diagrammed previously in Fig. 3(c). The other three
predicates are sketched in Fig. 19. Each profiles a process, i.e. a series of
relational configurations distributed through conceived time. Only a single
configuration is explicitly represented in each diagram, which must
nevertheless be interpreted as indicating the evolution of the profiled relationship
throughout the temporal profile (marked by the heavy-line portion of the
time arrow).
Cb) MAKE (c)
(a)
1m
WANT
1 •
I :
i ^
TNS
•
ii • |
raj
speech
event
Figure 19
The trajector of -vicu 'want' is a schematically-characterized thing, and
its landmark is a schematic process. I have used a broken-line arrow to
indicate the desire held by the trajector with respect to the landmark. Luiseno
-vicu differs from English want in that its trajector is obligatorily equated
with the trajector of the landmark process; hence the correspondence line
connecting the two trajectors in Fig. 19(a).

122
RONALD W. LANGACKER
The causative predicate -ni 'make' is diagrammed in 19(b). The trajec-
tor is again a thing, but there are two salient landmarks, one nominal and
the other relational. The relational landmark is the process induced by the
overall trajector; the causative relationship between this trajector and the
landmark process is represented by the double arrow. The nominal
landmark is the same entity that functions as the trajector of the landmark
process. [MAKE], in other words, is one of many predicates that has the
potential for taking two non-subject complements, one of them a direct
object nominal and the other a relational complement having the direct
object as one of its participants. (Cf. throw in He threw the clothes into the
closet — the clothes is construed as both the direct object of throw and the
subject (trajector) vis-a-vis the relational complement into the closet.)
Lastly, the tense morpheme -q marks present or recent-past time (and
also indicates that the subject is singular). In combination with -ni 'make',
it receives the recent-past construal, as indicated in Fig. 19(c). For reasons
beyond the scope of this paper (see Langacker 1985), it is analyzed as
profiling a schematic process, specified in the base as situated prior to the time
of speaking (but in temporal proximity to the speech event).
The full constituency tree for the semantic pole of nee-vicu-ni-q 'made
want to leave' is provided in Fig. 20, which assumes that nee-vicu represents
a familiar combination and has the status of a unit, while nee-vicu-ni and
nee-vicu-ni-q constitute novel expressions. Observe that the three
predicates described in Fig. 19 all function as profile determinants, each at its
own level of constituency. Moreover, each is conceptually dependent at its
own level of organization, for it contains a salient substructure that is
elaborated by the stem it combines with. Each suffix is therefore semantically a
head, and the stem it attaches to is a complement to that head.
At the first level of organization, the specific process profiled by
[LEAVE] elaborates the schematic process serving as the landmark of
[WANT], yielding the composite structure [LEAVE-WANT]. Recall that
[WANT], as part of its internal structure, establishes a correspondence
between its overall trajector and the trajector of the schematic process
functioning as its landmark. As a result, when the specifications of
[LEAVE] are superimposed on those of the processual landmark of
[WANT] to derive the composite structure [LEAVE-WANT], the overall
trajector of this composite predication is equated with that of the leaving,
i.e. what the trajector wants is for himself to leave (rather than some other
individual). The correspondence line internal to [WANT] therefore

THE NATURE OF GRAMMATICAL VALENCE
123
IZAVE-'WANT-MAKE- TNS
31
ltn
lmi
jL_1
@:&.'.o
Figure 20

124
RONALD W. LANGACKER
accomplishes what generative grammarians formerly achieved by their
transformation of Equi-NP Deletion. No derivation from a hypothetical
deep structure is required in this framework.
At the second level of organization, [LEAVE-WANT] elaborates the
relational landmark of [MAKE] to form the composite structure (LEAVE-
WANT-MAKE). This structure predicates a causal relationship between
the overall trajector and the process of desiring to leave. At the third level,
finally, the process (LEAVE-WANT-MAKE) elaborates the schematic
process profiled by the tense predication, so that the composite structure
(LEAVE-WANT-MAKE-TNS) situates just prior to the time of speaking
the profiled process of one individual causing another to want to leave.
Observe that both individuals are characterized only schematically at this
level of organization.
The complex structure rjee-vicu-ni-q 'made want to leave' is
categorized as a verb because its composite structure designates a process.
As with any other verb, it is eligible to participate in further valence
relations serving to elaborate the schematic things functioning as its trajector
and landmark. The probable constituency tree for these further levels of
composition is presented in Fig. 21 (it is exactly analogous to Fig. 9,
specifying the integration of subject and direct object nominals with a simple verb
of Hopi). At the first level of organization, [HIM] (i.e. a third-person
singular pronoun) elaborates the nominal landmark of (LEAVE-WANT-
MAKE-TNS); since the latter predication is the profile determinant, the
composite structure (HIM-LEAVE-WANT-MAKE-TNS) designates a
process, one equivalent to the composite structure of Fig. 20 except that its
landmark is now more specific. At the second level of organization, the
schematic trajector of (HIM-LEAVE-WANT-MAKE-TNS) is elaborated
by [I]. Again we have a complement relation (i.e. the processual
predication is the head), so the final composite structure (I-HIM-LEAVE-WANT-
MAKE-TNS) also profiles a process, as does any finite clause.
(1-HIM-IiEAVE-WANT-MAKE-TIB)
[IH <HIM-UAVE-WANT-MAKE-TNS )
[HIMh* (LEAVE-WANT-MAKE-TNS)
Figure 21

THE NATURE OF GRAMMATICAL VALENCE
125
Observe that the postulated constituency of this sentence, as reflected
in Figs. 17 and 21, is perfectly compatible with the one suggested by
phonological and morphological considerations. [I] and [HIM] are added
only after the entire complex verb has been assembled, as reflected by their
status as separate words: noo poy yee-vicu-ni-q. Nevertheless, the
correspondences which figure in the various valence relations (and those internal
to certain predications) properly establish the speaker as the trajector of
the causative relationship, and the third person singular individual as the
trajector with respect to both the wanting and the leaving. All of this is
accomplished without deriving the sentence from a hypothetical underlying
structure posited solely to accommodate its supposed predicate-argument
configuration (cf. Fig. 4(b)), and without special rules like Predicate
Raising and Equi-NP Deletion (Fig. 4(a)).
As a general matter, this framework promises to reconcile semantic
and grammatical constituency with the constituency one is led to posit on
phonological grounds. Basic grammatical relations are not defined with
respect to specific constituent-structure configurations, but rather through
correspondences and the trajector/landmark alignment internal to every
relational predication. The same correspondences, and hence the same
grammatical relations, can be established through alternate compositional
paths leading to the same overall composite structure. This is why the
conventions of a language often permit alternate word orders or phonological
phrasings for otherwise equivalent expressions. Correspondences are
essential to grammatical valence relations, but constituent structure is to some
degree incidental and variable.

A Usage-Based Model
Ronald W. Langacker
University of California, San Diego
Introduction
The generative tradition has strongly emphasized the importance of
"generality" in linguistic description. The quest for generality is of course
fundamental to the scientific enterprise; we may certainly accept its validity
for linguistics. It is not necessarily obvious, however, how this abstract
methodological imperative is best adapted and applied to the problems of
our discipline, with its own distinctive subject matter. Generative theorists
have applied it to linguistics in a specific manner that has had a powerful
impact on their conception of linguistic theory and description. I believe
their interpretation to be inappropriate for natural language, and its
influence to have been a continuing source of difficulty in dealing with linguistic
phenomena.
Summarized in (1) are three basic tenets of classic generative theory.
(1) (a) Economy:
A grammar should account for the widest possible array of
data with the fewest possible statements.
(b) Generativity:
A grammar is a set of statements specifying in full and
explicit detail how expressions are constructed; it gives a
well-defined set of expressions as "output".
(c) Reductionism:
If the rules of a grammar fully describe the composition of
a particular structure, that structure is not itself
individually listed in the grammar.
© Ronald W. Langacker

128
RONALD W. LANGACKER
The economy principle holds that the shortest grammar is the best
grammar, other things being equal; redundancy is therefore to be avoided. The
generativity principle construes the grammar of a language as a
self-contained algorithmic device, consisting primarily of rules for constructing
well-formed expressions. The third principle follows from the other two: if
a grammar is a set of rules for constructing expressions, and contains the
fewest statements possible, then any expression constructed by these rules
must itself be omitted from the grammar. Separately listing an expression
computable by general rules would be redundant (and redundancy is evil).
Though initially this seems quite reasonable, in practice it has had
some unfortunate consequences. One result is that virtually all research
activity has been dedicated to searching for general rules and universal
principles. Now obviously, the search for generalizations is a prime
objective — the question I raise is one of balance, and whether the
generalizations rest on adequate empirical foundations. It is apparent, for example,
that generative grammarians have never dealt seriously with lexicon in its
own terms. Nor do they often attempt, after stating a rule in general
fashion, to document the actual extent of its applicability and the various
factors that influence its felicity (cf. Gross 1979). The abstract systems of rules
and principles constructed by theorists seldom emerge organically from the
sensitive, fine-grained description of fully representative data (hence their
mortality rate is high and their lifetime often tragically short). In brief, all
the glory attaches to general principles and abstract theory; careful
attention to the minutiae of language data is left for those without the insight and
imagination to be good theorists.
These practical consequences are clearly a matter of judgment, and we
will pursue them no further. More significant is the issue of whether the
general methodological imperative of generality is appropriately
implemented in linguistics by the positions in (1). One can argue, for
instance, that generativity is purchased only at the price of arbitrary
assumptions and a priori boundaries that exclude from the description
substantial portions of the phenomenon that is putatively being described (a
case in point is the distinction between "semantics" and "pragmatics" — cf.
Haiman 1980; Langacker in press: ch. 4). One can also suggest that
economy and reductionism, as defined in (1), are of dubious validity in any
framework that makes a serious claim of psychological reality (cf.
Langacker 1982a). We will focus on this latter point.
The assumptions in (1) constitute a "minimalist" conception of Unguis-

A USAGE-BASED MODEL
129
tic knowledge; the grammar of a language is reduced to the smallest
possible set of statements, with all redundancy avoided. It is also a "top-down"
conception emphasizing computation: anything which follows from general
statements is omitted from the grammar, on the assumption that it is
computed rather than being represented individually. However, we have no a
priori reason to believe that the cognitive representation of a language
conforms to this conception. It is plausible, psychologically, to suppose that
speakers represent linguistic structures in different ways, with considerable
redundancy built in. It is also reasonable to assume that many structures are
learned as established units even when they also follow from general
principles — the computability of a structure does not in principle preclude its
learnability and inclusion as a distinct element in the cognitive
representation of the linguistic system. The generative grammarian might reply that
such considerations belong to the theory of performance, not a theory of
competence. But at best the competence/performance distinction is unclear
and problematic; as things stand, to invoke it in this manner is essentially
vacuous. In actual practice, the effect of this distinction has been to insulate
the framework from any possible attack based on its obvious psychological
implausibility.
If claims of psychological reality are taken seriously, questions of
economy assume the status of empirical issues, as opposed to
methodological ones. Is it in fact true that a speaker arrives at any kind of redundancy-
free representation of linguistic structure? Do speakers in fact avoid
learning specific structures as separate units if they happen to conform to general
rules? A description of linguistic ability that answers these questions
negatively cannot legitimately be attacked on the grounds that the description
fails to achieve maximum simplicity: the question of simplicity only arises
for two descriptions of the same range of data, but the issue at stake is
precisely that of determining what the relevant data is (i.e. what are the
cognitive structures that we are trying to model?). One could just as well omit
phonology from the grammar on the grounds that a grammar without a
phonological component is simpler than a grammar with one.
For a specific example, consider such expressions as dogs, trees, toes,
pins, bells, walls, and eyes, which instantiate a regular pattern of plural-
noun formation. Clearly, a regular rule of plural formation can be given for
English, and these expressions conform to the rule. According to the
principle of reductionism, incorporating this rule in the grammar of English
precludes the listing of individual plural forms like dogs, trees, toes, etc. The

130
RONALD W. LANGACKER
rules allow their computation from the noun stems, hence their inclusion in
the grammar would be redundant.
The goal of cognitive grammar is to characterize those psychological
structures that constitute a speaker's linguistic ability, i.e. his grasp of
established linguistic convention. This notion inspires an alternate
approach to forms like dogs, trees, toes, etc. A typical speaker uses
frequently-occurring expressions like these on countless occasions; at least
some of them must attain the status of "units" (i.e. familiar, thoroughly
mastered structures — cognitive routines). In fact, the pattern itself can
only be learned through the observation of instantiating expressions, some
of which most likely become units before the pattern is extracted; it is
implausible to suppose that these plural forms suddenly lose their status as
familiar units when the rule is acquired, and must henceforth be computed
from scratch.
An alternative conception is sketched in Fig. 1. The grammar of a
language is defined as a "structured inventory of conventional linguistic units".
Specific expressions are included in this inventory provided that they have
the status of units — a reasonable assumption for dogs, trees, etc. Also
included in the grammar are "schemas" extracted to represent the
commonality observed in specific expressions (both units and non-units). The
schema corresponding to a grammatical pattern can be regarded as a
template for the construction of instantiating expressions. Hence the
schema for nouns like dogs, trees, etc. is a complex structure whose internal
organization is exactly parallel to these plural forms, except that a
schematic noun stem (given as [THING/X]) occurs in lieu of a specific noun stem.
The schema therefore captures the pertinent generalization, and its
categorization of instantiating expressions constitutes their "structural
description". The coexistence in the grammar of the schema and
instantiations affords the speaker alternate ways of accessing a complex but regular
expression with unit status: it can simply be activated directly, or else the
speaker can employ the schema to compute it. Moreover, the schema is
available for the computation of novel instantiations (e.g. quagmires)', if
such an expression is frequently employed, it may very well become
established as a unit and thus be incorporated per se in the grammar. The
specific array of instantiations having the status of units doubtless varies
from speaker to speaker (and changes with experience for an individual
speaker), but this is not considered problematic.

A USAGE-BASED MODEL
131
GRAMMAR
SCHEMA
THING
[dog]
|dog|
[PPl
l-s 1
[TREE
|jtree
PLjl
-s J
FIXED EXPRESSIONS
NOVEL
EXPRESSION
Figure 1
Generative grammarians have normally resisted the idea that regular
expressions should be listed in a grammar, on the presumption that listing
entails a failure to capture significant generalizations. I refer to this attitude
as the "rule/list fallacy". It is fallacious because it tacitly presupposes only
two options: rules vs. lists. But nothing in principle prevents a third option,
namely positing both rules (i.e. schemas) and lists, as shown in Fig. 1. By
their very nature, schemas embody generalizations. The implicit
assumption that rules and lists are mutually exclusive merely reflects the generative
conception of economy, as described in (1), whose appropriateness for
natural language is precisely what is at issue.
In describing cognitive grammar as a "usage-based" model of language
structure, I have in mind the "maximalist", "non-reductive", and "bottom-
up" character of the general approach (as compared to the minimalist,
reductive, and top-down spirit of the generative tradition). The full import
of this description will gradually become apparent. For now, let us briefly
examine the basic thrust of each term.
The minimalist spirit of generative theory reflects an archetypal
conception of the linguistic system as a self-contained and well-behaved set of
general rules; though nobody believes that a language consists solely of
general rules, this archetype strongly influences virtually every aspect of
generative theory and descriptive practice. By contrast, the maximalist
conception views the linguistic system as a massive, highly redundant inventory
of conventional units. These units run the gamut from full generality to
complete idiosyncrasy, and no special significance attaches to any distinc-

132
RONALD W. LANGACKER
tions one might draw along this scale. Valid generalizations are sought and
captured (represented in the grammar by schematic units), but fully general
statements are probably a distinct minority: rather than being prototypical
for language, exceptionless rules are special, atypical cases. Moreover, the
maximalist outlook leads one to anticipate a gradation between linguistic
and extra-linguistic knowledge (cf. Langacker in press: ch. 2); while it does
not deny the possibility of innate structures specific to language, neither
linguistic ability nor the grammar of a particular language is conceived as a
discrete "module" with well-defined boundaries (pace Fodor 1983).
Cognitive grammar is non-reductive by virtue of recognizing both rules
or patterns and the individual knowledge of specific structures that conform
to them. A schema and its instantiations represent different facets of
linguistic knowledge, and if they have the status of units, both are included in
the grammar of a language. One advantage of this approach is its ability to
accommodate, with no special apparatus, instances where a fixed
expression is more detailed and elaborate than the structure that a rule or schema
would allow one to compute (e.g. an eraser is a particular type of object
with specific properties, not just 'something that erases'). The grammar is
not conceived as a constructive device giving expressions as "output'1, but
simply as providing a speaker with an inventory of symbolic resources that
he — the speaker — can employ for the construction of novel expressions,
using all the information and abilities at his disposal. As one such resource,
the schema describing a pattern of composition is not itself responsible for
actually constructing an expression. Instead it serves a categorizing
function: it furnishes the minimal specifications an expression must observe to
be categorized as a valid instantiation of the pattern it embodies. An
expression may satisfy these specifications, and thus be judged compatible
with the schema, even if its characterization is more precise and fully
articulated than anything predictable just from the schema and the component
morphemes.
Finally, the model is said to take a "bottom-up" (rather than a top-
down) approach. What this amounts to is a redistribution of emphasis:
instead of being almost solely concerned with general rules and principles,
we must also give substantial weight to their arrays of conventional
instantiations, investigating the actual extension of the patterns in question and the
factors that influence it. Furthermore, since patterns are abstracted from
specific instances, we need to investigate the schematization process. We
know, for example, that speakers learn and manipulate specific expres-

A USAGE-BASED MODEL
133
sions; but we do not know, in any direct way, precisely what degree of
schematization they achieve, i.e. how abstract and general the rules are that
they manage to extract from more specific structures. I suspect that
speakers differ somewhat in this regard, and do not invariably arrive at the
highest-level schemas that the data would support. In any event, the
omnipotence of high-level generalizations is not a matter of a priori necessity.
Though regularities are obviously noted and employed in the computation
of novel expressions, it is quite conceivable that low-level schemas are more
important for this purpose than highly-abstract schemas representing the
broadest generalizations possible. If high-level schemas are extracted, they
may be of only secondary significance, serving more of an organizing
function than an active computational one.
The Network Conception
Critical to the formulation of a usage-based theory is a coherent view
of linguistic categorization. A particular model of categorization, the "crite-
rial attribute" model, has generally been accepted without serious question
in the Western intellectual tradition. More recently, the "prototype" model
has been advanced as an alternative with greater claims to cognitive
plausibility. My own proposal, the "network" model, represents a synthesis of
prototype theory and categorization based on schemas. (For general
discussion, see Lakoff 1982b, in press; Langacker in press: chs. 10-11.)
In a strict formulation of the criterial attribute model, a category is
defined by a fixed set of properties or features. These attributes are
necessary and sufficient conditions for category membership, affording absolute
predictability in this regard: if an entity possesses all the criterial properties,
it is a member of the class; otherwise it is not. Class membership is
consequently an all-or-nothing affair; there are no degrees of membership, nor
does a category display any significant internal structure.
The prototype model was pioneered by Eleanor Rosch, and has been
presented and supported in numerous publications (e.g. Rosch 1973b,
1975, 1977, 1978). In this model, a category is defined with reference to a
prototype, i.e. a schematized representation of typical instances. Entities
that conform to this prototype are accepted unproblematically as "central"
members of the class. Non-conforming members may nevertheless be
assimilated to the category as "peripheral" members provided that they are
judged as being similar to the prototype in certain respects. A class is struc-

134
RONALD W. LANGACKER
tured internally by virtue of its organization into central and peripheral
members; moreover, class membership is a matter of degree, reflecting the
distance of a member from the prototype. Because there is no specific
checklist of criterial features, membership in a category is not subject to
absolute predictability (indeed, there need be no significant features that
are shared by all class members). Whether an entity qualifies depends on
the judgment of the categorizer, and his tolerance in accepting members
that diverge from the prototype. There is no fixed limit on how far
something can depart from the prototype and still be assimilated to the class, if
the categorizer is perceptive or clever enough to find some point of
resemblance to typical instances.
The model I have adopted for cognitive grammar incorporates the
prototype model as a special case. The members of a category are analyzed as
nodes in a network, linked to one another by various sorts of "categorizing
relationships". One kind of categorizing relationship is "extension" from a
prototype, which may either be a "local" prototype or the "global"
prototype for the category as a whole. The notion of extension, symbolized by
a dashed arrow, implies some conflict in specifications between the basic
and extended values; hence [A] > [B] indicates that [B] is incompatible
with [A] in some respect, but is nevertheless categorized by [A]. A second
kind of categorizing relationship holds between a "schema" and a structure
that "elaborates" or "instantiates" the schema. Symbolized by a solid
arrow, e.g. [A] —> [B], the relationship amounts to one of specialization:
[B] conforms to the specifications of [A], but is characterized with finer
precision and detail. (Note that a schema is not viewed as a set of features or
criterial attributes — it is an integrated structure whose internal
organization is parallel to that of its instantiations.) A third kind of categorizing
relationship, [A] <— > [B], amounts to a perception of mutual similarity;
it differs from extension only by lacking directionality. For any type of
categorizing relationship, we can speak of the "distance" between [A] and
[B], i.e. how much modification or elaboration of [A] is required to arrive
at [B]. Beyond this, the nodes and categorizing relationships comprising the
network vary greatly in their cognitive salience and degree of
entrenchment.
The network model is conveniently illustrated by the alternate senses
of a polysemous lexical item. A fragment of the network required for the
English verb run is presented in Fig. 2. With heavy lines (to indicate special
cognitive salience) I have singled out the global category prototype, namely

A USAGE-BASED MODEL
135
the conception of rapid bipedal human locomotion; this is presumably the
meaning that is acquired first, and also the one most likely to be activated
in a neutral context. We cannot be certain how far "upward" a speaker
extends this network through the process of abstraction (schematization),
and in particular, whether he extracts a "superschema" having all other
nodes as direct or indirect instantiations. Nor do we know how far
"downward" a speaker articulates the network into progressively more specialized
notions. Speakers may very well differ in these respects, and also in the
specific set of categorizing relationships they establish between nodes. But
although the precise configuration of the network is variable and even
indeterminate, the need to postulate some type of network is seemingly beyond
dispute: the meanings of a commonly-used lexical item define a "complex
category", i.e. one that is not reducible to a single structure (node).
[""rapid ""j
]_MOTIONj
RAPID
n-LEGGED \
LOCOMOTION!
COMPETITIVE
POLITICAL
ACTIVITY
(candidate)
s
RAPID
[MECHANICAL
MOTION
s*r\ (engine)
RAPID
FLUID
MOTION I
(vater)
S
RAPID
2-LEGGED
[LOCOMOTION
(person)
\
RAPID
k-LEGGED
[LOCOMOTION
(animal)
\
COMPETITIVE
2-LEGGED
LOCOMOTION
(race)
RAPID
k-LEGGED
[LOCOMOTION
(dog)
RAPID
k-LEGGED
LOCOMOTION
(horse)
Figure 2
A strict reductionist approach would seek maximum economy by
positing only a single structure to represent the meaning of a lexical category.
However, if our goal is to properly characterize a speaker's knowledge of
linguistic convention, any such account is unworkable. From neither the
category prototype alone, nor from an all-subsuming superschema (should
there be one), is it possible to predict the exact array of extended or

