Corpus Methods for Semantics. Quantitative studies in polysemy and synonymy

Corpus Methods for Semantics Human Cognitive Processing (HCP) Cognitive Foundations of Language Structure and Use Th...

Author: Glynn D. | Robinson J. A. (eds.)

120 downloads 1115 Views 3MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form

DOWNLOAD PDF

Corpus Methods for Semantics

Human Cognitive Processing (HCP)

Cognitive Foundations of Language Structure and Use This book series is a forum for interdisciplinary research on the grammatical structure, semantic organization, and communicative function of language(s), and their anchoring in human cognitive faculties. For an overview of all books published in this series, please see http://benjamins.com/catalog/hcp

Editors Klaus-Uwe Panther

Nanjing Normal University & University of Hamburg

Linda L. Thornburg

Nanjing Normal University

Editorial Board Bogusław Bierwiaczonek

Jan Dlugosz University, Czestochowa, Poland / Higher School of Labour Safety Management, Katowice

Mario Brdar

Josip Juraj Strossmayer University, Croatia

Barbara Dancygier

University of British Columbia

N.J. Enfield

Max Planck Institute for Psycholinguistics, Nijmegen & Radboud University Nijmegen

Elisabeth Engberg-Pedersen University of Copenhagen

Ad Foolen

Radboud University Nijmegen

Raymond W. Gibbs, Jr.

University of California at Santa Cruz

Rachel Giora

Tel Aviv University

Elżbieta Górska

University of Warsaw

Martin Hilpert

University of Neuchâtel

Zoltán Kövecses

Eötvös Loránd University, Hungary

Teenie Matlock

University of California at Merced

Carita Paradis

Lund University

Günter Radden

University of Hamburg

Francisco José Ruiz de Mendoza Ibáñez University of La Rioja

Doris Schönefeld

University of Leipzig

Debra Ziegeler

University of Paris III

Volume 43 Corpus Methods for Semantics. Quantitative studies in polysemy and synonymy Edited by Dylan Glynn and Justyna A. Robinson

Corpus Methods for Semantics Quantitative studies in polysemy and synonymy Edited by

Dylan Glynn University of Paris VIII

Justyna A. Robinson University of Sussex

John Benjamins Publishing Company Amsterdamâ•›/â•›Philadelphia

8

TM

The paper used in this publication meets the minimum requirements of theâ•¯American National Standard for Information Sciences – Permanence of Paper for Printed Library Materials, ansi z39.48-1984.

Library of Congress Cataloging-in-Publication Data Corpus Methods for Semantics : Quantitative studies in polysemy and synonymy / Edited by Dylan Glynn and Justyna A. Robinson. p. cm. (Human Cognitive Processing, issn 1387-6724 ; v. 43) Includes bibliographical references and index. 1. Semantics. 2. Cognitive grammar. 3. Computational linguistics. 4. Polysemy. 5. Corpora (Linguistics) I. Glynn, Dylan. II. Robinson, Justyna A. P325.C595 2014 401’.43--dc23 isbn 978 90 272 2397 5 (Hb ; alk. paper) isbn 978 90 272 7033 7 (Eb)

2014004752

© 2014 – John Benjamins B.V. No part of this book may be reproduced in any form, by print, photoprint, microfilm, or any other means, without written permission from the publisher. John Benjamins Publishing Co. · P.O. Box 36224 · 1020 me Amsterdam · The Netherlands John Benjamins North America · P.O. Box 27519 · Philadelphia pa 19118-0519 · usa

Table of contents

Contributors Outline

vii 1

Section 1.â•‡ Polysemy and synonymy Polysemy and synonymy: Cognitive theory and corpus method Dylan Glynn Competing ‘transfer’ constructions in Dutch: The case of ont-verbs Martine Delorge, Koen Plevoets, and Timothy Colleman Rethinking constructional polysemy: The case of the English conative construction Florent Perek Quantifying polysemy in Cognitive Sociolinguistics Justyna A. Robinson The many uses of run: Corpus methods and Socio-Cognitive Semantics Dylan Glynn

7 39

61 87 117

Visualizing distances in a set of near-synonyms: Rather, quite, fairly, and pretty Guillaume Desagulier

145

A case for the multifactorial assessment of learner language: The uses of may and can in French-English interlanguage Sandra C. Deshors and Stefan Th. Gries

179

Dutch causative constructions: Quantification of meaning and meaning of quantification Natalia Levshina, Dirk Geeraerts, and Dirk Speelman

205

vi


The semasiological structure of Polish myśleć ‘to think’: A study in verb-prefix semantics Małgorzata Fabiszak, Anna Hebda, Iwona Kokorniak, and Karolina Krawczak

223

A multifactorial corpus analysis of grammatical synonymy: The Estonian adessive and adposition peal ‘on’ Jane Klavan

253

A diachronic corpus-based multivariate analysis of “I think that” vs. “I think zero” Christopher Shank, Koen Plevoets, and Hubert Cuyckens

279

Section 2.â•‡ Statistical techniques Techniques and tools: Corpus methods and statistics for semantics Dylan Glynn

307

Statistics in R: First steps Joost van de Weijer and Dylan Glynn

343

Frequency tables: Tests, effect sizes, and explorations Stefan Th. Gries

365

Collostructional analysis: Measuring associations between constructions and lexical elements Martin Hilpert

391

Cluster analysis: Finding structure in linguistic data Dagmar Divjak and Nick Fieller

405

Correspondence analysis: Exploring data and identifying patterns Dylan Glynn

443

Logistic regression: A confirmatory technique for comparisons in corpus linguistics Dirk Speelman

487

Name index

535

Subject index

541

Contributors

Timothy Colleman Ghent University [email protected]

Dylan Glynn University of Paris VIII [email protected]

Hubert Cuyckens University of Leuven [email protected]

Stefan Th. Gries University of California, Santa Barbara [email protected]

Sandra Deshors New Mexico State University [email protected]

Anna Hebda Adam Mickiewicz University, Poznań [email protected]

Martine Delorge Ghent University [email protected]

Martin Hilpert University of Neuchâtel [email protected]

Guillaume Desagulier Université Paris 8 Vincennes Staint Denis Université Paris Ouest Nanterre La Défense UMR 7114 MoDyCo [email protected]

Jane Klavan University of Tartu [email protected]

Dagmar Divjak University of Sheffield [email protected]

Karolina Krawczak Adam Mickiewicz University, Poznań [email protected]

Małgorzata Fabiszak Adam Mickiewicz University, Poznań [email protected]

Natlia Levshina Université catholique de Louvain [email protected]

Nick Fieller University of Sheffield [email protected]

Florent Perek University of Freiburg [email protected]

Dirk Geeraerts University of Leuven [email protected]

Koen Plevoets Ghent University [email protected]

Iwona Kokorniak Adam Mickiewicz University, Poznań [email protected]

viii Corpus Methods for Semantics

Justyna A. Robinson University of Sussex [email protected]

Dirk Speelman University of Leuven [email protected]

Christopher Shank Bangor University [email protected]

Joost van de Weijer Lund University [email protected]

Outline

1. Aim of the volume It could be argued that Cognitive Linguistics is undergoing a paradigm shift. Originally, the field sought to show the inadequacies of earlier models of language and the theories of linguistic structure based upon them. Today, the emphasis has changed to testing the various theories about how language works (Geeraerts 2006; Gries and Stefanowitsch 2006; Stefanowitsch and Gries 2006; Gonzalez-Marquez et al. 2008; Glynn and Fischer 2010). This has brought analytical methods, based on observable and quantifiable data, to the fore. In the light of these developments, this volume systematises, reviews, and promotes a range of research techniques and theoretical perspectives that currently inform work across the field of linguistics, with a particular focus on Cognitive Semantics. More precisely, the aim of this book is twofold: i. Didactic: To broaden the understanding and application of the state-of-the-art corpus linguistic techniques for the study of conceptual structure in Cognitive Semantics. Scientific: To advance the state-of-the-art of those techniques through a collecii. tion of studies applied to the description of the conceptual structures of polysemy and synonymy. This publication grew out of the belief that there exists a strong desire in the research community to understand and learn how quantitative corpus methods work and how to apply them to research questions that are basic to the cognitive project. Instead of a rift between linguists using corpus data and those using traditional introspective analysis, constructive communication between the methodologies should be encouraged. Both the descriptive research and the explanations of the statistical techniques included in this book seek to promote such communication. The chapters that describe the statistical techniques are written to help linguists using traditional methods both understand how these new methods work and how to apply them. The research chapters, in turn, showcase the methods described. Their aim is not only to advance corpus-driven quantitative research in Cognitive Semantics, but also to promote the possibilities that these methodologies offer. Observational data and quantitative corpus-driven methods cannot inform all research questions. However, it is hoped that this volume will advance the current state-of-the-art in their use as well as promote their application in the broader linguistic research community.

2


2. Structure and summary The book divides into two sections. The first section begins with eleven chapters, arranged according to their object of study. These chapters begin with an overview of the field in “Polysemy and synonymy: Corpus method and cognitive theory” (Glynn). This chapter includes the analytical justification for approaching both lexis and morpho-syntax in terms of polysemy and synonymy as well as a justification of extending the traditional uses of the terms to cover any variation or similarity in use. The analytical chapters begin with morpho-syntactic polysemy, move to lexical polysemy, then on to lexical synonymy, and finally turn to morpho-syntactic synonymy. Beginning with research on the polysemy of morpho-syntactic semantics, the first descriptive chapter, “Competing ‘transfer’ constructions in Dutch” (Delorge, Plevoets, and Colleman), considers the polysemy of a morpheme-based construction. The ontprefix in Dutch combines with a range of verbs to express dispossession. Using correspondence analysis, the study seeks to capture the lexical semantic morpho-syntactic interplay associated with the construction. The next chapter, “Rethinking constructional polysemy” (Perek), also examines a grammatical construction. The syntactically encoded conative construction in English combines with a range of lexemes. Through the application of collostructional analysis, the author attempts the task of teasing out and identifying the semantic variation associated with the construction. Turning to lexical semantics, “Quantifying polysemy in cognitive sociolinguistics” (Robinson) examines the usage of polysemous adjectives in a community of speakers from South Yorkshire, UK. The study applies cluster analysis, logistic regression, and decision tree analysis in order to examine the extent to which individual conceptualisations are non-random and can be related to the socio-demographic characteristics of the speaker. “The many uses of run” (Glynn) is a repeat analysis of Gries’ (2006) study. Employing a combination of cluster analysis, correspondence analysis and logistic regression, it confirms Gries’ findings but argues that sociolinguistic dimensions should be included in the study of polysemy. Remaining with lexical semantics, but focusing on near-synonymy, “Visualizing distances in a set of near-synonyms” (Desagulier) examines rather, quite, fairly, and pretty in English, combining collostructional analysis and multivariate statistics such as correspondence analysis and cluster analysis. “The uses of may and can in French-English interlanguage” (Deshors and Gries) treats a lexical alternation in first language and second language use. With the use of cluster analysis and logistic regression, the authors seek to identify not only the relationship in use between may and can, but to compare this with French native speakers using English. Moving towards morpho-syntactic semantics, “Dutch causative constructions” (Levshina, Geeraerts, and Speelman) examines a lexeme-based grammatical construction alternation. Focusing on the expression of causation in Dutch, the study employs logistic regression analysis to determine both the semantic and extralinguistic factors

Outline 3

that determine the choice and difference in conceptualisation between the two constructions. The next study, “The semasiological structure of Polish myśleć ‘to think’” (Fabiszak, Hebda, Kokorniak, and Krawczak) continues to move from lexical to syntactic semantics with an analysis of the near-synonymy of a set of prefix-verb combinations. The study combines introspective methods and usage-feature analysis, examined with cluster analysis, correspondence analysis and logistic regression. “A multifactorial analysis of grammatical synonymy” (Klavan) studies a lexical-morphological alternation. Employing logistic regression, the study attempts to determine the conceptual differences that motivate speakers’ choice of a preposition over a grammatical case to express the spatial relation of on in Estonian. “A diachronic corpus-based multivariate analysis of ‘I think that’ vs. ‘I think zero’” (Shank, Plevoets, and Cuyckens) is an analysis of well-known complementiser alternation in English. Logistic regression is used to test a wide range of language factors proposed in the literature to motivate the omission of the complementiser. The second section consists of seven chapters that explain some of the tools and methods for quantitative corpus-driven research. It is designed to introduce the application and interpretation of the statistical methods used in the first section for researchers completely new to the field. It also serves as a ‘cookbook’, or is a quick reference, for intermediate users of the statistical techniques and the programming environment R. The first chapter, “Techniques and tools: Corpus methods and statistics for semantics” (Glynn), is an overview of the field. It examines two corpus methods that are commonly used in Cognitive Semantics and summarises many of statistical techniques currently used in the field. The second chapter introduces the statistical environment of R, used throughout the book (van de Weijer and Glynn). Readers with no experience in R will find this chapter useful when applying the techniques described in the previous chapters to data analysis. The following chapter, “Frequency tables: Tests, effect sizes, and explorations” (Gries), covers many of the essential and basic analytical concepts and how to apply them in R. Building on the statistical basics, the next chapter, “Collostructional analysis: Measuring associations between constructions and lexical elements” (Hilpert), explains the application of collostructional analysis, one of the most popular quantitative techniques in Cognitive Linguistics. This family of techniques are used to quantify the degree of association between linguistic forms. The next three chapters each consider a different multivariate statistical technique. The three techniques in question have proven popular in recent Cognitive Linguistic research. The chapter “Cluster analysis: Finding structure in linguistic data” (Divjak and Fieller), focuses on a method for sorting a given set of phenomena, such as lexemes, constructions, or senses, into categories of similar and dissimilar, relative to some other set(s) of linguistic phenomena such as meanings, argument types, case marking, and so forth. This is followed by the chapter “Correspondence analysis: Exploring data and identifying patterns” (Glynn), which considers a technique similar

4


to cluster analysis, but one that looks for correlations between different phenomena rather than categorising them. It is useful for identifying structure in complex multidimensional data. The third chapter on multivariate techniques, “Logistic regression: A confirmatory technique for variant comparison in corpus linguistics” (Speelman), considers an advanced form of statistical modelling. Logistic regression is a powerful and popular tool in the social sciences, including Cognitive Linguistics. As a confirmatory technique, regression analysis represents a level of statistical analysis that is more complex than the previous techniques covered. This chapter charts the basics of its application, interpretation and verification. Where the first section represents a broad, yet coherent, picture of the cutting edge in the application of these techniques, the second section seeks to offer an introduction to the different statistical techniques employed in corpus-driven semantics. Focusing on the study of polysemy and synonymy of both lexical morpho-syntactic forms, these empirical analyses represent the vanguard of corpus-driven Cognitive Linguistics.

The editors

References Geeraerts, D. (2006). Methodology in Cognitive Linguistics. In G. Kristiansen, M. Achard, R. Dirven, & F. J. Ruiz de Mendoza Ibañez (Eds.), Cognitive Linguistics: Current applications and future perspectives (pp. 21–50). Berlin & New York: Mouton de Gruyter. Glynn, D., & Fischer, K. (Eds.). (2010). Quantitative Cognitive Semantics: Corpus-driven approaches. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110226423 Gonzalez-Marquez, M., Mittelberg, I., Coulson, S., & Michael S. (Eds.). (2008). Methods in Cognitive Linguistics. Amsterdam & Philedelphia: John Benjmains. Gries, St. Th. (2006). Corpus-based methods and Cognitive Semantics: The many senses of to run. In St. Th. Gries, & A. Stefanowitsch (Eds.), Corpora in Cognitive Linguistics: Corpus-based approaches to syntax and lexis (pp. 57–99). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110197709 Gries, St. Th., & Stefanowitsch, A. (Eds.). (2006). Corpora in Cognitive Linguistics: Corpus-based approaches to syntax and lexis. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110197709 Stefanowitsch, A., & Gries, St. Th. (Eds.). (2006). Corpus-based approaches to metaphor and metonymy. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110199895

Section 1

Polysemy and synonymy

Polysemy and synonymy Cognitive theory and corpus method Dylan Glynn

University of Paris VIII

This chapter introduces the field of polysemy and synonymy studies from a Cognitive Linguistic perspective. Firstly, the discussion explains and defines the object of research, showing that the study of semantic relations, traditionally restricted to the description of lexical semantics, needs to be extended to include all formal structures, including morpho-syntax. Secondly, given the theoretical assumptions of Cognitive Linguistics, it is argued that quantitative corpus-driven methods are essential for the description of semantic structures. Lastly, the chapter charts the development of Cognitive Semantic research in polysemy and synonymy and demonstrates how the current corpus-driven research in the field is inherently linked to the traditions of radial network analysis and prototype semantics. It is argued that instead of an empirical revolution (as has been suggested in recent commentaries), the current trends in the use of observational data are a natural extension of the Cognitive Semantic research tradition. Keywords: Cognitive Linguistics, corpus linguistics, polysemy, prototype semantics, quantification, radial network analysis, synonymy

1. Introduction: Theory and method The idea of ‘corpus semantics’, just like the possibility of ‘quantifying meaning’, is not self-evident. This introduction to the field of corpus-driven Cognitive Semantics attempts to explain how semantic analysis can, and indeed, should, turn to corpus methods. It also explains why quantitative techniques are needed in this endeavour. Assuming the Usage-Based Model (Hopper 1987; Langacker 1987), how can we identify and explain the semantic structuring of language empirically? Post-Generativist and post-Structuralist approaches to language avoid positing, a priori, analytical constructs to explain the structuring of language, rather treating it holistically as a dynamic and varied result of use. However, without a structurally independent langue or an ‘ideal’ speaker’s competence, against what predictive model can we test

8

Dylan Glynn

our hypotheses or attempt to falsify our claims about language structure? Without constructs such as langue and ideal competence, linguistic research, whether Functional or Cognitive, must adopt an inductive, sample-based, methodology. To these ends, experimental techniques for the analysis of semantics have been developed, yet corpus methods remain poorly represented within the field. Moreover, the theory of Cognitive Linguistics recognises no internal language modules, such as syntax, lexis, semantics or pragmatics. From a non-modular perspective, the study of meaning must account for the integration of all these components of language structure and do so simultaneously in a functionally and conceptually plausible manner. Corpus-driven methods, and multivariate statistics more specifically, are perfectly suited for such a task. A usage-based semantics, therefore, must take two fundamental steps. Firstly, it must adopt inductive research methods. Whether elicited through experimentation, extracted from electronic corpora, or collated from questionnaires and field research, generalisations based on samples of data present the only possibility for hypothesis testing. Importantly, acknowledging this fact entails validating sample-based results through statistical confirmation. Secondly, it must develop corpus-driven semantic analysis. If we are to account holistically for the integrated complexity of the various dimensions of language structure, it is essential that we examine natural contextualised language production. Samples of natural language large enough to permit inductively valid claims are what we term corpora. Here again, statistics comes to the fore, though for a different reason. If we are to identify structure, sensitive to its usage context, multivariate statistics is a powerful, if not essential, tool due to the sheer complexity of the data. The aim of this book is both to introduce quantitative corpus-driven semantic methodology to the broader research community and to advance the state-of-theart. The methods in focus are those that are especially applicable to the lexical and constructional semantic relations of similarity and difference. Linguistic forms are used in different ways, and capturing this semasiological variation is what we term the study of polysemy. Likewise, speakers choose between different linguistic forms to express similar concepts. Explaining this onomasiological variation is what we term the study of synonymy. The reader should be aware that the term polysemy is not restricted to ‘true’ polysemy, where distinct referents are indicated by a single form. Instead, meaning is understood from a usage-based perspective, where any systematic variation in use represents semasiological structure. In the same vein, synonymy is not restricted to absolute similarity since, from a Cognitive Linguistic perspective, one assumes that any variation in form is motivated by some variation in use and that ‘true’ synonymy is rare, if it exists at all. Lastly, it must be added that the term semantics is used to indicate encyclopaedic semantics and pragmatics, as opposed to linguistic semantics in its narrow sense.


Moreover, the term corpus methodology should be understood to indicate research that is corpus-driven, as opposed to corpus-exemplified (q.v. Tummers et al. 2005). Where corpus-exemplified research identifies occurrences to explain or support a theory of language structure, corpus-driven research examines large samples of natural language in order to test theories about language structure (often previously proposed a priori in corpus-exemplified research). In this, we focus on quantitative techniques, and, more specifically, statistical methods for the exploration of data and the falsification or confirmation of quantitatively testable hypotheses. We begin by developing operational definitions of polysemy and synonymy (Section 2). The discussion then demonstrates, given the above definitions, why we need corpus methodology (Section 3). Finally, we consider the argument that the quantitative corpus tradition and the prototype-based Cognitive Semantic tradition are not only analytically compatible, but, in fact, are inherently entwined. It is argued that the introspection-based research in prototype structuring of linguistic categories and the results of the radial network tradition should be understood as the theoretical modelling of semantic structure, an essential first step in empirical research (Section 4).

2. Polysemy and synonymy: Definition, object and operationalisation Given the linguistic heritage of the 20th century, it is upon us to begin with two simple questions. Firstly, what exactly constitutes polysemous and synonymous semantic relations in usage-based semantics? Secondly, is it theoretically and analytically possible to speak of the polysemy and synonymy of a syntactic construction? This section answers each question in turn. In doing so, the discussion will offer operational definitions of the concepts ‘polysemy’ and ‘synonymy’ and will justify extending the study of such relations to morpho-syntax. With meaning defined as encyclopaedic and with all form being semantically motivated, how do we understand the notions of polysemy (difference in sense of a form) and synonymy (similarity in sense between forms)? Given a traditional understanding of the terms, usage-based approaches, such as Cognitive Linguistics or Functional Linguistics, are not interested in polysemy or synonymy per se. Much of what cognitivists and functionalists examine would be called ‘semantic vagueness’, ‘word fields’ or ‘syntactic alternation’ in other theoretical paradigms. Moreover, for Structuralism, which holds a distinction between the langue and its use, our position, that the meaning of a word can be operationalised as its use, is nonsense. For Generativism, which argues for an autonomous formal system, examining the semantic structure of syntactic patterns is equally nonsense. In Structuralist terms, polysemy was identified using the definitional test and/ or the ambiguity test. These tests were designed to distinguish polysemous relations from vague relations, on the one hand, and from a monosemic form, on the other (for

9

10

Dylan Glynn

further discussion, see Geeraerts 1993a). This modular understanding of semantic structure assumes two theoretical constructs – firstly, the notion of truth conditional semantics and, secondly, the notion of semantic categories determined by necessary and sufficient conditions. The same assumptions determined the definition of synonymy – any two lexemes were considered synonymous if replacing one lexeme with the other did not change the ‘truth semantic’ meaning of the phrase (Lyons 1968:â•›428). Cognitive Linguistics categorically refutes both assumptions.1 Indeed, cognitive-functional approaches to meaning do not recognise the notion of necessary and sufficient conditions nor any strict division between linguistic semantics and context pragmatics. Without such assumptions, the meaning of the terms polysemy and synonymy can be loosened and defined as: Polysemy – different concepts-functions of a form Synonymy – different forms for a concept-function Note that these definitions of polysemy and synonymy do not exclude taxonomic relations such as hyponymy and hyperonymy (for polysemy) or basic level, superordinate and subordinate relations (for synonymy). The definitions need to include these relations because, in a given situation-context, a given form may be used at different levels of specificity (‘vertical’ polysemy), just as a choice is made between different words signifying different levels of specificity (‘vertical’ synonymy). Such a broad understanding of semantic relations could be more accurately described as semasiological and onomasiological variation (Geeraerts 1993b; Grondelaers and Geeraerts 2003). Nevertheless, we will continue with the terms polysemy and synonymy since they enjoy wider currency.

1. See Geeraerts (1987, 1993a, 1994) and Tuggy (1993) for a more detailed explanation of the theoretical questions at stake here. For a summary of the questions that led to the original debates on truth conditional semantics and necessary and sufficient conditions, see Verschueren (1981) and Lakoff (1982). Footnote 10 lists the principal works that established Cognitive Linguistics’ position on these questions. The contemporary Anglo-Saxon Structuralist position is summarised in Cruse (2000) and Murphy (2003). The aforementioned debates largely ignored the contemporary French, German and Russian traditions. Examples of current French Structuralist research on semantic relations include Rastier’s (1987, 1991, 2011) context sensitive approach and Victorri and Fuchs’s (1996) construal sensitive framework. The German Structuralist understanding of semantic relations lies close to the Anglo-Saxon tradition (q.v. Coșeriu 1980; Kastovsky 1982; Lutzeier 1985; and Lipka 1992). The Leibnizian tradition of semantic primitives, important to the Moscow school of semantics, is represented by Apresjan (1974, 2000), Wierzbicka (1985, 1996), and Mel’čuk (1989). From a Leibnizian functional perspective, Wierzbicka (1989, 1990) offers responses to issues raised in Verscheuren (1981) and Lakoff (1982). Geeraerts (1999b) offers, in turn, responses to Wierzbicka.


Can we now justify applying the notions of polysemy (semasiological variation of a single form) and synonymy (onomasiological variation between more than one form) to the study of grammar? If we assume that all form is conceptually or functionally motivated and if we agree that, in the study of polysemy and synonymy, we are effectively studying variation in concept-function relative to form and variation in form relative to concept-function, then this must necessarily be extended to non-lexical meaning and form. Therefore, in blunt terms, the study of polysemy and synonymy includes the study of schematic forms and their meanings such as those typical of syntax and prosody just as much as it does the study of words and morphemes. We can, therefore, identify our object of study more precise terms: Polysemy – the functional-conceptual variation of any symbolic form Synonymy – the functional-conceptual relation between any symbolic forms2 This is a simple statement about an object of study that many linguists would take as obvious, yet others take as ludicrous. The division is a result of the fact that even if we accept that all formal structure is motivated, there certainly exist different types of form just as there exist different types of meaning. Phonological, gestural, syntactic, morphological and lexical forms all tend to possess different characteristics. There is no doubt in this – the referential meaning of a lexeme like chair is far from the intersubjective meaning of a request implicature, and these two ‘types’ of meaning differ from the abstract relational meaning of the Transitive Construction. Notwithstanding the belief that such linguistic devices are inextricably interwoven structurally, the characteristics of these different types of form and meaning, just as the tools needed to describe them, differ profoundly. If we accept that, analytically (as opposed to theoretically), such differences exist, then we can distinguish different lines of research. Given this, it is still possible to speak of lexical research or syntactic research as long as this is seen as an analytical emphasis, not an object of study. In order to avoid conjuring the theoretical modules of earlier theories of language, we can term the different lines of research – schematic 2. This definition would also cover antonymy. Although it may seem a little far-fetched to consider antonymy as a synonymy relation, this is merely a result of the terminology. Ideally, onomasiological structure would be a better term than synonymy, but the term is not established in the Anglo-Saxon tradition and would, therefore, add considerable terminological weight to the discussion. Nevertheless, it must be noted that antonymy should not be seen as an antonym (in a non-technical sense) for synonymy. It is well known from the study of lexical fields and, more recently, word space modelling in computational linguistics, that antonyms are, in fact, closely related to each other semantically and, therefore, are relatively synonymous. For example, hate is much closer in meaning to love than car or run. Generally, antonyms are synonyms semantically opposed by one culturally or perceptually salient feature. Therefore, in fact, antonymy is an antonym to synonymy, in other words, very close in its meaning. See Jones (2002:â•›51) and Murphy (2003:â•›37) for examples of the issue at hand.

11

12

Dylan Glynn

and concrete. Such distinctions are abstract enough to avoid leading us into the trap of modularising language structures, while also being transparent enough to help us easily differentiate between lines of research and the tools they necessitate. Therefore, although semantic structure and its relations exist equally for morpho-syntax and lexis, the kind of semantics typically associated with lexical forms is of a much more concrete nature than that of more schematic formal structures. This is, of course, a tendency, but an important one because the different kinds of formal structure and the semantics associated with them may warrant different analytical techniques. For these practical purposes, let us identify four objects of study: Polysemy – Concrete meaning – Schematic meaning Synonymy – Concrete form – Schematic form Seen in these terms, Lakoff ’s (1987) analysis of over is a study of concrete, or non-schematic, polysemy, but his study of the Deictic Construction is schematic polysemy. On the other side of the coin, Lakoff ’s (1987) analysis of anger is concrete (non-schematic) synonymy and the analysis of the Dative alternation by Goldberg (2002) is an instance of schematic synonymy. Regardless of whether we speak of words or syntax, the analytical object of the relations between different yet functionally-conceptually similar forms versus the relations between different functions-concepts of a single form should be evident. A few further examples should demonstrate the distinction and its value for identifying the fundamental objects of study for all usage-based linguistics. By way of example for schematic polysemy, consider Halliday’s (1967) study of the English Transitive Construction, Langacker’s (1982) analysis of the English Passive Construction or Bondarko’s (1983) analysis of the Perfective Aspect. For concrete polysemy, take Culioli’s (1990) analysis of the French lexeme donc ‘so, thus’ or Fillmore’s (2000) lexeme crawl. For synonymy, the same diversity exists. At the schematic level, we have Givón’s (1982) evidential markers, Halliday’s (1985) English Grammatical Conjunctions, Talmy’s (1988) Causative Constructions, Culioli’s (1990) English Negation Constructions, and Langacker’s (1991) English Nominal Constructions.3 Obviously, the lexical research is equally diverse: from Fillmore’s (1977) buy-sell frame, the 3. Seen from this perspective, the entire tradition of conceptual analysis (Wierzbicka 1985; Lakoff 1987; Stepanov 1997; Vorkachev 2004; and Bartmiński 2008 inter alia) is based upon synonymy. Such research begins with a concept and examines what words or expressions are available for its linguistic representation. On the grammatical front, a similar notion holds. Understood in these terms, Talmy’s (1988) Force Dynamics or Langacker’s Causative Constructions (1991:â•›408–411) are essentially onomasiological fields of near-synonymous forms. Of course, regardless of whether it is lexis or syntax, in order to understand the relationship between these different forms, we must investigate each form semasiologically, that is, its


lexical field of say-speak by Dirven et al. (1982), and Lehrer’s (1982) study on adjectives for the description of wine to the full swath of conceptual metaphor and metonymy studies. Indeed, in usage-based linguistics, most descriptive linguistic studies can be classified as one of these four lines of research. Despite the fact that a wide and varied range of linguistic analyses can be understood as the study of polysemy or synonymy, not all research can be characterised by this typology. Within both the Functional and Cognitive paradigms, there exist objects of study that are not readily characterised in this manner. Such research lies beyond the realm of polysemy and synonymy. Within Cognitive Linguistics, ad hoc, or non-entrenched, categorisation typical of conceptual integration (Fauconnier and Turner 1998), just as with the entire field of language processing, are beyond the purview of polysemy and synonymy research. Within Functional Linguistics, the detailed, context dependent analysis of conversation or the wide-ranging research on the characteristics of genre, register, and stylistics seem to lie outside the realm of semantic relations per se. This is not to say that such research, both the cognitive and the functional, does not inform the study of semantic relations, but it is distinct from this object of study. Having established the importance and limits of these two objects of study – the conceptual-functional structure of the form and the forms available to express a concept-function – we can now ask how best to operationalise these notions? In other words, how can we define the object of study in a measurable way?4 With no langue or independent syntax against which we can test hypotheses, what exactly are we analysing in the study of semantic relations? How can we falsify or verify results? We need an operationalisation of the object of study that either offers stability to the system or a means of capturing the dynamic nature of that system. The answer lies in Langacker’s (1987:â•›59–60) theory of the entrenched form-meaning pair.5 The theory of entrenchment can provide a frequency-based operationalisation of grammaticality – the more often a form-meaning pair is used, the more automated its processing becomes and the more ‘grammatically acceptable’ it is according to the speaker’s intuition.6

polysemy. Bondarko (1991) argues convincingly that the semasiological–onomasiological divide is fundamental to any theory of conceptually or functionally motivated language. 4. For a discussion on the importance of operationalisation in language science, see Stefanowitsch (2010). 5. See also Givón’s (2005:â•›48ff.) more detailed investigation into what he terms automated processing in the attention system. The idea of form-function mapping is extended to include perceptual and conceptual issues such as prototype structure. 6. The theory of entrenchment is based on the notion of automisation, a widely accepted theory in psychology. Schneider and Shiffrin (1977) and Shiffrin and Schneider (1977) were the first to develop the hypothesis.

13

14

Dylan Glynn

Generalised across a speech community, that is the ensemble of individual speakers’ linguistic knowledge, we have an operationalisation of grammar. If we accept these hypotheses and assume that through repeated contextualised use, the relation between a concept-function and a form becomes stable, then we have an identifiable object of study. At a theoretical level, therefore, the study of semantic relations is the study of variation of entrenched form-meaning pairs: Polysemy – entrenched functional-conceptual variation of a schematic or non-schematic form Synonymy – entrenched functional-conceptual relation between schematic and non-schematic forms It is important to note that Langacker’s theory of entrenchment is determined by relative frequency of use. This can be re-stated as an operational definition: the degree of entrenchment is determined by the frequency of association of a given form and a given use of that form. Of course, what constitutes ‘a form’ and ‘a use’ is open to debate, but the notion of entrenchment per se is operationalised in a way that permits quantified analysis of semantic structure – the aim of this volume. The relative frequency of a form-meaning pair determines its entrenchment in a speaker’s knowledge. This, when extended to an entire speech community, means that the frequency of a form-meaning pair in language can indicate the degree of its stability in the intersubjective system of language.7 Therefore: Polysemy and synonymy can be measured in terms of the relative frequency of association of form and meaning Given this understanding of semantic relations, questions as diverse as prototype effects and sociolinguistic variation are neatly explained by a single principle. This claim deserves a brief explanation. A central question for the study of polysemy is: which ‘senses’ are more ‘central’, or more prototypical, than others. A similar question exists in synonymy studies: which forms are more basic taxonomically (as in basic-level terms, Lakoff 1987)? Both the concept of prototype meaning and basic-level form can be operationalised in terms of frequency. Although it is not claimed that frequency alone can explain prototype or taxonomic structure, it is, nonetheless, one important operationalisation of these phenomena (see Arppe et al. 2010). The operationalised definition is straightforward: the more frequent a given meaning, the more ‘typical’ it is categorically. This is a frequency-based understanding of prototypicality. The same can be 7. Even if this operationalisation of entrenchment is adequate, which is questionable (see Note 8), this does not, in turn, entail that corpora are representative of frequency of use. Given the dynamic complexity of language, it is a reasonable argument that no corpus will be truly representative for the foreseeable future.


posited, mutatis mutandis, of basic level categories in taxonomic structure: if for polysemy, typicality is operationalised as the most frequent concept-function (relative to a form), for synonymy, basicness is operationalised as the most frequent form (relative to a concept-function). Therefore, for synonymy, the more frequent a form, the more basic it is taxonomically. Importantly, this frequency-based understanding of the system integrates the varied and dynamic nature of language into our model of semantic relations. For a usage-based understanding of language, the system is emergent and entirely dependent on context – context of situation, context of speaker, context of time, context of region. A given form-meaning pair will be more frequent in one city than another, in one register than another, for one gender more than another, and at one period of time more than another. Relative frequency, at a theoretical level, eloquently incorporates this complexity into the object of study. Therefore, the operationalisation – context-sensitive frequency-based typicality – simultaneously captures semantic structure both categorically (prototype effects) and taxonomically (basic-level effects) but also relative to social variation. Lastly, it must be stressed that this frequency-based approach to entrenchment is only an operational definition. Other operationalisations of the relationship between form and meaning may be equally valid. Despite Langacker’s concern with frequency, something that holds well for corpus linguists, perceptual and conceptual salience surely also have a hand in the learning process, and therefore in entrenchment. Langacker’s explanation of entrenchment assumes that all input has the same ‘weight’ in or ‘impact’ upon the system. This, it would seem, is a simplification. There is no reason to suppose that every occurrence of, or exposure to, language events has the same value in the process of entrenchment. The implications, especially for prototype and taxonomic structures are far reaching. We must suppose, therefore, that the frequency-based account of language cannot give us the full picture. It does, however, offer an operationalised and quantifiable object of study, one that will permit the testing of hypotheses, the verification of results, and one that will provide clear benchmarks for the comparison of results, using other methodologies and other operationalisations.8

3. Complexity and sampling: The need for quantification Why are quantitative corpus methods needed for a cognitive approach to polysemy and synonymy? There are two answers to that question. The first answer takes us back to our object of study and the second, to our model of language. 8. The possibility of using frequency as an operationalisation is central to a diverse range of discussions in usage-based linguistics. Schmid (2000, 2010), Tomasello (2003), Bybee (2007) and Geeraerts (2010a) are examples of research that consider this possibility.

15

16

Dylan Glynn

The first reason we need quantitative techniques can be summarised as a question of complexity. We defined our object of study as “relative frequency of association of form and meaning”. The discussion has, thus far, ignored an important issue – what constitutes a given form and what constitutes a given meaning? We have stated that a form is any form and that meaning is anything we know of the world. In fact, both the terms ‘form’ and ‘meaning’ are entirely misleading. All forms exist in a formal context and are, therefore, composite. Theoretically, one should not speak of a form but of a composite form. Similarly, meaning is situated contextually and therefore one should not speak of a meaning as a reified sense, but as an intersubjective result of communication.9 The importance and implications of these two points cannot be underestimated. Formal structure is complex. Even the simplest utterance is a composite form at some level. Moreover, dialect and sociolect variation means that even phonetic components can be indicative of usage variation. Prosody, syntax, morphology, lexis and even gesture, all come together in effectively every utterance as composite forms. Since it is a fundamental tenet of Cognitive Linguistics that language must be analysed holistically, we must treat forms as composite structures, always. If formal structure is complex from a cognitive perspective, meaning is more so. Even lexical meaning cannot be divided into discrete senses. It is a dynamic, context dependent, multi-dimensional and intersubjective social phenomenon. Geeraerts’ (1993a) and Kilgarriff ’s (1997) studies on polysemy and word meanings mark important milestones in the study of semantic relations. Their work shows that lexical senses, just like any functionally or conceptually determined category, cannot be assumed to be discrete or reifiable. Arriving at this point, both theoretically and descriptively, was a long road. Via theoretical research on prototype structures (Geeraerts 1993a; Zlatev 2003), on the one hand, and via corpus research in lexicology (Kilgarriff 1997; Glynn 2010b) on the other, the proposal that a reified and discrete understanding of semantic structure, including individuated senses, will not produce adequate descriptions of that structure has gained currency. Encyclopaedic semantics entails that all of sociolinguistic structure is included in the semantic analysis. The situation context, the gender, the age, the socio-economic class, the geographical region, and the social status are all dimensions of language use, dimensions of world knowledge in how to use language, how to communicate successfully. This world knowledge must, therefore, be integrated into semantic analysis. The result is a complex multidimensional ‘form’ coupled with a complex multidimensional ‘meaning’. It may be possible, using intuition and introspection, to consider

9. It is not the point of this discussion to enter into the debates on sign and communication theory. Suffice it is to say that any sign theory to which a Cognitive Linguist would ascribe would also see meaning as a result of a communicative act and not an inherent objectifiable phenomenon.


Table 1.â•‡ Variation in the results of introspection-based polysemy analysis Lexeme

Senses identified

Study

over (English)

6 basic senses, 21 sub-senses 17 senses 9 of Lakoff ’s sub-senses as 1 sense 3 basic senses, > 40 sub-senses 6 basic senses, 12 senses sub-senses 3 basic senses 1 basic sense, 15 senses sub-senses 13 senses 3 basic senses, 11 sub-senses 3 basic senses, 8 sub-senses 3 basic senses, inconclusive sub-senses 3 basic senses, 14 sub-senses 5 basic senses, 8 sub-senses

Lakoff (1987) Taylor (1989a) Vandeloise (1990) Deane (1993a, 2006) Dewell (1994) Kreitzer (1997) Tyler and Evans (2003) Cuyckens (1991) Geeraerts (1992) Bellavia (1996) Dewell (1996) Meex (2001) Liamkina (2007)

over (Dutch) über (German)

all of these dimensions, but to understand how they all interact, or how they could all interact, is an effectively impossible feat. Quantifying the analysis permits the use of multivariate statistics, which is designed for modelling and capturing structure in precisely this kind of complex system. The second reason we need to develop quantitative methods for the study of semantic relations lies in the model of language propounded by Cognitive Linguistics. Both the Structuralist and Generativist traditions assumed models of language that permitted the falsification of claims made about its structure. Necessary and sufficient conditions, for establishing semantic categories, and grammatical acceptability tests, for checking proposed grammatical rules, both allow an analyst to falsify hypotheses. What possibility does Cognitive Linguistics have for falsifying propositions made about language structure? How can we test for the descriptive or explanatory adequacy of a conceptual metaphor or a reference-point construction? Neither makes any predictions that can be falsified. The lexical polysemy studies of early Cognitive Semantics were excellent examples of this problem. Lakoff (1987) proposed 21 senses for the lexeme over, but his work was challenged (see Table 1). Although this, it would seem, is good scientific procedure – a study proposes a given number of senses, the results are challenged – there is, ultimately, no way of resolving the issue because there is no way of disproving his original analysis. Table 1 lists the different proposals of the number of senses for over in English, Dutch and German. This debate could effectively continue ad nauseum, since using one’s intuition to determine a category, especially a category that can have both a fuzzy boundary and better or worse exemplars, has no possibility for falsification. It is, ultimately, a matter of opinion.

17

18

Dylan Glynn

In earlier theoretical paradigms, truth conditional semantics and predictive rules could be tested using deductive proofs. Usage-based research has no such options. A counter example, even many, does not contradict a proposal about relative structure. From our perspective, linguistic structure is emergent; it is dynamic and varied. However, a quantification of the study of semantic relations permits inductive research. The only way to test our hypotheses is to take a sample of natural language use or a sample of elicited language use and make generalisations based on that sample. Once we speak in terms of samples and populations, then we speak of inductive research, and, in brief, statistics. Statistical analysis gives us the probability that a given finding in a given sample is not chance, it gives the possibility of modelling the variation in our data and testing the accuracy of analyses by using these models to predict language use. Quite simply, moving towards quantification and statistical analysis appears inevitable for all usage-based language research.

4. Modelling meaning. Multidimensional patterns and prototype effects How does the usage-based approach contribute to the prototype structure and radial network tradition of polysemy and synonymy in Cognitive Semantics? The study of both (vague) polysemy and (near) synonymy has a long tradition in Cognitive Linguistics. Indeed, many of the seminal works were devoted to such questions. This section places quantitative corpus-driven research, such as that presented in the current volume, in Gries and Stefanowitsch (2006) and Glynn and Fischer (2010), in this ‘historical’ context. At the end of the last century, what was often called the ‘network’ or ‘radial category’ approach to meaning included a wide range of polysemy and synonymy studies of especially spatial prepositions and grammatical cases. These forms were particularly interesting because they linked the theoretical research on perception-based construal and image schemata with the culturally determined lexico-grammatical structure. The aim of this research was to model prototype structure and encyclopaedic semantics. In effect, presented with the boundless considerations of encyclopaedic semantics as well as relative categorisation due to prototype effects on structure, theoretically and analytically, much of the work can be seen as an attempt to identify order in what is an immensely varied and complex system. While theoretical models such as Frame Semantics (Fillmore 1985) and Idealised Cognitive Models (Lakoff 1987) were proposed in an attempt to identify structure in encyclopaedic semantics and the prototype effects in language, radial network analysis, in its various guises (Barthélemy 1991; Rice 1993; Geeraerts 1995), was essentially a representational formalism designed to visualise and summarise systematicity in semantic complexity. This formalism was used, to various extents, by different authors, but the principle


of (i) employing encyclopaedic semantic features (ii) without the notion of necessary and sufficient conditions for category membership (iii) in order to distinguish senses and relate forms, underlies all radial network research. The theoretical research on semantics and categorisation began with Fillmore (1975, 1977) and Lakoff (1975, 1977). Although theoretical discussion on these topics continues to this day, the field was established by Verschueren (1981), Fillmore (1985), Vandeloise (1986), Geeraerts (1987), and Lakoff (1987).10 The application of these theories in descriptive analysis blossomed into what was termed the radial network approach. Early case studies included Lindner (1983), Brugman (1983a, 1984), Rudzka-Ostyn (1983, 1989), Hawkins (1985), Vandeloise (1986), Janda (1986, 1990), Norvig and Lakoff (1987), and Taylor (1988). This line of research proved so popular that the mainstay of Cognitive Semantics in 1990s could be argued to be based upon it. A string of anthologies largely devoted to the application of the method include Geeraerts (1989), Tsohatzidis (1990), Dubois (1991), Lehrer and Kittay (1992), Zelinsky-Wibbelt (1993), Schwarz (1994), Dirven and Vanparys (1995), Taylor and MacLaury (1995), Pütz and Dirven (1996), Ravin and Leacock (2000), Cuyckens and Zawada (2001), Cuyckens and Radden (2002), Nerlich et al. (2003), Cuyckens et al. (2003), and Rakova et al. (2007). At first sight, this highly abstracted and introspective research tradition would seem distinct from, even at odds with, the bottom up approach of corpus-driven methodology. However, upon closer inspection, we see that the very origins of corpus-driven, indeed, quantitative corpus-driven, semantic research lie in the radial network studies. The contemporary methodology directly inherits from and builds upon this tradition. For both the study of synonymy and polysemy, many of the earliest studies were entirely empirical. Moreover, as we will see below, the two approaches are theoretically linked. Let us, however, begin with an overview of the radial network research. For the practical concerns of brevity, it is impossible to offer even a snippet of this immense field. However, Table 2 offers a selection of studies, chosen to represent the depth and variation of the radial network approach to semasiological structure. The object of study, its general part of speech or grammatical category, whether this is schematic or concrete in form, as well as the method of analysis and reference аrе listed. It must be stressed that the distinction between schematic and concrete forms is only designed to show tendencies and no theoretical distinction is intended. The decision as to what to include is based on a subjective evaluation of the impact of the study and the author upon the field, as well the extent of the study, priority being given to monographs. 10. Beyond these early studies, the following research represents key points in the discussion on semantics and categorisation in Cognitive Semantics: Deane (1988), Geeraerts (1988, 1989, 1993a, 1997), Taylor (1989a), Wierzbicka (1989, 1990), Vandeloise (1990), Kleiber (1990, 1999), Lehrer (1990a, 1990b), Dunbar (1991, 2001), Tuggy (1993, 1999), and Croft (1998). Recent discussion includes Tyler and Evans (2003), Zlatev (2003) and Evans (2005, 2006).

19

20 Dylan Glynn

Table 2.â•‡ Prototype-encyclopaedic analysis of polysemy in Cognitive Semantics Object

Form

Schematicity Method

Reference

lie (English) talk, say, tell, speak (English) over (English) on the go Cx. (English) up, out (English) idea (English) kind of Cx. (English) uit (Dutch), wy (Polish) spatial preps. (English) za-, pere-, do-, ot- (Russian) spatial preps. (French) over (English) let alone Cx. (English) down (English) tall (English) ask (English) Genitive (English) Dative (Czech) vers (Dutch) Middle Voice Cx. (French) Resultative Cx. (English) Dative (Polish) over (Dutch) Ditransitive Cx. (English) in (Dutch) Instrumental (Russian) at, on, in (English) ‘give’ (Mandarin) at, by, to (English) (a)round (English) Genitive (Polish) in (English) Instrumental (Polish) off (English) op (Dutch) over (English) answer, respond (English) door, langs (Dutch) Caused-Motion Cx. (English) at, on, in (English) Dative (Polish) over (English) figure out (English)

verb verb prep. constr. particle noun constr. prep. prep. prefix prep. prep. constr. prep. adj. verb case case adj. constr. constr. case prep. constr. prep. case prep. verb prep. prep. case prep. case prep. prep. prep. verb prep. constr. prep. case prep. verb

concrete concrete concrete schematic concrete concrete schematic concrete concrete concrete concrete concrete schematic concrete schematic concrete schematic schematic concrete schematic schematic schematic concrete schematic concrete schematic concrete concrete concrete concrete schematic concrete schematic concrete concrete concrete concrete concrete schematic concrete schematic concrete concrete

Coleman and Kay (1981) Dirven et al. (1982) Brugman (1983a) Brugman (1983b) Lindner (1983) Brugman (1984) Kay (1984) Rudzka-Ostyn (1985) Herskovits (1986, 1988) Janda (1986) Vandeloise (1986) Lakoff (1987) Fillmore et al. (1988) Schulze (1988) Dirven and Taylor (1988) Rudzka-Ostyn (1989) Taylor (1989b) Janda (1990, 1993) Geeraerts (1990) Melis (1990) Goldberg (1991, 1995) Rudzka-Ostyn (1992, 1996) Geeraerts (1992) Goldberg (1992, 1995) Cuyckens (1993) Janda (1993) Rice (1993) Newman (1993) Deane (1993b) Schulze (1993) Rudzka-Ostyn (1994, 2000) Vandeloise (1994) Dąbrowska (1994) Schulze (1994) Cuyckens (1994) Dewel (1994) Rudzka-Ostyn (1995) Cuyckens (1995) Goldberg (1995) Sandra and Rice (1995) Dąbrowska (1997) Kreitzer (1997) Morgan (1997)

elicitation observation introspection introspection introspection introspection introspection introspection introspection introspection introspection introspection introspection introspection elicitation observation introspection introspection introspection observation introspection introspection introspection introspection introspection introspection introspection introspection introspection introspection introspection introspection introspection introspection introspection introspection observation introspection introspection elicitation introspection introspection introspection


Table 2.â•‡ (continued) Object

Form

Schematicity Method

Reference

at, on, in (English) Causative Cx. (English) Dative (Dutch) straight (English) on (English) to, for (English) What’s X doing Y Cx. (Engl.) crawl (English)

prep constr. constr. adj. prep. prep. constr. verb

concrete schematic schematic concrete concrete concrete schematic concrete

Cuyckens et al. (1997) Lemmens (1998) Geeraerts (1998) Cienki (1998) Rice et al. (1999) Rice (1999) Kay and Fillmore (1999) Fillmore (2000)

elicitation observation introspection introspection elicitation observation introspection introspection

Research in synonymy, though it received less attention and possibly produced less in terms of quantity, was equally important in the development of Cognitive Semantics. Studies such as Fillmore (1977), Lehrer (1982), Dirven et al. (1982), Janda (1986), and later Schmid (1993), Geeraerts et al. (1994), and Rudzka-Ostyn (1995) represent seminal work in the field. Table 3 offers a summary of Cognitive Linguistic case studies in synonymy, again up until the turn of the century. There is some redundancy with Table 2 because what may be a set of individual case studies on polysemy were also combined to present a study in near-synonymy. Again, the table is designed to offer an overview and is no way complete. It should be noted that although there was less work on synonymy per se, the conceptual metaphor studies, which were extremely numerous at the time, are, in effect, synonymy studies. Although such research was primarily interested in figurative lexemes, they remain, nonetheless, studies of near-synonymy (as noted by Kittay and Lehrer 1981 at the time). There was much discussion about what constitutes a source domain and/or target domain and whether certain expressions were in fact examples of the concept in question. From a lexical semantic point of view, these questions are, of course, questions of near-synonymy. Expressions such as to have the hots versus to be head over heels were said to profile different aspects of the target domain, in other words, they were near-synonyms. In order to appreciate the trends and heritage of contemporary methods in Cognitive Semantic research, it is helpful to make a quantified summary of the research output.11 Since, for reasons of space, full coverage of this research history is impossible, only the results of the investigation are presented. A survey of 126 studies was conducted from roughly the beginning of the Cognitive Linguistic research community with the publication of Paprotté and Dirven’s (1985) anthology The Ubiquity of

11. See Geeraerts (2005; 2006b), Croft (2009) and Glynn (2010c) for other discussions on methodological trends in Cognitive Linguistics.

21

22

Dylan Glynn

Table 3.â•‡ Prototype-encyclopaedic analysis of synonymy in Cognitive Semantics Object

Form

Schema- Method ticity

Reference

buy (English) speak (English) wine terms (English) trade names (English) za-, pere-, do-, ot- (Russian) Deictic Cx. (English) risk (English) house (English) start – begin (English) clothing terms (Dutch) see (English) contact (English) answer (English) up – down (English) front – back (English) Perfectivisers (Polish) run (English) por – para (Spanish) liegen – stehen (German) epistemic modifiers (Dutch) causatives (French) negation (English) kill (English) verb particle Cxs (English) beer terms (Dutch, French) football terms (Dutch) anaphoric nouns (English)

verb verb adj. nouns prefix constr. verb noun verb noun verb verb verb prep. prep. prefix verb prep. verb verb constr. various. verb constr. noun noun noun

concrete concrete concrete concrete schematic schematic concrete concrete concrete concrete concrete concrete concrete concrete concrete schematic concrete concrete concrete concrete schematic concrete concrete schematic concrete concrete concrete

Fillmore (1977) Dirven et al. (1982) Lehrer (1982) Vorlat (1985) Janda (1986) Lakoff (1987) Fillmore and Atkins (1992) Schmid (1993) Schmid (1993) Geeraerts et al. (1994) Atkins (1994) Dirven (1994) Rudzka-Ostyn (1995) Boers (1996) Boers (1996) Dąbrowska (1996) Taylor (1996) Delbeque (1996) Bornetto (1996) Sanders and Spooren (1996) Archard (1996) Lewandowska-Tomaszczyk (1996) Lemmens (1998) Gries (1999) Geeraerts (1999a) Geeraerts et al. (1999) Schmid (2000)

introspection observation elicitation observation introspection introspective observation elicitation observation observation observation introspection observation introspection introspection introspection introspection introspection introspection introspection introspection introspection observation observation observation observation observation

Metaphor to the turn of the century, where the field becomes extremely diverse and empirical methods begin to become the norm. In order to operationalise the criterion of what constitutes ‘cognitive’ research in the field, the survey is restricted to three publication avenues. Only full articles that addressed polysemy or synonymy in the three following kinds of sources are considered: The official journal of the International Cognitive Linguistics Association, Cognitive Linguistics, between 1990 and 1999. ii. Five foundational anthologies within the paradigm, three of which are proceedings of the first three conferences of the International Cognitive Linguistics Society (Rudzka-Ostyn and Geiger 1993; de Stadler and Eyrich 1993; and Casad i.


1996) and two of which predate the society but, at the time, were constitutive of the community (Paprotté and Dirven 1985 and Rudzka-Ostyn 1988). iii. Nine anthologies largely devoted to the Cognitive Linguistic research on polysemy and synonymy, published between 1989 and 1996 (Geeraerts 1989; Tsohatzidis 1990; Dubois 1991; Rauh 1991; Lehrer and Kittay 1992; Zelinsky-Â�Wibbelt 1993; Schwarz 1994; Taylor and MacLaury 1995; and Pütz and Dirven 1996). Monographs are not included since it is difficult to gauge the impact and importance they contributed to the field but also because they often contain multiple case studies. Another issue is how to determine what constitutes an example of the radial network approach to semantic relations. This is determined using the three-part definition offered above: employing encyclopaedic semantic features, without the notion of necessary and sufficient conditions for category membership, in order to distinguish senses or relate forms. Each of the 126 studies are categorised for year of publication, type of publication and author. The object of study for each study is also categorised as: – schematic vs. concrete (lexemes vs. constructions/grammatical categories) – polysemy vs. synonymy – linguistic phenomenon (actual form(s) under investigation) – language Lastly, the studies are categorised for their method of analysis. Three kinds of method are distinguished: introspection, observation, and elicitation. Two of these methods are further distinguished. For observational data (corpus based) methods, three methods are distinguished: – corpus-driven with statistical verification – corpus-driven with raw counts – corpus-illustrated (introspection exemplified with natural data) For elicited methods, another three methods are distinguished: – quantified direct elicitation (questionnaires etc.) with raw counts – quantified direct elicitation with statistical verification – experimental elicitation with statistical verification Distinguishing introspective, experimental and observational methods is unproblematic, save when a study uses more than one method, as in Dirven and Taylor (1988). In this case, for instance, the study is categorised as elicited, since these data feature more prominently than the corpus data in the analysis. The most striking result is just how balanced and broad the range of studies is. Although obviously Eurocentric, a surprisingly wide range of languages is considered (in total 30 different languages, with English making up 36% of the studies). Moreover,

23

24

Dylan Glynn

18 16 14 12

Concrete

10

Schematic

8

Polysemy Synonymy

6 4 2 0 1985 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999

Figure 1.â•‡ Object of study in Cognitive Semantics, 1985–1999

despite the predilection for spatial prepositions, particles, and morphemes, these parts of speech represent ‘only’ 18 (14%) of the studies. Although 14% represents a sizable proportion of the research, considering that the approach is often termed preposition research and considering that two of the anthologies were devoted to prepositions and another two to spatial representation, this figure is not as overwhelming as one might expect. However, the best indicator of the diversity of the research is found if we take a more coarse-grained perspective and consider schematic vs. concrete and synonymous and vs. polysemous objects of study. Figure 1 represents the numbers of such studies over the 15-year period. We see here how evenly dispersed the four different objects of study are over the period. Only in 1993 do we see a divergence, where the number of studies examining schematic forms such as grammatical constructions and grammatical categories drop when other objects of study remain steady or increase in number. Indeed, divided in this manner, not even the analysis of concrete instances of polysemy, such as prepositions, is even vaguely dominant. Although this tells us nothing of the methodological heritage, it is important to note that radial network analysis was not restricted to polysemy studies of spatial prepositions, and that the diverse range of lexical and grammatical analysis of both near-synonymy and vague-polysemy visible today is directly descendent from this tradition. Turning to the methodological trends, we see an important and consistent presence of empirical studies, even if the use of introspection dominates. Figure 2 depicts these trends. Two levels of granularity are summarised in a single plot: a simple distinction between empirical and introspective as well as a breakdown of empirical into elicited and observational. Figure 2 shows how empirical methods followed the trends in introspective studies and were, although less common, far from irrelevant. The


18 16 14 12

Empirical

10

Introspection

8

Observation

6

Elicitation

4 2 0

1985 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999

Figure 2.â•‡ Method of study in Cognitive Semantics, 1985–1999

exception to this is, again, 1993, where we see a large number of introspective studies and no published observation or elicitation-based research in the sample. At a more fine-grained level, there is a striking lack of statistical sophistication in the observation-based research. In the sample of studies, only a single corpus analysis, Gries (1999), showed any sophistication, the entire body of corpus-driven research being restricted to raw counts. Although we know this was not completely the case since studies outside the narrow range under consideration employed statistical techniques from the beginning of the research paradigm, the relative trend is clear – the use of statistical techniques to consider results of corpus-driven research was not the norm. The use of statistical techniques for elicited data, however, was more common, though far from typical and appearing quite late in the sample, Schulze (1991), Chaffin (1992), and Myers (1994) being the earliest instances. Although based on a limited sample of 126 studies from the official journal and a selection of anthologies, the survey hopefully demonstrates that the semantic tradition within Cognitive Linguistics is not restricted to prepositional studies nor is its methodology overwhelmingly introspective. Perhaps the image that Cognitive Linguistics is exclusively orientated towards introspective methodology results from the most widely known theoretical works that founded the paradigm. Certainly the seminal publications of Fillmore (1985), Talmy (1985), Lakoff (1987), and Langacker (1987) restrict themselves to introspective investigation. The importance of these theoretical and foundational works notwithstanding, we see above that the research community as a whole always included significant methodological diversity. Quantitative corpus-driven methods are, therefore, not a new turn or a new direction, but the natural development of an existing tradition. The rich descriptive heritage summarised in Tables 2 and 3 represents the Cognitive Semantic approach to polysemy and synonymy research. This volume continues the tradition, but instead of turning to prototype set theory as the analytical framework to capture the complexity of semantic relations, the corpus-driven research presented

25

26 Dylan Glynn

here uses multivariate statistical modelling and collocation association measures. Does this mean that we have done away with radial network analysis, prototype category structure and fuzzy set structure? It would be possible to dismiss the introspective radial network analyses of the early period of Cognitive Linguistics as little more than exercises in prototype set theory. After all, it has been shown that the method of analysis was largely ad hoc (Sandra and Rice 1995) and the last decade has seen few important studies employing the approach. Indeed, some in the empirical research community would seek to distance themselves from such a research tradition.12 However, to turn our back on this tradition or seek to demonstrate the superiority of the current methodology would do injustice to this previous research. Firstly, it is precisely this research tradition that freed the study of semantic relations from the notions of discrete senses and context independent semantics. Radial network studies were the first and essential step towards this realisation – both theoretically and analytically. Theoretically, it set the stage for an understanding of semantics as emergent structure and, analytically, it produced the idea of understanding meaning as a network of interacting factors. On both fronts, the corpus-driven and experimental study of semantic relations is a direct heir to radial network research. Indeed, within the tradition, as early as Geeraerts (1993a:â•›260), the idea that we need to move away from a reified and mono-dimensional understanding of meaning was being overtly mooted.13 Secondly, such studies are an essential step in empirical research. They represent hypothetical models of language structure, based on careful and systematic introspection-based analysis of language. Rather than ignore, or worse still, dismiss such theoretical research, empirical analysis needs to treat it as foundational. The results of introspection-based research are theoretical models, models that are, most likely, reasonably accurate descriptions of language structure. The new generation of experimental and observational linguistic analysis needs to test the accuracy and explanatory power of those models, modifying them where needed. Seen in this light, the configuration of the argument-structure of the verbs of ‘buying and selling’ by 12. There is much discussion about an empirical turn/revolution in Cognitive Linguistics (Geeraerts 2006b, 2010b; Levshina 2011; Klavan 2012 inter alios). Indeed, certain authors have expressed concerns about “empirical imperialism” (Lampert and Lampert 2010; Schmid 2010), a term coined by Geeraerts (2006a). Some of the discussion on the topic (Stefanowitsch 2008; Fischer 2010) represents the ‘turn’ as a break from the past and indeed Gries (forthc.) considers corpus-driven and experimentation-based research in polysemy as an entirely different discipline from research using prototype set theory. Whether it is a result of the gradual emergence of empiricism or a radical methodological revolution, there is little doubt that the analytical landscape of Cognitive Semantics has changed substantially over the last 25 years. 13. This line of thinking is not as radical as one might expect. Lehrer and Lehrer (1994) and Victorri and Fuchs (1996) represent further examples of early discussions on how a non-reified understanding of semantic structure needs to be developed.


Fillmore (1977), the schema specifications in Lakoff ’s (1987) study of over, or the image-schematic grammatical features of Janda’s (1993) study of the Dative are theoretical models and it is our task to test their descriptive accuracy and improve them. There are, of course, differences between the corpus-driven quantitative research and the introspection-based radial network studies. No matter how large a corpus, found data will always be biased towards what is common rather than what is possible. Introspection is a vital methodology for proposing hypotheses about what is possible in a language. It follows that a lack of corpus evidence of a given form-meaning pair does not mean it is not possible or does not occur. This is simply because even the largest corpus in the world is but a microscopic fraction of actual language use. This difference affects the results profoundly. Corpus-driven research is exclusively frequency-based, and this, in turn, will prioritise typical structures over less typical structures. It is for this reason that it would be difficult to simply carry out a corpus-driven study on over and compare the results with Lakoff ’s (1987) results. A corpus-driven study may or may not confirm parts of the network analysis, but it is unlikely that it would paint the same picture anymore than lack of confirmation would negate his hypothesis. This is simply because only the most frequent usage patterns, or schema configurations to use Lakoff ’s terminology, would be found. Another apparent difference would be the lack of interest in prototype effects. In the corpus-driven research, what has become of the analytical apparatus – prototype categorisation and fuzzy sets? Although there is little overt reference to such notions, the results of multifactorial analysis and collocation analysis are, in fact, structured as fuzzy-bounded prototype categories. Since the results are based upon relative frequency, they are, therefore, necessarily ‘prototype’ structured (at least if we accept a frequency-based operationalisation of prototypicality) and are not discrete. Moreover, multifactorial feature analysis identifies ‘meanings’ as tendencies, where a tendency is a multidimensional pattern of use. This, quite literally, produces networks of different uses – a frequency-based and complex multidimensional network of sense relations. The radial network analysis produced prototype maps of meaning upon one ‘semantic’ dimension where the ‘nodes’ were discrete reified senses. Today, multifactorial feature analysis produces multidimensional networks of usage patterns that can be interpreted as emergent language structure. The difference between the two is a natural progression, not a methodological schism. We can, therefore, deduce that, despite important differences, there is a direct line of descent in the methodology from the radial network approach described above to the contemporary corpus-driven research represented by studies such as Gries (2003, 2006), Heylen (2005), Divjak and Gries (2006), Divjak (2006, 2010a, 2010b), Wulff (2006), Gronderlaers et al. (2007, 2008), Wulff et al. (2007), Janda and Solovyev (2009), Hilpert (2008), Gilquin (2009), Glynn (2009, 2010a, 2010b, 2014a, 2014b, forthc.), Speelman and Geeraerts (2010), Krawczak and Kokorniak (2012), Krawczak (2014a, 2014b), and the research in this volume. Armed with contemporary corpus-driven

27

28

Dylan Glynn

methods, the task ahead of us now is, arguably, to return to the research questions that gave rise to Cognitive Semantics in an effort to understand the conceptual and functional structures that motivate language. Although contemporary quantitative research may be gaining descriptive adequacy, our ultimate goal remains explanatory adequacy.

References Apresjan, J. D. (1974). Лексическая Семантика. Синонимические средства языка [Lexical Semantics: Synonymous foundations of language]. Moscow: Nauka. Apresjan, J. D. (2000). Systematic lexicography. Oxford: Oxford University Press. Arppe, A., Gilquin, G., Glynn, D., Hilpert, M., & Zeschel, A. (2010). Cognitive corpus linguistics: Five points of debate on current theory and methodology. Corpora, 5, 1–27. DOI: 10.3366/cor.2010.0001 Atkins, B. (1994). Analyzing the verbs of seeing: A frame semantics approach to corpus lexicography. Proceedings of the Twentieth Annual Meeting of the Berkeley Linguistics Society, 42–56. Barthélemy, J.-P. (1991). Similitude, arbres, et typicalité. In D. Dubois (Ed.), Sémantique et cognition: catégories, prototypes, typicalité (pp. 205–224). Paris: Centre national de la recherche scientifique. Bartmiński, J. (2008). Aspects of cognitive ethnolinguistics. London: Equinox. Bellavia, E. (1996). The German über. In M. Pütz, & R. Dirven (Eds.), The construal of space in language and thought (pp. 73–107). Berlin & New York: Mouton de Gruyter. Boers, F. (1996). Spatial prepositions and metaphor: A Cognitive Semantic journey along the updown and front-back dimensions. Tübignen: Gunter Narr. Bondarko, A. V. (1983). Принципы функциональной грамматики и вопросы аспектологии [Principles of functional grammar and questions of aspectology]. Lenningrad: Nauka. Bondarko, A. V. (1991). Functional grammar: A field approach. Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/llsee.35 Brugman, C. (1983a). The story of over: Polysemy, semantics, and the structure of the lexicon. Trier: LAUT. Brugman, C. (1983b). How to be in the know about on the go. Proceedings of the Chicago Linguistics Society, 19, 64–76. Brugman, C. (1984). The very idea: A case study in polysemy and cross-lexical generalizations. Proceedings of the Chicago Linguistics Society, 20, 21–38. Bybee, J. (2007). Frequency of use and the organization of language. Oxford: Oxford University Press. DOI: 10.1093/acprof:oso/9780195301571.001.0001 Casad, E. (Ed.). (1996). Cognitive Linguistics in the redwoods. The expansion of a new paradigm. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110811421 Chaffin, R. (1992). The concept of a semantic relation. In A. Lehrer, & E. Kittay (Eds.), Frames, fields, and contrasts: New essays in semantic and lexical organisation (pp. 253–288). London: Lawrence Erlbaum. Cienki, A. (1998). Straight: An image schema and its metaphorical extensions. Cognitive Linguistics, 9, 107–150. DOI: 10.1515/cogl.1998.9.2.107 Coleman, L., & Kay, P. (1981). Prototype semantics: The English word lie. Language, 57, 26–44. Coșeriu, E. (1980). Textlinguistik. Tübingen: Gunter Narr.


Croft, W. (1998). Linguistic evidence and mental representations. Cognitive Linguistics, 9, 151– 173. DOI: 10.1515/cogl.1998.9.2.151 Croft, W. (2009). Toward a social Cognitive Linguistics. In V. Evans, & S. Pourcel (Eds.), New directions in Cognitive Linguistics (pp. 395–420). Amsterdam & Philadelphia: John Benjamins. Cruse, A. (2000). Aspects of the micro-structure of word meanings. In Y. Ravin, & C. Leacock (Eds.), Polysemy: Theoretical and computation approaches (pp. 30–51). Oxford: Oxford University Press. Culioli, A. (1990). Pour une linguistique de l’énonciation: Opérations et représentations. Paris: Ophrys. Cuyckens, H. (1991). The semantics of spatial prepositions in Dutch. Unpublished PhD dissertation, University of Antwerp. Cuyckens, H. (1993). The Dutch spatial preposition “in”: A cognitive-semantic analysis. In C. Zelinsky-Wibbelt (Ed.), The semantics of prepositions: From mental processing to natural language processing (pp. 27–72). Berlin & New York: Mouton de Gruyter. Cuyckens, H. (1994). Family resemblance in the Dutch spatial preposition op. In M. Schwarz (Ed.), Kognitive Semantik: Ergebnisse, Probleme, Perspektiven (pp. 179–196). Tübingen: Gunter Narr. Cuyckens, H. (1995). Family resemblance in the Dutch spatial prepositions Door and Langs. Cognitive Linguistics, 6, 183–207. DOI: 10.1515/cogl.1995.6.2-3.183 Cuyckens, H., Sandra, D., & Rice, S. (1997). Towards an empirical lexical semantics. In B. Smieja, & M. Tasch (Eds.), Human contact through language and linguistics (pp. 35–54). Frankfurt/Main: Peter Lang. Cuyckens, H., & Zawada, B. (Eds.). (2001). Polysemy in Cognitive Linguistics. Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/cilt.177 Cuyckens, H., & Radden, G. (Eds.). (2002). Perspectives on prepositions. Tübignen: Max Niemeyer. DOI: 10.1515/9783110924787 Cuyckens, H., Dirven, R., & Taylor, J. (Eds.). (2003). Cognitive approaches to lexical semantics. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110219074 Dąbrowska, E. (1994). Radial categories in grammar: The Polish instrumental case. Linguistica Silesiana, 15, 83–94. Dąbrowska, E. (1996). Temporal structuring of events: A study of Polish perfectivizing prefixes. In R. Dirven, & M. Pütz (Eds.), The construal of space in language and thought (pp. 467– 490). Berlin & New York: Mouton de Gruyter. Dąbrowska, E. (1997). Cognitive Semantics and the Polish dative. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110814781 de Stadler, L., & Eyrich, C. (Eds.). (1993). Issues in Cognitive Linguistics. Berlin & New York: Mouton de Gruyter. Deane, P. (1988). Polysemy and cognition. Lingua, 75, 325–361. DOI: 10.1016/0024-3841 (88)90009-5 Deane, P. (1993a). Multimodal spatial representation: On the semantic unity of ‘over’ and other polysemous prepositions. Duisburg: LAUD. Deane, P. (1993b). At, by, to, and past: A study in multimodal image theory. Proceedings of the Berkeley Linguistics Society, 19, 112–124. Deane, P. (2006). Multimodal spatial representation: On the semantic unity of over. In B. Hampe (Ed.), From perception to meaning: Image schemas in Cognitive Linguistics (pp. 235–284). Berlin & New York: Mouton de Gruyter.

