Editorial page
Title page
Copyright page
Table of contents
Meaning and Semantic Structure
The Nature of a Grammar
Grammar as a Symbolic Phenomenon
Distribution and Predictability
Arbitrary Grammatical Markings
Grammaticality Judgments
A View of Linguistic Semantics
Cognitive Domains
Profile and Base
Scale and Scope of Predication
Relative Salience of Substructures
The Nature of Grammatical Valence
The Problem
Grammatical Valence: Canonical Instances
Non-Canonical Valence Relations
Further Departures from Canonical Valence Relations
Scope and Morphological Layering
A Usage-Based Model
The Network Conception
General Applicability
1. The Nature of Grammatically Specified Notions
2. Categories of Grammatically Specified Notions
2.2 Plexity
2.3 State of Boundedness
2.4 State of Dividedness
2.5 The Disposition of a Quantity: A System of Categories
2.6 Degree of Extension
2.7 Pattern of Distribution
2.8 Axiality.
2.9 Perspectival Mode
2.10 Level of Synthesis
2.11 Level of Exemplarity
2.12 Other Categories and Processes
2.13 Nesting: An Interaction of Categories
2.14 Four Imaging Systems
3. Further Cognitive Connections
Where Does Prototypicality Come From?
2. A Case Study in Synonymics
3. Usage as Evidence for Prototypicality
4. Introspective Evidence of Prototypicality
5. The Functional Explanation of Prototypicality
6. Onomasiological and Semasiological Aspects of Cognitive Semantics
Appendix. A translation of the quotations.
The Natural Category MEDIUM: An Alternative to Selection Restrictionsand Similar Constructs
2. Co-occurrence and Categorization
2.1.2 Ascription Features
2.1.3 Transfer Features
2.2 Toward a Viable Theory of Categorization in the Lexicon
3. The Natural Category MEDIUM
3.2 The Categorial Periphery of MEDIUM
4. Conclusions and Inconclusions
Spatial Expressions and the Plasticity of Meaning
2. Insufficiencies of the \
2.2. Geometric Conceptualization
2.3. Synecdoche
2.4. Divergence from the Simple Relations
2.5. Added Constraints
2.6. Unexplained Context Dependencies
2.7. Unexplained Restrictions
3. Meaning and Use of Spatial Prepositions
3.2. Geometric Descriptions
3.3. Pragmatic Near-Principles
3.4. Use Types
4. Discussion
Contrasting Prepositional Categories: English and Italian
1. Prepositions and Polysemy
2. The Nature of Prepositional Meaning
3. The Prepositions
4. Conclusions
The Mapping of Elements of Cognitive Space onto Grammatical Relations: An Example from Russian Verbal Prefixation
Goto 337 /FitH 555.1 The Role of the Prefix and How It Is Captioned
1.0 The pere-Network
1.1 Configuration 1
1.2 Configuration 2
1.3 Configuration 3, <Interchange>
1.4 Configuration 4,  <Division>
1.5 Configuration 5,  <Thorough>
1.6 Configuration 6,  <Over>
1.7 Configuration 7,  <Bend>
1.8 Configuration 8,  <Turn>
1.9 Configuration 9,  <Mix>
2.0 The Cognitive Model in Semantic Description: A Summary
3.0 Mapping the Cognitive Configuration onto the Verbal Arguments
Conventionalization of Cora Locationals
2.0 Conventional Morphology
3.0 Choice of Verbal Prefixes
3.1 The Tree Trunk
3.2 All Burned Up
3.3 Down to the Ground
3.4 Bottles, Sticks and Pots
3.5 On the Edge of Things
3.6 High in the Sky
3.7 Comparative Anatomy
3.8 A Trip into Time
4.0 Conclusion
The Conceptualisation of Vertical Space in English: The Case of Tall
2. English Adjectives Denoting Vertical Space
2.2 High and Low
2.3 Tall
3. Collocational Preferences of Tall
4. The Status of Collocation
5. The Concept of TALLNESS
Length, Width, and Potential Passing
2. The Distribution of Length and Width
2.2 The Length and Width of Mobile Entities
2.3 The Length and Width of Immobile Multidimensional Entities
2.4 Metric Definition of Length/Width
2.5 Constraints
3. Synchronic Interpretation of the Distribution of Length and Width
3.2 Length/Width and Potential Passing
4. Interpretation of Length/Width in Logical Diachrony
5. Discrepancies between Logical and Historical Time. Evidence fromFrench
6. Conclusion