136
RONALD W. LANGACKER
specialized values conventionally associated with a lexeme (out of all those
values that are cognitively plausible). A speaker must learn specifically, for
instance, that run is predicated of people, animals, engines, water, hosiery,
noses, and candidates for political office; the conventions of English might
well be different. Equally deficient is the atomistic approach of treating the
individual senses as distinct and unrelated lexical items. The claim of
massive homonymy implied by such an analysis is simply unwarranted — it is
not by accident, but rather by virtue of intuitively-evident semantic
relationships, that the meanings are symbolized by the same form. A
network representation provides all the necessary information: an inventory of
senses describing the expression's conventional range of usage; the
relationships these senses bear to one another; schemas expressing the
generalizations supported by a given range of values; and specifications of distance
and cognitive salience.
Some classic problems of lexical analysis are readily addressed in terms
of the network model, namely the distinction between "polysemy" and
"homonymy", and that between "ambiguity" and "vagueness". The first
distinction hinges on whether the various senses associated with a given
form are semantically related. In the network model, semantic relatedness
is a matter of degree, so polysemy vs. homonymy does not reduce to a
simple dichotomy. Two senses may be related directly, by a categorizing
relationship, or else indirectly, through a chain of such relationships. Direct
relationships range continuously along the distance parameter (e.g. the
'rapid 4-legged locomotion' sense of run lies closer to the prototype 'rapid
2-legged locomotion' than does 'rapid mechanical motion'); moreover,
relationships at a particular distance may differ in salience (consider the
gradual "fading" of conventional metaphors). Since speakers are very adept
at perceiving semantic relationships, and since comparison is encouraged by
common symbolization, it is seldom safe to assume that no connection
whatever is established between the alternate senses associated with a form.
Even a tenuous relationship, dimly perceived, is still a relationship, and
though certain phenomena may presuppose a close or salient connection
(so that, for instance, Tom and his dog are both running is more felicitous
than Tom and the water are both running), there is no basis for positing a
specific cut-off point along the scale of semantic relatedness where
polysemy abruptly stops and homonymy begins. Homonymy is better
analyzed as the endpoint along the cline of relatedness — it is the limiting
or degenerate case of polysemy, where the only relationship between two

A USAGE-BASED MODEL
137
senses consists in their common phonological realization. Hence the actual
descriptive problem is not to distinguish between homonymy and polysemy,
but rather to characterize semantic networks as fully and accurately as
possible.
Whereas the issue of polysemy pertains to whether two or more senses
are semantically related, the distinction between "ambiguity" and
"vagueness" depends on whether multiple senses should be posited in the first
place. A commonly-employed test concerns the number of interpretations
supported by sentences like those in (2):
(2) (a) Tom has an uncle, and Bill does too.
(b) Tom has two ears, and Bill does too.
A possible interpretation of (2)(a) is that Tom has a maternal uncle (i.e. a
mother's brother) while Bill's uncle is paternal (a father's brother), or
conversely; the anaphoric relationship between clauses does not require that
the two uncles be of the same type. By contrast, (2)(b) cannot easily be
construed as meaning that Tom has two organs of hearing, while Bill has
two cobs of corn — both clauses must be interpreted as referring to the
same type of ear. From such observations, it is normally concluded that
uncle displays vagueness rather than ambiguity: it has the single meaning
'parent's brother' (unspecified for the gender of the linking relative) rather
than the two distinct senses 'mother's brother' and 'father's brother'. On
the other hand, ear is ambiguous, with anaphora sensitive to the difference
between the two meanings.
David Tuggy (1981) has argued that the felicity of interpretations
involving a mixture of types is often a matter of degree. Thus (3)(a) seems
peculiar if Tom is an artist doing a portrait while Bill is putting a new coat
of paint on a fence, but the sentence is much less infelicitous if Tom is
painting a massive mural instead.
(3) (a) Tom is painting, and Bill is too.
(b) Tom is talking, and Bill is too.
(c) Tom is writing, and Bill is too.
If Tom is giving a formal lecture, and Bill is outside in the corridor chatting
with a friend, (3)(b) is less than perfect, but much better than (2)(b) under
the mixture-of-type interpretation; the same is true of (3)(c) if Tom is a
novelist at work while Bill is simply writing a letter to his mother.
Moreover, while (4)(a) is marginal (unless the bacon has burst into flames), the
same mixture-of-type interpretation seems relatively natural in (4)(b).

138
RONALD W. LANGACKER
(4) (a) ?The fire is burning, and the bacon is too.
(b) Well, the fire is still burning. Oh my god! The bacon is
too!
Examples like these can be multiplied indefinitely, and suggest the
inadequacy of a fully discrete analysis that posits either a single vague sense or two
distinct senses.
Tuggy further suggests a promising way of handling these graded
phenomena in a usage-based approach employing the network model. It is
assumed, first, that schemas and instantiations may coexist in the grammar
as different facets of a speaker's linguistic knowledge (cf. Fig. 2). Thus, if
the semantic network for uncle contains the schematic node [PARENT'S
BROTHER], this does not preclude its also containing the instantiations
[MOTHER'S BROTHER] and [FATHER'S BROTHER]. Moreover, the
units of a grammar differ in their cognitive salience, which correlates with
the likelihood of their activation. For seemingly clear instances of
vagueness, we can hypothesize that the schema is quite salient relative to its
instantiations; the reverse is postulated for clear-cut instances of ambiguity.
This type of analysis is illustrated in Fig. 3, where the supposed difference
in salience is indicated by the use of heavy-line vs. broken-line boxes.
Because the gender of linking relatives is not a significant factor in English
kinship terminology, it is plausible to assume that [PARENT'S
BROTHER] is more salient in the network for uncle than either of its
instantiations. For those speakers who associate the two senses of ear, a
schema may well be extracted to represent their perceived commonality.
However, this schema must be quite abstract (since the similarity is so
tenuous), and would seem to be of little cognitive or communicative utility, so
its salience is presumably quite low.
(a) -uncle ■^^■^■m fa) ear
(PARENT'S |
BROTHER |
rMOESFs1^ .JfaSSts] I EARi L J m2
j^BPOTFffiRJ L5?2™??J |(body part)| |{ear of corn)]
Figure 3
?art)| |(
Let us assume that the felicity of anaphoric expressions like those in (2)
depends on the same semantic structure being activated in the construal of

A USAGE-BASED MODEL
139
each clause. In the case of uncle, the large disparity in cognitive salience
ensures that [PARENT'S BROTHER] is virtually always activated in
preference to its instantiations; hence the common construal of uncle in the
two clauses renders (2)(a) felicitous. By contrast, the specific senses of ear
are far more likely to be activated than their schema, with the consequence
that the felicity of (2)(b) varies depending on whether the same sense is
activated for both clauses. For relatively clear examples like these, the
network analysis is roughly equivalent to one that simply posits a single sense
for vague expressions and two senses for ambiguous expressions. Its
advantages are more apparent for examples of intermediate status, like those in
(3) and (4). For these we need only assume that the specific and schematic
senses have sufficient cognitive salience to compete for activation. Talk, for
instance, has a specific value pertaining to formal oral presentations,
another relating to informal conversation, and a schematic sense which
abstracts away from the social context and focuses on the verbal activity per
se. The felicity of a sentence like (3)(b) depends on which of these
structures prevails when it is constructed or interpreted on a particular occasion.
If Tom is giving a lecture and Bill is chatting with a friend, (3)(b) can
nevertheless be judged felicitous when one is concerned primarily with the
activity itself, so that the schema is activated for the construal of both
clauses. When, instead, the emphasis lies on the social aspects of the two
events, the specific values are activated in lieu of the schema; this
difference in the construal of talk in the two clauses renders (3)(b) problematic
(its degree of deviance reflecting the cognitive distance between the two
senses). In principle, then, the fluidity of well-formedness judgments for
sentences like (3)-(4) can be explicated in terms of a network whose
elements range along a continuous scale of salience and ease of activation.
Let us conclude this initial discussion with some brief remarks on the
growth and development of networks. There is an intimate connection
between the "outward" growth of a network through extension, on the one
hand, and its "upward" growth through schematization, on the other. The
process of extension occurs because a speaker perceives some similarity
between the basic value (i.e. the local or global prototype) and the
extended value. This similarity-perception represents the commonality of
the basic and extended values, so it constitutes a schema having the two for
instantiations, as depicted in Fig. 4(a). The similarity perception per se
need not be cognitively salient or achieve the status of a unit — it may be
only a fleeting occurrence. Still, a category's outward extension from the

140
RONALD W. LANGACKER
prototype should tend to be accompanied by a certain amount of upward
growth, as schemas are extracted to generalize over a more diverse array of
category members.
(a) [SCHEMA}
[PROTOTYPE]- -^EXTENSION)
In fact, the growth of a network from the category prototype probably
involves a variety of phenomena, as depicted abstractly in Fig. 4(b). Likely
sorts of development include at least the following: (i) extension, with or
without the implied schema achieving unit status (compare the extensions
[A] > [B] and [B] > [G]); (ii) the "downward" articulation of a
category as finer distinctions are made (e.g. the differentiation of [A] into [D],
[E], and [F]); (iii) the extraction and interpolation of subschemas for nodes
already present (note [H]); (iv) the incorporation of additional categorizing
relationships, as direct comparisons happen to be made between indirectly-
associated nodes (observe [C] —> [G]); and (v) adjustments in the
entrenchment and salience of elements, as determined by the vicissitudes of
usage and experience.
General Applicability
The semantic pole of a polysemous lexical item has served thus far as
our prime example of a complex category, i.e. one whose characterization
cannot be reduced to any single structure. I have proposed that a complex
category be described as a network, where nodes with varying degrees of
cognitive salience are connected by categorizing relationships. The network
model subsumes the prototype model as a special case, and further
accommodates taxonomic relationships based on schematization; it offers an

A USAGE-BASED MODEL
141
integrated account of these modes of categorization, and holds considerable
promise of empirical adequacy. I now suggest that the utility of these
notions extends beyond the realm of lexical polysemy: linguistic categories
are in general complex, and networks are required for their proper
description. Clearly, the network model conforms to the maximalist,
non-reductive, and bottom-up spirit of the usage-based approach. In fact, it is by
adopting this model of categorization that cognitive grammar achieves and
implements its usage-based character.
For a non-lexical example, consider the analysis of a phoneme as a
complex category. Let us suppose that the phoneme /a/ (in a particular
language) occurs only in the syllables /a/, /pa/, /ta/, and /ka/. Each preceding
consonant induces some phonetic modification of /a/, however minor it
might be. This phoneme consequently has at least four allophones, namely
[a], [pa], [la], and [ka] (where [pa] is the allophone induced by /p/, and so
on). The allophone [a], which stands alone as a syllable, is plausibly
regarded as the "basic allophone" and equated with the category prototype;
the others then function as context-induced extensions from this prototype,
as diagrammed in Fig. 5. Moreover, speakers may well extract a schema to
represent the commonality of the various allophones. Shown as [xa] in the
diagram, the schema is neutral as to whether and how the basic vocalism of
[a] is modified by a preceding consonant.
Figure 5
The network model therefore reconciles two classic views on the
nature of a phoneme: that which analyzes a phoneme as a set of allophones;
and that which treats it as a unitary but necessarily abstract entity (i.e. a
schema). The non-reductionist character of the analysis also accords with

142
RONALD W. LANGACKER
traditional phonemic descriptions, which provide a list of allophones for
each phoneme, and state the environments that condition each "derived" or
"non-basic'1 allophone. The necessity for a non-reductionist account is
readily apparent in this domain, since a speaker's phonetic ability does not
reside in any single structure. A speaker who fully controls the phonetics of
his language is able to pronounce not only the basic allophone, but also the
full array of derived allophones, properly distributed. Each implies an
articulatory (also an auditory) routine that a speaker masters as part of his
internal representation of the linguistic system. These units are properly
included in the grammar of a language, for they constitute one facet of a
speaker's grasp of linguistic convention.
The inventory of conventional units comprising the grammar of a
language is "structured", in the sense that some units function as components
of others. Often, in fact, a unit owes its specific character to a more
inclusive structure in which it occurs, and is therefore confined to this structure
(at least initially). For instance, the notation of Fig. 5 should not be
interpreted as implying that [pa], [la], and [ka] are free-standing units that can
occur independently; they occur only in the context of the respective
syllabic units [[p][pa]], [[t][ta]], and [[k][ka]], since the preceding consonant
induces their distinguishing phonetic properties. The categorizing
relationship [a] ---> [pa] of Fig. 5 is thus more adequately represented in Fig. 6(a),
which shows the extended variant in the environment that determines and
supports it.
(a) [Tf (b)
[a]
|m!J
0
Figure 6
DOG RUN'
The process of "accommodation", whereby a structure is adjusted to
make it compatible with its context, is obviously a major factor in the
evolution of complex categories. In this maximalist and non-reductive
framework, a variant arising through accommodation is recognized as a
distinct entity, a separate node in the network describing the category,
regardless of how fully or automatically the context determines its specific
properties. Let us suppose, for instance, that run is first learned with reference to

A USAGE-BASED MODEL
143
the canonical, upright, 2-legged locomotion of humans. Hence the semantic
pole of this verb is limited initially to a single value, [RUN] (which serves as
the prototype of the fully-articulated category that eventually develops).
What if the learner now observes the rapid locomotion of a dog? He may
himself decide to use the verb run to describe this activity, or he may hear
someone else describe it in this fashion. In either event, this usage implies
and induces a sense that diverges from the prototype in readily-observable
respects: the actor is canine rather than human, it uses four legs rather than
two, and so on. When predicated of a dog, run thus accommodates to its
subject through the semantic extension [RUN] > [RUN'], as sketched
in Fig. 6(b). The accommodation is easily made, and essentially predictable
from the specifications of the subject, but the extended value [RUN'] is
nonetheless distinct from [RUN] and takes its place in the network
constituting the semantic pole of the lexical item.
In similar fashion, I would posit numerous semantic variants of eat,
accommodated to the nature of the food consumed and the specific activity
required for its consumption. To be sure, I have no way of knowing just
how finely articulated this category is, and it is doubtful that the lowest-
level variants have any substantial measure of cognitive salience. It is
nevertheless a conventional fact of English usage — not something a
speaker must decide anew on each occasion — that eat is employed for such
diverse activities as the consumption of meat, bananas, peanuts, and soup.
There are standard objections to this analysis; they pertain to the
proliferation of meanings, the mixture of pragmatic and semantic considerations,
and the failure to extract a unifying generalization. However, all these
objections reflect the minimalist, reductive bias of contemporary semantic
theory, which is precisely what is at issue, and have little force from the
standpoint of cognitive grammar. In particular, the postulation of specific
variants does not prevent us from expressing the unifying generalization by
means of a coexisting schema, which may very well have greater salience.
Like its semantic pole, the phonological pole of a lexical item is a
complex category revealingly described as a network. Even an expression not
generally thought of as having multiple allomorphs can nevertheless assume
a variety of specific values (some of which may establish themselves as
units) depending on such factors as tempo and prosody. The applicability of
the network model is of course more obvious for expressions that do show
allomorphic variation. Consider the noun leaf, whose phonological pole is
diagrammed in Fig. 7(a). The basic allomorph, [lif], functions as the cate-

144
RONALD W. LANGACKER
gory prototype (it possibly subsumes more specific variants, as just noted).
Since the plural is leaves, we must also recognize the allomorph [liv]
(occurring only in the context of the plural construction), which constitutes an
extension from the basic allomorph; representing the commonality of [lif]
and [liv] is the schema [liF] (where [F] neutralizes the voicing contrast of [f]
and [v] — i.e. it is equivalent to an "archiphoneme"). Though Fig. 7(a)
depicts only the phonological pole, observe that the context for [liv]
requires bipolar characterization: the suffix triggering the phonological
extension must be the plural morpheme [PL/z] in particular, and not the
possessive.
(a)
uf|
IM---J5tvUz1
E3~-fPl
Figure 7
5karf[ iU)skarvl [z]
The voicing of final [f] in the plural is of course not limited to leaf.
Numerous other nouns display the same alternation, e.g. hoof/hooves, life/
lives, elf/elves, knife/knives, scarf/scarves, shelf/shelves, wife/wives, loaf/
loaves. Although the.pattern is not productive (consider///es, reefs, puffs,
etc.), it does constitute a regularity that speakers may incorporate as part of
their cognitive representation of the linguistic system. In the present
framework, regularities are expressed by schemas. It is assumed that any
configuration of structures — even a categorizing relationship — is
potentially subject to schematization, should the proper circumstances arise.
Here the conditions for schematization are indeed met: we find a series of
nouns, all conforming to the schematic characterization [...f], which
undergo parallel modifications in the context of the plural construction.
The phonological network for each of these nouns includes a categorizing
relationship between the basic allomorph ending in [f] and a secondary

A USAGE-BASED MODEL
145
allomorph ending in [v], this latter occurring only in combination with the
plural morpheme; thus, alongside [lif] —» [liv] we find [nayf] >
[nayv], [skarf] ——» [skarv], etc. A schema can be extracted to represent
the commonality of these categorizations, as shown in Fig. 7(b). Though
abstract, this schema is itself a categorizing relationship of phonological
extension, like any of its instantiations. Also like its instantiations, it
contains a specification of the conditioning environment: the extended variant
[...v] occurs only with [PL/z] (again, the semantic pole is not shown).
The schema depicted in Fig. 7(b) is the cognitive-grammar equivalent
of a morphophonemic rule. It expresses a systematic relationship between
the basic form of a noun and the special form it assumes in a particular
morphological context. However, the asymmetry implied by the direction of the
arrow, in either [...f] > [-v] °r its instantiations, is not interpreted
derivationally, i.e. as the relationship between an "underlying" and a
"surface" representation. It is interpreted instead as the asymmetry inherent to
comparison and categorization: [...f] is the standard of comparison
(category prototype), while [...v] is the target of comparison (a peripheral
member assimilated to the category through resemblance to the prototype).
Apart from the type of structures that figure in the categorization, there is
no fundamental difference between [...f] > [...v] (or [lif] — —» [liv],
etc.) and the semantic extension of a lexical item.
The general applicability of the network model is starting to become
apparent. With a limited set of constructs, this model offers a unified
account of many facets of linguistic organization that are normally
approached using very different techniques and descriptive devices. What
distinguishes the various domains of linguistic structure is not the
prevalence of complex categories requiring networks for their description, but
rather the types of structures that function as nodes in these networks. In
the case of a phoneme, these structures are single phonological segments
(allophones and the schemas extracted from them). For a morpheme, the
network at the semantic pole has individual senses as nodes, whereas
allomorphs serve as nodes at the phonological pole. The network model is
also adopted for the description of rules and their conventional
instantiations, in which case the individual nodes of the network have a complex
internal structure. In Fig. 1, which depicts a combinatory rule (one pattern
of plural-noun formation), the nodes include a schema and various
instantiations; each node incorporates two symbolic units as component structures,
together with the composite structure (not separately shown) resulting from

146
RONALD W. LANGACKER
their integration. The morphophonemic rule of Fig. 7(b) is also modeled as
a network; it is a different type of rule because the nodes comprising this
network have a different type of internal structure: rather than a
combinatory relationship between symbolic units, each node consists in a
categorizing relationship between two phonological sequences (allomorphs).
The rule of Fig. 7(b) is considered morphophonemic because it
depends on a morphological context. The phonological extension [...f]
— -> [...v] is limited to the plural-noun construction, which therefore
figures in the characterization of each node. What if a pattern of phonological
extension is not restricted to any particular morphological environment? In
that case, no such environment is specified in the schema or its
instantiations, and we recognize the rule as being purely phonological. For example,
in words like kitten, button, sentence, etc., where [t] precedes the syllabic
nasal [n], the [t] is commonly replaced by a glottal stop. Hence each word
has two phonological variants linked by a categorizing relationship, e.g.
[kitn] — —> [ki^n]. From some array of specific categorizations of this sort,
a schema is extracted to capture the regularity, as shown in Fig. 8. This
particular schema (unlike the previous one) is sufficiently salient or accessible
to be readily invoked for the computation of novel instantiations. When
activated in the context of a specific structure containing [tn], the schematic
extension [...tn...] > [...?n...] yields a variant having [,?n] in its stead;
even nonsense words are likely to be pronounced in this fashion.
[kxtnl Jki^ri
]...tn...
*-
1
l...'n...|
1 '
r
IbAtn
[■-
H
"bA^nl
s£ntns—
s£n9ns
Figure 8
A phonological rule is therefore analyzed in cognitive grammar as a
pattern of phonological extension. Phonological and semantic extension are
viewed as being directly analogous, with any differences between them
attributable to inherent properties of the semantic and phonological
domains. I will pursue their parallelism only to the point of noting that
semantic extension is also subject to schematization and often follows con-

A USAGE-BASED MODEL
147
ventional patterns. One such pattern, illustrated in (5), involves extending
the term for an animal to indicate a person who resembles that animal in
certain respects.
(5) (a) He's a pig.
(b) You're a rat.
(c) That lawyer is a real fox.
Pig, rat, and fox are among the numerous animal names conventionally
employed in this fashion; thus the semantic network associated withp/g, for
example, incorporates the categorizing relationship [PIG] —»
[PIGLIKE PERSON]. Specific relationships of this sort give rise to the schema
[ANIMAL] > [ANIMAL-LIKE PERSON], which is freely used for the
computation of novel instantiations. When I call somebody an ostrich, a
fennec, or even a veritable brontosaurus, I am thereby conforming to the
conventions of English, even if these particular terms have never before
been applied to people.
Distribution
Questions of distribution and productivity figure prominently in
contemporary linguistic theory. Many theoretical constructs have at one time
or another been invoked to deal with these matters (e.g. diacritics, rule
features, major vs. minor rules, disjunctive ordering). Moreover, they are
often cited in support of particular analyses and claims concerning the
organization of linguistic systems. For example, it is not always possible to
predict, on the basis of purely semantic or phonological properties,
precisely which lexical items are eligible to undergo a given rule; this absence
of absolute predictability is commonly taken as establishing the autonomy
of grammar and the need for special, "grammatical" features and classes. It
is also sometimes assumed that full generality is criterial for syntactic rules;
any rule displaying only partial productivity is assigned to some other
component of the grammar (e.g. lexicon or morphology).
Cognitive grammar's approach to these issues reflects its emphasis on
naturalness, conceptual unification, and austerity in the adoption of
theoretical constructs. The highly restrictive "content requirement" allows
only three types of units in the grammar of a language: (i) semantic,
phonological, and symbolic units that occur overtly; (ii) schemas for such
structures; and (iii) categorizing relationships. A separate domain of speci-