29

30

Dylan Glynn

Delbeque, N. (1996). Towards a cognitive account of the use of the prepositions por and para in Spanish. In E. Casad (Ed.), Cognitive Linguistics in the Redwoods: The expansion of a new paradigm in linguistics (pp. 249–318). Berlin & New York: Mouton de Gruyter. Dewell, R. (1994). Over again: On the role of image–schemas in semantic analysis. Cognitive Linguistics, 5, 351–380. DOI: 10.1515/cogl.1994.5.4.351 Dewell, R. (1996). The separability of German über: A cognitive approach. In M. PuÌ‹tz, & R. Dirven (Eds.), The construal of space in language and thought (pp. 109–133). Berlin & New York: Mouton de Gruyter. Dirven, R. (1994). Cognition and semantic structure: The experiential basis of the semantic structure of verbs of body contact. In M. Schwarz (Ed.), Kognitive Semantik: Ergebnisse, Probleme, Perspektiven (pp. 131–145). Tübingen: Gunter Narr. Dirven, R., & Taylor, J. (1988). The conceptualisation of vertical space in English: The case of tall. In B. Rudzka-Ostyn (Ed.), Topics in Cognitive Linguistics (pp. 379–402). Amsterdam & Philadelphia: John Benjamins. Dirven, R., Goossens, L., Putseys, Y., & Vorlat, E. (1982). The scene of linguistic action and its perspectivization by speak, talk, say, and tell. Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/pb.iii.6 Dirven, R., & Vanparys, J. (Eds.). (1995). Current approaches to the lexicon. Frankfurt/Main: Peter Lang. Divjak, D. (2006). Ways of intending: A corpus-based Cognitive Linguistic approach to near-synonyms in Russian. In St. Th. Gries, & A. Stefanowitsch (Eds.), Corpora in Cognitive Linguistics: Corpus-based approaches to syntax and lexis (pp. 19–56). Berlin & New York: Mouton de Gruyter. Divjak, D. (2010a). Structuring the lexicon: A clustered model for near-synonymy. Berlin & New York: Mouton de Gruyter. Divjak, D. (2010b). Corpus-based evidence for an idiosyncratic aspect-modality relation in Russian. In D. Glynn, & K. Fischer (Eds.), Quantitative Cognitive Semantics: Corpus-driven approaches (pp. 305–331). Berlin & New York: Mouton de Gruyter. Divjak, D., & Gries, St. Th. (2006). Ways of trying in Russian: Clustering behavioral profiles. Corpus Linguistics and Linguistic Theory, 2, 23–60. DOI: 10.1515/CLLT.2006.002 Dubois, D. (Ed.). (1991). Sémantique et cognition: Catégories, prototypes, typicalité. Paris: Centre national de la recherche scientifique. Dunbar, G. (1991). The cognitive lexicon. Tübingen: Gunter Narr. Dunbar, G. (2001). Toward a cognitive analysis of polysemy, ambiguity, and vagueness. Cognitive Linguistics, 12, 1–14. DOI: 10.1515/cogl.12.1.1 Evans, V. (2005). The meaning of time: Polysemy, the lexicon and conceptual structure. Journal of Linguistics, 41, 33–75. DOI: 10.1017/S0022226704003056 Evans, V. (2006). Lexical concepts, cognitive models and meaning-construction. Cognitive Linguistics, 17, 491–534. DOI: 10.1515/COG.2006.016 Fauconnier, G., & Turner, M. (1998). Conceptual integration networks. Cognitive Science, 22, 133–187. DOI: 10.1207/s15516709cog2202_1 Fillmore, C. (1975). An alternative to checklist theories of meaning. Proceedings of the Berkeley Linguistics Society, 1, 123–131. Fillmore, C. (1977). Topics in lexical semantics. In P. Cole (Ed.), Current issues in linguistic theory (pp. 76–138). Bloomington: Indiana University Press. Fillmore, C. (1985). Frames and the semantics of understanding. Quaderni di Semantica, 6, 222–254.


Fillmore, C. (2000). Describing polysemy: The case of ‘crawl’. In Y. Ravin, & C. Leacock (Eds.), Polysemy: Theoretical and computation approaches (pp. 91–110). Oxford: Oxford University Press. Fillmore, C., Kay, P., & O’Connor, M. (1988). Regularity and idiomaticity in grammatical constructions: The case of let alone. Language, 64, 501–538. DOI: 10.2307/414531 Fillmore, C., & Atkins, B. (1992). Toward a frame-based lexicon: The semantics of risk and its neighbours. In A. Lehrer, & E. Kittay (Eds.), Frames, fields, and contrasts: New essays in semantic and lexical organisation (pp. 75–102). London: Lawrence Erlbaum. Fischer, K. (2010). Quantitative methods in cognitive semantics. In D. Glynn, & K. Fischer (Eds.), Quantitative Cognitive Semantics: Corpus-driven approaches (pp. 43–61). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110226423 Geeraerts, D. (1987). On necessary and sufficient conditions. Journal of Semantics, 5, 275–291. DOI: 10.1093/jos/5.4.275 Geeraerts, D. (1988). Where does prototypicality come from? In B. Rudzka-Ostyn (Ed.), Topics in Cognitive Linguistics. Amsterdam & Philadelphia: John Benjamins. Geeraerts, D. (1989). Prospects and problems of prototype theory. Linguistics, 27, 587–612. DOI: 10.1515/ling.1989.27.4.587 Geeraerts, D. (1990). The lexicographical treatment of prototypical polysemy. In S. Tsohatzidis (Ed.), Meanings and prototypes: Studies in linguistic categorization (pp. 195–210). London: Routledge. Geeraerts, D. (1992). The semantic structure of Dutch over. Leuvense Bijdragen, 81, 205–230. Geeraerts, D. (1993a). Vagueness’s puzzles, polysemy’s vagaries. Cognitive Linguistics, 4, 223– 72. DOI: 10.1515/cogl.1993.4.3.223 Geeraerts, D. (1993b). Generalised onomasiological salience. In J. Nuyts, & E. Pederson (Eds.), Perspectives on language and conceptualization (Special edition of the Belgian Journal of Linguistics, 8) (pp. 43–56). Brussels: Editions de l’Université de Bruxelles. Geeraerts, D. (1994). Classical definability and the monosemic bias. Rivista di Linguistica, 6, 149–172. Geeraerts, D. (1995). Representational formats in Cognitive Semantics. Folia Linguistica, 39, 21–41. Geeraerts, D. (1997). Diachronic prototype semantics: A contribution to historical lexicology. Oxford: Clarendon Press. Geeraerts, D. (1998). The semantic structure of the indirect object in Dutch. In W. Van Langendonck, & W. Van Belle (Eds.), The dative. Vol. 2. Theoretical and contrastive studies (pp. 185–210). Amsterdam & Philadelphia: John Benjamins. Geeraerts, D. (1999a). Beer and semantics. In L. De Stadler, & C. Eyrich (Eds.), Issues in Cognitive Linguistics (pp. 35–55). Berlin & New York: Mouton de Gruyter. Geeraerts, D. (1999b). Idealist and empiricist tendencies in Cognitive Semantics. In T. Janssen, & G. Redeker (Eds.), Cognitive Linguistics: Foundations, scope, and methodology (pp. 163–194). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110803464.163 Geeraerts, D. (2005). Lectal data and empirical variation in Cognitive Linguistics. In F. José Ruiz de Mendoza Ibáñez, & S. Peña Cervel (Eds.), Cognitive Linguistics: Internal dynamics and interdisciplinary interactions (pp. 163–189). Berlin & New York: Mouton de Gruyter. Geeraerts, D. (2006a). Words and other wonders: Papers on lexical and semantic topics. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110219128

31

32

Dylan Glynn

Geeraerts, D. (2006b). Methodology in Cognitive Linguistics. In G. Kristiansen, M. Achard, R. Dirven, & F. J. Ruiz de Mendoza Ibañez (Eds.), Cognitive Linguistics: Current applications and future perspectives (pp. 21–50). Berlin & New York: Mouton de Gruyter. Geeraerts, D. (2010a). Theories of lexical semantics. Oxford: Oxford University Press. Geeraerts, D. (2010b). Recontextualizing grammar: Underlying trends in thirty years of Cognitive Linguistics. In E. Tabakowska, M. Choinski, & L. Wiraszka (Eds.), Cognitive Linguistics in action: From theory to application and back (pp. 71–102). Berlin & New York: Mouton de Gruyter. Geeraerts, D., Grondelaers, St., & Bakema, P. (1994). The structure of lexical variation: Meaning, naming, and context. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110873061 Geeraerts, D., Grondelaers, St., & Speelman, D. (1999). Convergentie en divergentie in de Nederlandse woordenschat. Amsterdam: Meertens Instituut. Geeraerts, D. (Ed.) (1989). Prospects and problems of prototype theory (Special edition of Linguistics, 27). Berlin & New York: Mouton de Gruyter. Givón, T. (1982). Evidentiality and epistemic space. Studies in Language, 6, 23–39. DOI: 10.1075/sl.6.1.03giv Givón, T. (2005). Context as other minds: The pragmatics of sociality, cognition and communication. Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/z.130 Glynn, D. (2009). Polysemy, syntax, and variation: A usage-based method for Cognitive Semantics. In V. Evans, & S. Pourcel (Eds.), New directions in Cognitive Linguistics (pp. 77– 106). Amsterdam & Philadelphia: John Benjamins. Glynn, D. (2010a). Synonymy, lexical fields, and grammatical constructions: A study in usage-based Cognitive Semantics. In H.-J. Schmid, & S. Handl (Eds.), Cognitive foundations of linguistic usage-patterns: Empirical studies (pp. 89–118). Berlin & New York: Mouton de Gruyter. Glynn, D. (2010b). Testing the hypothesis: Objectivity and verification in usage-based Cognitive Semantics. In D. Glynn, & K. Fischer (Eds.), Quantitative Cognitive Semantics: Corpus-driven approaches (pp. 239–270). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110226423 Glynn, D. (2010c). Corpus-driven Cognitive Semantics: An overview of the field. In D. Glynn, & K. Fischer (Eds.), Quantitative Cognitive Semantics: Corpus-driven approaches (pp. 1– 42). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110226423.1 Glynn, D. (2014a). The conceptual profile of the lexeme home: A multifactorial diachronic analysis. In J. E. Díaz-Vera (Ed.), Metaphor and metonymy across time and cultures (pp. 265– 293). Berlin & New York: Mouton de Gruyter. Glynn, D. (2014b). The social nature of anger: Multivariate corpus evidence for context effects upon conceptual structure. In I. Novakova, P. Blumenthal, & D. Siepmann (Eds.), Emotions in discourse (pp. 69–82). Frankfurt/Main: Peter Lang. Glynn, D. (Forthcoming). Mapping meaning: Corpus methods for Cognitive Semantics. Cambridge: Cambridge University Press. Glynn, D., & Fischer, K. (Eds.) (2010). Quantitative Cognitive Semantics: Corpus-driven approaches. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110226423 Goldberg, A. (1991). A semantic account of resultatives. Linguistic Analysis, 21, 66–96. Goldberg, A. (1992). The inherent semantics of argument structure: The case of the English ditransitive construction. Cognitive Linguistics, 3, 37–74. DOI: 10.1515/cogl.1992.3.1.37 Goldberg, A. (1995). Constructions: A construction grammar approach to argument structure. London: University of Chicago Press.


Goldberg, A. (2002). Surface generalization: An alternative to alternations. Cognitive Linguistics, 13, 327–356. DOI: 10.1515/cogl.2002.022 Gries, St. Th. (1999). Particle movement: A cognitive and functional approach. Cognitive Linguistics, 10, 105–145. DOI: 10.1515/cogl.1999.005 Gries, St. Th. (2003). Multifactorial analysis in corpus linguistics: A study of particle placement. London & New York: Continuum Press. Gries, St. Th. (2006). Corpus-based methods and Cognitive Semantics: The many senses of to run. In St. Th. Gries, & A. Stefanowitsch (Eds.), Corpora in Cognitive Linguistics: Corpus-based approaches to syntax and lexis (pp. 57–99). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110197709 Gries, St. Th. (Forthcoming). Polysemy. In E. Dąbrowska, & D. Divjak (Eds.), Handbook of Cognitive Linguistics. Berlin & New York: Mouton de Gruyter. Gries, St. Th., & Stefanowitsch, A. (Eds.). (2006). Corpora in Cognitive Linguistics: Corpus-based approaches to syntax and lexis. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110197709 Grondelaers, St., & Geeraerts, D. (2003). Towards a pragmatic model of cognitive onomasiology. In H. Cuyckens, R. Dirven, & J. Taylor (Eds.). Cognitive approaches to lexical semantics (pp. 67–92). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110219074.67 Halliday, M. (1967). Notes on transitivity and theme in English. Journal of Linguistics, 3, 37–81. DOI: 10.1017/S0022226700012949 Halliday, M. (1985). An introduction to Functional Grammar. London: Edward Arnold. Hawkins, B. (1985). The semantics of English spatial prepositions. Trier: LAUT. Herskovits, A. (1986). Language and spatial cognition: An interdisciplinary study of the prepositions in English. Cambridge: Cambridge University Press. Herskovits, A. (1988). Spatial expressions and the plasticity of meaning. In B. Rudzka-Ostyn (Ed.), Topics in Cognitive Linguistics (pp. 271–297). Amsterdam & Philadelphia: John Benjamins. Hopper, P. (1987). Emergent grammar. Berkeley Linguistics Society, 13, 139–157. Janda, L. (1986). A semantic analysis of the Russian verbal prefixes za-, pere-, do-, and ot-. Munich: Otto Sanger. Janda, L. (1990). Radial network of a grammatical category – its genesis and dynamic structure. Cognitive Linguistics, 1, 269–288. DOI: 10.1515/cogl.1990.1.3.269 Janda, L. (1993). A geography of case semantics: The Czech dative and the Russian instrumental. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110867930 Janda, L., & Solovyev, V. (2009). What constructional profiles reveal about synonymy: A case study of the Russian words for sadness and happiness. Cognitive Linguistics, 20, 367–393. DOI: 10.1515/COGL.2009.018 Jones, S. (2002). Antonymy: A corpus-based approach. London: Routledge. Kastovsky, D. (1982). Wortbildung und Semantik. Düsseldorf: Francke. Kay, P. (1984). The kind of/sort of construction. Proceedings of the Berkeley Linguistics Society, 10, 128–137. Kay, P., & Fillmore, C. (1999). Grammatical constructions and linguistic generalizations: The What’s X doing Y? construction. Language, 75, 1–33. DOI: 10.2307/417472 Kilgarriff, A. (1997). I don’t believe in word senses: Computers and the Humanities, 31, 91–113. DOI: 10.1023/A:1000583911091 Kittay, E., & Lehrer, A. (1981). Semantic fields and the structure of metaphor. Studies in Language, 5, 31–63. DOI: 10.1075/sl.5.1.03kit

33

34

Dylan Glynn

Klavan, J. (2012). Converging and diverging evidence: Corpus-linguistic and experimental methods for studying grammatical synonymy. Unpublished PhD dissertation, University of Tartu. Kleiber, G. (1990). Sémantique du prototype: Catégorie et sens lexical. Paris: Presses Universitaires de France. Kleiber, G. (1999). Problèmes de sémantique: La polysémie en questions. Villeneuve-d’Ascq: Presses universitaires du Septentrion. Krawczak, K. (2014a). Epistemic stance predicates in English: A quantitative corpus-driven study of subjectivity. In D. Glynn, & M. Sjölin (Eds.), Subjectivity and epistemicity: Corpus, discourse, and literary approaches to stance (pp. 355–386). Lund: Lund University Press. Krawczak, K. (2014b). Shame and its near-synonyms in English: A multivariate corpus-driven approach to social emotions. In I. Novakova, P. Blumenthal, & D. Siepmann (Eds.), Emotions in discourse (pp. 84–94). Frankfurt/Main: Peter Lang. Krawczak, K., & Kokorniak, I. (2012). A corpus-driven quantitative approach to the construal of Polish think. Poznań Studies in Contemporary Linguistics, 48, 439–472. DOI: 10.1515/psicl-2012-0021 Kreitzer, A. (1997). Multiple levels of schematization: A study in the conceptualization of space. Cognitive Linguistics, 8, 291–325. DOI: 10.1515/cogl.1997.8.4.291 Lakoff, G. (1975). Hedges: A study in meaning criteria and the logic of fuzzy concepts. Journal of Philosophical Logic, 2, 458–508. Lakoff, G. (1977). Linguistic gestalts. Proceedings of the Chicago Linguistics Society, 13, 236–287. Lakoff, G. (1982). Categories: An essay in Cognitive Linguistics. In Linguistic Society of Korea (Ed.), Linguistics in the morning calm (pp. 139–194). Seoul: Hanshin. Lakoff, G. (1987). Women, fire, and dangerous things: What categories reveal about the mind. London: University of Chicago Press. DOI: 10.7208/chicago/9780226471013.001.0001 Langacker, R. (1982). Space grammar, analysability, and the English passive. Language, 58, 22–80. DOI: 10.2307/413531 Langacker, R. (1987). Foundations of Cognitive Grammar. Vol. 1. Theoretical prerequisites. Stanford: Stanford University Press. Langacker, R. (1991). Foundations of Cognitive Grammar. Vol. 2. Descriptive application. Stanford: Stanford University Press. Lehrer, A. (1982). Wine and conversation. Bloomington: Indiana University Press. Lehrer, A. (1990a). Polysemy, conventionality, and the structure of the lexicon. Cognitive Linguistics, 1, 207–246. DOI: 10.1515/cogl.1990.1.2.207 Lehrer, A. (1990b). Prototype theory and its implication for lexical analyses. In S. Tsohatzidis (Ed.), Meanings and prototypes: Studies in linguistic categorization (pp. 368–381). London: Routledge. Lehrer, K., & Lehrer, A. (1994). Fields, networks, and vectors. In F. Palmer (Ed.), Grammar and meaning: A festschrift for John Lyons (pp. 26–47). Cambridge: Cambridge University Press. Lehrer, A., & Kittay, E. (Eds.). (1992). Frames, fields, and contrasts: New essays in semantic and lexical organization. Hillsdale: Lawrence Erlbaum. Lemmens, M. (1998). Lexical perspectives on transitivity and ergativity: Causative constructions in English. Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/cilt.166 Levshina, N. (2011). A usage-based study of Dutch causative constructions. Unpublished PhD dissertation, University of Leuven. Lewandowska-Tomaszczyk, B. (1996). Depth of negation: A cognitive semantic study. Łódź: Łódź University Press.


Liamkina, O. (2007). Semantic structure of the German spatial particle über. Journal of Germanic Linguistics, 19, 115–160. DOI: 10.1017/S1470542707000050 Lindner, S. (1983). A lexico-semantic analysis of English verb-particle constructions with up and out. Trier: LAUT. Lipka, L. (1992). An outline of English lexicology. Tübingen: Max Niemeyer. Lutzeier, P. (1985). Linguistische Semantik. Stuttgart: J. B. Metzler. Lyons, J. (1968). Introduction to theoretical linguistics. Cambridge: Cambridge University Press. DOI: 10.1017/CBO9781139165570 Meex, B. (2001). The spatial and non-spatial sense of the German preposition über. In H. Cuyckens, & B. Zawada (Eds.), Polysemy in Cognitive Linguistics (pp. 1–36). Amsterdam & Philadelphia: John Benjamins. Mel’čuk, I. A. (1989). Semantic primitives from the viewpoint of meaning-text linguistic theory. Quaderni di Semantica, 10, 65–102. Melis, L. (1990). La voie pronominale: La systématique des tours pronominaux en français moderne. Paris: Duclot. Morgan, P. (1997). Figuring out figure out: Metaphor and the semantics of the English verb particle construction. Cognitive Linguistics, 8, 327–358. DOI: 10.1515/cogl.1997.8.4.327 Murphy, L. (2003). Semantic relations and the lexicon: Antonymy, synonymy, and other paradigms. Cambridge: Cambridge University Press. DOI: 10.1017/CBO9780511486494 Myers, D. (1994). Testing for prototypicality: The Chinese morpheme gong. Cognitive Linguistics, 5, 261–280. DOI: 10.1515/cogl.1994.5.3.261 Nerlich, B., Todd, Z., Herman, V., & Clarke, D. (Eds.). (2003). Polysemy: Flexible patterns of meaning in mind. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110895698 Newman, J. (1993). The semantics of giving in Mandarin. In B. Rudzka-Ostyn (Ed.), Topics in Cognitive Linguistics (pp. 433–486). Amsterdam & Philadelphia: John Benjamins. Norvig, P., & Lakoff, G. (1987). Taking: A study in lexical network theory. Proceedings of the Berkeley Linguistics Society, 13, 195–206. Paprotté, W., & Dirven, R. (Eds.). (1985). Ubiquity of metaphor: Metaphor in language and thought. Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/cilt.29 Pütz, M., & Dirven, R. (Eds.). (1996). The construal of space in language and thought. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110821611 Rakova, M., Pethő, G., & RaÌ†kosi, C. (Eds.). (2007). The cognitive basis of polysemy: New sources of evidence for theories of word meaning. Frankfurt/Main: Peter Lang. Rastier, F. (1987). Sémantique interprétative. Paris: Presses universitaires de France. Rastier, F. (1991). Sémantique et recherches cognitives. Paris: Presses universitaires de France. Rastier, F. (2011). La mesure et le grain: Sémantique de corpus. Paris: Honoré Champion. Rauh, G. (Ed.). (1991). Approaches to prepositions. Tübignen: Gunter Narr. Ravin, Y., & Leacock, C. (Eds.). (2000). Polysemy: Theoretical and computational approaches. Oxford: Oxford University Press. Rice, S. (1993). Far afield in the lexical fields: The English prepositions. Tübingen: Gunter Narr. Rice, S. (1999). Patterns of acquisition in the emerging mental lexicon: The case of to and for in English. Brain and Language, 68, 268–276. DOI: 10.1006/brln.1999.2105 Rice, S., Sandra, D., & Vanrespaille, M. (1999). Prepositional semantics and the fragile link between space and time. In M. Hiraga, C. Sinha, & S. Wilcox (Eds.), Cultural typology and psycholinguistic issues in Cognitive Linguistics (pp. 107–127). Amsterdam & Philadelphia: John Benjamins.

35

36

Dylan Glynn

Rudzka-Ostyn, B. (1983). Cognitive Grammar and the structure of Dutch uit and Polish wy. Linguistic Agency University of Trier: Trier. Rudzka-Ostyn, B. (1985). Metaphoric processes in word formation. In W. Paprotté, & R. Dirven (Eds.), Ubiquity of metaphor: Metaphor in language and thought (pp. 209–241). Amsterdam & Philadelphia: John Benjamins. Rudzka-Ostyn, B. (1989). Prototypes, schemas, and cross-category correspondences: The case of ask. In D. Geeraerts (Ed.), Prospects and problems of prototype theory (pp. 613–661). Berlin & New York: Mouton de Gruyter. Rudzka-Ostyn, B. (1992). Case relations in Cognitive Grammar: Some reflexive uses of the Polish dative. Leuvense Bijdragen, 81, 327–373. Rudzka-Ostyn, B. (1994). The structure of the genitive category in Polish. Proceedings of the LAUD International Symposium Language and Space, Duisburg. Republished in Rudzka-Â� Ostyn (2000: Chapter 6). Rudzka-Ostyn, B. (1995). Metaphor, schema, invariance: The case of verbs of answering. In L. Goossens, P. Pauwels, B. Rudzka-Ostyn, A.-M. Simon-Vandenbergen, & J. Vanparys (Eds.), By word of mouth: Metaphor, metonymy, and linguistic action from a cognitive perspective (pp. 205–244). Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/pbns.33 Rudzka-Ostyn, B. (1996). The Polish dative. In W. van Belle, & W. van Langendonck (Eds.), The dative. Vol. 1. Descriptive studies (pp. 341–394). Amsterdam & Philadelphia: John Benjamins. Rudzka-Ostyn, B. (2000). Z rozważań nad kategorią przypadka [Considerations on the category of case]. Kraków: Universitas. Rudzka-Ostyn, B. (Ed.). (1988). Topics in Cognitive Linguistics. Berlin & New York: Mouton de Gruyter. DOI: 10.1075/cilt.50 Rudzka-Ostyn, B., & Geiger, R. (Eds.). (1993). Conceptualizations and mental processing in language. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110857108 Sanders, J., & Spooren, W. (1996). Subjectivity and certainty in epistemic modality: A study of Dutch epistemic modifiers. Cognitive Linguistics, 7, 241–264. DOI: 10.1515/cogl.1996.7.3.241 Sandra, D., & Rice, S. (1995). Network analyses of prepositional meaning: Mirroring whose mind – the linguist’s or the language user’s? Cognitive Linguistics, 6, 89–130. DOI: 10.1515/cogl.1995.6.1.89 Schmid, H.-J. (1993). Cottage and co., idea, start vs. begin. Die Kategorisierung als Grundprinzip einer differenzierten Bedeutungsbeschreibung. Tübingen: Max Niemeyer. DOI: 10.1515/ 9783111355771 Schmid, H.-J. (2000). English abstract nouns as conceptual shells: From corpus to cognition. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110808704 Schmid, H.-J. (2010). Does frequency in text instantiate entrenchment in the cognitive system? In D. Glynn, & K. Fischer (Eds.), Quantitative Cognitive Semantics: Corpus-driven approaches (pp. 101–135). Berlin & New York: Mouton de Gruyter. Schneider, W., & Shiffrin, R. (1977). Controlled and automated human information processing, I: Detection, search and attention. Psychological Review, 84, 1–66. DOI: 10.1037/ 0033-295X.84.1.1 Schulze, R. (1988). A short story of down. In W. Hüllen, & R. Schulze (Eds.), Understanding the lexicon: Meaning, sense, and world knowledge in lexical semantics (pp. 395–414). Tübingen: Niemeyer. DOI: 10.1515/9783111355191


Schulze, R. (1991). Getting round to (a)round: Towards the description and analysis of a ‘spatial’ predicate. In G. Rauh (Ed.), Approaches to prepositions (pp. 253–74). Tübingen: Gunter Narr. Schulze, R. (1993). The meaning of (a)round: A study of an English preposition. In A. Geiger, & B. Rudzka-Ostyn (Eds.), Conceptualizations and mental processing in language (pp. 399– 432). Berlin & New York: Mouton de Gruyter. Schulze, R. (1994). Image schemata and the semantics of off. In M. Schwarz (Ed.), Kognitive Semantik: Ergebnisse, Probleme, Perspektiven (pp. 197–213). Tübingen: Gunter Narr. Schwarz, M. (Ed.). (1994). Kognitive Semantik: Ergebnisse, Probleme, Perspektiven. Tübingen: Gunter Narr. Shiffrin, R., & Schneider, W. (1977). Controlled and automatic information processing, II: Perception, learning, automatic attending and a general theory. Psychological Review, 84, 127–190. DOI: 10.1037/0033-295X.84.2.127 Speelman, D., & Geeraerts, D. (2010). Causes for causatives: The case of Dutch ‘doen’ and ‘laten’. In T. Sanders, & E. Sweetser (Eds.), Causal categories in discourse and cognition (pp. 173– 204). Berlin & New York: Mouton de Gruyter. Stefanowitsch, A. (2008). Negative entrenchment: A usage-based approach to negative evidence. Cognitive Linguistics, 19, 513–531. DOI: 10.1515/COGL.2008.020 Stefanowitsch, A. (2010). Empirical cognitive semantics: Some thoughts. In D. Glynn, & K. Fischer (Eds.), Quantitative Cognitive Semantics: Corpus-driven approaches (pp. 355– 380). Berlin & New York: Mouton de Gruyter. Stepanov, J. S. (1997). Константы: Словарь русской культуры [Constants: A dictionary of Russian culture]. Moscow: Shkola Jezyki Russkoj Kul’tury. Talmy, L. (1985). Lexicalization patterns: Semantic structure in lexical forms. In T. Shopen (Ed.), Language typology and syntactic description (pp. 57–149). Cambridge: Cambridge University Press. Talmy, L. (1988). Force dynamics in language and cognition. Cognitive Science, 12, 49–100. DOI: 10.1207/s15516709cog1201_2 Taylor, J. (1988). Contrasting prepositional categories: English and Italian. In B. Rudzka-Â�Ostyn (Ed.), Topics in Cognitive Linguistics (pp. 299–326). Amsterdam & Philadelphia: John Benjamins. Taylor, J., & MacLaury, R. (1995). Language and the cognitive construal of the world. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110809305 Taylor, J. (1989a). Linguistic categorization: Prototypes in linguistic theory. Oxford: Clarendon Press. Taylor, J. (1989b). Possessive genitives in English. In D. Geeraerts (Ed.), Prospects and problems of prototype theory (Special edition of Linguistics 27) (pp. 663–686). Berlin & New York: Mouton de Gruyter. Taylor, J. (1996). On running and jogging. Cognitive Linguistics, 7, 21–34. DOI: 10.1515/cogl.1996.7.1.21 Tomasello, M. (2003). Constructing a language: A usage-based theory of language acquisition. London & Cambridge (Mss): Harvard University Press. Tsohatzidis, S. (Ed.). (1990). Meanings and prototypes: Studies on linguistic categorization. London: Routledge. Tuggy, D. (1993). Ambiguity, polysemy, and vagueness. Cognitive Linguistics, 4, 273–290. DOI: 10.1515/cogl.1993.4.3.273

37

38

Dylan Glynn

Tuggy, D. (1999). Linguistic evidence for polysemy in the mind: A response to William Croft and Dominiek Sandra. Cognitive Linguistics, 10, 343–368. Tummers, J., Heylen, K., & Geeraerts, D. (2005). Usage-based approaches in Cognitive Linguistics: A technical state of the art. Corpus Linguistics and Linguistic Theory, 1, 225–261. DOI: 10.1515/cllt.2005.1.2.225 Tyler, A., & Evans, V. (2003). Reconsidering prepositional polysemy networks: The case of over. In B. Nerlich, Z. Todd, V. Herman, & D. Clark (Eds.), Polysemy: Flexible patterns of meaning in mind and language (pp. 95–160). Berlin & New York: Mouton de Gruyter. Vandeloise, C. (1986). L’espace en français. Paris: Seuil. Vandeloise, C. (1990). Representation, prototypes, and centrality. In S. Tsohatzidis (Ed.), Meanings and prototypes: Studies on linguistic categorization (pp. 403–437). London: Routledge. Vandeloise, C. (1994). Methodology and analysis of the preposition in. Cognitive Linguistics, 5, 157–184. DOI: 10.1515/cogl.1994.5.2.157 Verschueren, J. (1981). Problems of lexical semantics. Lingua, 53, 317–351. DOI: 10.1016/ 0024-3841(81)90046-2 Victorri, B., & Fuchs, C. (1996). La polysémie: construction dynamique du sens. Paris: Hermès. Vorkachev, S. G. (2004). Счастье как лингвокультурный концепт [Happiness as a cultural-linguistic concept]. Moscow: Gnozis. Vorlat, E. (1985). Metaphors and their aptness for trade names in perfumes. In W. Paprotté, & R. Dirven (Eds.), Ubiquity of metaphor: Metaphor in language and thought (pp. 263–294). Amsterdam & Philadelphia: John Benjamins. Wierzbicka, A. (1985). Lexicography and conceptual analysis. Ann Arbor: Karoma. Wierzbicka, A. (1989). Prototypes in semantics and pragmatics: Explicating attitudinal meanings in terms of prototypes. In D. Geeraerts (Ed.), Prospects and problems of prototype theory (pp. 731–769). Berlin & New York: Mouton de Gruyter. Wierzbicka, A. (1990). Prototypes ‘save’: On the uses and abuses of the notion of ‘prototype’ in linguistics and related fields. In S. Tsohatzidis (Ed.), Meanings and prototypes: Studies on linguistic categorization (pp. 347–367). London: Routledge. Wierzbicka, A. (1996). Semantics: Primes and universals. Oxford: Oxford University Press. Wulff, S., Stefanowitsch, A., & Gries, St. Th. (2007). Brutal Brits and persuasive Americans: Variety-specific meaning construction in the into-causative. In G. Radden, K.-M. Köpcke, Th. Berg, & P. Siemund (Eds.), Aspects of meaning construction (pp. 265–281). Amsterdam & Philadelphia: John Benjamins. Wulff, S. (2006). Go-V vs. go-and-V in English: A case of constructional synonymy? In St. Th. Gries, & A. Stefanowitsch (Eds.), Corpora in Cognitive Linguistics: Corpus-based approaches to syntax and lexis (pp. 101–126). Berlin & New York: Mouton de Gruyter. Zelinsky-Wibbelt, C. (Ed.). (1993). The semantics of prepositions. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110872576 Zlatev, J. (2003). Polysemy or generality? Mu. In H. Cuyckens, R. Dirven, & J. Taylor (Eds.), Cognitive approaches to lexical semantics (pp. 447–494). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110219074.447

Competing ‘transfer’ constructions in Dutch The case of ont-verbs Martine Delorge, Koen Plevoets, and Timothy Colleman Ghent University

This paper zooms in on the semantic relations between the constructions of “possessional transfer” (i.e. constructions used to encode events of possessional transfer) in Dutch by zooming in on a specific morphological class of dispossession verbs, viz. verbs with the prefix ont- ‘away’, such as ontnemen ‘take away’, ontfutselen ‘fish out of ’, onttrekken ‘extract, withdraw’, ontheffen ‘relieve’, etc. A database with several thousand attested ont-examples from various corpora of present-day written Dutch will serve as the starting point for an investigation of their constructional possibilities and preferences: the ont-verbs will be shown to cluster into a number of subclasses in terms of alternation possibilities. In addition, a comparison of these present-day Dutch results with data from a diachronic corpus of 19th century Dutch will reveal a number of lexico-grammatical shifts: the use of the double object construction and (especially) of the aan-dative with ont-verbs is more heavily constrained now than it was in earlier stages of the language. Keywords: aan-dative, alternations, dispossession verbs, Dutch, ont-verbs

1. Introduction1 As shown in (1) to (5) below, the grammar of present-day Dutch includes several different argument structure constructions that can be used for the encoding of three-participant events of ‘possessional transfer’. The constructions in (1) and (2) have received by far the most linguistic attention, since they constitute the wellknown dative alternation: (1) exemplifies the double object construction in which 1. All three authors are associated with the BOF/GOA research project on ‘Meaning in-between structure and the lexicon’ in the Linguistics department at Ghent University. We would like to thank the editors as well as two anonymous referees for their helpful comments and suggestions. The usual disclaimers apply.

40 Martine Delorge, Koen Plevoets, and Timothy Colleman

the verb is combined with a subject and two bare NP objects, (2) exemplifies the socalled aan-dative, in which the theme is coded as a direct object but the recipient is marked with the preposition aan (cognate with on, but relevantly similar to to). Existing studies of these constructions and their semantic relations include Van Belle and Van Langendonck (1996), Janssen (1997), Geeraerts (1998), Colleman (2009a) and Colleman and De Clerck (2009). In addition to the aan-dative, there are a number of structurally similar constructions formed with other prepositions that can also be used to encode certain subtypes of ‘possessional transfer’ events. One of these is the construction in (3), in which the preposition van marks the source of the transfer in a ‘dispossession’ event. At first sight, the example in (4) – which also denotes an event of dispossession – exemplifies the very same construction. However, there is a crucial difference in the linking of argument roles between (3) and (4): in the latter clause, the direct object codes the possessional source, and van marks the theme. In terms of the distinction between the basic types of ditransitive alignment put forward in typological research by Haspelmath (2005) and Malchukov and colleagues (2010), the constructions in (2) and (3) represent indirective alignment (i.e. the theme is coded like the monotransitive patient; i.e. as a bare NP object, and the recipient/possessor is coded differently) but the construction in (4) represents secundative alignment (i.e. the recipient/possessor argument is coded like the monotransitive patient while the theme gets special marking). The met-construction in (5) is also secundative, but it denotes a fairly prototypical event of ‘caused possession’ rather than ‘dispossession’. (1) De man heeft de vrouw een boek gegeven. ‘The man has given the woman a book.’ (2) De man heeft een boek aan de vrouw gegeven. ‘The man has given a book to the woman.’ (3) De man heeft een boek gestolen van de vrouw. ‘The man has stolen a book from the woman.’ (4) De man heeft de vrouw beroofd van al haar boeken. ‘The man has robbed the woman of all her books.’ (5) De man heeft de vrouw begiftigd met een boek. ‘The man presented the woman with a book.’ In terms of constructional polysemy, the first two constructions cover a wider region in semantic space than the latter three. While the instances in (1) and (2) denote events of ‘caused possession’ with agent, theme and recipient participants, both the double object construction and the aan-dative can also be used to encode events of dispossession with agent, theme and (human) source participants, albeit with rather limited lexical possibilities (see Geeraerts 1998 and Colleman & De Clerck 2009 for examples and further discussion). (6) and (7) present examples with the dispossession verb ontnemen ‘take away’.

Competing ‘transfer’ constructions in Dutch

(6) De man heeft de vrouw een boek ontnomen. ‘The man has taken a book from the woman.’ (lit. has away-taken the woman a book) (7) De man heeft een boek aan de vrouw ontnomen. ‘The man has taken a book from the woman.’ The van-constructions in (3) and (4), by contrast, are limited to events of dispossession, whereas the met-construction in (5) is limited to a handful of infrequent and formal verbs of giving and cannot be used to encode the reverse orientation of transfer.2

2. Introducing the Dutch ont-verbs In this paper we will try to shed more light on the synchronic and diachronic semantic relations between the argument structure constructions introduced in section 1 by zooming in on a specific morphological class of dispossession verbs, viz. complex verbs with the prefix ont- ‘away’. While the ont-verbs are not particularly frequent in everyday language, they nevertheless constitute an interesting class in that many of them are found in several of the above constructions. Moreover, even though, semantically, ont-verbs of dispossession seem to form a rather homogeneous class at first sight, they nevertheless display very different constructional preferences. As such, a quantitative investigation of the degree of attraction between these ont-verbs and the argument structure constructions in question can shed more light on the issue of constructional competition between constructions with partly overlapping semantic ranges: how do the constructions illustrated in section 1 divide up the semantic domain of ‘dispossession’, which is itself a sub-domain of ‘possessional transfer’? Moreover, since we will not only look into the constructional preferences of the selected ont-verbs in present-day data, but also in data from a corpus of 19th century Dutch, it can also be investigated whether the semantic relation between the ‘dispossession’ constructions in question has changed in the course of the last century and a half. As such, the present study adds to the growing body of diachronic investigations of constructional semantics (see, e.g., Barðdal 2007 and Colleman & De Clerck 2011 for earlier studies). The following 15 ont-verbs were selected for the case study, listed in alphabetical order: ontdoen ‘strip’, ontfutselen ‘filch, fish out of ’, ontheffen ‘release’, ontlenen ‘take, derive, borrow’, ontladen ‘unload, be released’, ontlasten ‘relieve’, ontlokken ‘elicit

2. There is a single lexical exception in the case of the construction in (4): next to a set of verbs of dispossession, this pattern also accommodates the giving verb voorzien ‘provide’, as in Ze voorzagen hem van voedsel (‘They provided him with food’). Most probably, this is a calque of the French pattern pourvoir quelqu’un de quelque chose.

41

42

Martine Delorge, Koen Plevoets, and Timothy Colleman

(from)’, ontnemen ‘take away’, ontroven ‘steal away’, ontrukken ‘snatch (away)’, ontstelen ‘steal away’, onttrekken ‘withdraw, take away, derive’, ontvreemden ‘steal, thieve’, ontworstelen ‘tear, wrest from’ and ontwringen ‘wrench from’. These verbs were chosen on the basis of (i) their semantic status as (prototypical) dispossession verbs and (ii) frequency considerations. Semantically, ditransitive ont-verbs belong to two categories, good examples of which are ontnemen ‘take away from’ and ontzeggen ‘deny’, respectively. Whereas ontnemen denotes an event of dispossession, as in (6) and (7) above, ontzeggen instead denotes a blocked ‘caused possession’ event, as in (8). Put differently, ontnemen and ontzeggen are verbs of taking and not-giving, respectively. In terms of Colleman’s (2009b) multidimensional analysis of the semantic structure of the Dutch double object construction, the examples in (6) and (8) represent distinct subsenses of the construction, departing from the core ‘caused reception’ meaning along the semantic dimensions of ‘orientation of the transfer’ and ‘polarity of the transfer’, respectively. (8) Ze hebben vrouwen lange tijd het stemrecht ontzegd. ‘For a long time, they denied women the right to vote.’ In this case study, we focus on ont-verbs of the ‘dispossession’ type. The verbs under investigation select agent, theme and possessional source participants. Note that this is not to say that they occur exclusively in three-argument constructions: on the contrary, several of them occur far more often in two-argument constructions without an expressed source (see below). In terms of Goldbergian construction grammar, the selected verbs select a possessional source participant, but this participant need not be lexically profiled – though it is for the most prototypical members of the category, such as ontnemen (‘take away’) or ontfutselen (‘filch, fish out of ’) (for further details on the notion of lexical profiling, see Goldberg 1995:â•›43–48). The formation of dispossession verbs with the prefix ont- is not a productive word-formation pattern anymore, but dictionaries include quite a lot of examples, many of which are highly infrequent in present-day language (e.g. ontsjacheren ‘to barter sth. away from s.o.’ or ontschaken ‘to take s.o. away from s.o. by abduction’, to give but two examples). In order to avoid the selection of all too infrequent verbs for the case study, we only selected verbs which are included in the CELEX lexical database of Dutch. In addition, preference was given to verbs with a CELEX frequency of at least 3 occurrences per one million words of running text.3 However, this frequency criterion was applied liberally: an exception was made for a number of infrequent verbs that we wanted to include in the investigation for semantic reasons, such as ontroven and ontstelen, the verb bases of which are the semantically prototypical and highly frequent dispossession verbs roven ‘rob’ and stelen ‘steal’. 3. The CELEX frequency was checked using the WordGen tool developed by Wouter Duyck and colleagues (Duyck et al. 2004).


3. Methodology of the case study A database consisting of several thousands of attested ont-examples from various corpora of present-day written Dutch serves as the starting point for an investigation of the constructional flexibility of these verbs: the newspaper component of the present-day CONDIV-corpus, the 27 and 38 Million Words Corpora of the Dutch Institute for Lexicology (INL) and the Twente Nieuws Corpus (TNC). The selected corpora all represent relatively formal genres of written Dutch (newspapers, magazines, fictional and non-fictional prose, etc.). All forms of the 15 test verbs were automatically retrieved from the corpora, and the results were manually filtered and labelled according to syntactic construction. We used more than one corpus in order to be able to retrieve enough examples of the selected verbs to make well-grounded judgements about their constructional behaviour. After all, some of the verbs from the selection, as already mentioned, are quite infrequent. The different corpora are similar enough in their contents to justify combining them together. Section 4 presents and discusses the main findings from the synchronic investigation. In addition, we compared the present-day Dutch results with data from a corpus of 19th-century Dutch in order to investigate possible lexicogrammatical shifts. The diachronic part of the study is based on a corpus consisting of 50 volumes from the periodical De Gids (1850-1899). De Gids (‘The Guide’) was an influential literary and general cultural journal, and as such represents formal written Dutch. The results of the diachronic investigation are presented in Section 5.

4. The results of the present-day investigation 4.1

Overall distribution

As was noted in Section 2 above, the overall aim of the investigation is to investigate the constructional competition between the various Dutch argument structure constructions that can be used to encode ‘dispossession’ events, through an identification of the constructional preferences of a sample of ont-verbs in (a) present-day data and (b) 19th century data. Table 1 presents the results of the synchronic part of the corpus investigation. We distinguish six different three-argument constructions, four of which have already been introduced in Section 1: the double object construction (DOC), the aan-dative, and two constructions with van, which are labeled van-I (“indirective”) and van-S (“secundative”), respectively. The remaining constructions distinguished in the table are two fairly infrequent indirective patterns in which the (possessional) source argument is marked with the prepositions bij ‘at’ and uit ‘out (of)’, respectively. In addition, most of the investigated verbs occur in other constructions, too, i.e. in various two-argument constructions. In five cases, two-participant

43


constructions even account for the large majority of occurrences, viz. with ontvreemden ‘steal, thieve’, ontworstelen ‘wrest from’, onttrekken ‘withdraw, take away’, ontlasten ‘relieve’, and ontladen ‘unload’. In fact, the verb ontladen turned out to display just a single three-argument instance in the corpus data, from a total of 331 occurrences, viz. a single instance of the secundative van-construction. Because of this sparsity of data, this verb is excluded from further discussion. The rather frequent occurrence of verbs of dispossession in structures without a possessional source object is not very surprising. Newman (1996:â•›57) has already observed that “there is no giver necessarily present in the base of take […] There is only one person necessarily involved in the characterization of the basic meaning of TAKE”. As shown in Delorge (2010), simplex verbs of dispossession, such as stelen ‘steal’, roven ‘rob’, etc. typically display the same characteristic.4 In the remainder of this section, we remove the monotransitive and other two-argument constructions from consideration and focus on the distribution of the three-argument constructions. The distribution of the 14 remaining ont-verbs over the 6 three-argument constructions distinguished in Table 1 is statistically significant (χ² = 46741.05, df = 65, p < 2.2e-16). Table 4 in the Appendix lists the Pearson residuals. The Cramér’s V effect size is 0.7125667, which suggests a strong association. Note that effect sizes tend to increase with the number of rows and columns, however, and that there are quite a lot of cells with an expected frequency of less than 5, which reduces the value of the chi-square test. This is why we turn to an exploratory statistical tool in the next sub-section.

4.2

Four clusters of ont-verbs

As a first observation, it can be seen from the frequencies in Table 1 that there is a nearly complete lexical split between the secundative van-construction on the one hand and the other five three-argument constructions included in the table on the other. While many verbs occur in both the DOC and several of the indirective constructions in the corpus data, the secundative construction is associated with a number of verbs that occur in this three-argument construction (virtually) exclusively, viz. ontlasten ‘relieve’, ontdoen ‘strip’, and ontheffen ‘release’ (in addition, ontladen ‘unload’ was also attested in this construction only, see Section 4.1). Corpus examples are listed in (9) to (11). As can be seen from these examples, the theme of ontheffen and 4. It should be noted that constructions with a reflexive pronoun, such as the ontworstelen example in (i) below, were included in the rest category, too, as they do not contain three arguments. (i) Een deel van de Otavalo-Indianen probeert zich aan die armoede te ontworstelen. [TNC] ‘Part of the Otavalo Indians try to wrest themselves out of that poverty.’


Table 1.â•‡ Observed frequencies in the present-day data Verb ontdoen ontfutselen ontheffen ontlasten ontlenen ontlokken ontnemen ontroven ontrukken ontstelen onttrekken ontvreemden ontworstelen ontwringen Total

Three-argument constructions DOC

aan

0 458 0 0 0 587 3993 2 6 104 2 8 7 11 5178

0 146 5 0 6087 342 207 3 189 3 2248 3 26 11 9270

van-I van-S 0 19 0 0 2 23 15 0 0 0 7 33 0 0 99

2575 0 697 111 0 0 0 0 0 0 0 0 0 0 3383

uit

bij

Total

0 4 0 0 17 2 2 0 7 0 46 298 1 0 377

0 1 0 0 4 53 3 0 0 0 0 43 0 0 104

2575 628 702 111 6110 1007 4220 5 202 107 2303 385 34 22 18411

Other cxs

Total

1888 28 321 956 65 256 270 0 7 3 3562 653 859 1 8869

4463 656 1023 1067 6175 1263 4490 5 209 110 5865 1038 893 23 27280

ontlasten is typically an abstract entity, such as a task or duty. In the case of ontdoen, the lexical possibilities are wider. None of these verbs is a very prototypical representative of the class of ‘dispossession’ verbs semantically, unlike some of the other verbs which will be discussed below. Ontlasten and ontheffen imply a positive effect on the original possessor, who is relieved of something conceptualized as a burden – needless to say, in prototypical dispossession events, the transfer has a negative effect on the original possessor. As for ontdoen, this denotes an event of physical removal or separation rather than actual dispossession: typically, as in (11), the object NP does not refer to a human participant.5 (9) Inmiddels zijn creatieve oplossingen gevonden om hen te ontlasten van hun zware taak. [TNC] ‘In the meantime, creative solutions have been found to relieve them of their heavy task.’

5. The object NP can refer to a human participant, as in (ii) below. However, even such instances seem to denote an event of physical removal rather than of actual dispossession. (ii) Het Hof van Cassatie achtte het onmogelijk dat de man de 18-jarige van haar broek had kunnen ontdoen als zij zich daar tegen had verzet. [TNC] ‘The court of cassation judged that it would have been impossible for the man to strip the 18-year-old of her pants if she had offered resistance.’

45


(10) De bisschop onthief haar van haar functie. [TNC] ‘The bishop released her from office.’ (11) Met snijbranders ontdoen de slopers de molens van hun wieken. [TNC] With cutting torches, the demolishers are stripping the mills of their sails.’ Returning to the syntactic possibilities, ontheffen is the only one of these verbs that is also attested in another three-argument construction, viz. the indirective aan-construction, but only very sporadically so (a mere 5 out of 1023 corpus instances). Conversely, none of the remaining 11 ont-verbs are attested in the secundative van-construction a single time. This situation is reminiscent of the lexical split found with verbs of giving, where the secundative pattern is also limited to a handful of verbs that are not eligible for use in the DOC or the aan-dative, including begiftigen, as in (5) above (see Delorge & De Clerck 2007). Most of the remaining 11 verbs are used in the DOC as well as in several indirective constructions. However, a manual inspection of the frequencies in Table 1 (and of the Pearson residuals in Table A) suggests that they cluster into a number of classes with distinct constructional preferences. We used an exploratory statistical technique, viz. Correspondence Analysis (CA), to visualize such associations in the data: see the plot in Figure 1, in which the constructions are represented in capitals and the verbs in lower case (on CA, see Greenacre 2007, as well as the articles by Glynn in the present volume). The eigenvalues for the first two dimensions are 53.8% and 44.33%, respectively, indicating that the analysis presented in Figure 1 explains 98.13% of the variation (inertia). The table used to produce the correspondence analysis includes many small cells, and so the numerical summary of the analysis is included in the Appendix, Table 5. To assure a reasonable degree of accuracy in the analysis, the quality score (qlt) should be over 500. A figure of 500 indicates that 50% of the inertia for that data point lies off principle axes and therefore that points are less accurately displayed in the plot (see Glynn, this volume, and Greenacre 2007 for more details). The remainder of this sub-section interprets the visualization in Figure 1. The only outlier among the 11 verbs is ontvreemden ‘steal, thieve’, which is grouped with the uit-construction on the right-hand side of the plot. Indeed, ontvreemden is the only verb which is combined with a uit-phrase in the majority of its three-participant uses. Unlike the “secundative” verbs discussed above, ontvreemden, which is a hyponym of stelen ‘steal’, denotes a prototypical dispossession event involving a human possessor: if something is stolen, it is by definition stolen from someone. However, ontvreemden differs from ontnemen ‘take away from’ and other verbs to be discussed below in that this human possessor participant is not lexically profiled: in the majority of instances, the possessor role is not expressed. In fact, this is a characteristic ontvreemden shares with its hypernym (see Goldberg’s 1995:â•›45–46 semantic discussion of English steal and the frequency data for Dutch stelen in Delorge 2010). The uit-construction differs from the other three-argument constructions in Table 1 in that the PP refers to a locational source rather than to a possessional source. In (12),


Correspondence analysis graph ontstelen ontnemen DOC

Dimension1 (53.80%)

1.0

ontfutselen ontlokken

0.5

VAN.I BIJ

ontwringen ontroven

0.0

ontvreemden UIT

ontworstelen

–0.5 AAN ontrukken onttrekken ontlenen

0

1

2

3

4

5

Dimension2 (44.33%)

Figure 1.â•‡ Correspondence analysis of the present-day data (without the secundative construction)

for instance, the jewellery store is the place where the necklace was stolen, though this of course refers indirectly to the person from whom it was stolen, viz. the jeweller. While this is a pattern typically found with ontvreemden, the van-example in (13) shows that the verb does also occur in three-argument constructions with a genuine possessional source argument. (12) Fadiga ontvreemdde het collier, met een waarde van 300 euro, zondag uit een juwelierszaak. [TNC] ‘On Sunday, Fadiga stole the necklace worth 300 € from a jewelry store.’ (13) Muijtstege had de twee wel vaker financieel geholpen en ze hadden ook al eerder geld van hem ontvreemd. [TNC] ‘Muijtstege had given financial help to the two [petty criminals] before, and they had stolen money from him before as well.’ On the left-hand side, we find a kind of cline of verbs positioned in-between the DOC and the aan-dative. Starting at the top, the two verbs that are most closely associated with the DOC are ontnemen ‘take away from’ and ontstelen ‘steal away from’. Indeed, these are found with double object syntax in the large majority of their instances across the four corpora: the DOC accounts for 88.9% of the ontnemen instances in the database and for 94.5% of the ontstelen instances. Good examples of this are shown in (14) and (15).

47


(14) Verder mag de leiding verslaafden hun paspoort en geld niet meer ontnemen. [TNC] ‘Furthermore, the leaders are no longer allowed to dispossess addicts of their passports and money.’ (lit. to away-take addicts their passport and money) (15) 10000 frank kan de miljoenen niet vervangen die de staat mij ontstolen heeft. [CONDIV] ‘10,000 francs cannot replace the millions that the state has stolen from me.’ (lit. that the state has me away-stolen) In both cases, the observed preference for the DOC can be straightforwardly linked to the verbs’ lexical semantics. The verb bases nemen ‘take’ and stelen ‘steal’ are among the most prototypical simplex verbs of taking away. Similarly, their prefixed variants ontnemen and ontstelen denote prototypical dispossession events in which a human participant is deprived of an item in his/her possession, typically by another human participant. This item need not be a concrete object – it is in (14), but not, or less so, in (15) – but, in any event, there was a fairly prototypical possession relation between the indirect and direct object referents before the event, which is broken by instigation of the subject referent. Note that, unlike many of the other verbs in Table 1, ontnemen and ontstelen occur in two-argument constructions in only a small fraction of their instances (6% and 2.7%, respectively): these are typical three-participant verbs, with a lexically profiled possessional source participant. Van Belle and Van Langendonck (1996:â•›245) have also observed the preference of ontnemen for the double object construction, albeit on an introspective basis. They ascribe this to the factor [+/– involvement], which, in their account, is one of the two major semantic determinants of the dative alternation in Dutch, next to [+/– (material) transfer]. Ontnemen lexicalizes a transfer of possession which has a fundamental effect on the original possessor, and the same applies to ontstelen: the ‘transfer’ events denoted in (14) and (15) above (negatively) affect the indirect object referent, i.e. the loss of the direct object referent has clear consequences for the indirect object and his/her further actions. Since the DOC is hypothesized to highlight the relationship between the direct and the indirect object referents, the lexical semantics of verbs such as ontnemen and ontstelen tallies well with the constructional semantics of the double object construction (cf. De Schutter 1974:â•›205; Verhagen 1986 and Colleman 2009b; inter alia). According to Van Belle and Van Langendonck (1996:â•›245), the only reason why verbs such as ontnemen can appear in the aan-dative at all, despite their strong lexical focus on the affectedness of the possessional source participant, is that there is also a ‘caused motion’ event involved: the theme moves from the domain of possession of the source to that of the agent. An often advanced semantic explanation for the dative alternation posits that the aan-structure emphasizes the spatial aspects of the denoted transfer scene, i.e. the movement of the theme along a path from the agent participant towards the recipient participant (or, in this case, from the source participant


Table 2.â•‡ DOC and aan-frequencies in instances with concrete vs. abstract themes theme = concrete entity theme = abstract entity Total

DOC

aan

Total

â•⁄411 2786 3197

â•⁄19 110 129

â•⁄430 2896 3326

toward the agent participant) (see, e.g. Goldberg 1992 and Langacker 1991 for similar semantic hypotheses about the English to-dative, and Colleman & De Clerck 2009 for further discussion). This would seem to suggest that material transfers generally prefer the aan-construction, or, at least, that the proportion of aan-instances to DOC instances is larger when the theme is a concrete object than when it is some kind of abstract commodity. The ontnemen data, however, show that no significant difference can be attested between cases with a concrete theme on the one hand, and cases with an abstract theme on the other hand; see the distribution in Table 2 (χ² = 0.386, df = 1, p = 0.5342).6 This warns us against an all too literal interpretation of the ‘caused motion’ hypothesis: even when the theme is a concrete entity which undergoes a spatial transfer as it changes ownership, the DOC is still the preferred construction in the large majority of cases. The general semantic hypothesis put forward in Colleman (2009b) is that the aan-dative highlights the changing agent-theme relationship, whereas the DOC highlights all three participants and their interrelations, including the recipient participant. We leave it to future research to test the validity of this hypothesis in a more systematic way for verbs with a possessional source rather than a recipient, but, in any event, ontnemen and ontstelen fit the picture, with their strong lexical focus on the affectedness of the possessor participant. As we move from the top to the bottom of the left-hand side in Figure 1, the lexical preferences shift towards the aan-dative. Ontfutselen ‘filch, fish out of ’ and ontlokken ‘elicit from’ are the only two remaining verbs for which the DOC is the most frequently attested construction in the database (accounting for 69.8% and 46.5% of all occurrences, respectively). There is a twofold reason why they are positioned lower in the plot than ontnemen and ontstelen: (i) because their proportion of aan-examples is relatively higher (22% and 27%, respectively, as opposed to less than 5% for both ontnemen and ontstelen) and (ii) because they are the only verbs, next to ontvreemden ‘steal’, which are found in the indirective van-construction (in the case of ontfutselen) or in the bij-construction (in the case of ontlokken) somewhat more than sporadically, though it should be stressed that such uses still account for only a small fraction of

6. The frequencies in Table 2 do not add up to the overall frequencies of the DOC and aan-dative with ontnemen listed in Table 1, as a good number of cases were not straightforwardly classifiable into either concrete or abstract categories but represented various in-between uses.

49

50


their total number of occurrences. We will briefly return to the constructions with van and bij at the end of this sub-section. Still further down, in the lower-left corner, are those verbs which occur with aan in the majority of their (three-argument) uses. The verbs associated most closely with the aan-dative are ontlenen ‘take, derive, borrow’, onttrekken ‘withdraw from, derive’, ontrukken ‘snatch away, pull away’ and ontworstelen ‘wrest from’: this construction accounts for 99.6%, 97.6%, 93.6% and 76.5% of the overall number of three-argument instances, respectively. The first thing to observe is that these verbs cover a much wider region in semantic space than just ‘possessional transfer’: typical examples are shown in (16) to (19) below, none of which involves a human possessor participant that is ‘dispossessed’ of something in the strict sense of the word: rather, these instances denote various kinds of (metaphorical) ‘separation’ or ‘withdrawal’ events. (16) Of Jeroen Bosch inspiratie ontleende aan al dat water in de stad, is niet bekend. [TNC] ‘Whether Hieronymus Bosch derived inspiration from all that water in the town is unknown.’ (17) Tegenwoordig is [dat] de meest gebruikte methode om geur aan natuurlijke producten te onttrekken. [TNC] ‘Presently, that is the most usual method for extracting the odour of natural products.’ (18) Bourguiba mag zijn land hebben ontrukt aan de Franse overheersing, hij was niettemin een francofiel. [TNC] ‘Bourguiba may have wrenched his country from French rule, he was a francophile nonetheless.’ (19) Dat spoort met een tendens om de VS te ontworstelen aan internationale afspraken. [CONDIV] ‘That confirms a tendency to wrestle the USA out of international agreements.’ Many occurrences of onttrekken and ontrukken, especially, are more or less fixed expressions, such as iets aan de vergetelheid onttrekken/ontrukken ‘to save something from oblivion’ or iets aan het zicht onttrekken ‘to hide something from sight’. This is the case in more than 50% of the ontrukken examples and 26.6% of the onttrekken examples. Again, such examples are not prototypical instances of dispossession events. Still, the aan-construction can be used to express events with a human deprivee, too: see (20) and (21) for examples of ontrukken and onttrekken. (20) In het holst van de nacht ontrukten zwaarbewapende Amerikaanse immigratieagenten een verschrikt Cubaans jongetje van zes aan luid lamenterende verwanten. [TNC] ‘In the middle of the night, heavily armed American immigration agents snatched a startled Cuban boy of six from loudly lamenting family members.’


(21) De partij vraagt voorts aan procureur-generaal Van Oudenhove om alle Vlaams Blok-dossiers aan Dejemeppe te onttrekken. [CONDIV] ‘Furthermore, the party asks attorney general Van Oudenhove to withdraw all cases related to the Vlaams Blok from Dejemeppe.’ Only in such cases is the DOC a (marked) option as well, hence the low frequencies in the DOC column in Table 1. (22) is one of the few DOC clauses attested with ontworstelen, for instance: in this case, the verb is used in its literal, compositional sense ‘wrest from’, and the relation between the object referents is one of fairly prototypical possession. (22) Ik dacht aan de talloze keren dat ik in de supermarkt had moeten wachten omdat moeders vooraan in de rij hun gillende broedsel vergeefs een reep chocola probeerden te ontworstelen. [TNC] ‘I thought of the many occasions when I had found myself waiting in the supermarket because mothers at the front of the row fruitlessly try tried to wrest away a chocolate bar from their yelling brood.’ For a final observation, the constructions with bij and van can be seen to occupy an isolated position in the middle of Figure 1. Indeed, none of the verbs under investigation is particularly attracted to these constructions. The verbs with the relatively highest frequencies of indirective van-instances are ontvreemden ‘steal’ and ontfutselen ‘filch, fish out of ’, but even in these cases, the van-construction only accounts for about 3% of the occurrences. See (13) above for an example with ontvreemden and (23) below for one with ontfutselen. (23) Ze moeten 80.000 frank boete betalen omdat ze geld en andere goederen hebben ontfutseld van enkele gasten. [CONDIV] ‘They have to pay an 80,000 francs fine because they have filched money and other goods from a number of guests.’ In this regard, the ont-verbs differ from the pattern attested with simplex verbs of dispossession, for with stelen ‘steal’, nemen ‘take’, etc., the human possessor participant is marked with van in present-day Dutch (as in (3) above, see Delorge 2010 for details). It is sometimes suggested that it would be natural for ont-verbs of dispossession to substitute the “default” source preposition van for aan (e.g. Schermer-Vermeer 1991:â•›216–217). However, our data show that, so far, the van-construction has not really caught on with these verbs, at least not in formal registers of written Dutch. The “top verbs” of the bij-construction are ontvreemden ‘steal’, again, and ontlokken ‘elicit (from)’, but, similarly to the van-construction, this construction with bij accounts for only a small fraction of occurrences in both cases (in-between 4 and 5%). An ontlokken example is shown in (24). Again, such examples do not denote prototypical dispossession.

51

52


(24) Alleen het bootje met daarop Sint en zijn pieten die sinterklaasliedjes zongen, ontlokte applaus bij het publiek. [TNC] ‘Only the boat on which the saint and his servants were singing St-Nicholas songs drew applause from the public.’ To summarize, we can distinguish four clusters of ont-verbs in the present-day data: a set of verbs that are (virtually) exclusively used in the secundative van-construction: ontlasten ‘relieve’, ontdoen ‘strip’, ontheffen ‘release’ (and ontladen ‘unload’); ii. the verb ontvreemden ‘steal’, which marks the source participant with uit, bij, or van, if it is expressed at all; iii. a set of verbs with a strong lexical preference for the double object construction, most notably ontnemen ‘take away from’ and ontstelen ‘steal away from’; iv. a set of verbs with a strong lexical preference for the aan-construction, most notably ontlenen ‘take, derive, borrow’, onttrekken ‘withdraw, derive’, ontrukken ‘snatch away’ and ontworstelen ‘wrest, wrench away’. i.

There is a cline from (iii) to (iv), with verbs such as ontfutselen ‘filch, fish out of ’, ontroven ‘rob from’ and ontwringen ‘wrench from’ occupying intermediate positions. The next section explores the constructional preferences of ont-verbs in an older sub-stage of the Dutch language.

5. A diachronic perspective An important trend in construction grammar is the growth of interest in issues of diachronic and synchronic language variation in the syntax and semantics of schematic constructions – see Colleman and De Clerck (2011) for references and discussion. In this section, we will explore a number of differences and similarities in the constructional behaviour of the selected ont-verbs between present-day Dutch and 19th-century Dutch, as represented by a 50-year sample of the corpus De Gids (1850–1899). First, Table 3 presents the distribution of the 14 verbs under investigation over the six three-argument constructions in the 19th-century data – ontladen ‘unload’ was left out again. This distribution is statistically significant (χ² = 5659.828, df = 65, p-value < 2.2e-16). Table 6 in the Appendix lists the Pearson residuals, and the Cramér’s V effect size is 0.5195202. The data show that the virtually complete lexical split between the secundative construction on the one hand and the other three-argument constructions on the other was already present in the 19th century: the three verbs attested in the secundative van-construction are not attested in any of the other three-argument constructions, with the exception of ontheffen ‘release’, which has a small number of aan and DOC examples. Conversely, the other 11 verbs do not enter into the secundative construction.


Table 3.â•‡ Observed frequencies in the 19th-century data Verb ontdoen ontfutselen ontheffen ontlasten ontlenen ontlokken ontnemen ontroven ontrukken ontstelen onttrekken ontvreemden ontworstelen ontwringen Total

Three-argument constructions DOC

aan

0 30 2 0 9 179 1023 29 113 78 7 4 4 43 1521

0 11 4 0 873 144 697 11 287 13 39 5 6 41 2131

van-I van-S 0 0 0 0 34 0 0 0 0 0 0 4 0 2 40

89 0 304 30 0 0 0 0 0 0 0 0 0 0 423

uit

bij

Total

0 0 0 0 74 2 0 0 0 0 0 1 0 0 77

0 0 0 0 0 2 0 0 0 0 0 0 0 0 2

89 41 310 30 990 327 1720 40 400 91 46 14 10 86 4194

Other cxs

Total

87 3 23 67 12 17 57 0 32 2 52 17 87 16 472

176 44 333 97 1002 344 1777 40 432 93 98 31 97 102 4666

Figure 2, below, shows the plot from a Correspondence Analysis of the observed frequencies of these remaining 11 verbs in the DOC and the four indirective constructions. The two-dimensional analysis accounts for 96.65% of the inertia (Dim. 1: 72.95%, Dim. 2: 23.7%). Again, the numerical output is supplied in the Appendix, Table 7. We will not explore this distribution to the same degree of detail as in the previous section, but we will focus on the relation between the DOC and the aan-dative, the two constructions involved in the dative alternation. A difference with the visualization in Figure 1 is that, in Figure 2, the DOC and the aan-construction occupy less extreme positions in the cluster of points on the left-hand side of the plot: the distance between the two constructions is smaller than in Figure 1. This suggests that in the 19th-century data, there is more overlap in their distributions over the 11 verbs than in the present-day data. In order to further test this, we conducted two-by-two comparisons of the observed DOC and aan-dative frequencies of the 11 non-secundative verbs in 19th-century vs. present-day language. In six cases, this revealed a significant shift. Three verbs display a significantly stronger preference for the DOC in the present-day data compared to the 19th-century data: ontnemen (χ² = 1194.934, df = 1, p < 2.2e-16; OR = 0.07611464), ontstelen (χ² = 8.7284, df = 1, p = 0.003133; OR = 0.1744933) and ontlokken (χ² = 6.0898, df = 1, p = 0.0136; OR = 0.7244433). The odds ratios (ORs) are included as a measure of the effect size: the effect is stronger for ontnemen and ontstelen than for ontlokken (the more the OR differs from 1, the stronger the effect). Three other verbs display the reverse tendency, i.e. they have a significantly stronger

53


Correspondence analysis graph UIT

1.5

1.0 Dimension1 (87.61%)

54

0.5

0.0

–0.5

–1.0

VAN

ontlenen

ontvreemden

onttrekken AAN onturkken ontheffen ontworstelenontwringen ontlokken ontnemen BIJ ontfutselen DOC ontroven ontstelen

0

0.5 1 1.5 Dimension2 (8.948%)

2

2.5

Figure 2.â•‡ Correspondence analysis of the 19th C data (without secundative construction)

preference for the aan-dative in the present-day data compared to the older data: ontlenen (Fisher Exact p = 8.038e-09; OR = Infinite), onttrekken (Fisher Exact p = 2.828e11; OR = 198.7147) and ontrukken (χ² = 51.9195, df = 1, p-value = 5.782e-13; OR = 12.36263).7 In this case, the odds ratios are larger than 1 because the effect is reversed: the odds for the aan-dative are larger in the present-day data than in the old data. The remaining five verbs do not display a significant diachronic shift: the distribution attested with ontwringen, for instance, is remarkably constant, with a proportion of DOC to aan-dative instances of (about) 1:1 in both periods. The diachronic shifts observed in individual verbs can be seen as indicators of a tendency towards polarization or constructional specialization: in those verbs which already displayed a strong preference for one of the two alternating constructions in the 19th-century data, this lexical preference has become even stronger in the present-day data. That is, ontnemen, ontstelen and (to a somewhat lesser extent) ontlokken have become even more closely associated with the DOC, whereas ontlenen, onttrekken, and ontrukken have become even more closely associated with the aan-construction. All in all, there seem to be few verbs left in this morphological class which can be said to alternate more or less freely. An interesting follow-up question for further 7. In the case of ontlenen and onttrekken, we used the Fisher Exact test rather than the Pearson chi-square test because at least one of the cells in the 2-by-2 table had an expected frequency of less than 5.


diachronic research into the Dutch dative alternation is whether this polarization tendency observed for ont-verbs is also found in other ditransitive verb classes.

6. Conclusion Whereas the large majority of existing studies on the grammatical encoding of ‘possessional transfer’ events in Dutch – or English, for that matter – deal with verbs of giving, either primarily or exclusively, the present study has focused on verbs lexicalizing the reverse direction of transfer, i.e. verbs of dispossession. More specifically, we have presented a corpus-based case study of a specific morphological class of dispossession verbs, viz. prefixed verbs with ont- ‘away’. Starting with the frequency data included in a database with over 18,000 three-argument examples from four corpora of present-day written Dutch, we have distinguished four clusters of ont-verbs according to their constructional preferences. These are (i) verbs which are (virtually) exclusively found in the secundative van-construction, (ii) verbs with a preference for indirective constructions with uit, bij, or van, or for two-argument constructions without a source participant, (iii) verbs with a strong lexical preference for the double object construction, and (iv) verbs with a strong lexical preference for the aan-dative. These clusters provide valuable information about the semantic relations between the argument structure constructions at stake. As for the distinction between clusters (iii) and (iv), for instance, we have observed that, compared to the verbs strongly attracted to the aan-dative in cluster (iv), the verbs with a strong attraction to the DOC in cluster (iii) denote more prototypical events of dispossession in which a human original possessor is (strongly) negatively affected by the transfer. This suggests that the DOC is the preferred construction for the encoding of scenes in which the source of the transfer is a prototypical human deprivee – a finding which is in line with a general hypothesis known from the literature on the dative alternation with verbs denoting a possessional transfer in the canonical direction (i.e., verbs of giving and the like), viz. that compared to the aan-dative, the DOC highlights the affectedness of the indirect object referent. We have not found corroborating evidence, however, for another wellknown hypothesis about verbs of giving, viz. that the use of the aan-dative highlights the spatial aspects of the scene: in clauses with ont-verbs, it does not seem to matter for the choice between the DOC and the aan-dative whether or not the theme actually moves along a spatiotemporal path as it changes ownership. Furthermore, we have also observed that there is a cline from verbs with a strong DOC preference to verbs with a strong aan-preference, with a number of verbs occupying intermediate positions. This is different for the secundative van-construction, in that there is a nearly complete lexical split between this construction on the one hand and the other five constructions included in the table on the other: the secundative construction is frequently found with a number of ont-verbs which hardly occur in the DOC or in any

55

56


of the investigated indirective constructions. Semantically, these verbs depart from prototypical ‘dispossession’ in that they imply a positive rather than negative effect on the original possessor, who is relieved of something conceptualized as a burden, or denote an event of physical removal or separation. The secundative construction, it turns out, is hardly an option for the encoding of more protypical ‘dispossession’ events with a human possessional source maleficially affected by the subject’s action. The indirective constructions with van, bij, and uit, finally, account for a small fraction of the ont-examples only; ontvreemden ‘steal’ is the only verb in the data with a preference for one of these constructions, namely for the locative construction with uit – it remains to be investigated how exactly simplex verbs of dispossession such as stelen ‘steal’ or roven ‘rob’ behave with respect to these various indirective constructions. The most interesting finding from the diachronic part of the investigation is that the relation between the DOC and the aan-dative seems to characterized by an interesting tendency of polarization or semantic specialization: the lexical preferences displayed in the 19th-century data tend to become even stronger in the present-day language.

References Barðdal, J. (2007). The semantic and lexical range of the ditransitive construction in the history of (North) Germanic. Functions of Language, 14, 9–30. DOI: 10.1075/fol.14.1.03bar Colleman, T. (2006). De Nederlandse datiefalternantie: Een constructioneel en corpusgebaseerd onderzoek [The Dutch dative alternation. A constructionist and corpus-based investigation]. Unpublished Ph. D. dissertation, Ghent University. Colleman, T. (2009a). The semantic range of the Dutch double object construction: A collostructional perspective. Constructions and Frames, 1, 190–220. DOI: 10.1075/cf.1.2.02col Colleman, T. (2009b). Verb disposition in argument structure alternations: A corpus study of the Dutch dative alternation. Language Sciences, 31, 593–611. DOI: 10.1016/j.langsci. 2008.01.001 Colleman, T., & De Clerck, B. (2009). Caused motion? The semantics of the English to-dative and the Dutch aan-dative. Cognitive Linguistics, 20, 5–42. DOI: 10.1515/COGL.2009.002 Colleman, T., & De Clerck, B. (2011). Constructional semantics on the move: On semantic specialization in the English double object construction. Cognitive Linguistics, 22, 183–210. DOI: 10.1515/cogl.2011.008 Delorge, M., & De Clerck, B. (2007). A contrastive and corpus-based study of English and Dutch provide-verbs. Phrasis, 48, 121–142. Delorge, M. (2010). De relatie tussen betekenis en structuur bij privatieve en receptieve werkwoorden in het Nederlands [The relation between meaning and structure in verbs of dispossession and reception in Dutch]. Unpublished Ph. D dissertation, Ghent University. De Schutter, G. (1974). De Nederlandse zin: Poging tot beschrijving van zijn structuur [The Dutch clause: An attempt at describing its structure]. Brugge: De Tempel.