On Bounding in Ik
2. Change, Boundedness and Directionality
2.2 Directionals with Stative Verbs
2.3 Directionals with Other Verbs
3. Deictic Center with Other than Motion Verbs
4. Summary and Perspective
A Discourse Perspective on Tense and Aspect in Standard Modern Greek and English
1. Tense and Aspect in SMG: Establishing the Relevant Categories
2. The Semantics of Perfective and Imperfective
3.1 The Psychosemantic Basis of Aspect
3.3 The Iconicity Principle, Aspect and Tense
4.1 Packaging and Grinding with a Discourse Purpose in a Narrative
4.3 Tense and Discourse Representations
4.4 Aspectual Distinctions of English: Simple Tense versus Progressive Aspect
4.5 Progressives and Imperfectives
4.6 Packaging and Grinding in an English Text and Its SMG Translation
5. Conclusion
Appendix 1
Appendix 2
Semantic Extensions into the Domain of Verbal Communication
2. Underlying Concepts and Assumptions
3. Verbal Communication: Its Multi-Level Components and Its Donor Domains
4. Spatial Motion
4.2. The Components of Motion Events and Their Linguistic Expression
5. Extensions of Spatial Motion
5.2. Phrasal Verbs
5.2.2. Including a SURFACE Landmark
5.2.3. Including a POINT Landmark
5.2.4. Including a CHANNEL Landmark
5.2.5. Including an Agentive Co-Trajector or an Instrumental Trajector
5.3. Non-Phrasal Verbs
5.4. Major Patterns of Cross-Domain Correspondences
5.5. Domain Intersections
6. Conclusion
Spatial Metaphor in German Causative Constructions
2. Causative vs. Non-causative
3. Prototypical Causation
4. Events, Agents, Patients, and Instruments
5. Foregrounding and Syntactic Marking
6. Pedagogic Grammars vs. Actual Usage
7. A Survey of Prepositional Usage in German Causative Constructions
8. Prototype Combinations
9. Rare or Difficult Combinations
10. Non-prototypical Combinations
11. Causatives vs. Non-causative Transitives
12. Spatial Metaphor in Case Grammar
13. A Localist Basis for von, mit, and durch
14. A Cognitive-Spatial Account of German Causative Constructions
15. Conclusion
Náhuatl Causative/Applicatives in Cognitive Grammar
1. Preliminaries
1.2 Causation
2. Causatives
3. Applicatives
4. The Causative/Applicative Schema
5. Cases Intermediate between Causatives and Applicatives
5.2 Multiple and Mixed Construals
6. Verbalizing Usages
6.2. Verbalizations of Nouns
7. Summary and Conclusion
1. Language and Cognition: Aristotle's Approach
2. Mediaeval Speculative Grammar
3. Conclusion
Cognitive Grammar and the History of Lexical Semantics
2. Historical-Philological Semantics
3. Cognitive Semantics
4. Comparing the Cognitive and the Historical-Philological Tradition
5. A Survey of the History of Lexical Semantics
6. Main Lines in the History of Lexical Semantics
7. Conclusion
Subject Index
The series Current Issues in Linguistic Theory

Author: Rudzka-Ostyn Brygida  

Tags: linguistics   english language  

ISBN: 0304-0763

Year: 1988

                    Cur nt Issues in Linguisti Theor o
Br> n'dj Rud ka- si n

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
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
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
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
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.
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 considered on its own), but that the onomasiological point of view (in which it is studied how several items may express similar or identical concepts) should be taken into account as well. Conceptual expressivity is basically a factor connected with the semasiolog-
228 DIRK GEERAERTS ical explanation of prototypicality, whereas the onomasiological influences on prototype formation seem to refer to other kinds of expressivity, as was suggested by our study of vernielen and vernietigen. The incorporation of the onomasiological approach does not mean that cognitive semantics moves away from the functional perspective advocated in the previous secti