148
RONALD W. LANGACKER
fically "grammatical" structure is thereby precluded, as are any descriptive
elements (e.g. features or diacritics) devoid of both semantic and
phonological content. A unified treatment is proposed for lexicon, morphology, and
syntax: all are described by means of symbolic units exclusively; these
traditionally-recognized areas form a gradation of symbolic structures divided
only arbitrarily into distinct "components". To assume that productivity is
coextensive with a particular structural domain, or delimits a coherent, self-
contained body of phenomena, is at best gratuitous.
We must now consider how such a theory is capable of handling the
problems of variable productivity and non-predictable distribution.
Granted that certain rules are applicable only to a limited class of
structures, and granted further that the membership in this class is not always
predictable on semantic or phonological grounds, how does one specify the
proper restrictions? As we will see, the answer lies in the usage-based
character of the framework, together with the network model of complex
categories.
Linguistic expressions are symbolic, each residing in the relationship
between a semantic and a phonological "pole". Grammar consists of
patterns for combining simpler symbolic expressions to form progressively
larger ones. In cognitive grammar, these patterns (i.e. grammatical rules)
are analyzed as schematized expressions — they are themselves complex
symbolic structures, parallel in formation to the expressions they
schematize, but characterized at a level of abstraction that neutralizes the
differences among them. These combinatory patterns are equivalent to
grammatical constructions, so we can refer to the schemas describing them
as "constructional schemas". Each specifies the way in which two or more
"component" structures are integrated, at the semantic and phonological
poles, to form a bipolar "composite" structure. Constructional schemas
capture generalizations, and serve as templates for the assembly of novel
expressions.
I have already argued that constructional schemas coexist in the
grammar of a language with those of their instantiations that have the status of
units (cf. Fig. 1). These instantiations need not be limited to specific
expressions — we can also posit subschemas at various levels of abstraction,
corresponding to subpatterns discernible in the data. In fact, these structures
form a network, as they are linked to one another through relationships of
schematicity, and possibly through other types of categorization as well. A
grammatical construction can therefore be regarded as a complex category:

A USAGE-BASED MODEL
149
it does not reside in any single structure, but rather in a family of structures
connected by categorizing relationships. Internally, each node of this
network is quite complex, comprising an entire constructional schema or
subschema, or else a specific composite expression.
These notions are illustrated in Fig. 9, which sketches the network
corresponding to one pattern of past-tense marking in English. The maximal
generalization is captured by the topmost schema, which says, in effect,
that a verb stem containing [1] may form its past tense by ablauting this
vowel to [<e]. In this abbreviatory notation, only the two component
structures are explicitly indicated: [PROCESS/... 1...] is the verb stem, and
[PAST/...1... — -> ...ae...] is the appropriate allomorph of the past-tense
morpheme; a more complete representation would also show the nature of
their integration at the two poles, and the composite structure that results
(in this case [PROCESS-PAST/...ae...]). The two immediate instantiations
of this schema are subschemas representing special cases of the general
pattern, that in which the stem vowel [1] is followed by [t], and that in which it
is followed by a nasal ([N]); note that the component structures
[PROCESS/... it] and [PROCESS/...iN] elaborate the stem of the higher-level
schema (the added detail is also reflected in the past-tense morpheme,
which makes internal reference to the stem in specifying the ablaut
pattern). Instantiating the first of these subschemas are the specific verb forms
sat and spat, complex structures that I have rendered orthographically to
simplify the diagram. The second subschema is instantiated by the specific
verbs swam and began, and also by a lower-level subschema which
identifies the nasal consonant of the stem as the velar nasal in particular. This
subpattern in turn has two special cases (each with a number of
instantiating expressions) that differ in whether the nasal is final or followed by [k].
The analysis is complicated, but I would argue for its cognitive and
linguistic plausibility. It is certainly reasonable to suppose that forms like sat,
swam, began, rang, sank, etc. are learned by speakers as familiar units
(some, of course, are more frequent and deeply entrenched than others).
All significant generalizations are captured, both the global generalization
expressed by the topmost schema, and also the more limited generalizations
reflecting the prevalence of certain types of stems as participants in this
morphological construction. Whether speakers extract all available
generalizations is an open question, but there is no particular reason to
suppose that they proceed directly from specific forms to the highest-level
schemas supported by the data, or that a category, should it develop in this

150
RONALD W. LANGACKER
(PROCESS
PAST [
II PROCESS
1 ...It
PAST 1
.. .It--*-.. .aet]
sat
spat
A
I PROCESS
[...IN...
PAST I1
...X^.-j-^-^
.aeN...[
j PROCESS
l"-*9-"l
PAST 1
.. .Itj. .♦--»-.. .aeij.. .1
swam
"began
1 PROCESS
1 -:-*3 ,
PAST ||
...ip--^. ..aer)|
PROCESS
.irfa
PAST
.xrjk--*-.. .aerjk
Figure 9
manner, necessarily fails to undergo subsequent differentiation through the
emergence of subschemas.
A network like Fig. 9 brings out clearly the maximalist, non-reductive,
and bottom-up nature of a usage-based approach. Revealing its maximalist
character is the incorporation of structures representing all levels of
generality, from specific forms to an all-subsuming schema. The analysis is non-
reductive because it posits as units both schemas and specific expressions
computable from those schemas. Its bottom-up orientation is reflected in
the emphasis on low-level structures that provide the basis for higher-level
schematization. By contrast, linguists are more accustomed to a minimalist,
reductionist, top-down approach that achieves greater economy by
including in the grammar only a rule equivalent to the topmost schema.
Either type of approach must somehow specify precisely which verbs
participate in this construction, since most stems containing [i] form their
past tense in some other manner. A minimalist analysis would typically
mark the proper verb stems with some kind of diacritic (e.g. a rule feature,
or an indication of class membership), and condition the application of the
ablaut rule on the presence of this marking. Because the diacritic has no

A USAGE-BASED MODEL
151
intrinsic semantic or phonological content, it is considered a "grammatical"
construct; and because the set of verbs in question is not semantically or
phonologically predictable, it is taken as constituting a grammatical
(specifically, a morphological) class. This apparent need to posit purely
grammatical entities is then invoked as an argument for the autonomy of
grammatical structure.
Diacritics are rendered unnecessary by the maximalist analysis of Fig.
9. For example, the information that sing follows the [...1... > ...ae...]
ablaut pattern is provided directly, by (i) the inclusion of sang among the
conventional units of the grammar, and (ii) its categorization by the
schemas that define the pattern. Observe that the analysis obeys the
content requirement, as it employs only symbolic units (both specific and
schematic) and categorizing relationships. Though we can perfectly well
speak of a grammatical (or morphological) construction, it is fully
characterized in terms of symbolic relations between semantic and phonological
structures — there is nothing that represents a separate domain,
autonomous vis-a-vis semantics, of purely "grammatical" structure. Nor is the
existence of an independent set of grammatical entities established by the
impossibility of predicting the participating verb stems on semantic or
phonological grounds. To assume so is to embrace the "type/predictability
fallacy", i.e. the failure to distinguish between two issues that are in
principle distinct: the types of structures that occur, and the predictability of their
behavior.
We can now observe that the apparent argument cited above for the
autonomy of grammar is in reality simply an artifact of the minimalist
approach to grammatical description. If specific forms that instantiate a
pattern are excluded from the grammar by the thesis of reductionism, they
must be "constructed" by application of the relevant rule. If the lexical
items that undergo this rule are limited to an arbitrary class, they must
somehow be marked to undergo it. And since the requisite marking has no
intrinsic semantic or phonological content, it must represent some other
realm or dimension of linguistic organization. We have seen, however, that
symbolic units are capable of furnishing the requisite distributional
information provided that we take a maximalist, non-reductionist approach
allowing schemas and instantiations to coexist in the grammar of a
language.
In this usage-based framework, grammatical constructions are
analyzed as complex categories. A speaker's conventional knowledge of a

152
RONALD W. LANGACKER
construction is not limited to a single, abstract rule or schema standing in
isolation — it further embraces his knowledge of how the rule is
"implemented" with respect to more specific structures. By its very nature, a
high-level schema is compatible with a broad and structurally-diverse array
of potential instantiations. Typically, however, conventional usage carves
out for exploitation only limited regions within this field of structural
possibilities. A full linguistic description must identify these regions, i.e. it
must characterize the construction's conventional domain, as articulated by
a hierarchy of lower-level structures. Providing this information are
subschemas and expressions with the status of units: they specify the actual
array of subcases and specific instances that support and give rise to the
high-level generalization.
In the network for a grammatical construction, the individual nodes
and categorizing relationships presumably differ in their cognitive salience
and likelihood of activation, as they do in any complex category. There is
no reason to assume that the highest-level schema is necessarily the most
salient, or even that an all-subsuming schema is always extracted (indeed,
the abstractness of a high-level schema is probably inimical to its
prominence). For the network of Fig. 9, we can plausibly suppose that the
highest-level schema is less readily activated (has a lesser degree of prominence)
than specific forms like sat, swam, began, rang, sank, etc.; this would imply
that such forms are generally accessed as units (as opposed to being
computed from the stems by means of the schema). The opposite is no doubt
true for other constructions: greater salience attaches to schemas (though
not necessarily those at the highest levels), and relatively few instantiating
expressions have the status of units; computation must therefore
predominate. At least in principle, it is possible for behavioral evidence to be
brought to bear on such matters — cognitive salience and accessibility are
neither inherently mysterious nor beyond the reach of empirical inquiry.
This notion of accessibility is crucial to a usage-based account of
distribution and productivity. The general problem can be formulated as
follows: granted that a construction is a complex category, and properly
represented as a network, how is this network invoked for the assembly (or
evaluation) of a particular instantiating expression? We cannot assume that
access to the network is random, or that all nodes are simultaneously and
equally activated for this purpose — the resulting chaos would afford no
basis for clear judgments of well-formedness or the assignment of structural
descriptions. Instead, I suggest a working hypothesis that is basically com-

A USAGE-BASED MODEL
153
patible with a "connectionist" or "interactive-activation*' model of cognitive
processing (cf. Elman and McClelland 1984; Rumelhart and Zipser 1985;
Waltz and Pollack 1985). This two-part proposal is formulated in (6).
(6) (a) Uniqueness:
When an expression is assessed relative to a
grammatical construction, a single node (from the network
representing the construction) is activated for its
categorization; if this "active node" is schematic for the
expression, the latter is judged well-formed (conventional).
(b) Selection:
The likelihood that a given node will be chosen as the
active node for categorizing a target expression
correlates positively with its degree of entrenchment and
cognitive salience, and negatively with its "distance"
from the target (i.e. how far the target diverges from it
by elaboration or extension).
The thrust of (6)(a) is that an expression's well-formedness depends on how
it is structurally construed (i.e. what it is taken as being an instance of), and
that a single "episode" of categorization (structural description) construes it
in a particular way. Some factors that influence the choice of categorizing
structure (active node) are suggested in (6)(b); they are matters of degree,
and possibly antagonistic.
Let us consider these matters with reference to Fig. 10, which
represents the overall past-tense verb construction of English. At the semantic
pole, both the topmost schema and all the subschemas specify the
integration of [PROCESS] and [PAST] to form the composite structure
[PROCESS-PAST] (not individually shown). However, at the phonological pole
the topmost schema is so abstract that it is almost contentless; essentially, it
characterizes the stem and inflection only as having "some phonological
value". The reason for this extreme schematicity is that the various patterns
of past-tense formation have very little in common phonologically. Schemas
corresponding to four of these patterns are shown in the diagram. At the
left is the subschema for the ablaut pattern [...i... > ...ae...]; this is the
same structure that functions as the highest-level node in Fig. 9 (in this way
the entire network of Fig. 9 fits into the more inclusive network of Fig. 10).
The second subschema corresponds to the "regular" pattern of past-tense
formation. At the phonological pole of the past-tense morpheme, [-D]

154
RONALD W. LANGACKER
stands for a schematized suffix having [-d], [-t], and [-3d] as instantiations
(i.e. it specifies an alveolar stop and is neutral as to the presence of a
preceding schwa). The three subcases of the regular pattern are represented by
lower-level subschemas, which incorporate specific suffixes and specify the
phonological characteristics of the stems they attach to; elaborating each
low-level subschema is some array of instantiating expressions having unit
status. The third and fourth major subschemas describe respectively the
ablaut pattern of brought, caught, fought, sought, taught, etc., and the zero
pattern of verbs like cut, hit, slit, bet, spread, and bid.
Also represented in Fig. 10 are rough hypotheses concerning the
cognitive salience of the individual nodes. Among the schemas, the structures
corresponding to the regular pattern are attributed the greatest
prominence, as indicated by the heavy-line boxes; they can be regarded as the
category prototype. Most of the specific expressions learned as units
presumably have substantial salience as well. The schemas describing minor
patterns are analyzed as having a lesser degree of prominence, while the
topmost schema — considering the vacuity of its phonological pole — may
well be the least prominent of all. Such differences in salience (likelihood of
activation) are the device employed in this framework to implement the
distinction between productive and non-productive patterns (or between
major and minor rules). Since prominence is a relative matter and varies
continuously, we should expect in general to find a gradation between the
two types instead of a strict dichotomy.
The contrast may nevertheless be quite pronounced in particular
instances; a pattern that is distinctly more prominent than any potential
competitor will almost invariably be selected for the construction and
evaluation of novel expressions. The "regular" pattern of English past-tense
formation is so identified precisely because it has this type of advantage.
Consider a speaker who needs the past-tense form of flit, but happens not to
have learned it as a unit. He must therefore select one of the schemas in the
network for the past-tense construction to employ as the active node (basis
for categorization) in assessing possible alternatives. Though all four
patterns in Fig. 10 are potentially applicable, he will almost certainly choose
the regular pattern by virtue of its distinctive prominence. Within the
regular pattern, moreover, the subschema [[PROCESS/...T]-[PAST/-9d]] will
be chosen in preference to [[PROCESS/...C]-[PAST/-t]] on the basis of
distance: the former characterizes the stem with greater specificity (alveolar
stop vs. voiceless consonant) and hence is ''closer" to the stem (namely flit)

jlffiOCESS
i| ...
PAST
• ••_]
1 PROCESS
.. .1. . .
PAST I
... I.. . - -»-. . . ae. . . 1
1 PROCESS
1 ...T
past n1
...T--—. ..t|i
ll PROCESS
|| ...T
PASTll
-?a ||
>
G
>
O
m
03
>
m
D
o
D
Figure 10

156
RONALD W. LANGACKER
of potential target expressions. With [[PROCESS/...T]-[PAST/-sd]]
selected as the active node, the target flitted is judged well-formed; it is fully
compatible with the active node's schematic specifications. Other target
expressions, e.g. flat, flaught, and flit, conflict with these specifications and
are consequently judged ill-formed, despite being computable from other
schemas. Given the selection of [[PROCESS/...T]-[PAST/-ad]] as the
active node, their deviance is predicted by (6)(a), the uniqueness
hypothesis.
The situation is quite different for specific expressions that are
mastered as units, e.g. sat, taught, and hit. As units, they are themselves part of
the network representing conventional knowledge of the past-tense
construction, making them eligible for selection as the active node representing
this construction. Suppose, for instance, that a speaker wishes to express
the past tense of hit. If the past-tense expression hit has any substantial
cognitive salience (which it must, as a frequently-occurring form), the distance
factor virtually assures its being chosen as active node: its distance from the
target (i.e. the desired past-tense form of hit) is essentially zero. Only hit
itself is compatible with this categorizing unit — alternative expressions
such as hitted, haught, and hat are deviant as the past-tense of hit, though
computable via established schemas. In short, specific expressions with the
status of conventional units sanction themselves (or occurrences of
themselves) as being conventional, and thereby preempt the process of selecting
an active node.
Though the selection process is described only vaguely at best in
(6)(b), its general import can now be appreciated. The well-formedness
(conventionality) of an expression is not absolute: it depends on what unit
the expression is construed as instantiating, and on its compatibility with
that unit (as evaluated by a categorizing judgment). The selection of a
categorizing unit hinges on the dynamic interplay of factors that are
frequently antagonistic. If specificity is held constant, the determining factor is
entrenchment or cognitive salience; given a range of options, one of them
may establish itself as the "regular" choice (the default-case option).
Holding salience constant, the determining factor is cognitive "distance", i.e. the
extent to which the categorizing unit (active node) must be elaborated or
extended to "reach" the target. The tendency for specific structures to
prevail over patterns of greater generality has been noted in a variety of
theoretical frameworks; examples include disjunctive rule ordering
(Chomsky and Halle 1968), proper-inclusion precedence (Sanders 1974),

A USAGE-BASED MODEL
157
and Hudson's priority-to-the-instance principle (1984). In the present
framework, this tendency is naturally accommodated as one facet of a
broader theoretical perspective emphasizing a non-reductive, bottom-up,
usage-based account of language structure. Adopting this general
orientation enables one to countenance with equanimity a far-reaching implication
of (6)(b): for the most part, specific structures and low-level schemas are
more significant than high-level schemas expressing the most inclusive
generalizations. Even certain linguists who would resist this implication
have, in formulating the precedence principles cited above, at least partially
acknowledged its empirical force.
This conception gives rise to certain empirical predictions. One
pertains to "irregular" forms, i.e. those that are idiosyncratic or follow a minor
pattern: the more salient and deeply entrenched they are (as reflected in
frequency of occurrence), the more resistant they should be to regulariza-
tion. This correlation is in fact so firmly established that documentation
would be otiose — let us focus instead on the basis for predicting it.
Consider the past participle of drive. The regular pattern predicts drived, which
a speaker immediately recognizes as being incorrect. The proper form,
driven, exemplifies a minor pattern of little cognitive salience, but the form
itself occurs frequently and constitutes a well-entrenched, easily accessible
unit of English. If we attribute comparable salience to the unit driven and
the schema describing the regular pattern of past-participle formation, the
far greater specificity of the former determines its selection as active node
whenever the past participle of drive is required. Driven thus sanctions
itself as the correct expression, but drived is judged ill-formed (despite its
regularity) when measured against this standard. By contrast, the verb
thrive is relatively infrequent, and a typical speaker hardly ever has
occasion to use its past-participial form. Those who use thriven must know it as
a unit, but its status as such is only marginal; its rarity ensures its lack of
prominence. Consequently, neither this unit nor the schema representing
the regular pattern has an overwhelming advantage in the competition for
selection as active node: the former is more specific, but the latter is far
more salient. Hence the schema may well be selected, and if so, an
occurrence of thrived will slip by unnoticed and be accepted as well-formed
(which is quite unlikely in the case of drived).
A second prediction is that lower-level schemas should predominate in
the computation of novel expressions. This prediction is based on the
distance factor: because they are "closer" than high-level generalizations to

158
RONALD W. LANGACKER
the target expression, lower-level schemas should in general be selected as
the basis for computation (active node). For illustration, let us return to the
past-tense schemas of Fig. 9, and consider the relative likelihood of
innovative past-tense forms involving the ablaut of [i] to [ae]. Though brought is
well-entrenched as the past tense of bring, the sporadic occurrence of brang
is at least conceivable (it is attested dialectally and in child language). We
can explain this with reference to the low-level schema [[PROCESS/...ig]-
[PAST/...IQ ——> ...aerj]] of Fig. 9, extracted to represent the commonality
of rang, sang, and sprang; not nearly so salient as the schemas for the
regular pattern, it is nevertheless quite specific concerning the phonological
shape of the verb stem, and thus stands a decent chance of being activated
to compute the past-tense form of another stem meeting its specifications.
Similarly, shat is sometimes encountered as a jocular or euphemistic past-
tense form, and is computed from the low-level schema that expresses the
commonality of sat and spat. On the other hand, the innovative past-tense
forms san (for sin) and hass (for hiss) are unattested, uninterpretable, and
inconceivable. Observe that the lowest-level schemas available for the
computation of these forms are in fact quite abstract. San (if it occurred) would
invoke the schema [[PROCESS/...iN...]-[PAST/...iN... > ...aeN...]],
which identifies the consonant following the ablauted vowel only as a nasal
(not a specific segment), and is further neutral as to whether or not this
consonant is stem-final. Hass would be computable only from the topmost
schema in Fig. 9, which is maximally schematic concerning the environment
of the ablauted vowel.
To make the same point in another way, consider the topmost schema
in Fig. 10, which represents the maximal generalization concerning past-
tense verbs in English. We have already noted that this schema is
essentially vacuous at the phonological pole, because the various patterns of
past-tense formation have virtually nothing in common. If this high-level
schema were the one to be invoked for the computation or evaluation of
novel expressions, it would have the effect of allowing any kind of stem to
be marked for past tense in any way whatever — a maximally schematic
characterization of the construction imposes no significant constraints on its
possible instantiations. The actual work is obviously done by lower-level
schemas.
A final example should reinforce and further clarify this notion.
Postpositional endings in Classical Nahuatl vary in the type of noun stems to
which they attach. Certain postpositions, among them -ko 'in' and -tew

A USAGE-BASED MODEL
159
'like', attach only to lexical nouns (e.g. siwaa-tew 'like a woman1). Others,
including -waan 'with' and -pampa 'because of, occur only with pronouns
(e.g. no-waan 'with me'). A third class of postpositions, exemplified by
-pan 'on' and -caalan 'among', suffix to nouns of either sort (e.g. to-caalan
'among us', kwaw-caalan 'among the trees'). Each of these distributional
possibilities implies a low-level constructional subschema, depicted in
abbreviated form along the bottom row in Fig. 11. For instance, the box at
the lower left in this diagram stands for the constructional subschema
expressing the generalization that -ko occurs on lexical nouns; the box at
the lower right similarly expresses the generalization that -pampa combines
with pronouns. Note that -pan and -caalan figure in two such subschemas
each, one for each type of noun.
N -caalan N -vaan N -parcpa
pro pro pro * *^
Figure 11
From these low-level schemas, certain broader generalizations can be
extracted. Since -ko, -tew, -pan, -caalan, and others occur on lexical nouns,
a speaker could extract the intermediate-level subschema shown on the left
in Fig. 11; it specifies the existence of a compositional pattern whereby
postpositions suffix to lexical nouns. Similarly, the intermediate-level
subschema on the right in Fig. 11, specifying the attachment of postpositions to
pronouns, is supported by the occurrence on pronouns of -pan, -caalan,
-waan, -pampa, etc. Moreover, since postpositions occur on both lexical
nouns and pronouns, there are grounds for extracting the higher-level
schema depicted at the top, which specifies that postpositions suffix to
nouns. Each schema expresses the commonality of its immediate
instantiations. It is clear, however, that only the lowest-level schemas could be
invoked for the computation of novel expressions — it is only at this level
that the distributional restrictions are apparent. If the topmost schema were

160
RONALD W. LANGACKER
activated for this purpose, it would sanction the occurrence of any
postposition with either type of noun. The intermediate-level schemas would fare
no better, for they specify the occurrence of any postposition with either a
lexical noun or a pronoun.
Conclusion
Accepting the general principles of scientific inquiry does not itself
resolve the more specific but nonetheless crucial issue of how these
principles are appropriately applied to the problems of a particular discipline at a
given stage of its development. I have argued, both on methodological and
on empirical grounds, that the principle of generality has received in
linguistics a commonly accepted interpretation that is in fact not appropriate
to its subject matter. Current doctrine favors a minimalist account of
linguistic knowledge, described in accordance with a complex array of
theoretical apparatus featuring specialized devices for the various "components" of
the linguistic system. By contrast, cognitive grammar pursues a maximalist
account of linguistic knowledge, and tends toward austerity in the adoption
of theoretical constructs; it seeks a unified treatment of the various facets of
linguistic structure, attributing their differences to the content of the
domains in question rather than the basic constructs invoked to handle
them.
Prominent among these constructs are those comprised by the network
model of complex categories. The network model affords an integrated
account of categorization for the varied domains of linguistic structure. It
accommodates not only those phenomena generally thought of as involving
categorization, but also the nearest cognitive-grammar analogs of rules,
derivations, and structural descriptions. It is further responsible for the
usage-based character of the framework; by tolerating the coexistence in a
single network of specific expressions and schemas at varying levels of
abstraction, it implements the maximalist, non-reductive, bottom-up
orientation of the usage-based approach.
Problems of distribution and productivity are addressed by treating
grammatical constructions as complex categories. A single, high-level
generalization does not exhaust a speaker's conventional knowledge of a
construction. A full description must also specify how this generalization is
articulated through the supporting hierarchy of subpatterns and specific
expressions. In this maximalist account, the structures that occur in a given

A USAGE-BASED MODEL
161
construction are identified without the use of diacritics or other arbitrary
devices. Moreover, specific expressions and low-level schemas are seen to
be at least as important as higher-level schemas capturing the broadest
generalizations. We have no assurance that speakers invariably arrive at
high-level schemas, whose abstractness may render them essentially useless
for the computation and evaluation of novel expressions; there is reason to
think that lower-level schemas figure more prominently in this role. A
major advantage of the usage-based conception is its ability to
accommodate structures at this level of organization without the loss of valid
generalizations.