Duyck, W., Desmet, T., Verbeke, L., & Brysbaert, M. (2004). WordGen: A tool for word selection and nonword generation in Dutch, English, German, and French. Behavior Research Methods, Instruments, & Computers, 36, 488–499. DOI: 10.3758/BF03195595 Geeraerts, D. (1998). The semantic structure of the indirect object in Dutch. In W. Van Langendonck & W. Van Belle (Eds), The Dative. Volume 2. Theoretical and contrastive studies (pp. 185–210). Amsterdam & Philadelphia: John Benjamins. Goldberg, A. E. (1992). The inherent semantics of argument structure: The case of the English ditransitive. Cognitive Linguistics, 3, 37–74. DOI: 10.1515/cogl.1992.3.1.37 Goldberg, A. E. (1995). Constructions: A construction grammar approach to argument structure. Chicago: University of Chicago Press. Greenacre, M. (2007). Correspondence analysis in practice (2nd ed.). Boca Raton: Chapman & Hall/CRC. DOI: 10.1201/9781420011234 Haspelmath, M. (2005). Argument marking in ditransitive alignment types. Linguistic Discovery, 3, 1–21. DOI: 10.1349/PS1.1537-0852.A.280 Janssen, T. (1997). Giving in Dutch: An intra-lexematical and inter-lexematical description. In J. Newman (Ed.), The Linguistics of Giving (pp. 267–306). Amsterdam & Philadelphia: John Benjamins. Langacker, R. W. (1991). Concept, image, and symbol: The cognitive basis of grammar. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110857733 Malchukov, A., Haspelmath, M., & Comrie, B. (2010). Studies in ditransitive constructions: A comparative handbook. Berlin & New York: Mouton de Gruyter. Newman, J. (1996). Give. A Cognitive Linguistic study. Berlin: Mouton de Gruyter. Schermer-Vermeer, I. (1991). Substantiële versus formele taalbeschrijving: Het indirect object in het Nederlands [Substantial versus formal language analysis: The indirect object in Dutch]. Amsterdam: University of Amsterdam, Dutch Department. Van Belle, W., & Van Langendonck, W. (1996). The indirect object in Dutch: In W. Van Belle & W. Van Langendonck (Eds.), The dative. Volume I: Descriptive studies (pp. 217–250). Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/cagral.2 Verhagen, A. (1986). Linguistic theory and the function of word order in Dutch. Dordrecht: Foris.

57

58


Appendix Table 4.â•‡ Pearson’s residuals of the χ² of Table 1 ontdoen ontfutselen ontheffen ontlasten ontlenen ontlokken ontnemen ontroven ontrukken ontstelen onttrekken ontvreemden ontworstelen ontwringen

DOC

aan

van-I

van-S

uit

bij

–26.9111 â•⁄21.17231 –14.0511 â•⁄–5.58732 –41.4537 â•⁄18.05141 â•⁄81.45388 â•⁄â•⁄0.50072 â•⁄–6.7413 â•⁄13.47256 –25.3715 â•⁄–9.63693 â•⁄–0.82861 â•⁄â•⁄1.93476

–36.0072 â•⁄–9.57147 –18.5346 â•⁄–7.47589 â•⁄54.27883 â•⁄–7.32894 –41.6047 â•⁄â•⁄0.304086 â•⁄â•⁄8.65564 â•⁄–6.93123 â•⁄31.96341 –13.7075 â•⁄â•⁄2.146425 â•⁄–0.02316

–3.72107 â•⁄8.501756 –1.94289 –0.77257 –5.38299 â•⁄7.557048 –1.61472 –0.16397 –1.04221 –0.75853 –1.52988 21.49648 –0.42758 –0.34395

â•⁄96.62736 –10.7422 â•⁄50.01195 â•⁄20.06194 –33.5068 –13.6028 –27.8464 â•⁄–0.95851 â•⁄–6.09239 â•⁄–4.43409 –20.5712 â•⁄–8.4109 â•⁄–2.49949 â•⁄–2.01059

â•⁄–7.2614 â•⁄–2.47057 â•⁄–3.79141 â•⁄–1.50763 â•⁄–9.66559 â•⁄–4.10051 â•⁄–9.08068 â•⁄–0.31998 â•⁄â•⁄1.40804 â•⁄–1.48021 â•⁄–0.16867 103.3261 â•⁄â•⁄0.364079 â•⁄–0.67119

–3.81388 –1.35253 –1.99135 –0.79184 –5.19401 19.83698 –4.26796 –0.16806 –1.0682 –0.77745 –3.60682 27.68344 –0.43825 –0.35252

Table 5.â•‡ Numerical output of the correspondence analysis of the present-day data > Principal inertias (eigenvalues): dim 1 2 3 4

value % cum% 0.828914 53.8 53.8 0.683026 44.3 98.1 0.025129 1.6 99.8 0.003653 0.2 100.0 -------- ----Total: 1.540722 100.0 Rows: 1 2 3 4 5 6 7 8 9 10 11

| | | | | | | | | | |

scree plot ************************* ********************* *

name mass qlt inr k=1 ontf | 42 896 21 | -829 ontln | 407 1000 161 | 760 ontlk | 67 492 28 | -563 ontn | 281 995 292 | -1252 ontrv | 0 917 0 | -87 ontrk | 13 982 4 | 679 onts | 7 990 8 | -1301 ontt | 153 996 56 | 745 ontv | 26 1000 429 | 38 ontwrs| 2 901 0 | 308 ontwrn| 1 957 0 | -299

cor 891 946 486 978 158 981 972 987 0 898 686

ctr 35 283 26 531 0 7 15 103 0 0 0

k=2 cor ctr | 59 5 0 | | 181 53 19 | | -60 5 0 | | 163 17 11 | | 191 759 0 | | -10 0 0 | | 177 18 0 | | 71 9 1 | | -5078 1000 968 | | 17 3 0 | | 188 271 0 |


Table 5.â•‡ (continued) Columns: name mass qlt 1 | DOC | 345 1000 2 | AAN | 617 1000 3 | VANI| 7 824 4 | UIT | 25 995 5 | BIJ | 7 698

inr 348 204 22 383 43

k=1 | -1239 | 694 | -438 | 165 | -314

cor 986 946 37 1 10

ctr 638 358 2 1 1

k=2 cor | 146 14 | 166 54 | -2011 787 | -4835 994 | -2562 687

ctr 11 25 39 859 67

| | | | |

Table 6.â•‡ Pearson residuals of the χ² on Table 3 ontdoen ontfutselen ontheffen ontlasten ontlenen ontlokken ontnemen ontroven ontrukken ontstelen onttrekken ontvreemden ontworstelen ontwringen

DOC

aan

van-I

van-S

uit

bij

â•⁄–5.68127 â•⁄â•⁄3.923941 –10.4144 â•⁄–3.29846 –18.4732 â•⁄â•⁄5.547325 â•⁄15.98458 â•⁄â•⁄3.805351 â•⁄–2.66221 â•⁄â•⁄7.832866 â•⁄–2.37058 â•⁄–0.47808 â•⁄â•⁄0.196071 â•⁄â•⁄2.114915

â•⁄–6.72469 â•⁄–2.15422 –12.2317 â•⁄–3.90425 â•⁄16.49591 â•⁄–1.71846 â•⁄–5.9854 â•⁄–2.06827 â•⁄â•⁄5.875097 â•⁄–4.88802 â•⁄â•⁄3.232374 â•⁄–0.79243 â•⁄â•⁄0.407667 â•⁄–0.40802

–0.92132 –0.62533 –1.71948 –0.5349 â•⁄7.992057 –1.766 –4.05023 –0.61765 –1.9532 –0.93162 –0.66236 10.58121 –0.30883 â•⁄1.302676

â•⁄26.70959 â•⁄–2.03352 â•⁄48.77557 â•⁄15.50718 â•⁄–9.99249 â•⁄–5.74288 –13.171 â•⁄–2.00857 â•⁄–6.35164 â•⁄–3.02954 â•⁄–2.15395 â•⁄–1.18828 â•⁄–1.00428 â•⁄–2.94514

–1.27828 –0.86761 –2.38568 –0.74215 13.094 –1.63397 –5.61947 –0.85696 –2.70995 –1.29256 –0.91899 â•⁄1.46546 –0.42848 –1.25655

–0.20601 –0.13983 –0.38449 –0.11961 –0.6871 â•⁄4.66983 –0.90566 –0.13811 –0.43675 –0.20832 –0.14811 –0.08171 –0.06906 –0.20251

Table 7.â•‡ Numerical output of the correspondence analysis of the 19th-century data Principal inertias (eigenvalues): dim 1 2 3 4

value 0.319872 0.032668 0.007285 0.005248 -------Total: 0.365073

% 87.6 8.9 2.0 1.4 ----100.0

cum% 87.6 96.6 98.6 100.0

scree plot ************************* **

59


Table 7.â•‡ (continued) Rows: 1 2 3 4 5 6 7 8 9 10 11 12

| | | | | | | | | | | |

name ontf | onth | ontln | ontlk | ontn | ontrv | ontrk | onts | ontt | ontv | ontwrs| ontwrn|

mass 11 2 263 87 456 11 106 24 12 4 3 23

Columns: name mass qlt 1 | DOC | 403 999 2 | AAN | 565 995 3 | VAN | 11 979 4 | UIT | 20 895 5 | BIJ | 1 31

qlt 981 693 996 647 993 982 749 970 830 941 691 789 inr 470 239 132 144 15

inr 14 0 547 38 209 13 32 57 11 75 0 5

| | | | |

k=1 -658 69 870 -320 -408 -646 162 -887 400 828 -52 -192

| | | | | | | | | | | |

k=1 -649 383 1437 1483 -566

cor 991 951 454 856 30

cor 952 77 995 646 993 954 241 917 485 93 82 503 ctr 531 259 69 140 1

ctr k=2 15 | -116 0 | 195 622 | -26 28 | 15 237 | -9 14 | -110 9 | 235 59 | -214 6 | 337 8 | -2501 0 | 143 3 | -144 k=2 | -59 | 82 | -1547 | -317 | 82

cor 8 44 525 39 1

cor 29 616 1 1 0 28 508 53 344 848 609 285 ctr 43 117 777 63 0

ctr 4 2 6 1 1 4 179 34 42 711 2 15

| | | | |

| | | | | | | | | | | |

Rethinking constructional polysemy The case of the English conative construction Florent Perek

Freiburg Institute for Advanced Studies and Université Lille 3

This chapter examines the conative construction, e.g., I kicked at the ball, using collexeme analysis. Previous studies report that strong collexemes of a construction provide an indication of its central meaning, from which polysemic extensions are derived. However, the conative construction does not seem to attract a particular kind of verb that could be used to characterize its central meaning. To address this problem, a variant of collexeme analysis is suggested that consists in splitting the verbal distribution into semantic classes and consider “verb-class-specific” constructions independently. For the three classes tested, the most significant collexemes are found to be verbs whose inherent meaning contains the semantic contribution of the construction in that class. Hence, the most attracted collexemes do provide an indication of the constructional meaning, albeit specific to each verb class. Keywords: collexeme analysis, semantic classes, verb-class-specific constructions

1. Introduction1 In constructional approaches to grammar, argument structures are taken to be symbolic pairings of a syntactic structure with a schematic meaning independent of the verbs instantiating them (cf. Goldberg 1995, 2006). For example, the ditransitive construction (e.g., John built the children a new merry-go-round) is a pairing of the double-object syntactic pattern with a core meaning of ‘caused possession’. An increasingly large body of evidence from experiments (Goldberg et al. 2004) and corpus 1. This chapter is based on material presented at the 4th International Conference of the German Cognitive Linguistics Association on October 8th 2010 in Bremen. I would like to thank the audience of my talk for their interest and comments. I am also indebted to Dylan Glynn, Adele Goldberg and Martin Hilpert for their comments on earlier versions of this chapter.

62

Florent Perek

studies (Stefanowitsch and Gries 2003) suggests that there is a close relation between constructional meaning and constructional usage, in that the meaning of a construction closely corresponds to the meaning of the elements that typically occur in it. In the case of argument structure constructions, this means that the meaning of verbs occurring in a given syntactic pattern determines to a large extent the meaning that will be associated with this syntactic pattern. Along the same lines, previous corpus-based studies on the interaction of syntax and lexis using the method of collostructional analysis show that “strong collexemes of a construction provide a good indicator of its meaning” (Stefanowitsch and Gries 2003:â•›227); for example, the ditransitive is biased towards verbs lexicalizing its core meaning of caused possession, such as give. Collexeme analysis is thus considered as a valid approach to the analysis of constructional meaning. This chapter presents an attempt to use collostructional analysis to describe the meaning of the conative construction, in which a typically transitive verb is followed not by a direct object, but by a prepositional phrase headed by at (e.g., The waiter wiped at the counter). As shown by the literature review presented in Section 2, previous research indicates that the meaning of the conative construction is difficult to grasp with a single semantic generalization that would be both accurate and maximally general, which points to a polysemy analysis. Along the lines of Stefanowitsch and Gries (2003), Section 3 considers whether collostructional analysis can inform a polysemy analysis of the conative construction by identifying its central meaning(s), from which other meanings could be derived. However, a collexeme analysis of the construction reveals that no single verb type clearly stands out as prototypical, as is the case with previously studied constructions. These results challenge the claim that collexeme analysis is a good way to characterize the meaning of the construction from the verbs that most prominently occur in it. In Section 4, a solution to this problem is presented that restores the relation between constructional meaning and verbal use. Drawing on an earlier proposal by Croft (2003) that constructional polysemy is better viewed as generalizations over several semantic classes of verbs rather than extensions from a prototype, a slightly different implementation of collexeme analysis is suggested, whose basic idea is to split the verbal distribution into semantic classes and consider each of these thus-defined “verb-class-specific” constructions independently. The method is applied to three classes of verbs: verbs of striking, verbs of cutting and verbs of pulling. In each class tested, the most significant collexemes are verbs whose meaning inherently contains precisely those aspects of meaning that are arguably contributed by the construction when it is used with other verbs. Hence, the most attracted collexemes do provide an indication of the constructional meaning, albeit specific to each verb class. The conclusion of this study is two-fold. At the theoretical level, it shows that the polysemy of the conative construction is better seen not as a unified network, but rather as a conglomerate that can be explained by local lexical generalizations

Rethinking constructional polysemy

over classes of verbs. Such clusters of low-level generalizations are arguably, at least in this case, a more psychologically valid mental representation of constructional meaning than general schemata deriving from prototypical verbs. At the methodological level, this study shows that looking at the level of verb classes is a useful adaptation of collexeme analysis that can appropriately deal with cases which would otherwise yield results that are difficult to interpret. It allows us to see more clearly what the semantic contribution of a grammatical construction is, albeit for each semantic class separately.

2. The conative construction The conative construction is most naturally discussed with reference to the conative alternation, whereby the direct object of a transitive verb is realized as a prepositional phrase headed by the preposition at, as in John shot at the burglar. As we will see in this section, the meaning contributed by this syntactic construction is highly variable, which makes a maximally general semantic characterization of the construction challenging, if possible at all. One of the most cited semantic characterizations of the construction is that of Levin (1993), who suggests that the construction “describes an ‘attempted’ action without specifying whether the action was actually carried out”. Pinker’s (1989:â•›104) description, viz. “the subject is trying to affect the oblique object but may or may not be succeeding”, basically refers to the same idea while further specifying the origin of the “attempted action” interpretation, namely that the conative variant lacks the entailment that the referent of the at-phrase is affected by whatever activity the agent is engaged in. Finally, Goldberg (1995:â•›63–64) formulates a construction grammar account of the construction based on these earlier observations, in which she posits that the central meaning contributed by the construction is roughly ‘x directs action at y’, and accounts for the “attempted action” interpretation reported by Levin and Pinker by stipulating that in such cases “the verb designates the intended result of the act denoted by the construction”. The common idea behind all three analyses is the notion that the conative counterpart leaves the affectedness of the at-phrase referent unspecified, whereas it is strongly (if not necessarily) implied by the transitive variant. More recent work on the conative construction shows that this characterization is not by itself sufficient to account for the interpretation of all conative sentences. As Van der Leek (1996:â•›367) notes, “the conative does not, in its own right, guarantee an intended result reading when featuring otherwise transitive verbs”. Both Van der Leek (1996) and Broccias (2001) note that many conative sentences do entail that the patient is affected, albeit to a lesser extent than the transitive counterpart. Verbs of ingestion provide a good example thereof. Indeed, such expressions as James Bond sipped at his Martini do entail that at least some of the designated substance was

63

64 Florent Perek

ingested; what they prevent is a holistic interpretation where the whole substance would be consumed. Non-affectedness thus cannot be a relevant reading for verbs of ingestion, which rather involve a ‘bit-by-bit’ interpretation. Van der Leek (1996:â•›367) also notes that “usage of verbs of ingestion in the conative often seems to be motivated by a desire to signal that no real attempt is (or even can be) made to carry out the action to completion”. Example (1) (taken from Van der Leek 1996:â•›367, originally from the Longman Dictionary of Contemporary English) exemplifies such a case: (1) [Sandy was] sipping at her drink just to be polite Sentence (1) explicitly specifies the actual goal of the sipping (to be polite), and thus entails that Sandy has no real intention to consume the whole drink. In other words, the conative construction can be found in cases where there is apparently no intention on behalf of the agent to affect the target, and hence where actual affectedness is not only unlikely but also (and more importantly) irrelevant. Subscribing to a constructional approach whereby clausal meaning results from the fusion of a verb’s meaning with an abstract schema conveyed by the syntactic construction, Broccias (2001) presents a new analysis of the conative construction. To account for instances not covered by the “attempted action” generalization and to tackle several other issues with previous studies, Broccias argues that the conative construction conveys either one of three schemas: the allative schema, the ablative schema, and the allative/ablative schema, which combines aspects of the first two. The allative schema is described in purely locative terms as involving translational motion towards a target with which contact is not necessarily made, which more or less corresponds to the aforementioned analyses in terms of “attempted action”; note, for example, that Pinker (1989) describes the output of his conative lexical rule in a similar locative fashion as ‘X goes towards X acting-on Y’. This is also reminiscent of Goldberg’s description of the construction’s central meaning in terms of “directed-action”. Broccias’ ablative schema, contrary to the former one, does imply that contact is made but does not bring about the intended effect and is open to repetition; this schema is involved, for example, with verbs of ingestion, as mentioned earlier. It should be clear from the previous discussion that the semantic contribution of the conative construction is highly variable, and is, if anything, difficult to grasp with a single generalization. What could stand as the common motivation behind all these uses is the very abstract notion that the conative construction moves the focus to what the agent is doing, regardless of whatever effect this action brings about. This proposal echoes Dixon’s (1991:â•›280) analysis, who notes that “the emphasis is not on the effect of the activity on some specific object […] but rather on the subject’s engaging in the activity”. While this account seems reasonable at first blush, such an abstract characterization must still go a long way towards the actual semantic contribution with individual verbs, leaving a heavy burden to processes of meaning construction. In addition, such a general meaning could not account for why some verbs (such as


break and bend) cannot occur in this construction, since a priori any verb meaning involving an agent subject could, in theory, undergo a focus on the agent’s activity. Thus, the syntactic frame [NP V at NP] more likely corresponds to several different abstract schemas. Whether or not these schemas can be related in a polysemic network is a matter of debate, but it seems to be a reasonable position. Indeed, the various semantic contributions sketched above can be shown to share family resemblances, which gives credence to a polysemy analysis. For example, both the ‘intended-result’ and ‘bit-by-bit’ readings share the notion that, whatever else is going on in the sentence, there is in both cases some goal which is not reached by the agent: bringing about a result on the second entity for the former, and leading an incrementally unfolding event to its completion in the latter. In the next section, the polysemy of the conative construction is examined on the basis of corpus data. Specifically, it is proposed that the central meaning (or meanings) of the construction can be identified from an examination of its verbal distribution, using the method of collexeme analysis.

3. A collexeme analysis of the conative construction Previous discussions of constructional polysemy consider that a construction gains additional meanings through semantic extensions from a central meaning. For example, the central meaning of the ditransitive construction is ‘actual change of possession’, as instantiated by, e.g., the verb give. Several semantic extensions are derived from this central meaning, such as ‘enabled change of possession’ (as with, e.g., allow) or ‘intended change of possession’ (as with many verbs of creation, e.g. bake). All these meanings are related in that they all share the notion of some change of possession, but ‘actual transfer’ is the prototypical meaning since it is both concrete and “basic to human experience”, according to Goldberg’s (1995:â•›39) scene encoding hypothesis. How do we identify the central meaning of a construction? In quantitative corpus linguistics, it has been proposed that the verbal distribution of a construction reveals a great deal about its meaning. More precisely, the most frequent verbs occurring in a construction would be those instantiating its central meaning. This section presents an attempt to identify the central meaning of the conative construction on the basis of its verbal usage, using the method of collexeme analysis. Collexeme analysis is one of the specific implementations of the more general method of collostructional analysis suited to the identification of the central meaning of a construction. This section starts with an outline of what the method consists in (cf. Hilpert’s contribution (this volume, 391–404) for a more thorough introduction). Drawing on previous research, it is then shown how this method is useful for the study of grammatical constructions. The remainder of this section presents a collexeme analysis of the conative construction.

65

66 Florent Perek

Table 1.â•‡ Contingency table for collexeme analysis Lexeme L Other lexemes

3.1

Construction C

Other constructions

F(L in C) F(other L in C)

F(L in other C) F(other L in other C)

Collexeme analysis

Collexeme analysis was first introduced by Stefanowitsch and Gries (2003) as “an extension of collocational analysis specifically geared to investigating the interaction of lexemes and the grammatical structures associated with them” (ibid.:â•›209). Collexeme analysis is concerned with the words occurring in a given slot of a chosen construction, and more particularly with “determining the degree to which particular slots in a grammatical structure prefer, or are restricted to, a particular set or semantic class of lexical items” (ibid.:â•›211). The method starts with the identification of a particular construction in a corpus, and of a particular slot of that construction that can be filled with different lexical items. For each lexeme occurring in the slot, the following contingency table must be calculated, as in Table 1. This contingency table is then submitted to a distributional statistic (often the Fisher-exact test2) to calculate the collostruction strength of the lexeme. This value gives an index of the degree of statistical association between the lexeme and the construction, given their frequency of co-occurrence, the frequency of the lexeme elsewhere, and the frequency of other lexemes in the construction. The verbs in the distribution are then ranked according to their collostruction strength. The final step (interpretation) consists in using this ordered list of collexemes to inform a description of the meaning of the grammatical construction, which is essentially guided by the theoretical assumptions of the constructional approach. In construction grammar, the occurrence of a lexeme in a construction is to a large extent determined by the degree of semantic compatibility (cf. Goldberg 1995) between the meaning of the lexeme and that of the construction (or more precisely, the meaning assigned by the construction to the particular slot under study). In collexeme analysis, collostruction strength is assumed to correlate with semantic compatibility: lexemes are more attracted to some constructional slot (i.e. occur in that slot more often than expected) if they are more semantically compatible with the slot. It thus follows that the strongest collexemes of a construction, as the most semantically compatible lexemes, are a potential source of information about the meaning of the construction. 2. Despite the wide range of available distributional statistics, Stefanowitsch and Gries (2003:â•›218) argue that the Fisher exact test is a perfect choice for collostructional analysis: it “neither makes any distributional assumptions, nor does it require any particular sample size”.


The task of the analyst is thus to track down the origin of semantic compatibility from the lexical semantics of these collexemes, so as to deduce a characterization of the constructional meaning. Stefanowitsch and Gries (2003) illustrate their claims with a few case studies showing the usefulness of the method for the description of grammatical constructions. Two of these are of particular interest for us here: the into-causative construction (Subj V Obj into V-ing) and the famous ditransitive construction (Subj V Obj1 Obj2). For the into-causative, Stefanowitsch and Gries (2003) looked at the first verb slot of the construction and found that the top collexemes are verbs “instantiating the two major sub-senses of the construction, namely ‘trickery’ (as exemplified by trick/fool […]) and ‘force’ (as exemplified by coerce/force […])” (p. 226), while verbs instantiating senses of the construction that are intuitively less central (such as ‘verbal coercion’ and ‘persuasion through a positive or negative stimulus’) appear much further down the list. As to the ditransitive construction, the verb give turns out to be by far its strongest collexeme, which is to be expected given the principle of semantic compatibility: among the many ways in which a verb can be compatible with a construction, give and the ditransitive exemplify the optimal case where there is semantic identity. In other words, since the verb give is maximally compatible with the ditransitive construction, it comes as no surprise that it is its strongest collexeme. Yet, the authors argue that, contrary to what happens with the into-causative construction, the basic ‘transfer’ sense of the ditransitive is not overwhelmingly dominant in the collexemes of the construction, in that there are relatively few significant collexemes instantiating the central sense in the whole list (6 out of 30, 10 including metaphorical uses such as tell, show and teach). Rather, the high diversity of verbs provides, according to Stefanowitsch and Gries , evidence for the polysemy analysis of the construction put forward by Goldberg. It is indeed true that instances of the central sense are a minority among the collexemes in terms of the number of types, but these few types are clearly clustered towards the top of the list: at least four of them (eight including the metaphorical uses) are among the top ten collexemes. Thus, for both constructions, there seems to be a strong tendency for the top collexemes to instantiate the most central meaning(s). Both case studies thus present evidence that collexeme analysis is a valid quantitative method to profile the meaning of constructions from their prominent verbal collocates. As Stefanowitsch and Gries (2003:â•›227) conclude, “strong collexemes of a construction provide a good indicator of its meaning”. Therefore, the method should be helpful in identifying the elusive meaning of the conative construction.

67

68 Florent Perek

3.2

Data collection

The verbal distribution of the conative construction was extracted from the prose fiction part of the BNC, containing about 16 million words in 431 texts primarily drawn from novels. The choice of this corpus was neither arbitrary nor unmotivated. Intuitively, the conative construction seems to convey a complex descriptive function which makes it more at home in narrative genres, and probably not to be found so frequently in spontaneous spoken language. The latter intuition is actually borne out by an earlier attempt at finding conative sentences in the conversation part of the corpus, revealing that the construction is extremely rare in that register (only 17 tokens in 4 million words). The corpus was queried for all verbs followed by the preposition at (with an optional intervening adverb) in the same sentence, with the exclusion of frequent verbs that cannot support a conative reading and for which at can only be used in a purely locative sense (e.g. be, stay, live, arrive, etc.).3 The resulting set of sentences was manually annotated to select only conative sentences, which were defined according to two criteria: (1) the verb has to be transitive, and (2) the interpretation of the sentences has to fall somehow into one of those described in the previous section. Sentences with coordinated verbs were duplicated in the dataset (one duplicate per verb). This yielded a final set of 2,563 instances, distributed over 159 verb types.

3.3

Results

The collostruction strength of each verb in the construction was computed by Coll.analysis 3, an R program written and kindly provided by Stefan Gries, with the Fisher exact test as a distributional statistic.4 Following Stefanowitsch and Gries (2005), Coll.analysis applies a log transformation to the p-values yielded by the Fisher exact test, and changes the sign to a plus if the association is one of attraction (i.e. the actual verb’s frequency exceeds the expected frequency) and to a minus in case of repulsion (i.e. the actual verb’s frequency is below the expected frequency). This gives a more readable value than the p-values, often expressed in powers of ten. A collostruction strength above 1.301 means that the verb is significantly attracted to the construction; a collostruction strength below –1.301 means that the verb is significantly repelled by

3. The Corpus Query Processor program, part of the Corpus Workbench suite developed at the University of Stuttgart (http://cwb.sourceforge.net/), was used to query the corpus. The corpus was assembled from the XML version of the BNC with a script that parsed all texts of the corpus and copied only those with the “prose-fiction” genre attribute. Another script then converted the corpus into a format readable by CQP. 4. Available at: http://www.linguistics.ucsb.edu/faculty/stgries/teaching/groningen/.


Table 2.â•‡ The thirty strongest collexemes of the conative construction in BNC-prose-fiction Rank Verb

f(conative:all) coll.strength Rank Verb

f(conative:all) coll.strength

â•⁄1 â•⁄2 â•⁄3 â•⁄4 â•⁄5 â•⁄6 â•⁄7 â•⁄8 â•⁄9 10 11 12 13 14 15

226:661 179:823 â•⁄72:166 â•⁄53:156 â•⁄43:97 â•⁄73:643 â•⁄36:121 â•⁄71:689 â•⁄29:87 â•⁄31:107 â•⁄44:300 â•⁄91:1363 â•⁄36:291 â•⁄76:1217 â•⁄22:140

29:263 43:567 24:180 18:112 13:56 35:524 17:149 â•⁄9:32 â•⁄8:26 26:364 35:656 17:190 51:1186 11:112 23:466

tug clutch dab claw gnaw sniff nibble sip peck nag pluck tear stab grab hack

209.92 127.13 â•⁄75.74 â•⁄49.14 â•⁄46.02 â•⁄32.05 â•⁄31.26 â•⁄28.56 â•⁄26.95 â•⁄26.62 â•⁄24.13 â•⁄22.51 â•⁄17.41 â•⁄17.29 â•⁄13.08

16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

hammer snatch jab scrabble paw scratch slash swipe niggle poke suck prod kick lap strain

12.87 12.86 12.58 11 10.23 â•⁄9.13 â•⁄8.07 â•⁄8.07 â•⁄7.58 â•⁄7.55 â•⁄6.7 â•⁄6.52 â•⁄6.44 â•⁄4.82 â•⁄4.13

the construction. As noted above, the verbs at the top of the distribution ordered by collostruction strength provide an indication of the constructional meaning. The thirty strongest collexemes of the conative construction are reported in Table 2. As it turns out, the construction attracts a great variety of verbs. Almost all verb classes allowed in the construction are represented in that list: verbs of pulling (tug, pluck), verbs of seizing and holding (clutch, claw, grab, snatch), verbs of hitting and touching (dab, claw, peck, stab, hammer, jab, paw, swipe, poke, prod, kick), verbs of ingestion (gnaw, nibble, sip, peck, suck, lap), verbs of cutting (tear, hack, slash), etc. This result is not surprising in itself, as constructions are often associated with several related senses, and therefore several classes of verbs. Again, this points to a polysemy analysis, as indeed the collexemes presented in Table 2 arguably instantiate different senses of the construction. For example, assuming Broccias’ (2001) distinctions (cf. Section 2), clutch, stab and kick mostly instantiate the allative schema, while nibble, hack and suck rather instantiate the ablative schema. While it is, a priori, not problematic that the construction attracts different classes of verbs, the list of collexemes is, however, not particularly helpful in characterizing the construction’s meaning. Moreover, contrary to what happens in the case studies reviewed above, there does not seem to be a class of verbs that the construction attracts in particular. The list presents alternations of very different types of verbs, and no particular class seems to be more strongly attracted than the others. For example, the five most attracted collexemes exemplify precisely five different verb classes: tug (verb of pulling), clutch (verb of seizing/holding), dab (verb of touching/hitting), claw

69

70 Florent Perek

(verb of hitting or seizing/touching) and gnaw (verb of eating/chewing). Moreover, these verbs exemplify various semantic aspects of the construction: tug at entails no change of location and an inherent repetition of the attempt, clutch at and claw at entail either missed contact or prolonged exertion of a force, dab at entails little or no affectedness, gnaw at entails no completion. Thus, contrary to what Stefanowitsch and Gries (2003) found with the ditransitive construction, collexeme analysis is not helpful in identifying one (or more) particular sense of the conative construction which would be central and from which the other senses would be derived. As a matter of fact, there need not be an identifiable verb class corresponding to each constructional sense: since the senses of the conative construction are so highly abstract, they are liable to be combined with a great variety of verbs from different semantic classes. Hence, it is not particularly surprising that the collexeme list of the conative construction (or probably of any abstract construction) is not as easily interpretable as that of the ditransitive construction. As one reviewer suggests, this might be because the semantics of the conative construction is less directly related to basic bodily experience than that of the ditransitive or of the caused-motion construction; as such, it is less likely to correspond to patterns of lexicalization in the language in general. This means that collexeme analysis in its present form would not be able to identify the senses of many constructions, at least not as neatly as those of the ditransitive construction (for example).

3.4

Towards a solution

In the face of such results, this chapter suggests another approach based on a refinement of collexeme analysis, which might be more informative in the case of the conative construction, and probably many other constructions. This approach is motivated by an earlier proposal by Croft (2003), who criticizes the concept of constructional polysemy, and thus the related notion of a “central” meaning. According to Croft, the very concept of constructional polysemy is problematic in several respects. The main problem can be roughly summarized as follows: how can a construction be considered truly polysemous if its meaning in context only depends on the verb it is instantiated with? In the case of the ditransitive construction, Croft (ibid.:â•›55) notes that “each semantic class is associated with only one sense of the ditransitive construction”. This seems to be in part semantically motivated: for example, the fact that the modal extension (‘conditions of satisfaction imply that X causes Y to have Z’) is the only one occurring with promise (for instance) is expected since it is the only extension whose specifications do not conflict with the meaning of the verb. However, why the extension ‘X intends that Y have Z’ is the only one compatible with verbs of creation appears to be completely arbitrary, since there is nothing in the verb’s meaning that blatantly conflicts with a number of the other extensions, whose instantiation with the verb would thus make perfect sense.


A polysemic analysis of the conative construction runs into exactly the same problem: while there can be several different readings of a single conative sentence, not all interpretations are equally available in all instances. For example, in no case would conative sentences with verbs of ingestion mean ‘X moves towards Y in order to ingest Y’. Conversely, verbs of rubbing could never be used in the conative construction to convey the meaning ‘X rubs a part of Y and goes towards having Y totally rubbed’, let alone an allative interpretation (i.e. ‘X goes towards Y to rub Y’).5 Sometimes the unavailability of some readings is straightforwardly explained by intrinsic properties of the verbs themselves: for example, the impossibility of an incremental reading with semelfactives such as hit and kick can be explained by the aspectual properties of these verbs and more particularly the absence of an incremental theme. However, there are still perfectly sensible combinations that, nonetheless, are disallowed, which would not be the case if the construction was truly polysemous. Croft suggests that such cases are more appropriately accounted for not by considering the construction as genuinely polysemous, but by treating it as several “verb-class-specific constructions”, i.e. lower-level generalizations of a constructional meaning over a clearly delimited semantic verb class, instantiated only with verbs of that class. The remainder of this chapter presents evidence that this view might also be more appropriate for the conative construction. As observed in Table 2, no particular meaning stands out in the whole distribution of the construction. However, if we look again at Table 2 by focusing on verbs from a specific semantic field, a clearer picture emerges. A class that is fairly easy to delimit is that of verbs of eating. Table 3 reports the distribution of verbs of eating in the conative construction (the significantly attracted and significantly repelled collexemes appear on a gray background). Table 3.â•‡ Verbs of eating in the conative construction Verb

f(conative:all)

coll.strength

nibble peck suck lick gulp gobble munch pick eat

36:121 29:87 35:656 20:488 â•⁄9:267 â•⁄1:60 â•⁄1:84 79:4678 12:4089

â•⁄31.26 â•⁄26.95 â•⁄â•⁄6.7 â•⁄â•⁄2.68 â•⁄â•⁄1.07 â•⁄–0.18 â•⁄–0.3 â•⁄–1.1 –21.53

5. Conative uses of rub and other similar verbs (wipe, brush, …) do receive a form of “non-affectedness” interpretation which is not ‘X tries to rub Y’ but rather corresponds to a scenario in which some entity remains unaffected; this entity might be mentioned (as in rub at the stain) or might remain implicit or unspecified (as in rub at the counter, which most likely entails that the agent’s goal is to clean the counter and that this goal is not achieved).

71

72

Florent Perek

The most strongly attracted verb in that class, nibble, denotes an event of eating where only a small amount of some substance is ingested, and is therefore inherently compatible with the “bit-by-bit” reading supported by the construction. In fact, this verb is similar to give in the ditransitive construction: assuming a more specific eating-conative construction instantiated by verbs of eating only and whose meaning would be ‘eat in a bit-by-bit fashion’, the meaning of nibble is identical to the meaning of that construction, which largely motivates the prominent occurrence of that verb in the construction. The other significantly attracted collexemes also support the ‘bit-by-bit’ interpretation. Peck typically refers to how birds eat, by moving their beak forward repeatedly; in the conative construction, it is also frequently used to refer to people eating only a small amount of their meal. Suck and lick are not purely verbs of eating but rather describe a kind of action that an agent performs on another entity; when they are used to describe events of eating (as they very often are in the corpus), both typically refer to a slow and gradual means of ingestion through the progressive dissolution of a substance. Finally, the sole collexeme repelled by the construction is eat; this again reflects the semantic preferences of the construction, as eat is a maximally neutral verb of ingestion which is more commonly used to denote total consumption and lends itself less easily to a ‘bit-by-bit’ interpretation.6 This simple example shows that focusing on a particular class of verbs clearly captures what the semantic contribution of the construction is for this particular class. Thus, a collexeme analysis at the level of individual verb classes seems to be a promising approach. The next section elaborates on this proposal and presents a version of collexeme analysis based on semantic classes.

4. A collexeme analysis of verb-class-specific constructions In the previous section, it was found that a collexeme analysis performed on the whole distribution of the conative construction is not very helpful in characterizing its constructional meaning and does not clearly support a polysemy analysis either. 6. As Dylan Glynn notes in a review of an earlier version of this chapter, it is somehow unexpected that pick does not appear among the attracted collexemes of the construction, let alone that it almost reaches the threshold of repulsion, since pick at indeed seems to be a prime example of the ‘bit-by-bit’ reading induced by the construction. This result is explained by the fact that the verb pick is highly polysemous and at the same time highly frequent, and that it is not primarily a verb of eating: in fact, it probably occurs in this sense in the conative construction only. This asymmetry in the semantic distribution of pick thus appears to obscure its contribution to our understanding of the meaning of the construction. The general issue of the relation between frequency of verb forms and frequency of verb senses is taken up again in Section 4.1.2.


It was observed that a clearer picture emerges if we look only at verbs from a specific semantic class (in that case, verbs of eating): the meaning of the strongest collexemes clearly reflects the semantic contribution of the construction for this semantic class. This section outlines a more principled and systematic formulation of this approach and then presents its application to three classes of verbs: verbs of cutting, verbs of pulling and verbs of striking.

4.1

Method

This section first explains how verbs in this study were classified into semantic classes. It then turns to some statistical issues posed by the present approach.

4.1.1 Determining verb classes The present approach first requires that the verbs from the distribution of the conative construction are sorted into several classes. Of course, a given verb form can correspond to several meanings, and these meanings can belong to different semantic classes. For example, in Table 2, peck and pick can function as verbs of eating but also as verbs of striking (albeit more rarely). However, the frequencies obtained from the corpus are frequencies of verb forms, not of verb meanings, and thus some of these frequencies may actually be distributed over several semantic classes. All instances of a verb form cannot just be assigned to a single class or be counted in several classes simultaneously: it must be determined for each token to which semantic class it belongs. For the example of verb-class-specific collexeme analysis presented in the last section, the field of verbs of eating was relatively easy to select from the whole distribution. However, it might not be so easy to identify, on the sole basis of intuition, the verb classes found in the distribution and the semantic class each verb token belongs to. To facilitate this process, an external lexicographic source was relied on: WordNet (Fellbaum 1998), a lexical database of the English language which was created and is being maintained at the Cognitive Science Laboratory of Princeton University. It groups English words into sets of synonyms (called synsets) and provides lists of the various meanings of each word form that can be looked up to perform semantic annotation. Starting with an established list of sense distinctions, instead of building it during the annotation process, is not only convenient, it also allows the achievement of a crucial feature of empirical studies of meaning: overt operationalization (cf. Glynn 2010), in the sense that the analytical criteria are overtly identified. This makes the analysis falsifiable, since it enables it to be repeated on the same data or on another dataset (e.g. for the purpose of comparison). The list of verb senses could be drawn from any dictionary, but WordNet presents another useful feature for this approach: it records relations between synsets such as

73

74

Florent Perek

hyponymy, hyperonymy, part-whole relations, entailments, etc. Of particular interest to us, the relations of hyponymy (and conversely, hyperonymy) connect the synsets into a type hierarchy, which can be used to define verb classes: a verb class includes the verbs of a given synset and all of its hypernyms, i.e. verbs whose meaning includes (and often, elaborates) the meaning of the synset. Hence, co-hyponyms belong to the same class. In sum, WordNet can be used both to annotate for verb senses and to define verb classes on the basis of the annotated data and hyponymy/hyperonymy relations between senses recorded in the database. It has been noted elsewhere that WordNet sense distinctions are somehow arbitrary and sometimes so fine-grained that it is practically impossible to apply the classification to naturally occurring examples (not to mention the theoretical vacuity and actual impracticability of the very notion of sharp sense boundaries, cf. Kilgarriff 1997; Glynn 2010). While this is true in many cases, in the context of this study it is often unproblematic to ignore some sense distinctions as long as they do not extend over different verb classes. For example, drag has two senses in WordNet that may apply to conative uses of the verb: (i) ‘pull, as against a resistance’ and (ii) ‘draw slowly or heavily’. It is not clear from the glosses what the semantic difference is supposed to be, and if anything it is very subtle and therefore not easily applicable to the annotation of examples in context. This distinction can, however, be ignored, since both senses have pull as their direct hyperonym: they can thus be conflated into a single entry, drag, subsumed by the class of verbs of pulling. Even though the fine-grained sense distinctions posited in WordNet might not always be well-grounded, the coarser-grained distinctions imposed by verb classes are more reliable and more easily noticeable. This strategy thus avoids the pitfalls of drawing strict sense boundaries. The original dataset was manually annotated for WordNet senses with the help of an interactive program.7 As it turns out, while some verbs are highly polysemic according to WordNet’s classification, the conative construction is usually restricted to one or two senses of these verbs, and most verbs can belong to only one semantic class when they occur in the construction. The verb sense distribution was built by calculating the frequency of each word sense in the construction. Each verb sense in this distribution was then annotated with the synset ID of its direct hyperonym, or with its own synset ID if the verb sense is a hyperonym of other verbs in the distribution. This ID identifies both the class to which the verb belongs, and the most general verb (i.e. hyperonym) of that class. In the case of classes subsumed by another class, which can be diagnosed by the hyperonym of one class being a member of another class, the lower class was merged into the higher one. As a last step, in each class, senses of the

7. This tool was written in Java and uses the JWNL API to read the WordNet 3.0 files (http:// sourceforge.net/projects/jwordnet/), downloaded from the website (http://wordnet.princeton. edu/wordnet/download/).


same verb form were collapsed into one cell summing all frequencies of the verb form. With this method, maximally large and distinctive verb classes were obtained.

4.1.2 Statistical matters In the collexeme analysis of verbs of ingestion in Section 3.4, verbs were just filtered out on the basis of their belonging to the semantic class under study. However, if the verb-class-specific constructions hypothesis is taken seriously, a collexeme analysis of a specific semantic class only makes sense if the collostruct under consideration is not the general construction but a more specific one taking only verbs of this semantic class, and since such constructions have a lower frequency than the more general one, the actual collostruction strength values could be slightly different, hence changing the significance of some collexemes and possibly the order of the collexeme list. The frequency of a verb-class-specific construction is obtained by summing the frequency of all verb senses in the class. There is, however, still one missing set of frequencies: the frequency of each verb sense in other constructions. Unfortunately, except with a semantically annotated corpus, there is no easy way to determine this frequency, as it is practically intractable to manually annotate the whole corpus for verb senses. It must be acknowledged that this is an inherent weakness of this approach. However, as serious as it might be, this problem can be attenuated using two methods. First, in each verb class, only those verb senses that were by far the most frequent instance of their verb form are kept in the analysis. For example, catch occurs only seven times as a verb of striking in the conative construction versus fifty times in other senses (mainly as a verb of seizing); it was thus removed from the list of verbs of striking and does not appear in Table 5. The rationale behind this decision is that a verb form occurring clearly less prominently in a given verb-class-specific construction than in the other ones should be a weak collexeme of the construction anyway and is not likely to tell us much about the constructional meaning.8 Second, the overall frequency of the verb form was used for each verb sense, which makes the assumption that every occurrence of each verb form in the corpus has the meaning that the verb has in the conative construction. This is, of course, surely false for polysemous verbs, though not overly problematic for this study since it will merely downplay the collostruction strength of verbs. Indeed, the frequency of a verb sense is at least as high as the frequency of the verb form, and for polysemous forms it is a priori lower. The approximate collostruction strength calculated with the frequency of the verb form will thus be lower than the theoretical collostruction strength that would be calculated with the frequency of the verb sense, thus probably narrowing the range of significant collexemes. As it turns out, this 8. The deleted verbs include: scrape, scratch and slash for the cutting-conative construction, catch, pick, tweak and twitch for the pulling-conative construction, and catch, jab, peck, pick and poke for the striking-conative construction.

75

76

Florent Perek

possible downplaying of the attraction of the verbs to the construction does not prevent the identification of a number of interesting collexemes in each class.

4.2

Results

This section reports the collexeme analysis performed on the cutting-conative, pulling-conative and striking-conative constructions, defined as elaborations of the conative construction instantiated, respectively, by verbs of cutting, verbs of pulling and verbs of striking.

4.2.1 Verbs of cutting Events of cutting involve an agent moving a suitable instrument over the surface of an object, and causing a rupture in the physical integrity of that object as a result. With verbs of cutting, the conative construction does not support the allative interpretation (or at least not literally): contact is necessarily made between some instrument and the referent of the at-phrase, but this contact does not bring about the effect that the transitive use of the verb would entail: the cutting either fails entirely, or is too minimal for one to consider that the object is indeed cut. Hence, conative uses of verbs of cutting often convey the implicature that the action performed to do the cutting is repeated. Table 4 presents the collexemes of the cutting-conative construction. The analysis reveals three significantly attracted collexemes: hack, saw and chip. All three collexemes are particularly suited to the semantic contribution of the cutting-conative construction. The lexemes hack and saw are inherently repetitive: an event of hacking or sawing always consists of several identical actions. Moreover, a single movement (a stroke of a hacking tool or of a saw) generally does not by itself bring about the intended effect on the patient, e.g. cutting something to bits or sawing a piece of wood apart; the Table 4.â•‡ Collexemes of the cutting-conative construction Verb

f(conative:all)

coll.strength

WordNet gloss

hack saw chip chisel snip chop slice nick cut

22:140 â•⁄6:74 â•⁄4:93 â•⁄2:39 â•⁄2:54 â•⁄3:174 â•⁄3:237 â•⁄2:163 â•⁄4:3075

â•⁄19.76 â•⁄â•⁄3.69 â•⁄â•⁄1.63 â•⁄â•⁄1.11 â•⁄â•⁄0.87 â•⁄â•⁄0.47 â•⁄â•⁄0.27 â•⁄â•⁄0.23 –22.71

cut with a hacking tool cut with a saw break a small piece off from carve with a chisel sever or remove by pinching or snipping cut into pieces make a clean cut through cut a nick into separate with or as if with an instrument


movement must be repeated until the desired effect is obtained. Hence hack and saw naturally support the semantic contribution of the cutting-conative construction in their conceptual semantics, i.e. both ‘no-significant-effect’ and ‘repetition’. The item chip inherently features only one of these two aspects. In any event of chipping, only a small piece of the patient is broken off, and chip does not in any case support a truly holistic interpretation, i.e. an object that is chipped is only minimally affected and keeps its overall physical integrity, compared to what happens with true verbs of change of state like break. Events of chipping must be repeated if the patient is to be considered significantly affected. The only significantly repelled collexeme in the list is cut. Its repulsion can be explained by its status as a maximally neutral verb of cutting (and indeed the hyperonym of the whole class), which thus does not carry any semantic elaboration that would promote its use in the conative construction. In addition, cut lends itself to a holistic interpretation to a much larger extent than the attracted collexemes.

4.2.2 Verbs of pulling Events of pulling consist in an agent exerting a force on a patient, usually in order to move the patient towards self or to affect it in some other way (e.g. open a door). The effect on the patient is not an inherent feature of these verbs, but is rather a frequent implicature of their transitive use. The conative construction prevents this implicature of change of location/state, thus bringing the interpretation towards an ‘attempted action’ reading. Such uses also easily allow an interpretation of repeated actions, since a single iteration of pulling does not bring about a significant effect. Table 5 lists the collexemes of the pulling-conative construction. The construction has two significantly attracted collexemes: tug and pluck. According to the Oxford English Dictionary, tug applies to events where the puller puts a lot of energy in the pulling, or exerts a force during an extended period. Hence, tug focuses on the effort the agent puts into the act of pulling, and not so much on the dynamics of the event itself, i.e. whether the patient is set in motion or not. Table 5.â•‡ Collexemes of the pulling-conative construction Verb

f(conative:all)

coll.strength

WordNet gloss

tug pluck wrench yank haul jerk drag pull

226:661 â•⁄42:300 â•⁄12:314 â•⁄â•⁄1:122 â•⁄â•⁄5:411 â•⁄â•⁄8:717 â•⁄25:1528 138:6024

153.73 â•⁄10.31 â•⁄–0.49 â•⁄–1.64 â•⁄–3.9 â•⁄–7.02 –10.49 –38.41

pull hard pull or pull out sharply twist or pull violently or suddenly pull, or move with a sudden movement draw slowly or heavily pull, or move with a sudden movement draw slowly or heavily apply force so as to cause motion towards the source of the motion

77

78

Florent Perek

Pluck as a verb of pulling is often used to refer to the removal of some object from where it grows, e.g. fruit, plants, hair, or feathers. To overcome the inherent resistance of the ground to which the object is attached (e.g. skin, branch, earth), acts of plucking frequently involve a sharp and sudden pull so as to abruptly separate the object from its ground (as alluded to by WordNet’s gloss). The more general use of this verb to refer to other kinds of pulling keeps this ‘sharp and sudden’ aspect. Due to their short duration, acts of plucking are particularly prone to repetition. As indicated earlier, the repelled collexemes may have slightly overestimated repulsion scores; thus, the values of the five repelled collexemes (yank, haul, jerk, drag and pull) have to be interpreted with caution. However, the last two (drag and pull) provide some interesting insight into the construction’s meaning. Drag is more appropriately described as a verb of accompanied motion (i.e. where both agent and theme move along the same path, like bring) rather than a pure verb of pulling: it strongly presupposes the motion of the patient, which makes it at odds with the conative construction. Pull is, of course, the hyperonym of the semantic class, i.e., it is arguably the most neutral verb of pulling. Since it has no inherent semantic traits that particularly favor the conative reading(s), its appearance as a repelled collexeme is expected.

4.2.3 Verbs of striking Verbs of striking represent the largest of the three semantic classes under study. It comprises verbs that have either hit or strike as their hyperonym in WordNet. Events of striking consist in an agent performing some movement in the direction of a patient, aiming at forceful contact with the patient, usually with the intention of affecting it in some way (doing it harm or damage). In the conative construction, verbs of striking typically assume an allative interpretation: some effort is directed towards a goal (here, bringing about an effect on the patient) that is not reached. Table 6 lists the collexemes of the striking-conative construction. The significantly attracted collexemes include dab, hammer, swipe, buffet, kick, pummel and swat, and all of these verbs feature one or more particular semantic traits favored by the construction.9 The verb dab, by far the strongest collexeme, is categorized by WordNet as a verb of striking, though it is a very peculiar one. Contrary to more typical members, dabbing involves little energy and is normally not aimed at affecting the target, or at least not negatively. Rather, typical instances of dabbing include using a cloth to gather and remove a substance (like blood or tears) from a surface, or gently applying a substance

9. Table 5 also lists buffet as a significant collexeme. However, it is a very rare verb in our reasonably large corpus, occurring only twice, yet each time in the conative construction, which probably explains why it reaches the significance threshold. Since its rarity makes it a poor candidate as a relevant and telling collexeme of the construction, it was removed from the discussion.


Table 6.â•‡ Collexemes of the striking-conative construction Verb

f(conative:all) coll.strength

dab hammer swipe buffet kick pummel swat batter slap

71:166 29:263 â•⁄9:32 â•⁄2:2 51:1186 â•⁄4:31 â•⁄3:27 â•⁄7:161 16:510

â•⁄66.44 â•⁄â•⁄9.56 â•⁄â•⁄6.81 â•⁄â•⁄3.1 â•⁄â•⁄2.89 â•⁄â•⁄1.98 â•⁄â•⁄1.41 â•⁄â•⁄0.78 â•⁄â•⁄0.44

tap lash whack scuff whip bat bash punch pound

24:802 â•⁄8:265 â•⁄1:37 â•⁄1:44 â•⁄9:350 â•⁄1:71 â•⁄1:85 â•⁄5:278 â•⁄4:245

â•⁄â•⁄0.4 â•⁄â•⁄0.33 â•⁄–0.14 â•⁄–0.19 â•⁄–0.32 â•⁄–0.39 â•⁄–0.51 â•⁄–0.69 â•⁄–0.75

thump

â•⁄4:322

â•⁄–1.31

hook beat bang smash pat strike

â•⁄2:228 27:1372 â•⁄8:602 â•⁄4:421 â•⁄6:545 34:1990

â•⁄–1.37 â•⁄–1.62 â•⁄–1.96 â•⁄–2.14 â•⁄–2.3 â•⁄–3.39

hit

â•⁄7:2007

–17.96

WordNet gloss hit lightly beat with or as if with a hammer strike with a swiping motion strike against forcefully strike with the foot strike, usually with the fist hit swiftly with a violent blow strike against forcefully hit with something flat, like a paddle or the open hand strike lightly strike as if by whipping hit hard poke at with the foot or toe strike as if by whipping strike with, or as if with a bat hit hard deliver a quick blow to hit hard with the hand, fist, or some heavy instrument hit hard with the hand, fist, or some heavy instrument hit with a hook hit repeatedly strike violently hit hard hit lightly deliver a sharp blow, as with the hand, fist, or weapon deal a blow to, either with the hand or with an instrument

on a surface (e.g. for medical or cosmetic purposes). This typical lack of affectedness of the patient in an act of dabbing is in line with the meaning of ‘non-effective action’ that the conative construction is often claimed to convey. The verb hammer originally refers to an act of hitting involving a hammer or a similar tool as instrument; in that restricted use, typical things that can be hammered include nails, metal sheets and other metallic goods. If anything, this use of hammer typically entails repetition, i.e., just as with hack in Section 4.2.1, any event of hammering normally involves multiple blows on the patient, since a single blow does

79

80 Florent Perek

not suffice in affecting the patient in the intended way. For example, nails are rarely properly hammered into a wall with a single blow, but rather inserted only partly, and the hammering must be repeated as many times as necessary. Similarly, a sheet of metal can never be shaped into any appropriate form with a single blow; it has to be worked until the intended shape is arrived at. Of course, the verb in its modern use is not restricted to describe exclusively acts of striking with a hammer, but the aspects of ‘minimal effect’ and ‘repetition’ found in the original meaning of the verb arguably subsist (as comfirmed by modern dictionaries), and the instrumental component is echoed by the notion of a forceful and violent striking usually accompanied by loud noise (which many dictionaries gloss as ‘as if with a hammer’). The verbs swipe, kick and swat are similar cases in that they refer to a precisely defined shape of motion in space. In other words, what makes an event of swiping, kicking or swatting, is, above all, a particular movement performed by the agent, respectively a swinging blow10 (of the arm or of an instrument), an outward motion of the foot, and the motion of a flat surface (an open hand or an instrument with the appropriate shape) through the air so that the surface hits a target (often an insect, crushing it). This makes these verbs agent-centered, i.e., they focus on describing what the agent is doing rather than the effects that its action may have. In addition, kick specifies the body part involved (a leg), further reinforcing its agent-centered character. Strikingly, there turn out to be much fewer verbs with a focus on the shape of motion among the other (i.e. non-attracted) collexemes. Possible candidates include lash, whip, slap and possibly punch; however, the shape evoked by the former two is due to the kind of instrument used rather than the action performed itself, and the latter two less obviously refer to a fully described shape. The other verbs rather focus on the manner of impact or on its effects. It thus seems that this semantic property (‘precisely defined shape’) is highly correlated with the striking-conative construction. Finally, pummel combines aspects of hammer and of the agent-centered verbs. It is slightly agent-focused since it refers to a particular body part (the fists). But more importantly, it is inherently repetitive, as all consulted dictionaries indicate: pummeling consists of a succession of small blows, most often dealt with the fists. As for the repelled collexemes, the usual cautioning remarks apply. Let us, however, note that, just like with the cutting- and pulling-conative constructions, the maximally neutral verbs hit and strike are, as expected, the most repelled collexemes of the striking-conative construction.

10. As a confirmation of this analysis, the OED notes that swipe is chiefly used in the context of cricket.

4.3


Discussion

As should be clear from the preceding discussion, Stefanowitsch and Gries’ (2003) claims about the relation between the collexemes attracted to a construction and that construction’s meaning are clearly borne out for these three verb-class-specific instantiations of the conative construction. Namely, the attracted collexemes all prominently profile in their inherent semantics one or more semantic trait(s) that the construction contributes by itself when it occurs with other verbs. The semantic generalizations that each collexeme supports are reported in Table 7. The collexeme list clearly exemplifies the principle of semantic compatibility and how this principle bears on usage; namely, verbs with a meaning that lends itself particularly well to the interpretation sanctioned by the construction are “attracted” by it: they are much more frequent in that construction than chance would predict. Conversely, the hyperonym of the semantic class is the most repelled collexeme in each case, which the principle of semantic compatibility also predicts since such verbs are supposedly the most neutral verbs in their class, and thus do not profile any particular semantic trait that would attract them to the construction. In conclusion, it seems possible to characterize the meaning of the conative construction, or more precisely, the meaning the construction contributes when it Table 7.â•‡ Semantic generalizations supported by the collexemes of verb-class-specific constructions Verb-class-specific Construction

Collexemes

Semantic generalization(s)

cutting-conative

hack saw

event consisting of several identical movements with a minimal individual effect; hence it is inherently unbounded and repetitive

chip

minimal effect; no holistic interpretation

tug

focus on the effort (energy and duration) that the agent puts into the action rather than its effects

pluck

idem, plus a short duration which makes it prone to repetition

dab

lowly energetic; patient often not directly affected

hammer

inherently consists of several repeated blows; a single blow does not produce a sufficient effect

swipe kick swat

agent-centered: they profile a precisely defined motion that the agent performs, as well as information on the entity set in motion

pummel

profiles a body part (fists), inherently repetitive

pulling-conative

striking-conative

81

82

Florent Perek

combines with verbs of each semantic class under study, simply by attending to the salient semantic properties of the collexemes in each class. Of course, these collexemes do not lexicalize one of the meanings of the conative construction per se, as is the case with give and the ditransitive construction. But there is still arguably some abstract semantic quality shared between the collexemes and the constructional meaning as it occurs with other verbs. Such a semantic characterization would be much more difficult (if possible at all) to arrive at by looking at the entire distribution, i.e. at the level of the general construction vs. the more specific verb-class-specific constructions. The methodological and theoretical implications of this finding are elaborated on in the concluding words of the next section.

5. Conclusion As the first large-scale corpus-based investigation of the conative construction, this study contributes to the documentation of the construction’s usage. Its initial goal was to see what the verbs most frequently used with the construction could tell us about its meaning, drawing on the method of collexeme analysis. As it turns out, a collexeme analysis of the construction based on data from the prose-fiction part of the BNC fails to highlight its central meaning(s), since there does not seem to be a particular kind of verb that the construction attracts. Hence, while the collexeme list is not totally at odds with the meaning of the construction as it has been characterized introspectively, in this case collexeme analysis does not seem to be helpful in characterizing it precisely. To solve this problem, a different kind of analysis was proposed. Instead of considering the conative construction as a whole, the focus was shifted to verb-class-specific constructions, i.e. elaborations of a construction instantiated by verbs from a specific semantic class. A collexeme analysis was performed on three verb-class-specific constructions, respectively instantiated by verbs of cutting, verbs of pulling and verbs of striking, identified on the basis of the lexical database WordNet. The collexemes of each of these lower-level constructions feature in their inherent meaning the semantic traits that are characteristic of verbs of that class when they occur in the construction. In other words, collexeme analysis profiles the constructional meaning much better at the level of each verb class than at the most general level. Of course, it does not mean that collexeme analysis is ineffective for the conative construction taken as a whole; it is just not particularly telling. The collexemes found for the overarching construction are attracted because they are more compatible with the constructional meaning. But the conative construction is so multifaceted when taken at the most general level that it is much easier to understand why these verbs are collexemes and what this tells us about the meaning of the construction if we go down to the level of verb classes.


On the theoretical side, these results shed some light on the nature of constructional generalizations. Namely, a long-standing debate in constructional approaches to grammar is concerned with which level of generalization best reflects speakers’ knowledge of constructions. In the case of argument structure, earlier constructional approaches (cf. Fillmore and Kay ms.; Goldberg 1995) sought to cast the broadest generalizations possible by positing one single very abstract meaning accounting for all instances of the construction, either directly or through an extension of the constructional meaning. However, more recent research questions this commitment and emphasizes the importance of lower levels of generalizations to appropriately account for the distribution and meaning of constructions; see, for example, Boas’ (2003) concept of “mini-constructions” to account for English resultatives, and of course Croft’s (2003) proposal for “verb-class-specific constructions” (cf. also Fillmore 2001; Glynn 2004). Of course, the debate “general vs. local” might appear null and void in a truly constructional account, in which both abstract schemas and their various elaborations can be stored at any level of generality. But if a number of local generalizations alone account for what appears at first sight to be a single general construction, this casts the question of whether the overarching construction is needed at all, all the more so if the local generalizations provide a better account in terms of accuracy and coverage. This is precisely what happens with the conative construction: to the extent that speakers attend to frequently occurring verbs in some syntactic context, and use that information to “get a ‘fix’ on the construction’s meaning” (Goldberg 2006:â•›92), they can usefully exploit this lexical semantic information only at the level of verb-class-specific-constructions. Under this view, a verb appears to be a collexeme of the general construction only because it is, first and foremost, a collexeme of a verb-class-specific construction. In sum, the results of this study suggest a different view of the polysemy of the conative construction, which can plausibly be extended to other constructions. The various meanings of the conative construction are better seen not as a network of related senses, but as a cluster of low-level generalizations over similar verb meanings, in line with Croft’s (2003) proposal. As a reviewer points out, it thus would seem as if we are actually dealing with a case of constructional homonymy, i.e. several constructions sharing the same form but conveying different meanings. However, the possibility that these verb-class-specific constructions might be, at least to some extent, unified under a higher-level generalization should not be entirely rejected. The fact that low-level generalizations can determine the semantic contribution of the syntactic pattern for verbs of the semantic class does not exclude the possibility of cross-generalizations between different classes. First, if several distinct verb classes receive the same semantic contribution (which is plausible, since the conative construction conveys a wide yet still limited range of meanings), they could form a single higher generalization, which in turn could be used to produce new combinations. Second, patterns of analogy between different classes might well play a major role in determining the distribution and in helping speakers get at the correct

83

84

Florent Perek

interpretation, forming generalizations of intermediate scope. The generalizations accounting for the conative construction could well be centered on a few classes first, from which an abstract meaning could be extracted and applied to other verbs and classes. Such a scenario is probably necessary to explain the inclusion of “orphans”, i.e. verbs whose semantic class does not have any other representative in the distribution. Obviously, there is still much to learn about the workings of constructional generalizations. I hope, however, to have presented in this chapter a promising application of collexeme analysis to understand the mechanisms of constructional abstraction and the possible underlying representations on the basis of corpus data.

References Boas, H. (2003). A constructional approach to resultatives. Stanford: CSLI Publications. Broccias, C. (2001). Allative and ablative at-constructions. In M. Andronis, C. Ball, H. Elston, & S. Neuvel (Eds.), CLS 37: The main session: Papers from the 37th meeting of the Chicago linguistic society. Volume 1 (pp. 67–82). Chicago: Chicago Linguistic Society. Croft, W. (2003). Lexical rules vs. constructions: A false dichotomy. In H. Cuyckens, T. Berg, R. Dirven, & K. Panther (Eds.), Motivation in language: Studies in honour of Günter Radden (pp. 49–68). Amsterdam: John Benjamins. Dixon, R. (1991). A new approach to English grammar: On semantic principles. Oxford: Clarendon Press. Fellbaum, C. (1998) (Ed.). WordNet: An electronic lexical database. Cambridge, MA: MIT Press. Fillmore, C., & Kay, P. (MS). Construction Grammar (course reader). University of California, Berkeley. Fillmore, C. (2001). Mini-grammars of some time-when expressions in English. In J. Bybee, & M. Noonan (Eds.), Complex sentences in grammar and discourse: Essays in honor of Sandra A. Thompson (pp. 31–60). Amsterdam: John Benjamins. Glynn, D. (2004). Constructions at the crossroads: The place of construction grammar between field and frame. Annual Review of Cognitive Linguistics, 2, 197–233. DOI: 10.1075/arcl.2.07gly Glynn, D. (2010). Testing the hypothesis: Objectivity and verification in usage-based cognitive semantics. In D. Glynn, & K. Fischer (Eds.), Quantitative cognitive semantics: Corpus-driven approaches (pp. 239–270). Berlin: Mouton de Gruyter. DOI: 10.1515/9783110226423 Goldberg, A. (1995). Constructions: A construction grammar approach to argument structure. Chicago: University of Chicago Press. Goldberg, A. (2006). Constructions at work: The nature of generalization in language. Oxford: Oxford University Press. Goldberg, A., Casenhiser, D., & Sethuraman, N. (2004). Learning argument structure generalizations. Cognitive Linguistics, 15(3), 289–316. DOI: 10.1515/cogl.2004.011 Kilgarriff, A. (1997). I don’t believe in word senses. Computers and the Humanities, 31(2), 91– 113. DOI: 10.1023/A:1000583911091 Levin, B. (1993). English verb classes and alternations. Chicago: Chicago University Press. Pinker, S. (1989). Learnability and cognition: The acquisition of argument structure. Cambridge, MA: MIT Press.


Stefanowitsch, A., & Gries, St. Th. (2003). Collostructions: Investigating the interaction between words and constructions. International Journal of Corpus Linguistics, 8(2), 209–243. DOI: 10.1075/ijcl.8.2.03ste Stefanowitsch, A., & Gries, St. Th. (2005). Covarying collexemes. Corpus Linguistics and Linguistic Theory, 1(1), 1–43. DOI: 10.1515/cllt.2005.1.1.1 Van der Leek, F. (1996). The English conative construction: A compositional account. In L. Dobrin, K. Singer, & L. McNair (Eds.), CLS 32: The main session: Papers from the 32th meeting of the Chicago linguistic society (pp. 363–378). Chicago: Chicago Linguistic Society.

85

Quantifying polysemy in Cognitive Sociolinguistics Justyna A. Robinson University of Sussex

This chapter uses various statistical techniques to explore the extralinguistic grounding of individual conceptualisations of polysemous adjectives in English, such as awesome, gay, wicked. It considers the extent to which individual conceptualisations are non-random and can be related to the socio-demographic characteristics of the speaker. The experimental survey data collected from 72 speakers is analysed via hierarchical agglomerative clustering, decision tree analysis, and logistic regression analysis. The results reveal that not only individual adjectives, as indicated in Robinson (2010a), but whole groups of polysemous adjectives currently undergoing semantic change form usage patterns that can be explained by a very similar sociolinguistic distribution. This study demonstrates that employing a socio-cognitive perspective when researching polysemy is hugely advantageous. Keywords: adjectives, decision tree analyses, hierarchical agglomerative clustering, logistic regression, semantic variation, semantic change

1. Polysemy Polysemy, which is usually defined as one form that has several related1 yet distinct2 meanings, has been widely explored in various areas of linguistics (for an overview, 1. Polysemous meanings are related historically (as opposed to homonymous meanings which are not). From the methodological point of view there are different approaches as to whether a particular sense is a valid reading of a polysemous category. While certain studies include historically-related polysemous meanings in their dataset (e.g. Sweetser 1990), other studies consider only senses that are conceptually related in a given point in time (e.g. Beretta et al. 2005; Klein and Murphy 2002). Some discussion of diachronically-related senses that can be perceived as unrelated is available in Geeraerts (1997) and Blank (2003). 2. Challenges and potential solutions for determining the number and boundaries between polysemous readings are discussed in Dunbar (2001), Geeraerts (1993), Gries (2006), Hanks

88

Justyna A. Robinson

see Allan and Robinson 2012; Cuyckens and Zawada 2001; Geeraerts and Cuyckens 2007; Lewandowska-Tomaszczyk 2007; Nerlich et al. 2003; Rakova et al. 2007; Ravin and Leacock 2000; Vanhove 2008). Although polysemy was traditionally associated with the lexicon, much of the research has shown that polysemy also emerges when syntactic, morphological, and phonological usage is considered (see e.g. Brugman 1981; Taylor 1995:â•›142). It has become apparent that polysemy is not just a feature of certain words, but that it is a form of categorisation (see e.g. Taylor 1995:â•›99; Lakoff 1987:â•›12). Therefore, much of the research on polysemy has since focused on learning more about patterns of human categorisation. Some of the key observations of polysemous usage indicate that knowledge is categorised in terms of family resemblance models with less central meanings clustered around a prototype (see e.g. Geeraerts 1989, 1993; Janda 1990; Lakoff 1987:â•›379). Since the majority of these observations are drawn from the analysis of intralinguistic data only, there is little known about the extent to which the categorisation of specific linguistic events varies between speakers in the same community. For instance, can we assume that the prototypical centre of a polysemous category will be the same for each speaker in a given community? Much research on language variation indicates that linguistic usage does indeed differ between speakers in a community (Chambers et al. 2002; Coulmas 1997; Fought 2004; Labov 2001). These studies demonstrate that the way people speak may be predicted, for example, from their profession, the area in which they live, their social networks (Milroy 1980, 1987), practices they engage in (Eckert 2000) or identities they construct and adopt (Bucholtz 2011). Theoretically, Cognitive Linguistics agrees with the social grounding of linguistic variation by arguing for the experiential and perspectival nature of meaning (Geeraerts 1993:â•›60). However, little research has been done to relate linguistic usage patterns directly to extralinguistic categories and to account for the socio-cultural grounding of categorisation. Notable exceptions are represented in the literature by Geeraerts et al. (1994), Geeraerts et al. (2010), Kristiansen and Dirven (2008), Pütz et al. (2012a), Pütz et al. (2012b), Pütz et al. (2014), Reif et al. (2013), Robinson (2012a), and Robinson (2012b).

2. Scope of the study The current chapter contributes to the discussion of the extralinguistic grounding of individual conceptualisations of polysemous categories. It considers the extent to which individual conceptualisations are non-random and can be related to the socio-demographic characteristics of the speaker. The current chapter also demonstrates (2000), Kilgariff (1997), Krishnamurthy and Nicholls (2000), Lehrer (1990), and summarised in Lewandowska-Tomaszczyk (2007), and Ravin and Leacock (2000).

Quantifying polysemy in Cognitive Sociolinguistics

how various statistical techniques (i.e. hierarchical agglomerative clustering, logistic regression, and decision tree analyses) may be employed in order to enhance the way investigations of polysemy are carried out. The current chapter elaborates on my earlier study (Robinson 2010a) that demonstrated the benefits of implementing a sociolinguistic perspective in cognitive research on polysemy. In that study, I found out that the usage of innovative or conservative senses of the adjective awesome can be predicted from the socio-demographic characteristics of a speaker. For example, the use of awesome ‘terrible’ can be predicted from the speech of people of 60 years or older. Not only do these findings support cognitive linguistic understanding of polysemy but they also indicate that significant differences exist in the extent to which the different meanings of the same semantic category are salient for different speakers. Although these findings provide compelling evidence for the existence of socio-semantic usage patterns, the conclusions regarding systematic, socially-grounded usage can be only generalised as far as the adjective awesome is concerned. What remains to be verified is whether speakers’ usage of other similar words (e.g. other adjectives currently undergoing change) is structured in a similar way. The current chapter addresses this issue by investigating the usage of eight polysemous adjectives that are presently undergoing change.3 Firstly, I establish whether any meaningful patterns can be detected when the usage of several polysemous adjectives is considered simultaneously. Provided that meaningful usage patterns emerge, I then determine whether these are non-randomly related to the socio-demographic characteristics of the speakers who use these categories. In order to achieve the aims of the research, various exploratory and confirmatory statistical techniques are implemented. Thus, after introducing the data (Section 3), I summarise the aims of the hierarchical agglomerative cluster analysis (Section 4) and apply this exploratory method to analyse the dataset (Section 5). This analysis is supplemented with confirmatory analyses involving logistic regression (Section 6) and a decision tree analysis (Section 7) before the conclusions are drawn (Section 8).

3. This chapter is based on my doctoral research (Robinson 2010b). I would like to thank the University of Sheffield for funding this research project and Joan Beal, Ewa Dąbrowska, Philip Durkin, and Susan Fitzmaurice for generous advice and guidance on many aspects of this research. I would also like to thank Christopher S. Butler, Dagmar Divjak, and anonymous reviewers for comments on the earlier version of the current chapter. All other shortcomings are mine.

89

90 Justyna A. Robinson

3. Data and method Initial steps in the research follow those presented in Robinson (2010a). Eight polysemous adjectives currently undergoing change (Table 1) and five controlling adjectives have been chosen for the study. Available corpora (the British National Corpus (henceforth, BNC) and the Oxford English Corpus (henceforth, OEC)) and dictionaries (the Oxford English Dictionary (henceforth, OED)) indicate that the investigated adjectives have recently developed a distinctive meaning in British English. For a few of them, a potentially disappearing meaning has also been identified (see Table 1). In order to determine the usage patterns of these adjectives in a speech community, I carried out interviews with 72 speakers from South Yorkshire, UK. The speaker sample was equally representative of both men and women, different age groups (11–94 years old), and socio-economic backgrounds. Each of the speakers was asked a series of questions aimed at eliciting the most salient usage of polysemous adjectives. These questions followed a schema of asking for a referent that could be best described with an adjective in question, as shown in the following example: (1) Interviewer: Who or what is wicked? Participant: My mum. (referent) Interviewer: Why is your mum wicked? Participant: Because she lets me play my music as loudly as I want to. (justification for use) Each participant provided, on average, three instances of use of the investigated adjectives, which yielded more than seventeen hundred cases for analysis (excluding counts of those for controlling adjectives). The information obtained on both the referents and justification for the use of each adjective allowed me to put individual responses into groups of similar usage. Each of these usage groups was then given a sense label. This sense label was mainly Table 1.â•‡ The summary of potentially disappearing and emerging senses of the investigated adjectives Adjective

Incoming meaning

Potentially disappearing meaning

awesome chilled cool fit gay wicked solid skinny

great good good/trendy attractive lame good hard, tough ‘latte’/low fat

terrible

happy

mean


Table 2.â•‡ Example of the database structure Participant

wicked ‘good’

wicked ‘evil’

awesome ‘great’

awesome ‘terrible’

Speaker A Speaker B Speaker C Speaker D

2 3 0 0

1 0 2 3

3 2 0 1

0 0 2 3

derived on the basis of the match between the usage and the citation of senses used in the dictionaries. For instance, Example (1) above would be generalised with the meaning wicked ‘good’. Another sense group that emerged for the adjective wicked was wicked ‘evil’. Occasionally, speakers indicated that they were aware of a certain use of an adjective but they clearly distanced themselves from using this sense. In such cases, a category of a ‘reported’ sense was introduced.4 A category labelled ‘N/A’ was introduced in order to account for overlapping senses that could not be reliably assigned to any of the above groups or for other problematic answers.5 The raw frequency of use of each sense for each participant was recorded in the database. The next step in the analysis was to verify if any common usage patterns emerge across all senses used by participants. Usage patterns can be determined by identifying clusters in which two or more senses are used similarly (frequently or infrequently) by a number of speakers. Let us examine Table 2 to illustrate this. Table 2 presents a mini database that is structured in a similar way to the one used in the current study. One can observe that speakers A and B use awesome ‘great’ and wicked ‘good’ more frequently than other senses and that speakers C and D use awesome ‘impressive’ and wicked ‘evil’ more frequently than other senses of the adjectives. Thus, one may conclude that two distinct usage patterns emerge for two different groups of speakers: one for speakers A and B and another one for speakers C and D. In order to establish usage patterns in a large database, like the one used in the current study, I used the exploratory technique of hierarchical agglomerative cluster analysis (hereafter, HAC). Once semantic usage patterns are established, I consider the question of why these senses cluster together.6 Two variants are considered. Since the adjectives used 4. This separation of ‘reported’ senses from ‘non-reported’ senses is performed for practical reasons of potentially showing where in a community constraints on usage emerge. This is not to suggest that these are two different senses. 5. For instance, when the adjective gay used as a female name Gay. 6. Before one starts to analyse the visual output of a cluster analysis, it must be stressed that every HAC will always cluster elements together even if the clustering makes no sense. In other words, HAC will impose a structure on any data even if such a structure does not exist (also see Divjak and Fieller, this volume).

91

92

Justyna A. Robinson

in this study are undergoing change, one option is to assess whether senses that cluster together are similar in terms of the historical information we have about them. Do recently developed senses group separately from historically older senses? Another possibility is to consider whether speakers who share the same socio-demographic characteristics (in terms of age, gender, education, etc.) use the same combinations or clusters of senses. The findings of this analysis help to answer the question of whether there are any systematic semantic usage patterns and whether these are socially grounded. In order to find out more information about the users of the senses, I employed two confirmatory techniques: logistic regression and answer tree analysis.

4. Hierarchical agglomerative clustering There are a number of exploratory techniques used by social scientists that can help to find groups in data, such as principal component analysis, factor analysis, and different types of cluster analysis. These techniques differ in respect to their aims. Principal component analysis and factor analysis are methods for reducing the dimensionality of data (summarising the information in a complete set of variables using fewer variables), whereas cluster analysis aims to organise observations in meaningful structures. I use the exploratory technique of HAC in the current study because I am not interested in data reduction (as the dataset contains a comparatively small number of dimensions) and I intend to investigate groups in the data which are non-randomly similar. HAC belongs to a family of multivariate exploratory statistical methods (i.e. non-hypothesis-testing) for finding groups in data based on measured characteristics. HAC starts with each case in a separate cluster and then combines the clusters sequentially, reducing the number of clusters in each step until only one cluster is left. This hierarchical clustering process can be represented as a dendrogram, where each step in the clustering process is illustrated by a fork in the tree diagram. A detailed discussion of HAC as well as other types of cluster analysis is presented in Divjak and Fieller (this volume). The dendrogram in Figure 1 is an example of how senses from Table 1 cluster7 when their usage evidence for awesome ‘great’ (meaning 1), awesome ‘terrible’ (meaning 2), wicked ‘good’ (meaning 3), and wicked ‘evil’ (meaning 4) is inputted into the cluster analysis. The numbers visible on the junctions of the dendrogram represent the measurement of the distance at which clusters fuse together in the hierarchical cluster analysis. This example illustrates meanings 1 and 3 being combined at a fusion value of 2, whereas meanings 2 and 4 are combined at a fusion value of 1. The elements that are clustered earlier (represented by lower fusion values) are more closely related than 7. Distance measure used: phi-square; amalgamation strategy used: Ward.


Awesome ‘great’

2

Wicked ‘good’ Awesome ‘terrible’ 1 Wicked ‘evil’

Figure 1.â•‡ HAC of senses of presented in Table 1

elements that are clustered later (represented by higher fusion values). The cluster analysis groups senses that are used in a similar way by different people. This dendrogram indicates that different people in the research sample are mostly using the same combination of senses 1 and 3 (1+3) and 2 and 4 (2+4), rather than other combinations such as (1+4), (2+3), or (1+2+3).

5. Hierarchical agglomerative cluster analysis of collected data Having briefly outlined what constitutes HAC, I move on to discuss details of the computational steps of performing HAC on the current dataset. The analysis was performed using software called ClustanGraphics 7.05 (hereafter, Clustan).

5.1

Selection of polysemous adjectives

Table 3 presents the eight adjectives that are included in the HAC together with relevant meanings (a total of thirty-five meanings). Milligan and Cooper (1986, cited in Everitt et al. 2001:â•›179) suggest that only variables that are expected to be forming clusters should be included in the analysis. Irrelevant or masking variables should be excluded, if possible. Therefore, controlling adjectives and meanings grouped in the category ‘N/A’ are excluded from the HAC. The raw frequencies of the use of different senses are recorded in an SPSS database (following the format of the example of Table 2). Certain types of variables need transforming or standardising before the HAC is run (e.g. standardizing to z-scores). However, there is no need to standardise the variables in the current study as their scales do not differ. For more information on standardising variables for cluster analysis, see Divjak and Fieller (this volume).

5.2

Dissimilarity matrix

The next step involves generating a dissimilarity matrix which shows the distances between items. This procedure calculates either the similarities or dissimilarities (also

93


Table 3.â•‡ Adjectives and their meanings explored in the HAC Adjective Meanings awesome chilled cool fit gay skinny solid wicked

great relaxed

terrible cold

impressive calm, collected cold

reported relaxed good, calm, reported trendy collected good, trendy attractive athletic healthy suitable reported attractive lame unmanly homosexual happy reported reported reported lame homosexual happy latte/ thin showing mean tight fitting low fat skin hard of one hard dependable (person) substance (object) good evil reported good

referred to as distances or proximities), either between pairs of variables or between pairs of cases. Between-variable dissimilarities are generated using SPSS 18 and they are copied into Clustan. Distances can be measured by using different metrics according to different data types (see Divjak and Fieller, this volume). For instance, proximities in count data in the current dataset are measured with phi-square.

5.3

Amalgamation strategy

Once distances between items are calculated, an amalgamation strategy is applied. This procedure involves using one of a few available algorithms (for a summary, see Divjak and Fieller, this volume) that define how separate elements are to be clustered. In the current dataset, the algorithm Increase in Sum of Squares (sometimes called Ward)8 is chosen. This method merges the two elements whose merging least increases their sum of squared deviations from their mean. The Ward algorithm is considered to be a robust method for data classification, which is sensitive to outliers (Gries

8. Other algorithms have also been considered (single linkage, complete linkage, and average linkage). The basic operation of such methods is similar as they fuse individuals or groups of individuals who are closest or most similar. Differences arise because various methods define the distance between individuals or groups of individuals in different ways (see Everitt et al. 2001:â•›Chapter 3). Scholars agree that there is no single best amalgamation method (Everitt et al. 2001:â•›Chapter 8; Moisl 2009; and Tan et al. 2006:â•›639–642).


2007). This method also appears to have affinity with other linguistic research, e.g. Beitel et al. (2001), Divjak and Gries (2006), Gries and Stefanowitsch (2006). There is one more step to complete before we can start analysing the dendrogram. The order of the original data can also influence the amalgamation of values. In order to obtain the optimal ordering of the cases, I follow the ‘serialize procedure’. This procedure yields the best order of variables that can be obtained from the current proximity matrix.

5.4

Dendrogram

An HAC of the meanings of the adjectives resulted in the dendrogram presented in Figure 2. The vertical layout of the dendrogram in Figure 2 means that we analyse it from left to right. The HAC of the current dataset clustered senses that are used in similar ways by the same people. For example, the fact that wicked ‘good’ and awesome ‘great’ clustered together means that a number of people who used wicked ‘good’ in the interview were also likely to use awesome ‘great’. Skinny ‘thin’ and awesome ‘great’ are fused later and belong to different clusters. It does not mean that no participant exhibited the use of these two senses. Instead, it just means that people using skinny ‘thin’ were less likely to also use awesome ‘great’ in the same interview.

5.5

Best-cut

After generating a dendrogram, one needs to decide which subclusters are meaningful and therefore should be highlighted for analysis. There are many “rules of thumb” as far as the analysis of cluster levels is concerned, but I will briefly talk about two approaches to this process. First of all, one may delimit borders of individual clusters based on the structure of the dendrogram and perceived (dis)similarities between variables. This procedure involves considering whether various subclusters make sense from the point of the intuitive (dis)similarities between data and the scope of the investigation. Another possibility is to employ statistical measures to determine the best number and size of clusters in the dendrogram (also called best cut). Ideally, conclusions from introspection and statistical analysis should overlap (although this is not always the case). The last scenario is that none of the possible divisions of the dendrogram make sense in the context of a given research question (statistically and through introspection). This is always a possibility since “cluster analysis can create as well as reveal structure” (Breckenridge 2000:â•›261) (cf. Divjak and Fieller, this volume). Initial inspection of the dendrogram in Figure 2 indicates that three large clusters could be delimited (see highlighted clusters in Figure 3), the top one being more independent from the remaining two. This three-cluster solution seems sensible in the

95

96 Justyna A. Robinson TokensGayLame TokensWickedGood TokensAwesomeGreat TokensCoolGoodTrendy TokensChilledRelaxed TokensFitAttractive TokensGayReportedLame TokensSolidHardPerson TokensFitReportedAttractive TokensSolidOfOneSubstance TokensAwesomeImpressive TokensGayUnmanly TokensSkinnyShowingSkin TokensSkinnyTightFitting TokensFitAthletic TokensSkinnyThin TokensGayHomosexual TokensSolidHardObject TokensWickedEvil TokensGayReportedHappy TokensChilledCalmCollected TokensWickedReportedGood TokensCoolReported TokensSolidDependable TokensSkinnyLatte TokensFitSuitable TokensCoolCold TokensGayHappy TokensCoolCalmCollected TokensFitHealthy TokensGayReportedHomosexual TokensSkinnyMean TokensAwesomeTerrible TokensChilledCold TokensChilledReportedRelaxed

Figure 2.â•‡ Dendrogram of clustered senses

ClustanTM


TokensGayLame TokensWickedGood TokensAwesomeGreat TokensCoolGoodTrendy TokensChilledRelaxed TokensFitAttractive TokensGayReportedLame TokensSolidHardPerson TokensFitReportedAttractive TokensSolidOfOneSubtance TokensAwesomeImpressive TokensGayUnmanly TokensSkinnyShowingSkin TokensSkinnyTightFitting TokensFitAthletic TokensSkinnyThin TokensGayHomosexual TokensSolidHardObject TokensWickedEvil TokensGayReportedHappy TokensChilledCalmCollected TokensWickedReportedGood TokensCoolReported TokensSolidDependable TokensSkinnyLatte TokensFitSuitable TokensCoolCold TokensGayHappy TokensCoolCalmCollected TokensFitHealthy TokensGayReportedHomosexual TokensSkinnyMean TokensAwesomeTerrible TokensChilledCold TokensChilledReportedRelaxed

Figure 3.â•‡ Dendrogram with three and seven-cluster division

ClustanTM

97


light of the current research project. Taking into consideration diachronic information on the usage of individual meanings (cf. Table 1), one can notice that each of the three clusters groups senses that are of different historical depth. Thus, the top cluster includes novel senses, the bottom cluster largely includes senses that are considered to be disappearing, and the middle cluster mostly represents diachronically ‘middle’ senses (neither recent innovations, nor necessarily disappearing senses). Moreover, the most recent and the oldest senses are grouped into two clusters positioned at the extreme ends of the dendrogram. These visual characteristics indicate that these two clusters are substantially different from each other in terms of usage/people who use them. The statistical delineation of clusters was carried out by following the best-cut procedure. Best cut in the data was established by using a significance test (upper tail rule, cf. Mojena 1977) on the fusion values at every stage in which the clusters join together in the dendrogram. Best cut indicates the level at which the change in fusion values is significant for most groups. This is then displayed by highlighting partitions on the dendrogram (these clusters are delimited with squares on Figure 3). The partition corresponds to the largest number of clusters, which is significant at the level of 5%. The best-cut procedure indicates that the seven-cluster solution turns out to be statistically significant. The seven-cluster solution breaks Cluster 2 down into two further clusters and Cluster 3 into four further clusters, whereas Cluster 1 remains unchanged. At first sight, it is less apparent why these sub-clusters would be separated. One potential explanation could involve historical information on the usage of some of these subclusters. Therefore, one could suggest that the cluster (skinny ‘mean’ and awesome ‘terrible’) contains disappearing senses. At this point, a decision needs to be made as to the level (seven or three clusters) at which to carry out further analysis. This initial exploratory analysis shows that the three-cluster solution already seems to exhibit interesting sense groupings. Going deeper into subclusters (seven clusters) might yield more detailed, but not necessarily as relevant, information (from the point of view of the current research project) on groups of variables. Besides, one rule of thumb says that one should distinguish as few clusters as possible. In the current chapter, I present the analysis at the level of three clusters only.9

9. The analysis of clusters at the level that a researcher intuitively considers appropriate (three clusters) may lead to ignoring interesting nuances in use that can be revealed at a more detailed ‘best-cut’ level of seven clusters. Therefore, the analysis of the seven-cluster solution of the current data is presented in Robinson (2010b).

5.6


Validation of clusters

The validation of a given clustering involves a series of procedures that determine the robustness of a present solution for making predictions. Different studies suggest various ways of assessing the validity of a given clustering (see Duda et al. 2001:â•›557–559; Everitt et al. 2001: Chapter 8; Moisl and Jones 2005; and Tan et al. 2006:â•›532–555). In the current study, I examine both the internal stability and the external validity of the present cluster solution following suggestions presented by Clatworthy et al. (2005).

5.6.1 Confirmatory analysis: Internal stability This first step in the validation procedure aims at answering the question of whether a given cluster solution can be replicated on new data. In order to validate the clustering in Figure 2, I run a tree validation procedure (also called bootstrap validation), which involves a series of random trials on randomised proximities. Each trial generates a different dendrogram for the given data and the series of trials provide a mean dendrogram and confidence intervals. The validation is achieved by comparing the initial clustering with the clustering in the randomised dendrogram. The validation procedure confirms that the obtained dendrogram and the best-cut division of the dendrogram can be replicated even on randomised data. 5.6.2 Confirmatory analysis: External validity (sociolinguistic analysis) As Clatworthy et al. (2005:â•›333) point out, the internal stability of clusters is not sufficient evidence to determine the value of a cluster solution. External validity procedures are employed to determine whether a set of external predictors (i.e. variables which were not included in the clustering process) can be associated with the obtained cluster solution and whether the same cluster solution can be replicated from a new independent sample. This approach is considered to be one of the better ways to validate cluster solutions (Aldenderfer and Blashfield 1984:â•›66). In Sections 6 and 7, the external validity of the cluster solution is examined. I verify whether any information about the speakers (their age, gender, etc.) who provided the usage data for this study could explain the way in which the senses in Figure 2 are grouped. This practically means employing other statistical measures to validate the use of clusters (linguistic variables) against external variables (socio-demographic variables). In order to perform external validation, multivariate techniques (logistic regression modelling and decision tree analysis) are employed. From a statistical point of view, verifying the external stability is a necessary element of cluster analysis. From the viewpoint of the current study, this procedure can be considered as a way of testing whether any meaningful semantic usage patterns emerge from analysing sociolinguistic variation. This naturally leads to gaining more insights as to whether conceptualisations of polysemous categories are non-randomly grounded in socio-cognitive contexts.

99


5.6.3 Summary of cluster-variables Each of the three main clusters represents a usage pattern which is treated as a linguistic variable. In order to carry out statistical calculations, each of these conceptual patterns-variables needs to be presented as a binary variable to suit further statistical analyses. In order to do this, counts of senses belonging to a given cluster are added up. Following this, by the means of a visual bander10 (SPSS 18), a new variable is created that reflects the characteristics of a whole cluster. This is done through collapsing a large number of ordinal categories into a smaller set of categories, representing low and high usage of a given cluster of senses. Table 4 presents the summary of the dependent variables used, along with associated coding and category information. Table 4.â•‡ Linguistic variables (dependent variables) used to investigate their association with socio-demographic variables (independent variables) Dependent variables

Coded as

Categories

Cluster 1 Cluster 2 Cluster 3

1, 0 1, 0 1, 0

High use High use High use

Low use Low use Low use

The following independent variables are considered in the analyses: age group, gender, education, National Statistics Socio-Economic Classification score for a participant’s profession11 (hereafter, NSEC), and a postcode or a neighbourhood variable, Table 5.â•‡ Socio-demographic variables (independent variables) used to investigate their association with the use of different clusters of senses (dependent variables) Independent Coded as variables

Categories

Age group Gender NSEC Education

(1, 2, 3, 4) (1, 2) (1, 2, 3) (1, 2, 3, 4, 5)

Postcode

(1, 2, 3)

Up to 18 Male Higher Schooling prior the age of 16 Lower property prices

19–30 Female Medium Secondary school Middle property prices

31–60

Over 60

Lower College

University

Current student

Higher property prices

10. A visual bander is an SPSS tool to recode values of a variable into groups. Data frequently need to be manipulated before analyses are conducted. Data may need to be recoded, computations may need to be made, new variables may need to be created, or certain records may need to be selected (for more information, see Einspruch 2005). 11. See Office for National Statistics: Standard Occupational Classification (2000) for details.

Quantifying polysemy in Cognitive Sociolinguistics 101

which is based on property values in areas defined by the postcode of a participant’s residence. For a summary of the coding of the independent variables, see Table 5.

6. Logistic regression Having explored the data via HAC, confirmatory statistical analyses are carried out in order to validate emergent groups in the data. More specifically, I aim to assess the overall effect of socio-demographic categories on the use of particular clusters of meanings, hypothesising that the clustering solution can be explained by the categories of age, gender, NSEC, education and/or postcode value. In order to address the above-mentioned aim, a multifactor statistical model is employed, i.e., a model that considers several external factors simultaneously and measures their effect on the use of each cluster. In addition, the appropriate statistical approach needs to allow us to check for confounding variables. Socio-demographic factors may constitute such cases, i.e., education and occupation may be confounded, as people who are more educated are likely to have better jobs. Logistic regression analysis fulfils these requirements. Logistic regression can be used to test hypotheses about the relationship of several independent variables to a dichotomous dependent variable (see Hosmer and Lemeshow 1989; Kleinbaum 1994; Speelman (this volume); Tabachnick and Fidell 2001 for introductions to logistic regression). Logistic regression is increasingly being used in linguistic studies (e.g. Benki 1998; Bresnan et al. 2007; Kallel 2007; Levshina et al. (this volume); Tummers et al. 2004). The logistic regression model also allows for estimating odds ratios for each of the independent variables in the model. For instance, one may establish how many times a given meaning is more likely to be used by age group than by age group . Logistic regression also provides information on variance (the percentage to which an independent variable is explained by the dependent ones) and is used to determine the importance of independent variables. In the current study, logistic regression is performed using SPSS 18 to assess the overall effect of socio-demographic factors (independent variables) on the use of clusters. All responses (including missing values) for seventy-two participants are included in the analysis. All sociolinguistic factors are entered into the model, and then the factors are examined to verify whether they meet removal criteria using a forward stepwise method. The final model is established once no further variables are eligible for removal. The final model is then reported. The resultant fitted model informs us about significant changes in regression coefficients (expressed as B) between predictors. In cases where a stable regressive model could not be established and a final solution could not be found (even by modifying model criteria, such as increasing the


number of iterations: i.e. the series of approximations used by the logistic regression), I obtain insights into investigated variation by using multivariate statistical modelling based on decision trees (for more details, see Section 7).

6.1

Logistic regression of Cluster 1

Logistic regression analysis is performed to verify the hypothesis that the likelihood of high use of Cluster 1 can be modelled from speakers’ age, gender, NSEC, education and/or postcode (neighbourhood). Logistic regression analysis on Cluster 1 yields an unstable solution, so one cannot make predictions regarding the use of this cluster. Nevertheless, interesting findings can be obtained from examining decision trees (see Section 7.1).

6.2 Logistic regression of Cluster 2 Logistic regression analysis is performed to verify the hypothesis that the likelihood of high use of Cluster 2 can be predicted from speakers’ age, gender, NSEC, education and/or postcode (neighbourhood) value. The summary of the logistic regression analysis of the use of Cluster 2 is presented in Table 6. The final model reported includes variables that best account for the observed variation. Insignificant variables are excluded from the model. Table 6 shows the coefficients of regression Beta (hereafter, B), their standard errors, the Wald chi-square statistics, associated p-values, and odds ratios.12 The resultant fitted model indicates which independent variables are included in the final logistic model. It also informs us about significant changes in regression coefficients (B) between predictors. B determines the direction of the relationship between a given predictor and the dependent variable (the use of Cluster 2). If B is positive, the odds for the use of Cluster 2 are increased; when B is negative, then the odds are decreased; B equalling 0 leaves the odds unchanged. Explanations of the indicator variables can be found directly under each regression table. Model summary. According to the model, a high use of Cluster 2 can be modelled from speakers’ age (p = .005) and NSEC (p = .005). Age. The most significant differences of use exist between age groups and (p = .001, B = –5.752), and also between age groups and (p = .009, B = 3.407). This means that ‘middle’ age groups speak more similarly to each other. The analysis of probability measures (hereafter, P) presented in

12. For further discussion of stepwise regression, regression coefficient, iterations and the output of the logistic regression analysis in SPSS, see Brace et al. (2006) and Norušis (1999).


Table 6.â•‡ Summary of the logistic regression analysis of Cluster 2 Variables AgeGroup AgeGroup(1)a AgeGroup(2)b AgeGroup(3)c NSEC NSEC(1)d NSEC(2)e Constant a:

B

S.E.

–5.752 â•⁄–.313 â•⁄3.407

1.709 â•⁄.841 1.301

â•⁄3.905 â•⁄1.010 â•⁄–.056

1.467 â•⁄.929 â•⁄.380

Wald

df

p

Odds ratio

12.907 11.333 â•⁄â•⁄.138 â•⁄6.855 10.564 â•⁄7.085 â•⁄1.180 â•⁄â•⁄.021

3 1 1 1 2 1 1 1

.005 .001 .710 .009 .005 .008 .277 .883

â•⁄â•⁄.003 â•⁄â•⁄.732 30.179 49.672 â•⁄2.745 â•⁄â•⁄.946

change between the age group in relation to the age group

b:


c:


d:

change between the NSEC group in relation to the NSEC group <Medium>

e:

change between the NSEC group <Medium> in relation to the NSEC group

Table 7.â•‡ Probability values used for logit of Cluster 2 Age group

Age group probability

NSEC group

NSEC probability

Up to 18 19–30 31–60 Over 60

â•⁄2.5% 88.9% 91.6% 26.6%

NSEC 1 NSEC 2 NSEC 3

94.7% 26.5% 11.6%

Table 7 indicates that speakers of age groups and are most likely to be high users of Cluster 2 (P = 88.9% and 91.6% respectively). NSEC. The significant ‘jump’ in B-coefficients exists between NSEC2 and NSEC1 (p = .008, B = 3.905). Speakers who occupy higher occupations (NSEC1) are most likely to exhibit higher use of Cluster 2 (P = 94.7%), in comparison to speakers of middle and lower occupations (P (NSEC2) = 26.5%, P (NSEC3) = 11.6%). These findings are graphically presented in the form of logistic function estimate values (logit) in Figures 4 and 5. The bars with positive values (above 0) represent categories of independent variables (age group and NSEC, respectively), the occurrence of which correspond to a probability of the use of Cluster 2. In the figures, the taller the bar above 0, the higher the probability of high use of Cluster 2. The bars with negative values (below 0) represent categories of independent variables, the occurrence of which correspond to a probability of the use of Cluster 2. The lower the bar, the higher the probability of ‘not high’ use of Cluster 2. In the logistic regression analysis, the predictive and explanatory power of the fitted model needs to be assessed. In order to validate predicted probabilities, the


3 2 1 0 –1 –2 –3 –4

upto 18

19–30

31–60

over 60

Figure 4.â•‡ Logistic function estimate values for age group in Cluster 2

4 3 2 1 0 –1 –2 –3

NSEC1

NSEC2

NSEC3

Figure 5.â•‡ Logistic function estimate values for NSEC in Cluster 2

c-statistic is used (see Peng et al. 2002:â•›6). The c-statistic analyses the proportion of observed to initially-predicted probabilities of occurrences of Cluster 2. In the case of Cluster 2, the fitted model (one that includes socio-demographic variables) achieves a success rate of 80.6%, which is an improvement over the intercept model (51.4%), i.e., a model that does not include any of the socio-demographic variables to account for the observed variation, but includes a constant term only and the model that does not take into consideration NSEC (73.6%). The explanatory power of the calculated model refers to how effectively it fits the actual data for estimating the outcome variable (Moss et al. 2003:â•›925). This could be assessed by a number of ‘goodness-of-fit’ measures. –2 Log Likelihood (hereafter, –2LL) indicates the overall fit of the model. It reflects the significance of the unexplained variance in the model. Its lowering values indicate improvement of a model fit (increasing the likelihood of the observed results). R-square measurements (Cox and Snell, Nagelkerke tests) indicate how much variation the model actually explains. Sometimes these measures may yield different results (for further discussion, see


Field 2005:â•›239–240). The Hosmer and Lemeshow test is another measure that is considered by some researchers to be a more accurate measure for assessing the goodness-of-fit of the model (Peng et al. 2002:â•›6). It tells you how closely the observed and predicted probabilities match and insignificant results in the Hosmer and Lemeshow test signify a model that fits the data well. In the case of Cluster 2, –2LL (53.740) and an insignificant Hosmer–Lemeshow test indicate that the model fits the data well and is more adequate for explaining variation than models that do not consider socio-demographic factors. R-square measurements (Cox and Snell = .472, Nagelkerke = .630) indicate that the variation in the outcome variable is explained moderately well by the logistic regression model. Logistic regression analysis evidences that the use of Cluster 2 can be satisfactorily modelled from the age and NSEC of speakers, although age group has a more significant overall effect on the use of the given variable than NSEC. Logistic regression analysis confirms our hypothesis and validates the external stability of Cluster 2.

6.3

Logistic regression analysis of Cluster 3

Logistic regression analysis is performed to verify the hypothesis that the likelihood of high use of Cluster 3 is modelled from speakers’ age, gender, NSEC, education and/or postcode (neighbourhood). Logistic regression on this cluster yields an unstable solution, so predictions regarding the use of this cluster cannot be made. Nevertheless, interesting findings can be observed from examining decision trees (see Section 7.3).

7. Decision tree analysis In cases where the logistic regression model cannot be established, I use the results of another multivariate technique. A decision tree analysis is a technique based on separating cases into segments that are as different from each other as possible. For instance, with a decision tree analysis one can easily detect segments and patterns such as ‘female bridge players with at least 5 years’ experience are likely to win a game’, or ‘students who miss more than 40 days of school a year are twice as likely to drop out’. This procedure uses appropriate algorithms that predict the class (belonging) of a dependent variable from the values of predictor variables. The choice of algorithms largely depends on the type of data. The most appropriate algorithm chosen for our analysis is Chi-square Automatic Interaction Detection (hereafter, CHAID).13 This is

13. Other algorithms have also been considered: CandRT, QUEST (for a summary, see SPSS White Paper, Answer Tree Algorithm Summary, 2005). All p-values in the CHAID algorithm


a non-parametric stepwise regression procedure that produces splits until it gets a significant p-value for each split. The CHAID algorithm (available via Answer Tree 3.0) is used to examine factors predicting the use of senses in a cluster. It supplements logistic regression analysis, especially in cases where a logistic regression model cannot be determined. However, it does not mean that both techniques are different ways of answering the same questions. The CHAID algorithm identifies groups of speakers that use similar meanings in a similar way. Logistic regression estimates an overall effect of an independent variable (i.e. age, gender, or social class of speakers) on the use of a particular meaning cluster. Therefore, the two methods are two different ways of looking at the same data. Decision trees have been widely used in database marketing research (Chaturvedi and Green 1995:â•›245; Magidson 1994; and Rao and Steckel 1995) and in clinical science (Barrio et al. 2006:â•›595; Boscarino et al. 2003:â•›303; and Saltini et al. 2004:â•›737) for performing classification or segmentation. However, the use of decision trees in linguistics is rare (but cf. Heylen 2005; Robinson 2012a; Schmid 2010).

7.1

Decision tree of Cluster 1

I run the decision tree analysis in order to verify the importance of socio-demographic factors (summarised in Table 5) in predicting high use of Cluster 1. More specifically, this analysis shows whether there are any significant socio-demographic groups (e.g. age) or subgroups (age by gender) that use the senses in Cluster 1. The output of the analysis is presented in the form of a decision tree. The decision tree presenting a multivariate analysis of Cluster 1 is presented in Figure 6. The output presents several levels of significant splits (here two levels). Each split is based on the rule of the lowest p-value. In other words, if two splits are significant, the actual split in the decision tree follows the split according to the independent variable for which the p-value is the lowest. In the case of a tie, the rule with the higher chi-square value is listed first. In the case of another tie, the rule with lower degrees of freedom (hereafter, df) is listed first. Both chi-square and df values are displayed in the tree for each split. In the decision tree in Figure 6, the square at the top (Node 0) represents the characteristics of a variable to be analysed. There are two categories in this variable: low uses of Cluster 1 and high uses of Cluster 1. There are 39 cases of the former and 33 cases of the latter, accounting for, respectively, 54.17% and 45.83 % of the variable. These frequencies are visually presented in the form of bars at the bottom of each square (node). Low use of the cluster is represented by a darker shade of grey, whereas a high use of the cluster is represented by a lighter shade of grey. analysis were adjusted for multiple comparisons using the Bonferroni method (SPSS Answer Tree 3.0).


WARDCluster1 (Binned) Node 0 Category % low 54.17 high 45.83 Total (100.00)

n 39 33 72

AGEGROUP Adj. P-value=0.000, Chi-square=45.9182, df=3 upto 18 Node 1 Category % low 5.56 high 94.44 Total (25.00)

19–30 n 1 17 18

Node 2 Category % low 38.89 high 61.11 Total (25.00)

31–60

over 60


n 7 11 18


n 13 5 18

n 18 0 18

NSEC Adj. P-value=0.0412, Chi-square=6.0710, df=1 3;2 Node 5 Category % low 54.55 high 45.45 Total (15.28)

1 n 6 5 11


n 7 0 7

Figure 6.â•‡ Decision tree of Cluster 1

The first significant split takes into consideration the age group to which speakers belong. The statistics for the significance of this spilt are described just above the split. The statistics summary indicates that age group is the most significant predictor of using Cluster 1 (p < .001, χ2 = 45.91, df = 3). Speakers who use this cluster most frequently belong to the two youngest generations; age group exhibits high use in 94.44% of cases and age group does so in 61.11% of cases. The results are also presented graphically: light grey bars at the bottom of Nodes 1 and 2 (squares in Figure 6 representing age groups and , respectively) indicate the large frequency of the category representing high usage of Cluster 1. Nodes 3 and 4 (representing age groups and , respectively) indicate the high proportion of speakers who exhibit a low usage of the senses grouped in Cluster 1. This is represented graphically by the light grey bars at the bottom of nodes 3 and 4 (Figure 6). The second significant split is based on the NSEC of speakers (p < .0412, χ2 = 6.07, df = 1). The multivariate analysis combined together NSEC2 and NSEC3 (medium


and lower occupations) and separated them from NSEC1 (higher occupations). All speakers who are years old and occupy higher professional positions indicate low usage of Cluster 1 (100% of low usage responses). The risk estimate for this decision tree is 0.18, which indicates that if I use the decision rule based on the current decision tree I correctly classify 82% (100% minus 18%) of cases (the calculations of risk are not presented on the decision tree). The multivariate analysis via Answer Tree 3.0 externally validates the use of Cluster 1, showing that the age of participants and, in the case of middle age speakers, their occupation, predicts high use of the senses grouped in Cluster 1.

7.2

Decision tree analysis of Cluster 2

The overall effect of socio-demographic factors in modelling high use of Cluster 2 is established using logistic regression analysis. Decision tree analysis is run in order to verify whether I could obtain any further insights into the use of Cluster 2 in relation to socio-demographic dimensions, especially in the context of determining significant subgroups of use. Figure 7 illustrates the relative importance of socio-demographic factors in predicting the use of Cluster 2 meanings. The most important factor in predicting usage is the age of participants (p = .0002, χ2 = 20.8, df = 2). Speakers of age are grouped together as the highest users of Cluster 2. Moreover, multivariate analysis shows that in every age group, speakers living in the most affluent neighbourhoods (Postcode 3, i.e. above £142,795) or holding professional positions (NSEC1) most frequently exhibit high use of the meanings in the cluster (p < .05). Additionally, there is a distinction at the level of gender (p = .01) in the youngest age group of speakers living in the most affluent areas. In this group males are all ‘high users’ of Cluster 2 whereas females are all ‘low users’ of that cluster. Risk estimate indicates that 85% of variation can be correctly classified when applying the decision rule based on the current decision tree. To conclude, decision tree analysis validates the findings of logistic regression and provides additional evidence for the external stability of Cluster 2.

7.3

Decision tree of Cluster 3

Decision tree analysis is run in order to assess the relative importance of socio-demographic factors in predicting high use of Cluster 3 (see Figure 8). The multivariate statistical analysis shows that age group is the most significant predictor of using Cluster 3 (p < .001, χ2 = 44.43, df = 2). Speakers who use this cluster most frequently belong to the two oldest generations. All of the speakers are high users of the cluster. Speakers of age group use this cluster in 77.78% of

Quantifying polysemy in Cognitive Sociolinguistics 109 WARDCluster2 (Binned) Node 0 Category % low user 54.17 high user 45.83 Total (100.00)

n 39 33 72

AGEGROUP Adj. P-value=0.002, Chi-square=20.6041, df=2 upto 18

19–30; 31–60

Node 1 Category % n low user 88.89 16 high user 11.11 2 Total (25.00) 18

POSTCODE Adj. P-value=0.0466, Chi-square=5.8557, df=1 ;below £117,543 Node 4 Category % n low user 100.00 13 high user 0.00 0 Total (18.06) 13

3:2 Node 6 Category % low user 38.46 high user 61.54 Total (36.11)

n 3 2 5

GENDER Adj. P-value=0.0105, Chi-square=6.5399, df=1 male

female n 0 2 2

Figure 7. Decision tree of Cluster 2


NSEC Adj. P-value=0.0153, Chi-square=7.8391, df=1

above £142,795 Node 5 Category % low user 60.00 high user 40.00 Total (6.94)

Node 10 Category % low user 0.00 high user 100.00 Total (2.78)

over 60



1 n 10 16 26


NSEC Adj. P-value=0.0020, Chi-square=11.6157, df=1 3:2 Node 8 Category % n low user 90.91 10 high user 9.09 1 Total (15.28) 11

1 Node 9 Category % n low user 14.29 1 high user 85.71 6 Total (9.72) 7


WARDCluster3 (Binned) Node 0 Category % low 45.83 high 54.17 Total (100.00)

n 33 39 72

AGEGROUP Adj. P-value=0.0000, Chi-square=44.4320, df=2 31–60

upto 18;19–30 Node 1 Category % low 80.56 high 19.44 Total (50.00)

over 60


n 29 7 36

n 4 14 18

Node 3 Category % n low 0.00 0 high 100.00 18 Total (25.00) 18

POSTCODE Adj. P-value=0.0200, Chi-square=7.3649, df=1 below £117,543;above £142,795



Figure 8.â•‡ Decision tree of Cluster 3

cases (especially when they live in the highest and lowest postcodes (p = .02, χ2 = 7.36, df = 1)). Risk estimate indicates that 87.5% of variation can be correctly classified when applying the decision rule based on the current decision tree. Decision tree analysis confirms the hypothesis and provides evidence for the external stability of Cluster 3. Overall, the external validity of the cluster solution has been confirmed. The use of each of the three clusters can be predicted from the language of speakers who differ in socio-demographic terms.

8. Summary and discussion of results Having carried out HAC on the usage data, it has become apparent that each of the main three clusters can be most satisfactorily predicted from the speech of different generations (see summary in Table 8). The use of Cluster 1 (innovative speech) is best predicted from the speech of the youngest speakers, the use of Cluster 3 (historically older senses) is best predicted from the speech of older speakers, and the use of


Table 8.â•‡ Summary of the exploratory and confirmatory analyses of the use of polysemous adjectives Cluster 1 Cluster 2 Cluster 3

Exploratory analysis

Confirmatory analysis

More recent senses Middle senses Oldest senses

Younger generations Middle generations especially NSEC1 Older generations

Cluster 2 (historically neither old nor recent senses) from the speech of middle age groups. Additionally, the results of the statistical analysis of Cluster 2 show that speakers in professional occupations are mostly ‘high users’ of the senses grouped here. The HAC reveals that sociolinguistically meaningful semantic usage patterns emerge when usage evidence from several polysemous words is considered. It becomes apparent that the use of a selected group of senses can be most typical for a socio-demographically defined group of speakers. In other words, there are speakers for whom the same senses (e.g. fit ‘attractive’, gay ‘lame’, and wicked ‘good’) are the most salient readings of polysemous categories (fit, gay, and wicked). This finding suggests that not only individual words, such as awesome, but whole groups of polyÂ� semous adjectives currently undergoing semantic change form usage patterns that can be explained by a very similar sociolinguistic distribution. This study validates Robinson (2010a) by providing further evidence for the social grounding of polysemous conceptualisations and suggests that employing a socio-cognitive perspective in linguistic research is clearly advantageous. This study also showcases the benefits of engaging various statistical techniques to explore lexical meaning.14

References Aldenderfer, M. S., & Blashfield, R. (1984). Cluster analysis. Sage Publications: Newbury Park. Allan, K. & Robinson, J. A. (Eds.). (2012). Current methods in historical semantics. Berlin & Boston: Mouton de Gruyter. Barrio, G., De La Fuente, L., Toro, C., Vicente, T. M., Vallejo, F., & Silva, T. (2006). Prevalence of HIV infection among young adult injecting and non-injecting heroin users in Spain in the era of harm reduction programmes: Gender differences and other related factors. Epidemiology and Infection, 135(4), 592–603. DOI: 10.1017/S0950268806007266

14. Although in this chapter I carry out statistical analyses by employing software packages such as ClustanGraphics 7.05, SPSS 18, and Answer Tree 3.0, the same analysis can also be performed with the help of R. The cluster analysis can also be performed with more recent versions of SPSS.

111


Beitel, D. A., Gibbs, R., & Sanders, P. (2001). The embodied approach to the polysemy of the spatial preposition on. In H. Cuyckens, & B. Zawada (Eds.), Polysemy in Cognitive Linguistics (pp. 241–260). Amsterdam: Benjamins. Benki, J. R. (1998). Evidence for phonological categories from speech perception. Unpublished PhD dissertation, University of Massachusetts. Beretta, A., Fiorentino, R., & Poeppel, D. (2005). The effects of homonymy and polysemy on lexical access: An MEG study. Cognitive Brain Research, 24(1), 57–65. DOI: 10.1016/j.cogbrainres.2004.12.006 Blank, A. (2003). Polysemy in the lexicon and in discourse. In B. Nerlich, Z. Todd, V. Herman, & D. D. Clarke (Eds.), Polysemy: Flexible patterns of meaning in mind and language (pp. 267–293). Berlin & New York: Mouton de Gruyter. Boscarino, J. A., Galea, S., Ahern, J., Resnick, H., & Vlahov, D. (2003). Psychiatric medication use among Manhattan residents following the World Trade Center disaster. Journal of Traumatic Stress, 16(3), 301–306. DOI: 10.1023/A:1023708410513 Brace, N., Kemp, R., & Snelgar, R. (2006). SPSS for psychologists, 3rd edition. New York: Palgrave Macmillan. Breckenridge, J. N. (2000). Validating cluster analysis: Consistent replication and symmetry. Multivariate Behavioural Research, 35(2), 261–286. DOI: 10.1207/S15327906MBR3502_5 Bresnan, J., Cueni, A., Nikitina, T., & Baayen, R. H. (2007). Predicting the dative alternation. In G. Boume, I. Kraemer, & J. Zwarts (Eds.), Cognitive foundations of interpretation (pp. 69– 94). Amsterdam: Royal Netherlands Academy of Science. The British National Corpus, Version 3 (BNC XML Edition). Accessed via http://www.sketchengine.co.uk [March 2008]. Brugman, C. (1981). Story of over. MA thesis, University of California, Berkley. Bucholtz, M. (2011). White kids: Language, race and styles of youth identity. Cambridge: Cambridge University Press. Chambers, J. K., Trudgill, P., & Schilling-Estes, N. (2002). The handbook of language variation and change. Oxford: Blackwell. Chaturvedi, A., & Green, P. E. (1995). Software review: SPSS for Windows, CHAID 6.0. Journal of Marketing Research, 32, 245–254. DOI: 10.2307/3152056 Clatworthy, J., Buick, D., Hankins, M., Weinman, J., & Horne, R. (2005). The use and reporting of cluster analysis in health psychology: A review. British Journal of Health Psychology, 10(3), 329–358. DOI: 10.1348/135910705X25697 Coulmas, F. (Ed.). (1997). The handbook of sociolinguistics. Oxford: Blackwell. Cuyckens, H., & Zawada, B. (Eds.). (2001). Polysemy in Cognitive Linguistics. Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/cilt.177 Divjak, D., & Gries, St. Th. (2006). Ways of trying in Russian: Clustering behavioral profiles. Corpus Linguistics and Linguistic Theory, 2(1), 23–60. DOI: 10.1515/CLLT.2006.002 Duda, R., Hart, P., & Stork, D. (2001). Pattern classification. New York: Wiley Interscience. Dunbar, G. (2001). Towards a cognitive analysis of polysemy, ambiguity, and vagueness. Cognitive Linguistics, 12(1), 1–14. DOI: 10.1515/cogl.12.1.1 Eckert, P. (2000). Linguistic variation as social practice. Oxford: Blackwell. Einspruch, E. L. (2005). An introductory guide to SPSS® for Windows®, 2nd edition. Thousand Oaks, London & New Delhi: Sage Publications. Everitt, B. S., Landau, S., & Leese, M. (2001). Cluster analysis. London: Arnold. Field, A. (2005). Discovering statistics using SPSS. London: Sage. Fought, C. (2004). Sociolinguistic variation: Critical reflections. Oxford: Oxford University Press.


Geeraerts, D. (1989). Prospects and problems of prototype theory. Linguistics, 27(4), 587–612. DOI: 10.1515/ling.1989.27.4.587 Geeraerts, D. (1993). Vagueness’s puzzles, polysemy’s vagaries. Cognitive Linguistics, 4(3), 223– 272. DOI: 10.1515/cogl.1993.4.3.223 Geeraerts, D. (1997). Diachronic prototype semantics: A contribution to historical-lexicology. Oxford: Clarendon Press. Geeraerts, D., Grondelaers, S., & Bakema, P. (1994). The structure of lexical variation. Meaning, naming, and context. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110873061 Geeraerts, D., & Cuyckens, H. (Eds.). (2007). Oxford handbook of Cognitive Linguistics. Oxford: Oxford University Press. Geeraerts, D., Kristiansen, G., & Piersman, Y. (Eds.). (2010). Advances in Cognitive Sociolinguistics. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110226461 Gries, St. Th. (2006). Corpus-based methods and Cognitive Semantics: The many meanings of to run. In St. Th. Gries, & A. Stefanowitsch (Eds.), Corpora in Cognitive Linguistics: Corpus-based approaches to syntax and lexis (pp. 57–99). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110197709 Gries, St. Th. (2007). Cluster analysis: A practical introduction with R (for Windows). Paper presented at Departmental Research Seminar, University of Sheffield. Gries, St. Th., & Stefanowitsch, A. (2006). Cluster analysis and the identification of collexeme classes. In S. Rice, & J. Newman (Eds.), Empirical and experimental methods in cognitive/ functional research. Stanford: CSLI. Hanks, P. (2000). Do word meanings exist? Computers and the Humanities, 34(1–2), 205–215. DOI: 10.1023/A:1002471322828 Heylen, K. (2005). A quantitative corpus study of German word order variation. In S. Kepser, & M. Reis (Eds.), Linguistic evidence: Empirical, theoretical and computational perspectives (pp. 241–264). Berlin: Mouton de Gruyter. DOI: 10.1515/9783110197549.241 Hosmer, D. W., & Lemeshow, S. (1989). Applied logistic regression. New York: Wiley. Janda, L. (1990). The radial network of a grammatical category – its genesis and dynamic structure. Cognitive Linguistics, 1(3), 269–288. DOI: 10.1515/cogl.1990.1.3.269 Kallel, A. (2007). The loss of negative concord in Standard English: Internal factors. Language Variation and Change, 19(1), 27–49. DOI: 10.1017/S0954394507070019 Kilgariff, A. (1997). I don’t believe in word senses. Computers and the Humanities, 31(2), 91– 113. DOI: 10.1023/A:1000583911091 Klein, D., & Murphy, G. L. (2002). Paper has been my ruin: Conceptual relations of polysemous senses. Journal of Memory and Language, 47(4), 548–570. DOI: 10.1016/S0749-596X (02)00020-7 Kleinbaum, D. G. (1994). Logistic regression: A self-learning text. New York: Springer-Verlag. DOI: 10.1007/978-1-4757-4108-7 Krishnamurthy, R., & Nicholls, D. (2000). Peeling an onion: A lexicographer’s experience of manual sense-tagging. Computers and the Humanities, 34(1–2), 85–97. DOI: 10.1023/A:1002407003264 Kristiansen, G., & Dirven, R. (Eds.). (2008). Cognitive Sociolinguistics: Language variation, cultural models, social systems. Berlin: Mouton de Gruyter. DOI: 10.1515/9783110199154 Labov, W. (2001). Principles of linguistic change: Social factors. Oxford: Blackwell. Lakoff, G. (1987). Women, fire, and dangerous things: What categories reveal about the mind. Chicago: Chicago University Press. DOI: 10.7208/chicago/9780226471013.001.0001


Lehrer, A. (1990). Polysemy, conventionality, and the structure of the lexicon. Cognitive Linguistics, 1(2), 207–246. DOI: 10.1515/cogl.1990.1.2.207 Lewandowska-Tomaszczyk, B. (2007). Polysemy, prototypes and radial categories. In D. Geeraerts, & H. Cuyckens (Eds.), Oxford handbook of Cognitive Linguistics (pp. 139– 169). Oxford: Oxford University Press. Magidson, J. (1994). The CHAID approach to segmentation modelling. In R. P. Bagozzi (Ed.), Advanced methods of marketing research (pp. 118–159). Cambridge, MA.: Blackwell. Milligan, G. W., & Cooper, M. C. (1986). A study of comparability of external criteria for hierarchical cluster analysis. Multivariate Behavioural Research, 21(4), 41–58. DOI: 10.1207/s15327906mbr2104_5 Milroy, L. (1980). Language and social networks. Oxford: Blackwell. Milroy, L. (1987). Observing and analysing natural language: A critical account of sociolinguistic method. Oxford & New York: Blackwell. Moisl, H. L. (2009). Exploratory multivariate analysis. In A. Lüdeling, & M. Kytö (Eds.), Corpus linguistics: An international handbook (pp. 874–898). Berlin: Mouton de Gruyter. DOI: 10.1515/9783110213881.2.874 Moisl, H. L., & Jones, V. (2005). Cluster analysis of the Newcastle Electronic Corpus of Tyneside English: A comparison of methods. Literary and Linguistic Computing, 20(1), 125–146. DOI: 10.1093/llc/fqi026 Mojena, R. (1977). Hierarchical grouping methods and stopping rules: An evaluation. Computer Journal, 20(4), 359–363. DOI: 10.1093/comjnl/20.4.359 Moss, M., Wellman, D. A., & Cotsonis, G. A. (2003). An appraisal of multivariable logistic models in the pulmonary and critical care literature. Chest, 123(3), 923–928. DOI: 10.1378/chest.123.3.923 Nerlich, B., Todd, Z., & Clarke, D. D. (2003). The acquisition of get between four and ten years. In B. Nerlich, Z. Todd, V. Herman, & D. Clarke (Eds.), Polysemy: Flexible patterns of meaning in mind and language (pp. 333–357). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110895698 Nerlich, B., Todd Z., Herman, V., & Clarke, D. D. (2003). Polysemy: Flexible patterns of meaning in mind and language. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110895698 Norušis, M. J. (1999). SPSS regression models 10.0. Chicago: SPSS Inc. Office for National Statistics.(2000). Standard Occupational Classification. Volume 2. Coding Index. London: The Stationery Office. The Oxford English Corpus. Accessed via http://www.sketchengine.co.uk [March 2008]. Oxford English Dictionary Online. Oxford University Press. Accessed via http://www.oed.com [March 2008]. Peng, C.-Y. J., Lee, K. L., & Ingersoll, G. M. (2002). An introduction to logistic regression analysis and reporting. Journal of Educational Research, 96(1), 3–14. DOI: 10.1080/00220670209598786 Pütz, M., Robinson, J. A., & Reif, M. (Eds.). (2012a). Cognitive Sociolinguistics: Variation in cognition and language use. Special issue of Review of Cognitive Linguistics, 10(2). DOI: 10.1075/rcl.10.2.01int Pütz, M, Robinson, J. A., & Reif, M. (2012b). The emergence of Cognitive Sociolinguistics: An introduction. Annual Review of Cognitive Linguistics, 10(2), 241–263. DOI: 10.1075/rcl.10.2.01int


Pütz, M., Robinson, J. A., & Reif, M. (Eds.). (2014). Cognitive Sociolinguistics. Social and cultural variation in cognition and language use. Benjamins Current Topics, 59. Amsterdam/ Philadelphia: John Benjamins. Rakova, M., PethÅ‚, G., & Rákosi, C. (Eds.). (2007). The cognitive basis of polysemy: New sources of evidence for theories of word meaning. Frankfurt/Main: Peter Lang. Rao, V. R., & Steckel, J. H. (1995). Selecting, evaluating, and updating prospects in direct mail marketing. Journal of Direct Marketing, 9(20), 20–31. DOI: 10.1002/dir.4000090205 Ravin, Y., & Leacock, C. (Eds.). (2000). Polysemy: Theoretical and computational approaches. Oxford: Oxford University Press. Reif, M., Robinson, J. A., & Pütz, M. (Eds.). (2013). Variation in language and language use: Sociolinguistic, socio-cultural and cognitive perspectives. Frankfurt/Main: Peter Lang. Robinson, J. A. (2010a). Awesome insights into semantic variation. In D. Geeraerts, G. Kristiansen, & Y. Piersman (Eds.), Advances in Cognitive Sociolinguistics (pp. 85–109). Berlin & New York: Mouton de Gruyter. Robinson, J. A. (2010b). Semantic variation and change in present-day English. Unpublished PhD dissertation, University of Sheffield. Available via http://etheses.whiterose.ac.uk/2232/ Robinson, J. A. (2012a). A sociolinguistic perspective on semantic change. In K. Allan, & J. A. Robinson (Eds.), Current methods in historical semantics (pp. 191–231). Berlin & Boston: Mouton de Gruyter. Robinson, J. A. (2012b). A gay paper: Why should sociolinguistics bother with semantics? English Today, 28(4), 38–54. DOI: 10.1017/S0266078412000399 Saltini, A., Mazzi, M. A., Del Piccolo, L., & Zimmermann, C. (2004). Decisional strategies for the attribution of emotional distress in primary care. Psychological Medicine, 34(4), 729– 739. DOI: 10.1017/S0033291703001260 Schmid, H. J. (2010). Decision trees. In A. Clark, C. Fox, & S. Lappin (Eds.), The handbook of computational linguistics and natural language processing (pp. 180–196). Oxford: Blackwell. DOI: 10.1002/9781444324044.ch7 SPSS (2005). Answer tree algorithm summary; SPSS white paper series. Sweetser, E. (1990). From etymology to pragmatics: Metaphorical and cultural aspects of semantic structure. Cambridge: Cambridge University Press. DOI: 10.1017/CBO9780511620904 Tabachnick, B. G., & Fidell, L. S. (2001). Using multivariate statistics. Boston: Allyn and Bacon. Tan, P.-N., Steinbach, M., & Kumar, V. (2006). Introduction to data mining. Boston: Pearson Addison-Wesley. Taylor, J. R. (1995). Linguistic categorization: Prototypes in linguistic theory. Oxford: Oxford University Press. Tummers, J., Speelman, D., & Geeraerts, D. (2004). Quantifying semantic effects: The impact of lexical collocations on the inflectional variation of Dutch attributive adjectives. In G. Purnelle, C. Fairon, & A. Dister (Eds.), Le poids des Mots: Actes des 7emes journées internationales d’analyse statistique des données textuelles (pp. 1079–1088). Louvain-la-Neuve: Presses Universitaires de Louvain. Vanhove, M. (Ed.). (2008). From polysemy to semantic change: Towards a typology of lexical semantic associations. Amsterdam: John Benjamins. DOI: 10.1075/slcs.106

The many uses of run Corpus methods and Socio-Cognitive Semantics Dylan Glynn


Multifactorial usage-feature analysis (profile-based approach) has been successfully applied to polysemy research (Gries 2006; Glynn 2009, 2010). This chapter represents a repeat analysis of Gries (2006). The study has three aims: (i) to verify the results of the previous study; (ii) to identify limitations in the application of the statistical technique employed (hierarchical cluster analysis) in the previous study; and (iii) to demonstrate the need to account for sociolinguistic dimensions in polysemy research. The study is based on a sample of 500 occurrences of the lexeme to run, extracted in even proportions from British English and American English and from online personal journals (blogs) and conversations (American National Corpus and British National Corpus). Keywords: cluster analysis, Cognitive Semantics, corpus linguistics, polysemy, multifactorial usage-feature analysis, sociolinguistics

1. Introduction Gries’ (2006) study ‘Corpus-based methods and cognitive semantics: The many senses of to run’ counts amongst the most influential contributions to the description of polysemy in Cognitive Linguistics. It is important not because its methodology is original, nor because it is complete or extensive, nor even because of the theoretical claims it makes, but because it simply and overtly shows how corpus-driven methods can be applied to the study of polysemy. Its contribution is the combination of theory and method: two pieces of a puzzle that establish the foundations of a theoretically and empirically coherent approach to the description of semasiological structure. The current study does not challenge the theory, the method, or the results of Gries (2006), but seeks to refine each. The chapter divides into two parts. The first part considers three theoretical issues. Firstly, how can current corpus-driven methods of semantic analysis inform the description of prototype effects on conceptual

118 Dylan Glynn

structure? Secondly, how is the notion of lexical sense operationalised in multifactorial usage-feature analysis? Thirdly, it is argued that the study of prototype structured lexical senses must also integrate the social dimensions of language for descriptive adequacy. The second part of the chapter takes the form of a case study that repeats the semasiological analysis of run presented in Gries (2006). This case study treats three issues. Firstly, Gries’ results are largely confirmed, but it is shown that cluster analysis, the statistical method employed, produces unstable representations. It is argued that its use needs further research before it can be considered reliable. Secondly, correspondence analysis, an alternative statistical technique, is introduced. By adding a sociolinguistic dimension to the analysis, it is shown that this statistical technique is capable of representing a more complex semasiological structure. Thirdly, with the use of confirmatory statistical modelling, it is demonstrated that in order to obtain descriptive adequacy, semasiological analysis must account for sociolinguistic structure.

2. Usage-based Cognitive Semantics 2.1

Corpus-driven radial network analysis

How can quantitative corpus-driven analysis inform our understanding of polysemy and prototype structures in lexical semantics? In order to answer this question, we need to identify the aims of the endeavour. Lakoff (1990) presented the commitment to empiricism and inductive research as the gold standard of Cognitive Semantics. Radial network analysis (Lakoff 1987) and Frame Semantics (Fillmore 1985) are analytical models designed to offer an empirical means for describing meaning structure, assuming both prototype effects and encyclopaedic semantics.1 The research of the era demonstrated that: i. linguistic semantics, in the strict sense, cannot adequately account for meaning structure in language – instead, it demonstrated the need for ‘encyclopaedic semantics’; ii. necessary and sufficient conditions cannot adequately determine socio-conceptual categories – instead, it demonstrated the need for ‘prototype effects’. Although prototype category theory and the radial network analysis that employs it (Lakoff 1987) represented a necessary and substantial step toward an empirical 1. Lakoff (1987) was an early protagonist of both the theory of prototype categorization and the model of radial network analysis. Prototype theory was developed and refined by Geeraerts (1989, 1993, 1997, 2000), Taylor (1989), and Kleiber (1990). Radial network analysis was developed and formalised by, especially, Rudzka-Ostyn (1989), Cuyckens (1993), and Janda (1993).

The many uses of run 119

Cognitive Semantics, it did not quite attain that goal. Theoretically and empirically, the shortcomings are well established (Geeraerts 1993; Sandra and Rice 1995). Methodologically, radial network analysis often continued the Structuralist and Generativist tradition: i. analytically, radial network analysis assumed this structure to take the form of discrete senses or ‘nodes’; ii. empirically, radial network analysis employed introspection to determine semasiological structure. Seen from this perspective, the Cognitive Semantics era of radial network analysis made the first important steps towards empiricism, but ultimately fell short of the mark. We will now briefly consider the shortcomings of radial network analysis and then move on to consider how corpus-driven methods, especially usage-feature analysis (profile-based) corpus methods (Geeraerts et al. 1994; Gries 2003), resolve these shortcomings. There is no need to re-cover well-trodden ground. Let us assume the theoretical models of encyclopaedic semantics (Fillmore 1985; Lakoff 1987) and prototype categorisation (Rosch 1975; Lakoff 1987). The first replaces truth-conditional semantics with world knowledge. The second replaces necessary and sufficient conditions with prototype structure. These two criteria for semasiological structure accepted, we can focus on the two problems: (i) the analytical assumption of discrete senses and (ii) the methodological technique of introspection. Firstly, the assumption of discrete senses is something that is intuitively attractive. Indeed, like it is obvious that the world is flat, it would seem obvious that words have meanings and that we choose between those meanings in communication. Understood in these terms, senses are reified as discrete units. This naïve operationalisation may aid in language learning, dictionary writing and, typically, only comes amiss in inter-personal disputes. However, the evidence for discrete lexical senses is as naïvely sound as the horizon is evidence for the flatness of the world. At a broader conceptual level, Fuzzy Set Theory, a distinct yet related part of Prototype Set Theory, disposes of the notion of discrete conceptual categories. If we do not suppose that the concept associated with a lexeme is discrete, why then do we continue to assume that the sub-categories of lexical senses are discrete? Lakoff ’s (1987) study of over identifies a list of usage-features in terms of minimal perceptual distinctions expressed as image schemata. Yet rather than seeing meaning construction as the relative correlation of schema features, Lakoff continues to hunt for ‘senses’ as reified configurations of those schema features. Indeed, there appears no reason to assume the existence of discrete lexical senses and there is a growing body of research that refutes them (Geeraerts 1993; Kilgarriff 1997; Zlatev 2003; Glynn 2010 inter alia). Let us assume, therefore, that discrete senses are a useful heuristic in discussing

120 Dylan Glynn

semasiological structure in lexicography, but let us not assume that reified senses actually exist. Secondly, radial network analysis employed introspective methodology. Although introspection has an essential and inarguable role in language research, both for proposing hypotheses and performing analyses, with no truth-conditional tests to help determine conceptual structure, it will only ever be one part of an empirical science. Tyler and Evans (2001) have attempted to develop a ‘principled approach’ to identifying semasiological structure. This goes a long way towards minimising the risk of ad hoc categorisation using introspection. It does not, however, offer the possibility of result falsification. It is this second point that is essential. According to their own models of language, Generativists and Structuralists both had means for falsification using introspection. First, a proposed structure could be falsified by the intuition of a native speaker, whose linguistic knowledge was thought to exactly represent the grammar of a language. Second, truth-conditional semantic tests could be ‘failed’, thus establishing the lack of membership of a discrete category. However, if we assume a usage-based model of language, then neither of these possibilities is open to us, making introspection severely limited in terms of its ability to test hypotheses or falsify results. Accusing the tradition of radial network analysis of relying solely on introspection and of assuming the existence of discrete senses is, perhaps, unfair. Lakoff (1987) is careful in his wording to avoid the issue and Geeraerts (1989, 1995) explicitly develops a representational format that permits the description of polysemy without employing ‘nodes’. Moreover, there exist both corpus-driven and experimental studies in the tradition (cf. Glynn, this volume 7–38). Indeed, the corpus-driven usage-feature approach propounded by Gries (2006) goes back to the very origins of Cognitive Semantics (Dirven et al. 1982), just as elicitation-driven usage-feature analysis (Lehrer 1982) also finds itself at the origins of the theoretical paradigm. Moreover, as early as Geeraerts (1993), Geeraerts et al. (1994), and Lehrer and Lehrer (1994), in both theoretical and empirical terms, the argument for a non-reified approach to lexical senses was put forward. Therefore, even if it did not represent the main drive of research, Cognitive Semantics can be argued to have been slowly moving towards empirical methods and some in the field have long held that meanings cannot be understood as reified objects. Seen in this light, Gries (2006) is but one empirical Cognitive Semantic study in a long history. Its step was to explicitly apply corpus-driven usage-feature and multivariate statistics to the question of prototype-structured polysemy. This step, interpreting multivariate results in terms of prototype effects, is important. Empirical methods and non-reified senses may be theoretically sound, but with no way of modelling non-discrete results or coherently representing the structuring of language, the application of the method will struggle to gain ground. Therefore, a corpus-driven usage-feature methodology will build on the radial network tradition by fulfilling four aims:


a. b. c. d.

2.2

identifying encyclopaedic semantic structure; identifying prototype effects in that conceptual structure; positing non-discrete lexical sense; positing results that can be empirically falsified.

Operationalising prototype structured non-discrete lexical senses

It is one thing to demonstrate that lexical senses are not discrete in nature, it is another question altogether to develop a rigorous means for identifying and describing non-discrete semasiological structures. Multifactorial usage-feature analysis achieves this while simultaneously and empirically accounting for prototype effects in the structure. This section explains how the method represents non-discrete senses but also how it offers corpus evidence for the prototype structuring of those senses. Is it possible to make claims about the prototypicality of conceptual structure with corpus data? Prototype structure is an analytical model; it is not an object of study. It can be used to explain different structures in language, depending on how it is operationalised. Gries (2006:â•›75) offers a “non-exhaustive” list of different operationalisations of the notion of prototype structured polysemy: intuition determined judgements of similarity and ‘goodness’; elicitation ease; diachronic evidence; centrality/predominance in a radial network, and so forth. Geeraerts (1987:â•›288) argues that there are two basic operationalisations of prototypicality. He terms these the analytic and introspective criteria. Although his debate was with the proponents of truth-conditional semantics, we can rephrase this, mutatis mutandis, as frequency-based versus salience-based prototypicality. There are many different approaches to prototype structure, but it is likely that they will all be based on one of these two operationalisations: perceptual – conceptual ‘prominence’ versus relative frequency ‘commonness’. For the semasiological variation of a term such as run, we can suppose there would be little debate that ‘fast pedestrian motion’ is the prototypical ‘sense’. From synchronic frequency of use, diachronic evidence of earliest uses, and intuition-based conceptual salience, to widely accepted theories of embodiment and primacy of perception, all evidence points unanimously to ‘fast pedestrian motion’ as the ‘central’ meaning. However, as Geeraerts (1987) shows, theoretically, there is no reason a priori to assume that prototype models using one or the other operationalisation would offer the same results. Of course, this is not to say they will not. For an example as conceptually basic as run, it is likely they will and this is why, in Gries’ (2006:â•›76) comparison of different methods, each method indicates the same prototype structure. However, if we are developing a methodology for identifying semantic structure, it is important we do not make the assumption that these different methods should necessarily offer convergent results.

122 Dylan Glynn

Schmid (2000) and Gries (2003) have both made claims about the relationship between frequency and conceptual structure. These claims have yet to be confirmed empirically and the authors appear to have distanced themselves from their earlier position (Schmid 2010; Gries p.c). Although frequency of occurrence surely has an important role in determining conceptual structuring, conceptual and perceptual salience are also likely to have an impact. It is, therefore, unlikely that there is a oneto-one index where more frequent equates more central. Arppe and Järvikivi (2007), Arppe et al. (2009), Tribushinina (2009) and Gilquin (2010) are recent examples of research seeking to understand how these two fundamentally different operationalisations of prototypicality interact. Eventually, we may understand how their interaction impacts upon language structure and learning, but for the moment, this has not been determined. Having established the fact that we are restricting the notion of prototypicality to one based on frequency, let us turn to the non-discrete senses that we attempt to structure in these prototype terms. It is important to understand that the two senses that Gries identifies as the most frequent, and therefore (proto)typical, were predicted by a set of usage-features (Gries 2006:â•›85). They were not identified in the data as senses per se. For this reason, it is not, in fact, these two senses, but two ‘configurations of features’, to use the terminology of Geeraerts et al. (1994), or the ‘behaviour profiles of ID tags’ to use Gries’ terminology, that are the (proto)typical structures. This is why we should not speak of the many ‘senses’ of run but of the many ‘uses’ of run, where ‘use’ is understood as a re-occurring configuration of features (or ID-tags). This does not contradict Gries’ results. On the contrary, it further emphasises their theoretical and methodological implications. Given the context of tense (past), transitivity (transitive), complement syntax (to + infinitive), and agent type (Human), Gries is able to predict, with 100% accuracy, occurrences of the lexeme that would be traditionally defined as the ‘fast pedestrian motion’ sense of run. However, what Gries has identified as a ‘sense’ is merely a tendency for this configuration of features to occur together. In a given example, one or two of these features may not be applicable, but, relatively, it would still be an example of this usage/sense (as opposed to discrete lexical sense). What this gives us is a non-discrete operationalisation of ‘lexical sense’. It is important to understand that this is independent from prototypicality. The fact that this configuration of features is so stable and predictive only means that this particular sense is relatively discrete. In other words, it has a clear behavioural profile and its usage pattern can be accurately identified. This is independent from the relative frequency of this configuration, which determines its (proto)typicality. In terms of frequency, the features of ‘past tense’, ‘transitive’, ‘finite to infinitive syntax’, and ‘human agent’ were also the most frequent. Therefore, we have a non-discrete definition of a sense of run, but also a quantification of its typicality, or frequency-based prototypicality.


This discussion has sought to show how lexical senses can be operationalised through usage-feature analysis and how this brings in, seamlessly, the notion of frequency-based prototypicality. Yet the implications go further than providing an operational definition of prototypicality and lexical sense. These configurations of usage-features are, in fact, usage-contexts. This brings us to the question of the sociolinguistic dimension of meaning.

2.3

Multidimensional prototype effects: Form, meaning and context

Having established that corpus-driven usage-feature analysis can attain the goals set out in the era of radial network analysis, we now turn to what is argued to be an essential element of the object of study that is often side-lined. If we accept the usage-based model of language and the two operationalisations above ((i) sense – the configuration of features and (ii) prototypicality – the frequency of those configurations), the description of semasiological structure must integrate the social dimensions of language into its analysis. To appreciate why this is the case and why it is important, we must return to the question of prototype structuring. The operationalisation of semasiological structure in terms of frequency has an inherent limitation. Frequency, as a tool for determining linguistic structure, must necessarily be treated relatively. It is precisely this mistake Chomsky makes with the infamous argument that I live in New York is more common than I live in Dayton, Ohio tells us nothing of language structure. Chomsky (1964:â•›215) holds that frequency is external to language structure and will give us information about the world, or the context, instead of language itself. Although it is possible to argue that language is a mirror of the world and therefore, at some level, I live in New York is, in fact, more important in language than I live in Dayton, it would be difficult to demonstrate this to be the case for the identification of semasiological prototypicality. Instead, we can simply examine frequency relative to context. To understand how this notion applies to the polysemy, let us consider the example of Gries’ two frequency-based (proto)typical senses ‘fast pedestrian motion’ and ‘manage’. If the corpus had been children’s literature or sports magazines, ‘fast pedestrian motion’ would likely be frequent and ‘manage’ infrequent. By contrast, in the context of economic news press the ‘fast pedestrian motion’ is likely to be extremely infrequent, especially compared to the ‘manage’ sense. It should be obvious how essential the notion of context is to frequency-based studies of meaning. Are we trying to determine a typical meaning that is true for all language in all contexts? If this were possible (especially for a language as diverse as English), it is surely not possible with any corpus currently available or that will be available in the foreseeable future. Even taking a single context distinction, spoken versus written, the largest and most ‘balanced’ corpus in existence is non-representative to an unimaginable degree. The is because, in reality, the amount of spoken language greatly

124 Dylan Glynn

outweighs the quantity of written langauge, where the reverse is currently true of electronic corpora. In brief, the study of frequency-based prototype effects must be relative to context. We, therefore, must posit (proto)typicality structures, not for an entire language but for a language context, a specific place and time. We must avoid employing usage-based methods to describe the reductionist notion of the langue of Structuralism or the ideal speaker competence of Generativism. Our object of study is synchronically and diachronically varied – our models of conceptual structure must be sensitive to this.