PART II
ASPECTS OF A
MULTIFACETED RESEARCH PROGRAM

The Relation of Grammar to Cognition1
Leonard Talmy
University of California, Berkeley
0. Introduction
A fundamental design feature of language is that it has two subsystems
which can be designated as the grammatical and the lexical (as these are
characterized below). Why is there this universal bifurcation when, in
principle, a language could be conceived having only a single system, the
lexical? The explanation in this paper is that the two subsystems have distinct
semantic functions, ones that are indispensable and complementary. To
develop this account further, we must first note that we take a sentence (or
other portion of discourse) to evoke in the listener a particular kind of
experiential complex, here to be termed a "cognitive representation" or
"CR".2 Now, the grammatical and lexical subsystems in a sentence seem
generally to specify different portions of a CR. Together, the grammatical
elements of a sentence determine the majority of the structure of the CR,
while the lexical elements together contribute the majority of its content.
Lexical elements do incorporate some of the same structural indications
that grammatical elements express, but when the two are in association or
in conflict within a sentence, it is generally always the grammatical
elements' specifications of structure that are determinative.3 The grammatical
specifications in a sentence, thus, provide a conceptual framework or,
imagistically, a skeletal structure or scaffolding, for the conceptual material
that is lexically specified.
More generally, across the spectrum of languages, the grammatical
elements that are encountered, taken together, specify a crucial set of
concepts. This set is highly restricted: only certain concepts appear in it, and
® Leonard Talmy

166
LEONARD TALMY
not others, as seen below. The purport of the present paper is that this set
of grammatically specified notions collectively constitutes the fundamental
conceptual structuring system of language. That is, this cross-linguistically
select set of grammatically specified concepts provides the basic schematic
framework for conceptual organization within the cognitive domain of
language.
Thus, grammar, broadly conceived, is the determinant of conceptual
structure within one cognitive domain, language, and as such is the main
object of this paper's study. But such a study directly opens out into a
broader investigation across other cognitive domains, such as visual
perception and reasoning, as discussed at the end of the paper. That is, the greater
issue, toward which the present study ultimately aims, is the general
character of conceptual structure in human cognition.
The present investigation into the semantics of grammar is of a scope
that follows in a progression from previous types of study. These have
mostly been either an in-depth semantic analysis of a selected grammatical
element (or class of elements) of particular interest within a language, e.g.,
the Turkish evidential suffix -mi$ (Slobin & Aksu 1982); or an exposition
of the meanings and functions of all the grammatical elements of a single
language, say, as in a grammar of Dyirbal (Dixon 1972); or a
cross-linguistic typology of the different kinds of grammatical devices used for a single
semantic function, say, to indicate the interrogative (Ultan 1978). Some
previous work has also treated broader issues of grammatical meaning
(Sapir 1921, Boas 1938, Whorf 1956, Jakobson 1971). But the present study
is perhaps the first to address grammatical expression in language at the
superordinate level, with the aim of determining the semantic and cognitive
properties and functions of this structural component of language as a
whole.4
The terms "grammatical" and "lexical" as employed here require some
immediate elaboration. The distinction between the two is made formally
— i.e., without reference to meaning — on the basis of the traditional
linguistic distinction between "open-class" and "closed-class". A class of
morphemes is considered open if it is quite large and readily augmentable
relative to other classes. A class is considered closed if it is relatively small and
fixed in membership. We can identify the particular classes belonging to
these two types. The open classes of elements — i.e., the lexical classes —
are the roots of nouns, verbs, and adjectives.5 Everything else is closed-
class — and is here considered to be, quite generally, "grammatical".
Among the overt elements of this type are such bound forms as inflections

THE RELATION OF GRAMMAR TO COGNITION 167
and derivations, such free forms as determiners, prepositions,
conjunctions, and particles, and perhaps also such suprasegmental forms as
intonation patterns. Included among abstract, or implicit, closed-class forms are
grammatical categories and grammatical relations, word order, and perhaps
also paradigms and "zero" forms. Additionally here are regular
combinations of simpler closed-class forms, tending to have a unified or integrated
semantic function — what are below called "grammatical complexes",
including grammatical constructions and syntactic structures.6
The issues presented in this introduction are treated below in three
sections. Section 1 examines the notions specified by a heuristic sampling of
grammatical elements, outlines the kinds of constraints on such notions,
proposes a property held in common by such notions but largely absent
from excluded notions, and contrasts such grammatically specified notions
with ones that are lexically specified. Section 2 presents a number of
categories in which grammatically specified notions are seen to pattern, as
well as broader conceptual systems in which these categories in turn
participate, ending with the identification of four comprehensive "imaging
systems". This section, further, examines the interactions of grammatical
specifications with lexical specifications within categories and the nesting of
such interactions across categories, as well as the cognitive processes that
accompany these interactions. And Section 3 presents an explanation of the
function of grammatical specification, as well as possibilities of its relations
to other cognitive systems.
1. The Nature of Grammatically Specified Notions
In this section we examine a small sampling of grammatical forms for
the particular component notions that they specify. The sample will give a
heuristic indication of the kinds of notions that get grammatically specified
as well as of the kinds of notions that possibly never do. By contrast, it will
be seen that the excluded kinds can be readily specified by lexical elements.
A particular property will be seen to run through most of the grammatical
notions. To indicate this property at the outset, it is preponderantly the
case that grammatical specifications of structure are relativistic or
topologylike, and exclude the absolute or the metrically Euclidean. Finally, a
systematic difference is shown between the characteristics of grammatically
specified notions and of lexically specified ones.
We begin with a simple demonstration that the concepts specified by
grammatical forms are constrained in two ways: as to their categories and as

168
LEONARD TALMY
to the membership of these categories. Many languages have inflections on
the noun that specify the "number" of the object referred to by the noun,
for example its 'singularity* or 'plurality', like the English -0 and -s. By
contrast, no languages appear to have inflections that specify the "color" of the
object referred to by a noun, e.g., its 'redness' or 'blueness'. Here, single
quotes enclose "notions", while double quotes enclose categories of
notions. The "number" category can be specified grammatically and in that
form is readily seen to play a structuring role in a CR.7 The "color"
category is perhaps never found specified by grammatical elements, though it is
readily found specified by lexical elements, e.g., English red and blue.
Further, though, even within a conceptual category acceptable for
grammatical expression, there are great constraints on the particular notions that
can be specified. Thus, "number" notions that are expressed grammatically
include little more than 'singular', 'dual', 'trial', 'plural', and 'paucal'. They
apparently never include, say, 'even', 'odd', 'dozen', or 'numerable',
whereas such notions, again, can be specified lexically, as shown by the
words just used.
Given such constraints on grammatically specifiable notions, we can
seek properties that hold in common for included notions but need not
apply to excluded notions. In this regard, consider a deictic like the English
this or that as in This/That chair is broken. A closed-class element of this
type specifies the location of an indicated object as being, in effect, on the
speaker-side or the non-speaker-side of a conceptual partition drawn
through space (or time or other qualitative dimension). This integral
specification can be analyzed as containing the component notions enclosed
by quotes in (1):
(1) (a,b) a 'partition' that divides a space into 'regions'/'sides'
(c-e) the 'locatedness' (a particular relation) of a 'point' (or
object idealizable as a point) 'within' a region
(f,g) (a side that is the) 'same as' or 'different from'
(h,i) a 'currently indicated' object and a 'currently
communicating' entity
Notions that might at first be ascribed to such deictics, such as of distance
or perhaps size, prove not to apply, on the evidence of sentence-pairs like
(2):
(2) a. This speck is smaller than that speck,
b. This planet is smaller than that planet.

THE RELATION OF GRAMMAR TO COGNITION 169
The scenes referred to by (2a) and (b) differ greatly, involving tiny objects
millimeters apart or huge objects parsecs apart. But the sentences differ
only lexically, not grammatically. Hence, the scenes' differences as to the
magnitude of size or distance must arise from the lexical elements, they
cannot be traced to the deictics (or other grammatical elements) in the
sentences. Thus, the notions specified by a this or a that are abstracted away
from any particularities of magnitude and so, to this extent, are genuinely
topological. Their specification of a conceptual partition remains constant,
but this partition's distance can — by the characterization of topology as
"rubber-sheet geometry" — be "stretched" indefinitely without challenge
to any semantic constraints of the deictics. This finding about deictics alerts
us to noticing whether any grammatical elements make specifications about
magnitude. A spot check through English and various other languages
suggests that — while there are grammatical specifications for relative
magnitude8 — there are possibly never any for absolute or quantified
magnitude, whether of size, distance, or other parameters. We can
provisionally conclude that the referents of grammatical elements have the
topological property of being "magnitude-neutral".
For another case, consider the type of adposition that specifies, for a
moving object, certain characteristics of path and of reference-point or
-frame. An example of this type is English through as used, for instance, in
/ walked through the woods. In this usage, through specifies, broadly,
'motion along a line that is within a medium'. The component notions
contained here include those in (3):
(3) (a) 'motion'
(b-e) which can be understood as 'one-to-one correspondences'
between 'adjacent' points of 'space' and adjacent points of
'time'
(f) motion that describes a 'line' (i.e., a 'linear extent')
(g) the locatedness of a line within a 'medium'
(h,i) a medium, i.e., a region of three-dimensional space set
apart by the locatedness within it of 'material' that is in a
'pattern of distribution' with properties and a range of
variation still to be determined
It can be first observed, from a sentence-pair like (4), that the concept
specified by through is indifferent to particulars of shape or contour in the
linear path described by the moving object. This is evident here because, as
before, the two sentences differ only lexically, not grammatically — they

170
LEONARD TALMY
both use through while referring to different path contours. Another cross-
linguistic spot check of closed-class elements suggests that they largely have
this further topological property of being "shape-neutral".
(4) a. I zig-zagged through the woods,
b. I circled through the woods.
With a sentence pair like (5), it can be further determined that the 'rate' of
motion is not specified by through, a finding that also appears quite general
among grammatical elements. And (6) shows that through, again like
grammatical elements generally, excludes specification of the 'kind of material'
involved — here, comprising the "medium" — and of the 'sensorimotor
characteristics' attendant on executing the action involved — as, here,
those attendant on wading in liquid vs. weaving amidst obstacles. Thus, it
can be further held that grammatical elements are generally rate-neutral,
material-neutral, and sense/motor-neutral.
(5) (a/b) I crept / dashed through the woods.
(6) (a/b) I walked through the water / woods.
In the aim of ascertaining any properties common to grammatically
specified notions, the notions examined above are gathered together in (7).
For heuristic purposes, the notions are provisionally divided into two
groups on the basis of their relation to topology. In group (a) are combined
the notions that properly belong to the specific mathematical system of
topology and, with them, the intuitively comparable notions that might
belong to a language-based system of topology — one that perhaps could
serve as the model for the construction of a new topology-like mathematical
system.9 In group (b) are the notions that fall outside any usual conception
of topological properties. The number of notions in the first group is
fourteen, while the second has six — an indication of a preponderant propensity
for grammatical elements to specify quasi-topological notions. The ratio in
this direction is in fact improved if we consider that even several notions in
group (b) — the bottom three — resemble topological ones in the sense of
involving relativistic relationships between quantities rather than absolutely
fixed quantities.

THE RELATION OF GRAMMAR TO COGNITION
171
(7) some notions found to be specified by grammatical elements
a. topological
point
linear extent
locatedness
within
region
side
partition
or topology-like
singularity
plurality
same
different
"adjacency'
' of points
one-to-one correspondence
pattern of distribution
b. non-top ological
material
space
time
motion
medium
entity currently
indicated/communicating
In the complementary aim of ascertaining any properties excluded
from grammatical specification, the categories of notions found above not
to be specified by the elements investigated are listed in (8). Rather than
topological, topology-like, or relativistic, these notions involve Euclidean-
geometric concepts — e.g., fixed distance, size, contour, and angle — as
well as quantified measure, and various particularities of a quantity: in sum,
characteristics that are absolute or fixed.
(8) some categories of notions seemingly rarely or never
specified by grammatical elements
absolute/quantified magnitude kind of material
(of distance, size, etc.) sensorimotor characteristics
shape/contour of line color
rate
The provisional conclusion to be drawn from these findings is that, if
grammatical specifications generally correspond to (linguistic-) cognitive
structuring, then the nature of that structure is largely relativistic or topological
rather than absolute or Euclidean.
This preponderant requirement for conceptual neutralities among
closed-class elements is in sharp contrast with the referential freedom of
lexical items, which can express not only structural abstractions but also
wide-ranging specificities. For example, specificity as to magnitude is seen
in nouns like inch and mile; as to shape, in nouns like circle, adjectives like
square, and verbs like ricochet; as to rate, in verbs like dawdle and hurry, in
material, in a noun and verb like iron and bleed; as to sensorimotor
characteristics in watch and wade; and, of course, as to color by such adjectives as
red and blue.
To elaborate further the contrast between the grammatical and the
lexical type of specification, consider the full complement of both element-

172
LEONARD TALMY
types in a single whole sentence, viz., that selected in (9):
(9) A rustler lassoed the steers.
We first list the grammatical elements present in the sentence and the
notions that they specify in (10):
(10)
a.
-ed
b. the
d. -s
e. a...-0
f. the grammatical category
of "verb" for lasso
g/h. the grammatical category
of "noun" for rustler/steer
i/j. the grammatical relations
of "subject"/"object" for
rustler/steer
k. active voice
1. intonation, word order,
pattern of auxiliaries
'occurring at a time before that of
the present communication'
'has ready identifiability for the
addressee'
'not before in discussion or
otherwise readily identifiable for the
addressee'
'multiple instantiation of object'
'unitary instantiation of object'
'event character'
'entity character'
'agent'/'patient' (among the
possibilities)
'point-of-view at the agent'
'the speaker "knows" the situation
to be true and asserts it'
The lexical items in the sentence have specifications that can be
characterized as in (11):
(11) a complex of concepts involving:
a. rustler: a person, property ownership, illegality, mode of
activity
b. steer: object of particular appearance, physical makeup, etc.
relation to animal kingdom
castration
institution of breeding for human consumption
c. lasso: certain objects (a body and a lasso) in particular
configurations

THE RELATION OF GRAMMAR TO COGNITION
173
certain movement sequences
accompanying cognitive intending, directing,
monitoring, etc.
In surveying the two lists, we can see these differences emerge: The
grammatical elements are more numerous, and their specifications seem
more spare and simpler, and more structural in function. Together, their
specifications seem to establish the main delineations of the scene
organization and communicative setting of the CR evoked by the sentence. The
lexical elements are fewer in number, but their specifications are greater in
quantity and complexity, and function more to contribute content than
structure. The lexical specifications are greater in three ways: compared to
a grammatical specification, each has a) more total information, b) greater
intricacy of information, and c) more different types of information
together. Taken together, their specifications comprise most of the
conceptual content of the CR scene that is evoked by the sentence.
These grammatical-lexical differences can be set into further relief by
in turn varying each element-type while keeping the other constant. Thus,
varying only the grammatical elements of (9), as is done in (12), seems to
alter the scene organization and discourse properties of the referent event
but to leave its basic contents intact:
(12) Will the rustlers lasso a steer?
By contrast, varying only (9)'s lexical elements, as in (13), shifts us to a new
scene altogether, and yet the basic breakup of the scene and of its
communicative setting seems to remain the same:
(13) A machine stamped the envelopes.
2. Categories of Grammatically Specified Notions
The preceding sampling of grammatical elements has yielded a set of
notions helpful toward discovering common semantic properties. But the
set has been small and unstructured. With a broader and more systematic
investigation, patterns of organization among the notions become evident.
Grammatically specified notions can be seen to pattern in categories, and
the categories, in turn, in integrated systems, as presented below. And
within these notional patterns can be seen certain regularities of function
and process. These patterns and regularities constitute principal features of

174
LEONARD TALMY
conceptual organization in language.
Several such features are brought forward below. One feature is an
extensive homology between the representation of space and that of time.
The first category, "dimension", includes this space-time homology, and
largely crosscuts the remaining categories. These categories will, in the
majority, apply to both space and time, and parallel examples from each
dimension will be presented side by side.
Another feature is that, of the member notions of any category
represented in a language, often each notion will be incorporated in at least
some lexical items. Correlatively, the language will often contain
grammatical forms that interact with each lexicalization type in a way that yields the
expression of another notion of the category. Each such type of interaction
can be regarded as a type of cognitive operation that converts the indication
of one notion to that of another within the same category. A corollary
feature is that a language with grammatical forms for converting from notion
A to notion B frequently has forms as well for conversion in the reverse
direction — that is, it can also trigger the reverse cognitive operation.10
Some of the grammatical forms in a language function specifically to
perform a particular conversion operation. Others simply make structural
specifications that can come into conflict with the specification of a
neighboring lexical item. In such cases, the basic pattern is that the
grammatical form's specification always takes precedence, and triggers a kind of
operation, a "shift", in the lexical item's referent that brings it into
accord.11
As a note on methodology in what follows, efforts were made to
determine categories on the basis of particular grammatical meanings
encountered, rather than to posit the categories as part of an a priori schema which
then sought corroborative examples. In the research leading to this paper,
grammatical forms were sampled from a range of languages, but an effort
has been made to take most of the exemplification from English.
2.1 Dimension
The category of "dimension" has two principal member notions,
'space' and 'time'. The kind of entity that exists in space is — in respectively
continuous or discrete form — 'matter' or 'objects'. The kind of entity
existing in time is, correspondingly, 'action' or 'events' — terms here used
neutrally as to whether the entity is static or changing. These notions thus
relate as in (14):12

THE RELATION OF GRAMMAR TO COGNITION
175
(14) dimension
space:
time:
continuous
matter
action
discrete
objects
events
Homologies between the linguistic structuring of space and of time will
be indicated in the categories that follow. But here we can indicate
operations of conversion between these two main members of the dimension
category. Thus, a verb root that lexicalizes expression of an event or of action
as a temporal quantity can be associated with grammatical forms, including
nominalizations, that signal a cognitive operation of "reification". By the
semantic effect of this operation, the referent becomes conceptualized as an
object or a mass, one that can participate in many of the same activities
(such as being given or gotten) as a physical quantity, as well as in many of
the corresponding syntactic constructions (including pluralization and
modification: ...gave me two quick calls), as exemplified in (15). (A way of
representing the grammatical complexes involved here and in the next
operation is presented in connection with the following category.)
(15) an event: reified as an object:
John called me. John gave me a call.
I was called by John. I got a call from John.
action: reified as mass:
John helped me. John gave me some help.
I was helped by John. I got some help from John.
The reverse conversion also occurs. A noun referring to an object or mass
can be associated with grammatical forms, including verb-forming
derivations, that signal a cognitive operation of "actionalizing". By this operation,
the physical referent is melded together with some of the activity in which
it participates, with the semantic effect that much of the referent's tangible
concrete character is backgrounded, subordinated to a conceptualization in
terms of a process of occurrence, as illustrated in (16):
(16) object(s)/mass: actionalized as:
Hail(stones) came in through It hailed in through
the window the window.

176
LEONARD TALMY
Ice is forming over the windshield. It is icing up over
the windshield.
I removed the pit from the cherry. I pitted the cherry.
2.2 Plexity
The category here to be termed "plexity" is a quantity's state of
articulation into equivalent elements. Where the quantity consists of only one
such element, it is "uniplex", and where it consists of more than one, it is
"multiplex". When the quantity involved is matter, plexity is, of course,
equivalent to the traditional linguistic category of "number" with its
component notions "singular" and "plural". But the present notions are intended
to capture the generalization from matter over to action, which the
traditional notions do not do.13
Specifications as to plexity are made by both lexical items and
grammatical elements, and there is interplay between the two when they are
both in association. Example English lexical items that basically specify a
uniplex referent are — for matter and action, respectively — bird and (to)
sigh. They can occur with grammatical elements that themselves specify a
uniplexity, like those italicized in (17a) (many languages have here a more
regular, overt system of markers than English). But they can also occur
with grammatical elements that specify a multiplexity, as in (17b). In this
association, such elements can be thought to trigger a particular cognitive
operation, one of "multiplexing". By this operation, an original solo
referent is, in effect, copied onto various points of space or time.
(17) matter action
a. uniplex A bird flew in. He sighed (once).
b. multiplex Birds flew in. He kept sighing.
The reverse of the preceding pattern is also found in language. First,
there are lexical items that intrinsically specify a multiplexity. English
examples are furniture or timber (i.e., 'standing trees') for matter and
breathe for action, as used in (18a). And, too, there are grammatical forms
able to appear in association with these, as in (18b), that signal an operation
the reverse of multiplexing — one that can be called "unit-excerpting". By
this operation, a single instance of the specified equivalent units is taken
and set in the foreground of attention.

THE RELATION OF GRAMMAR TO COGNITION
177
(18) matter
action
a. multiplex Furniture overturned in the earthquake.
She breathed without pain.
b. uniplex A piece 0/furniture overturned in the earthquake.
She took a breath / breathed in without pain.
The English grammatical forms seen above that signaled multiplexing s
and keep Ang— consisted solely of explicit morphemes. The forms that
signal unit-excerpting differ in that they also include abstract elements:
particular grammatical categories that require the insertion of one out of a
certain set of lexical items, as represented in (19c,d). The forms can,
moreover, contain two or more independent elements. These forms are here
considered to be "grammatical complexes", comparable to other
grammatical constructions or indeed to lexical complexes (collocations): they
combine distinct elements within a structural whole serving a single overall
semantic function. Actually, by one analysis, all grammatical forms are
complexes, merely ranked along a cline of elaborateness. Under this
analysis, a grammatical form includes not only any explicit and generic
elements, but also the semantic and syntactic category memberships of its
input and output forms, as represented throughout (19). Thus, the English
multiplexing forms, in (19a,b), are merely at the simpler end of a
continuum:
(19)
»
r
[ In + s
upx
N
mpx
e.g., bird: birds
(b)
keep + [ ]v + -ing
upx
V
mpx
e.g., sigh: keep sighing

178
LEONARD TALMY
(c)
Nunilo/ + [_]N
N
J UpX
e.g., furniture: a piece of furniture
(d)
Vdummv. + [[ ]v +DERIV]N
mpx upx
upx
e.g., breathe: take a breath
(<n
+ PTC
V
upx
e.g., breathe: breathe in
Support is lent to the thesis that a more elaborate grammatical complex can
have a semantic unity by the existence, within the same or another
language, of a simpler form with the same semantic function. As an example of
just this circumstance, the English unit-excerpting complex for nouns,
which is rather elaborate, is paralleled in function by a simple suffix in
Yiddish, either - " / or - " ele (otherwise indicating diminutives), as illustrated
in (20):
(20) zamd'sand': zemdl'grain of sand'
groz 'grass': grezl 'blade of grass'
shney 'snow': shneyele 'snowflake'
2.3 State of Boundedness
When a quantity is specified as "unbounded", it is conceived as
continuing on indefinitely with no necessary characteristic of finiteness intrinsic
to it. When a quantity is specified as "bounded", it is conceived to be
demarcated as an individuated unit entity. In application to nouns, these
notions largely correspond to the traditional linguistic distinction between
"mass" and "count", and in application to verbs they can correspond to

THE RELATION OF GRAMMAR TO COGNITION 179
"imperfective" and "perfective", among other terms (the closeness of these
correspondences varies with different usages of the traditional terms).
However, as with plexity, the concepts designated by the new terms are
intended to capture the commonality between the space and time
dimensions and to generalize over their usually separate analyses.
Among English examples of lexical items, water and (to) sleep basically
specify unbounded quantities, whereas sea and (to) dress basically specify
bounded ones. These specifications are demonstrated by the fact that these
words are, respectively, unacceptable and acceptable in construction with
the grammatical complex "in NP _ . ", which itself specifies bounded-
° r extent-of-time ' r
ness, as seen in (21):
(21) matter
action
a. unbounded *We flew over water in 1 hr.
*She slept in 8 hrs.
b. bounded We flew over a sea in 1 hr.
She dressed in 8 mins.
As with plexity, there exist grammatical elements that can, in
construction with a lexical item, shift its basic specification for state of boundedness
to the opposite value. Those acting in this way on an unbounded-type
lexical item, in effect, trigger a cognitive operation of "bounding", or "portion-
excerpting". By this operation, a portion of the specified unbounded
quantity is demarcated and placed in the foreground of attention. Examples of
such grammatical elements in English are shown in (22).
The reverse of the preceding pattern also exists. The English nouns
shrub and panel each refer intrinsically to a bounded entity. But the
grammatical elements -ery and -ing can be added to them, yielding shrubbery
and paneling, forms which now refer to unbounded quantities. In effect, the
grammatical elements have triggered a cognitive operation of "debound-
ing" whereby the quantity formerly within bounds is now conceptualized in
a form with indefinite extension. In English, however, such elements are
not productive; they cannot, for example, be used with sea to yield the
meaning 'pelagic water', nor with (a) tear to yield 'lachrymal fluid'.14

180
LEONARD TALMY
(22)
matter:
N-
uantity °f + I Jn
bounded-quantity
e. g., water: body of water
action:
[ 'Vunbd + f°T Nextent-of-time
e.g., sleep: sleep for an hour
unbd
IN
bd
'bd
2.4 State of Dividedness
The category of "state of dividedness" refers to a quantity's internal
segmentation. A quantity is "discrete" (or "particulate") if it is
conceptualized as having breaks, or interruptions, through its composition.
Otherwise, the quantity is conceptualized as "continuous".15
Both lexical and grammatical elements are sensitive, in their
specifications, to the distinctions of this category. But there appear to be no
grammatical elements that solely specify discreteness or continuity for a
quantity, nor any that signal an operation for reversing a quantity's lexically
specified state of dividedness. If forms of the latter type existed, we can
describe how they would behave. A grammatical form for a continuous-
type lexical item would signal an operation of "discretizing", whereby the
originally continuous referent would become conceptualized as a
particulate aggregation. Conversely, a grammatical form for a discrete-type lexical
item would trigger an operation of "melding", whereby the separate
elements of the original referent would be conceptualized as having fused
together into a continuum.
Although such grammatical forms seem lacking, there do exist certain
indirect or inexplicit mechanisms for these operations. Thus, the continuity
specified by the noun water can be reconceptualized as discrete with the
locution particles of, as in: Water I Particles of water filled the vessel.
However, the grammatical complex used here does not directly specify this shift
but, like the complexes in Sections 2.5 and 2.13, comprises a several-stage
sequence of other cognitive operations. In the reverse direction, there

THE RELATION OF GRAMMAR TO COGNITION 181
appears to be a general conceptual tendency for a basically discrete-type
referent of a lexical root to undergo at least some degree of spontaneous
melding, without the addition of any explicit grammatical forms. Thus,
foliage, timber, and furniture, as contrasted with leaves, trees, and pieces of
furniture, tend to evoke referents with a degree of blurring and fusion
across their component elements.
Because the category of dividedness has limited realization by itself,
further treatment of it will be deferred until the next section, where it can
be seen in interaction with the other categories.
2.5 The Disposition of a Quantity: A System of Categories
The preceding four categories of attributes — dimension, plexity,
boundedness, and dividedness — all pertain to a quantity simultaneously
and, taken together, can be considered to constitute a system of attributes
that may be termed a quantity's "disposition". The intersections of these
categories form an array that can be schematized as in (23).
(23) discrete continuous
• • * • * ^::£ip:£::iii:JL unbounded
A' B' " "l:"
multiplex
bounded
uniplex
+ the distinction between matter and action,
which crosscuts all of the above16

182
LEONARD TALMY
Each intersection of attributes indicated here is specified by various lexical
items (although one, a bounded multiplexity for action, is quite minimally
represented in English). An example or two (most seen earlier) is given for
each intersection in (24):17
(24) A': timber/furniture B': water
(to) breathe (to) sleep
A: (a) family B: (a) sea/panel
(to) molt (to) empty
(The bird molted.) (The tank emptied.)
a: (a) bird
(to) sigh
Now if the particular contentful referent for which one chooses a
lexical item happens to be wedded, by that lexical item, to an unwanted set of
structural specifications, there generally are grammatical means available
for converting this to a desired set. Such means range in directness from
specifying the single relevant operation to involving a circuitous sequence
of operations (cf. Section 2.13 on "nesting"). A number of starting- and
ending-points for such conversions, and the means for accomplishing them,
are indicated in (25):
(25) A'—»A a stand of timber B' —»B a body of water
breathe for an hour sleep for an hour
A'—»a a piece of furniture —
take a breath / breathe in
A—»a a member of a family —
?molt a single feather
A—»A' members of a family B—»B' paneling
(A-»a-»A#)
molt and molt empty and empty
a—»A' trees
keep sighing
a—»A a stand of trees —
(a-»A'-»A)
sigh for a while

THE RELATION OF GRAMMAR TO COGNITION 183
2.6 Degree of Extension
Implicit in the vertical dimension of the schematic arrangement in (23)
is a further category that can be called "degree of extension". This category
has three principal member notions, terms for which are given in (26)
together with schematic representations of the notions for the linear case.
Lexical items referring to either matter or action may be taken to
incorporate specifications as to their referent's basic degree of extension, and three
examples of these for the linear spatial case are also shown in (26):18
(26) point bounded extent unbounded extent
• • • • •
speck ladder river
Now a lexical referent that is perhaps most basically conceived as of
one particular degree of extension can, by various grammatical
specifications that induce a shift, be reconceptualized as of some other degree of
extension. For a first example, consider the event referent of climb a
ladder, which seems basically of bounded linear extent in the temporal
dimension, as is in fact manifested in (27) in conjunction with the grammatical
element "m + NP , . ":
extent-of-time
(27) She climbed up the fire-ladder in 5 minutes.
With a different accompanying grammatical form, like the "at + NP .
tjmc" in (28), (as well as different contextual specifications), the event
referent of the preceding can be shifted toward a conceptual schematization as
a point of time — i.e., as being point-durational:
(28) Moving along on the training course,
she climbed the fire-ladder at exactly midday.
This shift in the cognized extension of the event can be thought to
involve a cognitive operation of "reduction" or, alternatively, "adoption of
a long-range perspective". This shift can also go in the other direction. The
event referent can be conceptually schematized as an unbounded extent by
the effect of grammatical forms like "keep -ing'\ "-er and -ef\ and "as +
S", as in (29):
(29) She kept climbing higher and higher up the fire-ladder as we
watched.
Here there would seem to have taken place a cognitive operation of "mag-

184
LEONARD TALMY
nification", or "adoption of a close-up perspective". By this operation, a
perspective point is established from which the existence of any exterior
bounds falls outside of view and attention — or, at most, is asymptotically
approachable.
The preceding event referent was continuous, but a discrete case can
exhibit the same shifts in extension. One such case, perhaps to be
considered as most basically of bounded extent, is shown with that degree of
extension in (30a). But the referent can also be idealized as a point, as in
(30b) (clearly, the cows would not all have died at the same moment, and
yet the spread of their death times is conceptually collapsed into such a
single moment). Or, the referent can be schematized as an unbounded
extent, as in (30c):
(30) a. The cows all died in a month.
b. When the cows all died, we sold our farm.
c. The cows kept dying (and dying) until they were all gone.
The alternative schematizations of extension just seen as specifiable for
an event referent are generally also available for an object referent. Thus,
e.g., the referent of (a) box can be specified for idealization as a point or as
a bounded extent (of area or volume). Some grammatical elements making
such specifications are illustrated in (31). Also set forth here are the
homologies between these and the event-specific elements:
(31) point The box is 20 feet away from
the wall.
I read the book 20 years ago.
bounded extent The box is 2 feet across.
I read the book in 2 hours.
(point within) bounded extent The ball is in the box.
She left while I read the
letter.
2.7 Pattern of Distribution
The pattern of distribution of matter through space or of action
through time is a further category of notions that can be both grammatically
and lexically specified.19 For action through time — the only dimension we
will be looking at here — this category together with the preceding one
largely constitute the traditional category of "aspect".

THE RELATION OF GRAMMAR TO COGNITION 185
Several of the main patterns of distribution for action through time are
shown schematically in (32) (the dots here, which represent situatedness in
complementary states, should really be adjacent, but they are sketched
apart with a connecting line to show the crossing of state-interfaces). Also
shown are illustrative English verbs, both non-agentive and agentive, that
incorporate these patterns.
(32) one-way one-way full-cycle multiplex steady-state gradient
non-resettable resettable
die fall flash breathe sleep widen (intrans)
kill drop hit beat carry widen (trans)
One can determine that these verbs incorporate the specifications indicated
by noting the grammatical forms with which they can and cannot occur (or,
to put the latter case in our terms: ...grammatical forms toward whose
specifications they will not [readily] shift). A full demonstration is not in
order here, but a few examples show the principle: The resettable type of a
one-way event is distinguished from the non-resettable type by its
compatibility with iterative expressions, as in: He fell 3 times; the non-resettable
type cannot occur here: *He died 3 times. This same one-way form is
distinguished from a full-cycle form by its ability to appear in sentences like: He
fell and then got up, which the latter cannot do: *The beacon flashed and
then went off. A gradient type can appear with adverbs of augmentation, as
in The river progressively widened, unlike a steady-state type: * She
progressively slept. And so on.
Grammatical elements can, of course, also specify differing patterns of
temporal distribution, and the present form of diagramming can readily
reveal some of their distinctions. Thus, the closed-class elements back and
again, singly and in combination, can indicate versions of full-cycle, sesqui-
cycle, and double-cycle patterns, as shown in (33):

186 LEONARD TALMY
(33) one-way full-cycle sesqui-cycle double-cycle
go to sleep go back to sleep go to sleep again go back to sleep again
Now consider the circumstance where a verb of one distribution type
appears with grammatical forms of another type. The resultant seems
invariably to be that the verb shifts its specifications into conformity with those
of the grammatical forms. For an example we again take die, whose basic
specifications can be adjudged as point-durational one-way non-resettable
— schematizable, now more precisely, as: % . This verb is used with its
basic specifications in a sentence like (34a). But in a sentence like (34b), the
grammatical form "be + -ing" induces a shift. In effect, the infinitesimal
interval between the two states involved for die — viz., 'aliveness' and
'deadness' — is spread out, with the creation thereby of an extent-dura-
tional gradient. This is the shift in the distribution pattern's structural type.
But concomitantly, a shift in the basic contentful referent is engendered.
Instead of 'dying', the new gradient refers to 'moribundity'. The distinction
becomes clear in noting that, as the conception is structured linguistically,
one can have been dying without having died, and, correlatively, one can
have died without having been dying.20
(34) a. He died as she looked on.
b. He was (slowly) dying as she looked on.
2.8 Axiality.
The adjectives in a pair like well/sick behave contrarily when in
association with grammatical forms specifying degree like slightly and almost, as
seen in (35a), and they select for different readings of temporal forms like
"in + NPcxtcnt_of_time"> as seen in (35b). In this, perhaps surprisingly, they
parallel the behavior of certain kinds of expressions that specify spatial
relations, e.g., at the border I past the border:
(35)
a- He\ sliphtlv / S'Ck ' PaSt thC b°rder- \
He s slightly | *weU / *at the bordei, j
awake
asleep

THE RELATION OF GRAMMAR TO COGNITION 187
TT , , f well / at the border. )
He s almost < 0 . . . 0 , . , >
( ysick / ?past the border. J
b. He got well / to the border in 5 days. — i.e., progressively in the
course of
He got sick / past the border in 5 days. — i.e., first after the
elapse of
This behavior can be accounted for by positing that such adjectives, in
referring to a more generic notional parameter, such as that of 'health', are
not simply "opposites" but, rather, presuppose a schematic axis that is
structured and directed in a particular way. Each adjective, then, labels a
different portion of that axis. The adjectives here seem in particular to
presuppose a directed line bounded at one end; well refers to the end-point
while sick refers to the remainder of the line, correlating greater magnitude
with greater distance along the line. These are the "axial properties", or
"axiality", of the lexical items, i.e., the specific relations each has to a
particular conceptual axis and to other lexical items with referents along the
same axis. It is the lexicalization of such axiality that can align adjectives
with expressions of spatial relation. Grammatical forms like the ones just
above also have axial properties, and these can function in consonance with
those of a lexical item, as in the acceptable cases of (35), now schematized
as to axiality in (36):
(36) past the at the
border border
is slightly sick / past the border
is almost well / to the border
got well / to the border in 5 days
got sick / past the border at dawn (after
5 days)
sick well

188
LEONARD TALMY
In other cases, though, the axiality of a grammatical form can conflict with
that of a lexical item and, accordingly, can cause the latter to shift. Thus,
sick in (37) — now associated with grammatical forms that refer to an end-
point — shifts from its basic "directed shaft" type of axiality, and indeed
from its reference to an axis of 'health'; it now specifies the end-point of an
axis pertaining to 'feeling physically bad'.
(37) (After exposure to the virus, he felt worse and worse and)
he was almost sick at one point. / he finally got sick in 3
days.21
2.9 Perspectival Mode
As seen earlier, a particular event, whether static or changing, can
have a pattern of distribution through time that is perhaps most basically
associated with or intrinsic to it in its own right. But, in addition, language
has the means for specifying an independent schema as to how one is to
attend to the event. This schema includes the location and deployment of
the perspective point one adopts from which to regard the event and the
distribution of one's attention over the event. This category of
specifications, here called the "perspectival mode", can either conform with or
diverge from the event's own basic pattern of distribution. Two principal
members of the category are characterized in (38):
(38) the assuming of:
a. a steady-state long-range perspective point with global
scope of attention
b. a moving close-up perspective point with local scope of
attention
For illustration, consider first an example with a basically steady-state
referent, viz., objects in location. The (38a) type of perspectival mode —
the one more congruent with such a referent — is invoked in (39a), multiply
specified there by the set of grammatical forms shown underlined, namely,
plural forms, an adverbial expression of spatial dispersion, and the locative
preposition in. But these can be replaced by grammatical forms coding for
the (38b) perspectival mode — as in (39b) with singular forms, an adverbial
expression of temporal dispersion, and the motion preposition through.
Thereby, the evoked CR is converted to one where one's perspective and
attention or one's own projected location shifts in turn from object to
object. In effect, a steady-state multiplexity of objects has been converted

THE RELATION OF GRAMMAR TO COGNITION 189
to a sequential multiplexity of events consisting of conceptualized
encounters with the objects.
(39) a. There are houses at various points in the valley.
b. There is a house every now and then through the valley.
For representing certain static spatial configurations, the
moving-perspective mode, though non-congruent in character, is greatly favored over the
steady-state mode. Thus, the ready colloquial formulation of (40b) for
moving-perspective is matched in the global steady-state mode of (40a) only by
a stilted scientific style:
(40) a. The wells' depths form a gradient
that correlates with their locations on the road.
b. The wells get deeper the further down the road they are.
The reverse of the preceding circumstances also exists. That is, a
sequential multiplexity of events, an example of which is represented in
(41a) with the more congruent moving-perspective mode, can also become
the object of a fixed global viewing, as represented in (41b).
Metaphorically, the effect here is as if the vertical time line is tilted up into present-
moment horizontality for integrated or summational assessment.
(41) a. I took an aspirin time after time during / in the course of
the last hour.
b. I have taken a number of aspirins in the last hour.22
2.10 Level of Synthesis
The category to be considered now pertains to bounded quantities, like
those schematized in the A/B row in (23). One form of locution already
seen to specify such quantities is the particular type of "NP of NP"
construction illustrated in (42a). Here the second NP specifies the identity of
the quantity involved, itself conceptualized as without intrinsic bounds,
while the first NP specifies the bounding, or "portion-excerpting"', per se of
the quantity. Moreover, in addition to such a pure operation of bounding,
the first NP can further specify the particular form or configuration that the
excerpted portion has, as in (42b):23
(42) a. a set of trees a body of water
b. a cluster of trees a puddle/drop of water
The two NPs here can be seen as coding for two different "levels of
synthesis". Describing this for the internally discrete case, e.g., a cluster of trees,

190
LEONARD TALMY
we can say that the second NP specifies an unsynthesized multiplexity of
independent elements, while the first NP specifies a particular Gestalt
synthesized out of that multiplexity.
Furthermore, language can mark an additional cognitive distinction
here. Either level of synthesis can be placed in the foreground of attention
while the other level is placed in the background. One grammatical device
for marking this is the placement of the foregrounded NP at the head of the
larger nominal construction (in English, placing it first), as shown in (43a).
With the use of this device, moreover, predications can be made that
pertain solely to one level of synthesis or the other, as seen in (43b):
(43) a. the cluster of trees / the trees in the cluster
b. That cluster of trees is small.
The trees in that cluster are small.
There are certain forms, furthermore, whose referents are keyed to
applying to only one or the other level of synthesis. Thus, together (toward each
other) tends to correlate with multiple objects at large, while in upon -self
tends to correlate with a composite formed therefrom, as seen in (44):
(44) The bricks in the pyramid came crashing together / *in upon
themselves.
The pyramid of bricks came crashing in upon itself / together.
The preceding phenomena have involved the shift of attention from a
multiplexity to a Gestalt that it can constitute, a process that can be called
"Gestalt formation". Also encountered in language are means for
specifying the reverse: shifting attention from a Gestalt to components seen as
constituting it, in a process of "componentializing". This procedure can
take place when the starting lexical item specifies an entity taken to be
already at the more synthetic level, as is the case with iceberg in (45a). By
grammatical devices like those in (45b), such an entity can be analytically
converted from conception as a coherent whole to one of component parts
and their interrelations. Again we encounter a surface form — in two —
that correlates with only one level of synthesis and not the other:
(45) a. The iceberg broke in two.
b. The two halves of the iceberg broke apart (*in two).
The two levels of synthesis with the two directions of conceptual shift
applicable to them define four notional types, as indicated in (46). The term

THE RELATION OF GRAMMAR TO COGNITION 191
Figure is used here as described in Talmy (1978b, 1983).
(46) example type operation
cluster of trees: "composite Figure" 1
trees: "multiple Figures" I
_Gestaltjf orm ation _
iceberg: "meta-Figure" I
two halves of iceberg: "component Figures" T
componentialization
2.11 Level of Exemplarity
A further cognitive distinction can be specified for a multiplexity of
objects- This distinction does not affect the basic reference to all the
members of the multiplexity, but addresses how attention is directed and
distributed within that multiplexity. Either the/w// complement of the
multiplexity is in the foreground of attention, with perhaps individual items here and
there singled out in the background of attention. Or a single exemplar out
of the multiplexity is placed in the foreground of attention, with the
remaining items more dimly conceived in the background of attention. This
distinction as to "level of exemplarity" is specified by grammatical devices in
perhaps most languages. But English stands out in the extensiveness of its
specifications: there are different pairs of grammatical forms that mark the
distinction for a number of different types of multiplexity. A rather full list
of these pairs is indicated in (47), with examples showing first the
full-complement form and then the counterpart exemplar form:
(47) a. Oysters have siphons / a siphon.
An oyster has siphons / a siphon.24
b. All oysters have siphons / a siphon.
Every oyster has siphons / a siphon.
c. All the members raised their hand(s).
Each member raised his hand(s).25
d. Many members raised their hand(s).
Many a member raised his hand(s).
e. Some members here and there raised their hand(s).
A member here and there raised his hand(s).