3. Case study: run in America and Britain in diaries and conversation 3.1

Two corpus-based studies on ‘run’

Our current study imitates Gries (2006) in the set of usage-features (ID tags) analysed as closely as possible. The aim is not to test the results or to improve upon them through more advanced statistical analysis or a larger, more diverse sample. The aim is merely to show that even for a lexeme as culturally ‘simple’ and as socially ‘neutral’ as run, one must account for the social dimension of language in semantic analysis.2 In doing so, we will see why the statistical method he employs faces issues of reliability and we will introduce a different statistical technique. We begin with a summary of Gries’ (2006) study. Gries’ analysis is based on 815 occurrences of the lemma to run, extracted from the British component of the International Corpus of English and the Brown Corpus of American English. Approximately 400 occurrences were taken from each. These occurrences were manually analysed and categorised (using intuition) as belonging to one of 48 senses. These senses were taken from the Collins Cobuild E-Dictionary, the Merriam Webster’s American online dictionary, and the WordNet project. This categorisation in terms of dictionary senses is the first factor of the analysis. Although it is normally the goal of usage-feature analysis to determine different ‘senses’ through the identification of ‘feature configurations’ (Geeraerts et al. 1994) or ‘behavioural profiles’ (in Gries’ terminology), being able to match such configurations against

2. Kudrnáčová (2010) has also followed up Gries’ (2006) study with a more fine-grained corpus-based semantic analysis. Her study is not quantitative, but her corpus-illustrated insights will inform future research. In descriptive terms, the next step is to apply a more detailed usage-feature analysis and begin, not with dictionary senses, but a range of subtle semantic features. The senses should then be clusterings of those semantic features rather than simply matches between dictionary entries and observed occurrences.


dictionary definitions is a useful heuristic. In Gries (2006), it is used to show how a frequency-based study can inform an understanding of prototype structure in polysemy. The 815 occurrences in Gries’ data set are analysed for a range of factors, or usage dimensions. These factors consist of the usage-features typical in this kind of methodology – formal and semantic features ranging from syntax and collocation, tense and aspect, to the semantics of the argument structure and participants. In this study, the formal factors include tense, aspect, voice, transitivity, mood, and clause type. The semantic factors include subject type, object type, and complement type. These ‘type’ features are categories such as human, concrete countable object, concrete mass noun, machines, abstract entities, organisations, locations, quantities, events, processes, etc. The dictionary senses found are exemplified and enumerated. The most frequent dictionary senses identified are that of ‘fast pedestrian motion’ (203 occurrences / 25%, exemplified on p. 63) and ‘manage’ (101 occurrences / 12% exemplified on p. 71). The analysis and subsequent categorisation of the occurrences as dictionary senses is systematically explained by example. It is this systematic explanation that is used in the current study to repeat the analysis and to categorise the occurrences as dictionary definitions. The current study is based on 500 occurrences of run, 250 each of British and American English, subdivided again into 125 examples each from conversation and online personal diaries. The sample was restricted to this relatively small number due to practical reasons – usage-feature analysis is laborious and resource consuming. The point of the study being to investigate the need to include sociolinguistic parameters in polysemy research, the improved descriptive accuracy afforded by increasing this number would not substantially improve the ability to demonstrate the point. Also, the methods under investigation must be shown to produce coherent results with small numbers, since, for the same practical reasons, the usage-feature (or profile-based) method tends to deal with small samples. The British and American diary examples were taken from the LiveJournal corpus, developed by Dirk Speelman, at the University of Leuven, and the conversation examples were taken from the British National Corpus and the American National Corpus. The usage-feature analysis is replicated using the same dictionary senses employed by Gries and the same range of formal and semantic usage-features. An aside should be made here. Despite the fact that Gries more than adequately demonstrates the principle of the method, descriptively, the study is preliminary (Gries 2006:â•›81). The obvious question of why one would focus on dictionary senses (instead of solely usage-features, or ID-tags, to use Gries’ terminology) can be answered by the fact that the study’s aim is to show how prototype structure can be handled with the method. Nevertheless, in terms of descriptive adequacy, this option is far from ideal. Moreover, as the author stresses himself, the size of the sample is too small to properly apply multivariate statistical analysis. It is not that the sample is small in itself, but the type-token (or perhaps ‘sense-token’) ratio is not acceptable

126 Dylan Glynn

for multifactorial analysis. Gries repeatedly stresses this point, but it should be added that this problem is compounded by the fact that the study is not restricted to run, but includes all the verb particle constructions based on run. Arguably, this makes the study partially one of near-synonymy instead of polysemy. Many of the senses identified are determined formally by the combination of the verb and the particle. Verb particle constructions in Germanic, just like the prefixed verb constructions in Slavic (see Fabiszak et al., this volume 223–252), challenge the distinction between synonymy and polysemy. In any case, many of the senses in question are both formally and semantically distinct. A true test of the usage-feature method for the study of semasiological variation is when that variation is not linked to any overt, or obvious, formal distinction. By excluding the verb particle construction, Gries’ study would have included less semantic variation but also less formal variation for ‘automatically’ determining it. This does not detract from the goal or the results of Gries’ study, but future work should take such questions into account. Note that the current study also uses dictionary senses as one of its analytical factors and includes the particle constructions. This is done solely to permit a comparison with Gries. Table 1 lists the most common senses in the current study, compared with the figures from Gries (2006). The list of senses applied in this study was determined by the senses submitted to the hierarchical cluster analysis in Gries (2006:â•›82). For some of these senses, the number of occurrences (supplied in the preceding section, Gries 2006:â•›63–73) are not known. Although the reasoning behind the categorisation of the examples as dictionary definitions is reasonably clear, taxonomical issues of hyperonymy in the discussion occasionally mean that the number of occurrences for a given sense is not stated. This is the case for ‘function’ vs. ‘execute’ and ‘manage’ and for ‘free motion’ versus ‘motion’ and ‘fast motion’. The application of Gries’ dictionary senses to our data was reasonably straightforward, using the examples and explanations included in the study. There were, of course, some classification issues. For example, what constitutes ‘fast’ in ‘fast motion’? Table 1.â•‡ Most frequent dictionary senses Dictionary sense

Current study

Gries (2006)

‘fast pedestrian motion’ ‘escape’ ‘motion’ ‘fast motion’ ‘free motion’ ‘execute’ ‘in charge of ’ ‘manage’ ‘function’ ‘become used up’

160 (32%) â•⁄ 57 (11.5%) â•⁄ 23 (4.5%) â•⁄ 17 (3.5%) â•⁄ 17 (3.5%) â•⁄ 18 (3.5%) â•⁄ 16 (3%) â•⁄ 25 (5%) â•⁄ 17 (3.5%) â•⁄ 26 (5%)

203 (25%) â•⁄ 32 (4%) â•⁄ 24 (3%) â•⁄â•⁄4 (0.5%) â•⁄â•›– â•⁄ 28 (3.5%) â•⁄ 24 (3%) 101 (12%) â•⁄â•›– â•⁄ 14 (2%)


The large difference in the number of occurrences on this point suggests that there may have been a difference in coding for this sense. Nevertheless, assuming there is bound to be some analytical variation, the results are reasonably comparable. This is especially true given the small size of the samples and the differences between the corpora. The principal differences are ‘become used up’ and ‘escape’, which are more frequent in this study, and ‘manage’, which is substantially more frequent in Gries’ study. For this final difference, even if we allow for some confusion over the semantically similar categories of ‘execute’, ‘in charge of ’, ‘function’, and ‘manage’, the difference is marked. We can suppose that such differences are a result of register. Indeed, this is precisely the problem with frequency-based studies addressing (proto)typicality. Thematic variation, or variation in ‘topic of discourse’, can have a substantial effect, even upon coarse-grained analysis of semasiological structure. There is no need to examine such differences and similarities further. Both samples are small, with a high type-token ratio, which means statistical significance would tell us little. Remembering that the ultimate point is to show that frequency-based prototype structures are context dependent, it is sufficient to show that the overall study is comparable to that of Gries’. Below is an exemplified list of the most common senses. The examples are all extracted from the LiveJournal Corpus sample under investigation. For further exemplification and discussion, see Gries (2006:â•›63–73). (1) ‘fast pedestrian motion’ I want to like run into a bathroom at school and cry my eyes out whenever i see him. (2) ‘escape’ does anyone have about $400 laying around, i think i want to run away to Las Vegas for a few days, lol. (3) ‘motion’ Action Cat is really starting to like the new kitty, who I call Buddy cause he has yet to receive a formal name. They run around and play all the time now and it’s really cute. (4) ‘fast motion’ Hang on, till I get the brake on, or you’ll run into the river. (5) ‘free motion’ we’ve made three different trips … the group of friends that i run around with. (6) ‘execute’ you know like it’s easier for you to go and run a program you know through the disk.

128 Dylan Glynn

(7) ‘in charge of ’ there’s er it was for the er cat scanner and it was run by the Co-Op it was, it was just oh I saw that sign outside. (8) ‘manage’ I am now the new landlord of the rose and crown pub which mama used to run. (9) ‘function’ they said that uh cars would cost two dollars and they would run forever. (10) ‘become used up’ Well, it doesn’t do so bad. It’s usually cigs we run out of not petrol.

3.2

Semasiological clustering without social dimensions

Gries (2006:â•›81–82) submits all the senses (minus ‘idiomatic’ ones) to an agglomerative hierarchical cluster analysis (see Divjak 2010; Divjak and Fieller, this volume, 405–442, for an explanation of the technique). The senses are clustered using the full range of features (Gries 2006:â•›fn. 19, p. 94). The results of the cluster analysis are reasonably coherent, especially given the number of senses versus the number of examples and the number of usage-features. There is some degree of intuitively sound clustering, which could be re-interpreted as prototype structuring. Nevertheless, there is also a large amount of clustering that does not appear semantically motivated. Gries (2006:â•›81, 83) accepts this and suggests that the data sparseness is, at least partially, to blame. Replicating the procedure gives similar results – a reasonable degree of intuitively sound clustering but also a reasonable amount of ‘noise’ in the dendrogram where clusters make little or no sense. For the sake of brevity, we will not present the dendrogram, but instead present the cluster results obtained by simplifying the data and limiting the usage-features used to cluster them. In order to obtain a more coherent clustering of senses, rare senses were omitted. Also, the two most frequent senses, ‘fast pedestrian motion’ and ‘escape’, were omitted. These two senses were found to systematically dominate the clustering, rendering the relations between the other senses difficult to discern. We can suppose that these two senses were so distinct in usage that the clustering could not model their relationship and more subtle relations of the other senses simultaneously. This effect was found, regardless of the distance measure used. A combination of three factors is used to cluster the dictionary senses. Following Gries’ study, these factors consist of Transitivity, Subject ‘Type’ Semantics and Object ‘Type’ Semantics. Figure 1 presents the results using the simplified range of senses and these three factors. It is produced using the Euclidean distance measure (the simplest


free motion

motion

fast motion

in charge of

manage

execute

become used up

meet free motion metaphoric

increase

motion metaphoric

check/rehearse

motion into difficulty

broadcast

campaign

caused motion

be

copy

extend temporarily

flow

exist in abundance

function

diffuse

10 0

5

Height

15

20

Cluster dendrogram

hclust (*, "average")

Figure 1.â•‡ Hierarchical cluster analysis of dictionary senses. Distance matrix – Euclidean; agglomeration method – ‘average’

distance measure) and ‘average’ as the agglomeration method (a common agglomeration method). We must interpret such plots with caution. Even having removed the rarely occurring senses, some of the remaining senses are still infrequent, for example – ‘caused motion’, ‘motion into difficulty’ and ‘campaign’. Nevertheless, the overall picture seems reasonably coherent. Examining the dendrogram, two broad sense clusters emerge, clustered by the right and the left branches. The left branch includes most of the abstract senses, with perhaps the exception of ‘function’, which is less abstract. Note, however, that the analysis has ‘function’ and ‘diffuse’ as quite distinct from this abstract cluster. It appears that the analysis has trouble incorporating these senses. The intuitive adequacy of the model is left up to the reader, but it is worth pointing out that the literal motion senses are coherently grouped together as well as the control senses (‘manage’, ‘in charge of ’, ‘execute’). However, the place of ‘become used up’ with these two groupings of senses is not clear, nor is the relationship between the ‘control’ senses and the literal motion senses. There does exist an internal logic to the cluster of abstract senses. The metaphoric motion senses are grouped together, just as are the ‘spread’ senses of ‘flow’, ‘exist in abundance’, and ‘extend temporarily’. The other groupings are not illogical, but apart from representing abstract or metaphoric meanings of run, they share little semantically.

130 Dylan Glynn

Gries (2006:â•›fn. 19, pp. 93–94) found that changing the distance matrix and/or the agglomerating method did not alter the results. This was not the case with the current data set. Experimenting with different agglomeration methods greatly improved or worsened the interpretability of the dendrogram and, occasionally, the actual results of the cluster analysis. Likewise, different distance measures also produced different results. This could, perhaps, be a sign of the instability of the analysis – attempting to cluster 23 senses based on a sample of 500 is far from an ideal condition in multivariate statistics. Figure 2 presents the results of the Canberra distance matrix. It is clustered with the Ward agglomeration method. The different agglomeration methods did not change the results for the Canberra matrix, only legibility. The Ward method gave the clearest dendrogram. In Figure 2, again we see two main branches. At first, the overall branching and clustering of the senses appears more coherent than those produced using the Euclidean distance measure. However, if we inspect the clustering more closely, intuitive semantic coherence is not wholly systematic. At the coarse-grained level, we have lost the clear distinction between relatively concrete uses such as ‘manage’ and ‘literal motion’ versus ‘metaphoric motion’ as well as the extending and disseminating senses. For the four sub-clusters, there is a little more semantic coherence. The first sub-cluster of ‘execute’ and ‘manage’ is intuitively sound. The second is also coherent,

execute manage free motion motion motion metaphoric fast motion in charge of difficulty campaign meet exist in abundance extend temporarily be copy flow rehearse caused motion free motion metaphoric diffuse function become used up broadcast increase

10 0

5

Height

15

20

25

Cluster dendrogram

dist_mat hclust (*, "ward")

Figure 2.â•‡ Hierarchical cluster analysis of dictionary senses. Distance matrix – Canberra; agglomeration method – ‘Ward’


save for the sense ‘in charge of ’. This is semantically related to ‘execute’ and ‘manage’ and, therefore, given the small sample, is more or less in the ‘correct’ branch. The next sub-cluster of ‘difficulty’ (run into difficulty), ‘campaign’ (run for election), and ‘meet’ (run into a friend) is semantically coherent, given a broad interpretation of ‘campaign’ that includes meeting people and difficulties. This is not as unlikely an interpretation as one might first suppose. Recall that the different semantic types of objects and subjects determine these sense clusters. Moving to the right across the clusters, the next sub-cluster of ‘exist in abundance’ and ‘extend temporarily’ is intuitively coherent. However, the rest of the group appears semantically heterogeneous. The last cluster on the right, although distinct with a long branch stemming from the rest of the dendrogram, also lacks obvious semantic coherence. Although one is able interpret semantic structure here, it is not self-evident why ‘diffuse’ and ‘function’ or ‘broadcast’ and ‘increase’ should group together. The point of both this small study and Gries’ is merely to consider two methodological possibilities. In light of this, the fact that the two distance matrices produced different clusterings raises important methodological questions. Standards and checks for appropriateness need to be developed before the use of cluster analysis can be relied upon to determine frequency-based semasiological structure.3

3.3

Semasiological clustering with social dimensions

Before we consider the effects of social variation on semantic structure, it must be stressed that one would not expect to find substantial variation with these data and for this lexeme. Therefore, even a small degree of variation is a sign of the extent of the issue. There are four reasons for this: 1. In terms of cultural variation, run is a ‘simple’ lexeme. It is the kind of lexeme where one would not expect variation across dialects. 2. In terms of register, run is a ‘neutral’ lexeme, not belonging to either formal or informal registers. It is the kind of lexeme where one would expect relatively little variation across text types. One exception to this might be the two central senses of ‘fast pedestrian motion’ versus ‘manage’, where text type would be expected to show variation in use.

3. Divjak and Gries (2006:â•›37) state that the Canberra distance matrix is best suited to small cell counts, such as we have here. Gries (2009:â•›317) says the choice is subjective. Gries and Stefanowitsch (2010:â•›79) employ the Manhattan distance matrix, citing Levy et al. (1999) as justification. Levy et al.’s study compares five distance matrices but not Canberra. It seems that the question of how the choice of distance matrix affects the results needs to be investigated systematically. Divjak and Fieller (this volume, 405–442) also discuss the range of methods.

132 Dylan Glynn

3. Although the differences between American and British English are substantial, the dialects remain mutually intelligible for most speakers of both varieties, especially in written language and educated speech. In other words, the difference between American and British English is not that great, making dialect a good test case. 4. Although there are certainly differences between the registers of spoken conversation and online personal diaries, the style of the latter is also extremely informal and is also dialogic. Unlike traditional diaries, authors here engage in discourse with readers and the style of the genre is conversational and casual. Therefore, just as for dialect variation, one would not expect substantial differences in the text type variation. We could repeat the clustering presented in the previous section for the two dialects and the two registers and compare the clustering. However, the cluster analyses on the full data set are obviously unstable and halving the data would make any multivariate analysis impossible. Let us begin, rather, with a Chi-squared test of independence that identifies statistically significant differences along the lines of register and dialect. A Pearson’s Chi-squared test of independence for dialect identifies significant differences between the British and American data for the dictionary senses (p = 0.001263). The residuals show that ‘become used up’ but also ‘escape’ and ‘fast motion’ are more typical of the British use, and ‘meet’, ‘increase’ but also ‘execute’ and ‘diffuse’ of the American use. Register also reveals a significant difference (p = 6.376e-05) with the residuals showing that ‘escape’, ‘fast pedestrian motion’, ‘metaphoric motion’ are associated with the diaries, and ‘caused motion’, ‘diffuse’, ‘execute’, ‘function’, and ‘increase’ with the conversation data. Having established that there is significant variation, let us move to trying to capture how that variation interacts with the semasiological structure. Although cluster analysis is a powerful tool for identifying how the different senses are related, it cannot show how register and dialect affect those relations. Ideally, given enough data, we could label the occurrences of the different senses for dialect and register and even both simultaneously. The cluster analysis would then show the relations between the different senses relative to the social factors, clustering, for instance, ‘fast pedestrian motion BrEng’ and ‘fast pedestrian motion AmEng’ etc. Although a straightforward procedure, for the number of senses involved, this would require a much larger data set. Another statistical technique, explained in Glynn (this volume, 443–485), is correspondence analysis. A multivariate and exploratory technique similar in many ways to cluster analysis, it visualises relations between all the factors considered rather than just one factor. Figure 3 presents the results of a binary correspondence analysis, which examines the interaction of dialect, register, and dictionary sense.


Cause Motion Flow Extend Time

0.5

Use up

BrE.Conv

Function Motion Manage

Escape

Rehearse Fast Ped. Motion

Increase

AmE.Conv Diffuse Execute Difficulty

AmE.Blog Met. Motion

–0.5

0.0

Fast Motion BrE.Blog Free Motion Become Broadcast

Meet

Campaign

Extend Space

–1.0

Copy

–0.5

0.0

0.5

1.0

1.5

Figure 3.â•‡ Binary correspondence analysis of register, dialect, and dictionary sense

The first two dimensions of the analysis explain 87% of the variation (inertia), which is a relatively stable analysis. Immediately, it is visible that American Conversation (AmE.Conv) is distinct in use relative to the dictionary senses, dominating the right two quadrants of the plot on the central axis line. The senses ‘increase’, ‘diffuse’, and ‘motion into difficulty’ (Difficulty) are distinctly and highly associated with the American conversation data point on the right of the plot. In the bottom half of the plot, we find a range of senses distinctly associated with the American diary genre (AmE.Blog). The senses ‘campaign’, ‘copy’, and perhaps ‘metaphoric motion’ (Met. Motion) are highly and distinctly associated with American diary use. ‘Meet’ and ‘extend space’ are likely to be associated with American English but are not distinct to either register, lying between the two data points for American Diary and American Conversation. Moving to the British uses, the plot becomes more difficult to interpret. The analysis suggests that there is less register variation in the British sample, the two data points British Conversation (BrE.Conv) and British Diary (BrE.Blog) both lying in the same top left quadrant. Nevertheless, the dialect variation is clear – the senses ‘flow’ and ‘extend time’ are highly and distinctly associated with the British use. Other senses, such as ‘use up’, ‘cause motion’ and ‘escape’, are also relatively associated with British use, but this association is not distinctive.

134 Dylan Glynn

0.5

S.Cause Motion

S.Flow S.Use up

S.Function

0.0

S.Increase

S.Diffuse

S.Execute

Reg.Conversation Lang.BrEng S.Manage S.Free Motion S.Motion S.Broadcast S.Become S.Fast Motion

S.Extend Time

S.Fast Ped. Motion

S.Difficulty

S.Escape S.Rehearse

Lang.AmEng

Reg.Blog

S.Meet

–0.5

S.Extend Space

S.Metaphoric Motion S.Campaign

S.Copy

-1.0

-0.5

0.0

0.5

1.0

Figure 4.â•‡ Multiple correspondence analysis. Burt matrix, method ‘adjusted’

In order to obtain a clearer picture of the interactions at hand, let us submit the same data to a multiple correspondence analysis. The binary analysis, in Figure 3, gives us a reliable and stable representation of the associations, but it cannot capture interactions between dialect and register. This is because these two factors were concatenated in order to produce a two-dimensional contingency table for the analysis. We can expand that table into a three-dimensional table and apply multiple correspondence analysis. The results are more difficult to interpret and can be less stable (less accurately capturing and representing associations in the data). However, the plot in Figure 4 was produced using the recently developed ‘adjusted’ method which addresses both issues of stability and clarity. Fortunately, the results are clear and the explained inertia is 86.7% (Dim. 1: 61.2%, Dim. 2: 25.5%), which for a adjusted multiple correspondence analysis, using Burt matrices, is a stable result (Greenacre 2007). Further details on and an explanation of the technique of correspondence analysis, and its limitations and strengths, can be found in Glynn (this volume, 443–485). The results presented in Figure 4 largely reflect the binary correspondence analysis, but by treating the factors of dialect and register independently, the analysis affords us a clearer depiction of their interaction. Each of the four quadrants is characterised by one of the four sociolinguistic features: the top right – British dialect (Lang. BrEng); the bottom right – diary register (Reg.Blog); the bottom left – American dialect (Lang.AmEng); and the top left – conversation register (Reg.Conversation).


We see that senses, such as ‘execute’ and ‘diffuse’ between the American data point and the Conversation data point, are common to these two usage dimensions. The senses ‘campaign’ and ‘metaphoric motion’, lying between the American data point and the register of diary (Reg.Blog), are common to these dimensions. The senses ‘beyond’ the American data point, relative to the British dialect data point in the top right-hand quadrant, are neutral with regard to register, but are distinctly American in contrast to British. These senses include ‘extend in space’, ‘copy’, and ‘meet’. Repeating the interpretation, beginning from the top right-hand quadrant and the British data point, we see that ‘use up’ is distinctly typical of British conversation and that ‘fast motion’ is typical of British diaries. The senses ‘flow’ and ‘extend time’ are less associated with a given register, but are distinctly British, relative to the American data. Again we see that register variation for the British use is less important. Finally, note the position of ‘manage’ and ‘fast pedestrian motion’. These data points, along with some other senses, are in the centre of the plot. The senses located in the centre are the senses that are not affected by either of the two sociolinguistic usage factors. These senses are central, but not just in the way that Gries (2006) argued. Although still understood in terms of frequency, we now also have two usage dimensions, dialect and register. Not only are these senses among the most frequent, they are among the senses least affected by context. This is a crucial refinement to the frequency operationalisation of (proto)typicality – uses that that are common (frequent) across all contexts are more central to the meaning of a lexeme. This finding is equally as important as discerning which senses are typical of specific contexts. Gries (2006) stresses that the small sample means that the study can only be seen as a methodological test, rather than a fully descriptive analysis. For these reasons, the statistical techniques employed are only exploratory. He suggests the use of configural frequency analysis to identify statistical significance in the results, allowing one to determine which correlations are not chance, and which may be simply a result of the small sample. Although configural frequency analysis would be an excellent choice for this, it requires more data than is available in either study. It also follows that, with more data, log-linear analysis or multinomial logistic regression would be even better, giving not only statistical significance but also predictive strength to the model. Such analyses are now within the capabilities of corpus-driven research, but require a larger scale analysis. Moreover, before such an analysis is undertaken, the identification of senses must be better operationalised. The analysis of the usage-features must be found to cluster into senses and then these multivariate senses must be shown to be statistically significant. With senses based on clusters of usage-features (ID Profiles), rather than revealed by matching occurrences with dictionary entries, we can then return to the clustering. This step in corpus-driven polysemy research has begun (Glynn 2009, 2010, in press), but remains at the initial stages. Once we are armed with the analytical tools to identify multivariate senses (rather than dictionary senses), then we

136 Dylan Glynn

need to progress to modelling the semasiological structure and the prototype effects, using more advanced statistical procedures such as configural frequency analysis and log-linear analysis. The present purposes are to demonstrate that sociolinguistic effects must be integrated into the study of prototype structuring. To these ends, let us submit the data to binary logistic regression. Explained in Speelman (this volume, 487–533), logistic regression is a confirmatory multivariate technique that allows us not only to determine which of the usage features and/or dictionary senses are significantly associated with either of the sociolinguistic factors, but it also enables us to determine how important that association is.

Logistic Regression – Dialect

Let us begin with dialect. Three logistic regression models are reported: a multiple model based on usage-features excluding dictionary senses (Model 1); a second multiple model that includes dictionary senses (Model 2); and a simple model with the dictionary senses as a sole predictor variable (Model 3). The models are all checked for multicollinearity, and factors producing a variance inflation of more than 2.5 are removed.4 Moreover, the models are checked for singularity with a Kappa calculated condition number – any model with a value higher than 6 is rejected.5 The strict check on variance inflation and singularity assure an orthogonal model. The models are also checked for influential observations as well as overfitting, neither of which is a problem. Outliers are not removed. In a backward elimination of factors, model selection was based on significance values and the Akaike’s information criterion (AIC), not on predictive strength.6 For readers unfamiliar with logistic regression, the testing of the model and criteria for acceptability were extremely strict, making the results as conservative as possible. For the sake of brevity, some non-significant levels are omitted, indicated by ‘…’. Positive coefficients predict British English and negative coefficients (“–”) predict American English. Since we are comparing models, only the coefficients and some

4. Some authorities indicate a variance inflation factor of 10 to be acceptable (DeMaris 2003:â•›517; Dodge 2008:â•›96; Chatterjee and Hadi 2006:â•›238; Marques de Sá 2007:â•›307; Speelman p.c.), other authorities are non-committal (Faraway 2002:â•›117–120; Maindonald and Braun 2003:â•›201–203). Glynn (2010) and Speelman (this volume, 487–533) opt for a maximum inflation value of 4. Szmrecsanyi (2006:â•›215) notes that even values as low as 2.5 can be a cause for concern. Multicollinearity is a serious issue in regression and can lead to Type I errors. Since we do not necessarily understand the relationship between many of the factors in our model, a maximum VIF of 2.5 is used to determine which factors can be combined in the model. 5. Baayen (2008:â•›182) states that a condition number between 0 and 6 indicates no multicollinearity and 15 indicates a medium degree. 6. The AIC score helps compare the parsimony of different models. The scores are relative and a lower number indicates a more parsimonious model.


Table 2.â•‡ Logistic regression models for dialect Coefficients

Model Statistics

Transitivity – Transitive Tense – Past Tense – Present Aspect – Progressive Aspect – Simple Mood – Imperative Mood – Interrogative Clause Type – SubPronoun Clause Type – SubNP Clause Type … Subject – Human Subject – Locations Subject – Machine Subject … Sense – Use Up Sense – Diffuse Sense – Execute Sense – Campaign Sense – Fast Motion Sense – Flow Sense – Increase Sense – Meet Sense … d.f. G2 C Nagelkerke R2 Bootstrapped R2

Model 1

Model 2

Model 3

â•⁄0.596619* â•⁄0.658282* â•⁄0.387003 â•⁄– â•⁄– â•⁄0.533600 â•⁄1.048808* –1.807732º –0.490190 â•⁄… –1.446052º –0.289667 –1.564679* â•⁄… â•⁄– â•⁄– â•⁄– â•⁄– â•⁄– â•⁄– â•⁄– â•⁄– â•⁄– 20 41.39** â•⁄0.668 â•⁄0.112 â•⁄0.0138

â•⁄0.487405º â•⁄– â•⁄– â•⁄0.178450 â•⁄0.542254º â•⁄0.392821 â•⁄1.094234* –2.225015 –2.225015º â•⁄… â•⁄– â•⁄– â•⁄– â•⁄– â•⁄0.955200º –1.364175º –1.196454* –1.289732 â•⁄0.792536 â•⁄1.405831 –2.287706* –1.566828* â•⁄… 27 70.71*** â•⁄0.716 â•⁄0.186 â•⁄0.0435

â•⁄– â•⁄– â•⁄– â•⁄– â•⁄– â•⁄– â•⁄– â•⁄– â•⁄– â•⁄– â•⁄– â•⁄– â•⁄– â•⁄– â•⁄0.94852* –1.65945* –1.14862* –1.65945 â•⁄0.82546 â•⁄1.55943 –2.12945* –1.59046* â•⁄… 20 60.16*** â•⁄0.671 â•⁄0.156 â•⁄0.057

essential model statistics are reported.7 In Table 2, the coefficients for each of the levels (usage-features) are listed with the alpha levels (º p sequence (1987:â•›19–21). However, Allerton gives no corpus evidence to support his claims and, although convincing, his model is not empirically grounded. Paradis’s taxonomy is finer-grained. Like Quirk et al., Paradis (1997:â•›27–28) proposes a general bipartition between modifiers that scale upwards (‘reinforcers’) and modifiers that scale downwards (‘attenuators’). She postulates the existence of a cline between these two poles. She unifies each set on the cline by means of the principle of cognitive synonymy (Cruse 1986). She further subdivides ‘reinforcers’ and ‘attenuators’ into ‘totality modifiers’ and ‘scalar modifiers’. Unlike Quirk et al., but like Allerton, Paradis is well aware that some degree modifiers are polysemous, semantically fuzzy, and therefore paradigmatically unstable. Unlike Allerton’s, Paradis’s typology is empirically founded. She devotes a chapter of her monograph to the distribution of degree modifiers of adjectives across the London-Lund Corpus of spoken English (1997:â•›Chapter 2). Her goal is both to examine the distribution of degree modifiers of adjectives across modes (spoken vs. written) and to inspect the detail of collocational preferences in the hope of establishing a subtle classification. She concludes that intonation is the key to understanding the meaning and the grading force of degree modifiers (1997:â•›Chapter 4). Previous corpus-based approaches to degree modifiers make extensive use of raw counts, or relative frequencies that do not filter away overrepresented adjectives. Other 4. Strictly speaking, Quirk et al.’s two-branch model is restricted to degree modifiers used as subjuncts, but as Paradis (1997:â•›23–24) observes, this model applies equally well to degree modifiers of adjectives.

Visualizing distances in a set of near-synonyms 151

approaches may also suffer from the kind of corpus that they use, the lack of statistical methods, or the selection of inadequate statistics. If one uses a corpus that is relatively small, statistical methods that are invalid for low counts, such as (pointwise) mutual information or χ2, are inappropriate, and one ends up being trapped in a vicious circle. This is a problem if one wants to investigate collocations, some of which might be relatively rare but by no means insignificant, both statistically and semantically.

2.2

Corpus-based Cognitive Linguistics

Cognitive Linguistics is a usage-based approach to language that makes no principled distinction between language use and language structure. A linguistic unit is entrenched and stored in grammar when a usage pattern generalizes across recurring instances of language use. The more frequently speakers encounter a linguistic unit, the more the linguistic unit is entrenched, i.e. established as a cognitive routine. In Cognitive Linguistics, entrenched patterns of usage provide privileged access to speakers’ knowledge of their language. The meaning of a linguistic unit “involves both conceptual content and the construal of that content” (Langacker 2008:â•›44). Conceptual content is referred to as a domain: a consistent knowledge representation that serves as a basis for the construal of other conceptual units. For instance, adjectives such as damp, dampish, moist or dry relate to conceptual units to be construed with respect to the domain of wetness. The content that the domain designates can further be construed with respect to the more basic domain of property. Construal is the way a speaker presents a conceptual representation through the choice of a linguistic expression. While lexemes and constructions bring with them a conventional meaning, this coded meaning is modified by the context in which linguistic units occur. Far from being a fixed, pre-assigned feature, meaning is negotiated locally and socially. By adopting a specific focal adjustment, the speaker linguistically organizes a scene and influences the way the conceptual representation that the expression evokes will be received by the hearer. This is consistent with the theoretical framework of Cognitive Construction Grammar (CCxG), as laid out in Langacker (1987), Goldberg (1995), and further refined in Goldberg (2003, 2006, 2009) and Langacker (2008, 2009). One of the tenets of CCxG concerns the link between constructions and construal.5 Most Construction Grammar supporters recognize that a given construction is both a product and a vector of conceptualization (see Note 1). Goldberg (2003:â•›219) notes that a linguistic pattern counts as a construction “as long as some aspect of its form or function is not strictly predictable from its component parts or from other constructions recognized to exist”. Productive or semi-productive constructions such as day after day, twistin’ 5. The fact that the noun construction has two corresponding verbs, construct and construe, is not innocent.

152 Guillaume Desagulier

the night away, or boy, was she in trouble! are idiosyncratic and non-compositional. They must be learned on the basis of the input. Likewise, it may well be the case that entrenched types of the <moderator + adjective> construction provide access to conceptual structures in ways that are not necessarily compositional. Competence-based versions of Construction Grammar might deny the <moderator + adjective> pattern the constructional status or at least warn that we should not treat de facto all collocation patterns as constructions. In Fillmorean Construction Grammar, for example, only collocation patterns that are truly productive count as constructions, other kinds of collocation patterns being consigned the ‘meta-grammar’ together with other ‘patterns of coining’ (Kay 2013). In contrast, usage-based models of Construction Grammar such as CCxG have a broader understanding of the notion of a construction. In a neo-Saussurean fashion, they consider that all form-meaning pairings count as constructions: “it’s constructions all the way down” (Goldberg 2003). To put it plainly, according to competence-based models of Construction Grammar, the linguist has no a priori reason to think that a unit counts as a construction, whereas according to usage-based models, linguists have no a priori reason to believe that a unit is not a construction. In this paper we adopt a usage-based perspective because it leaves it to empirical analysis to decide whether a unit counts as a construction or not. Our choice is reinforced by the fact that the quantitative apparatus designed for constructions can also handle regular collocation patterns (i.e. patterns that competence-based models consider as not necessarily constructional). The entrenchment of some instances of the <moderator + adjective> construction explains why moderators scale in different ways depending on the adjective they modify. For example, the scaling force of pretty depends on the adjective it modifies. As Athanasiadou points out, “pretty small is not very small, but pretty straight is very/ quite straight” (Athanasiadou 2007:â•›557). More recently, Goldberg (2006:â•›45) allowed that “individual patterns that are fully compositional are recorded alongside more traditional linguistic generalizations”. This means that all types of the <moderators + adjective> construction, i.e. entrenched as well as non-entrenched types, are worth investigating because each is the trace and the vector of construal mechanisms in the domain of degree modification. Since corpus-based linguistics provides a comprehensive array of methods to capture context and knowledge, it is not surprising that it has become central in the investigation of cognitive patterns of usage in Cognitive Linguistics (Gries & Stefanowitsch 2006). Paradis (1997:â•›62) claims that the semantic relation between degree modifiers and adjectives is bidirectional. More precisely, the adjective selects a degree modifier, which in turn restricts the interpretation of the adjective. This means two things. First, we should not examine probabilistic co-occurrences of words regardless of their morphosyntactic environment. Instead, we should inspect co-occurrences of moderators and adjectives within constructional patterns because both moderators and adjectives contribute to the meaning of the <moderator + adjective>


construction. Second, the semantic interaction between a moderator and an adjective is not necessarily compositional.

2.3

Cognitive synonyms or near-synonyms?

According to Paradis (1994:â•›160; 1997:â•›71), moderators are ‘cognitive synonyms’ because substituting one for the other does not change the truth-value of the proposition (Cruse 1986:â•›270ff.).6 Cognitive synonyms share identical conceptual content and differ only in style (die vs. pass away), register (drunk vs. pissed), and connotation (firm vs. stubborn). In other words, cognitive synonymy is a matter of both sameness and difference. Near-synonyms are sometimes referred to as ‘plesionyms’ (Hirst 1995; Edmonds & Hirst 2002; Storjohann 2009). Plesionyms differ from cognitive synonyms because the former involve a slight change in denotational reference, be it in terms of degree (drunk vs. hammered), fuzzy boundary (forest vs. woods), viewpoint (slim vs. skinny), intensity (break vs. destroy), etc. Substituting near-synonyms alters truth conditions but the sentences where they appear remain semantically similar. Cognitive synonymy and near-synonymy are often hard to tell apart, and an opposition between these two concepts is somehow unhelpful. According to Edmonds and Hirst, near-synonyms bring with them a finer-grained representation than cognitive synonyms. Their conclusion is that “it should not be necessary to make a formal distinction between cognitive synonyms and plesionyms” (Edmonds & Hirst 2002). The conceptual content that moderators share is essentially functional: rather, quite, fairly, and pretty moderate the qualities denoted by the adjectives they modify. But moderators are not completely interchangeable in all contexts. Two reasons have been put forth. First, moderators do not express the same degree of moderation. Second, they involve distinct modes of construal (Paradis 2000, 2008). I will put forth another reason: moderators do not always operate within the same conceptual domains.

3. Method I combine two broad types of statistics: analytical statistics and multifactorial methods. Analytical statistics uses measures such as frequencies in the domain of hypothesis testing. Rather than formulate and test hypotheses, multifactorial methods aim to formulate a statistical model, i.e. the statistical description of variation and complex relations in a multi-variable dataset.

6. Cruse’s definition of cognitive synonymy echoes Quine’s (1951).


To describe and interpret the distribution of the four moderators, we shall start with a study of collocation patterns. The information concerning the distribution of moderators will serve as input for two methods of multifactorial analysis. The paragraphs that follow will justify the choice of the Corpus of Contemporary American English, explain how the data was extracted, and present the measures of lexico-grammatical co-occurrence that are used, namely collexeme analysis, and multiple distinctive collexeme analysis. Finally, two multifactorial analyses will be introduced: hierarchical cluster analysis and correspondence analysis. Their input will be provided by the results of the collexeme and multiple distinctive collexeme analyses.

3.1

The corpus

At the time of writing, the Corpus of Contemporary American English (Davies, 1990-present), henceforth COCA, consists of 414,771,808 words of spoken and written American English divided among 169,140 texts. The spoken part contains approximately 85 million words and consists of transcripts of conversation from TV and radio programs. The written part is divided evenly between four genres: fiction (80 million words), popular magazines (86 million words), newspapers (82 million words), and academic journals (82 million words). The whole corpus spreads across the period from 1990 to 2010, and 20 million words are added each year. One advantage of COCA is that it is probably the largest publicly and freely available annotated corpus of English. Its size and sampling scheme increase the reliability and validity of observations of relatively rare linguistic phenomena. However, one major disadvantage with COCA is that the texts themselves are not available for download for copyright reasons. The only way to run queries is via the native search interface.7 Nevertheless, it is simple to copy and export query results into a text editor or a spreadsheet, clean up the data, cross-tabulate, and run statistical tests. Davies claims that COCA is balanced. To some extent this is true because words are evenly distributed across genres. Yet, as is the case with all major corpora, balance is more an ideal than a reality. COCA also suffers from the fact that speaker-related metadata are not computed relative to the overall size of the corpus, making them very difficult to measure statistically. For example, COCA indicates the identity of the speakers, but gender is left implicit. However, for the current purposes, the fact that speaker-related information is hard to obtain is of little importance since the main

7. This could be a major issue in light of Leech’s “standards of good practice” for corpus users and corpus compilers (Leech 1997:â•›6). At the top of Leech’s list is the following recommendation: “[t]he raw corpus should be recoverable.” Even if the whole texts are unavailable to the COCA end-users, it is nevertheless possible to “dispense with the annotations, and to revert to the raw corpus” (ibid.), although partially.


focus of the present study is on function-related distributions and collocation patterns, regardless of dialectal issues. Last but not least, queries on the COCA yield duplicates. This is largely due to automatic text collection, a small price to pay for a corpus of this size. I have not removed duplicates for two reasons. First, duplicates that result from a genuine verbatim quotation are very hard to distinguish from duplicates that result from an error in automatic text compilation. Second, the number of duplicates does not affect the statistics in any significant way. All corpora come with restrictions, and it is important to bear these in mind when interpreting results. All things considered, the advantages of COCA outnumber its disadvantages.

3.2

From collocates to collexemes

A common assumption in corpus linguistics is that the context of a variable (lexical or phrasal) reveals important aspects of its syntactic and semantic properties (Sinclair 1991; Biber et al. 1998). The easiest way to analyze the context of a variable is to extract its collocates and determine those that most typically combine with the variable. According to Sinclair (1991:â•›170), collocation is “the occurrence of two or more words within a short space of each other in a text”. Traditionally, the conventional size of what Sinclair calls ‘a short space’ varies from 1 to 5 words on both sides of the variable, depending on the case study. Syntactic studies tend to examine wider spans, while lexical studies (e.g. those that focus on idiomatic compounds such as bread and butter) examine shorter spans. Some authors also assume that statistical methods can easily filter out the noise generated by a wide concordance span, but once again the wider the span, the weaker the syntactic claims the linguist can make. In the literature on intensifiers, little is said as to the optimal span to investigate. In his study on amplifiers, Kennedy inspects “a window of two words” on each side so as to “retrieve collocations that may have been separated by intervening words” (2003:â•›472–473). Amplifiers can modify elements that precede or follow. This fact alone justifies a two-way search. Amplifier collocates can be adjectives, other adverbs, or verbs, and given their multifunctionality, “intervening words” are frequent. Moderators are slightly different. Insofar as they are ‘word modifiers’ (Stoffel 1901), there is no need to inspect too wide a range of words, and since these adverbs are premodifiers it is useless to inspect the left context.8 Bearing these specificities in mind, I extracted 8. In many collocation-based studies, node words or phrases are often clustered on the basis of their collocates within a relatively large span. One big problem with this method is that the semantic relation between many of these collocates and the node(s) is loose, which brings about noise. By adopting a constructional approach, and therefore treating the sequence <moderator + adjective> as a construction, I restrict the semantic investigation to the syntactic frame of the construction and inevitably minimize the risk of obtaining noise in the data points.


all adjectives that occur in the first two slots to the right of rather, quite, fairly, and pretty. I also adopted the “paradigmatic reduction” outlined in Lorenz (2002:â•›144) and considered only moderators of adjectives so as to focus on subtle semantic relations. Collocation extraction is a preliminary step to determining sequences of moderators and adjectives that co-occur more frequently than would be expected by chance. Determining collocation strength is a particularly helpful way of spotting differences between semantically similar expressions. Even though we assume that rather, quite, pretty and fairly are near-synonyms, and therefore map onto similar content domains in a quality-related conceptual space, we may expect significantly distinct collocation patterns for each moderator. A related assumption is that distinct collocation patterns reflect subtle differences in schematic-domain profiling. As mentioned earlier, the semantic relation between degree modifiers and adjectives has been described as bidirectional. While this is a convincing assumption, a simple collocate-based inquiry based on raw counts and/or basic relative frequencies is unsatisfactory because it fails to distinguish those adjectives that are significantly attracted to a degree modifier from those that are frequent in the corpus regardless of the context where they appear.9 In this study of the <moderator + adjective> construction, we shall turn to collexeme analysis (Stefanowitsch & Gries 2003), a method that can both handle two-way semantic influences and filter out adjectives that have a high overall token frequency in the corpus. Collexeme analysis is part of a family of methods known as collostructional analysis (Hilpert, this volume, 391–404).10 It investigates which lexemes typically occur in a given slot in a single grammatical construction (e.g. the X-er, the better). It takes as input the frequency of a lexeme in a given construction, the frequency of the same lexeme in all other constructions, the frequency of the construction with other lexemes, and the frequency of all other constructions with other lexemes (Stefanowitsch & Gries 2003:â•›218). The data is tabulated and the 2x2 table is submitted to the Fisher-Yates Exact test (Pedersen 1996). Unlike χ2-statistics, the Fisher-Yates Exact test does not presuppose or violate any distributional assumptions (Kilgarriff 2005). Unlike χ2-statistics and Mutual Information, the Fisher-Yates Exact test is not invalid when counts are low. This means that rare collocations in COCA (such as quite medical or pretty social) can nevertheless be included in the overall calculation of collostruction strength. The p-value that one obtains through the Fisher-Yates Exact test is an indication of the association strength between the lexeme and the construction: 9. Raw counts provide the simplest association measure for pair types. When the construction under investigation displays low token frequency, raw counts are often the only option. However, they should be regarded as a last resort. 10. Each of these methods measures the degree of attraction or repulsion between lexical items and constructions. In collostructional analysis, a construction is defined as a conventional pairing of form and meaning, in light of Goldberg (1995).


the smaller the p-value, the stronger the association between a lexeme and a construction. In recent versions of collexeme analysis, the p-value is transformed in an inverse logarithmic function so as to make the distinction between very small p-values easier to identify. In the present study, this operation will be repeated for all adjective types that co-occur with each moderator. Collexeme analysis will first help determine which adjectives are most strongly attracted to the construction <moderator + adjective> by quantifying the bidirectional attraction between a moderator and the adjective it modifies. Second, it will help confirm that rather, quite, pretty, and fairly are functional synonyms by revealing significant overlap in the selection of adjectives. The more significant the overlap, the more these four moderators map onto similar content domains, and the stronger the claim that moderators are cognitive synonyms. Cognitive synonyms also display differences. To make these differences apparent, one can look for the distinctive collexemes of each construction. This is where a second method of collostructional analysis is needed: distinctive collexeme analysis (Gries & Stefanowitsch 2004). Insofar as we want to compare four near-synonymous constructions, , , <pretty + adjective>, and , and contrast them in their respective collexeme preferences, we should implement multiple distinctive collexeme analysis (Hilpert 2006; Gilquin 2007). This method differs from simple collexeme analysis because it filters away overlapping collocates and retains only those adjectives that are specific to each moderator. The output makes it possible to classify the distinctive adjectives according to their function and meaning, and to get a better grasp of the individual functional specificities of rather, quite, pretty, and fairly.

3.3

Collexemes as input for multivariate statistics

Moderator collexemes can serve as input for two usage-based techniques whose aim is to capture semantic relations between near-synonyms on the basis of multiple factors: hierarchical cluster analysis and correspondence analysis. Both methods are exploratory. Instead of testing a hypothesis in relation to pre-assigned categorical clusters, multifactorial analyses follow a bottom-up approach. Indeed, clusters are determined by the similarity of the members of the same groupings and their dissimilarity to the members of other groupings. In theory, these methods dispense with the linguist’s preconception of how the data is categorized. But in practice the clustering results in large part from the criteria that the linguist adopts to combine points into clusters. Recent research on lexical near-synonymy in Cognitive Linguistics has made extensive use of cluster analysis (Divjak 2006, 2010; Divjak & Fieller this volume, 405–441; Divjak & Gries 2008). Hierarchical cluster analysis is the generic name of a family of statistical techniques for clustering data (i.e. for structuring observed data into groups) and representing them graphically in a tree-like format. Among these clustering techniques, I use hierarchical agglomerative clustering, whereby individual


data points are successively agglomerated into similar clusters, and similar clusters are merged iteratively into bigger clusters until one last cluster is obtained.11 Results appear in the form of a dendrogram, which facilitates an objective and accurate identification of semantic classes and subclasses for a given lexical set. Hierarchical agglomerative cluster analysis will be used to justify the existence of the paradigm of moderators within the broader paradigm of degree modifiers of adjectives. The collostruction strength of the collexemes of 23 English degree modifiers of adjectives will serve as input to compute the distance matrix. Correspondence analysis is another distance-based clustering technique that represents the structure of cross-tabulations graphically on a multi-dimensional plane (Benzécri 1973; Benzécri 1984; Greenacre 2007; Glynn, this volume, 443–486). In lexical semantics, correspondence analysis offers a convenient way of mapping correlations between lexical items graphically. Like hierarchical cluster analysis, correspondence analysis has become a popular method in Cognitive Semantics (Glynn 2010a). Interpreting a correspondence map is relatively simple: the closer the data points on the map, the stronger the correlation between these data points. However, interpreting a map can also be tricky, since it flattens multi-dimensional distances onto a two-dimensional plane. Nevertheless, correspondence analysis yields promising results in areas such as exemplar-based semantics or cognitive sociolinguistics (Glynn 2010b). Cognitive linguists like the fact that a correspondence map may provide access to the complex conceptual maps that structure language knowledge and language use. Correspondence analysis is used to show how moderators imply a specific construal depending on the adjectives they correlate with. Input is provided by the cross-tabulation of the frequencies of 25 distinctive collexemes for each moderator. The list of 25 distinctive collexemes will be obtained via the multiple distinctive collexeme analysis outlined above. This paper attempts to show that near-synonymy is a complex phenomenon whose study can only benefit from a combination of different statistical methods. These methods range from frequencies and collocations – including (multiple distinctive) collexeme analysis – to exploratory multifactorial techniques – namely cluster analysis and correspondence analysis. If cognitive synonymy is indeed a matter of sameness and difference, only a combination of statistical techniques can provide a sound-enough basis for a complete picture of semantic relations to emerge regarding moderators.

11. Hierarchical divisive clustering proceeds in reverse order: it starts at the root and successively splits the clusters.


4. Results 4.1

Collexeme analysis

Because rather, quite, pretty, and fairly are alternate ways of expressing moderation, we should expect them to display similarities and differences. Each feature will be illustrated in turn, starting with similarities, which will be captured by means of a collexeme analysis. Collexeme analysis generates a ranked list of attracted lexemes (i.e. ‘collexemes’) and rejected lexemes. This list can be used to determine what meanings are congruent with the semantics of the construction and what meanings are incongruent. For our current purpose, we shall focus on congruent meanings.

Table 1.â•‡ Top 10 adjectival collocates of rather, quite, fairly, and pretty in COCA rather

quite

adjective

frequency in corpus

frequency in construction

adjective

frequency in corpus


large different small difficult unusual simple limited good high strange

119,992 â•⁄17,021 165,348 â•⁄â•⁄6,672 â•⁄17,736 â•⁄48,134 â•⁄â•⁄3,533 378,826 191,591 â•⁄23,457

260 231 189 129 116 â•⁄96 â•⁄95 â•⁄89 â•⁄85 â•⁄80

different sure clear good right possible similar ready common simple

â•⁄17,021 137,372 â•⁄81,553 378,826 â•⁄â•⁄4,558 â•⁄88,919 â•⁄60,967 â•⁄55,338 â•⁄63,239 â•⁄48,134

2,247 1,347 â•⁄â•‹805 â•‹â•⁄ 578 â•‹â•⁄ 548 â•‹â•⁄ 458 â•‹â•⁄ 328 â•‹â•⁄ 300 â•‹â•⁄ 278 â•‹â•⁄ 271

fairly

pretty

adjective

frequency in corpus


adjective

frequency in corpus


good easy large common high simple new certain small typical

378,826 â•⁄59,914 119,992 â•⁄63,239 191,591 â•⁄48,134 â•⁄64,824 â•⁄71,149 165,348 â•⁄20,483

346 337 281 278 247 243 202 201 190 154

good sure bad clear big tough cool close hard strong

378,826 137,372 â•⁄90,297 â•⁄81,553 187,641 â•⁄33,746 â•⁄â•⁄3,235 â•⁄94,845 123,894 â•⁄69,137

7,492 1,241 â•‹â•⁄ 743 â•‹â•⁄ 729 â•‹â•⁄ 583 â•‹â•⁄ 486 â•‹â•⁄ 481 â•‹â•⁄ 471 â•‹â•⁄ 436 â•‹â•⁄ 383


Let us postulate that each moderator of adjectives forms a construction that attracts certain adjectival collexemes and rejects others. To calculate the collostruction strength of a given adjective A for a given <moderator + adjective> construction C, collexeme analysis needs four frequencies: the raw frequency of A in C, the raw frequency of A in all other constructions (i.e. ¬C), the frequency of C with adjectives other than A (i.e. ¬A), and the frequency of ¬C with that of ¬A (Stefanowitsch & Gries 2003:â•›218, Hilpert this volume). To generate a ranked list of attracted adjectives based on collostruction strength, this operation is repeated for each adjective that co-occurs with each moderator in the corpus. Table 1, above, summarizes the frequencies of the 10 most frequent adjectives in the four <moderator + adjective> constructions in COCA, i.e. , , and <pretty + adjective>. Each of the four subcomponents of the table are submitted to a collexeme analysis, along with the following information: corpus size: 415M words frequency of : 12,574 frequency of : 29,735 frequency of : 10,834 frequency of <pretty + adj.>: 35,949 Table 2 presents the ten most strongly attracted adjectives of the four moderators based on collostruction strength.12 Table 2.â•‡ Top 10 collexemes of rather, quite, fairly, and pretty rather

quite

adjective

coll. adjective strength

large different unusual small difficult odd remarkable limited vague strange

1025.19 â•⁄709.26 â•⁄705.49 â•⁄521.34 â•⁄480.6 â•⁄436.5 â•⁄415.88 â•⁄413.37 â•⁄400.68 â•⁄384.69

different sure clear possible similar good ready simple remarkable common

fairly coll. adjective strength 15082.43 â•⁄8231.32 â•⁄4923.96 â•⁄2216.18 â•⁄1614.27 â•⁄1482.02 â•⁄1480.81 â•⁄1357.46 â•⁄1297.76 â•⁄1259.48

easy common simple large good straight-forward certain typical high consistent

pretty coll. adjective coll. strength strength 2686.55 2078.88 1883.49 1751.18 1523.3 1433.43 1326.03 1315.09 1250.31 1112.44

good sure clear bad tough cool amazing big close strong

12. It was performed thanks to the script Coll.analysis 3.2 for R (Gries 2007).

61820.34 â•⁄8148.39 â•⁄4764.46 â•⁄4734.36 â•⁄3635.13 â•⁄3628.34 â•⁄2848.64 â•⁄2604.22 â•⁄2531.19 â•⁄2139.59


One might object that there is little difference between the ranked list based on raw frequencies (Table 1) and the ranked list based on collostruction strength (Table 2). However, the latter is more reliable since collostruction strength is the product of an association measure – here Dunning’s log-likelihood ratio – which uses absolute and collocate frequencies to determine which lexemes co-occur more frequently than expected in a given construction. All the above collexemes are attracted to each construction at the very significant level of p < 0.001 since coll.strength > 3.13 Upon inspection, collexeme analysis informs a semantic analysis of moderator constructions. The collexemes of rather split into sets of adjectives that denote spatial dimension (large, small, limited), atypicality (unusual, odd, vague, strange, remarkable), difference, and difficulty. The collexemes of quite split into the following classes: difference/similarity (different, similar), epistemic modality (sure, possible), dynamic modality (ready), positive value (good, clear), (a)typicality (common vs. remarkable), and simplicity. As regards fairly, collexemes divide up into adjectives that denote simplicity (easy, simple, straightforward), typicality (common, typical, consistent), spatial dimension or position (large, high), positive value (good), and epistemic modality (certain). The collexemes of pretty fall into the following sets: positive/negative values (good vs. bad, cool, clear, strong), difficulty (tough), nonstandard identification (amazing), spatial dimension or position (big, close), and epistemic modality (sure). As expected, collexeme analysis reveals that the behavioral profiles of moderators are both similar and different. Similarity is evidenced by the significant degree of overlap in the selection of collexemes. Many adjectives co-occur with more than one moderator (certain, clear, common, different, difficult, good, large, remarkable, simple, sure). Some semantic classes are shared among moderators. For example, both rather and quite are compatible with the expression of (a)typicality. Moderating qualities that belong to the class of spatial extension can be done by means of fairly or rather. The expression of epistemic modality is a feature common to quite, pretty, and fairly. This confirms that moderators are synonyms to some extent. However, at a finer-grained level of analysis, semantic correspondences between moderators are not that straightforward. Even though the expression of atypicality is observed with quite, it is in fact a distinctive feature of rather because adjectives denoting that semantic class are more varied with rather. Even though the expression of epistemic modality is common to the three moderators, it is actually characteristic of quite. The adjective different co-occurs with both rather and quite. However, a quick look at the collostruction strength shows that different is far more distinctive of quite (coll. strength = 15082.43) than it is of rather (coll. strength = 709.26). Good is attracted to pretty, fairly, and quite, but it is definitely more distinctive of pretty

13. As a rule, if collostruction strength is based on p-values, the following equivalences hold true: coll.strength > 3 = p < 0.001; coll.strength > 2 = p < 0.01; coll.strength > 1.3 = p < 0.05.


(coll. strength = 61820.34) than it is of fairly (coll. strength = 1523.3), and quite (coll. strength = 1482.02).14 Collexeme analysis is not the easiest way of spotting differences in collocational preferences because it does not filter away overlapping adjectives and it requires a tedious comparison of collostruction strengths to determine relevant thresholds of attraction. This can be done by means of a multiple distinctive collexeme analysis (MDCA), which contrasts constructions in their distinctive collocational preferences by getting rid of overlapping collexemes. As explained above, this method is best suited for comparing related constructions, preferably alternations. Whether rather, quite, pretty, and fairly are alternative ways of expressing moderation is undeniable. However, the internal structure of the paradigm that these four adverbs belong to is problematic because some of these adverbs (e.g. quite) can be used as maximizers. To make sure the paradigm of moderators is internally coherent and thus to maximize the interpretation of MDCA, one interesting option is to conduct a hierarchical cluster analysis over a pool of 23 degree modifiers in English and see if the four adverbs cluster together on the basis of their preferred collexemes.

4.2

Collexeme analysis as input for hierarchical cluster analysis

Hierarchical cluster analysis describes a range of multifactorial methods for investigating structure in data, with the goal of identifying subgroups of similar objects. Following Gries & Stefanowitsch (2010), I use hierarchical agglomerative clustering (Everitt et al. 2011:â•›Section 4.2) to see how English degree modifiers cluster on the basis of their preferred collexemes. The 23 degree modifiers, which include the four moderators under investigation, are: a bit, a little, absolutely, almost, awfully, completely, entirely, extremely, fairly, frightfully, highly, jolly, most, perfectly, pretty, quite, rather, slightly, somewhat, terribly, totally, utterly, very. Originally, Paradis selected them because they epitomize the degree modifier paradigm in most lexicographic works (1997:â•›15–17). If rather, quite, pretty, and fairly cluster together, then these four adverbs form a homogeneous paradigm despite their multifunctional behavior. For each of the 23 adverb types listed above, I first extracted all adjectival collocates in COCA, amounting to 432 adjective types and 316,159 co-occurrence tokens. Then, I conducted a collexeme analysis for each of the 23 degree modifiers. To reduce the data set to manageable proportions, the 35 most attracted adjectives were selected on the basis of their collostruction strength. For these 23 adverb types and their 432 adjective types, a 23-by-432 co-occurrence table containing the frequency of each adverb-adjective pair type was submitted to a hierarchical agglomerative cluster analysis, which requires a distance object as input. The distance object is a dissimilarity matrix

14. It is also a collexeme of rather, although to a much lesser extent (coll. strength = 34.34).


au bp edge #

88 3 20

JOLLY

ALMOST

FRIGHTFULLY

86 11 72 26 89 44 17 15 14 85 68 6 TERRIBLY

HIGHLY

VERY

MOST

EXTREMELY

QUITE

FAIRLY

83 24 12 88 59 4 91 73 2 PRETTY

A_BIT

SOMEWHAT

A_LITTLE

UTTERLY

SLIGHTLY

ENTIRELY

PERFECTLY

90 64 9 86 43 7 88 73 3

92 12 18

75 36 16 73 48 82 84 10 5

TOTALLY

96 100 1

ABSOLUTELY

88 16 13 81 29 1162 36 8

RATHER

79 7 19

AWFULLY

84 7 21

COMPLETELY

Height

200 250 300 350 400 450 500 550 600

Figure 1.â•‡ Cluster dendrogram of 23 degree modifiers of adjectives in English, clustered according to their adjectival collexemes (distance: Canberra; cluster method: Ward)

that one obtains by converting tabulated frequencies into distances with a user-defined distance measure. When variables are ratio-scaled, the linguist can choose from several distance measures (Euclidean, City-Block, correlation, Pearson, Canberra, etc.).15 For our purpose, the measure of dissimilarity of the adverb types in the columns was computed using the Canberra distance metric, because it handles the relatively large number of empty occurrences best (see Divjak & Gries [2006:â•›37] for further methodological details). Finally, one needs to apply an amalgamation rule that specifies how the elements in the distance matrix get assembled into clusters. Here, clusters were amalgamated using Ward’s method (Ward 1963), which evaluates the distances between clusters using an analysis of variance. This method has the advantage of generating clusters of moderate size.16 Figure 1 shows the resulting dendrogram. The plot should be read from bottom to top. There are three numbers around each node. The number below each node specifies the rank of the cluster (here, from 1 to 21, i.e. from the 1st generated cluster to the 21st). The two numbers above each node

15. For reasons of space, I cannot discuss the reasons why one should prefer a distance measure over another. A description of some distance measures can be found in Gries (2010:â•›313–316). 16. All computations were performed with R 2.13 (R Development Core Team 2011) and the package pvclust (version 1.2-2, www.is.titech.ac.jp/~shimo/prog/pvclust/). This package allows the user to include confidence estimates through multiscale bootstrap resampling, a possibility missing in other packages, such as hclust.


indicate two types of p-values,17 which are calculated via two different bootstrapping algorithms: AU and BP. The number to the left indicates an ‘approximately unbiased’ p-value (AU) and is computed by multiscale bootstrap resampling. The number to the right indicates a ‘bootstrap probability’ p-value (BP) and is computed by normal bootstrap resampling. The number to the left is a much better assessment of how strongly the cluster is supported by the data. In both cases, the closer the number is to 100, the stronger the cluster. AU p-values suggest the clusters we obtain represent the data accurately. Indeed, the plot shows the standard values of most clusters are significantly high, with AU p-values ranging from 79 to 96. An AU p-value of 96 implies that the hypothesis that the cluster is invalid is rejected with a significance level of 0.04. The dendrogram displays several homogeneous clusters:18 a. cluster 19 groups together maximizers; it breaks down into cluster 1 (completely, totally) and cluster 13 (perfectly, absolutely, entirely, utterly); b. cluster 9 groups together diminishers (slightly, a little, a bit, somewhat); c. cluster 12 groups together moderators (rather, pretty, fairly, quite); d. cluster 18 groups together boosters and breaks down into cluster 16 (most, very, extremely, highly), cluster 6 (awfully, terribly), and cluster 14 (frightfully, jolly); the presence of an approximator (almost) within the cluster of boosters (cluster 15) is surprising but may be due to its intensive use as a sentential adverb, more than a modifier of adjectives.19 The cluster analysis based on collexemes yields functionally and semantically motivated groups. As Paradis (1997:â•›27) observed, rather, quite, pretty, and fairly do cluster together under the moderator paradigm (cluster 12) despite their multifunctionality. However, the internal structure of this cluster still needs explaining. It is not clear why fairly and quite cluster together (cluster 2), and why rather is not part of cluster 4, which groups together pretty and cluster 2. Furthermore, the stratification of cluster 12 does not follow their conventional distribution in terms of grading force (rather> quite> pretty> fairly), as found in Paradis (1997:â•›148–155). For now, suffice it is to say that moderators form a functionally coherent class. Performing MDCA to explore internal distinctions is therefore relevant.

17. The term “p-value” is the one that the authors of the pvclust package have adopted. Actually, it seems that these p-values are confidence estimates. 18. In the classification of degree modifiers that follows, I adopt Paradis’s terminology. 19. See Paradis (1997:â•›37) for confirmation.


4.3

Multiple distinctive collexeme analysis

We saw above that rather, quite, pretty, and fairly display similarities and differences. One way to amplify these differences is to conduct a multiple distinctive collexeme analysis. Instead of computing the degree of attraction between a lexical item and a construction, distinctive collexeme analysis contrasts constructions in their respective collocational preferences (Gries & Stefanowitsch 2004). This method has proved useful when it comes to distinguishing minimal semantic and functional differences between near-synonymous constructions (e.g. the ditransitive vs prepositional dative alternation). The input is slightly different from what we have in collexeme analysis. This time, one needs to tabulate the type frequency of the collexeme in the first construction, the type frequency of the same collexeme in the second construction, and the frequencies of the two constructions with words other than the collexeme under investigation (Gries & Stefanowitsch 2004:â•›102). Again, 2x2 tables are submitted to the Fisher-Yates Exact test for each relevant lexeme. However, when one wants to compare more than two constructions and input more complex tables, such as Table 3 below, the Fisher-Yates Exact test cannot be used. Instead, one needs to carry out a one-tailed exact binomial test, and the method goes under the name of multiple distinctive collexeme analysis.20 The same script as the one used for collexeme analysis was used. Below, Tables 4 to 7 list, for each moderator, the ten most distinctive adjectives. MDCA compares the observed frequency of each adjective with its expected frequency. If adjectives were distributed at random over the different moderator constructions, we would not find any significant deviation between observed and expected frequencies because the distribution of each adjective would follow the frequencies of the moderators. For each construction token, the script performs a binomial test Table 3.â•‡ Input for a multiple distinctive collexeme analysis of an adjective in four moderator constructions Construction

Adjective A

Other adjectives

Row totals

rather + adj quite + adj pretty + adj fairly + adj column totals

a c e g a+c+e+g

b d f h b+d+f+h

a+b c+d e+f g+h a+b+c+d+e+f+g+h

20. Gilquin (2007) illustrates how multiple distinctive collexeme analysis determines the verbs that are distinctively associated with the non-finite verb slot of English periphrastic causative constructions.


Table 4.â•‡ The 10 most distinctive adjectives of rather rather + adj

observed frequency

expected frequency

pbin rather

SumAbsDev

odd unusual strange vague difficult simplistic lengthy peculiar bizarre curious formal

â•⁄74 116 â•⁄80 â•⁄56 129 â•⁄28 â•⁄36 â•⁄31 â•⁄51 â•⁄29 â•⁄29

14.51 37.58 25.70 12.71 65.35 â•⁄4.70 â•⁄9.26 â•⁄5.94 14.09 â•⁄5.94 â•⁄5.94

38.99 30.40 21.54 24.75 13.89 18.31 13.81 17.21 17.46 14.91 14.91

56.75 46.39 38.70 35.86 35.86 28.42 27.53 27.08 25.83 25.09 23.49

Table 5.â•‡ The 10 most distinctive adjectives of quite quite + adj

obs freq

exp freq

pbin quite

SumAbsDev

different right possible ready true likely similar capable willing correct

2247 â•⁄548 â•⁄458 â•⁄300 â•⁄235 â•⁄206 â•⁄328 â•⁄152 â•⁄142 â•⁄106

825.30 184.75 150.95 104.02 â•⁄91.23 â•⁄76.46 148.32 â•⁄51.19 â•⁄49.55 â•⁄36.10

Inf 238.72 216.96 120.33 â•⁄70.29 â•⁄69.12 â•⁄66.10 â•⁄66.87 â•⁄56.31 â•⁄45.22

Inf 410.01 374.49 206.00 120.24 118.13 116.57 114.97 â•⁄96.25 â•⁄78.33

Table 6.â•‡ The 10 most distinctive adjectives of fairly fairly + adj

obs freq

exp freq

pbin fairly

SumAbsDev

common easy new constant recent certain typical consistent straightforward regular

278 337 202 117 122 201 154 130 123 â•⁄70

â•⁄83.86 108.56 â•⁄43.38 â•⁄16.63 â•⁄20.36 â•⁄66.03 â•⁄35.55 â•⁄33.26 â•⁄30.73 â•⁄12.05

76.85 83.94 88.58 84.16 72.50 49.00 61.99 45.98 44.94 40.51

145.26 130.39 123.43 123.41 109.92 â•⁄98.80 â•⁄89.24 â•⁄68.47 â•⁄65.25 â•⁄59.78


Table 7.â•‡ The 10 most distinctive adjectives of pretty pretty + adj

obs freq

exp freq

pbin pretty

SumAbsDev

good bad cool tough big hard scary smart amazing close

7731 â•⁄758 â•⁄488 â•⁄489 â•⁄591 â•⁄447 â•⁄212 â•⁄196 â•⁄347 â•⁄498

3613.04 â•⁄343.37 â•⁄214.45 â•⁄226.02 â•⁄295.44 â•⁄225.61 â•⁄â•⁄98.34 â•⁄â•⁄91.32 â•⁄195.03 â•⁄314.03

Inf 201.75 144.61 122.02 113.65 â•⁄83.52 â•⁄52.78 â•⁄48.17 â•⁄44.89 â•⁄40.52

Inf 349.84 252.23 213.70 204.40 143.97 â•⁄93.43 â•⁄83.66 â•⁄82.05 â•⁄77.60

to determine the probability of a particular observed frequency given the expected frequency.21 This probability is then log-transformed. The resulting value (pbin) captures distinctiveness.22 It is used to determine whether a given adjective is distinctive for a particular construction or not, and whether the co-occurrence between the adjective and the moderator construction is statistically significant or not. The co-occurrence is statistically significant if the absolute distinctiveness value is higher than 1.3, p < 0.05. Finally, SumAbsDev gives the sum of all absolute pbin values for a particular adjective. The higher the figure, the more the adjective deviates from its expected distribution. MDCA makes patterns of attraction more visible. It confirms that rather attracts adjectives that denote atypicality/deviation from a norm (odd, unusual, strange, vague, peculiar, bizarre, curious). Additionally, rather attracts adjectives that denote difficulty/simplicity. By far, the most distinctive collexeme of quite is different (pbin and SumAbsDev = infinite). Its antonym (similar) is also among the 10 most distinctive collexemes. Modal meanings are well represented: possible and likely denote epistemic meaning, and ready, capable, and willing denote dynamic meaning. Also distinctive of quite are adjectives that denote factuality (right, true, correct). Fairly attracts some sets that are semantically close, such as typicality (common, typical), similarity/stability (constant, consistent, regular), and epistemicity (certain). Other sets include easiness (easy, straightforward), and time location (new, recent). Lastly, the most distinctive collexeme of pretty is good (pbin and SumAbsDev = infinite). Good belongs to the 21. For instance, the probability to find 74 occurrences of odd in when you would have expected it 14.51 times. 22. It receives a positive sign when the verb occurs more frequently than expected in the construction and a negative sign when the verb occurs less frequently than expected. In short, positive values indicate attracted collexemes whereas negative values indicate repelled collexemes. For reasons of space, I have selected positive values only.


category of positive values, along with cool and smart. Pretty also attracts an antonym such as bad, which denotes a negative value. Other distinctive semantic sets include difficulty (tough, hard), spatial dimension or location (big, close), deviation from a norm (amazing), and psychological stimulus (scary). Since MDCA excludes overlap (i.e. those adjectives which collexeme analysis revealed as common to at least two moderators), it makes some tendencies that collexeme analysis revealed more apparent: a. atypicality as well as difficulty/simplicity are the most distinctive features of rather; b. difference/similarity and modal meanings are the most distinctive features of quite; c. typicality is a distinctive feature of fairly, along with similarity/stability; d. whatever their polarity, value judgments are the most distinctive features of pretty, along with difficulty and dimension/position. But MDCA also reveals tendencies that were harder to grasp with collexeme analysis. Indeed, moderators follow a division of labor in the expression of some functions. Atypicality is a distinctive feature of rather, whereas typicality is a distinctive feature of fairly. Easiness is a distinctive feature of fairly, whereas the expression of difficulty is distinctive of both rather and pretty. There is a difference in register though: it seems that pretty is less formal than rather (rather difficult vs. pretty tough, pretty hard). All in all, MDCA shows that even though moderators are functionally close, they do not profile the same conceptual domains. Even though collexeme analysis and MDCA reveal tendencies that were much harder to capture with only raw frequencies, the above observations are partial because of the limited number of selected collexemes. For a deeper assessment of the synonymy of moderators and the division of labor that they follow, we should increase the level of granularity of our analysis. One obvious solution is to investigate more collexemes, but the more data we have, the more difficult it is to make generalizations. Rather than inspect and compare collostruction-based frequency tables manually, we should also be able to compute and visualize the relative attraction between (a) moderators, (b) adjectives, (c) moderators and adjectives. With this goal in mind, we can use the output of MDCA as input for correspondence analysis.

4.4 Multiple distinctive collexeme analysis as input for correspondence analysis Correspondence analysis (henceforth CA) is an exploratory statistical technique that takes the frequencies of multiway tables as input, then summarizes and visualizes distances between the variables. It determines the probability of global association


Table 8.â•‡ Input for correspondence analysis (sampled) Adjective (distinctive collexeme)

fairly

pretty

quite

rather

able abstract accurate amazing aware awful awkward bad beautiful big …

â•⁄3 14 83 â•⁄6 â•⁄1 â•⁄0 â•⁄0 19 â•⁄1 56 …

â•⁄â•⁄4 â•⁄â•⁄7 â•⁄66 347 â•⁄11 100 â•⁄12 758 â•⁄â•⁄2 591 â•⁄…

â•⁄67 â•⁄â•⁄5 109 â•⁄97 109 â•⁄12 â•⁄â•⁄7 â•⁄32 129 â•⁄45 â•⁄…

â•⁄0 22 â•⁄4 22 â•⁄0 â•⁄6 27 22 21 23 …

between rows and columns, and tests this association using the χ2 test.23 Two rows/ columns will be close to each other if they associate with the columns/rows in the same way. Table 8 shows a sample of the input used for CA. It brings together the 25 most distinctive collexemes of each moderator and the raw frequency of each collocation type. The whole table contains 400 cells. CA uses these frequencies to compare (a) line profiles, i.e. adjectives, (b) column profiles, i.e. moderators, (c) line profiles and column profiles, i.e. moderators and adjectives. This method reintroduces overlap because the table contains raw frequencies of adjectives that co-occur with at least two moderators. As we saw above, overlap is a characteristic of the paradigm of moderators. Taking overlap into account is therefore a way of mapping co-occurrence patterns more realistically than if we simply ignored it. CA transposes the multidimensional distances to a two-dimensional plane that maps the correlations between the variables. More precisely, it transforms the input table (i.e. a table of numerical information) into a graphic display in which each row and each column is represented as a point in a Euclidean space. Figure 2, below, is the graphic output of CA.24

23. Since CA is an exploratory technique, one does not need to check whether the conditions of use of χ2-statistics are met. For our current purpose, the hypothesis of independence can be rejected because χ2 = 33623.82, df = 297, and p-value < 2.2e-16. 24. To conduct CA and output the graph, I used R with the packages FactoMineR (http:// cran.r-project.org/web/packages/FactoMineR/index.html) and dynGraph (http://cran.rproject.org/web/packages/dynGraph/index.html).


Correspondence analysis graph 2,0

Dimension2 (28.14%)

1,5

1,0

0,5

0,0

–0,5

–1.5

–0.75

0

0.75

1.5

Dimension1 (52.77%)

Figure 2.â•‡ CA biplot of the <moderator + adjective> construction in COCA

The plot is built along two axes, which are the principal axes of inertia.25 Their intersection defines the average profile of all the points in the cloud. CA decomposes the overall inertia by identifying a small number of representative dimensions. Each axis corresponds to a dimension. The plot displays only two dimensions, which are selected according to their eigenvalues. The eigenvalue of a dimension measures how much information is present along the axis of that dimension. The first axis (dimension 1, eigenvalue = 0.533) represents 52.77% of the inertia, whereas the second axis (dimension 2, eigenvalue = 0.284) represents 28.14% of the inertia. There is a third dimension, whose eigenvalue is 0.193. Even though dimension 3 accounts for 19.08% of the inertia, it is not taken into account in the plot. This is not a problem because the first two dimensions already explain 80.91% of the information contained in the input table, and the results can be interpreted with enough accuracy without dimension 3. Because the plot contains a lot of data, we should examine each dimension in turn. On the horizontal axis, dimension 1 contrasts pretty and quite. Each of them attracts 25. In CA, “inertia” is very similar to the “moment of inertia” in applied mathematics. It measures the total variance of the data table.


Correspondence analysis graph

1,5

Dimension2 (27.06%)

1,0

0,5

0,0

–0,5

–0.5

0

0.5

0.75

1.5

Dimension1 (57.51%)

Figure 3.â•‡ CA biplot of the <moderator + adjective> construction in COCA (with semantic annotation)

its own cloud of adjectives, and each cloud is clearly delimited. On the vertical axis, dimension 2 opposes fairly and rather (at the top of the cloud) to pretty and quite (at the bottom of the cloud). This goes against the implicit assumption that moderators split up between rather and quite on the one hand, and pretty and fairly on the other hand (see, for example, Downing & Locke 2006). The proximity between fairly and rather is evidenced by the continuum formed by their distinctive collexemes (both horizontally and vertically). Comparatively, only two adjectives (surprising and difficult) stand halfway between rather and quite. In all likelihood, the boundary between fairly and rather can be drawn above a cluster of adjectives with negative connotations (obscure, crude, mundane, lengthy, simplistic), which are distinctive of rather. At this stage, it is still difficult to spot any division of labor among moderators because of the granularity of the plot. Figure 3, above, presents the graphic output of CA once all adjectives have been semantically annotated. The relative position of moderators in the cloud is very similar to the configuration displayed in Figure 2.


Annotating adjectives makes it considerably easier to identify the functional specificities of each moderator as well as the division of labor among them. The specificities of each moderator are listed below: rather: dimension or position in space (ex. long, high), atypicality/oddity (ex. odd, bizarre), negative attitudes (ex. ironic), unclearness (ex. vague, obscure); quite: epistemic, dynamic, and factual meanings (ex. likely, able, true), difference (ex. different, separate), psychological states (ex. surprised, concerned, content); fairly: location in time (ex. recent, new), typicality (ex. typical, common, standard); pretty: appreciative and unappreciative values (ex. good, great vs. bad, awful), cleverness and stupidity (ex. smart vs. stupid, dumb), difficulty (ex. difficult, tough, hard), psychological stimuli (ex. scary, funny). The above list shows that moderators follow a division of labor in the intensification of some complementary meanings: – rather modifies spatial location and atypicality whereas fairly modifies time location and typicality; – rather modifies the expression of negative attitude whereas quite modifies the expression of positive attitude; – pretty modifies the expression of difficulty whereas fairly and rather modify the expression of simplicity; – pretty modifies the expression of psychological stimuli whereas quite modifies the expression of psychological states; – quite modifies the expression of difference whereas fairly and rather modify the expression of similarity/stability. Lastly, some meanings are not distinctive of any moderator in particular: – modifying the degree of surprise/salience and atypicality/extraordinariness is common to pretty, rather, and quite; – modifying the degree of simplicity and similarity/stability is common to both fairly and rather. To summarize, we have three major configurations: – first configuration: moderators operate within one conceptual content; – second configuration: two complementary aspects of a conceptual domain are intensified by two distinct moderators; – third configuration: one conceptual content can be intensified indiscriminately by different moderators.


5. Discussion and conclusion In this paper, we have proposed and combined several statistical methods to provide a bidirectional semantic modeling of the <moderator + adjective> construction. We have made three points. Firstly, we have reasserted the need for better statistics in the collocation-based study of degree modifiers. Collostructional analysis is superior to most techniques based on raw counts and/or percentages because it filters away co-occurring pairs that are unrealistically too frequent or too rare, regardless of the size of the corpus. Secondly, we have shown that combining univariate and multivariate statistics can help map usage patterns and conceptual structure in a set of near-synonyms. The relationship between moderators and adjectives is indeed bidirectional, and it can be represented spatially. Thirdly, my results partly support Paradis (1997) regarding the cognitive synonymy of moderators, which are both similar and different. Moderators are similar because they have a functional basis in common, namely modifying the degree of a property denoted by an adjective. Moderators are also different because they do not modify the same classes of adjectives. In Cognitive Grammar terms, moderators do not always operate within the same conceptual domains. If they do, they follow a division of labor. These findings are of great significance to the study of constructions since two items that co-occur significantly are likely to be entrenched as a constructional unit.26 Figure 2 shows that some pairs are more entrenched than others. For example, quite surprised is more entrenched than quite surprising; pretty crazy is more entrenched than pretty silly; rather vague is more entrenched than rather abstract; and fairly straightforward is more entrenched than fairly easy. Once a pairing of lexemes is sufficiently entrenched, it is likely to acquire a meaning/function of its own. Figure 3 shows that the division of labor of moderators is not limited to the expression of intensification. It includes the expression of various meanings, such as the expression of modality, value judgments, dimension, position in time or in space, etc. Cognitive Construction Grammar takes an inventory approach to the mental representation of grammar. Such an approach assumes that grammar predominantly stores language structure in a complex constructional network instead of building structure “on demand”. In a section on partial productivity, Goldberg postulates that the candidates for the verb slot in the ditransitive construction are stored in speakers’ memories as similarity clusters on the basis of their type frequencies (1995:â•›133–136). The higher the type frequency, the bigger the cluster, and the more productive the 26. However we should be wary of not establishing too strong a correspondence between high frequencies and entrenchment (Geeraerts 2000). Studies in cognitive semantics have shown that what determines the entrenchment of a linguistic unit is not so much its high frequency as its absolute frequency as its frequency of occurrence relative to the frequency of similar units in similar contexts (Geeraerts, Grondelaers & Bakema 1994). Also, some linguistic units are entrenched not because they occur frequently, but because they are salient (Schmid 2007, 2010).


verb class is. In an effort to map usage patterns, Goldberg provides a two-dimensional representation where verbs cluster spatially according to similarity (1995:â•›135). Accordingly, give, pass, bequeath or grant are good candidates for the verb slot in the ditransitive construction, whereas envy or forgive are poor (but by no means impossible) candidates. Goldberg’s map is a theoretical abstraction because it is not based on actual corpus data or similarity metric, as opposed to Figure 2 and Figure 3 in Section 4. Although the latter bear on a different case study, they can be considered as corpus-driven and statistically grounded extensions of Goldberg’s graphic intuition. Figure 2 is flexible enough to represent both entrenched collocations (e.g. rather vague, quite different, fairly new, pretty good) and collocations that are improbable, yet possible (e.g. rather neat, quite cool, fairly stupid, pretty right). It reflects the fact that speakers tend to use certain adjectives with certain degree-modifiers, but can also extend moderators idiosyncratically to other classes of adjectives. The existence of dense, neat clusters of adjectives around pretty and quite suggests that speakers are more conservative in their use of adjectives with these two moderators. Figure 3 confirms that adjectives cluster around moderators on the basis of semantic similarity. In sum, I have presented evidence that shows that types of the <moderator + adjective> construction form a network structured by similarity clusters. Recent studies on near-synonymy in the Cognitive Linguistic framework have concluded that multifactorial techniques can help map usage patterns (Glynn 2010b) and spot “clusters in the mind” (Divjak & Gries 2008; Divjak 2010). Given how representative the corpus I have used is, the same kind of conclusions can be drawn regarding the <moderator + adjective> construction, pending experimental verification. Perhaps in comparison to other less macroscopic approaches, the results presented in this paper seem conditional. Nevertheless, these results take usage-based representations of linguistic units seriously and are verifiable. One can test these findings with a different corpus and compare the results. Confirmatory statistics such as logistic or log-linear regression will have to corroborate the claim that these findings are not due to chance and provide a faithful representation of the reality of the data. For reasons of space, I have deliberately left aside three aspects of the <moderator + adjective> construction, some of which have received much attention in the past, such as syntactic idiosyncrasies (Allerton 1987; Gilbert 1989), grading force (Paradis 1997), and subjectivity (Nevalainen & Rissanen 2002; Athanasiadou 2007). However, I believe that the methodology presented in this paper can shed new light on each of these aspects. Regarding syntactic idiosyncrasies, distinctive collexemes can be used as input for correspondence analysis to obtain a clearer picture of alternations such as vs. , vs. , or vs. . Regarding grading force, adjectives can be annotated according to gradability in correspondence analysis following the categories proposed in Paradis (1997:â•›49), namely non-gradable, scalar, extreme, limit. Regarding subjectivity, the methodology I have proposed can be applied to a diachronic corpus (see


also Hilpert 2006), so as to conduct a quantitative assessment of subjectification over the long-term history of rather, quite, fairly, and pretty. I have made use of multifactorial methods to visualize data that is not properly speaking multifactorial. Instead of clustering lexical co-occurrence data alone, it will be in the interest of future research to expand what I have presented in Figure 3 and integrate more strata of richly annotated data within the same plot (e.g. information concerning grading force, boundedness, and semantic classes of adjectives). The resulting two-dimensional map will provide a finer-grained representation of the scale of synonymy of moderators. It will also help explain why collostructions featuring the same adjective do not have the same connotation depending on which moderator is used.27 To summarize, using multiple distinctive collexemes as input for correspondence analysis has three assets. First, it enables the linguist to ignore collocates that would be frequent or infrequent whatever the context (because they have a high overall frequency throughout the corpus) and focus on relevant pairs. Once distinctive collexemes have been identified, their raw frequencies can be used safely in a multi-way table to map correlations between moderators and adjectives by means of a multivariate statistical technique. Second, one graphic output is enough to synthesize similarities and differences in a set of near-synonyms. Similarities between moderators (e.g. rather and fairly) are evidenced by their relative proximity on the map. Differences (e.g. pretty vs. quite) are made apparent by the relative distance between items. This is also true of adjectives, which tend to cluster according to meaning. Third, visualizing distinctive collexemes in a correspondence analysis plot can do more than depict proximities and distances within separate paradigms. It is also a potentially accurate means of determining entrenchment continua along the two dimensions that structure the Euclidean space. Hopefully, the methodology I have proposed can be used to represent the complex inventory of constructions that shapes speakers’ grammars.

References Allerton, D. J. (1987). English intensifiers and their idiosyncrasies. In R. Steele & T. Threadgold (Eds.), Language topics: Essays in honour of Michael Halliday (pp. 15–31). Amsterdam & Philadelphia: John Benjamins.

27. For example, even though good is a distinctive collexeme of pretty, raw frequencies show that it also co-occurs with fairly and quite. Presumably, it does not have the same meaning in each construction. It will be in the interest of future research to include context-based variables to clarify these variations in meaning.


Altenberg, B. (1991). Amplifier collocations in spoken English. In S. Johansson & A. Stenström (Eds.), English computer corpora: Selected papers and research guide (pp. 127–149). Berlin & New York: Mouton de Gruyter. Athanasiadou, A. (2007). On the subjectivity of intensifiers. Language Sciences, 29, 554–565. DOI: 10.1016/j.langsci.2007.01.009 Benzécri, J-P. (1973). L’analyse des données, 2. L’analyse des correspondances. Paris: Dunod. Benzécri, J.-P. (1984). Analyse des correspondances, exposé élémentaire (2nd ed.). Paris: Dunod. Biber, D., Conrad, S., & Reppen, R. (1998). Corpus linguistics – Investigating language structure and use. Cambridge: Cambridge University Press. DOI: 10.1017/CBO9780511804489 Bolinger, D. L. M. (1972). Degree words. The Hague: Mouton. DOI: 10.1515/9783110877786 Church, K. W., & Hanks, P. (1990). Word association norms, mutual information, and lexicography. Computational Linguistics, 16, 22–29. Cruse, D. A. (1986). Lexical semantics. Cambridge: Cambridge University Press. Davies, M. (1990–present). The Corpus of Contemporary American English (COCA): 410 + million words. http://corpus.byu.edu/coca Divjak, D. (2006). Ways of intending: Delineating and structuring near-synonyms. In St. Th. Gries & A. Stefanowitsch (Eds.), Corpora in Cognitive Linguistics: Corpus-based approaches to syntax and lexis (pp. 19–56). Berlin & New York: Mouton de Gruyter. Divjak, D. (2010). Structuring the lexicon: A clustered model for near-synonymy. Berlin & New York: Mouton de Gruyter. Divjak, D., & Gries, St. Th. (2006). Ways of trying in Russian: Clustering behavioral profiles. Corpus Linguistics and Linguistic Theory, 2, 23–60. DOI: 10.1515/CLLT.2006.002 Divjak, D., & Gries, St. Th. (2008). Clusters in the mind? Converging evidence from near synonymy in Russian. The Mental Lexicon, 3, 188–213. DOI: 10.1075/ml.3.2.03div Downing, A., & Locke, P. (2006). English grammar: A university course (2nd ed.). London: Routledge. Edmonds, P., & Hirst, G. (2002). Near-synonymy and lexical choice. Computational Linguistics, 28, 105–144. DOI: 10.1162/089120102760173625 Everitt, B. S, Landau, S., Leese, M., & Stahl, D. (2011). Cluster Analysis (5th ed.). Oxford: Wiley-Blackwell. DOI: 10.1002/9780470977811 Evert, S. (2004). The statistics of word cooccurrences: Word pairs and collocations. Unpublished doctoral dissertation. Institut für maschinelle Sprachverarbeitung. University of Stuttgart. Firth, J., R. (1957). A synopsis of linguistic theory, 1930–1955. In J. R. Firth (Ed.), Studies in linguistic analysis. Special volume of the Philological Society (pp. 1–32). Oxford: Blackwell. Geeraerts, D. (2000). Salience phenomena in the lexicon: A typology. In L. Albertazzi (Ed.), Meaning and Cognition (pp. 79–101). Amsterdam & Philadelphia: John Benjamins. Geeraerts, D., Grondelaers, S., & Bakema, P. (1994). The structure of lexical variation: Meaning, naming, and context. Berlin, New York: Mouton de Gruyter. Gilbert, E. (1989). Quite, rather. Cahiers de recherche en grammaire anglaise, 4, 4–61. Gilquin, G. (2007). The verb slot in causative constructions. Finding the best fit. Constructions, SV1-3/2006. www.elanguage.net/journals/index.php/constructions/article/view/18/23 Glynn, D. (2010a). Corpus-driven Cognitive Semantics. An introduction to the field. In D. Glynn & K. Fischer (Eds.), Corpus-driven Cognitive Semantics: Quantitative approaches (pp. 1–42). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110226423.1 Glynn, D. (2010b). Synonymy, lexical fields, and grammatical constructions. A study in usage-based Cognitive Semantics. In H. Schmid & S. Handl (Eds.), Cognitive foundations of linguistic usage-patterns: Empirical studies (pp. 89–118). Berlin & New York: Mouton de Gruyter.


Goldberg, A. E. (1995). Constructions: A construction grammar approach to argument structure. Chicago: University of Chicago Press. Goldberg, A. E. (2003). Constructions: A new theoretical approach to language. Trends in Cognitive Sciences, 7, 219–224. DOI: 10.1016/S1364-6613(03)00080-9 Goldberg, A. E. (2006). Constructions at work: The nature of generalization in language. Oxford: Oxford University Press. Goldberg, A. E. (2009). The nature of generalization in language. Cognitive Linguistics, 20, 201– 224. DOI: 10.1515/COGL.2009.013 Greenacre, M. J. (2007). Correspondence analysis in practice (2nd ed.). Boca Raton: Chapman & Hall/CRC. DOI: 10.1201/9781420011234 Gries, St. Th. (2007). Coll.analysis 3.2. A program for R for Windows 2.x. Gries, St. Th. (2010). Statistics for linguistics with R: A practical introduction. Berlin & New York: Mouton de Gruyter. Gries, St. Th., & Stefanowitsch, A. (2004). Extending collostructional analysis: A corpus-based perspective on ‘alternations’. International Journal of Corpus Linguistics, 9, 97–129. DOI: 10.1075/ijcl.9.1.06gri Gries, St. Th. & Stefanowitsch, A. (2010). Cluster analysis and the identification of collexeme classes. In J. Newman, & S. Rice (Eds.), Empirical and experimental methods in cognitive/ functional research (pp. 73–90). Stanford, CA: CSLI. Gries, St. Th. & Stefanowitsch, A. (2006). Corpora in Cognitive Linguistics: Corpus-based approaches to syntax and lexis. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110197709 Hilpert, M. (2006). Distinctive collexeme analysis and diachrony. Corpus Linguistics and Linguistic Theory, 2, 243–256. DOI: 10.1515/CLLT.2006.012 Hirst, G. (1995). Near-synonymy and the structure of lexical knowledge. In J. Klavans (Ed.), AAAI Symposium on Representation and Acquisition of Lexical Knowledge: Polysemy, Ambiguity, and Generativity (pp. 51–56). Cambridge (Mass.): AAAI Press. Kay, P. (2013). The limits of (Construction) Grammar. In T. Hoffmann & G. Trousdale (Eds.), The Oxford Handbook of Construction Grammar (pp. 32–48). Oxford: Oxford University Press. Kennedy, G. (2003). Amplifier collocations in the British National Corpus: Implications for English language teaching. TESOL Quarterly, 37, 467–487. DOI: 10.2307/3588400 Kilgarriff, A. (2001). Comparing corpora. International Journal of Corpus Linguistics, 6, 97–133. Kilgarriff, A. (2005). Language is never, ever, ever, random. Corpus Linguistics and Linguistic Theory, 1, 263–276. Langacker, R. W. (1987). Foundations of Cognitive Grammar. Vol. 1. Theoretical prerequisites. Stanford: Stanford University Press. Langacker, R. W. (1991). Foundations of Cognitive Grammar. Vol. 2. Descriptive application. Stanford: Stanford University Press. Langacker, R. W. (2008). Cognitive Grammar: A basic introduction. Oxford: Oxford University Press. Langacker, R. W. (2009). Cognitive (construction) grammar. Cognitive Linguistics, 20, 167–176. Leech, G. (1997). Introducing corpus annotation. In R. Garside, G. Leech, & T. McEnery (Eds.), Corpus annotation: Linguistic information from computer text corpora (pp. 1–18). London: Longman.


Lorenz, G. (2002). Really worthwhile or not really significant? A corpus-based approach to the delexicalization and grammaticalization of intensifiers in Modern English. In I. Wischer, & G. Diewald (Eds.), Speech, place, and action: Studies in deixis and related topics (pp. 143–161). New York: Wiley. Manning, C. D., & Schütze, H. (1999). Foundations of statistical natural language processing. Cambridge, MA: MIT Press. Nevalainen, T., & Rissanen, M. (2002). Fairly pretty or pretty fair? On the development and grammaticalization of English downtoners. Language Sciences, 24, 359–380. Paradis, C. (1994). Compromisers – a notional paradigm. Hermes, 13, 157–167. Paradis, C. (1997). Degree modifiers of adjectives in spoken British English. Lund: Lund University Press. Paradis, C. (2000). It’s well weird: Degree modifiers of adjectives revisited: The nineties. In J. M. Kirk (Ed.), Corpora galore: Analyses and techniques in describing English (pp. 147–160). Amsterdam & Atlanta: Rodopi. Paradis, C. (2008). Configurations, construals and change: Expressions of degree. English Language and Linguistics, 12, 317–343. Pedersen, T. (1996). Fishing for exactness. In Proceedings of the South-Central SAS Users Group Conference (pp. 188–200). Austin, TX. Quine, W. V. O. (1951). Main trends in recent philosophy: Two dogmas of empiricism. The Philosophical Review, 60, 20–43. Quirk, R., Greenbaum, S., Leech, G., & Svartvik, J. (1985). A comprehensive grammar of the English language. London & New York: Longman. R Development Core Team. (2011). R: A language and environment for statistical computing. Vienna: Austria: R Foundation for Statistical Computing. Schmid, H.-J. (2007). Entrenchment, salience, and basic levels. In D. Geeraerts & H. Cuyckens (Eds.), The Oxford handbook of Cognitive Linguistics (pp. 117–138). Oxford: Oxford University Press. Schmid, H.-J. (2010). Does frequency in text instantiate entrenchment in the cognitive system? In D. Glynn & K. Fischer (Eds.), Quantitative methods in cognitive semantics: Corpus-Â� driven approaches (pp. 101–133). Berlin, New York: Mouton De Gruyter. Simon-Vandenbergen, A.-M. (2008). Almost certainly and most definitely: Degree modifiers and epistemic stance. Journal of Pragmatics, 40, 1521–1542. Sinclair, J. (1991). Corpus, concordance, collocation. Oxford: Oxford University Press. Stefanowitsch, A., & Gries, St. Th. (2003). Collostructions: Investigating the interaction of words and constructions. International Journal of Corpus Linguistics, 8, 209–243. Stoffel, C. (1901). Intensives and downtoners: A study in English adverbs. Heidelberg: Carl Winter. Storjohann, P. (2009). Plesionymy: A case of synonymy or contrast? Journal of Pragmatics, 41, 2140–2158. Traugott, E. C. (2008). The semantic development of scalar focus modifiers. In A. van Kemenade & B. Los (Eds.), The handbook of the history of English (pp. 335–359). Oxford: Blackwell. Ward, J. H. (1963). Hierarchical grouping to optimize an objective function. Journal of the American Statistical Association, 58, 236–244. Wiechmann, D. (2008). On the computation of collostruction strength: Testing measures of association as expressions of lexical bias. Corpus Linguistics and Linguistic Theory, 4, 253–290.

A case for the multifactorial assessment of learner language The uses of may and can in French-English interlanguage Sandra C. Deshors and Stefan Th. Gries

New Mexico State University / University of California, Santa Barbara

In this study, we apply Gries and Divjak’s Behavioral Profile approach to compare native English can and may, learner English can and may, and French pouvoir. We annotated over 3,700 examples across three corpora according to more than 20 morphosyntactic and semantic features and we analysed the features’ distribution with a hierarchical cluster analysis and a logistic regression. The cluster analysis shows that French English learners build up fairly coherent categories that group the English modals together followed by pouvoir, but that they also consider pouvoir to be semantically more similar to can than to may. The regression strongly supports learners’ coherent categories; however, a variety of interactions shows where learners’ modal use still deviates from that of native speakers. Keywords: Behavioral Profiles, hierarchical cluster analysis, logistic regression, modal verbs

1. Introduction and overview Acquiring a foreign language is one of the most cognitively challenging tasks, given how languages differ in every level of linguistic analysis. From a cognitively and psycholinguistically-oriented perspective, learning a language requires identifying a very large amount of co-occurrence data – tense t and number n require subject-verb agreement with morpheme m, idiom i consists of word w and word x, communicative function f is communicated with intonation curve c, etc. – as well as storing and retrieving them. Crucially, these types of co-occurrences are typically probabilistic only rather than absolute/deterministic and, thus, hard to discern and learn: usually, learners need to cope with many-to-many mappings between forms and functions,

180 Sandra C. Deshors and Stefan Th. Gries

and often it is only the confluence of differently predictive information on several levels of linguistic analysis that narrows down the search for a particular meaning (in comprehension) or a particular form (in production). In the Competition Model by Bates and MacWhinney (1982, 1989), for example, this situation is modeled on the assumption that forms and functions are cues to functions and forms, respectively, and many different cues of different strengths, validities, and reliabilities must be integrated to, say in production, arrive at natural-sounding choices. Semantics is a particularly tricky linguistic domain in this regard, in native language, but even much more so in foreign language learning. Not only do languages often carve up semantic space very differently (so that the categories of the language acquired first will influence category formation in the following), but semantic differences are also often much less explicitly noticeable (than, say, the presence or absence of a plural morpheme), which makes the identification of probabilistic co-occurrence patterns all the more difficult. In order to allow for a precise description of semantic, or more generally functional, characteristics of synonyms, antonyms, and senses of polysemous words, Gries and Divjak developed the so-called Behavioral Profile (BP) approach (cf. Gries and Divjak 2009). This approach, to be discussed in more detail below, is highly compatible with a psycholinguistic perspective of the type outlined above and involves a very fine-grained annotation of corpus data as well as their statistical analysis. The method of behavioral profiles has been successfully employed in a variety of contexts – synonyms, antonyms, and word senses of polysemous words have been studied both within one L1 or across two different L1s – as well as having received first experimental support, but so far there have been no studies that test the BP approach’s applicability to L1 and L2 data, which is what we will undertake here. The semantic domain we will explore is one that has proven particularly elusive, namely, modality. While many semantic phenomena can be clearly delineated and, to some degree, explained by the linguistic analyst, modality has been much more problematic; in fact, even the scope of the notion of modality has not really been agreed upon yet. In this chapter, we specifically focus on the semantic domain of possibility as reflected in: – the choices of can vs. may in essays written by native speakers of English; – the choices of can vs. may in essays written by French learners of English;1 – the use of pouvoir in essays written by native speakers of French. In Section 2, we discuss in what sense these modals pose a particular challenge to the analyst as well as present previous corpus-based work on can and may and highlight 1. Following Bartning (2009), the term “advanced learner” is henceforth assumed to refer to “a person whose second language is close to that of a native speaker, but whose non-native usage is perceivable in normal oral or written interaction” (Hyltenstam et al. 2005:â•›7, cited in Bartning 2009:â•›12).

A case for the multifactorial assessment of learner language 181

some of the shortcomings of such work. In Section 3, we discuss the BP approach in general as well as our own data and methods in particular. Section 4 presents the results of our exploration, and Section 5 concludes the chapter.

2. Setting the stage 2.1

What is problematic about the modals?

As near synonyms in the domain of modality, may and can have fueled much theoretical debate with regard to their semantic relations. As a pair, both forms have overlapping semantics which cover simultaneously the meanings of possibility, permission and ability (cf. Collins 2009). This means that both forms can be used to express epistemic, deontic and dynamic types of possibility. It follows that the semantic investigation of may and can triggers two problematic questions: first, to what extent the various senses of each form can be distinguished, and second, to what degree both forms are semantically equivalent? With regard to the first question, studies such as Leech (1969) and Coates (1983) have illustrated the difficulty in distinguishing between the senses of may and can. Leech (1969:â•›76), for instance, notes that “[t]he permission and possibility meanings of may are close enough for the distinction to be blurred in some cases”. Similarly, Coates (1983:â•›14) identifies a “continuum of meaning” – i.e. gradience – in which possible modal uses shade into each other. In the case of the meanings of can, for instance, Coates notes that while permission and ability correspond to the core of two largely intersecting fuzzy semantic sets, possibility, on the other hand, is found “in the overlapping peripheral area” (p. 86). With regard to the issue of the semantic equivalence of may and can, the literature reveals similarly debated standpoints. While some studies recognize the similarities of the two forms, others do not. In the former case, for instance, Collins (2009:â•›91) states that “[t]he two modals of possibility may and can, share a high level of semantic overlap” (despite their differing frequency of occurrence and different degrees of formality), and Leech (1969:â•›75) notes that “[i]n asking and giving permission, can and may are almost interchangeable”. Conversely, studies such as Coates (1983) have clearly distinguished the two forms. For instance, while Coates (1983) does recognize that the English modals share certain meanings and can be organized into semantic clusters, she generally denies the synonymy of may and can by classifying the two forms into two distinct semantic groups. Although she accepts that the two forms may have overlapping meanings in some cases, she claims that even then, the two forms do not occur in free variation. The occurrence of one form over the other has been shown to be influenced, to some extent, by its linguistic context. It has indeed been illustrated that particular


co-occurring grammatical categories interfere with the interpretation of the modals. Leech (2004:â•›77), for instance, notes that certain uses of may are only to be found in particular grammatical contexts: “only the permission sense, for instance, is found in questions (…) and the negation of the possibility sense is different in kind from the negation of the permission sense”. Generally, several grammatical categories have been recognized as interacting with the uses of may and can. While negation is one category that has commonly been identified (cf. Hermerén 1978; Palmer 1979; Coates 1980, 1983; De Haan 1997; Huddleston 2002; Radden 2007; Byloo 2009), voice and sentence types have also been shown to have similar influences on the forms. Overall, the above-mentioned studies all provide clear illustrations of the complexity of the semantic relations between may and can on the basis of empirically gathered evidence. However, they all tend to be based on generalized observations of idiosyncratic behavioral tendencies. In that respect, they all raise the issue of how to provide a more systematic account of the modals’ semantic characteristics and how to integrate qualitative findings into a quantitative and empirically-grounded approach.

2.2

Previous corpus-based work on the modals

2.2.1 Native English As already mentioned above, Hermerén (1978) has shown that the semantics of the modals in native English are morphosyntactically motivated to a considerable degree such that linguistic categories such as voice, grammatical person, type of main verb (action, state, etc.), aspect and sentence type influence the interpretation of the modals: “if these categories can be shown to modify the meaning of the modal […] it is important that this should be accounted for in the description of the semantics of the modals” (p. 74). While this claim calls for empirical validation, one implication of Hermerén’s (1978) argument is that the quantitative study of modal forms will require a powerful and versatile methodological approach. In a very similar fashion, Klinge and Müller (2005:â•›1) argue that, to capture the essence of modal meaning, “it seems necessary to cut across the boundaries of morphology, syntax, semantics and pragmatics and all dimensions from cognition to communication are involved”. A second corpus-based study of the modals in native English is Gabrielatos and Sarmento (2006). This study illustrates an attempt to account for syntactic contextual information while using a quantitative corpus-based approach to investigate core English modals (i.e. can, could, may, might, must, shall, should, will and would). Although their study does not involve the comparison of English varieties, it presents, however, a comparative analysis of the frequencies of uses of the modals in an aviation corpus and a representative corpus of American English. Generally, it raises the following questions:


– To what degree do syntactic structures and modal forms interact contextually? – To what degree does such interaction affect investigated modal forms semantically? – How can such interaction be quantitatively investigated in a corpus including cross-linguistic and interlanguage data? The authors acknowledge that the modals’ distribution varies as a function of their syntactic contexts and they show that frequencies of occurrence of core English modals reflect the type of syntactic environment in which they feature: “there is a great deal of variation in the use of modal verbs and the structures they occur in, depending on the context of use” (p. 234). However, their lack of a suitable cognitively-motivated theoretical framework prevents them from providing a meaningful interpretation of the data and to further explore their findings. To this date, Collins (2009:â•›1) presents: the largest and most comprehensive [study] yet attempted in this area [modality] based on an analysis of every token of the modals and quasi-modals (a total of 46,121) across the spoken and written data.

Collins (2009) investigates the meanings of the modals in three parallel corpora of contemporary British English, American English and Australian English. Despite the author’s recognition that a corpus quantitative approach “typically combined with a commitment to the notion of ‘total accountability’ may influence hypotheses applied to the data, or formulated on the basis of it” (p. 5) and despite the large size of his data set, his analysis is of limited informative value due to: – a theoretical framework that does not allow for the full exploitation of the linguistic context of the modals, and; – a statistical approach that inhibits rather than unveils linguistic patterns at play in the data. With regard to the first point, Collins (2009) restricts his approach to the identification of the forms’ lexical meanings. His theoretical framework consists of a traditional tripartite taxonomy including epistemic, deontic and dynamic senses. Regrettably, while he recognizes that some uses of the modals can yield preferences for particular syntactic environments, his analysis does not address that fact in a systematic quantitative fashion. As for the second point, while, statistically, Collins (2009) limits his investigation to providing frequency tables of modal forms, his overall approach is problematic because it is based on the erroneous assumption that the frequent occurrence of a modal form warrants its linguistic relevance. In the case of may and can, for instance, Collins uses raw frequencies to show that deontic may is the “least common” sense of the three as it is chosen 7% of the time over epistemic may (79%) and dynamic may (8.1%). However, he does not show whether the (low) frequency of deontic may is significantly different from the also low frequency of dynamic may, and our


analysis of his data shows that, excluding the indeterminate cases, the distribution of may’s senses across the American, Australian, and British data is highly significant (χ2 = 42.68; df = 4; p < 0.001). This, in turn, raises the questions of: – To what extent are Collins’ (2009) frequencies of the occurrences of modal forms in each corpus comparable? – Since the observed frequency discrepancies are not a matter of chance, then what motivates, linguistically, the different uses of each form in each independent corpus? So in sum, while studies such as Gabrielatos and Sarmento (2006) and Collins (2009) provide many descriptive results, they are often merely or largely form-based alone and are lacking in terms of determining which of the many frequencies are statistically and/or linguistically relevant. As a result, such studies do not come close to allow us to develop a characterization of modals that essentially allows us to classify/predict modal use.

2.2.2 Learner English and contrastive approaches From a cross-linguistic and an interlanguage perspective, investigating the modals raises two related issues, namely (i) the possibility of a lack of (direct) semantic equivalence between the modal forms in the learner’s native language (L1) and his/her target language (L2), and (ii), the fact that such cross-linguistic semantic dissimilarity will affect the uses of the forms in L2. The modals may and can and native French pouvoir illustrate the case in point. Despite the fact that all three forms contribute to the expression of the semantic notion of possibility, pouvoir synchronically covers the whole range of the modal uses of may and can. One corpus-based study of learners’ use of modals is Aijmer (2002), which is based on a corpus of Swedish L2 English writers. She compares (i) the frequencies of key modal words in native English and advanced Swedish-English interlanguage, as well as (ii) frequencies encountered in Swedish learner English with those from comparable French and German L2 English. Aijmer’s study indicates “a generalized overuse of all the formal categories of modality” and she further points out that “it is only at a functional level that any underuse was detected, with the learner writers failing to use may at all in its root meaning” (p. 72). Similarly, Neff et al. (2003) investigate the uses of modal verbs (can, could, may, might and could) by writers from several L1 backgrounds. Neff et al. (2003) use a learner corpus including Dutch-, French-, German-, Italian-, and Spanish-English interlanguage, which they contrast with a reference corpus of American university English. Neff et al. (2003:â•›215) identify the case of can as potentially interesting “since it is overused by all non-native writers”. They further report that the frequency of may by French native speakers stands out in comparison to the frequencies by all other non-native speakers included in the study, but since their study does basically nothing


but compare raw frequencies of occurrence regardless of any contextual features , it is not particularly illuminating. Generally, and similar to Gabrielatos and Sarmento (2006) and Collins (2009), both Aijmer (2002) and Neff et al. (2003) made the disadvantageous methodological decision to conveniently, but ultimately problematically, rely on information that is retrievable without human effort. In addition, even the studies that address learner use do not relate their findings to the wider context of (second) language acquisition. In a corpus-based contrastive study, Salkie (2004) investigates the nature of the semantic relations between the three forms in native English and native French. He uses a subpart of the parallel corpus INTERSECT (cf. Salkie 2000), and focuses on three working hypotheses, namely that: – “pouvoir corresponds more closely to one of the English modals rather than the other” (p. 169); – “pouvoir is less specific than the English modals” (p. 170); – “pouvoir has a sense which is different from both the English modals but is not just a general sense of possibility” (p. 170). While Salkie (2004) concludes in favour of the third hypothesis, it is worth pointing out, however, that his results were based on only 100 randomly extracted occurrences of each English modal form (i.e. may and can) and their respective French translations. By way of a more general summary, it is probably fair to say that corpus-based approaches to modality in L1 and L2s leave things to be desired. Some studies point to the immense complexity of the subject but do not choose multifactorial or multivariate methods that are capable of addressing this degree of complexity. In addition, some studies are based on large numbers of modals but, frankly, do not do very much with the vast amount of data other than present arrays of statistically under-analyzed frequency tables. On the other hand, the analytically much more interesting studies of the kind of Salkie (2004) are based on very small samples. Finally, many studies are largely if not exclusively form-based and focus only on learners’ over-/underuse of modals in particular examples or kinds of contexts.

2.3

Characteristics of the present study

2.3.1 Methodological considerations The above discussion fairly clearly indicates what kinds of steps would be desirable, an approach that: – can integrate linguistic information and patterning from many different levels of linguistic analysis in a way alluded to by Hermerén (1978), as well as Klinge and Müller (2005);


– involves not only a sample that is studied with regard to more linguistic parameters, but at the same time also larger than the previous studies that aimed at more than description; – explores similarities and differences of L1 uses of can and may, but also explores the way these English modals are used in L2 language (here from French learners) as well as how the same concept is used by the learners in their L1 (here pouvoir). Given these demands, we decided to use the so-called Behavioral Profile approach, which fits the above wish list very well. It combines the statistical methods of contemporary quantitative corpus linguistics with a cognitive-linguistic and psycholinguistic perspective or orientation (cf. Divjak and Gries 2006, 2008, 2009; Gries 2006, 2010b; Gries and Divjak 2009, 2010; and others). As such, it diverges radically from the above-mentioned more traditional corpus-based approaches to modality in both L1 and L2. Methodologically, it involves four steps: – the retrieval of all instances of a word’s lemma from a corpus in their context; – a manual annotation of a number of features characteristic of the use of the word forms in the data; these features are referred to as ID tags and typically involve morphosyntactic and semantic features in particular. Each ID tag contributes to the profiling of the investigated lexical item(s); – the generation of a table of co-occurrence percentages, which specify, for example, which words (from a set of near-synonymous words) or senses (of a polysemous word) co-occur with which morphosyntactic and/or semantic ID tags; it is these vectors of percentages that are called profiles; – the evaluation of that table by means of statistical techniques. Given how this approach is completely based on various kinds of co-occurrence information, it comes as no surprise that, just like much other work in corpus linguistics, the BP approach assumes that “the distributional characteristics of the use of an item reveals many of its semantic and functional properties and purposes” (Gries and Otani 2010:â•›3). While these previous studies have investigated a variety of different lexical relations (near synonymy, polysemy, antonymy) both within languages (English, Finnish, Russian) and across languages (English and Russian), the present study will add to the domains in which Behavioral Profiles have been used in two ways: (i) so far, no non-native language data have been studied, and (ii) we will add French to the list of languages studied. As the first BP study focusing on learner data, and only the second BP study that compares data from different languages, this paper is still largely exploratory. We will mainly be concerned with the following two issues: – To what degree can the Behavioral Profiling handle the kind of learner data that are inherently more messy and volatile than native data and provide a quantitatively adequate and fine-grained characterization of the use of can and may by


native speakers and learners, and how does that use compare to the use of French speakers’ use of pouvoir? – As a follow-up, and if meaningful groups of uses emerge, to what degree do the distributional characteristics that BP studies typically include allow us to predict native speakers’ and learners’ choices of modal verbs, and how do these speaker groups differ? The former question will be explored with the kind of cluster-analytic approach usually employed in BP studies; for the latter question, we will turn to a logistic regression (cf. Arppe 2008 for another BP approach using (multinomial) regression).

2.3.2 Theoretical orientation In previous studies, the BP approach was used for more than just the quantitative description of the data. Rather, it is firmly grounded in, and attempts to relate the results of the statistical exploration of the data to usage-based/exemplar-based approaches within Cognitive Linguistics and psycholinguistics. While this orientation is also compatible with our current goals, there is one particular earlier model in L2/ FLA research that is especially well-suited to, or compatible with, our current objectives, namely the Competition Model (CM) by Bates and MacWhinney (cf. Bates and MacWhinney 1982, 1989). This model is “a probabilistic theory of grammatical processing which developed out of a large body of crosslinguistic work in adult and child language, as well as in aphasia” (Kilborn and Ito 1989:â•›261). MacWhinney (2004:â•›3) himself characterized it as a “unified model [of language acquisition] in which the mechanisms of L1 learning are seen as a subset of the mechanisms of L2 learning”. The CM is characterized by the two following assumptions: – Linguistic signs map forms and functions onto each other (probabilistically) such that forms and functions are cues to functions and forms respectively. – In language production, forms compete to express underlying intentions or functions, and in language comprehension, the input contains many different cues of different strengths, validities, and reliabilities, which must be integrated: native speakers “depend on a particular set of probabilistic cues to assign formal surface devices in their language to a specific set of underlying functions” (Bates and MacWhinney 1989:â•›257). As a usage-based and probabilistic model, the CM assumes that both frequency and function determine the choice of grammatical forms in language production; as with most usage-based and/or corpus-linguistic approaches, we too consider frequency in a corpus as a proxy for frequency of exposure (in both comprehension and production). Cross-linguistically, this is an important assumption because across languages cues are instantiated in different ways and speakers assign them varying degrees of strength. It is therefore important to describe and explain L1 statistical regularities as


“[t]hey are part of the native speaker’s knowledge of his/her language, and they are an important source of information for the language learner” (Bates and MacWhinney 1989:â•›15). Overall, Kilborn and Ito (1989: 289) conclude that existing psycholinguistic studies have successfully demonstrated that the CM is appropriate for the characterization of learner language through cue distributions and they report “extensive evidence for the invasion of L1 strategies into L2 processing”. In addition, it is also obvious how much the CM is compatible with a BP approach. The main notions that drive the Competition Model are cue strengths, validities, and reliabilities, and all of these are essentially conditional probabilities, i.e. percentages. While the BP approach as such does not cover the full complexity of how conditional cue strengths, validities, and reliabilities can interact, it is a useful and experimentally validated (cf. Divjak and Gries 2008) approach employing a similar logic. A theory of language transfer requires that we have some ability to predict where the phenomena in question will and will not occur. In this regard contrastive (Gass 1996:â•›324) analysis alone falls short; it is simply not predictive.

3. Data and methods 3.1

Retrieval and annotation

The data are from three untagged corpora: the French subsection of the International Corpus of Learner English (henceforth ICLE-FR), the Louvain Corpus of Native English Essays (LOCNESS), and the Corpus de Dissertations Françaises (CODIF). All corpora included in the present work were collected by the Centre for English Corpus Linguistics (CECL) at the Université Catholique de Louvain (UCL) and made available to us by the Director of the Centre, Professor Sylviane Granger. ICLE-FR has a total of 228,081 words, including 177,963 words of argumentative texts and 50,118 words of literary texts. LOCNESS is a 324,304-word corpus that includes three sub-data sets: a 60,209-word-sub-corpus of British A-Level essays, a 95,695-word sub-corpus of British university essays and a sub-corpus of American university essays that has 168,400 words. The CODIF is a corpus of essays written by French-speaking undergraduate students in Romance languages at the Université Catholique de Louvain (UCL). CODIF also includes argumentative and literary texts and has a total of 100,000 words.2

2. Information on the total number of words featuring in each individual text type (i.e. argumentative, literary) is not available.


Table 1.â•‡ Excerpt of an annotation table including selected variables Case Match Corpus ClType 5 133 1760 1886 2876 3540 3645

may may may can cannot peut peuvent

native native native il il fr fr

coordinate main main coordinate subordinate main subordinate

Use

VerbSemantics

Neg

RefAnim

process state process process state process process

ment/cog/emotional copula ment/cog/emotional ment/cog/emotional abstract ment/cog/emotional abstract

affirmative affirmative negative affirmative negative negative negative

animate inanimate animate animate inanimate animate inanimate

Given the corpora’s compositions, the three corpora included in our study are highly comparable. They all consist of written data produced by university students (ICLE, CODIF, the LOCNESS British and American university sections) or by students approaching university entrance (i.e. the LOCNESS British A-Level section).3 All participants’ contributions are in the form of an essay of approximately 500 words long. In terms of content, all essays deal with similar topics such as: crime, education, the Gulf War, Europe, or university degrees. The data we subjected to the BP approach consist of instances of may and can in native English and French-English interlanguage as well as pouvoir in native French from the above corpora. Using scripts written in R (cf. R Development Core Team 2010), we retrieved 3,710 occurrences of the investigated modal forms from all sub-corpora, which were imported into a spreadsheet software and annotated for 22 morphosyntactic and semantic variables.4 Table 1 exemplifies this database with a very small excerpt of these data, and Table 2 presents the total range of variables included in the study and their respective levels. For each variable, an encoding taxonomy was designed prior to annotation. Due to the large number of variables included in this study and the absence of a number of them from previous studies on the English modals, not all encoding taxonomies were theoretically motivated. In cases where the annotation is not based on accounts from the existing literature, a bottom-up approach was adopted for the identification of recurrent features in the data. This procedure, for instance, was carried out in the case of the variable VerbSemantics where, prior to annotation, recurrent semantic features were identified as characteristic of the lexical verbs used alongside the modals.

3. The inclusion of the LOCNESS British A-Level section alongside sub-corpora solely including university participants is not judged problematic as LOCNESS only involves English native speakers whose level of English is not expected to develop any further. 4. Although the annotation process included a variable encoding the semantic role of the subject referent of the modals, this study does not account for that variable due to its high correlation with VOICE.


Table 2.â•‡ Overview of the variables used in the study and their respective levels Type

Variable

Levels

data

Corpus GramAcc (acceptability) Neg (negation) SentType (sentence type) ClType (clause type) Form SubjMorph: subject morphology SubjPerson: subject person SubjNumber: subject number Voice Aspect Mood SubjRefNumber: subject referent number Senses SpeakPresence Use

native, interlanguage, French yes, no affirmative, negated declarative, interrogative main, coordinate, subordinate can, may, pouvoir (and negated forms) adj., adv., common noun, proper noun, relative pronoun, date, noun phrase, etc. 1, 2, 3 singular, plural active, passive perfect, perfective, progressive indicative, subjunctive singular, plural

syntactic

morphological

semantic

VerbSemantics

RefAnim: subject referent animacy AnimType: subject referent animacy type

epistemic, deontic, dynamic weak, medium, strong accomplishment, achievement, process, state abstract, general action, action incurring transformation, action incurring movement, perception, etc. animate, inanimate animate, floral, object, place/time, mental/ emotional, etc.

Because of space restrictions, we are not able to provide a more comprehensive account of the annotation process (but cf. Deshors 2010 for details). However, three variables – Senses, VerbType, and VerbSemantics – require some brief explanatory comments.

3.1.1 The variable Senses As for Senses, the semantic category of modality includes a wide range of heterogeneous meanings that many scholars have attempted to unite under a variety of categorization systems (cf. Palmer 1979; Coates 1983; Bybee and Fleischman 1995; Huddleston 2002; Nuyts 2006; Byloo 2009). While Depraetere and Reed (2006:â•›277) note that “in classifying modal meanings, it is possible to use various parameters as criterial to their classification”, this study assumes a coding taxonomy based on a traditional tripartite distinction between epistemic, deontic and dynamic meanings.


Following Nuyts (2006:â•›6), epistemic senses concern “an indication of the epistemic estimation, typically, but not necessarily, by the speaker, of the chances that the state of affairs expressed in the clause applies in the world”. Consider (1) as an illustration of epistemic may: (1) indeed, Europe 92 may lead to the disappearance of cultural differences Following Palmer (1979:â•›58), deontic modality refers to cases where “[b]y uttering a modal, a speaker may actually give permission (may, can)”. (2) illustrates deontic can: (2) if all public schools started to say you can only come here if you are Hispanic or if you are Polish, our schooling system would be in great chaos Finally, dynamic meanings denote “an ascription of a capacity to the subject-participant of the clause (the subject is able to perform the action expressed by the main verb in the clause)” (Nuyts 2006:â•›3). Generally, dynamic modality expresses the potentiality of an event occurring. Nuyt’s type of dynamic modality includes ability/ capability cases where the possibility of event occurrence stems from the ability of the (grammatical) subject to carry out the event. In that regard, the term ability is not restricted to a ‘physical’ interpretation and equally applies to mental and technical types of ability. Example (3) illustrates dynamic can: (3) Mrs Ramsay is the central character because she can see the whole personality of the other ones Generally, our frequencies of use of may and can in their different senses match those previously encountered in existing studies solely concerned with the native use of the modals, such as Coates (1980) and Collins (2009). While Coates (1980:â•›218), for instance, reports that “by far the most common usage of may is to express epistemic possibility”, she stresses the distinctive nature of the uses of may and can: The patterns resulting from my analysis of the data (…) leads me to conclude that in normal everyday usage may and can express distinct meanings: may is primarily used to express epistemic possibility, while can primarily expresses root possibility.5

3.1.2 The variable VerbType The variable VerbType targets the lexical verbs with which the forms are used and characterizes their telicity. Conceptually, the variable VerbType follows Vendler (1967) in its recognition that the notion of time is crucially related to the use of a

5. Coates (1980, 1983) categorizes modal meaning according to a two-way distinction that includes epistemic and non-epistemic modality. She refers to the latter type as “root” modality.


verb and is “at least important enough to warrant separate treatment” (p. 143). This variable assesses: – whether may and can have preferences for lexical verbs denoting a state, a process, an accomplishment or an achievement,6 and if so, – it identifies in which type of corpus preferential patterns occur.

3.1.3 The variable VerbSemantics Similarly to the variable VerbType, VerbSemantics identifies the type of semantic information conveyed by the lexical verbs used with the modals. The internal organization of this variable results from a bottom-up approach and does not follow any particular theoretical framework. This variable consists of the levels denoting abstract process, physical actions, actions incurring movement, actions incurring some physical transformation, communicative processes, mental/cognitive/emotional processes, perception processes and verbal statement involving a copula verb. Example (4) illustrates a case where the lexical verb expresses a mental/cognitive/emotional process: (4) Her search for the final touch can be seen as a search for harmony Once all matches were annotated, the resulting data table was evaluated statistically.

3.2

The BP approach in this study: Statistical analysis

As mentioned above, the data were evaluated in two different ways.7 The first of these involved the type of cluster analysis that is characteristic of much work using the BP methodology. In this first part, we used Gries ’s (2010a) R script Behavioral Profiles 1.01 and computed five behavioral profiles, one for each modal form as occurring in each language variety, i.e. native can, native may, interlanguage (IL) can, IL may, and native pouvoir (FR). Such profiles consist of vectors of co-occurrence percentages of a single modal form with each level of all independent variables and provide form-specific summaries of their semantic and morphosyntactic behavior in each sub-corpus. In a second step, the profiles were assessed statistically with a hierarchical cluster analysis to explore the similarity and differences between the modal forms, and in keeping with previous studies (cf. Divjak and Gries 2006), we chose the Canberra metric as a measure of (dis)similarity and Ward’s rule as an amalgamation strategy. 6. Accomplishment verbs encode verbal statements that imply a unique and definite time period; achievement verbs encode verbal statements that imply a unique and definite time instant; process verbs identify statements that reflect non-unique and indefinite time periods; state verbs identify statements that reflect non-unique and indefinite time instants. 7. All statistical computations and plots were performed with R (for Linux), version 2.11.0 (see R Development Core Team 2010).


Following Gries and Otani (2010), we computed different cluster analyses, one involving all variables that the uses of the modals were annotated for, one for only the syntactic variables, and one for only the semantic variables. The second analytical step involved a binary logistic regression including the following variables and predictors: – Form as the dependent variable with only two levels here: can vs. may; – GramAcc, Neg, SentType, ClType, SubjMorph, SubjPerson, SubjNumber, Voice, Aspect, Mood, SubjRefNumber, Senses, SpeakPresence, Use, VerbSemantics, RefAnim, AnimType as independent variables in the form of main effects; – all these variables’ interactions with Corpus as additional predictors (to see which variables’ influence on modal use differs the most between L1 English and L2 English). The logistic regression was then performed with the model selection process during which insignificant predictors were discarded from the model: first insignificant interactions, then individual variables that were not significant and did not participate in a significant interaction.

4. Results and discussion 4.1

Cluster analysis

Our first cluster analysis yielded the results shown in Figure 1. The left plot is a dendrogram of the five modal forms that were clustered; the right plot represents average silhouette widths for assuming two, three, and four clusters. The average silhouette widths point to a two-cluster solution, maybe a three-cluster solution, but the difference is minor since the former would result in a French-vs.-English clustering, and the latter in a French-vs.-can-vs.-may clustering. This is compatible with Salkie’s analysis, who argued that pouvoir is very different from both can and may, and intuitively, both these solutions “make sense”, which provides first evidence in favor of the approach. To anticipate the potential objection that this may seem trivial, let us mention that it is in fact not. The data in Figure 1 show that the BP vectors are good and robust descriptors of how the modals behave because many other theoretically possible cluster solutions, such as the ones listed in (5), would not have made linguistic sense at all. (5) a. {{{canil maynative pouvoir} cannative} mayil} b. {{{cannative mayil pouvoir} canil} maynative} c. {{canil maynative} {pouvoir mayil} cannative}

1.0

0.03

0.4

0.6

0.8

0.07

0.2

(Average) silhouette widths

0.1

0.0

MAY.native

MAY.il

CAN.native

CAN.il

POUVOIR.fr

Height

700 750 800 850 900 950


0 1 2 3 4 Number of clusters in the solution

18 16 14

CAN.native

CAN.il

MAY.native

MAY.il

POUVOIR.fr

12 6

CAN.native

CAN.il

8

10

MAY.il

MAY.native

POUVOIR.fr

Height

0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5

Figure 1.â•‡ Dendrogram for all independent variables (il = interlanguage)

Figure 2.â•‡ Dendrograms for all morphosyntactic variables (left panel) and all semantic variables (right)

However, in what follows we show that a fine-grained comparative description of cross-linguistic language varieties can be obtained by focusing on differences between the independent variables used for clustering. Consider Figure 2, which shows the dendrograms for all morphosyntactic variables and all the semantic variables in the left and right panel, respectively. Interestingly, the results show that the intuitively very reasonable dendrogram in Figure 1 is not replicated by looking at morphosyntax or semantics alone, which to some extent at least contrasts with Gries and Otani’s results, where the results did not differ very much between the three clusterings. The reasonable similarities of Figure 1 emerge only when all variables are combined. In particular, in both panels of Figure 2 canil and cannative are grouped together, but then the remaining forms are grouped differently. In the morphosyntactic dendrogram, the two kinds of may are successively amalgamated and the French pouvoir is only added after all English forms have been


affirmative

er en ce s di ff o rn yo in lt se ve ra

negated

subord clauses

main clauses

0.4 0.2 0.0 –0.2 –0.4

Syntactic differences mat interlang - may native

Figure 3.â•‡ Snakeplot for most extreme differences between syntactic ID tags of may

clustered. In other words, morphosyntactically, we find a clear English-French divide, but interlanguage may is too different from native may to be grouped together. To identify the source of this difference, we used what in BP approaches has been called a snakeplot, namely a plot of the pairwise differences between the percentages for, in this case, mayil and maynative (cf. Divjak and Gries 2009 or Gries and Otani 2010 for more examples). As indicated in Figure 3, the main morphosyntactic ways in which learners deviate from native speakers are that learners underuse may in subordinate clauses and in negated clauses. This is in fact an interesting finding because it means that learners disprefer the rarer of the two modals – may – in those contexts which are already morphosyntactically more challenging, as if using can is the default they resort to when they are already under a higher processing load (cf. the so-called complexity principle). In the semantic dendrogram, by contrast, we find a different patterning. Semantically, canil and cannative are again very similar and grouped together early, but then the next clustering step groups the two forms of may together. However, interestingly, it is not the English forms that are then all grouped together – rather, contrary to Salkie’s earlier analysis, pouvoir is semantically more similar to can than may is.

4.2

Logistic regression

The model selection process involved thirteen steps during which insignificant predictors were discarded. The final and minimally adequate model includes 16 significant variables and 6 significant interactions and returned a highly significant correlation: loglikelihood chi-square = 3296.47; df = 60; p < 0.001; the correlation between the


Table 3.â•‡ Overview of the results of the final GLM model Predictor

Chi-square (df)

Predictor

Chi-square (df)

Corpus GramAcc Use Elliptic ClType VerbType VerbSemantics SubjPerson SubjNumber SubjMorph RefAnim

â•⁄ 24.9 (1) *** â•⁄ 13.8 (1) *** â•⁄ 67.9 (1) *** 100.0 (2) *** â•⁄ 10.9 (1) *** â•⁄ 97.4 (2) *** 384.9 (6) *** â•⁄ 26.6 (2) *** â•⁄â•⁄ 1.3 (1) ns â•⁄ 49.1 (4) *** â•⁄ 59.2 (1) ***

AnimType Voice SentType Negation SpeakPresence Corpus:ClType Corpus:VerbSemantics Corpus:SubjNumber Corpus:RefAnim Corpus:AnimType Corpus:Negation

â•⁄â•⁄â•⁄98.2 (11) *** â•⁄â•⁄â•⁄55.0 (1) *** â•⁄â•⁄â•⁄47.2 (1) *** â•⁄â•⁄â•⁄87.2 (1) *** 29905.9 (2) *** â•⁄â•⁄â•⁄60.0 (2) *** â•⁄â•⁄â•⁄32.2 (6) *** â•⁄â•⁄â•⁄37.4 (1) *** â•⁄â•⁄ 122.2 (1) *** â•⁄â•⁄ 118.2 (11) *** â•⁄â•⁄â•⁄12.0 (1) ***

observed forms – may vs. can – and predicted probabilities is very high: R2 = 0.955. Correspondingly, the model’s classificatory power was found to be very powerful with a classification accuracy of 99%. Table 3 summarizes all the significant variables and interactions yielded in the final model. Overall, the final model includes one significant interaction involving a morphological variable (out of seven morphological variables), two significant interactions involving syntactic variables (out of three syntactic variables) and three significant interactions involving semantic variables (out of eight semantic variables). But what do the interactions reflect? Let us begin with Corpus:ClType, as represented in Figure 4. The frequencies of may and can differ with regard to the type of clauses in which they occur in native and learner English. The (weak!) effect is that, in interlanguage

may

may

can

can

can

may

may

can

can

can

0.4 0.2

Coordinate

Main Subordinate

0.0

0.4 0.2 0.0

may

0.6

0.8

may

0.6

0.8

1.0

Native

1.0

Interlanguage

Coordinate

Figure 4.â•‡ Bar plots of relative frequencies of Corpus:ClType

Main Subordinate


0.8

may

0.4

0.6

may

can

can

0.2 Affirmative

Negative

0.0

0.4

can can

0.2 0.0

Native 1.0

may may

0.6

0.8

1.0

Interlanguage

Affirmative

Negative

Figure 5.â•‡ Bar plots of relative frequencies of Corpus:Neg

English, can is more strongly preferred over may in main clauses than it is in native English. While, as previously noted, existing literature concerned with the native use of the modals commonly recognizes negation as “an important aspect of modal meaning” (Hermerén 1978), our study not only confirms the need to include negation in an investigation of the uses of the modals but further recognizes its significance as a morphological criteria to assess interlanguage (dis)similarity. Consider Figure 5 for the interaction Corpus:Neg. Figure 5 shows that, while all speakers prefer to use can in negated clauses, the interlanguage speakers do so more strongly. This result does not come as a surprise: On the one hand, this is also compatible with the complexity principle – negated clauses are more complex and preferred with the more frequent modal. On the other hand, where epistemic may not would be used in English, French speakers would tend to use a lexical verb along with the adverb peut-être to indicate the speaker’s uncertainty, as illustrated in (6): (6) a. This may not be the case b. Ce n’est peut-être pas le cas Consider Figure 6 for the interaction Corpus:SubjNumber. While native speakers use can more often with singular subjects than with plural subjects, it is the other way round with the learners, again a result compatible with the complexity principle. While the native speakers’ choices of may and can do not vary much between animate and inanimate subjects, the learners’ choices do: with animate subjects, they prefer can much more strongly. Figure 7 represents the interaction Corpus:RefAnim.


may

can

can

pl

sg

may

0.4

0.6

may

can

can

0.2 0.0

0.0

0.2

0.4

0.6

0.8

may

0.8

1.0

Native

1.0

Interlanguage

pl

sg

Figure 6.â•‡ Bar plots of relative frequencies of Corpus:SubjNumber

1.0

may

can

0.4

0.6

may

can

can

0.2

can

Animate

Inanimate

0.0

0.4

0.6

may

0.2 0.0

Native

0.8

may

0.8

1.0

Interlanguage

Animate

Inanimate

Figure 7.â•‡ Bar plots of relative frequencies of Corpus:RefAnim

Consider Figure 8 for the interaction Corpus:VerbSemantics; the upper panel represents the interlanguage data, the lower panel represents the native speaker data, and the bars are sorted from large absolute pairwise differences (left) to small absolute pairwise differences (right). The learners and the native speakers differ most strongly with semantically more abstract verbs and time/place verbs, as in He thinks that if he can achieve one impossible act, then this will change everything. The learners prefer can with abstract verbs more strongly than the native speakers, but they prefer may more strongly with time/place verbs. However, there are also (less pronounced) differences for verbs that would typically have a human agent.


may

may may

may may

may may

can

can can

can can

can can

0.0

0.2

0.4

0.6

0.8

1.0

Interlanguage

abstr

templ

comm

act_transf act_gen/mot ment/perc

copula

may

may

may

may may

may may

can

can can

can can

can can

0.0

0.2

0.4

0.6

0.8

1.0

Native

abstr

templ

comm

act_transf act_gen/mot ment/perc

copula

Figure 8.â•‡ Bar plots of relative frequencies of Corpus:VerbSemantics

For instance, the learners prefer may with communication verbs and can with action-transformation verbs. Virtually no difference at all is found with copulas. As for the final interaction, Corpus:AnimType, we do not represent it here graphically. While it is significant, the large number of categories plus the fact that the most pronounced differences occur with a small number of very infrequent categories does not yield much in terms of interesting findings. As for the main effects, we will not discuss them here in detail. This is because these main effects by definition do not tell us anything about the can and may variables across languages (since these variables do not interact with Corpus). However, since they do tell us something about which modal verb is preferred by both native speakers and learners, we summarize them here visually in Figure 9. The x-axis lists the main effects, on the y-axis we show the percentage of can obtained for levels of these main effects, and then the levels are plotted at their observed percentage of can; the dashed line represents the overall percentage of can in the data.

weak

passive il native

yes

active

literal declarative no

two one prop_noun pr accmp-achv proos three com_noun

metaphorical

0.6 0.4

interrogative yes

no

state

rel-dem/pr other

strong

0.0

0.2

Percentage of can

0.8

1.0


medium

corpus speaker grm. voice presence accept

use

sent. type

elliptic

verb type

subj. subj. pers. morph.

Figure 9.â•‡ Main effects of the logistic regression

Finally, a brief look at the regression’s misclassifications seems to indicate that they did not occur randomly. While all 34 misclassifications occurred in the interlanguage data, 29 of them occurred with may in a form characteristic only of the French-English learner language. In the large majority of those misclassifications, may is found to express a possibility that results from some sort of theoretical demonstration. Consider the examples in (7) and (8). While the ones in (7) illustrate our current point, (8) provides an additional example of an atypical occurrence of learner may, which clearly denotes a strong sense of possibility and whose interpretation is heavily reminiscent of that of can. (7) a. So we may say that … b. To conclude, we may say that … c. As a conclusion, we may say that … d. This is why we may now speak of the stupefying effect e. This is the reason why we may say that … (8) “Dresden is an old town”, we may read of its history

5. Concluding remarks By way of a summary, the BP approach and the subsequent logistic regression allows us to recognize how can and may (in native and learner English), as well as pouvoir, relate to each other as well as what helps determine native speakers’ and learners’ choices. On the whole, distributionally we do find the expected groupings: the cans, then the mays, and only then pouvoir. However, it is interesting that, semantically,


English can is more similar to French pouvoir than to English may, and the subsequent regression results provided some initial information on why that is so. More specifically, the way learners choose one of the two verbs is often compatible with a processing-based account in terms of the complexity principle – they choose the more basic and frequent can over may when the environment is complex – but is also strongly influenced by the animacy of the subject and the semantics of the verb: can is overpreferred by learners with animate subjects and with abstract verbs, and underpreferred with time/place verb semantics. With regard to the modals per se, our results confirm previous studies’ recognition of the influential role of the linguistic context in the uses of may and can. Indeed, while the main effects included in our final logistic regression model support studies that have identified morphosyntactic components such as Voice and SentType as particularly influential categories (Leech 1969, 2004; Huddleston 2002; Collins 2009), our results reveal the necessity to also take the semantic context of modals more seriously, as reflected by the strong effects of VerbType and VerbSemantics. More generally speaking and in the parlance of the Competition Model, the cluster analysis and the high classification accuracy of the regression suggest that, on the whole, the learners have built up mental categories for can and may that are internally rather coherent. However, the interactions in the regression show that these cues are weighted incorrectly and sometimes trigger a verb choice that is not in line with native speaker choices, but that even this kind of incorrect choice is largely predictable (because the regression can still make the correct classifications (cf. Deshors 2010 for more detailed discussion as well as a distinctive collexeme analysis revealing additional verb-specific preferences). In other words, even though this is the first study involving learner data (and only the second involving different languages), the BP approach and especially the follow-up in terms of the logistic regression are therefore an interesting diagnostic: (i) the overall results can testify to the strength of the categories that are being studied, and (ii) the regression with its inclusion of the interactions of all variables with “native speaker vs. learner” exactly pinpoints where interactions become significant, i.e. where the categories of the learner are still substantially different from the native speaker. For further applications and extensions, see Gries and Wulff (2013) for a similar application to the choice of (of- and s-) genitives by native speakers and learners, and Gries and Deshors (to appear) for an even more advanced approach to precisely pinpoint where non-native speakers’ choices deviate from those of native speakers and how much so. Needless to say, more and more rigorous testing is necessary, but to our knowledge this is the first study proposing this kind of approach more generally and the use of a regression with a native-learner variable as a measure of L2 “proficiency”; the results illustrate that learners’ “non-nativeness” manifests itself at all linguistic levels simultaneously.


References Aijmer, K. (2002). Modality in advanced Swedish learners’ written interlanguage. In S. Granger, J. Hung, & S. Petch-Tyson (Eds.), Computer learner corpora, second language acquisition and foreign language teaching (pp. 55–76). Amsterdam: John Benjamins. Arppe, A. (2008). Univariate, bivariate and multivariate methods in corpus-based lexicography: A study of synonymy. Unpublished PhD dissertation, University of Helsinki. Available at: . Bartning, I. (2009). The advanced learner variety: 10 years later. In E. Labeau, & F. Myles (Eds.), The advanced learner variety: The case of French (pp. 11–40). Frankfurt/Main: Peter Lang. Bates, E., & MacWhinney, B. (1982). Functionalist approaches to grammar. In E. Wanner, & L. R. Gleitman (Eds.), Language acquisition: The state of the art (pp. 173–218). Cambridge: Cambridge University Press. Bates, E., & MacWhinney, B. (1989). Functionalism and the competition model. In B. MacÂ� Whinney, & E. Bates (Eds.), The cross-linguistic study of sentence processing (pp. 3–73). Cambridge: Cambridge University Press. Bybee, J., & Fleischman, S. (1995). Modality in language and discourse. Amsterdam: John Benjamins. DOI: 10.1075/tsl.32 Byloo, P. (2009). Modality and negation: A corpus-based study. Unpublished PhD dissertation, University of Antwerp. Coates, J. (1980). On the non-equivalence of may and can. Lingua, 50(3), 209–220. DOI: 10.1016/0024-3841(80)90026-1 Coates, J. (1983). The semantics of the modal auxiliaries. London: Croom Helm. Collins, P. (2009). Modals and quasi modals in English. Amsterdam: Rodopi. De Haan, F. (1997). The interaction of modality and negation: A typological study. New York: Garland. Depraetere, I., & Reed, S. (2006). Mood and modality in English. In B. Aarts, & A. MacMahon (Eds.), The handbook of English linguistics (pp. 268–287). London: Blackwell. Deshors, S. C. (2010). A multifactorial study of the uses of may and can in French-English interlanguage. Unpublished PhD dissertation, University of Sussex. Divjak, D. S., & Gries, St. Th. (2006). Ways of trying in Russian: Clustering behavioral profiles. Corpus Linguistics and Linguistic Theory, 2(1), 23–60. DOI: 10.1515/CLLT.2006.002 Divjak, D. S., & Gries, St. Th. (2008). Clusters in the mind? Converging evidence from near synonymy in Russian. The Mental Lexicon, 3(2), 188–213. DOI: 10.1075/ml.3.2.03div Divjak, D. S., & Gries, St. Th. (2009). Corpus-based cognitive semantics: A contrastive study of phasal verbs in English and Russian. In K. Dziwirek, & B. Lewandowska-Tomaszczyk (Eds.), Studies in cognitive corpus linguistics (pp. 273–296). Frankfurt/Main: Peter Lang. Gabrielatos, C., & Sarmento, S. (2006). Central modals in an aviation corpus: Frequency and distribution. Letras de Hoje, 41(2), 215–240. Gass, S. (1996). Second language acquisition and linguistic theory: The role of language transfer. In W. C. Ritchie, & T. K. Bhatia (Eds.), Handbook of second language acquisition (pp. 317–340). San Diego: Academic Press. Gries, St. Th. (2006). Corpus-based methods and cognitive semantics: The many meanings of to run. In St. Th. Gries, & A. Stefanowitsch (Eds.), Corpora in cognitive linguistics: Corpus-based approaches to syntax and lexis (pp. 57–99). Berlin: Mouton de Gruyter. DOI: 10.1515/9783110197709


Gries, St. Th. (2010a). Behavioural Profiles 1.01: A program for R 2.7.1 and higher. Gries, St. Th. (2010b). Behavioral profiles: A fine-grained and quantitative approach in corpus-based lexical semantics. The Mental Lexicon, 5(3), 323–346. Gries, St. Th., & Deshors, S. C. (To appear). Using regressions to explore deviations between corpus data and a standard/target: two suggestions. Corpora. Gries, St. Th., & Divjak, D. S. (2009). Behavioral profiles: A corpus-based approach to cognitive semantic analysis. In V. Evans, & S. Pourcel (Eds.), New directions in cognitive linguistics (pp. 57–75). Amsterdam: John Benjamins. Gries, St. Th., & Divjak, D. S. (2010). Quantitative approaches in usage-based cognitive semantics: Myths, erroneous assumptions, and a proposal. In D. Glynn, & K. Fischer (Eds.), Quantitative cognitive semantics: Corpus-driven approaches (pp. 333–354). Berlin: Mouton de Gruyter. Gries, St. Th., & Otani, N. (2010). Behavioral profiles: A corpus-based perspective on synonymy and antonymy. ICAME Journal, 34, 121–150. Gries, St. Th., &Wulff, S. (2013). The genitive alternation in Chinese and German ESL learners: Towards a multifactorial notion of context in learner corpus research. International Journal of Corpus Linguistics, 18(3), 327–356. Hermerén, L. (1978). On Modality in English: A study of the semantics of the modals. Lund: LiberLäromedel/Gleerups. Huddleston, R. D. (2002). The Cambridge grammar of the English language. Cambridge: Cambridge University Press. Hyltenstam, K., Bartning I., & Fant L. (2005). High Level Proficiency in Second Language Use. Research program for Riksbanken Jubileumsfond. (Stockholm university) http://www. biling.su.se/~AAA. Kilborn, K., & Ito, T. (1989). Sentence processing strategies in adult bilinguals. In B. MacÂ� Whinney, & E. Bates (Eds.), The cross-linguistic study of sentence processing (pp. 257–291). Cambridge: Cambridge University Press. Klinge, A., & Müller, H. H. (2005). Modality: Intrigue and inspiration. In A. Klinge, & H. H. Müller (Eds.), Modality studies in form and function (pp. 1–4). London: Equinox. Leech, G. (1969). Towards a semantic description of English. Bloomington, IN: Indiana University Press. Leech, G. (2004). Meaning and the English verb. London & New York: Longman. MacWhinney, B. (2004). A unified model of language acquisition. Retrieved from [Accessed 18 June 2010]. Neff, J., Dafouz, E., Herrera H., Martínez, F., & Rica, J. P. (2003). Contrasting the use of learner corpora: The use of modal and reporting verbs in the expression of writer stance. In S. Granger, & S. Petch-Tyson (Eds.), Extending the scope of corpus-based research: New applications, new challenges (pp. 211–230). Amsterdam: Rodopi. Nuyts, J. (2006). Modality: Overview and linguistic issues. In W. Frawley (Ed.), The expression of modality (pp. 1–26). Berlin: Mouton de Gruyter. Palmer, F. (1979). Modality and the English modals. London & New York: Longman. Radden, G. (2007). Interaction of modality and negation. In W. Chłopicki, A. Pawelec, & A. Pokojska (Eds.), Cognition in language: Volume in Honour of Professor Elżbieta TabaÂ� kowska (pp. 224–254). Kraków: Tertium. R Development Core Team (2010). R: A language and environment for statistical computing. Foundation for statistical computing. Vienna, Austria. .