192
LEONARD TALMY
f. Members one after another raised their hand(s).
One member after another raised his hand(s).
g. Hardly any members raised their hand(s).
Hardly a member raised his hand(s).
h. No members raised their hand(s).
No member (Not a member) raised his hand(s).
i. She held a gun in both hands.
She held a gun in either hand.26
2.12 Other Categories and Processes
A number of further notional categories and cognitive processes can be
discerned in language, but there is opportunity here to present briefly only
two additional examples:
Scene-Division Properties. A lexical item can have particular "scene-
division properties", that is, a principal breakup of its referent into parts
and participants. For example, the referent of the English verb serve breaks
up into an activity, an item served, and a social dyad involving the two roles
of 'host' and 'guest' — this last being its particular "personation" type
(Talmy 1985a) — as illustrated in (48a). But grammatical forms can also
have scene-division properties. Thus, a subject-plus-reflexive-object
complex has a single-role specification. When such a grammatical form occurs
with a dyadic verb like serve, it triggers an operation of
"monad-formation": the verb's referent shifts to one with monadic personation, as in
(48b). In this shifted state, its referent is equivalent to that of an
intrinsically monadic expression, like that in (48c):27
(48) a. The host served me some dessert from the kitchen.
b. I served myself some dessert from the kitchen.
c. I went and got some dessert from the kitchen.
Associated Attributes. Lexical expressions like apartment and hotel
room, in addition to their basic denotations, may be taken to have
"associated attributes" — here, respectively, those of 'permanent residence'
and 'temporary lodging.' Such attributes may mesh or conflict with the
specifications of another element in the same sentence. The attributes of
the above two nominals mesh and conflict respectively, e.g., with the
closed-class directional adverb home, which specifies a permanent residence.
In the case of conflict, as in (49b), a cognitive process operates on the lexical
item to leave its essential characteristics intact but replace its associated

THE RELATION OF GRAMMAR TO COGNITION 193
attributes with the closed-class element's specifications:
(49) a. He drove home to his apartment,
b. He drove home to his hotel room.
2.13 Nesting: An Interaction of Categories
A number of what can be considered "meta-properties" govern the
behavior of categories of grammatical notions, in general and with respect
to one another. One of these, the capacity for nesting, already illustrated in
Section 2.5, can be readily presented here: The operations and shifts
described throughout Section 2 need not take place singly. The output of
one can serve as the input to another, thereby building up hierarchical
levels of embedding. While there are a number of interesting examples of
this for different types of matter and action, we will go directly to
illustrating one of the longer cases in (50):
(50) a. The beacon flashed (as I glanced over).
b. The beacon kept flashing.
c. The beacon flashed 5 times in a row.
d. The beacon kept flashing 5 times at a stretch.
e. The beacon flashed 5 times at a stretch for 3 hours.
In (50a), the lexical verb flash appears with its basic structural specification
as a point-durational full-cycle uniplex event. This undergoes the operation
of multiplexing, to yield the unbounded multiplexity in (50b). This then
undergoes bounding in (50c). This bounded multiplexity then first goes
through the operation of reduction to become schematized as a new
pointlike uniplex quantity, and this is in turn multiplexed, yielding (50d). This
new unbounded multiplexity is finally then bounded in (50e). The nesting
of structural specifications in this last stage can be represented
schematically as in (51):
(51) [( )-( ) ( )-( )]
Quite analogous to this temporal nesting, except for the lack of specific
numerals, is the spatial example in (52):
(52) a. A duck landed on the pond.
b. Ducks landed on the pond.
c. A flock of ducks landed on the pond.
d. Flocks of ducks landed on the pond.
e. A group of flocks of ducks landed on the pond.

194
LEONARD TALMY
2.14 Four Imaging Systems
Most of the preceding categories of grammatically specified notions,
together with categories not discussed here, group together under four
much broader conceptual systems, ones that can be understood as principal
"imaging systems" of language. These are great complexes in language that
organize the structuring and the "viewing" of conceptual material. The four
systems outlined here (there are additional ones) are relatively independent
of each other in content, with each adding a distinct conceptual parameter
to those of the others, but their contributions can be coordinated and
linked, at times by individual grammatical forms.
The first imaging system is "structural schematization". This system
comprises all the forms of conceptual delineation that can be ascribed to a
quantity, or to the pattern in which two or more quantities are interrelated,
whether in space or time or some other conceptual dimension. A number of
the categories of notions presented above are part of this system. After
"dimension", all the categories pertaining to the disposition of a quantity and
its generalizations belong here, in particular, the categories of plexity, state
of boundedness, state of dividedness, degree of extension, pattern of
distribution, and axiality. Belonging here, too, are the category of
scene-division properties and that of the "partitioning" of space or time that is
specified by such deictics as this and that, described in Section 1. A further
major component of this imaging system is the spatial or temporal
"geometric" schematization, including the topology-like kind, that is specified
especially by the adpositional systems of languages. This was only touched on
here in Section l's discussion of the English preposition through, but it is an
extensive domain, one treated at length in such works as Bennett (1975),
Gruber (1965), Jackendoff (1977), Langacker (1986a), Talmy (1975, 1982,
1983), Herskovits (1986).
The second imaging system is the "deployment of perspective". Given
a structurally schematized scene, this system pertains to how one places
one's "mental eyes" to look out upon that scene, including the location, the
distance away, and the movement pattern of this conceptual perspective
point. Belonging to this system from the discussion above is the category of
perspectival mode, with its options of a steady-state or a moving
perspective point. Also here is the category of degree of extension, when its
alternatives are interpreted as "adopting a long-range vs. a close-up
perspective".
The third imaging system is "distribution of attention". Given a

THE RELATION OF GRAMMAR TO COGNITION 195
schematized scene and a vantage from which to regard it, this system
pertains to the allocation of attention that one can direct differentially over the
aspects of the scene. Belonging here from the discussion above are the
categories of level of synthesis and level of exemplarity, as well as the
component of the perspectival mode category that involves global vs. local
scope of attention. In addition, a major category, not treated here,
comprises the obligatory "Figure/Ground" distinctions that language imposes
on a referent scene. Here, within a scene, there is ascribed to one element
the status of "Figure", with its attentional primacy, and to another element
the status of "Ground", with its function in the background of attention as
a reference object for the localizing of the Figure (Talmy 1978b, 1978c,
1983). Additionally in this imaging system are such grammatically marked
discourse concepts as focus, topic, comment, given and new.
The fourth imaging system, not treated here at all, is "force dynamics",
which, given a structured scene, involves the forces that the elements of the
scene exert on each other. Comprehended here are the notions of force
exerted by one quantity on another, as well as notions of resistance to such
force, the overcoming of such resistance, blockage to the exertion of force,
and the removal of such blockage. The system of force dynamics includes
the traditional linguistic concepts of the "causative", but is a generalization
over those concepts (Talmy 1976, 1985b).
3. Further Cognitive Connections
Grammatically specified structuring appears to correspond, in certain
of its functions and characteristics, to the structuring in other cognitive
domains, such as that of visual perception, compared below, or those of
inference and memory. In particular, perhaps the principal overarching
function of the structuring common across cognitive domains is that of
providing conceptual coherence, that is, acting as a means for integrating and
unifying a body of otherwise disparate conceptual material. In language
and, as suggested below, in vision, this fundamental function has three
main global forms of realization: coherence across a conceptual inventory,
coherence within a scene, and coherence through time.
Across the inventory of notions available for expression within any one
language, grammatical specifications bring coherence principally by
constituting a classification of the vast variety of conceived and perceived
material. They gather different portions of the material together into subdivi-

196
LEONARD TALMY
sions distinct from each other. By this, any particular currently cognized
element is associated with its implicit "subdivision-mates". An illustrative
case here are the twenty-plus motion-related prepositions in English, such
as through and into, which together subdivide the conceptual domain of
'paths considered with respect to reference-objects'. This domain covers a
great and varied range, but any particular "path" generally falls within the
purview of one or another preposition, associated there with other "paths".
To a certain extent, such associations can be regarded as arbitrary or
idiosyncratic. Thus, as seen earlier, classed together by through are such
dissimilar cases as a straightforward liquid-parting course (walking through
water) and a zig-zag obstacle-avoiding course (walking through woods).
The question arises why such distinctions should be effaced by the
grammatical system, while they are observed by the lexical and other cognitive
systems. Why are grammatical elements — say, such prepositions — not a
large and open class marking indefinitely many distinctions? One may
speculate that the cognitive function of such classification lies in unifying
contentful material within a single conceptual system and in rendering it
manipulable — i.e., amenable to transmission, storage, and processing —
and that its absence would render content an intractable agglomeration.
Providing coherence within a cognized scene was the function of
grammatical structuring that was originally indicated in the Introduction. There
it was put forward that the grammatical elements of any particular sentence
together specify the structure of the cognitive representation evoked by that
sentence. Their specifications act as a scaffolding or framework across
which contentful material can, in effect, be splayed or draped. It can be
posited that such structuring is necessary for a disparate quantity of
contentful material to be able to cohere in any sensible way and hence to be
amenable to simultaneous cognizing as a Gestalt. That is, without such
structuring, not only does the inventory of concepts available for expression
in a language become less coherent, but also any selection of such concepts
concurrently juxtaposed by a sentence would tend to be only a collection of
elements, rather than elements assembled so as to convey an integrated
idea or thought complex.
In the course of discourse, a great welter of notions are expressed in
rapid succession, posing the potential problem of an unconnected sequence
of ideational elements. But grammatically specified structuring is a
principal contributor to the conceptual coherence through time that is requisite
here. Through such structuring, a cognitive continuity is maintained

THE RELATION OF GRAMMAR TO COGNITION 197
through this flux and a coherent Gestalt is summated over time. A language
can have a great stock of closed-class elements participating in this
function, for example, such English forms as "yes, but", moreover, nevertheless,
besides, instead, also. Such forms direct the illocutionary flow, specify the
"logical tissue" of the discourse, and limn out its rhetorical framework.
That is, these grammatical forms establish a structure that extends over a
span of time, and thus provides a conceptual level with temporal constancy
amidst more fleeting aspects of content.
The preceding three global forms of grammatically specified
structuring apply over the scope of any single language but, as indicated in the
Introduction, a fourth form must also be recognized that holds for language
in general. While each language has to some extent a different set of
grammatical specifications, there is great commonality across languages, so one
can posit that each set is drawn from an innate inventory of concepts
available for serving a structuring function in language.
Further, though, a qualifying property of this inventory can be
adduced. It can be observed that grammatically specified concepts range
cross-linguistically from ones that are of extremely widespread (perhaps
universal) occurrence and of broad application within a language, down to
ones appearing in a scant few languages with minimal application. Thus,
the innate inventory of available structuring notions that is posited here
seems to be graduated as to significance for the language faculty (cf. the
tabular listing of grammatical notions in Talmy (1985a: 126ff)). For
example, the notions 'entity' and 'occurrence' as expressed by the grammatical
categories "noun" and "verb" are of great application and probably
universal distribution, the notional category "number" seems of roughly middle
standing in the ranking, while notions like 'in the morning' and 'in the
evening' are expressed inflectionally on the verb in just a few languages.
Notably, compared to spatio-temporal structuring, the notional
category of "affect" is rather low in the graduated inventory of concepts that
language draws on for structuring purposes, a fact that is significant
considering its importance in other cognitive domains (cf. the other cross-domain
differences noted below). The affect category does have scattered
representation, for example 'affection' expressed by diminutive affixes, 'scorn' by
pejoratives, 'concern' by a conjunction like lest, and 'hurt' by the "adver-
sive" construction (as in the English: My flowers all died on me.). But
seemingly no language has a system of closed-class forms marking major
affect distinctions in the way that, say, the modal system in English specifies

198
LEONARD TALMY
distinctions of force opposition (Talmy 1985b). Such an affect system can
easily be imagined, however. Consider a parent addressing a child in
danger near an open window. Grammatical systems readily allow the
parent to refer to the spatial structure in this situational complex — Get
away from the window! — leaving the affectual component to be inferred.
But there is no closed-class form comparable to a modal, one that we could
represent as "FEAR", as in FEAR the window!, that would allow the
parent to refer to the affectual component of the complex and leave the
spatial component to be inferred. Comparably, to a child near a freshly
painted wall and about to harm it, a parent would likely again express the
spatial structure — Get away from the wall! — leaving the affect to be
inferred. There is no closed-class affect form for iike, be nice to1, which we
could represent as "FAVOR", that the parent could use instead — FAVOR
the wall! — thereby leaving the spatial component for inference.
Parallels can now be drawn between the structuring system operating
in language and that in visual perception (cf. Jackendoff in press).2*. The
principal function of structure to provide coherence appears common
across the two cognitive domains, and the three global forms of such
coherence just outlined for language correspond to comparable forms in the
operation of vision.
First, as proposed in cognitive psychology, the perception of any
particular object is mediated by its association with related objects in a schema
for that object type, and the set of such schemas constitutes a classificatory
system (Neisser 1967). This posited functioning of visual perception thus
parallels the classificatory function of linguistic structure across a language's
conceptual inventory.
Second, there is a parallel between the linguistic coherence within a
referent scene and the visual coherence within a perceptual scene. The
welter of optical sensations registered at any one moment from some whole
visual scene is rendered coherent by the perception of structural delineations
running through it. For example, one looking at, say, the interior of a
restaurant from one corner of the room does not see simply a pastiche of
color daubs and curves but, rather, perceives a structured whole that
includes the framework of the room, the spatial pattern of tables and
people, and the individuated tables and people themselves. And seeing a
person in some posture involves perceiving a structural framework in the
human figure, as Marr (1982) describes this in his treatment of the "3-D
model" in visual perception. Children's line drawings of scenes and stick-

THE RELATION OF GRAMMAR TO COGNITION 199
figure sketches of people, animals, and objects (Kellogg 1970) demonstrate
our early capacity to abstract structure from visual scenes and scene parts.
Third, one can observe a parallel between the coherence through time
in linguistic discourse and that in visual perception. If the viewer in the
illustrative restaurant now walks through the room, the patterns in which
visual stimuli and the perception of structure change give rise in turn to the
perception of a coherent continuity of path and view occurring within an
overall "scene-structure constancy".
It is reasonable to assume that, in addition to these language-vision
parallels in global structuring, a number of particular structuring devices
match across the two domains. Perhaps most of the grammatically specified
conceptual categories treated in this paper — including, for example, state
of boundedness and level of exemplarity — correspond to structuring
factors in visual perception. Further, the first three of the broader linguistic
systems for conceptual organization, the "imaging systems" outlined in
Section 2.14, seem also to correspond to broader systems of visual organization.
One can adduce still further parallels between language and vision as
to the properties of their structuring. The topology-like character of
grammatical specifications may have some parallel in the character of the
perceived delineations of a scene, or internal structure of a figure, or plan of a
path to be followed through obstacles. Such perceptions of structure seem
in certain respects to abstract away from Euclidean particularities of exact
magnitude, shape, or angle, and more to involve qualitative or approximate
spatial relationships. As a further parallel, the capacity of grammatical
specifications to nest, one within another, and form embedded structuring
seems to correspond to embedded structuring within a visual scene. The
restaurant scene above was described in terms of an overall framework that
embedded a spatial pattern, itself consisting of individuated objects. Marr's
(1982) analysis of an object like the human figure then continues the
embedding, with perceived structurings of the body ranked from its overall
linearity, to its stick-figure-like limb structure, and further to its
articulations of these components.
Whereas the preceding has outlined a set of parallels between language
and vision, significantly, each of these two cognitive domains has prominent
structuring devices that play little or no role in the other domain. Thus, in
visual perception, three major parameters that structure (parts of) a scene
are bilateral symmetry (moving or static), rotation, and dilation (expansion
or contraction) (Gibson 1966, Palmer 1983) and, if color can be treated as

200
LEONARD TALMY
structural, it is a fourth. In language, by contrast, grammatical specification
of symmetry is minimal, perhaps limited entirely to the notion 'reciprocal'.
Closed-class indication of rotation is limited in English to the prepositions
and verb particles around and over, and is barely augmented in other
languages. Dilation is grammatically expressed in English by the verb particles
in and out when referring to radial motion (spread out I shrink in) and,
again, such notions are not greatly more elaborated in other languages.
And color, of course, was this paper's original example of a conceptual
category not grammatically specified.
In the other direction, there are several prominent linguistic categories
of seemingly little structural function in visual perception. Examples are
"status of reality", as expressed, e.g., by inflections for mood, "status of
knowledge", as expressed by evidentials, and "comparison of alternatives",
as expressed by a category of particles that includes instead, only, and also.
Further possible examples are "relative temporal location", as expressed by
tense markings, "degree", as expressed by adjective inflections and
modifiers (e.g., English -er, -est, almost, too), and "force dynamics", as
expressed by modals (Talmy 1985b).
While language may not share these conceptual structuring categories
with visual perception, it may well do so with other cognitive domains.
Thus, its closed-class category "status of knowledge", which distinguishes
such notions as 'known as fact', 'inferred', 'deduced', and 'considered
probable' is very likely related to a set of basic parameters in our reasoning
faculty. And, significantly, certain conceptual categories in language have a
structuring apparently similar to that of conceptual models that form part of
our broader faculty for conceptualization, in particular, our naive or folk
models ("mental models" — cf. Gentner & Stevens 1982, Lakoff in press)
as well as models in early science. For example, Talmy (1985b)
demonstrates that the way in which language structures its concepts of
causation and force interactions greatly parallels the conceptual structuring of
naive physics and medieval physics (cf. diSessa 1986), while all three of
these forms differ conceptually in a similar way from modern physics.
Generalizing from all these findings, the possibility is that there is a
fundamental core to conceptual structure that is common across cognitive
domains — a core that thus epitomizes the nature of conceptual structure
for human cognition — but that each domain has features of structuring,
and perhaps also functions for structuring, that are not shared by others.
Determining the overall and particular character of conceptual structure is

THE RELATION OF GRAMMAR TO COGNITION 201
the aim of the research advanced in the present study, one requiring a
cooperative venture among the cognitive disciplines.
Notes
1. This paper is a moderately revised and fully rewritten version of Talmy (1978a). Since
1978, the amount of additional material on the present subject, both descriptive and
theoretical, has grown to be quite extensive. The present version incorporates a certain
amount of this new material, as well as bibliographic updating, but the remainder will be
reserved for an entirely new paper at a later date.
2. The word "evoke" is used because the relationship is not direct. The CR is an emergent,
compounded by various cognitive processes out of the referential meanings of the
sentence elements, understanding of the present situation, general knowledge, etc.
Our term "cognitive representation" is similar in purport to Fillmore's (1975)
"scene" but is chosen over that more specifically visual term. The linguistically evoked
complex can include much from other sense modalities (notably somesthetic/kinesthetic
and auditory) as well as abstract conceptual aspects. More recently, Lakoff's (in press)
notion of an "idealized cognitive model", or ICM, points toward a comparable mental
entity.
3. For their part, grammatical elements are generally more unalloyed in their indication of
structure. They can express more contentful concepts, but this is largely limited. An
example of it is in English upon as used in We marched/rode/sailed/rushed upon them
[e.g., the enemy]., where upon incorporates the notion of 'attack', seemingly equivalent
to the paraphrase 'into attack upon'.
4. More recently, research on different aspects of this broader scope has included work by
Jackendoff (1983), Bybee (1985), Slobin (1985), Morrow (1986), and Langacker (in
press).
5. Not included are regular adverbs, which seem in all languages to be derived from the
three open classes just mentioned, rather than to comprise in their own right an open
class of specifically adverbial roots. Of possible inclusion as a type of open class are the
systems of idcophones, or "expressive forms" found, for example, in a number of Asian
and African languages. Also includable, at a level above that of basic elements, are
"lexical complexes" (collocations) like English kick the bucket or have it in for.
6. More accurately, rather than a dichotomy between an open and a closed type of class,
there appears to be a cline. A class can range from having quite few members, like that
of number inflection in English, to having very many, like that of noun roots in English,
and the class's properties may correspondingly range from relatively more grammatical to
more lexical. There exist some mid-sized classes — e.g., the several score individual
classifiers of Chinese, or the three dozen or so instrumental prefixes in the polysynthetic verb
of Atsugewi (Talmy 1972. 1985a) — that appear to have properties part grammatical and
part lexical in character.
7. One can note, for example, the effect on one's cognitive representation in considering
first the sentence 1 looked at the dog and then / looked at the dogs. The addition of the
grammatical element -5 has a major effect on the delineational breakup of — to put it vis-

LEONARD TALMY
ually — the scene before the mind's eye.
For example, augmentative and diminutive elements, insofar as they refer to actual size,
seem to specify size relatively greater or lesser than the norm for the particular object
indicated. And closed-class elements specifying distance — like English just or way, as in
just/way up there — specify notions of 'near1 and 'far1 that are relative to the referent
situation.
The properties of the specifically linguistic form of topology require determination. In
this regard, consider the English preposition in, which in one main usage specifies a plane
so curved as to define a volume of space. The referent of this morpheme, as in
mathematical topology, is magnitude-neutral: in the thimble I volcano; and it is shape-neutral: in the
well I trench. But in addition, its referent is closure-neutral, i.e., indifferent to whether
the curved plane leaves an opening or is wholly closed: in the bowl I ball. And it is
discontinuity-neutral, i.e., indifferent to whether the curved plane is solid or gapped: in the bell-
jar I birdcage. These last two properties would form a proper part of language's
topological system, whereas they are strictly excluded from mathematical topology.
In many cases, a language favors only one such direction, having much lexicalization with
notion A and simple grammatical means for reaching notion B, but in the reverse
direction having only little lexicalization and complex grammatical forms. Languages differ
typologically in the directions they favor. This issue will not be taken up here, but is
treated at length in Talmy (1985a).
Shifts are actually one member of a set of "reconciliation processes" — including blends,
juxtapositions, schema-juggling, and blockage — that can be triggered by the association
of a grammatical and a lexical form with incompatible structural specifications. In the
non-shift processes, the grammatical specification does not take precedence over the
lexical one, but plays an equal role with it. Of all these processes, this paper treats mostly
shifts, but an additional number are discussed in Talmy (1977).
In addition to space and time, language represents other conceptual dimensions that also
belong to the present category. For an example, recall from Section 1 that this and that
specify a partition drawn through space — and can do so through time as well — and
indicate that a referent entity is on the same or the other side of the partition as the speaker.
Now consider the English pronouns you and they in their indefinite usage (akin to
German man or French on). These also specify a partition, but one drawn through "identifi-
cational space", understood as a new conceptual dimension. They indicate, respectively,
that 'the average person1 is or is not identified with the speaker in some relevant respect,
i.e., is on the same or the other side of the identificational partition as the speaker. Thus,
a person who smokes that is visiting a new neighborhood can ask a passer-by about the
purchase of cigarettes with you, but about the sale of cigarettes with they:
(i) Where can you buy cigarettes around here?
Where do they sell cigarettes around here?
But a person looking for a location to open a tobacco shop would ask a business
consultant in the neighborhood about purchases and sales with the reverse assignment of you
and they:
(ii) Where can you sell cigarettes around here?
Where do they buy cigarettes around here?
It is true that there are the traditional terms "semelfactive11 and "iterative" referring,
respectively, to one and more than one instantiation of an event. But there is no real