Salkie, R. (2000). Corpus linguistics: A brief guide to research in French language and linguistics. AFLS Cahiers, 6, 44–52. Salkie, R. (2004). Towards a non-unitary analysis of modality. In L. Gournay, & J.-M. Merle (Eds.), Contrastes: mélanges offerts à Jacqueline Guillemin-Flescher (pp. 169–182). Paris: Ophrys. Vendler, Z. (1967). Verbs and times. In Z. Vendler (Ed.), Linguistics in philosophy (pp. 97–121). New York: Cornell University Press.

Dutch causative constructions Quantification of meaning and meaning of quantification Natalia Levshina, Dirk Geeraerts, and Dirk Speelman

F.R.S. – FNRS, Université catholique de Louvain / University of Leuven

This chapter is a multivariate corpus-based study of two near-synonymous periphrastic causatives with doen and laten in Dutch. Using multiple logistic regression and classification trees, the study explores the conceptual differences between the constructions. The results support the existing definition of doen as the direct causation auxiliary, and interpretation of laten as the indirect causative (e.g. Verhagen and Kemmer 1997). However, the analyses also reveal more specific patterns: the most distinctive semantic pattern of doen is affective causation, whereas the contexts with the highest probability of laten refer to inducive causation. These differences remain valid when we control for geographic and thematic variation, as well as for the individual Effected Predicates treated as random effects in a mixed model. Keywords: classification trees, logistic regression, mixed model, periphrastic causatives

1. Introduction This chapter is a contribution to empirical Cognitive Semantics (e.g. Glynn and Fischer 2010).1 It is a corpus-based multivariate onomasiological study (cf. Tummers et al. 2005), which uses quantitative corpus evidence to describe, explain and predict the choices that speakers make between semantically related constructions when they categorize their experience. To do so, the linguist needs to identify the relevant semantic, pragmatic, social and other features that influence this choice. This kind of 1. This research was supported with a grant from the Flemish Research Fund – FWO (G033008). The authors would also like to thank Kris Heylen for his help in collecting the corpus data. The usual disclaimers apply.

206 Natalia Levshina, Dirk Geeraerts, and Dirk Speelman

study requires advanced statistical multivariate techniques, such as logistic regression, which allow the researcher to model the impact of each factor, while controlling for the others. The approach is relatively new, but a number of studies have been implemented already. It is interesting to note that a substantial share of these studies focus on constructions that differ from each other with respect to information structure and processing. Examples are the dative alternation in English (e.g. Bresnan et al. 2007), presence or absence of the presentative er-construction in Dutch (Grondelaers et al. 2007), particle placement in English (Gries 2003), word order variation in Dutch final verbal clusters (de Sutter 2009) and in the German ‘middle field’ (Heylen 2005). It seems that these alternations, a challenge for traditional linguistic descriptions, have benefited the most from the multifactorial probabilistic methods due to a variety of ways in which the underlying information-processing factors can be captured. However, more “semantic” constructional variation, like the one discussed here, can benefit from these methods too because highly abstract grammatical meaning can be captured in a corpus by a multitude of indirect indicators used as circumstantial evidence. The current chapter, which is an elaboration of the pilot study carried out by Speelman and Geeraerts (2009), focuses on the near-synonymous Dutch causative constructions with doen and laten. Our study incorporates several linguistic, thematic and geographical factors in a multivariate statistical model, which allows us to test the existing semantic hypotheses about doen and laten, keeping the conceptual factors apart from the other sources of variation. We argue that the distinctive conceptual features that emerge in the quantitative model constitute the distinctive prototypes of the constructions – the semantic configurations with the highest intercategorial cue validity. Although the Prototype Theory of categorization (Rosch 1975; Rosch and Mervis 1975) has been dominant in Cognitive Linguistics, many psychological and, more recently, linguistic studies (e.g. Medin and Schaffer 1978; Bybee and Eddington 2006) have demonstrated the crucial role of specific exemplars (or low-level schemata) in category organization and development. This is why we also test whether the abstract semantic differences between the constructions still hold if we take into account the lexemes that fill in the effected predicate slot, many of which display strong preference for doen or laten. The method applied for this purpose is mixed-effect modelling with the general semantic and other factors as fixed effects and the specific effected predicates as random effects. The chapter has the following structure. First, we give a brief introduction of the Dutch causative constructions. In Section 3, the data and the potentially relevant variables are presented. Section 4 reports the results of the multiple logistic regression analysis and additional tests, which are interpreted linguistically and cognitively in Section 5. The chapter ends with a summary of our findings.

Dutch causative constructions 207

2. Dutch causative constructions Modern standard Dutch has two periphrastic causatives with the infinitive: the constructions with doen ‘do’ and laten ‘let’. They share the same schematic pattern: an initiator causes another entity to acquire a state or perform an action. Consider example (1): (1) De politie deed/liet de auto stoppen. the police did/let the car stop ‘The police stopped the car’. Using the terminology from Kemmer and Verhagen (1994), de politie ‘the police’, is the causer of the event; de auto ‘the car’ is the causee that performs the action specified by the effected predicate stoppen ‘stop’. The forms deed and liet are the past forms of the causative auxiliaries doen and laten, respectively. The most striking feature of the Dutch causatives is that laten as a causative auxiliary can refer both to the enabling and coercive types of causation (see Verhagen and Kemmer 1997:â•›69). Compare the situations in (2a), (2b) and (2c): (2) a. De trainer liet de spelers loopoefeningen doen. [coercive] the coach let the players running-exercises do ‘The coach made the players do running exercises’. [ambiguous] b. Hij liet iedereen zijn roman lezen. He let everyone his novel read ‘He made/had/let everyone read his novel’. c. De politie liet de dader ontsnappen.[enabling] the police let the criminal escape ‘The police let the criminal escape’. There have been a number of usage-based studies that have tried to establish the differences between the constructions (Kemmer and Verhagen 1994; Verhagen and Kemmer 1997; Degand 2001; Stukker 2005; Speelman and Geeraerts 2009). Verhagen and Kemmer (1997) write about the semantic difference between doen and laten in terms of the speaker’s conceptualization of the situation as direct or indirect causation, respectively. Direct causation means that “there is no intervening energy source ‘downstream’ from the initiator: if the energy is put in, the effect is the inevitable result” (Verhagen and Kemmer 1997:â•›70). Indirect causation, which also includes the situations of enablement and permission, emerges when the situation “can be conceptualized in such a way that it is recognized that some other force besides the initiator is the most immediate source of energy in the effected event” (Ibid.:â•›67). Speelman and Geeraerts (2009) showed, in their multivariate analysis of the Corpus of Spoken Dutch, that there is also a substantial amount of geographic and register variation in the use of the constructions. From the conceptual point of view, it


was suggested that doen is an obsolescent form with a tendency towards semantic and lexical specialization, most probably in direct physical causation (Speelman and Geeraerts 2009:â•›200), although this hypothesis was not tested. The highly abstract conceptual patterns, such as direct and indirect causation, cannot be directly observed in a corpus-based study. Our aim is to explore a set of independent contextual factors (“diagnostic features”, according to Speelman and Geeraerts 2009) that can serve as indirect, or circumstantial, evidence of semantic differences between the constructions. These contextual factors, operationalized as independent variables in the logistic regression model, are listed in Section 3, as well as the extralinguistic (geographic and thematic) variables that are explored in this study.

3. Data and variables 3.1

Data

The study is based on an 8 million token corpus of Netherlandic and Belgian Dutch, compiled from the TwNC and LeNC newspaper corpora (2001–2002). The corpus was balanced with regard to four subject domains of the articles: politics, economy, football and music. We used a syntactically parsed version of the data, which was obtained with the help of the Alpino parser of Dutch (Bouma et al. 2001). This allowed us to extract the contexts with constructions automatically. The contexts were then checked manually to avoid spurious hits and formally similar but functionally different constructions, such as the adhortative laten in Laten we gaan ‘Let’s go’. We also excluded idiomatic expressions with effected predicates that do not occur independently, e.g. begaan, which only occurs in the set expression laten begaan ‘release, give freedom’. After the manual cleaning, we were left with 6,808 observations, which were then coded for seven semantic, syntactic, geographical and thematic variables presented in the next section.

3.2

The response variable

The speaker’s choice for doen or laten in the given context was used as the binary response variable. The distribution of the constructions in the data set was skewed towards laten, which occurred 5,636 times, while doen was used only in 1,172 contexts, which is approximately 5 times less.


3.3

The linguistic predictors

The variable CrSem refers to the semantic class of the causer: animate (humans and animals) or inanimate (material and abstract entities). All previous studies reported the more frequent use of animate causers with laten and inanimate ones with doen. Verhagen and Kemmer (1997) and Stukker (2005) studied the causer’s semantics only in combination with the semantic class of the causee. They found, however, that inanimate causers in combination with both animate or inanimate causees tend to be used more frequently with doen because these configurations correspond to physical and affective causation types, respectively, which imply direct causation. The most typical configuration for laten, which normally represents inducive causation, consists of animate causers and causees. This type of causation is indirect because humans cannot influence other humans’ minds directly, telepathy disregarded (Verhagen and Kemmer 1997:â•›71). The remaining possibility, the combination of animate causers with inanimate causees, allows both for direct and indirect interference of the causer. CeSem stands for the semantic class of the causee, which can also be animate or inanimate. If the causer is not the main source of energy in indirect causation, then it should most probably be the causee (cf. Stukker 2005). Thus, one could expect a higher degree of animacy of the causee in the laten-construction in comparison with doen. This variable has not been examined separately in any of the previous studies, although from Verhagen and Kemmer’s (1997) description of causation types it follows that the chances for inanimate causees to be used with doen are somewhat higher than for animate ones. Both explicit and implicit causees (see below) were classified, depending on the context and the semantics of the effected predicate. Nevertheless, we were unable to classify 13 cases with implicit causees, so we left those contexts out. CdEventSem describes the semantic class of the caused event. It can be mental or non-mental (physical or social). In case of metaphorical meaning, we assigned the semantic class that corresponded to the target domain. For example, in (3) the caused event was coded as mental: (3) Het doet het belletje rinkelen. it makes the bell_DIM ring ‘It rings a bell’. This variable was included to test whether doen is associated with the physical causation, as Speelman and Geeraerts (2009) suggested. EPTrans refers to the transitivity (including ditransitivity) or intransitivity of the effected predicate. The previous studies showed that laten is more favoured by transitive verbs, which was regarded as evidence for the indirectness of the causative situations indicated by this construction because it involved a longer causation chain with more participants.


The variable CeSynt was inspired by Kemmer and Verhagen’s (1994) observations about the laten-construction. In some contexts, the causee in Dutch allows not only for zero-marking as in (4a), but also for the prepositions aan and door, the dative and instrumental/agentive markers in Dutch, respectively, as in (4b) and (4c):2 (4) a. Hij liet zijn vrouw zijn nieuwe gedicht lezen. ‘He made/let his wife read his new poem’. b. Hij liet zijn nieuwe gedicht aan zijn vrouw lezen. ‘He let his wife read his new poem.’ c. Hij liet zijn nieuwe gedicht door zijn vrouw lezen. ‘He had his new poem read by his wife’. d. Hij liet zijn nieuwe gedicht lezen. ‘He had his new poem read’. Kemmer and Verhagen (1994) argue on the basis of cross-linguistic evidence that propositional or indirect-object marking of the causee implies a smaller degree of integration of the causee into the causative event and its lower affectedness in comparison with the default zero-marking (or, for personal pronouns, marking with the case of the direct object). This smaller integration and affectedness is typical of indirect causation. Therefore, we should expect prepositional marking to boost laten. On the other hand, Kemmer and Verhagen also suggested that implicitness of the causee, like in (4d), means even larger peripherality and non-affectedness of the causee (Kemmer and Verhagen 1994:â•›139). However, a more recent research study by Loewenthal (2003) has shown that implicit causees in the laten-construction have a moderate degree of affectedness, although the peripherality claim may still hold. Considering all this, and also the low frequencies of the prepositional marking, we distinguished two levels of the predictor: “Central” (the causee is explicit and unmarked) and “Peripheral” (the causee is implicit or marked with a preposition).

4. Statistical analysis Multiple logistic regression allows us to model the speaker’s behaviour by taking into account several factors that influence the speaker’s choice simultaneously. The analyses were carried out with the help of R statistical software (R Development Core Team 2010). The first step of multivariate analysis is the selection of variables that have an impact on the speaker’s choice. To select the relevant variables, we used the forward and backward stepwise selection procedures based on Akaike’s Information 2. Note that all three marking options are available only for one effected predicate, lezen ‘read’. The aan-marking is typical for verbs of perception and, consequently, mental causees, whereas the preposition door normally marks agentive causees.


Table 1.â•‡ Results of multiple regression (simple main effect model) Predictor

Estimate (log odds ratio)

(Intercept) CrSem = Inanimate EPTrans = Intransitive Country = BE CdEventSem = Mental CeSynt = Peripheral SubjectDomain = Football SubjectDomain = Music SubjectDomain = Politics

–4.38 (p < 0.001) â•⁄ 3.44 (p < 0.001) â•⁄ 1.48 (p < 0.001) â•⁄ 0.68 (p < 0.001) â•⁄ 0.79 (p < 0.001) –0.90 (p < 0.001) â•⁄ 0.12 (p = 0.38) â•⁄ 0.45 (p < 0.001) â•⁄ 0.35 (p = 0.009)

Criterion (AIC). This criterion helps strike a balance between the predictive power of a model and its parsimony. All predictors, except CeSem, entered the final model (see Table 1).3 The order of the variables in the table reflects their importance in predicting the speaker’s behaviour, as selected by the forward stepwise algorithm. The most important predictor is CrSem, and the least influential one is SubjectDomain. The column with the estimates provides the log odds ratios of the doen-construction for the given value of the predictor in comparison with the reference level (the values of the variables not mentioned in the table: the animate causer, the transitive effected predicate, the Netherlands, the explicit zero-marked causee, the non-mental effected predicate, and Economy as the article’s subject domain). If the log odds ratio is equal to 0, doen and laten have equal chances to occur, which means that the predictor is not informative. A positive value means that the chances of doen are higher for the given value in comparison with the reference level of the same predictor. A negative log odds ratio, conversely, stands for relatively higher chances of laten. For example, the inanimate causer increases the log odds ratio of doen in comparison with the reference level, the animate causer, by 3.40, which corresponds to the simple odds ratio of 29.96 (i.e. the chances for inanimate causers to occur in the doen-construction are almost 30 times as high as those of the animate causers). The p-values next to the estimate demonstrate how confident one can be that the estimate is not equal to 0: the lower the p-value, the more certain one can be. Conventionally, a value of α = 0.05 is used as a cut-off point for significant effects. The overall quality of the model is satisfactory, as the measurements in the lefthand column in Table 3 suggest. The most intuitive measure is the proportion of correct predictions of doen and laten by the model. We can correctly predict 90.1% of

3. Not all statisticians agree on the value of stepwise selection (e.g. Harrell 2001). However, analysis of the full model and single term deletion tests yield the same model structure.


the speakers’ choices (the cut-off probability is set to 0.5). However, this is not the most informative measure because laten is so frequent that if we simply predicted laten for all contexts, we would be correct in 82.8% of the cases, which serves as the baseline. There are special measures that neutralize the skewness, for example, the index of concordance C, also called the area under the ROC curve (see Hosmer and Lemeshow 2000:â•›160), which is 0.893 in our case. This number means that for all pairs of contexts with doen and laten, the model assigns a higher probability to the auxiliary actually observed in the context in almost 90% of the cases. C is equal to 0.5 if the predictions are random, and is equal to 1 if they are perfect. The related measures are Somers’ Dxy = 0.787 (rank correlation between the predicted probabilities and observed responses ranging from 0 to 1) and Goodman-Kruskal’s Gamma = 0.795 (with the range from –1 to 1). R2 is Nagelkerke’s generalized R2 index for logistic regression models, which is analogous to the measure of explained variation in linear models. It stands for a proportional reduction in the absolute value of the log-likelihood measure in comparison with the intercept-only model (see e.g. Menard 2001:â•›24–27 for more details). It ranges from 0 (no predictive power) to 1 (a perfect fit of the data). For this model, R2 = 0.523. All these values demonstrate that the model has a substantial predictive power. However, one more thing should be taken into account. The effect of some of the predictors on the response variable may be non-additive, i.e. it cannot be explained by the summary effect of the predictors taken separately. An example of such an interaction from health care is a situation when one and the same medical treatment produces different effects on patients depending on their sex or age. In this study, we focused on the interactions between the intralinguistic variables (semantic and syntactic ones). We selected a model with interaction terms on the basis of AIC. The procedure was performed in such a way that all five semantic and syntactic variables in all possible combinations had a chance to occur in the model. One three-way interaction CrSem:EPTrans:CeSynt turned out to be significant. Table 2 displays the model with the interaction (it also lists the relevant lower-order interaction terms). In a model with interaction terms, interpretation of the coefficients is less straightforward because we can no longer estimate an independent impact of a predictor that participates in an interaction without taking into account different levels of the other variables with which it interacts. For example, the estimate for CrSem = Inanimate in Table 2 should be interpreted as the combination of CrSem = Inanimate AND EPTrans = Transitive AND CeSynt = (the two latter terms are the reference levels of the corresponding variables). Table 3 lists the summary statistics for the two models. One can see that the predictive power of the model with interactions is slightly better in comparison with the main-effect only model, although not dramatically. This shows that our main-effect only model was informative, but too coarse to deal with some combinations of the predictor values.


Table 2.â•‡ Model with main effects and three-way interactions Predictor


(Intercept) CrSem = Inanimate (for EPTrans = Transitive and CeSynt = Central) EPTrans = Intransitive (for CrSem = Animate and CeSynt = Central) Country = BE CdEventSem = Mental CeSynt = Peripheral (for EPTrans = Transitive and CrSem = Animate) SubjectDomain = Football SubjectDomain = Music SubjectDomain = Politics EPTrans = Intransitive: CeSynt = Peripheral (for CrSem = Animate) CrSem = Inanimate: CeSynt = Peripheral (for EPTrans = Transitive) CrSem = Inanimate: EPTrans = Intransitive (for CeSynt = Central) CrSem = Inanimate: EPTrans = Intransitive: CeSynt = Peripheral

–3.59 (p < 0.001) â•⁄ 3.67 (p < 0.001) â•⁄ 0.41 (p = 0.051) â•⁄ 0.68 (p < 0.001) â•⁄ 0.78 (p < 0.001) –1.93 (p < 0.001) â•⁄ 0.16 (p = 0.27) â•⁄ 0.49 (p < 0.001) â•⁄ 0.34 (p = 0.014) â•⁄ 3.48 (p < 0.001) –0.31 (p = 0.437) â•⁄ 0.26 (p = 0.459) –2.60 (p < 0.001)

Table 3.â•‡ Summary statistics for two models Statistic

Model without interactions

Number of observations Proportion of correct predictions (baseline = 82.8%) C Dxy Gamma Generalized R2 AIC

6795 (doen: 1170, laten: 5625) 90.1% 90.3% 0.893 0.787 0.795 0.523 3695.8

Model with interactions

0.91 0.821 0.829 0.553 3525.2

Three-way interactions are hard to grasp intuitively, so we used another technique, named CART (Classification And Regression Trees), to visualize the interactions in a convenient way. The algorithm splits up the observations according to the values of each predictor, trying to separate the observations with doen from those with laten in the best possible way. It begins with the split that allows the cleanest separation, and then proceeds with the resulting subsets, choosing the next best split.

214 Natalia Levshina, Dirk Geeraerts, and Dirk Speelman CrSem=Anim

EPTrans=Tr

Laten 5111/318 CeSynt=Periph Laten 307/122

Doen 20/61

Doen 187/669

Figure 1.â•‡ Classification tree for semantic and syntactic variables

The procedure then cross-validates the resulting tree against different subsets of data, selecting the most parsimonious model with the purest “leaves”. Figure 1 shows the classification tree for our dataset (only the semantic and syntactic variables took part in the classification). It was implemented with the help of the rpart package in R.4 The minimum number of observations allowed in a split was 20; the algorithm performed 10 cross-validations. Under these conditions, only three of the five linguistic features take part in splits: CrSem, EPTrans and CeSynt. Recall that they are also the ones that interact significantly in the regression model. Each split is labeled with a decision rule, e.g. “CrSem=Anim”. If the condition is met, i.e. for all animate causers, one should follow the left branch; otherwise, the right branch should be explored. The names of the leaves display the predominant construction in the group; the numbers below stand for the number of laten- and doen-observations. The error rate of the classification was low: 9.5%, in comparison with 17.2% if we simply always predicted laten as the default auxiliary (cf. the baseline in Table 4). The first observation one can make is that doen needs more conditions to be met (inanimate causer AND intransitive effected predicate, or, in very few cases, inanimate causer AND transitive effected predicate AND explicit unmarked causee), whereas the animate causer is perfectly sufficient to obtain a leaf with a sufficient probability of laten, which also contains the largest number of observations. In addition, laten 4. An alternative solution is to use conditional inference trees. Their main advantage is that they neutralize the bias towards covariates with many possible splits (see Hothorn et al. 2006). However, all linguistic variables in the present analysis are binary, so this factor should not cause problems.


emerges in more specific situations with the inanimate causer, transitive effected predicate and peripheral causee. The classification also tells us that the features EPTrans and CeSynt are relevant for the classification only in the case of the inanimate causer. For animate causers, these features are not powerful enough to influence the outcome. In a similar way, the syntactic expression of the causee has a decisive effect only in the case of an inanimate causer and a transitive effected predicate.

5. Linguistic interpretation of the statistical models Some of our expectations based on the (in)direct causation hypothesis were confirmed by the marginal effects of the variables in the simple main effect model: inanimate causers, intransitive effected predicates and syntactically central causees do favour doen. However, we found no indication that the semantic class of the causee is significant, although one might expect that an animate causee is a better candidate for an indirect causation event because it is the main source of energy in the causation process. This lack of evidence creates a dilemma that is common in empirical studies (Geeraerts 1999). On the one hand, it may cast doubt on the indirect-direct causation hypothesis in the way it was formulated above. Alternatively, one could question our operationalization of the causee’s role in terms of animacy or inanimacy. The latter scenario seems to be more reasonable. The fact that doen is preferred by mental caused events suggests that the causation categorized with doen frequently involves animate causees as experiencers. In contrast, the laten-construction, preferred by more dynamic non-mental caused events, contains animate causees, who can play a more active, agentive role. Therefore, other, more sophisticated ways of determining the causee’s role could be helpful. The observed preference of doen by mental caused events is unexpected. In combination with inanimate causers, it seems that doen is highly associated with affective causation, which involves a stimulus (causer) that triggers a mental reaction of an experiencer (causee). This behaviourist-like causation type is very direct. However, Verhagen and Kemmer’s (1997) theory did not predict the predominance of affective causation within the semantics of doen; Speelman and Geeraerts (2009) even spoke about direct physical causation as doen’s specialization. Next, we calculated the probabilities of doen and laten for every configuration of the linguistic features, as predicted by the model with interactions. To do so, we first calculated the sum of the relevant estimates provided in Table 2, including the intercept value, and then transformed the resulting log odds ratios into probabilities.5 The

5. According to the formula P = exp(x)/(1+exp(x)), where x is the sum of the log odds ratios (coefficients) for all variables (the relevant values) and the intercept.


configurations with the highest probabilities of doen and laten can be illustrated by contexts (5) and (6), respectively. (5) Het artikel (…) deed mij terugdenken aan mijn ontmoeting met de Algerijnse ambassadeur in Brussel, voorjaar 1972 op een receptie in Den Haag. ‘The article (…) made me think back to my meeting with the Algerian ambassador in Brussels in the spring of 1972 at a reception in The Hague’. (6) Prinses Juliana zou in de jaren 60 liefdesbrieven aan haar dochters hebben laten analyseren door een grafoloog. ‘They say that in the 1960s Princess Juliana had love letters to her daughters studied by a graphologist’. Context (5), where the probability of doen is 89.7%, contains an inanimate causer, an intransitive effected predicate, an explicit unmarked causee and a mental caused event. This is a typical example of affective causation. Context (6) illustrates the configuration with the highest probability (99.2%) of laten, combining an animate causer, a transitive effected predicate, a peripheral causee and a non-mental caused event. The example evokes the service frame, when the causer uses the causee’s professional services to have some work done. Note that both examples contain animate causees, who play different semantic roles (an experiencer and agent, respectively). In his linear discriminant analysis of verb particle placement, Gries (2003) interprets the clusters of attributes with a highest distinctive load (highest discriminant scores of the sentences) as prototypes of each of the two constructions that he contrasts. Can we follow this approach and claim that contexts like (5) and (6) exemplify the prototypes of the constructions with doen and laten, respectively? Indeed, it was shown by Rosch and Mervis (1975) that the features that help maximally distinguish the given category from the others are also the features that are maximally shared between the members of the same category. The presence of these features, operationalized as a family resemblance score of a category member, also correlated positively with the member’s prototypicality in the category based on typicality ratings. This could lead us to the conclusion that the doen- and laten-observations such as (5) and (6), with the most distinctive features, should also be the best representatives of the categories from the intracategorial perspective. However, the more recent studies of natural language categories (e.g. Ceulemans and Storms 2010) show that these kinds of salience do not always correlate. The results of additional experiments (Levshina 2011) show that significant positive correlations between the intra- and intercategorial types of salience operationalized in several different ways are observed only in the case of doen, and not in the case of laten. A possible explanation is that laten, which is used much more frequently than doen, is also a more heterogeneous category. In a polysemous category, the sense that is the most semantically distant from a contrasting category may not be central intracategorially. In addition, one can imagine that the distinctive features of laten with regard to


another construction may be different from those with regard to doen. All this means that we should be very specific about the perspective and operationalization when using the term ‘prototype’. The configurations of the distinctive features exemplified by the sentences (5) and (6) are thus distinctive corpus-based ‘prototypes’ with regard to the choice between doen and laten modelled with the help of logistic regression. In addition, the analyses reveal that doen is indeed quantitatively and semantically restricted, as Speelman and Geeraerts (2009) wrote. This can be illustrated by the classification tree in Figure 1, which shows clearly that a larger number of semantic and syntactic conditions should be satisfied for doen to have more chances to occur in a context than laten. Therefore, doen seems to have more Gestalt-like semantics than laten, which has a looser set of semantic features. This conclusion can be supported by the fact that laten is a highly schematic auxiliary with a semantic range from permission to coercion. So far, we have not discussed the behaviour of the Subject Domain variable. The effect of the topic on the distribution of doen and laten was not predicted by any of the previous hypotheses. It would be natural to assume that different topics differ with regard to lexicon. This is why we looked at the top five most popular effected predicates in the four subject domains, which are shown in Table 4 with their relative frequencies. One can see that the four topics are dominated by the same highly frequent verbs: zien ‘see’, weten ‘know’, horen ‘hear’, denken ‘think’, liggen ‘lie’, vallen ‘fall’, gaan ‘go’ with different relative frequencies. This might not be a serious problem if these highly frequent verbs did not demonstrate an outspoken preference either for laten or doen. For instance, weten is a typical laten-verb, with 450 occurrences in our data, all of them with laten. Some previous research, e.g. Levshina et al. (2009), which was based on collostructional analysis (Stefanowitsch and Gries 2003), also showed that attraction between the effected predicates and the constructions (auxiliaries) is indeed very strong. This is why the differences in the relative frequencies of these influential predicates may have an effect on the distribution of doen and laten in the subject domains. One of the ways to capture this idiomatic difference is to incorporate the lexical “noise” in the model as random effects. This method, called mixed-effect modelling, has proved to be a powerful tool in linguistic research, especially in Table 4.â•‡ Top five most frequent effected predicates in the four subject domains Economy

Football

Music

Politics

zien ‘see’ 13% weten ‘know’ 11% liggen ‘lie’ 5% vallen ‘fall’ 3% stijgen ‘go up’ 3%

liggen ‘lie’ 9% zien ‘see’ 8% weten ‘know’ 5% vallen ‘fall’ 4% spelen ‘play’ 2%

horen ‘hear’ 11% denken ‘think’ 8% zien ‘see’ 4% klinken ‘sound’ 3% weten ‘know’ 2%

weten ‘know’ 10% zien ‘see’ 6% vallen ‘fall’ 4% gaan ‘go’ 2% denken ‘think’ 2%


psycholinguistic experiments with individual subject- and item-related noise (see a variety of case studies in Baayen 2008). It is also helpful in corpus-based studies when idiosyncrasies of individual words cannot be handled with the help of coarse-grained semantic classifications. Using the lmer package in R , we fit a mixed-effect with the effected predicates (1,165 types represented by 6,795 tokens) as random effects. By doing so, we “tell” the model that some effected predicates may inherently prefer doen, and that some verbs may prefer laten. The algorithm slightly lowers or increases the value of the intercept for each verb depending on its preferences in the data. It also takes into account the frequency of the verb in the data set. Ideally, we would have to do the same for the constructional slots of the causer and the causee. However, application of this method is less evident for nominal slots because of pronominal reference and lower type-token ratios. There is also evidence that the ties between nouns and constructions in which they appear are weaker than those between constructions and verbs (cf. Tomasello et al. 1997). Fitting a mixed-effect model (with main effects only) yields the results shown in Table 5. The factors and the tendencies that we had in the corresponding model without random effects remain very similar, with the exception of the subject domain, which ceases to contribute substantially to the model’s performance (according to AIC). Therefore, we can conclude that the difference in probabilities of doen and laten across different topics is due to the lexical effects. Also, most of the absolute values of the coefficients in the mixed model are slightly higher than those in the fixed-effect only model (see Table 1) because we have filtered out some part of the lexical “noise”, which caused overfitting in the initial model. The model demonstrates that the abstract features related to direct or indirect causation are still significant when conditioned on the lexical effects (cf. Bresnan et al. 2007:â•›87). At the same time, additional tests show that the random-effect model alone would allow the prediction of the choice of doen and laten correctly in a vast majority of cases (78% for the Netherlandic subcorpus and 74% for the Belgian data). This can be seen as evidence of strong exemplar effects at the level of the effected predicates. Table 5.â•‡ Logistic regression model with effected predicates as random effects Predictor


(Intercept) CrSem = Inanimate EPTrans = Intransitive Country = BE CdEventSem = Mental CeSynt = Peripheral SubjectDomain = Football SubjectDomain = Music SubjectDomain = Politics

–5.95 (p < 0.001) â•⁄ 4.22 (p < 0.001) â•⁄ 2.25 (p < 0.001) â•⁄ 0.76 (p < 0.001) â•⁄ 1.11 (p < 0.001) –0.83 (p = 0.003) Not Available


However, the best-performing model is the one with both the abstract and the lexical features. This finding is perfectly in line with the non-reductionist constructionist approach to language, which assumes that high-level generalizations coexist with low-level schemata in the speaker’s knowledge about constructions (Langacker 1987; Goldberg 1995, 2006).

6. Conclusion In this multivariate corpus-based onomasiological probabilistic study, we used logistic regression to find the factors that influence the choice between the causative constructions with doen and laten by speakers of Dutch. The analyses showed that the highest probability of doen is observed in the contexts of affective causation: an inanimate stimulus causing a conceptually and syntactically central cognizer to experience some mental state (an intransitive event). Conversely, the laten-construction has the highest chances of being observed when the causer is animate, the effected predicate is transitive, the causee is implicit or marked with a preposition (syntactically and conceptually peripheral), and the caused event is non-mental. This configuration can be exemplified by the service frame. The combinations of these features, which have the highest cue validity with regard to the choice between the categories, can be regarded as the distinctive prototypes of the constructions. Their relations with the other salience phenomena, such as family resemblance, goodness of membership, entrenchment, etc., are to be explored empirically, although there are indications that the inter- and intracategorial typicality measures tend to correlate for compact categories without rich polysemy (in our case, doen). We also found evidence of exemplar effects in categorization at the level of the lexemes that fill in the effected predicate slot, which can serve as powerful predictors of the speaker’s choice on their own. However, the best prediction is achieved when the model combines both the above-mentioned semantic generalizations and the lexemes. This supports the constructionist hypothesis that the mind stores linguistic knowledge at different levels of generalization. In addition, the results show that the doen-construction has more chances of being chosen by a Flemish speaker than by a Dutch one, which supports the previous findings by Speelman and Geeraerts (2009) for spoken data and can be explained historically (the Flemish variety is believed to retain more archaic features). The constructions also display different behaviour across the four subject domains, which, as the mixed-effect model demonstrates, can be explained by the domain-specific differences in the distribution of the effected predicates.


References Baayen, R. H. (2008). Analysing linguistic data: A practical introduction to statistics using R. Cambridge: Cambridge University Press. DOI: 10.1017/CBO9780511801686 Bresnan, J., Cueni, A., Nikitina, T., & Baayen, R. H. (2007). Predicting the dative alternation. In G. Boume, I. Kraemer, & J. Zwarts (Eds.), Cognitive foundations of interpretation (pp. 69– 94). Amsterdam: Royal Netherlands Academy of Science. Bouma, G., van Noord, G., & Malouf, R. (2001). Alpino: Wide-coverage computational analysis of Dutch. In W. Dalemans, K. Sima’an, J. Veenstra, & J. Zavrel (Eds.), Computational linguistics in the Netherlands 2000: Selected papers from the Eleventh CLIN meeting (pp. 45– 59). Amsterdam: Rodopi. Bybee, J. (2010). Language, usage and cognition. Cambridge: Cambridge University Press. DOI: 10.1017/CBO9780511750526 Bybee, J., & Eddington, D. (2006). A usage-based approach to Spanish verbs of ‘becoming’. Language, 82(2), 323–355. DOI: 10.1353/lan.2006.0081 Ceulemans, E., & Storms, G. (2010). Detecting intra and inter categorical structure in semantic concepts using HICLAS. Acta Psychologica,133 (3), 296–304. DOI: 10.1016/j.actpsy. 2009.11.011 De Sutter, G. (2009). Towards a multivariate model of grammar: The case of word order variation in Dutch clause final verb clusters. In A. Dufter, J. Fleischer, & G. Seiler (Eds.), Describing and modeling variation in grammar (pp. 225–254). Berlin & New York: Mouton de Gruyter. Degand, L. (2001). Form and function of causation. A theoretical and empirical investigation of causal constructions in Dutch. Leuven: Peeters. Geeraerts, D. (1999). Idealist and empiricist tendencies in cognitive semantics. In T. Janssen, & G. Redeker (Eds.), Cognitive linguistics: Foundations, scope and methodology (pp. 163– 194). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110803464.163 Geeraerts, D. (2006). Salience phenomena in the lexicon: A typology. In D. Geeraerts (Ed.), Words and other wonders: Papers on lexical and semantic topics (pp. 74–97). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110219128.1.74 Glynn, D., & Fischer, K. (2010). Quantitative methods in Cognitive Semantics: Corpus-driven approaches. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110226423 Goldberg, A. E. (1995). Constructions: A construction grammar approach to argument structure. Chicago: University of Chicago Press. Goldberg, A. E. (2006). Constructions at work: The nature of generalizations in language. Oxford: Oxford University Press. Gries, St. Th. (2003). Multifactorial analysis in corpus linguistics: A study of particle placement. New York: Continuum. Grondelaers, S., Geeraerts, D., & Speelman, D. (2007). A case for a cognitive corpus linguistics. In M. Gonzalez-Marquez, I. Mittleberg, S. Coulson, & M. Spivey (Eds.), Methods in cognitive linguistics (pp. 149–169). Amsterdam & Philadelphia: John Benjamins. Harrell, F. E. (2001). Regression modelling strategies with applications to linear models, logistic regression, and survival analysis. Heidelberg & New York: Springer. Heylen, K. (2005). A quantitative corpus study of German word order variation. In S. Kepser, & M. Reis (Eds.), Linguistic evidence: Empirical, theoretical and computational perspectives (pp. 241–264). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110197549.241


Hosmer, D. W. & Lemeshow, S. (2000). Applied logistic regression. New York: Wiley. DOI: 10.1002/0471722146 Hothorn, T., Hornik, K., & Zeileis, A. (2006). Unbiased recursive partitioning: A conditional inference framework. Journal of Computational and Graphical Statistics, 15(3), 651–674. DOI: 10.1198/106186006X133933 Kemmer, S., & Verhagen, A. (1994). The grammar of causatives and the conceptual structure of events. Cognitive Linguistics, 5(2), 115–156. DOI: 10.1515/cogl.1994.5.2.115 Langacker, R. W. (1987). Foundations of cognitive grammar: Volume I: Theoretical prerequisites. Stanford: Stanford University Press. Levshina, N. (2011). Doe wat je niet laten kan: A usage-based analysis of Dutch causative constructions [Do what you cannot let: A usage-based analysis of Dutch causative constructions]. Unpublished doctoral dissertation, University of Leuven. Levshina, N., Geeraerts, D., & Speelman, D. (2009). Collostructional analysis of Dutch causative constructions. Paper presented at the Third International AFLiCo Conference, 28 May, Paris. Loewenthal, J. (2003). Meaning and use of causeeless causative constructions with laten in Dutch. In A. Verhagen, & J. van de Weijer (Eds.), Usage-based approaches to Dutch (pp. 97– 130). Utrecht: LOT. Medin, D. L., & Schaffer, M. M. (1978). Context theory of classification learning. Psychological Review, 85(3), 207–238. DOI: 10.1037/0033-295X.85.3.207 Menard, S. (2001). Applied logistic regression analysis. Thousand Oaks: Sage. R Development Core Team (2010). R: A language and environment for statistical computing. Foundation for statistical computing. Vienna, Austria. . Rosch, E. (1975). Cognitive representation of semantic categories. Journal of Experimental Psychology, 104(3), 192–233. DOI: 10.1037/0096-3445.104.3.192 Rosch, E., & Mervis, C. B. (1975). Family resemblances: Studies in the internal structure of categories. Cognitive Psychology, 7(4), 573–605. DOI: 10.1016/0010-0285(75)90024-9 Speelman, D., & Geeraerts, D. (2009). Causes for causatives: The case of Dutch doen and laten. In T. Sanders, & E. Sweetser (Eds.), Causal categories in discourse and cognition (pp. 173– 204). Berlin & New York: Mouton de Gruyter. Stefanowitsch, A., & Gries, St. Th. (2003). Collostructions: Investigating the interaction between words and constructions. International Journal of Corpus Linguistics, 8(2), 209–243. DOI: 10.1075/ijcl.8.2.03ste Stukker, N. (2005). Causality marking across levels of language structure. Unpublished PhD dissertation, University of Utrecht. Tomasello, M., Akhtar, N., Dodson, K., & Rekau, L. (1997). Differential productivity in young children’s use of nouns and verbs. Journal of Child Language, 24(2), 373–87. DOI: 10.1017/S0305000997003085 Tummers, J., Heylen, K., & Geeraerts, D. (2005). Usage-based approaches in cognitive linguistics: A technical state of the art. Corpus Linguistics and Linguistic Theory, 1(2), 225–261. DOI: 10.1515/cllt.2005.1.2.225 Verhagen, A., & Kemmer, S. (1997). Interaction and causation: Causative constructions in modern standard Dutch. Journal of Pragmatics, 27(1), 61–82. DOI: 10.1016/S0378-2166 (96)00003-3

The semasiological structure of Polish myśleć ‘to think’ A study in verb-prefix semantics Małgorzata Fabiszak, Anna Hebda, Iwona Kokorniak, and Karolina Krawczak Adam Mickiewicz University, Poznań

The aim of the present chapter is to investigate the semasiological structure of the Polish verb myśleć ‘to think’ relative to the construal imposed by its prefixes. It juxtaposes the results of cognitive linguistic corpus-based introspective analysis with the results of a series of statistical tests of almost 4,000 manually coded example sentences. Multiple correspondence analysis employs the techniques presented in Glynn (this volume), hierarchical cluster analysis follows Divjak and Fieller (this volume) and logistic regression is based on Speelman (this volume). The study shows that the do- prefix has a strong preference for clausal, processual complements, while the wy- prefix opts for nominal complements, suggesting objectifiable results of the thinking process. Keywords: hierarchical cluster analysis, logistic regression, multiple correspondence analysis, Polish verbal prefixes, Polish verb myśleć ‘to think’

1. Introduction Given that it is a universal human capacity to think and to verbalize one’s thoughts in an intersubjective context, it is a reasonable assumption that a lexeme designating such a cognitive activity and state would be present in every language (Fortescue 2001:â•›15). However, despite its universality, an accurate explanation of the meaning and use of a cognition verb, such as myśleć in Polish or think in English, is a particularly challenging task. The difficulty in representing the semantic structure of such predicates stems from their conceptual schematicity. This schematicity is elaborated by instantiation in context to produce any one of the verb’s multiple senses. Along lexical and discursive lines, it encodes various loads of intentionality, modality, subjectivity, and in some of its uses, it is highly grammaticalized. Through detailed corpus-driven

224 Małgorzata Fabiszak et al.

conceptual analysis, our project takes up the challenge of systematically describing the semantic structure of prefixed think verbs in Polish.1 Specifically, the analysis examines the contribution of six prefixes to the semantics of the mental predicate relative to its object. The first stage of the study employs introspective analysis in order to establish a set of testifiable hypotheses about the semantic contribution of these highly schematic prefixes. The second stage turns to a quantitative analysis of a large number of occurrences, annotated manually for an array of grammatical and semantic features. The ‘behavioural profiles’ of the verbs that result from this analysis are identified through subsequent multivariate statistical modeling (Geeraerts et al. 1994; Gries 2006; Glynn 2010b). Thinking, in most general terms, can be perceived as “internalized (and abbreviated) speech”, which is thus tantamount to “self-awareness” (Fortescue 2001:â•›17f.). This “private, internal activity” can be further specified into at least three kinds of processes, which consist of “evaluating” someone or something, “believing in the truth of a proposition” or “‘mulling over’ some mental content” (Fortescue 2001:â•›30). On a metalinguistic plane, ‘think’ can be regarded as a language-independent semantic universal (Wierzbicka 1992, 1996, 1997, 1999), which has recently been corroborated in several cross-linguistic studies, e.g. by Amberber (2003) on Amharic or by Junker (2003) on East Cree, an Algonquian language. The question remains, however, as to which senses of ‘think’ should be universally constitutive of the core of human thought (Fortescue 2001:â•›32). It is, after all, indisputable that language-specific exponents of think are characterized by polysemy and that they enter onomasiological relations by way of “lexical elaboration” (Goddard 2003). Attempts have also been made to reduce such presumably irreducible concepts. Jackendoff (1983:â•›234ff., as cited in Fortescue 2001:â•›32) factorized such mental predicates to the formula [be [rep[x]][in y’s mind]]. This formula is similar to the metaphorical representation of the thought process, as analysed within Conceptual Metaphor Theory, where the mind is understood as a container for thoughts-objects (Lakoff and Johnson 1980; Szwedek 2007). Perhaps a good solution would be to view the mental verb ‘think’ as belonging to “a network of experiential categories” or as a “radial” or “polycentric” category, whose most focal sense should be “the lowest common denominator” shared by all senses of ‘think’ (Fortescue 2001:â•›32–35).

1. The EmBeR project focuses on the conceptualization of abstract categories, such as Em(otion), Be(lief), and R(eason) in Polish and English. It uses corpus-driven techniques and multivariate statistics in an effort to capture the complex semantic structure of the three categories. Team members include Małgorzata Fabiszak, Anna Hebda, Iwona Kokorniak, Karolina Krawczak, and Barbara Konat at Adam Mickiewicz University, Poznań, Poland. We are greatly indebted to Dylan Glynn for introducing us to the methods of multivariate statistics. Our thanks also extend to three anonymous reviewers and the editors of this volume for their constructive comments. Any remaining shortcomings remain our own.

The semasiological structure of Polish myśleć ‘to think’ 225

Linguistically speaking, ‘think’ is indeed characterized by semantic generality and impoverishment (Danielewiczowa 2002:â•›131), but its conceptual schematic structure is contextually expanded (Kustova 2000:â•›250). It is therefore of utmost importance to examine the structural and semantic characteristics of this mental verb, bearing in mind that even seemingly negligible differences in verb valency may be significant (Danielewiczowa 2000:â•›231). There is at least a twofold distinction to be drawn in the semasiological structure of ‘think’ between states and activities, which is formally realized in English in the use of aspects, where the progressive can only be applied to non-stative mental events (e.g. Vendler 1967:â•›110f.; D’Andrade 1987, 1995). This bipartite division leads to the treatment of the mind as a “processor” or “container” (D’Andrade 1987, as cited in Palmer 2003:â•›270). Goddard (2003:â•›112) goes beyond Vendler’s (1967) bifurcation into think about and think that, and specifies the semantic prime think into four subgroups, depending on their conceptual syntactic features, whereby: a. b. c. d.

X thinks about Y [topic of thought] X thinks something (good/bad) about Y [complement] X thinks like this: ——— [quasi-quotational complement] X thinks that [——— ]S [propositional complement]. (Goddard 2003:â•›112)

Danielewiczowa (2000) proposes a further distinction, showing that thinking can also be dynamic, intentional, knowledge-driven, factual or hypothetical. It can be conscious or overwhelmingly intensive and subconscious, it can induce change in the subject, be a point-of-time or period-of time event, refer to an object or be self-oriented. The object of thinking can be abstract or concrete, factual or hypothetical, verifiable or non-verifiable (Danielewiczowa 2002:â•›132f.). It can vary considerably in terms of content, length, cohesion, origin, and structure (Pawłowska 1981). Object semantics will also be affected by how the object is introduced. For example, the complementizer że [that] is likely to specify the object more than pronouns such as co [what], gdzie [where], dlaczego [why], which only hint at what the object relates to (Pawłowska 1981). Insofar as the grammatical form of mental verbs influences their “semantic potential” and “pragmatic effects” (Danielewiczowa 2000:â•›265), their aspect, object form and object semantics will naturally entail changes in the function and profiling of such predicates. It is the aim of the present study to identify the semasiological organization of and onomasiological relations between the instantiations of think relative to the construal imposed by prefixes. Obviously, the schematic semantics of the unprefixed predicate contributes to the content of each of the prefixed forms. It constitutes the superordinate meaning component, the “semantic primitive” or “building block” of the subordinate expressions (Wierzbicka 1992:â•›9). Their semantico-pragmatic load is specified and disambiguated by their prefixal modification. It is, therefore, expected


that the internal structure of such prefixed ‘think’ forms and the near-synonymous relations holding between them will be particularly distinct. Myśleć ‘think’, when unaccompanied by any prefix, is an imperfective verb. Its aspect may change into the perfective one by means of adding a prefix, such as do(domyślić się ‘guess sth’), po- (pomyśleć ‘think sth’), prze- (przemyśleć ‘think over’), roz- (rozmyślić się ‘change one’s mind’), wy- (wymyślić ‘come up with’), za- (zamyślić się ‘fall into deep thought’), and others. However, the ‘perfectivized’ verb form can be further ‘imperfectivized’ through suffixation (cf. Comrie 1976), e.g. domyślać się ‘guess’, rozmyślać ‘ponder’, wymyślać ‘invent’. In the present contribution, the semantics of the aforementioned prefixes relative to their perfective and imperfective aspects will be considered in quantitative terms in relation to their combination with objects. The procedure adopted here is in line with the tradition of corpus-driven Cognitive Linguistics, as practised by Dirven et al. (1982), Geeraerts et al. (1994), Gries (2003, 2006), Divjak (2006, 2010), and Glynn (2009, forthc.), to mention but a few. In accordance with this tradition, introspective conceptual analysis, if crucial and necessary, is viewed as insufficient, as it cannot produce falsifiable answers to research questions. It is, therefore, essential for a large number of examples of the object of study to be subjected to detailed manual conceptual analysis of their syntactic and semantic characteristics. Once this stage is completed, the annotated data can be submitted to exploratory and, subsequently, confirmatory multivariate statistical modelling. The purpose of this procedure is to find and verify “patterns of usage features” (Glynn 2009, 2010b), also known in the field as “feature configurations” (Geeraerts et al. 1994) or “behavioural profiles” (Gries 2006; Divjak and Gries 2006). Summing up, thinking in its simplicity and universality is a multidimensional mental event of states and processes, which can be dynamic, factual, hypothetical, consciously entertained by the subject, or unconscious, based on and expressing knowledge, or unfounded, oriented toward an object or objectless, intentional or unintentional. All these features contribute to the holistic semasiological structure of ‘think’, rather than portioning it indefinitely. In this article we seek to account for some of the active semantic components in Polish prefixed myśleć, mapping its meaning from usage events. Section 2 presents a conceptual linguistic analysis of the Prefix, Aspect and Object Form relations, illustrated with examples from the PWN Corpus of the Polish Language. Section 3 describes the corpus, Section 4 outlines the coding procedure, and Section 5 presents the multivariate analysis. The results of the usage-feature analysis are submitted to the exploratory statistical techniques: multiple correspondence analysis and hierarchical cluster analysis. Where applicable, the results are further supported by the confirmatory method of logistic regression analysis. The selection of these three statistical techniques is motivated by the different affordances they offer to researchers. Correspondence analysis has been developed for nominal categorical data and allows the visualisation of the correspondences between multiple factors – linguistic features. The most promising tendencies observed in the


correspondence analysis maps are then further tested with the confirmatory statistical method – logistic regression, which goes beyond the identification of correspondences and gives probability scores indicating which of the correspondences are significant and what is the contribution of the individual features to the overall pattern. Finally, cluster analysis offers a representation which can be best compared with the result of the introspective conceptual analysis offered in Section 2.

2. Introspective conceptual analysis of the prefixed forms of myśleć ‘to think’ in Polish At first sight, the grammatical constructions that the verb myśleć ‘think’ and its prefixed forms go into may appear to be arbitrary and unpredictable. However, as Langacker (1991:â•›294) points out, the unpredictability of grammar arises from the objectivist approach to semantics. By contrast, the Cognitive Grammar approach finds: semantic value in every one of its uses. Moreover, it is precisely because of their conceptual import – the contrasting images they impose – that alternate grammatical devices are commonly available to code the same situation. (Langacker 1991:â•›295)

Thus, a fine-grained analysis should elucidate the complex nature of grammatical structures (Langacker 1991:â•›294). Myśleć ‘think’ can be considered as an imperfective verb. Its aspect may change into the perfective one by means of adding a prefix, such as do- (domyślić się), po- (pomyśleć), prze- (przemyśleć), roz- (rozmyślić się), wy- (wymyślić), za- (zamyślić się), and others. However, these ‘perfectivized’ forms may undergo suffixal imperfectivization, resulting in the imperfective form, e.g. domyślać się, rozmyślać, wymyślać. The semantics of the aforementioned prefixes will be considered below and their combination with objects will be accounted for. The prefix do- may combine with verbs to indicate an approximation to a goal or result (Śmiech 1986:â•›90–91). In cognitive terms, it represents the path image schema in which the goal or final point of the path along which a movement takes place is in focus. Reaching the goal may involve encountering certain difficulties along the way, where the trajector (TR) makes every effort to achieve the goal regardless of any obstacles. The intensity of the effort is observed by Dickey (2009), who calls do- verbs “intensive-resultative”. In domyśla/ić się the TR corresponds with the subject of the main clause. What constitutes the landmark (LM) here is the metonymic gathering of information in the course of the thinking activity. Whether the activity is complete or not is indicated, in the case of domyśla/ić się, by the suffix, where -ać, as in (1a–b), construes an uncompleted process, and -ić , as in (2a–b), a completed process. The gathered information


in the LM position is usually rendered explicitly in sentences, i.e. either with a nominal phrase, as in (1a) and (2a), or with a clausal complement, as in (1b) and (2b). (1) a. Nie domyślał się przyczyny [Not DO-thought-masc.3.sg.impf refl cause-sg.gen sporu. [argument-sg.gen] ‘He did not guess what the reason for the argument was.’ b. Nie domyślał się, że ktoś jest w [Not DO-thought-masc.3.sg.impf refl that someone-nom is in] pokoju. [room-sg.loc]2 ‘He did not guess that someone was in the room.’ (2) a. Nietrudno było domyślić się przeznaczenia tego [easily was DO-think-inf refl purpose-gen this-gen] miejsca. [place-gen] ‘One could easily guess what the purpose of this place was.’ b. Zaskoczony, zupełnie nie kontrolowałem wtedy [Surprised totally not controlled-masc.3sg.impf then] twarzy i mogę się domyślić, co [face-gen and can-1sg.pres refl DO-think-inf what wyczytała z niej ta biedna out-read-fem.3sg.perf from it-fem.gen this poor-fem-3sg] brzydula [ugly girl-nom] ‘Surprised, I did not control my facial expression and I can guess what she read in it, this poor ugly girl.’ The nominal phrase that domyśla/ić się can be combined with takes the genitive case, which in one of its uses may conceptualize a goal that is being aimed at (Rudzka-Â� Ostyn 2000:â•›201). It is not due to chance, then, that do-predicates are called ‘attainment verbs’, as both their nominal complement and the prefix construe the attainment of a desired result (e.g. Dickey unpubl. manusc.:â•›15). Additionally, the reflexive pronoun in domyśla/ić się indicates that the experiencer of the process taking the subject position is both the instigator and the passive experiencer affected by the result of the process (Dąbrowska 1997:â•›93). The ‘beneficial’ or ‘favourable’ effect of the process on the subject has been observed by Przybylska

2. In the word-for-word gloss the prefixes are rendered with capital letters in their Polish form. We follow here Tabakowska (2003a).


(2006:â•›60–61), as well as by Tabakowska (2003b:â•›15), who considers it rather natural that “one does not, under normal circumstances, want to cause damage to oneself ”. Another prefix, po-, may form delimitative verbs to indicate (i) a short duration of an action; (ii) a limited nature of an action; or (iii) when combined with the dative reflexive pronoun sobie, an ‘affective’ state of the subject (Piernikarski 1975:â•›33). In combination with myśleć, po- tends to indicate sense (ii). It may change the verb into a perfective one, which according to Dickey’s (2000) ‘east-west aspect theory’ has the meaning of temporal definiteness, where a temporally definite event “is unique in the temporal fact structure of a discourse, i.e. (…) it is viewed as both (a) a complete whole and (b) qualitatively different from preceding and subsequent states of affairs” (Dickey and Hutcheson 2003:â•›27–28).3 When it comes to complementation, pomyśleć may take a clausal complement, as in (3a), or a prepositional phrase with o ‘about’, and a nominal phrase in the locative (LOC) case. It may refer to “dispersed motion in any direction relative to a landmark” in the abstract temporal domain (Radden and Dirven 2007:â•›321), thus indicating that the object of thinking is not directly focused on, as in (3b). In less frequent cases, the object of thinking is profiled by means of nad ‘over’ with the nominal in the instrumental (INSTR) case, where the preposition, in one of its spatial senses, construes a multiplex path with “a single point-like TR moving in a variety of directions (…) to trace a multiplex collection of paths, which in turn can be construed as a mass” (Dewell 1994:â•›375). Metaphorically, in relation to pomyśleć, this schema can be understood as considering alternative ways to follow before reaching a final solution, where only the solution is mentioned explicitly, as in (3c): (3) a. Z radością pomyślałaś, że [With joy-instr PO-thought-fem.2sg.perf that] spotkałaś mężczyznę, który lubi [met-em.2sg.perf man-gen who like-3sg.pres] cię taką, jaka jesteś. [you-acc such like be-2sg.pres] ‘You have thought with joy that you met a man who likes you the way you are.’ b. Gdyby miał normalne warunki do pracy, [if had-masc.3sg normal conditions-acc to work-gen] nawet nie pomyślałby o wyjeździe. [even not PO-thought-masc.3sg.cond about departure-loc] ‘If he had normal working conditions, he wouldn’t even think about going abroad.’ 3. However, Śmiech (1986:â•›18–19) adds that a verb with the po- prefix, although taking a perfective form, may behave like an imperfective one. In the case of pomyśleć, the process seems to be completed and the verb form is perfective.


c. Gdyby chociaż można było gdzieś tam [if at least can-3sg.pres was-3sg somewhere there] dostać pracę to może pomyślałbym [get-inf job-acc then perhaps PO-thought-masc.3sg.cond] nad powrotem [over comeback-instr] ‘If it were possible to get a job somewhere there, I would think about coming back.’ Prze-, yet another prefix, in the spatial domain may depict a three-dimensional and bounded LM, such as a tunnel in which the TR moves from one end to the other, where the TR “gradually fills the whole volume of the landmark” (Pasich-Piasecka 1993:â•›19). When extended metaphorically, prze- may carry the meaning of ‘entirely through’, as observed by Pasich-Piasecka (1993:â•›18). In przemyśleć, the prefix not only implies the in-depth nature of the mental activity, but also points to its completeness. The TR does not move in the physical, but in the temporal space here, covering a certain period of time to perform the activity. The entity depicted is in the accusative case, as in (4), constitutes the LM which is directly affected by the experiencer of the process: (4) Wszystko przemyślał, wszystko [all-acc PRZE-thought-masc.3sg.perf all-acc rozważył considered-masc.3sg.perf] ‘He thought everything over, he considered everything.’ As Janda (1993:â•›143–145) points out, the accusative is the case marker which shows the dynamicity of the process and perfectly reflects the fact that the entity taking it is directly influenced by the TR and, as a result, may undergo certain changes. The prefix in the verb przemyśleć points to the completeness of the temporally extended process, which quite often may be complemented by other elements in the sentence, such as cały ‘whole’, wszystko ‘all’, etc. Due to its extended and complete sense, przemyśleć belongs to the perdurative group of verbs (Grochowska 1979:â•›70; Przybylska 2006:â•›160; Dickey 2009, unpubl. manusc.). Roz-, in its basic image schema, represents the TR and LM constituting one entity before a change, which takes different forms afterwards. Thus, the comparison of the two states of the entity before and after the change profiles different senses of roz-, which are to a great extent determined by the verb they go with (Przybylska 2006:â•›201). When in combination with roz-, myśleć may take either an imperfective or a perfective form. In the former case, the observed change is in the intensity of the mental activity. As Przybylska (2002:â•›271) notes, the activity is represented in both the basic and the prefixed form of the verb; the difference may lie in the duration and intensity of the activity. The same observation is made by Dickey (unpubl. manusc.:â•›14),


who calls this type of verbs ‘procedural’ as they “do not alter the basic lexical meaning of the source verb” and modify only the aforementioned aspects. Compare the following sentences (invented examples): (5) a. Piotr myślał o Marysi. [Peter-masc.nom.sg was thinking-3sg.impf about Mary-fem.loc.sg] ‘Peter was thinking about Mary.’ b. Piotr rozmyślał o [Peter-masc.nom.sg through-was thinking-3sg.impf about Marysi. Mary-fem.loc.sg] ‘Peter mused/was musing about Mary.’ In both (5a) and (5b), the verbs take the imperfective form; in the latter case, however, the prefix emphasizes the extendedness of the process. The dispersed nature of the mental activity is additionally highlighted by the preposition which rozmyślać may go with. Just like pomyśleć, it may be combined with o ‘about’ and the nominal phrase in LOC, as in (5b), or with nad ‘over’ and the substantive in INSTR, as in (5c). The prepositions and the case markers show that the objects are not directly affected by the process. The prepositions, whose senses discussed earlier for pomyśleć are also relevant for rozmyślać, profile either proximity or a variety of directions to follow, respectively. The cases, on the other hand, indicate just the area, but not the direct focus, of attention. (5) c. Otarł je szybko i, zacisnąwszy [Wiped-masc.3sg.perf them fast and clenching] szczęki, rozmyślał gorączkowo nad [jaws-acc was ROZ-thought-masc.3sg.impf frantically over sposobami ratunku ways-INSTR help-gen] ‘He wiped them fast and, clenching his jaws, he frantically pondered over the ways to help.’ d. Zrób sobie kawę i przestań bezproduktywnie [Make-imp self-dat coffee-acc and stop-imp unproductively rozmyślać! ROZ-think-inf] ‘Make yourself a coffee and stop pondering unproductively.’ As can be observed, the sense of roz- is strongly related to the prepositional senses it goes with. However, rozmyślać does not have to take any object and it still implicitly shows that the TR’s (subject’s) thoughts go in more than one direction, as in (5d). Roz- can also take a perfective form with myśleć, resulting in a reflexive verb rozmyślić się ‘change one’s mind’. The observed change in the subject’s mental state is


between the ‘normal’ process of the mental activity represented by the unprefixed form into the ‘changed’ mental state represented by the prefixed form. The reflexive pronoun emphasizes the internal mental change of the subject, which may also bring about a change in the subject’s behaviour, frequently conceived of by observers as a negative change (Przybylska 2002:â•›279–280). The behavioural change may be considered to be the oblique result of the mental activity (Dickey unpubl. manusc.:â•›15). Consider (6), where Salisz’s change of mind may result in his change of behaviour and his calling another person back, reflected in (6) by zawoła go z powrotem ‘call him back’: (6) Pędził co sił w obawie, że Salisz [Rushed-masc.3sg.impf what strength-pl.gen in fear-loc that Salisz się rozmyśli, że zawoła go z refl] [ROZ-think-3sg.fut that call-3sg.fut him-acc with powrotem. return-instr] ‘He ran with all his might for fear that Salisz would change his mind and call him back.’ The prefix wy- in the spatial domain construes the TR’s emergence from the LM, or its coming into existence by leaving the bounded region of the LM, thus evoking the container image schema. Some correspondences can be found between wy- and English out of, which in its prototypical image schema depicts “removal or departure of one concrete object from within another object or place” (Lindner 1983:â•›81). In combination with myśleć, wy- metaphorically refers to a mental state as a result of which one, as in (7a), or more ideas, as in (7b), emerge from one’s mind: (7) a. Sama wymyśliłam przepis na moje [alone-fem WY-thought-fem.3sg.perf recipe-acc on my najukochańsze ciasto. lovliest-acc cake-acc] ‘I came up with the recipe for my favourite cake myself.’ b. Co pewien czas panny wymyślały takie [What some maidens-nom WY-thought-fem.3pl.impf such time wieloznaczne słowa. ambiguous-acc] [words-acc] ‘Every now and again girls came up with such ambiguous words.’ In the perfective form of the verb (7a), one focuses on the effect of the process of thinking, indicating also its thorough nature. The result is reflected in the direct object position, representing the LM, which takes the accusative marking, usually in the singular form, to profile a single completed event. In the imperfective form of the verb (7b), however, the distributive nature of wy- is more conspicuous (Śmiech 1986). As Dickey (unpubl. manusc.) observes, in


this case “closure is either not reached or not salient in the conceptualization, e.g. for states, events in process, [or] habitually repeated events”. Whether the mental activity is in constant process, producing regular results, or whether it can be considered as a habitually repeated event should not really matter; what counts in this case is that the effect, represented by the nominal phrase in the accusative case, frequently takes the plural form, emphasizing the distributive nature of the mental process. The prefix za-, among many other senses distinguished by Tabakowska (2003a:â•›166–172), can represent a construal of ‘excess’ with intransitive perfective verbs, being extended from the sense of ‘going beyond a boundary’. In the group of intensive-resultative verbs, Dickey (unpubl. manusc.:â•›15) classifies za… się (lit. ‘behind’… refl) verbs as absorbtives, as they construe a continuous process whose subject, by becoming deeply engrossed in the activity, loses control over it. As the mental activity occurs independently of the TR’s will, some adverse consequences, or, in other words, oblique results, may be expected to take place. In the analysis of the prefixed verbs and the objects that they take, one should be able to observe that the thinking activity may give either direct or oblique results, as distinguished by Dickey (2009, unpubl. manusc.), thus forming two endpoints of a continuum. The former – giving a direct, and usually positively loaded, effect of the mental process – is represented by the accusative case and is combined with wymyślić/ ać ‘think out’ or przemyśleć ‘think through’. Then comes domyślić/ać się ‘guess’, indicating the goal of the process in the genitive case. It is followed by pomyśleć o+LOC/ nad+INSTR ‘think about/over’ or rozmyślać o+LOC/nad+INSTR ‘ponder about’. Objects in both the locative, construing an indirect constant movement of the TR round the focal point, and the instrumental, not indicating the goal of the action, but serving wyprze-

do-

ACC

GEN

direct result

porozo+LOC/ nad+INST

podo-

rozzaREFL Intr

Clause

oblique result

Abbreviations: ACC – Accusative; GEN – Genitive; LOC – Locative; INST – Instrumental; REFL – Reflexive; and Intr – Intransitive.

Figure 1.â•‡ Relationship between prefixed verb and the affectedness of the object4

4. Domyślić się and zamyślić się are reflexive verbs, which may take complements, while rozmyślić się is reflexive and intransitive.


to manifest it (Janda 1993:â•›143), are not directly affected by the action, and thus appear to be more oblique. Neither is the clausal object that comes together with pomyśleć, że…‘think that…’ or domyślić się, że… ‘guess that…’ Rozmyślić się ‘change one’s mind’ and zamyślić się ‘be lost in thought’ constitute the end of the continuum, taking oblique objects. Their presence is not mentioned in sentences at all; however, the verbs themselves, as Dickey (2009:â•›27) points out, may indicate an adverse or harmful result of the process, which represents the most extreme degree of subjectification, as understood by Langacker (1999). Introspective and qualitative analysis of corpus examples, the results of which are represented above in Figure 1, allows us to formulate the following hypothesis: wy-, prze-, do- myślić (się) on the one hand, and po-, do-, roz-, za-myslić (się) on the other, are conceptually similar. This hypothesis will be explored with correspondence analysis and hierarchical cluster analysis in Section 5. First, however, we will devote some space to the description of the corpus and the coding procedure. At this juncture, it is also noteworthy that our research question is further specialized, as intransitive reflexive forms, which do not take any object, will be excluded from the present study due to its exclusive focus on the relation holding between the prefixed forms of Polish ‘think’ and the semantics of their objects.

3. The corpus The lexico-grammatical analysis of the lexemes under study has been conducted on the online version of the PWN Corpus of Polish. Glynn (2009:â•›84) stresses the importance of representativity and corpus size in lexical semantic studies: Content words repeat infrequently and lexical variation is typically sensitive to extra-linguistic factors. These two conditions mean that for a lexical semantic study to capture any degree of semantic subtlety of even the most common usages associated with a given lexeme, the corpus must be large and preferably representative of various types of language and register.

It is essential, however, that we be aware that “even the largest corpus in the world is but a microscopic fraction of actual language use” (Glynn, this volume). The PWN Corpus of Polish consists of 40 million words sampled from 386 books, 977 issues of 185 newspapers and magazines, 84 recorded (and transcribed) conversations, 207 websites, as well as several hundred leaflets. It is, therefore, relatively balanced both in terms of genre and topic distribution. It must be noted, however, that it is biased toward unspontaneous written language. Tables 1 and 2 show the percentage ratios of particular text types and topics in the corpus. Unfortunately, there is no dedicated tool that would allow the researcher to tag the genre and topic of specific word uses automatically. To avoid the misattribution


Table 1.â•‡ Genre distribution in the PWN Corpus of Polish Source

Percentage of corpus

Number of text samples

fiction non-fiction newspapers and magazines spoken leaflets Internet (websites, blogs, chats, forums)

20 21 45.5 â•⁄4.5 â•⁄5.5 â•⁄3.5

195 192 997 issues/185 titles â•⁄84 272 files 207

Table 2.â•‡ The PWN Corpus of Polish: Topic distribution Subject matter

Percentage

Subject matter

Percentage

philosophy, religion history, geography literary criticism, linguistics sciences

â•⁄7 17 â•⁄9 â•⁄9

politics, economy social sciences applied sciences arts other

14 â•⁄5 â•⁄8 â•⁄5.5 25.5

of examples to the specific topics on the basis of limited co-text, we did not include these variables in our analysis. It is thus strictly a semantic-syntactic analysis. The tables above are only provided in order to give the reader an idea of what the corpus represents. The study has been designed in such a way so as to examine the relationship and potential correlation between the suffixes do-, po-, prze-, roz-, wy- and za-, the aspect, object form and object semantics. With that purpose in mind, the dataset has been manually tagged in order to measure the extent to which the lemmas in question may be synonymous.