THE RELATION OF GRAMMAR TO COGNITION 203
equivalent to number: "aspect" includes too much else about the temporal structure of
action. And in any case, none of the traditional terms refers generically to both the
dimensions.
M. The mechanism actually resorted to in many such cases, including that of tear, is the use
of the plural, as in:
(i) Tears flowed through that channel in Hades.
There seems to be a sequence of cognitive operations here in getting from a bounded to
an unbounded quantity. Speculatively, the bounded quantity is first treated as a uniplex
entity, it is then multiplexed, the resultant entities are conceived as spatially juxtaposed,
and their boundaries are lastly effaced, thereby creating an unbounded continuum.
15. The present category may be prone to confusion with the preceding one. Contributory
here is the normal meaning range of continuous, which as easily covers 'boundlessness' as
it does 'internal seamlessness'. However, the two categories can vary independently.
Thus, in the preceding section, the lexical examples given for unboundedness, water and
sleep, happened also to be internally continuous; but the same demonstration of
unboundedness could have been made with internally discrete examples like timber and breathe.
In general, unbounded forms share many properties, whether continuous or discrete.
Thus, mass nouns and plural count nouns, both unbounded, share many syntactic
characteristics not shared by singular count nouns, e.g.:
(i) a / every — book / *ink / *books;
(ii) all / a lot of / more / some [unstressed] / 0 [generic] — ink / books / *book;
0 ['progressively more'] (e.g., The machine consumed ink I books I * book for an
hour.)
16. For schematizing action along the one-dimensional time axis, an adaptation of the two-
dimensional A', B\ A, and B diagrams would be necessary — and can be readily
visualized.
17. The lexical types for several of these intersections, it should be noted, do have traditional
terms. Thus, nominal forms of the a, A or A', and B' types, respectively, have been
called count nouns, collective nouns, and mass nouns. And verbal forms of the a and B'
types, respectively, have been called punctual and durative verbs. The matrix presented
here augments, systematizes, and generalizes the traditional notions.
18. This category can be considered a generalization over the earlier category of "state of
boundedness" by the inclusion of the "uniplexity" notion. It can in turn itself be
generalized — becoming the category "pattern of extension" — by the further inclusion of
such notions as a quantity bounded at one end but unbounded at the other (see Talmy
1983).
19. This category clearly patterns with the preceding five within a single system of notions,
one that would be an expansion or generalization over "disposition of a quantity".
20. Our main purpose here is to note the shift in structural distribution type. The shift in
content will doubtless prove part of a larger pattern as well, but this is not yet worked out.
21. The category of axiality can be seen as an extension of the preceding category, pattern of
distribution. The two categories address temporal stasis or change, involving both spatial
relations (e.g., fall there, past the border here) and qualitative states (e.g., die/flash,
awake/asleep there, sick/well here). But where the preceding category focused on discrete
states, the present category elaborates the notion of a scalar quantity functioning in con-

LEONARD TALMY
junction with a discrete state. Due to their structural character, these two categories
pattern together with all the categories after "dimension" as part of a single broad conceptual
system of "structural schematization", described below in Section 2.14, and are thereby
distinguished from the categories described next, which belong to different conceptual
systems.
The use of the perfect here points to a principal function of perfect forms in general: They
can indicate the temporal counterpart of matter located within a bounded extent of space,
of the type seen in (i). That is, a sentence containing the perfect, as in (ii), suggests a
paraphrase like that in (hi), which is homologous with (i):
(i) There were 5 aspirins on the table.
(ii) I have taken 5 aspirins in the last hour.
(iii) There were 5 aspirin-takings in the last hour.
(In support of this interpretation, as pointed out to me by Peyton Todd, the perfect seems
always to involve a temporal span bounded at both ends.)
All three notions here — (a) identity of a quantity, (b) portion-excerpting from that
quantity, (c) configuration of that portion — can be respectively represented by three distinct
NPs together in a construction, as in:
(i) a clustering (c) of a set (b) of trees (a).
Many lexical items conflate the specification of two or all of these notions at once. Thus,
conflating (c) and (b) is a cluster 'a clustering configuration of a set' and a drop 'a small
globular form of an amount [of a liquid]'. A lexical item conflating all three notions is a
tear 'drop of lachrymal fluid'. (See Talmy (1985a) for a general treatment of
"conflation").
For the plural form oysters, the plural form siphons is ambiguous as to whether there are
one or more siphons per oyster. All the other combinations unambiguously indicate the
number of siphons per oyster. Thus, the exemplar form is always unambiguous in this
regard — one of its advantages over the full-complement form. This same arrangement
holds through the list.
The difference between each and every arising in this analysis can now be added to those
observed elsewhere (e.g., Vendler 1968). Each is the exemplar counterpart of the full-
complement expression all the, but not of all without the. Thus, *Each oyster has a siphon
cannot function as a generic assertion. Every is not as unilaterally aligned in this way, but
does serve more naturally as the counterpart of all without the.
One more pair can be added to this list by adjoining two complementary unpaired forms
from two different languages. The English form some, as in some friends of mine, is a full-
complement form requiring the plural and has no exemplar counterpart in the singular.
The corresponding Italian form qualche, as in qualche amico mio, requires the singular
and lacks a plural counterpart.
Though the grammatical complex in (48b) is determinative in setting the role-number as
monadic, a trace of the verb's original scene-division type does remain. In the CR, the
metaphoric suggestion of a dyad is blended in, as if both 'host' and 'guest' are together
present in the "I", perhaps as separate subparts of the single person. For this reason,
(48b) is not the complete semantic equal of (48c). Such blending is, beside shifting,
another major process of reconciliation between incompatible specifications, referred to
in note 11.

THE RELATION OF GRAMMAR TO COGNITION
205
28. Clearly, the language-related faculty of the brain evolved to its present character in the
presence of other already existing cognitive domains, including that of vision, and no
doubt developed in interaction with their mechanisms of functioning, perhaps incorporat
ing some of these.

Where Does Prototypicality Come From?
Dirk Geeraerts
University of Leiden/University of Leuven
1. Hypotheses about the Sources of Prototypicality
Prototype theory is as it were part of the prototypical core of the
cognitive paradigm in semantics, particularly in lexical semantics. I think it is safe
to say that it is by now quite obvious that gradience and salience are among
the linguistically relevant aspects of semantic structure. One need only
recall the early experimental work by Rosch (1973b) and Labov (1973) to
appreciate the importance of graduality and vagueness for the adequate
description of word meaning. But what about explanatory adequacy? Can
we move beyond the descriptive level and explain why prototypicality exists
at all? There are at least four different hypotheses that have been proposed
to explain prototypical phenomena. Each of these hypotheses has been
formulated (or at least hinted at) by Eleanor Rosch herself; this is an
indication that the hypotheses might well be complementary rather than mutually
contradictory. I will call these four hypotheses the physiological, the
referential, the statistical, and the psychological one. Let us have a look at them.
The physiological hypothesis says that prototypicality is the result of
the physiological structure of the perceptual apparatus (Rosch 1973b). This
hypothesis has been formulated with regard to the prototypicality effects in
the domain of colour terms (the first major field in which prototypicality
phenomena have been observed). Particular colours are thought to be focal
because the human eye is more sensitive to certain light frequencies than to
others. The scope of the physiological explanation is probably fairly
limited; it may only be applicable to concepts immediately referring to
perceptual phenomena, or at least to bodily experiences that have a distinct
physiological basis. Since this is most likely not the majority of cases,
additional hypotheses will have to be invoked to explain the prototypical struc-

208
DIRK GEERAERTS
ture of concepts that have no immediate physiological basis.
The referential hypothesis states that prototypicality results from the
fact that some instances of a category share more attributes with other
instances of the category than certain peripheral members of the category
(or share attributes with more other instances than these peripheral cases).
The peripheral applications of a category share attributes with relatively
few other cases, or share only a relatively small number of attributes with
other, more central members of the category. This is the family
resemblance model of prototypicality (Rosch & Mervis 1975); in
psychological terms, it states that the prototypical instances of a category maximize
cue validity. I have dubbed this view 'referential' because it considers
prototypicality to be an automatic consequence of the structure of the range of
application of a concept. Once you know what objects, events etc. a
concept can refer to, you can compute differences in salience by comparing the
number of shared attributes among those things. One might even say that
prototypicality is a secondary phenomenon: it is a side-effect of the mutual
attribute relations among the instances in the referential range of
application of the concept.
Statistical explanations of prototypicality state that the most frequently
experienced member of a category is the prototype. At least, this is the
simple form of the frequency model. It can also be combined with the family
resemblance model; the weight of an attribute within a concept is then not
only determined by its role within the family of applications constituting the
category, but also by the relative frequency with which it is experienced
(Rosch 1975).
The psychological hypothesis is a functional one. It states that it is cog-
nitively advantageous to maximize the conceptual richness of each category
through the incorporation of closely related nuances into a single concept
because this makes the conceptual system more economic. Because of the
maximal conceptual density of each category, the most information can be
provided with the least cognitive effort (Rosch 1977).
In what follows, I would like to show that the functional explanation of
prototypicality is more general than the other ones because it can explain
cases of prototypicality that are counterexamples to the other models. I will
elaborate the psychological hypothesis by indicating some more functional
sources of prototypicality; I will try to make clear that prototypicality is the
outcome of some deep-seated principles of cognitive functioning.

WHERE DOES PROTOTYPICALITY COME FROM? 209
2. A Case Study in Synonymies
Dutch has a pair of synonyms vernielen and vernietigen, which both
roughly mean "to destroy". Though they exhibit some degree of phonetic
similarity, their origin is quite diverse. Vernielen is the older form. It is
already to be found in Middle Dutch, and it is formed by means of the
common verb-forming prefix ver- and the adjective niel, only a few examples of
which survive, but which probably meant "down to the ground". Etymolog-
ically, then, vernielen means "to throw down to the ground, to tear down".
Vernietigen, on the other hand, makes its first appearance in the 16th
century; it is formed by means of the same prefix ver- and the adjective nietig,
which is itself a derivation from the negation particle niet (English not) and
the suffix -ig (which corresponds with English -y). Vernietigen gradually
replaces a third form vernieten, which is a straightforward derivation from
niet with ver-, and which is extinct by the end of the 17th century.
Vernietigen literally means "to annihilate, to bring to naught".
The best way to study both words is to turn to the Woordenboek der
Nederlandsche Taal (hence WNT), the major dictionary of Dutch that
covers the period from 1500 up to 1920 and that, by the way, is still
uncompleted after about a century of editorial work. This dictionary is being
compiled on the basis of a huge corpus of quotations; there are as yet no equally
representative corpora for contemporary Dutch, so that it is rather more
difficult to get an adequate picture of 20th-century usage than it is to study
the semantic history of the vocabulary of Dutch. For the purpose of this
paper, this is not very important: it suffices to pick out one synchronic
period and to see how both words relate to each other in that period. For a
number of practical reasons (among others the amount of available
material), I will concentrate on the 19th century, stretching the temporal borders
of that period with approximately one decade at each end. In this way, a
period from 120 to 130 years can be considered, ranging from roughly 1790
to 1910. In the light of the history of culture, this seems quite justified; we
more or less envisage the cultural period from the French revolution up to
the First World War: the 19th century in the broadest sense.
To get a good picture of the development of vernietigen and vernielen,
it would be necessary to present and discuss the entire articles that I have
compiled for the WNT, and the complete set of quotations on which they
are based. For obvious lack of space, I will only give illustrative quotations;
translations of the quotations are given in the Appendix. The set of quota-

210
DFRK GEERAERTS
tions in the table shows that vernielen and vernietigen can be used
indiscriminately with the same range of application. Each numbered pair of
quotations gives examples of one particular kind of usage. These examples
should be studied from two points of view. On the one hand, the question
has to be asked whether vernielen and vernietigen exhibit any syntagmatic
differences, i.e. differences in their collocational properties. On the other
hand, the question arises whether they are paradigmatically different, i.e.
whether they exhibit purely conceptual differences. (More details on the
analysis of both verbs can be found in Geeraerts 1985a.)
Syntagmatically, we not only see that both words can be used by the
same author in the same context without noticeable differences (as in (1),
(2), (8) and (9)), but also that the range of application of each word can be
divided into three identical major groups, which can moreover be
subdivided along parallel lines. There is a set of applications in which the words
are used with regard to concrete, material objects (1, 2, 3); a set in which
they are used with regard to abstract objects (4, 5), and a set in which they
are used with regard to persons (6, 7, 8, 9). Within the first set, frequently
occurring applications relate to buildings (1), other human artifacts (2), and
natural objects, in particular plants and crops (3). With regard to the
abstract applications, we can distinguish between the annihilation of the
existence of certain abstract objects as such (4), and applications in which
the realisation or fulfillment of certain abstract notions that contain an
aspect of expectation or intention with regard to the future is prevented
(5). With regard to persons, (6) expresses their death as such; (7) and (8)
indicate how someone's bodily or mental health, respectively, are
undermined. (9) expresses how armies are beaten; this application is half-way
between the abstract group (the armies cease to exist as functional entities),
and the personal group (individual soldiers are killed). The existence of
analogous subdivisions within each of the major groups shows that the
syntagmatic equivalence of vernielen and vernietigen is not a coincidence, but
that it is an essential part of their relationship.
Furthermore, the examples also show that there is a paradigmatic,
strictly conceptual equivalence between both: they do not only have the
same collocational properties, but they also seem to express the same
concepts in the same contexts. (The distinction between syntagmatic and
paradigmatic meaning is used here for purposes of analysis only; it does not
imply any particular view with regard to the theoretical relation between

Table 1
i With regard to
1 concrete things
To demolish
buildings or parts
thereof
To destroy other
human artifacts
To destroy
natural objects
(D
(2)
(3)
VERNIELEN
Dat huis was...cvenmin als de naburige
tegen de verwoestendc veeten dier tijd
bestand. Reeds onder den zoon en
opvolgcr des stichtcrs werd hct ... tot
den grond toe vernield (Veegens,
Hist. Stud. 2, 282, 1869).
Er gaat dan stroom op den daarvoor
gevormden zijwcg over, waarbij
genoeg warmtc ontwikkcld wordt om
de draadwindingen in zecr korten tijd
te vernielen (Van Cappclle, Elcctr.
214, 1908).
Hoevcel het wild vernielt wordt door
een Engelschman zeer gocd uitecnge-
zet bij gelcgenheid van een aanval
op de beschcrming die het wild aldaar
...gcnict(V61ksvlijt 1872, 175).
VERMETIGEN
Allcen zijn de vroegcre kruisvensters
door vensterramen van nieuweren trant
vervangen en hebben de vrijhcidsmanncn '
van 1795...hct wapen des stichters in den
voorgevel met ruwc hand vernietigd
(Veegens, Hist. Stud. 1, 125, 1864).
Zonder dcze voorzorg zou het draadjc
door dc enorme hittc van den glocidraad
vernietigd worden (Van Cappclle, Elcctr.
295, 1908). |
Bij het vernietigen van dc onkruiden door
het bewerken dient op hunne voortplan- i
ting en ontwikkcling tc worden gclct
(Reinders, Landb. 1, 309, 1892).

Table 1 - continued
to
I—^
to
With regard to
abstract things
With regard to
persons
To annihilate
existing situations,
characteristics etc.
To prevent the
execution of plans,
hopes, intentions etc.
To kill people, to take
someone's life
(4)
(5)
(6)
Wei wat hamer ! Wordt door zulke
sentimentele zotternyen niet al de
inwendige kragt vernield ? (Wolff en
Deken, Blank. 3, 220, 1789).
De bergstroom in zijn grammen loop
Verscheurt zijn zoom, verdrinkt de
dalen: Alzoo vernielt Gij 's Menschen
hoop! (Ten Kate, Job 53, 1865).
Mij gendenkt ook nog dat Nicolaas
Gaal...mij placht te verhalen...dat
de oude man om deze ontstolen eer
zich zoo ontstelde en vergramde, dat
het ook scheen of hij dezen
dief wel had willen vernielen
(Fruin, Geschr. 1, 1974, 1888).
Stel mij niet zoo hoog, zei ze onthutst, ik
zou daaraan niet beantwoorden; ik zou
uw ideaal vernietigen (Vosmaer, Amaz.
175, 1880).
Zy is dan, van kindsbeen af, opgevoed
on mynheer Daniel's echtgenote te
worden, en nu is die hoop van een geheel
leven vernietigd ! (Conscience, Kwael d.
T. 2, 65, 1859). Dit toeval vernietigde ons
geheele plan (Haafner, Ceilon 103,
1810).
Intusschen heeft de Godin de Natuur
besloten nu voor altijd de Drijvende
Eilanden en al hun inwoners te
vernietigen (Quack, Soc. 1, 246, 1875).

Table 1 - continued
With regard to
persons —
continued
To undermine
someone's physical
health
To undermine
someone's
psychological
well-being
To defeat groups of
armed men, or armies
(7)
(8)
(9)
De beroerte, die haar zwakke '
levenskrachten in een half uur tijds
vernielde, had reeds in het eerste
oogenblik hare spraak verlamd (Beets,
CO. 206, 1840).
Dc vrouwen, Lus, zijn zonen, a( de
anderen blcven stom, vern'teld van
ontsteltenis, op hun stoelen genagcld
(Buyssc, Neef Perscyn 45, 1893).
De uitslag van den stryd was ditmael
hem niet gunstig: geheel zyn leger
werd vernield of uiteen geslagen
(Conscience, Gesch. v. Belgie 110, 1845).
Hy moet rusten. Zulke driften vernietigen
het sterkste gestel (Wolff en Deken,
Leev. 1,290, 1784).
Toen.-.antwoordde zij langzaam met een
doffe stem, als vernietigd door haar
eigene woorden: "Ja, indien het nog
mogelijk is" (Buysse, Mea Culpa 68,
1896).
Het gansch leger der Turken was
i vernietigd ! (Conscience, Gesch. v.
1 Belgie 352, 1845).
3
X
w
w
a
o
O
H
o
>
r
n
o
2
w
O
2

214
DIRK GEERAERTS
both aspects of lexical meaning and particularly with regard to the question
whether selectional restrictions are always an automatic consequence of a
concept's paradigmatic characteristics.) As a preliminary step, notice that
the concept "to destroy" does not only appear as the notion "to annihilate
the existence of someone or something, to cause someone or something to
disappear out of existence", but that it also exhibits the weaker nuance "to
undermine someone or something with regard to some aspect of his
existence" (without a complete destruction or a complete removal out of
existence being implied). The distinction can easily be discovered within the
personal group of applications. In (6), a person is killed, taken out of
existence, while in (8) (and most likely also in the second quotation from (7))
someone's existence is undermined from one point of view or another, but
not entirely annihilated. Likewise, we can see that within the abstract
group, (4) signifies the suppression ofihe existence of some abstract things
as such, whereas in (5), plans, hopes, and expectations are undermined
with regard to their realisation and fulfillment: the plan as such is not
removed (at least not to begin with), but it is reduced to ineffectiveness and
futility. In short, both vernielen and vernietigen express the notions of
complete destruction and partial damage, that is to say, the complete removal
out of existence of something or someone, and the less drastic undermining
in some respect, of the existence of people or objects. According to the syn-
tagmatic context, these notions receive further specifications. For instance,
with regard to persons, complete destruction means killing, but with regard
to concrete things, destruction signifies material demolition, and so on.
(For the sake of completeness it should be added that the equivalence of
vernielen and vernietigen is less straightforward in present-day Dutch than it
is in 19th-century Dutch. Some of the quotations discussed here are now
felt to be rather awkward; in particular, it would be difficult to use vernielen
with regard to persons.)
On the basis of the foregoing observations, one might be tempted to
conclude that the semantic structure of vernielen and vernietigen in 19th-
century Dutch is completely identical: both syntagmatically and paradig-
matically, they have the same range of application. However, a number of
facts testify that both words have different prototypical structures, i.e., that
they have different conceptual centres. There are two sets of facts to be
considered: corpus-based facts relating to the way in which both words are
used in our corpus of quotations, and introspective facts relating to the way
in which the words are perceived by the speakers of the language. In gen-

WHERE DOES PROTOTYPICALITY COME FROM? 215
eral, consideration of these facts will lead to the conclusion that the abstract
applications are central within the structure of vernietigen, and that the
material applications are central in the cases of vernielen. As such, each
verb has a different semantic structure in spite of the fact that the elements
of these structures appear to be the same.
3. Usage as Evidence for Prototypicality
Five observations support the prototypical hypothesis. In the first
place, the abstract group of applications is quantitatively more prominent
within the structure of vernietigen than the material set of applications,
while the reverse is true of vernielen, in which the material group is the
most frequently occurring one. In both cases, the major group is
represented by approximately three times the quotations of the less central
group.
In the second place, the differences in centrality show up in the fact
that the prominent applications exhibit specifications and particular
nuances that they do not have when they are peripheral within the structure
of the lexical item. Thus, the material group of vernielen contains a
metonymical extension of the application with regard to plants and crops,
towards an application in which the fields and gardens where these plants
and crops grow appear as the direct object of the verb. Likewise, the
application with regard to buildings receives a figurative extension towards an
application with regard to an allegorical "wall" that separates two people.
These extensions are probably not impossible within the concrete set of
applications of vernietigen, but the fact that they do not appear there is
statistically interesting: it indicates that the concrete application is more
productive in the case of vernielen than in the case of vernietigen.
Conversely, the abstract group has nuances and additional specifications in the
case of vernietigen that are lacking in the same group with vernielen,
although it is quite easy to imagine that they would in fact occur there. For
example, vernietigen has a fairly large set of applications in which social
movements, institutions, activities and so on are abolished, one quotation
in which it is said that railway transport destroys distances (obviously,
distances do not disappear as such, they are only functionally overcome), and
one quotation in which a philosopher is said to destroy the soul (again, the
soul is not destroyed as such, but the idea that the soul exists is metonymi-
cally abolished by the philosopher in question). None of these extensions of

216
DIRK GEERAERTS
the abstract use of the concept "to destroy" can be found in the case of
vernielen, which is indicative of the fact that the abstract use is less prominent
in the latter verb than in the semantic structure of vernietigen.
In the third place, the salience of the material kind of usage can be
derived indirectly from the nominalisations of both verbs. Both vernietiging
and vernieling have the verbal sense "the fact, the act or the process of
destroying or being destroyed", but only vernieling exhibits the metonymi-
cal extension towards the concept expressing the result of that process or
that act, i.e., the concrete damage that issues from it. (In the latter case,
the word is typically used in the plural: vernielingen more or less equals the
notion "damage".)
In the fourth place, the internal structure of the set of personal
applications reflects the differences in prototypical structure between both verbs.
To begin with, notice that the personal group contains concrete as well as
abstract applications; to kill someone is clearly more concrete than to
undermine someone's psychological well-being or his social position. If we
then have a look at the mutual relationship between the abstract and the
concrete subgroups of the application with regard to persons, we find that
the abstract subgroup is proportionally dominant in the case of vernietigen,
whereas the reverse is true in the case of vernielen. Also, we find that
extensions of the concrete subgroup of the personal application with regard
to other living beings than people or with regard to personifications, are not
as strongly present in the case of vernietigen than in the case of vernielen. (It
should be added that these observations have to be considered with more
care than the previously mentioned points, since there is a general tendency
throughout the centuries covered by the WTVr-material, to remove the
personal application from the structure of vernielen. There are relatively less
personal applications in the structure of 19th-century vernielen than in the
structure of either 19th-century vernietigen or 16th-century vernielen; as has
already been mentioned, it is even more difficult to use vernielen with
regard to persons in present-day Dutch. In any case, the 19th-century
material does seem to show that the material subgroup of the personal
application of vernielen is more resistant to the tendency in question than the
abstract subgroup, as can be predicted from our centrality hypothesis.)
Finally, the importance of prototypicality can be derived from the fact
that different nuances play a central role within the core of each concept,
whereas those nuances are not particularly important within the
corresponding group in the other concept. Thus, the destruction of buildings and

WHERE DOES PROTOTYPICALITY COME FROM? 217
other human constructions is prominent within the material use of
vernielen, but is only rarely present within the material group of vernietigen.
Within the structure vernielen as a whole, demolishing buildings is the
single most frequently represented kind of usage, but within the structure
of vernietigen, it is merely one among many equally important nuances of
the material set of applications. In the same way, the central, abstract
group within the structure of vernietigen is itself centred round applications
relating to the dissolution, the cancellation, the annulment of agreements,
commitments, engagements, obligations, permissions, rights, and so on,
and of the laws, orders, contracts etc. in which they are contained and
through which they come into existence. Whereas vernielen only rarely
exhibits this kind of usage, it is the most frequently occurring sense within
the abstract group of vernietigen as well as within that word as a whole.
In general, these facts of linguistic usage clearly favour the hypothesis
that the abstract applications of the concept "to destroy" are prototypical
within the structure of vernietigen, whereas the concrete applications are
prominent in the case of vernielen. Taking into account that each central
group is itself concentrated round a dominant kind of usage, it seems
plausible to say that the latter is the prototypical sense for each of the verbs in
question. It should furthermore be noted that these prototypical
phenomena seem to be connected with the etymology of the words. On the
one hand, the abstract prototype of vernietigen may well be connected with
the abstract character of the words niet "not", and nietig "null and void,
insignificant", on which it is based. Moreover, the common phrase nietig
verklaren "to declare something to be null and void, dissolve, annul
something" corresponds pretty closely with the central notion within the abstract
group of vernietigen. On the other hand, the centrality of the application
with regard to buildings in the structure of vernielen seems to correspond
with the etymological meaning "to tear down, to throw to the ground" that
we reconstructed above as the original meaning of the verb.
4. Introspective Evidence of Prototypicality
Before we can deal with the introspective evidence in favour of the
prototypicality hypothesis, two preliminary questions have to be answered.
In the first place, how trustworthy is the introspective methodology? The
paradoxical fact of the matter is that it is exactly the unreliability of
introspection that makes it interesting for our purposes. If introspection were

218
DIRK GEERAERTS
able to yield a completely adequate picture of the facts of linguistic usage
(which is doubtful), it would simply reduplicate the results reached in the
previous paragraphs on the basis of a direct examination of linguistic usage.
But given the presupposition that introspection yields only a partial insight
into the semantic structure of the words that are investigated, we can also
presuppose that it will be exactly the prototypical kinds of usage of those
words, that reach the introspective consciousness of the language user. We
can use the results of the introspective method as support for the
prototypical hypothesis if we presuppose that prototypical kinds of usage (precisely
because they are more salient than other applications) will more easily pass
the threshold of conscious attention. Given this presupposition, the
introspective judgements of native speakers may shed light on the question which
kinds of usage are predominant within a certain concept.
In the second place, how can the introspective method be used with
regard to historical material? There are no 19th-century speakers of Dutch
around to be asked what they think is the meaning of particular words, so
how are we going to get introspective judgements at all? The fact is that we
do have information on how the 19th-century speakers of Dutch perceived
the near-synonyms that we are investigating, viz. in the form of synonym
dictionaries. Synonym dictionaries (at least the older ones) are notoriously
unreliable as descriptions of actual patterns of usage; most of the time, the
compilers of synonym dictionaries rationalise away the actual identity of
words by imposing distinctions that cannot be discovered in the actual facts
of usage. However, these rationalisations need not always have proceeded
out of the blue: it seems quite plausible that they were guided by the
introspective judgements of the compilers. So, if we like to know something of
the introspective insights of the 19th-century speakers of Dutch, we can
have a look at the synonym dictionaries of that time to see whether the
distinctions they make between vernielen and vernietigen (however inadequate
as a picture of the complete set of possible kinds of usage) do indeed reflect
the differences in prototypical structure of both words.
And indeed, the 19th-century synonym dictionaries of Dutch do
distinguish between vernielen and vernietigen along lines that fit into our
hypothesis. On the one hand, there are those that draw the line syntagmat-
ically, such as Weiland & Landre (1825), who state that vernielen can only
be used with regard to "lighamelijke dingen" (material things), whereas
vernietigen is more widely used, in particular also with regard to
"menschelijke instellingen" (human institutions). De Beer (1897) expresses

WHERE DOES PROTOTYPICALITY COME FROM? 219
an analogous point of view. On the other hand, there are those that
describe the distinction along paradigmatic lines, so that there would be an
actual notional difference between the verbs in question, rather than
merely a distinction in selectional restrictions. Whereas vernietigen is
defined as "to bring to naught, to annihilate", vernielen is defined as "to
damage, to smash to pieces, to tear down". In this sense, vernietigen implies
a complete annihilation whereas there may be some pieces left of the
original object in the case of vernielen. It is easy to see that this paradigmatic
point of view, which can be found among others in Pluim (1894), is
connected with the previous, syntagmatic one: it is precisely because vernielen
relates to material things that the notion of remaining debris comes to the
fore. Likewise, a complete annihilation (in which the original objects
disappear completely) is less likely in the material world of concrete objects, so
that the restriction of vernietigen to abstract objects will tend to be related
to the notion of complete annihilation. This is in fact done by Weiland &
Landre (1825), though not all proponents of the paradigmatic distinction
adhere to the syntagmatic distinction. For instance, De Flines (1810)
mentions that vernietigen can in fact be used with regard to material objects, but
that there is a difference with vernielen in the degree of damage achieved.
By and large, these views faithfully reflect the insight into the
prototypes of vernielen and vernietigen that we have gained by considering the
actual facts of linguistic usage. Syntagmatically, it is recognised that the
material context is more important for vernielen, whereas abstract objects
are predominant in the case of vernietigen. Paradigmatically, this is
reflected by the fact that vernielen carries overtones of material destruction
and damage (think of the relationship between the prototypical usage of
vernielen with regard to buildings, and the definitions of that word that
bring to the fore the act of smashing and demolishing things), whereas
vernietigen calls forth the idea of complete annihilation (as it were, wiping
something off the face of the earth). As such, the stubborn efforts of the
compilers of synonym dictionaries to find semantic differences among near-
synonyms seem to be not entirely gratuitous. To the extent that they try to
capture the characteristics of the most salient kinds of usage of both lexical
concepts, they strenghten our own hypothesis about the differences in
prototypical structure among the verbs.

220
DIRK GEERAERTS
5. The Functional Explanation of Prototypicality
There are a number of interesting conclusions to be derived from the
above analysis of the near-synonyms vernielen and vernietigen. First,
prototypicality is an interesting new point of view in the study of synonyms. It
is traditionally well-known in lexical semantics that there are relatively few
true synonyms in natural languages, and the ways in which near-synonyms
differ can be very diverse. Our discussion of vernielen and vernietigen shows
that there is one more factor to be added to the list of differentiating
factors: near-synonyms may be distinct with regard to the prototypical
structure imposed on an otherwise identical range of application. Once again,
the importance of prototype theory for the traditional concerns of lexical
semantics becomes apparent (cf. Geeraerts 1983a, 1983b, 1984, 1985b);
prototype theory opens up new perspectives in the study of synonyms.
Secondly, there are some indications that introspective judgements in
lexical semantics relate to the prototypically salient instances of concepts
rather than to the full range of actual usage possibilities. If this can be
confirmed by additional comparisons between introspective perceptions of
lexical meanings and actual usage patterns, more will be known about the
value of both methodologies (introspective and corpus-based) in lexical
semantics. Also, if we maintain the classical view of modern linguistics that
it is one of the goals of linguistic theory to account for the introspective
judgements of native speakers, and if these judgements appear to be
influenced by prototypical phenomena, yet one more reason presents itself for
incorporating prototype theory into lexical semantics.
Thirdly, the fact that vernielen and vernietigen have the same
conceptual and collocational range of application, and yet differ with regard to the
core and the periphery of their categorial structure, indicates that there are
at least some cases of prototypicality that cannot be explained by means of
the referential model. Vernielen and vernietigen refer to the same set of acts
and processes; as such, the differences in their prototypical structure cannot
be the automatic consequence of their referential range, as is implied by the
family resemblance hypothesis. In addition, the physiological and the
statistical explanation will not be of much avail either. There is no particular
organ or mechanism for the perception of processes of destruction, and
even if there were, we would still need two different physiological structures
to explain the distinction between both verbs, which is beyond all intuitive
plausibility. The statistical explanation is inapplicable for the same reason

WHERE DOES PROTOTYPICALITY COME FROM? 221
as the referential hypothesis: since the range of application of both verbs is
the same, the frequency of occurrence of the processes referred to is the
same for both verbs. That is to say, the frequency with which the
demolishing of buildings occurs in reality, relative to the frequency with which, say,
agreements are cancelled in reality, has exactly the same effect on both
verbs, since these refer to the same objective reality. Because they denote
the same things in reality, the structure of reality (either with regard to the
frequency of occurrence of its elements, or with regard to the mutual
resemblances among those elements) cannot be invoked to explain the
distinction in semantic structure between vernielen and vernietigen.
In short, we can reject all materialistic explanations of the
prototypicality effects observed in the verbs under consideration. Indeed, the
physiological, the referential, and the statistical hypotheses have this in
common: that they try to explain prototypicality on the basis of
materialistic data, either the material structure of the human perceptual apparatus, or
the material characteristics (statistical or otherwise) of the referential range
of the concepts involved. Given that we have to reject these materialistic
hypotheses, we can provisionally choose, by elimination, for the
psychological, functional explanation of prototypicality. To support this choice, I
would like to make clear that the functional hypothesis has some additional
advantages, besides the fact that it avoids the problem of the materialistic
hypotheses. First, however, three remarks have to be made.
To begin with, it might be claimed that a statistical explanation of the
prototypicality effects in vernielen and vernietigen can indeed be given, if
we take into account, e.g., that the material sense occurs much more
frequently with vernielen than with vernietigen, or that the abstract
specification of the notion "to destroy" is statistically much more prominent in the
latter verb than in the former. However, the frequencies that are
mentioned here are linguistic frequencies, not referential frequencies, i.e., they
are frequencies of occurrence of words, not of the things those words refer
to. Because the frequency at stake here is linguistic rather than referential,
it can hardly be invoked to explain prototypicality; as an aspect of linguistic
usage, it is one of the things we have to explain, not one of the things that
are themselves part of the explanation. We can use linguistic frequencies to
determine what instances of a concept are prototypical (that is what we did
in section 3), but explaining prototypicality on the basis of linguistic
frequency is putting the cart before the horse. Some kinds of usage are not
prototypical because they are more frequent; they are more frequent

222
DIRK GEERAERTS
because they are prototypical. The apple is not a prototypical fruit because
we talk more about apples than about mangoes, but because we experience
apples more often than we encounter mangoes (and this fact, in turn, may
be the reason why we talk more about apples). Frequency of linguistic
occurrence may be a heuristic tool in the pinpointing of prototypes, but it is
not the source of prototypicality as meant in the statistical hypothesis.
The second remark has to do with the fact that criticism with regard to
the referential, family resemblance model of prototypicality has already
been formulated elsewhere. This has been the case in the work of Pulman
(1983) and — in more stringent fashion — in the well-known article by
Armstrong et al. (1983). They argue that gradience can be observed in
concepts with rigid boundaries (their examples relate to natural numbers), so
that family resemblances cannot be invoked to explain the differences in
salience among numbers. There are two reasons, however, why their
argumentation is less relevant than they assume. First of all, they more or
less equate prototype theory and the family resemblance model of the
sources of prototypicality, whereas it is quite clear that the family
resemblance model is merely one of a number of hypotheses concerning the
sources of prototypicality: ruling out one hypothesis does not mean that
one can ignore the others. And also, I do not think that Armstrong et al.
are successful in presenting a counter-example to the family resemblance
model. Even if a concept has rigidly defining characteristics, family
resemblances may exist among the non-defining characteristics of the
instances of that category. Since cognitive semantics is basically
encyclopaedist in its approach, these non-defining, "encyclopaedic" attributes
should be incorporated into the computation of degrees of shared
attributes. As Lakoff (1982a) has shown, such encyclopaedic, experiental
factors do indeed occur with regard to numbers, and they can be used to
explain the prototypicality ratings found by Armstrong et al.
My third remark is this: my criticism of the materialistic hypotheses
should not be overgeneralised. The fact that they do not work in the case of
vernielen and vernietigen clearly does not imply that they do not work in any
case, but merely makes clear that next to the physiological, the referential,
and the statistical model, there will have to be at least one other source of
prototypicality.
Let us now come back to the functional model of prototypicality and
try to elaborate it. Remember that the psychological hypothesis involves
requirements that the cognitive system is to comply with if it is to function

WHERE DOES PROTOTYPICALITY COME FROM? 223
efficiently: prototypicality exists because it is cognitively advantageous. As
we have seen, Rosch has specified this functional advantage in terms of the
economical effect of informational density; prototypical categories enable
one to reach the most information with the least cognitive effort. This
functional line of reasoning can be supplemented with some additional (and
perhaps even more fundamental) functional reasons for having prototypical
categories. We can base the discussion on one of the fundamental insights
of cognitive psychology, viz. that cognition should combine structural
stability with flexible adaptability. On the one hand, cognition should have a
tendency towards structural stability: the categorial system can only work
efficiently if it can maintain its overall organisation for some time, if it does
not change fundamentally any time new information has to be
incorporated. At the same time, however, it should be flexible enough to be easily
adaptable to changing circumstances. To prevent it from becoming chaotic,
it should have a built-in tendency towards structural stability, but this
stability should not become rigidity, lest the system stops being able to adapt
itself to new and unforeseen circumstances.
This necessity of flexibility is one of the aspects of lexical semantics
that was recognized by the prestructuralist tradition of historical semantics,
but that has been more or less lost in the meantime, as a result of the
structuralist attention for fixed synchronic structures. Be that as it may, it will be
clear that prototypically organised categories are particularly well suited to
fulfil the double demand for flexible adaptability and structural stability. On
the one hand, the fact that slightly deviant nuances can be developed within
a particular category indicates that categories have the dynamic ability to
cope with changing conditions and changing expressive needs. On the other
hand, the same fact (that marginally deviant concepts can be incorporated
into existing categories as peripheral instances of the latter) proves that
these categories have a tendency to maintain themselves as holistic entities,
thus maintaining the overall structure of the categorial system. Prototypical
categories maintain themselves by adapting themselves to changing
circumstances and new expressive needs; at the same time, they function as
expectational patterns with regard to reality: new facts are interpreted in
terms of information that is already at the disposal of the individual. The
flexibility of the cognitive system does not only show up in the fact that it
can adapt itself to new experiences, but this flexibility is supplemented with
the fact that existing categories have a formative influence with regard to
experience; new experiences are fitted into the expectational patterns pro-

224
DIRK GEERAERTS
vided by the existing categorial system.
Along these lines, prototypicality appears to be the outcome of some
fundamental, deep-seated principles of cognitive functioning. The form of
the conceptual system appears to be determined by a set of basic functional
requirements, and prototypically structured concepts admirably meet these
requirements. If this is correct, the same basic principles should also have a
role to play in other cognitive disciplines. That is to say, if prototypicality is
an emanation of some basic characteristics of all cognition, we should be
able to find analogies of the prototypical idea in other fields of cognitive
science, next to lexical semantics. I have tried to prove at length elsewhere
(1985b) that this is in fact the case: the importance of interpretative
schemata mediating between experience and existing knowledge is an idea
that can be traced in a number of cognitive disciplines. It is very much
apparent in Artificial Intelligence (Minsky's frame notion); it can be found
in cognitive psychology, particularly in the work of Bruner, and to some
extent in that of Piaget; it can be related to some of the views of the early,
Husserlian phenomenological movement in philosophical epistemology;
and it has some important similarities with the paradigmatic conception of
scientific enquiry inaugurated by Thomas Kuhn.
These are exciting parallels because they suggest that the functional,
psychological hypothesis concerning the sources of prototypicality can at
the same time be the basis for a truly integrated cognitive science in which
the insights of linguistics, Artificial Intelligence, cognitive psychology,
philosophical epistemology, and the philosophy of science can be brought
together under a common denominator. In this respect, the functional
model of prototypicality, even if it does not rule out the possible
importance of the physiological, the referential, or the statistical explanation,
does seem to be more general than the latter, not just because it is based on
fundamental principles of cognition, but also because similar views have
been put forward in other branches of cognitive science.
6. Onomasiological and Semasiological Aspects of Cognitive Semantics
Unfortunately, the optimistic perspective of the previous paragraph
does not solve everything. To round off the discussion, I would like to show
that a complete explanation of all questions to be raised with regard to ver-
nielen and vernietigen is far from available. The picture we have
reconstructed so far looks like this: apparently, the linguistic community at some

WHERE DOES PROTOTYPICALITY COME FROM? 225
point in its development finds it convenient to have two distinct categories
for the concepts of material and abstract destruction. Thus, a pair of
etymologically distinct words becomes available, originally vernielen and
vernieten, later on vernielen and vernietigen', their compound character
ensures that one of them signifies material destruction, the other abstract
annihilation. Gradually, the flexibility that is inherent in all human
categorisation extends these concepts beyond their etymological usage; as a
result, they have the same range of application in the 19th century.
But now consider the original situation in which these flexible
extensions have hardly begun taking place. Is it then not irrational to use
vernielen to express abstract annihilation, when you already have vernietigen
or vernielen to do so? The question can be put in terms of global and local
efficiency. As we have argued, the global efficiency of the conceptual
system commands its flexible, prototypical organisation. But there is, in the
case of vernielen and vernietigen, also a local efficiency principle that says
that it is uneconomic to have two terms expressing the same things. We are
then forced to ask: why does not the local efficiency principle stop the
application of the global principle? Why is not the prototypical extension of
vernielen towards abstract forms of destruction checked or prevented by the
consideration that you already have a lexical category expressing abstract
destruction? There is yet another way of formulating the problem:
prototype-based flexibility is necessary because of the expressive needs of the
speaker: he may want to express concepts for which no specific term is
available. But why then would he use these flexible mechanisms of
semantic extension if such a specific term is indeed available?
One kind of answer might simply be that the global principle is
stronger than the local principle; the global principle simply supersedes the
local principle to the extent that local inefficiencies are created. We are
then saying that the global principle is so general that its strength overrules
the local principle, and that it applies even where it is not strictly necessary.
Still, this does not tell us why the local principle is weaker than the global
principle. Also, it is rather awkward to explain a mechanism that is unfunc-
tionally overproductive on the basis of functional considerations. Couldn't
we therefore find a more rational explanation of the flexible extensions?
The way out, as far as I can see, is to take into account other kinds of
expressivity than the purely conceptual one. Using vernielen to express a
concept that is commonly expressed by vernietigen may be conceptually
superfluous, but that does not mean that doing so may not serve particular

226
DIRK GEERAERTS
expressive purposes. On the level of the linguistic form, for example, it may
be quite functional to use another word than the usual one. The varieties of
such a formally expressive synonymy are well-known in traditional lexical
semantics; near-synonyms may exhibit connotational and emotional
differences (as in euphemisms), stylistic differences (as in popular words versus
poetic terms), or sociolinguistic differences (as in learned words versus
common words). Perhaps we can even say that speakers have an urge for
stylistic variation as such, even if the formal variants do not carry specific
overtones; variation may well be governed by a straightforward desire to
avoid monotony, to create new ways of expressing oneself, to experiment
with unexpected innovations as a way of stressing one's own individuality.
Moreover, it may well be that the importance of metaphor in natural
language is determined precisely by its stylistic expressivity; metaphorical
expressions would then be created primarily to add expressive weight to the
message one wants to convey. (See the contribution of B. Rudzka-Ostyn to
this volume.)
It is quite plausible, then, that factors such as these have governed the
extension of vernielen and vernietigen beyond their original meanings and
into each other's etymological range of application. For instance, using
vernielen to express a process of abstract cancellation may have been
stylistically particularly expressive, because the process of material destruction
normally denoted by vernielen carried overtones of physical violence that
were less marked in the case of vernietigen. The extended use of vernielen
would then have been a case of metaphorical hyperbole. It is, however,
very difficult to pinpoint exactly which form of expressivity is the relevant
one with regard to the two verbs that we are concerned with here; our
historical material for the earliest (Middle Dutch) history of vernielen and
vernielen, for instance, is very hard to interpret with regard to such questions.
Still, some clear cases may in fact be found. For instance, the first quotation
of (5) in Table 1 clearly carries more overtones of violence, force, and
intensity than the second quotation in that pair of examples. (This is mainly
made apparent by the presence of a simile, marked by alzoo.) This suggests
that the verbs highlight slightly different aspects of the situation described,
or rather, represent the situation from different points of view (determined
by the prototypical core of each verb). (In Langacker's terminology, the
distinction between the two verbs, when used with regard to the same
process, might then be characterized as a figure/ground-distinction: vernielen
takes the violent process as figure, and vernietigen the destructive result.)

WHERE DOES PROTOTYPICALITY COME FROM? 227
This is not an altogether implausible hypothesis, but it is unfortunately hard
to confirm for the simple reason that the historical texts used here do not
give us enough clues to discern such subtle differences in stylistic or
emotional overtones.
On the whole, then, what can we conclude from our discussion of this
additional problem? On the one hand, it inspires caution with regard to our
attempts to explain prototypical phenomena: the linguistic materials at our
disposal do not always allow completely satisfactory answers with regard to
the questions at stake to be formulated. On the other hand (and this is, I
think, the more important conclusion), the discussion suggests that
prototype formation may be influenced by other factors than purely conceptual
ones. Stretching the meaning of a lexical item may be motivated by the
desire to use another form than the one that is usual to express the idea in
question; stylistic, sociolinguistic, connotational expressivity rather than
purely conceptual needs may determine the flexible use of a category. In
such a case, the conceptual coherence of the prototypically structured
category (i.e., the fact that the new, peripheral kinds of usage have to be
accessible from the prototypical core) constitutes a limit to the desire for formal
variation: you can use a particular lexical item to express an idea that is
usually signified by another word, but only on the condition that the idea in
question is part of the prototypical potentialities of that lexical item.
Basically, you stretch an item's meaning to express something conceptually
new, but you can also stretch it to express something conceptually old in a
formally new way. This is a very important suggestion, because it implies a
warning against a tendency that is a natural characteristic of cognitive
semantics: the tendency, in fact, to look for purely cognitive or conceptual
explanations of the facts one encounters. Taking the cognitive,
experiential, encyclopaedic nature of linguistic signs seriously should not imply
looking only for strictly conceptual explanations. Language is not just content:
it is also form, and its formal side has an expressivity of its own, which does
seem to create lexical configurations that can hardly be explained if we only
take into account the conceptual expressivity of language.
In the traditional terms of lexical semantics, this means that the
explanation of prototypicality should not restrict itself to the semasiological
perspective (in which each category is c