4. Feature annotation As observed by Divjak (2006:â•›22), a project aimed at establishing the degree of similarity between lexical items requires “precise syntactic and semantic data on the distribution of the potentially near-synonymous lexemes over constructions and of their collocates over the slots of those constructions”. Therefore, after all the contextualized instances of the investigated lexemes were extracted from the corpus, each of them was “translated into metalanguage”, as Divjak (2006:â•›21) puts it. In other words, every occurrence of domyślać, domyślić (się) pomyśleć, przemyśleć, przemyśliwać, rozmyślać, rozmyślić, wymyślać, wymyślić and zamyślić was analysed and annotated for a number of formal and semantic properties. Laborious and time-consuming as it was, manual


tagging of the data was necessary for securing the presence of all the elements crucial to the semantic description of a lexeme. This approach, referred to as ‘usage-feature’ analysis (Glynn 2010c:â•›8), has proven successful in corpus-driven Cognitive Linguistic research (Gries 2003, 2006; Divjak 2006, 2010; Gries and Stefanowitsch 2006; Grondelaers et al. 2007; Glynn 2009, 2010a, in press; Speelman et al. 2009; Glynn and Fischer 2010). The decision as to which variables and features to code was made on the basis of relevant literature, including, among others, Vendler (1967), Divjak (2006), Gries (2006), and Glynn (2009, 2010b). The annotation used by Divjak (2006), Gries (2006), and Glynn (2009, 2010b) proved particularly informative. According to Divjak (2006:â•›23), what is essential is that “the main participants of events or situations are encoded”. At first, the occurrences were annotated for a limited, yet diverse, variety of morphological, syntactic and semantic features, including: tense, aspect, and voice; transitivity, mood as well as the subject’s and complement’s form, humanness, animacy, abstractness, etc. (see Gries 2006:â•›73, 75; Divjak 2006:â•›34–36; and Glynn 2010b:â•›245–249 for more examples). With time (and the growth of the coders’ expertise), new categories were added to the coding schema. The description of the morphological features now also includes the prefix, person and number of the verb. As for syntactic properties, apart from mood and transitivity, negation has been incorporated into the adopted usage-feature set. Finally, with regard to the semantics of subjects and objects attested in the corpus, the following categories were added: subject visibility, competence and specificity, as well as some further formal features such as object case/person/number, and (in the case of clausal complements) clause semantics. Given that it is the mutual dependence between the prefix, the aspect and the complement type that lies at the heart of the present analysis, every direct object, be it a noun phrase, a clause, or a reduced clause, was decomposed into its formal and semantic characteristics. Consequently, NP complements (coded for case and number) were labelled as human (e.g. Piotr ‘Peter’/detektyw ‘detective’), concrete (e.g. potrawa ‘dish’), or abstract (e.g. plan ‘plan’/rozwiązanie ‘solution’), while clausal complements were grouped into achievements (i.e. single moment actions e.g. die), accomplishments (i.e. durative, heterogeneous processes, e.g. reach a verdict), states (e.g. be, have), activities (i.e. durative and uncompleted actions, e.g. to be hiding in the shed), or hypothetical (i.e. conditionals or clauses in the future tense).5 Pronominal direct objects were tagged for case, number and person (if applicable) and their co-referents were classified according to the same principles as NP complements. In the analysis, however, it turned out that object form (NP or clause) and object semantics were the most informative. NP object case is not a feature independent of the prefix (a particular prefixed form takes a

5. The first four clausal categories were distinguished on the basis of Vendler (1967).


Table 3.â•‡ Prefixed forms of myśleć and their co-occurrence with Object Forms Predicate

NP

Clause

Reduced Pronoun Narrative Intrans. Lexeme Lemma Clause total total

domyślać się domyślić się pomyśleć przemyśleć przemyśliwać rozmyślać rozmyślić się wymyślać wymyślić zamyślić się Total

â•⁄â•⁄70 â•⁄â•⁄28 â•⁄155 â•⁄115 â•⁄â•⁄12 â•⁄â•⁄56 â•⁄â•⁄â•⁄0 â•⁄168 â•⁄569 â•⁄â•⁄14 1187

â•⁄338 â•⁄237 â•⁄570 â•⁄â•⁄10 â•⁄â•⁄â•⁄5 â•⁄â•⁄38 â•⁄â•⁄â•⁄0 â•⁄â•⁄â•⁄0 â•⁄â•⁄78 â•⁄â•⁄20 1296

â•⁄3 â•⁄3 46 â•⁄0 â•⁄0 â•⁄0 â•⁄0 â•⁄0 â•⁄0 â•⁄1 53

â•⁄51 â•⁄22 â•⁄81 â•⁄86 â•⁄â•⁄7 â•⁄28 â•⁄â•⁄0 â•⁄36 373 â•⁄â•⁄5 689

â•⁄2 â•⁄1 57 â•⁄0 â•⁄1 â•⁄9 â•⁄0 â•⁄0 â•⁄0 â•⁄0 70

â•⁄78 â•⁄59 â•⁄89 â•⁄14 â•⁄â•⁄3 â•⁄49 â•⁄64 â•⁄59 â•⁄37 145 597

â•⁄542 â•⁄350 â•⁄998 â•⁄225 â•⁄â•⁄28 â•⁄180 â•⁄â•⁄64 â•⁄269 1057 â•⁄185 3892

â•⁄â•‹892 â•‹â•⁄ 998 â•‹â•⁄ 253 â•‹â•⁄ 244 1,326 â•‹â•⁄ 185 3892

particular object case). NP object number did not seem to interact in any regular ways with the prefixes. So that is why these two features were later disregarded. Altogether, a total of 3,982 tokens were extracted from the corpus as representative of the six investigated lemmas in their two aspectual forms. Table 3 illustrates the distribution of the lemmas in the dataset as well as the number of NP/pronominal/ clausal complements for every type. Table 3 shows that not all the prefixed forms are equally frequent, with only 28 examples of przemyśliwać and as many as 1,057 tokens of wymyślić. Relative to the representativeness of the corpus, we could claim that, visà-vis its relative frequency, wymyślić is the most conspicuous of the analysed lexemes. Moreover, the distribution of the same prefix between the two aspects is far from balanced. As many as 61% of the do- forms are Imperfective, as if the process of reaching the goal indicated by do- was more often focused on than the very act of having reached it. Even though the prefix prze- focuses on the period of time covered by the verb, it also emphasizes the completion and is thus far more commonly used in the Perfective Aspect (89% of cases). Since we focus on the verb-object interaction here, we disregard the intransitive uses of roz-. Considering its transitive uses, roz- is always Imperfective, po- is always Perfective, while za-, unlike po-, can potentially co-occur with either aspect. In the investigated corpus it is also restricted to the Perfective Aspect. For wy- the ratio is 80% Perfective to 20% Imperfective. This may suggest that it is only do- which can be either Perfective or Imperfective, while the other prefixes show a clear preference for one Aspect: po-, prze-, wy- and za- for the Perfective and roz- for the Imperfective. It can be explained by the processual, dispersed nature of roz-, as indicated in the introspective part of the study. As the aim of this paper is to explain the interaction of Prefix and Object grammar and semantics in the creation of meaning, the intransitive forms will be discarded from further analysis. Also, the Narrative selects almost exclusively for the po-prefix,


so there is no need to investigate their interaction with other prefixes. The Reduced Clauses are rare; hence, these occurrences are grouped with full Clause complements. In the cases where the semantics of the pronominal object could be identified and coded, the examples were included in further analysis. Those cases, however, whose semantics could not be analysed, were deleted. Following this, the pronominal objects were re-coded as NPs when they stood for NPs and as Clauses if they stood for clauses, as in the construction myśleć o tym, że … ‘to think about it that…’. As will be shown in the sections that follow, a careful analysis of all the annotated features and their subsequent submission to multivariate statistics have enabled us to identify “patterns of usage” (Glynn 2009). Apart from producing results corresponding to and, therefore, corroborating the hypothesis presented in Section 2, the exploratory statistics applied here will also reveal interesting variation in usage not predicted by the introspective analysis. This demonstrates that introspection does indeed need to be complemented by quantitative measures to produce more comprehensive results.

5. Multivariate analysis of the results of feature annotation In this section, our data are further explored with two multivariate techniques: correspondence analysis and cluster analysis (for descriptions of these methods see Glynn, this volume, and Divjak and Fieller, this volume). Then, the confirmatory method of logistic regression analysis is used to corroborate the results of the correspondence analyses. These will enable us to investigate the contribution of Object Semantics and other essential factors to the holistic meaning structure of the prefixed cognition verb myśleć. The section ends with a discussion of the interaction of all the analysed variables. First, a chi-square test is applied to the data to check if the variation between the lexemes is significant: Pearson’s chi-square test χ² = 3803.6, df = 55, p-value < 2.2e-16 The chi-square test demonstrates that there is significant variation between the lexemes under investigation. In Table 4, the inspection of Pearson’s residuals shows that the correlation between the prefix do- and Imperfective Aspect is the most important, as is that between po- and Perfective Aspect. In regard to Object Form, the strongest and most significant correlations hold between do- and Object Clause, and wy- and Object NP. Finally, with respect to Object Semantics, the prefix docorrelates importantly with Accomplishment, Activity and State, po- with Achievement and Hypothetical clauses, while wy- correlates with Concrete and Abstract, prze- with Abstract, and roz- with Human nominal objects.


Table 4.â•‡ Pearson’s residuals for the prefixed verb forms of Polish myśleć ‘to think’ Imperfective Perfective Object Clause Object NP Object ABSTRACT Object ACCOMPLISHMENT Object ACHIEVEMENT Object ACTIVITY Object CONCRETE Object HUMAN Object HYPOTHETICAL Object STATE

do-

po-

prze-

roz-

wy-

za-

â•⁄17.014 –10.499 â•⁄11.915 –11.959 â•⁄–9.362 â•⁄â•⁄3.871 â•⁄â•⁄3.91 â•⁄â•⁄6.095 â•⁄–6.118 â•⁄–2.588 â•⁄–3.637 â•⁄12.882

–14.556 â•⁄â•⁄9.125 â•⁄10.229 –10.95 –13.081 â•⁄–0.009 â•⁄â•⁄4.404 â•⁄â•⁄4.426 â•⁄–2.245 â•⁄â•⁄0.122 â•⁄10.093 â•⁄â•⁄4.57

–3.739 â•⁄2.251 –6.088 â•⁄6.305 â•⁄8.166 –0.671 –1.931 –1.246 –2.708 –1.401 –2.06 –4.853

14.704 –9.035 –0.772 â•⁄0.865 –0.273 –0.795 â•⁄0.039 –1.165 –1.902 â•⁄5.578 â•⁄3.509 –2.847

â•⁄–4.415 â•⁄â•⁄2.611 –16.923 â•⁄17.488 â•⁄16.819 â•⁄–2.628 â•⁄–6.608 â•⁄–8.353 â•⁄â•⁄9.222 â•⁄â•⁄0.633 â•⁄–6.625 –12.313

–3.134 â•⁄1.902 â•⁄0.639 –0.609 –0.775 –0.823 â•⁄0.074 –0.124 –0.454 â•⁄0.813 â•⁄1.322 â•⁄0.086

Let us now turn to correspondence analysis in order to see how these factors interact with one another. Figure 2 presents the results of a multiple correspondence analysis. It employs a Burt matrix/‘joint’ method to correct for low explained inertia (Greenacre 1993). The explained inertia is 81.2% (Dim. 1 = 68.4, Dim. 2 = 12.8), which is a stable result in multiple correspondence analysis. As was predicted by the analysis of the results of Table 4, the Imperfective Aspect correlates highly with the do- and roz- prefixes. The combination of the goal-indicating do- with the Imperfective Aspect shows the focus on the processual nature of

roz–

1.50

IMPERF do–

0.75 ABSTR NP

0.00

wy–

ACT

HUM

ACCOM CLAUSE

prze– PERF–

CONCR

–0.75

HYPO

za–

–1

STATE ACH

–0.5

0

po–

0.5

1

Figure 2.â•‡ Multiple correspondence analysis: Prefixes of the Polish verb myśleć ‘to think’ and their correlation with Aspect, Object Form and Object Semantics


the verb domyślać się in the conceptualization. The prefix roz- indicates the dispersed nature of the verb it is attached to and, in combination with the Imperfective, emphasizes the extensive nature of the thinking process. As for the hypothesized correlation between wy-, prze-, do- myślić (się), on the one hand, and po-, do-, roz-, za-myślić (się), on the other (see Section 2 and Figure 1), Figure 2 partly confirms these assumptions, which were arrived at via theoretical and introspective reasoning. Firstly, the prefixes prze- and wy- in Figure 2 are grouped together relative to nominal objects of thought. Secondly, za- and po- also correspond to one another relative to the Perfective Aspect and hypothetical clausal objects of thought. Finally, there is a third correlation discernible between do- and roz-, which were originally expected to belong together with the previous group; here, they are clustered together relative to the Imperfective Aspect. In the next step of data analysis, we will combine two factors: Aspect and Prefix, with a view to visualizing more clearly the perfective and imperfective uses of the relevant prefixed predicates. Accordingly, the next plot presents the binary correspondence analysis of Prefix combined with Aspect, Object Form and Object Semantics. The binary correspondence in Figure 3 clearly captures the dispersion of object types relative to the prefixes. It is a stable analysis, with 95% explained inertia (Dim. 1: 67.16, Dim. 2: 27.84). The plot shows that nominal Abstract objects correlate closely with Perfective prze- and Perfective and Imperfective wy-, which are plotted to the left of the centre. Above this grouping, we can see that Concrete nominal objects are most closely correlated with the Perfective wy-, while Human nominal objects are in distinct correlation with Imperfective roz-. This is made evident by the semantic feature being pushed away from the rest of the plot and located just above the prefix in question. All these correlations are illustrated in examples (8a–e) below: (8) a. Przemyślałam jeszcze raz całą sytuacje od początku i doszłam do tego samego. ‘I thought this situation over once more and came to the same [conclusion].’ (prze+Perf+ObjNP+ObjAbstr) b. Jeżeli zorientuje się, że wróciłyśmy, wymyśli jakieś usprawiedliwienie. ‘If he realizes we have come back, he will invent some excuse.’ (wy+Perf+ObjNP+ObjAbstr) c. Bo jak facet 500 lat temu mógł wymyślic helikopter lub rower? ‘How could a guy invent a helicopter or a bike 500 years ago?’ (wy+Perf+ObjNP+ObjConcr) d. Mówi, że fryzjerstwo jest twórcze: wymyśla własne patenty, na przykład układanie na krem Nivea, dla efektu i z braku żelu. ‘He says that hairdressing is creative: he invents his own patents, for example, doing hair with the use of the Nivea cream, for effect and for want of gel.’ (wy+Imperf+ObjNP+ObjAbstr)


e. Patrzyła na mnie i rozmyślała o Witku. ‘She was looking at me and thinking about Witek.’ (roz+Imperf+ObjNP+ObjHum) The mental predicate przemyśleć in (8a) designates a thorough, temporally extended activity coming to its end, which may bring a change to the object affected by the mental process. The choice of abstract objects of thought seems justifiable for such an in-depth analysis. It also stands to reason that the resultative, goal-oriented prefix wyattracts abstract and concrete nominals (8b, c), which profile the effect of the thinking process in a reified and comprehensive way. The imperfective form wymyślać, rather than focusing primarily on the effect of thinking, emphasizes the distributive nature of the process, in the example here additionally reinforced by the plural use of the nominal in the object position (8d). It also makes perfect sense (example (8e)) why the prefix roz-, accentuating the extensive nature of the mental activity, should tend towards objects designating humans, normally our loved ones. Now, when moving to the right-hand bottom quadrant of Figure 3, we can see how the Prefix do- combined with the Imperfective Aspect correlates with Activity and State, while do- combined with the Perfective Aspect is in close correlation with State and in distinct correlation with Accomplishment, as exemplified below (9a–d): (9) a. Albo nie chciała wtedy powiedzieć wprost, że się domyśla, że w tajemnicy przed nią popalał heroinę. ‘Or she didn’t want to say directly that she was coming to realize that he was clandestinely smoking heroin.’ (do+Imperf+ObjClause+ObjActivity) b. W pierwszej części filmu nikt się nie domyśla, że ten szczuplutki chłopiec o wąskiej twarzy i dużych ustach to kobieta. ‘In the first part of the film nobody realizes that this slim boy with a narrow face and lavish lips is a woman.’ (do+Imperf+ObjClause+ObjState) c. Twój rozmówca natychmiast domyśli się, że przeżywałeś jakiś dramat. ‘Your interlocutor will realize at once that you were experiencing some drama.’ (do+Perf+ObjClause+ObjState) d. W żaden sposób nie mogliśmy się domyślić , co ich tu sprowadziło. ‘We could in no way work out what brought them here.’ (do+Perf+ObjClause+ObjAccomplishment) The sentences above demonstrate that the verb myśleć, when coupled with the attainment prefix do- in either aspect, tends to co-occur with (1) objects designating states that affect other people in the surrounding world (examples (9b and c)), to which the conceptualiser is denied full access; (2) bounded or unbounded activities that happen beyond the speaker’s immediate field of attention (example (9a)); and (3) accomplishments whose genesis and course of development do not fall within the speaker’s known reality either (example (9d)). Therefore, it can be seen that the processual meaning of


Obj.Human Obj.Hypothetical

0.50 ROZ–Imperf Obj.Concrete

0.25

ZA–Perf WY–Perf ObjNP

0.00

Obj.Achievement ObjCLAUSE Obj.Activity

Obj.Abstract PRZE–Perf

–0.25

PO–Perf

DO–Imperf Obj.State DO–Perf

WY–Imperf

Obj.Accomplishment

–0.50

–1

–0.5

0

0.5

Figure 3.â•‡ Binary correspondence analysis: Prefix combined with Aspect, Object Form and Object Semantics

the verb and the semantic features of the said objects of thought, characterized by a certain degree of obscurity, are compatible. Considering the right-hand upper quadrant of the plot, we can see the correlation between the prefix po- and Hypothetical, Achievement and Activity clauses, which are illustrated in examples (10a–c): (10) a. Pomyślałem, że i ja mógłbym latać. ‘I thought that I could fly too.’ (po+Perf+ObjClause+ObjHypo) b. Pomyślałem, że niczego nie spostrzegł. ‘I thought that he hadn’t noticed anything.’ (po+Perf+ObjClause+ObjAch) c. Można było pomyśleć , że się o coś modli. ‘One could think that he was praying for something.’ (po+Perf+ObjClause+ObjActivity) This delimitative prefix is construed against the background of a variety of possible paths, one of which is selected and leads to the final point, becoming the focus of attention, to use Langacker’s parlance. The final point here is represented by clauses designating hypothetical events, achievements or activities. This shows that pomyśleć is more directly correlated with clauses, imposing upon them its delimitative nature. It is also noteworthy that activities and achievements are more directly observable and more easily enclosed in or confined to a single event of thought than accomplishments or states.


Hypothetical clausal objects are attracted by the prefixes roz- and za-, as exemplified below (11a–b): (11) a. “Widocznie każdy musi wszystko przeżyć od nowa”, rozmyślał na głos Olejniczak. ‘“Apparently everyone has to re-live everything”, Olejniczak was thinking aloud.’ (roz+Imperf+ObjClause+ObjHypothetical) b. “Może warto spróbować…”, zamyślił się. ‘“Maybe it’s worth trying”, he thought.’ (za+Perf+ObjClause+ObjHypothetical) It is noteworthy that in Section 2 both these prefixes were hypothesized to be conceptually similar. They both profile a mental process, whereby the subject is absorbed in thought about a given object. Due to its dispersed character, the procedural intensifying prefix roz- designates an expanse of mental space affected by the thinking process, without actually impacting whatever constitutes the object of pondering. Here, the mind seems to be construed in a more processual manner. The intensive-resultative za-, on the other hand, construes a situation in which either the subject seems to be behind the curtain of thought, or his/her mind, viewed as a container, appears to be completely covered by or filled in with the mental state (cf. Tabakowska 2003a).6 Both these mental processes engross the subject to such an extent that any other activity is impossible. Hypothetical objects which correlate closely with these prefixes concern the unknown, the possible, the unrealized, which naturally intrigues humans and invites extensive (roz-) as well as deep (za-) consideration. Let us now consider the difference in the use of nominal and clausal objects. Both correspondence analyses reveal that this distinction is basic in the data. These two object types divide the prefixes into two discrete groups represented in the plots by two separate clusters on the right and left halves. Correspondence analysis offers no statistical confirmation as to the significance or accuracy of what it identifies. In order to obtain such information, we have to turn to logistic regression (see Speelman, this volume). The model below produces accurate and predictive results, which can be assessed on the basis of the C score (0.846) and the estimated R2 (0.513). Although not expressed as a probability, the C score of 0.85 can be understood as approximating a 85% success rate in predicting the behaviour of the data in taking a nominal or clausal object (see Speelman, this volume, for a detailed discussion on how to interpret the model statistics). The estimated R2 score at 0.513 is a very strong indication of a stable and predictive model, any figure above 0.3 being accepted as a strong result. The model was checked for overfitting and multicollinearity, neither of which posed a problem. 6. The prefix za- is also present in such words as zasłonięty ‘covered with’ and zapełniony ‘filled with’.


Variance inflation factor values were beneath 2.5, which is well below accepted levels of tolerance (Dodge 2008:â•›96). Deviance Residuals:Â€ MinÂ€Â€Â€Â€Â€Â€1QÂ€Â€ MedianÂ€Â€Â€3QÂ€Â€Â€Â€Â€ MaxÂ€Â€ -2.2793Â€ -0.6885Â€ -0.5204Â€Â€0.5139Â€Â€2.0329Â€Â€ Coefficients: Â€Â€Â€Â€Â€Â€Â€Â€Â€Â€Â€ Estimate Std. Error z value (Intercept)Â€ 0.22609Â€Â€Â€ 0.37899Â€Â€ 0.597Â€ Aspect PerftiveÂ€ -0.56256Â€Â€Â€ 0.17132Â€ -3.284Â€ Prefix do-Â€Â€ -1.59438Â€Â€Â€ 0.37208Â€ -4.285 Prefix po-Â€Â€ -0.98245Â€Â€Â€ 0.34924Â€ -2.813Â€ Prefix prze-Â€ 1.86671Â€Â€Â€ 0.39706Â€Â€ 4.701 Prefix roz-Â€ -0.04704Â€Â€Â€ 0.42383Â€ -0.111Â€ Prefix wy-Â€Â€Â€ 2.29423Â€Â€Â€ 0.35414Â€Â€ 6.478 --‘***’ 0.001, ‘**’ 0.01, ‘*’ 0.05, ‘.’ 0.1

Pr(>|z|)Â€ 0.55080Â€Â€Â€Â€ 0.00102 **Â€ 1.83e-05 *** 0.00491 **Â€ 2.59e-06 *** 0.91162 9.27e-11 ***

Null deviance: 3795.0Â€ on 2738Â€ degrees of freedom Residual deviance: 2463.5Â€ on 2732Â€ degrees of freedom AIC: 2477.5 Number of Fisher Scoring iterations: 4 Obs 2739

Model L.R. 1331.48

d.f. 6

P 0

C 0.846

Dxy 0.692

R2 0.513

Brier 0.14

A closer look at the table of coefficients tells us how the model predicts the outcome as either clausal or nominal. In the table, negative coefficients predict the occurrence of clausal objects and positive coefficients predict the occurrence of a nominal object. The strongest predictor for the clausal object is the prefix do-. The second feature predicting clausal objects is po-. However, the coefficient score is close to zero, indicating that its role is minor, despite its statistical significance. We can now return to the the results of correspondence analysis in Figure 3 to better understand these findings. Although both po- and do- are present, we now know that in fact do- is the more important association in the cluster. Nominal complementation, in turn, is strongly predicted by the prefixes wy- and prze-, which again bears out the correlations visualized in Figure 3. The fact that the prefix roz- is the weakest predictor, neither significant nor important, also corresponds to the findings of the correspondence analysis, where it is located on the line separating the halves containing nominal and clausal objects, respectively.


Overall, the above findings indicate that the attainment prefix do-, in particular, but also the delimiting po- are more importantly associated with processual objects, as designated by clausal complements. The resultative, goal-oriented prefix wy-, on the other hand, and the prefix prze- correlate importantly with the abstract category of things, profiled by nominal objects. This corroborates the results of the exploratory correspondence analyses presented above. It is also interesting in itself that no predictive model could be generated with the inclusion of object semantics. This may indicate that the semantics of the object of thought for these mental predicates can only be distinguished at a higher and more schematic level of analysis. Rather than drawing a fine-grained distinction between particular types of objects relative to particular verbs, it appears that, at least for some prefixes, a more coarse-grained preference can be identified – either for processual or reifying construal of the object. The semantically coarse-grained results of the logistic regression analysis are compatible with the schematic semantic characterization of the prefixes in question here. Taking the strongest predictors for either response variable feature, we can see that the attainment, process-oriented do- is attracted to clausal objects, whereas the resultative wy-, bringing about more or less concrete but objectifiable results, is more likely to be associated with nominal phrases. Having analysed and interpreted the correspondence analyses, which brought to light a number of revealing interdependencies between various semantic and formal features, and displayed interesting continua thereof, some of which have been confirmed by the Logistic Regression Analysis, we may turn to Hierarchical Cluster Analysis. The insights from the correspondence analyses and the logistic regression analysis will allow us to interpret the results of the cluster analysis. The cluster analysis, in Figure 4, shows that roz- and za-, prze- and wy-Imperf as well as do- and po- cluster with each other, whereas wy-Perf stands out from the rest. This clustering only partly correlates with the hypothesis from Section 2, where wy-, prze-, and do- clustered at one end of the continuum, while po-, do-, roz- and za- clustered at the other. However, the correspondence analysis conducted above allows us to account for this discrepancy. As was shown in examples (8a–c), wy- and prze- correlate with NP objects, but docorrelates with Clausal objects. This is why prze- combined with Perfective Aspect and wy- combined with Imperfective Aspect cluster together – they both take Abstract nominal objects. Wy- combined with Perfective Aspect is most importantly correlated with concrete nominal objects. This is why it occupies a separate branch in the cluster. The introspective analysis, which relied on the native speaker’s competence of what is acceptable in the language, could not account for the difference in the frequency of use of do- with clausal vs. nominal objects. Similarly, the analysis based on intuition revealed that po- and roz- can both take nominal objects, yet it failed to ascertain that this potential is highly significant only for roz-. This is probably the reason why the HCA shows a clustering of po- and do-, and not of po- and roz-. This


wy-Perf

57 52 6

92 78 5 po-Perf wy-Imperf

prze-Perf

100 88 2 za-Perf

roz-Imperf

0

100 81 1

100 87 5 do-Imperf

200

96 81 4

do-Perf

400

600 800

1000

au bp edge #

Figure 4.â•‡ Hierarchical cluster analysis (Distance: Euclidean, cluster method: ward)

correlation is also clearly visible in the correspondence analysis and is a result of the propensity of these prefixes to correlate with clausal objects such as Achievement, Activity and State. They are conceptually similar, as they both construe thinking as a process, which is metaphorically understood as a journey on/along the road/path (po-) leading to a goal (do-). The roz-, za- cluster is caused by the tendency of both these prefixes to co-occur with Hypothetical clausal objects. Its presence can be further justified by the schematic meaning of the prefixes, which indicate that the subject is construed as being deep in thought to such an extent that any other activity is hindered.

6. Conclusion The present paper describes the semasiological structure of the onomasiological field of prefixed think verbs in Polish relative to their object form, object semantics and aspect. These factors have been selected on the basis of previous pilot studies that revealed their significance for the research question. The aim of providing an accurate semantic description has been attained in a twofold manner. First, the prefixes were subjected to a detailed introspective investigation in light of a number of relevant theoretical approaches. In this phase, we projected the inter- and intra-category conceptual relations, constituting the background against which crucial decisions could be made regarding the annotation schema and the hypotheses to be tested at the quantitative stage of the study.


The second step involved submitting the annotated data to multivariate statistical modelling in the form of two exploratory methods, correspondence analysis and hierarchical cluster analysis, and one confirmatory method, logistic regression analysis. The application of two mutually complementary phases of analysis ensured greater reliability of results. Employing pertinent exploratory methods enabled us to identify actual patterns of use in the object of our study, as well as patterns surfacing in multiple dimensions that no introspection could possibly bring to light. Importantly, the results thus obtained are replicable and falsifiable, which adds to the accuracy of the model produced. Further stages of research into the semantics of the Polish mental verbs should focus on the contribution of Subject grammar and Subject semantics as well as modality and negation to the construal of the mental process. So far, Fabiszak, Hebda and Konat (2012) have run a pilot study on the contribution of Subject form and Object form and semantics to the construal of the scene in the Polish verb wierzyć ‘to believe’, while Fabiszak and Hebda (2011) looked at the role of Subject, mood and negation in the meaning construction of wierzyć ‘to believe’. Kokorniak and Krawczak (2010), in turn, focus on nominal complementation of the verbs myśleć ‘think’ and sądzić ‘suppose’. The construal of prefixed think verbs in Polish relative to the grammatical subject person and adverbial modification is the object of study in Krawczak and Kokorniak (2012). Clausal that and zero complementation of the English verb think relative to an array of syntactic and semantic properties is analysed in Krawczak and Kokorniak (2010), while Krawczak and Fabiszak (2011) concentrate on the construal of the nominal as well as the clausal complementation of the verbs myśleć ‘think’ and wierzyć ‘believe’.

References Amberber, M. (2003). The grammatical encoding of ‘thinking’ in Amharic. Cognitive Linguistics, 14, 195–219. DOI: 10.1515/cogl.2003.008 Comrie, B. (1976). Aspect: An introduction to the study of verbal aspect and related problems. Cambridge: Cambridge University Press. D’Andrade, R. (1987). A folk model of the mind. In D. Holland, & N. Quinn (Eds.), Cultural models in language and thought (pp. 112–148). Cambridge: Cambridge University Press. DOI: 10.1017/CBO9780511607660.006 D’Andrade, R. (1995). The development of cognitive anthropology. Cambridge: Cambridge University Press. DOI: 10.1017/CBO9781139166645 Dąbrowska, E. (1997). Cognitive Semantics of the Polish dative. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110814781


Danielewiczowa, M. (2000). Główne problemy opisu i podziału czasownikowych predykatów mentalnych. [Main problems in the description and classification of verbal mental predicates.] In R. Grzegorczykowa, & K. Waszakowa (Eds.), Studia z semantyki porównawczej. [Studies in comparative semantics] Vol. 1 (pp. 227–247). Warszawa: Wydawnictwo UW. Danielewiczowa, M. (2002). Wiedza i niewiedza: Studium polskich czasowników epistemicznych. [Knowing and not knowing: A study of Polish epistemic predicates.] Warszawa: Katedra Lingwistyki Formalnej UW. Dewell, R. B. (1994). Over again: Image-schema transformations in semantic analysis. Cognitive Linguistics, 5, 351–380. DOI: 10.1515/cogl.1994.5.4.351 Dickey, S. M. (Unpublished manuscript). Subjectification and the Russian perfective. Dickey, S. M., & Hutcheson, J. (2003). Delimitative verbs in Russian, Czech and Slavic. In R. A. Maguire, & A. Timberlake (Eds.), American contributions to the Thirteenth International Congress of Slavists (pp. 23–36). Columbus: Ohio Slavca. Retrieved from [Accessed 9th November 2009]. Dickey, S. M. (2000). Parameters of Slavic aspect: A cognitive approach. Stanford: CSLI. Dickey, S. M. (2009). Subjectification and the East-West aspect division. Paper presented at the 9th Slavic Cognitive Linguistics Conference, 16th October 2009, Prague. Dirven, R., Goossens, L., Putseys, Y., & Vorlat, E. (1982). The scene of linguistic action and its perspectivization by SPEAK, TALK, SAY, and TELL. Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/pb.iii.6 Divjak, D., & Gries, St. Th. (2006). Ways of trying in Russian: Clustering behavioral profiles. Corpus Linguistics and Linguistic Theory, 2, 3–60. DOI: 10.1515/CLLT.2006.002 Divjak, D. (2006). Ways of intending: A corpus-based Cognitive Linguistic approach to near-synonyms in Russian. In St. Th. Gries, & A. Stefanowitsch (Eds.), Corpora in Cognitive Linguistics: Corpus-based approaches to syntax and lexis (pp. 19–56). Berlin & New York: Mouton de Gruyter. Divjak, D. (2010). Structuring the lexicon: A clustered model for near-synonymy. Berlin & New York: Mouton de Gruyter. Dodge, Y. (2008). The concise encyclopedia of statistics. Berlin: Springer. Fabiszak, M., & Hebda, A. (2011). Social and individual cognition, modality and negation in the use of the Polish verb wierzyć ‘to believe’. Paper presented at SLE 2011, Logrono, Spain. Fabiszak, M., Hebda, A., & Konat, B. (2012). Dichotomy between private and public experience: The case of Polish wierzyć ‘believe’. In Ch. Hart (Ed.), Online proceedings of UK-CLA meetings 1 (pp. 164–176). Hertfordshire: The UK Cognitive Linguistics Association. Retrieved from . Fortescue, M. (2001). Thoughts about thought. Cognitive Linguistics, 12, 15–45. DOI: 10.1515/cogl.12.1.15 Geeraerts, D., Grondelaers, S., & Bakema, P. (1994). The structure of lexical variation: Meaning, naming, and context. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110873061 Glynn, D. (2009). Polysemy, syntax, and variation: A usage-based method for Cognitive Semantics. In V. Evans, & S. Pourcel (Eds.), New directions in Cognitive Linguistics (pp. 77– 106). Amsterdam & Philadelphia: John Benjamins. Glynn, D. (2010a). Synonymy, lexical fields, and grammatical constructions: A study in usage-based Cognitive Semantics. In H.-J. Schmid, & S. Handl (Eds.), Cognitive foundations of linguistic usage-patterns: Empirical studies (pp. 89–118). Berlin & New York: Mouton de Gruyter.


Glynn, D. (2010b). Testing the hypothesis: Objectivity and verification in usage-based Cognitive Semantics. In D. Glynn, & K. Fischer (Eds.), Quantitative Cognitive Semantics: Corpus-driven approaches (pp. 239–270). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110226423 Glynn, D. (2010c). Corpus-driven Cognitive Semantics: An overview of the field. In D. Glynn, & K. Fischer (Eds.), Quantitative Cognitive Semantics: Corpus-driven approaches (pp. 1– 42). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110226423.1 Glynn, D. (Forthcoming). Mapping meaning: Corpus methods for Cognitive Semantics. Cambridge: Cambridge University Press. Glynn, D., & Fischer, K. (Eds.). (2010). Quantitative methods in Cognitive Semantics: Corpus-driven approaches. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110226423 Goddard, C. (2003). ‘Thinking’ across languages and cultures: Six dimensions of variation. Cognitive Linguistics, 14, 109–140. DOI: 10.1515/cogl.2003.005 Greenacre, M. (1993). Correspondence analysis in practice. London: Academic Press. Gries, St. Th. (2003). Multifactorial analysis in Corpus Linguistics: A study of particle placement. London: Continuum Press. Gries, St. Th. (2006). Corpus-based methods and Cognitive Semantics: The many senses of to run. In St. Th. Gries, & A. Stefanowitsch (Eds.), Corpora in Cognitive Linguistics: Corpus-based approaches to syntax and lexis (pp. 57–99). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110197709 Gries, St. Th., & Stefanowitsch, A. (Eds.). (2006). Corpora in Cognitive Linguistics: Corpus-based approaches to syntax and lexis. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110197709 Grochowska, A. (1979). Próba opisu reguł łączliwości przedrostka prze- z tematami czasowniÂ� kowymi. [An attempt at the description of the combinatory rules of the prefix prze- with verb roots.] Polonica, 5, 59–74. Grondelaers, S., Geeraerts, D., & Speelman, D. (2007). A case for a cognitive Corpus Linguistics. In M. Gonzalez-Marquez, I. Mittleberg, S. Coulson, & M. Spivey (Eds.), Methods in Cognitive Linguistics (pp. 149–169). Amsterdam & Philadelphia: John Benjamins. Jackendoff, R. (1983). Semantics and cognition. Cambridge, MA: MIT Press. Janda, L. (1993). A geography of case semantics: The Czech dative and the Russian instrumental. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110867930 Junker, M.-O. (2003). A native American view of the ‘Mind’ as seen in the lexicon of cognition in East Cree. Cognitive Linguistics, 14, 167–194. DOI: 10.1515/cogl.2003.007 Kokorniak, I., & Krawczak, K. (2010). Thinking about thinking: Constructions of Polish mental verbs in discourse. Paper presented at Syntax in Cognitive Grammar, Częstochowa. Krawczak, K., & Kokorniak, I. (2010). Verbs of cognition, their construal, and complementation in interactive events. Paper presented at the 4th International Conference of the German Cognitive Linguistics Association, Bremen. Krawczak, K., & Fabiszak, M. (2011). Cognition verbs in Polish, their construal and complement semantics. Paper presented at the 3rd Conference of the Scandinavian Association for Language and Cognition, Copenhagen. Krawczak, K., & Kokorniak, I. (2012). A corpus-driven quantitative approach to the construal of Polish think. Poznań Studies in Contemporary Linguistics, 48, 439–472. DOI: 10.1515/psicl-2012-0021


Kustova, G. (2000). Niektóre problemy opisu predykatów mentalnych. [Some problems in the description of mental predicates.] In R. Grzegorczykowa, & K. Waszakowa (Eds.), Studia z semantyki porównawczej [Studies in comparative semantics] Vol. 1 (pp. 249–263). Warszawa: Wydawnictwo UW. Lakoff, G., & Johnson, M. (1980). Metaphors we live by. Chicago: University of Chicago Press. Langacker, R. (1991). Foundations of Cognitive Grammar. Vol. 2. Descriptive application. Stanford: Stanford University Press. Langacker, R. (1999). Losing control: Grammaticalization, subjectification, and transparency. In A. Blank, & P. Koch (Eds.), Historical semantics and cognition (pp. 147–175). Berlin & New York: Mouton de Gruyter. Lindner, S. (1983). A lexico-semantic analysis of English verb particle constructions. Trier: LAUT. Palmer, G. (2003). Talking about thinking in Tagalog. Cognitive Linguistics, 14, 251–280. Pasich-Piasecka, A. (1993). Polysemy of the Polish verbal prefix prze-. In E. Górska (Ed.), Images from the cognitive scene (pp. 11–26). Kraków: Universitas. Pawłowska, R. (1981). Znaczenie i użycie czasownika ‘myśleć’. [The meaning and use of the verb ‘think’.] Polonica, 7, 149–160. Piernikarski, C. (1975). Czasowniki z prefiksem po- w języku polskim i czeskim: Na tle rodzajów akcji w językach słowiańskich. [Verbs with the po- prefix in Polish and Czech: In the background of Aktionsarten in Slavic languages.] Warszawa: PWN. Przybylska, R. (2002). Stru ktura schematyczno-wyobrażeniowa prefiksu czasownikowego roz-. [Image-schematic structure of the verbal prefix ‘roz-’.] Polonica, 21, 269–286. Przybylska, R. (2006). Schematy wyobrażeniowe a semantyka polskich prefiksów czasownikowych do-, od-, prze-, roz-, u-. [Image schemata and the semantics of Polish verb prefixes do-, od-, prze-, roz-, u-.] Kraków: Universitas. Radden, G., & Dirven, R. (2007). Cognitive English grammar. Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/clip.2 Rudzka-Ostyn, B. (2000). Z rozważań nad kategorią przypadka. [Ruminating on the category of case.] Kraków: Universitas. Schlesinger, I. M. (1998). Cognitive space and linguistic case: Semantic and syntactic categories in English. Cambridge: Cambridge University Press. Śmiech, W. (1986). Derywacja prefiksalna czasowników polskich. [Prefix derivation of Polish verbs.] Wrocław: Ossolineum. Speelman, D., Tummers, J., & Geeraerts, D. (2009). Lexical patterning in a construction grammar: The effect of lexical co-occurrence patterns on the inflectional variation in Dutch attributive adjectives. Constructions and Frames, 1, 87–118. DOI: 10.1075/cf.1.1.05spe Szwedek, A. (2007). An alternative theory of metaphorisation. In M. Fabiszak (Ed.), Language and meaning: Cognitive and functional perspectives (pp. 312–327). Frankfurt/Main: Peter Lang. Tabakowska, E. (2003a). Space and time in Polish: The preposition za and the verbal prefix za-. In H. Cuyckens, T. Berg, R. Dirven, & K.-U. Panther (Eds.), Motivation in language: Studies in honor of Günter Radden (pp. 153–177). Amsterdam & Philadelphia: John Benjamins. Tabakowska, E. (2003b). The notorious Polish reflexive pronouns: A plea for Middle Voice. Glossos 4. Retrieved from [Accessed 9th November 2008]. Vendler, Z. (1967). Linguistics in philosophy. Ithaca, NY: Cornell University Press.


Wierzbicka, A. (1992). Semantics, culture, and cognition: Universal human concepts in culture-specific configurations. Oxford: Oxford University Press. Wierzbicka, A. (1996). Semantics: Primes and universals. Oxford: Oxford University Press. Wierzbicka, A. (1997). Understanding cultures through their key words. Oxford: Oxford University Press. Wierzbicka, A. (1999). Emotions across languages and cultures. Cambridge: Cambridge University Press. DOI: 10.1017/CBO9780511521256

A multifactorial corpus analysis of grammatical synonymy The Estonian adessive and adposition peal ‘on’ Jane Klavan

University of Tartu

In the present study, both monofactorial and multifactorial corpus methods are applied to the alternative use of the analytic adpositional construction and the synthetic case construction in present-day written Estonian. 600 examples from the fiction sub-corpora of written Estonian were annotated for 16 different (seven morphosyntactic and nine semantic) factors. In order to determine which factors are most influential in determining the choice between the two constructions, a logistic regression model was built to fit the data. The analysis confirmed the statistical influence of the following factors: lexical complexity, type of verb, word order, word class, and mobility. The results reported in this study align with previous research, which has shown that case affixes are used to express more abstract relations and adpositions more concrete ones. Keywords: adpositional construction, case construction, Finno-Ugric languages, logistic regression

1. Introduction1 To express a spatial scene where a vase is located on a table, one can either use the adessive case (example (1a)) or the adpositional construction with the postposition peal ‘on’ (example (1b)) in Estonian: (1) a. vaas on laual vase.sg.nom be-prs.sg3 table.sg.ad ‘the vase is on the table’. 1. I am indebted to Dylan Glynn for his valuable comments and his extensive help with the logistic regression analysis. Methodologically, the present chapter relies heavily on Glynn (2007, 2010).

254 Jane Klavan

b. vaas on laua peal vase.sg.nom be-prs.sg3 table.sg.gen on.ad ‘the vase is on the table’. Among other things, I will be taking a closer look at one of the central claims made in the literature about the difference between the ways the two strategies, i.e. the synthetic vs. analytic way of expressing meaning, are used. It is claimed that case affixes are used to express more abstract relations and adpositions to express more concrete ones (e.g. Bartens 1978; Comrie 1986; Luraghi 1991:â•›66–67; Ojutkangas 2008; Hagège 2010:â•›37–38; Lestrade 2010). Adpositions are said to be semantically more specific than cases and they are used to express the less predictable spatial meanings; cases, on the other hand, are more abstract and used to express more frequent spatial meanings (Lestrade 2010). The chapter proceeds from the theoretical premises of both Construction Grammar (Goldberg 1995, 2006) and Cognitive Grammar (Langacker 1987, 2008), where one of the basic general assumptions is that of no-synonymy – when two constructions differ syntactically, then they also differ either semantically or pragmatically (Goldberg 1995:â•›67). At present, there is no detailed corpus study on this topic and the few previous studies that are available proceed from an approach based largely on introspection (Rannat 1991; Vainik 1995). Accordingly, the main aim of the chapter is to carry out a completely corpus-based analysis of the adessive case construction and the adpositional construction with the postposition peal ‘on’. The central focus will be on finding out how significant different morphosyntactic and semantic factors are in determining the use of either the adessive constructions or the adpositional constructions, and which factors are the most important ones. Both monofactorial and multifactorial statistical techniques are used to analyse the results. The monofactorial analysis is seen as a necessary and beneficial first step in the quantitative analysis of the results. The purpose is to move from more simple, univariate exploratory analysis to more complex but more powerful multivariate analysis. Monofactorial techniques make it possible to identify which variables are statistically significant, but by employing such a multivariate statistical technique as regression analysis, it is possible to determine the contribution of the different semantic and morphosyntactic variables to the alternation and calculate the relative strength of each individual variable. A number of researchers have successfully argued for the benefits of using multifactorial analyses (e.g. Gries 2003; Wulff 2003; Glynn 2007, 2010; Szmrecsanyi 2010). Gries (2003:â•›155), among others, emphasises that “variation phenomena are too multifaceted to be treated adequately by means of minimal-pair tests and the researchers own judgements”. The chapter was inspired by numerous other studies on different variation phenomena, including those specifically about the synthetic-analytic distinction (for dative alternation in English, see Bresnan et al. 2007, Bresnan and Ford 2010; for English comparative alternation, Mondorf 2003; for the English genitive

A multifactorial corpus analysis of grammatical synonymy 255

alternation, Rosenbach 2003, Szmrecsanyi 2010; for particle placement alternation in English transitive phrasal verbs, Cappelle 2009). The chapter consists of six sections. In the following section, an overview of the functions of the Estonian adessive case and the adposition peal is given. Section 3 describes the data sample; Section 4 gives an overview of the variables coded and presents the monofactorial results. The results of a multifactorial statistical (logistic regression) analysis are presented and discussed in Section 5; the chapter ends with a conclusion (Section 6).

2. The Estonian adessive case and the adposition peal ‘on’ 2.1

The Estonian adessive case

Estonian nouns and adjectives decline in fourteen cases; six of these cases are referred to as locative cases and can be divided into interior locative cases (illative, inessive, elative) and external locative cases (allative, adessive, ablative) (see Table 1). The Estonian adessive case belongs to the set of external locative cases and expresses, first and foremost, spatial or temporal relations. It normally takes the role of an adverbial or attribute in the clause (Erelt et al. 1995:â•›58). Estonian external locative cases express spatial relations of an open surface and they form a three-part series – allative, adessive, ablative – expressing direction, location and source, respectively (Erelt et al. 2007:â•›240; see Table 1). The present chapter focuses only on the adessive, although the other forms of external locative cases are also said to be synonymous with the respective forms of the adposition peal ‘on’ (direction: allative -le ~ peale ‘onto’; source: ablative -lt ~ pealt ‘off ’). Although the primary meaning of the locative cases was the expression of spatial relations, in modern Estonian they fulfil a number of abstract functions. For instance, it is more frequent for the Estonian adessive to mark the possessor or agent (examples (2c) and (2d)) than, for example, location (example (2a)); the following functions of the Estonian adessive case are relevant for the present analysis (Erelt et al. 2007:â•›250): Table 1.â•‡ Estonian locative cases as exemplified by the noun laud ‘table’ Interior Exterior

lative (direction)

locative (location)

separative (source)

illative laua-sse ‘into table’ allative laua-le ‘onto table’

inessive laua-s ‘in table’ adessive laua-l ‘on table’

elative laua-st ‘out of table’ ablative laua-lt ‘off table’

256 Jane Klavan

(2) a. Location Vaas on laual. vase.sg.nom be-prs.3sg table.sg.ad ‘The vase is on the table’. b. Instrument Mari mängib klaveril mõnd lugu. Mari.nom play-prs.3sg piano.sg.ad some tune.sg.part ‘Mari plays some tunes on the piano’. c. Possessor Maril on kaks last. Mari.ad be-prs.3pl two child.sg.part ‘Mari has two kids’. d. Agent with finite verb forms See asi ununes mul kiiresti. this thing.sg.nom forget-prs.3sg me.sg.ad quickly ‘I quickly forgot about that thing’.

2.2

The Estonian adposition peal ‘on’

In addition to the locative cases, location and change of location in Estonian can be expressed with adpositions, adverbs, and nouns declined in interior and exterior locative cases (Erelt et al. 1993:â•›71). In Estonian reference grammars, adpositions are treated as uninflected words that are used together with nouns and express similar meanings as case endings. In comparison with adpositions, the meaning of cases is said to be much more abstract and the usage range much broader (Palmeos 1985; Erelt et al. 1995:â•›33–34; Erelt et al. 2007:â•›191). This is in line with the general claims made concerning the differences between adpositions and case affixes (Comrie 1986; Hagège 2010; Lestrade 2010). Nevertheless, as is stressed in the following sub-section, there are still instances where both the adessive case and the adposition peal ‘on’ are seen as semantic alternatives in Estonian. A distinctive morphological characteristic of Estonian adpositions is that like locative cases they constitute three-member sets that are semantically and grammatically divided into the lative, locative, and separative forms (see Table 2). The adposition Table 2.â•‡ The three-member sets of Estonian postpositions sees ‘in’ and peal ‘on’ Interior Exterior

lative (direction)

locative (location)

separative (source)

illative si-sse ‘into’ allative pea-le ‘onto’

inessive ‘in’ see-s adessive pea-l ‘on’

elative ‘out of ’ see-st ablative pea-lt ‘off ’


peal ‘on’ takes external locative case endings: peale – peal – pealt. In the present chapter, only the locative form peal ‘on’ is discussed. At the clause level, the Estonian adpositional phrase has two basic functions – that of an adverbial and adverbial modifier (Erelt et al. 1993:â•›137). The adposition peal ‘on’ is polysemous; relevant for the present study are the following senses: (3) a. Location Leib on laua peal. bread.sg.nom be-prs.3sg table.sg.gen on.ad ‘Bread is on the table’. b. Place Turu peal oli suur sagimine. market.sg.gen on.ad be-pst.3sg big commotion.sg.nom ‘There was a big commotion on the market’. c. Instrument Mängi klaveri peal ette! play-imp.2sg piano.sg.gen on.ad ahead ‘Play something on the piano!’

2.3

The parallel use of the Estonian adessive and the adposition peal ‘on’

When comparing the meanings of the adposition peal ‘on’ (examples in (3)) to those of the adessive (examples in (2)), it can be seen that these two forms are used as alternatives to each other, especially in the functions of expressing location, place and instrument. According to Palmeos (1985:â•›15), the analytic construction – genitive together with the adposition peal ‘on’ – expresses the same meaning as the synthetic adessive. At the same time, it has been claimed in Estonian reference grammars that the meaning of adpositions is more concrete and specific than that of the cases (Erelt et al. 2007:â•›191). This has also been mentioned by Palmeos (1985:â•›18), who notes that the analytic construction conveys the meaning more clearly than the synthetic one. This clarity of expression is partly due to the grammatical homonymy inherent in the Estonian language – in some cases, when using the synthetic construction, it is not clear whether we are expressing location or possession and sometimes, the use of the adessive to express location is not possible because the possessive reading is too strong. Nevertheless, there are still numerous instances where both the adessive case and the adposition peal ‘on’ can express more or less the same meaning. A small-scale corpus analysis showed that in the 5-million-word fiction sub-corpus of the Balanced Corpus of Estonian (2008), there are 314 different Landmarks used with both the adessive case and the adposition peal ‘on’. Furthermore, similar results of the parallel use of these two constructions were obtained with an open production task that

258 Jane Klavan

studied the expression of different spatial relations in Estonian (Salm 2010). Klavan et al. (2011) have reported the results of a forced choice task and a production task, the aim of which was to determine which semantic factors play a role in the use of the adessive and adposition peal ‘on’. The results of these studies indicate that the adessive is used when there is an abstract relation between Trajector and Landmark and the Landmark is a place; the adposition peal ‘on’ is used when there is an unconventional spatial relation between Trajector and Landmark and when the Landmark is a thing. However, the two tasks also yielded results where there was no significant difference between the two locative constructions. The present study hopes to provide converging evidence and to shed new light upon this issue by building upon the previous work, but using a different methodology (mono- and multifactorial corpus analysis) and looking at other factors besides the semantic ones.

3. The data sample The data analysed in this chapter come from the corpus of present-day written Estonian. A data sample of 300 instances of the Estonian adessive case from the fiction sub-corpus of the Morphologically Disambiguated Corpus (2010; size 104,000 words) and 300 instances of the Estonian adposition peal ‘on’ from the fiction sub-corpus of the Balanced Corpus of Estonian (2008; size 5 million words) was collected. Based on the findings from the previous variation and spatial expression studies cited above, the data were manually coded for multiple variables or ‘predictors’, which are outlined in the following section. It is worth pointing out that in order to get the 300 suitable instances of the adessive case, it was necessary to manually work through 1,700 instances of this construction; the reason being that the adessive fulfils a number of other functions in Estonian besides expressing spatial relations.

4. Corpus-linguistic operationalizations and monofactorial results The original coding schema included more than 30 variables; in the present analysis only the sixteen most important ones are discussed. These sixteen variables can be divided into two groups and include the following: seven morphosyntactic factors (length of the Landmark phrase, lexical complexity of the Landmark, word order, verb lemma, syntactic function of the Landmark phrase, word class of Landmark and Trajector) and nine semantic factors (type of relation between Landmark and Trajector, type of Landmark, animacy, number and mobility of Landmark and Trajector, relative size between Landmark and Trajector). The following sub-sections give an overview of the operationalizations of these variable groups and their different levels.


The significance of each variable was also tested individually and these results are presented directly after the presentation of the respective variable. The monofactorial analysis of the results relies on Gries (this volume) and predominantly makes use of the Chi-squared test to evaluate the raw frequency counts encountered in the corpus. Following Gries (ibid.), I also computed the effect sizes and inspected the Pearson residuals for each variable in order to determine how strong the results of the Chisquare tests are and what exactly determines the difference between the frequency counts. As Gries (ibid.) points out, the effect size theoretically ranges from 0 (‘no effect’) to 1 (‘perfect correlation’).

4.1

Morphosyntactic factors

4.1.1 Length of the Landmark phrase Previous studies on grammatical alternation have shown that the more explicit form is favoured in cognitively complex environments (Mondorf 2003:â•›294). This phenomenon has also been referred to as Rohdenburg’s complexity principle (ibid.). Given thus the option of either using a synthetic or analytic locative construction, longer words have been found to opt for the analytical variant and shorter words/phrases for the synthetic variant (Cooper and Ross 1975; Hawkins 1994; Wasow 1997; Arnold et al. 2000; Mondorf 2003; Wulff 2003). Mondorf (2003:â•›251–253) argues for a presumably universal tendency, the phenomenon that she terms analytic support: In cognitively more demanding environments which require an increased processing load, language users – when faced with the option between a synthetic and analytic variant – tend to compensate for the additional effort by resorting to the analytic form.

Although Mondorf (2003) focuses on the English comparative construction, the claim is that this compensatory strategy can be extended to other kinds of variation phenomena that draw on the synthetic-analytic distinction. Since length constraints have been found to affect other syntactic variation phenomena as well (e.g. Cooper and Ross 1975; Hawkins 1994; Wasow 1997; Arnold et al. 2000; Wulff 2003), it was decided to code the present data for the length of the locative phrase. Measures of syntactic complexity can be efficiently operationalized by counting the number of graphemic words (Bresnan and Ford 2010). For the present analysis, length of the Landmark phrase was measured both in words and syllables. In line with Bresnan and Ford (2010:â•›9, fn. 8) length was transformed by the logarithm in order to compress extreme values and bring the distribution more closely into the logistic regression model assumption of linearity. The results show that the mean length of the Landmark phrase with the adessive (1.95) differed highly significantly from the mean length of the Landmark phrase

260 Jane Klavan

Table 3.â•‡ Length of the Landmark phrase in words Number of words in the Lm phrase

Adessive

peal

Total

1 word 2 words 3 words 4 words 5 words 6 words 7 words Total

114 123 â•⁄41 â•⁄15 â•⁄â•⁄2 â•⁄â•⁄2 â•⁄â•⁄3 300

192 â•⁄88 â•⁄13 â•⁄â•⁄5 â•⁄â•⁄2 â•⁄â•⁄0 â•⁄â•⁄0 300

306 211 â•⁄54 â•⁄20 â•⁄â•⁄4 â•⁄â•⁄2 â•⁄â•⁄3 600

with the postposition peal ‘on’ (1.46), t(78) = 4.78, p < .001. As seen from Table 3, the adessive was predominantly used with Landmark phrases that were two or more words long and the postposition peal with Landmark phrases that were one word long (examples (4a) and (4b)). If we take that a longer Landmark phrase implies a cognitively more complex environment (cf. Wasow 1997; Cappelle 2009:â•›149), Rohdenburg’s complexity principle does not hold here, i.e. the more explicit form (the postposition peal) was not favoured in a cognitively more complex environment. (4) a. ILU1980\stkt0001: Pihlakad kasvasid sellelsamal põllupeenral. rowan.pl.nom grow-pst.3pl this.sg.ad flowerbed.sg.ad ‘The rowans were growing on this flower bed’. b. MJ_A: Kai peal oli meeletu trügimine… pier.sg.gen on.ad be-pst.3sg mad.sg.nom pushing.sg.nom ‘On the pier people were pushing madly…’ According to Mondorf (2003:â•›254), who discusses the English comparative construction, one of the effects of the so-called more-support that reflects the general analytic support is that: a separate lexeme as degree marker rather than an inflectional suffix can serve both as an unambiguous signal indicating increased processing load to the reader and as a less condensed and more explicit way of structuring a complex phrase.

Since the synthetic variant in -er allows recognition only after the adjective and its inflection have been processed, complex environments should call for early recognition and hence the analytic variant would be used in English for the comparative construction (Mondorf 2003:â•›255). However, in the case of the Estonian adessive and the adposition peal ‘on’, this signalling argument does not work because both the adessive case marker and the postposition peal ‘on’ follow the locative phrase (see examples (4a)


Table 4.â•‡ Length of the Landmark phrase in syllables Number of syllables in the Lm phrase

Adessive

peal

Total

1 syllable 2 syllables 3 syllables 4 syllables 5 syllables more than 5 syllables Total

â•⁄â•⁄7 â•⁄64 â•⁄61 â•⁄45 â•⁄32 â•⁄91 300

â•⁄33 125 â•⁄31 â•⁄59 â•⁄17 â•⁄35 300

â•⁄40 189 â•⁄92 104 â•⁄49 126 600

and (4b)). Thus, quite the opposite effect may play a role in Estonian. Precisely due to the fact that in the case of Estonian locative phrases, the postposition only comes at the end of the phrase, the locative case is a better signal to indicate that what we have here is a long locative phrase expressing a support relation – in Estonian all of the words in the adessive locative phrase are marked for the adessive case, as sellelsamal ‘this’ in example (4a) above. In addition, the results also show that the mean length of the Landmark phrase in syllables was significantly different with the adessive (4.87) and the postposition peal (3.15), t(435) = 7.18, p < .001. The proportion of peal ‘on’ uses with Landmark phrases one or two syllables long was considerably higher than the proportion of adessive uses with Landmark phrases of the same length; when the Landmark phrase was more than 10 syllables long, the proportion of adessive occurrences was much bigger than that of peal ‘on’ (Table 4).

4.1.2 Lexical complexity of Landmark It was decided to also code the Landmark for what has been termed here as lexical complexity, with the levels of ‘compound’ and ‘single lexeme’, e.g. writing desk vs. desk. Table 5 shows that the number of occurrences where the Landmark word was a compound was significantly higher with the adessive (88 instances) than with the adposition peal ‘on’ (17 instances), χ²(1, N = 600) = 56.57, p < .001, φ = 0.31. This gives further indication of the adposition peal being preferred with shorter, less complex Landmark phrases, while the adessive tends to be used with longer and more complex Landmark phrases. Table 5.â•‡ Lexical complexity of Landmark Landmark word

Adessive

peal

Total

compound single lexeme Total

â•⁄88 212 300

â•⁄17 283 300

105 495 600

262 Jane Klavan

Table 6.â•‡ Position of the locative phrase within the clause Position of locative phrase

Adessive

peal

Total

final initial middle Total

129 â•⁄83 â•⁄88 300

111 â•⁄83 106 300

240 166 194 600

4.1.3 Word order: Position of the locative phrase Several researchers have discussed the principle of end-weight in relation to grammatical variation (Wasow 1997; Cappelle 2009:â•›149–150). This phenomenon states that “long, complex phrases tend to come at the ends of clauses” (Wasow 1997:â•›81). However, it is not entirely clear what is meant by ‘weight’. For example, it can either be length or complexity (cf. Wasow 1997; Cappelle 2009:â•›149). In the present chapter it is simply assumed that because the analytic adpositional construction with peal ‘on’ can create a heavier constituent effected by the extra lexeme (peal ‘on’), it weighs more than the synthetic adessive case. However, it should be noted that length is only one aspect of how the principle of end-weight can be operationalized. For the present study, the word order factor was coded according to the position of the locative phrase in the clause with the levels of ‘final’, ‘initial’ and ‘middle’. Estonian is a language with a relatively free word order and the locative phrases can come at the beginning of a clause (in which case the clause is referred to as an existential clause in Estonian reference grammars), at the end of a clause or in the middle of a clause. However, there was no significant difference between the two construction types in the present dataset (see Table 6), χ²(2, N = 600) = 3.02, p = .22. Contrary to the purported principle of end-weight, the locative construction with the adessive case was slightly more frequent in the final position. This can be, at least partly, explained by the fact that the mean length of the locative constructions with the adessive was longer than the mean length of the locative constructions with the postposition peal ‘on’ and that these longer Landmark phrases with the adessive predominantly occur in the final position within a clause (see Section 4.1.1 above). 4.1.4 Word order: Loc_Nom vs. Nom_Loc In addition to the position of the locative phrase within the clause, it was decided to code whether the locative phrase follows or precedes the Trajector phrase. This factor has two levels – ‘Nom_Loc’ and ‘Loc_Nom’. ‘Loc’ refers to the locative phrase and ‘Nom’ to the Trajector phrase. The relevant frequencies are given in Table 7. In general, it can be seen that the preferred word order is such that the locative phrase follows the Trajector (393 occurrences in total). Although the raw frequencies indicate that when the locative phrase precedes the Trajector, the adessive is used more often


Table 7.â•‡ Word order: Loc_Nom vs. Nom_Loc Loc_Nom vs. Nom_Loc

Adessive

peal

Total

Loc_Nom Nom_Loc Total

111 189 300

â•⁄96 204 300

207 393 600

Table 8.â•‡ Verb lemmas used with locative constructions Verb lemma

Adessive

peal

Total

action verbs existence verbs motion verbs posture verbs no verb Total

113 â•⁄40 â•⁄46 â•⁄51 â•⁄50 300

â•⁄91 â•⁄85 â•⁄35 â•⁄45 â•⁄44 300

204 125 â•⁄81 â•⁄96 â•⁄94 600

than the adposition peal ‘on’, statistically these results are not significant, χ²(1, N = 600) = 1.66, p = .19.

4.1.5 Verb lemma Every instance of the adessive and the adposition peal ‘on’ was coded for the verb lemma used in these sentences. In total, there were 212 different verbs used with these locative constructions. The verbs were subcategorised into different groups based largely on Levin (1993) and include the following levels: ‘action verbs’ (e.g. tegema ‘to do’), ‘existence verbs’ (e.g. olema ‘to be’), ‘motion verbs’ (e.g. jooksma ‘to run’), and ‘posture verbs’ (e.g. istuma ‘to sit’). In addition, this factor also had the level of ‘no verb’ which was used for elliptical sentences where no overt verb lemma was expressed. The raw frequencies of these verb groups are given in Table 8. The Chi-squared test revealed that the frequencies of the two constructions significantly differed by verb lemma (χ²(4, N = 600) = 22.76, p < .001, φ = 0.19). The Pearson residuals show that existence verbs determine the difference between these frequency counts – the adpositional construction is significantly more often used with existence verbs like olema ‘to be’, asuma ‘to be situated’, and asetsema ‘to be placed’. 4.1.6 Syntactic function of the Landmark phrase Both the adessive and the adpositional construction can fulfil two syntactic functions – that of an adverbial and a modifier. It was therefore decided to code the syntactic function of the locative phrase in the dataset with precisely these two levels: ‘adverbial’ and ‘modifier’. The results show that there is a modest, but significant difference (Table 9), χ²(1, N = 600) = 5.78, p = .016, φ = 0.09. Although both the adessive

264 Jane Klavan

Table 9.â•‡ Syntactic function of the Landmark phrase Syn. function of Lm phrase

Adessive

peal

Total

adverbial modifier Total

264 â•⁄36 300

281 â•⁄19 300

545 â•⁄55 600

case and the postposition peal ‘on’ predominantly fulfil the function of an adverbial, the adessive case is slightly more frequently used in the function of a modifier, as in example (5). (5) ILU1980\stkt0023: Paigad küünarnukkidel ja hõlmadel tükkisid lahti … patch.pl.nom elbow.pl.ad and flap.pl.ad start-pst.3pl apart ‘Patches on the elbows and flaps started to come apart…’

4.1.7 Word class of Landmark and Trajector Different expression types have been found to affect the choice of syntactic alternatives (see, for example, Bresnan and Ford 2010). Both Landmarks and Trajectors were coded for the following types: ‘noun’, ‘pronoun’, ‘verb phrase’. The majority of Landmarks used with the adessive case and the postposition peal ‘on’ are noun phrases (Table 10). However, when the Landmark is a pronoun, the postposition peal ‘on’ is used more frequently. This difference is significant – χ²(1, N = 600) = 10.10, p = .001, φ = 0.13. This result is related to the variable length of the Landmark (see Section 4.1.1 above). Pronouns are short words and these results reflect, once again, the tendency to use the postposition peal ‘on’ with shorter Landmarks. Table 10.â•‡ Word class of Landmark Word class of Landmark

Adessive

peal

Total

noun pronoun Total

292 â•⁄â•⁄8 300

274 â•⁄26 300

566 â•⁄34 600

Word class of Trajector

Adessive

peal

Total

noun pronoun verb phrase Total

210 â•⁄49 â•⁄41 300

166 â•⁄82 â•⁄52 300

376 131 â•⁄93 600

Table 11.â•‡ Word class of Trajector


The majority of Trajectors used with both locative constructions are also noun phrases (Table 11). Nevertheless, curiously, the same tendency to use the adpositional construction with pronouns occurs and the difference is also significant, χ²(2, N = 600) = 14.76, p < .001, φ = 0.16.

4.2

Semantic variables

Numerous cognitive-functional studies on spatial language expressions have shown that various properties of Trajector and Landmark participating in the locative construction influence the use of spatial expressions (e.g. Talmy 1983; Herskovits 1986; Vandeloise 1991; Feist and Gentner 2003; Coventry and Garrod 2004; Carlson and Van der Zee 2005). In the vein of this research tradition, various semantic properties of Landmarks and Trajectors were coded in the present data.

4.2.1 Type of relation between Landmark and Trajector It has been suggested in previous work on cases and adpositional constructions that cases are semantically more abstract than adpositions (Bartens 1978; Comrie 1986; Luraghi 1991:â•›66–67; Ojutkangas 2008; Hagège 2010:â•›37–38; Lestrade 2010). Both the Estonian adessive and the adposition peal ‘on’ can express spatial and abstract relations between a Trajector and a Landmark. This variable was coded in the dataset with the levels of ‘abstract’ and ‘spatial’ in order to see if this general assumption of cases expressing abstract relations was also borne out in the present data. A relation was coded abstract when either the Trajector or Landmark was abstract or the relation itself was abstract, i.e. if there was a meaning transfer. There was a marginal significant difference, χ²(1, N = 600) = 5.04, p = .02, φ = 0.10; indeed, the adposition peal ‘on’ is more frequent with abstract relations in the present dataset (Table 12). However, one must bear in mind here that for the present analysis, only such occurrences where the alternation between the adessive and the adposition peal ‘on’ is possible were looked at. If we compare the general usage of these two constructions, we can easily see that the adessive case expresses abstract functions, where the use of an adposition is not possible (cf. Section 2 above, examples (2c) and (2d)). Table 12.â•‡ Type of relation between Landmark and Trajector Relation type

Adessive

peal

Total

abstract spatial Total

â•⁄49 251 300

â•⁄71 229 300

120 480 600

266 Jane Klavan

4.2.2 Type of Landmark It can also be predicted that there is a general difference between what types of Landmarks are used together with either the Estonian adessive case or the adposition peal ‘on’. Accordingly, Landmarks in the dataset were coded for their type, the levels of which were ‘location’ (e.g. street) and ‘object’ (e.g. table). The rationale of making a distinction between small easily manipulable objects and large static objects or locations is that location should lend itself more easily for abstraction and hence is more likely to be used with the adessive (Bartens 1978). On the other hand, as has been put forward in previous studies, adpositions are more concrete and specific than cases, and they convey the meaning of spatial location of an object more clearly and should thus be more frequent with small easily manipulable objects such as Landmarks (Bartens 1978; Palmeos 1985:â•›18; Comrie 1986; Luraghi 1991:â•›66–67; Ojutkangas 2008; Hagège 2010:â•›37–38; Lestrade 2010). The results of the Chi-squared test indicate that the frequency counts of the adessive and the adposition peal ‘on’ significantly differed by the type of Landmark (χ²(1, N = 600) = 11.21, p < .001, φ = 0.14) – the adessive tends to be used when the Landmark is a location and the adposition peal ‘on’ when it is an object (Table 13). 4.2.3 Animacy of Landmark and Trajector Since animacy is considered to be a very important cognitive category and is discussed in numerous linguistic and psycholinguistic studies (for overviews, see, for example, de Vega et al. 2002:â•›121–122; Feist and Gentner 2003:â•›2; Bresnan and Ford 2010:â•›10), it was decided to code this category for the Estonian adessive and adposition peal ‘on’ dataset as well. This variable has only two levels – ‘animate’ and ‘inanimate’. Unsurprisingly, the results show that the adposition peal ‘on’, rather than the adessive, is used in cases of animate Landmarks (Table 14).

Table 13.â•‡ Type of Landmark Landmark

Adessive

peal

Total

location object Total

169 131 300

128 172 300

297 303 600

Animacy of Landmark

Adessive

peal

Total

animate inanimate Total

â•⁄â•⁄2 298 300

â•⁄17 283 300

â•⁄19 581 600

Table 14.â•‡ Animacy of Landmark


The results of the Fisher’s exact test revealed a significant difference – p < .001; the odds ratio is 0.11. However, here it must be, once again, emphasised that the adessive has another important function in Estonian besides expressing space – that of expressing possession (see example (2c) above). Since animate objects are very apt to possessing things, the combination of an animate Landmark and the adessive case fulfils the function of the possessive construction. Therefore, if there is need in Estonian to talk about an object placed on top of an animate Landmark, this would be expressed with the adposition peal ‘on’. There was no significant difference between the adessive and the adposition peal ‘on’ for the variable animacy of Trajector (Table 15), χ²(1, N = 600) = 0.44, p = .51. The data do, however, confirm the general tendency of Trajectors to be more frequently animate than inanimate.

4.2.4 Number of Landmark and Trajector Another cognitively and typologically important category in grammar is number (Greenberg 1966), which also plays a role in certain grammatical variations (e.g. Bresnan and Ford 2010:â•›179). In the dataset, both Landmarks and Trajectors were coded either ‘plural’ or ‘singular’. When context and formal plural marking were in conflict or when there was ambiguity, the analysis proceeded from the context. The results (Table 16) show that there is no difference between the two locative constructions. Although the proportion of plural Landmarks is a little higher with the adessive construction, this difference is not significant: χ²(1, N = 600) = 2.84, p = .09. Interestingly, there is a difference in the use of the adessive and the adposition peal ‘on’ according to whether the Trajector is singular or plural – the proportion of adessive occurrences with a plural Trajector is higher than the proportion of the adposition peal ‘on’ occurrences with a plural Trajector (Table 17) and this difference is significant: χ²(1, N = 600) = 4.03, p = .04, φ = 0.08. This may be due to an interaction Table 15.â•‡ Animacy of Trajector Animacy of Trajector

Adessive

peal

Total

animate inanimate Total

167 133 300

175 125 300

342 258 600

Number of Landmark

Adessive

peal

Total

plural singular Total

â•⁄40 260 300

â•⁄27 273 300

â•⁄67 258 600

Table 16.â•‡ Number of Landmark

268 Jane Klavan

Table 17.â•‡ Number of Trajector Number of Trajector

Adessive

peal

Total

plural singular Total

â•⁄94 206 300

â•⁄72 228 300

166 434 600

Mobility of Landmark

Adessive

peal

Total

mobile static Total

116 184 300

178 122 300

294 306 600

Table 18.â•‡ Mobility of Landmark

with another variable – type of Landmark.2 Since the adessive tends to be used with locations and since locations are large, immobile objects such as streets or fields, one would expect to find more than one object (Trajector) located on such large locations, e.g. people/cars on the street vs. a person/a car on the street. Indeed, a little more than half of the occurrences with a plural Trajector have location as the type of Landmark in the dataset.

4.2.5 Mobility of Landmark and Trajector Following de Vega et al. (2002), both Landmarks and Trajectors in the dataset were coded for the variable mobility, the levels of which were ‘mobile’ and ‘static’. Mobile objects are those that do not have a fixed position in the environment, either because they move by themselves (e.g. humans, animals) or can be moved by an external agent (e.g. a table). Static objects (the majority of which in the dataset are also locations, but not all) have a fixed position in the environment (e.g. streets, trees). It can be seen from Table 18 that the adessive very frequently occurs with a static Landmark and the adposition peal ‘on’ with a mobile Landmark; this difference is also highly significant: χ²(1, N = 600) = 25.64, p < .001, φ = 0.21. There was no significant difference between these two constructions for the mobility of Trajector, χ²(1, N = 600) = 0.15, p = .69 (Table 19). 4.2.6 Relative size between the Landmark and the Trajector In Cognitive Linguistic analyses of spatial expressions, it has been claimed that Landmarks tend to be larger than Trajectors (Talmy 1983; Herskovits 1986; Langacker 1987; Vandeloise 1991). In order to validate this claim and to see whether this factor influences the use of the Estonian adessive case and the postposition peal ‘on’, the 2. I am indebted to Krista Ojutkangas for this suggestion (p.c.).


Table 19.â•‡ Mobility of Trajector Mobility of Trajector

Adessive

peal

Total

mobile static Total

268 â•⁄32 300

265 â•⁄35 300

533 â•⁄67 600

Table 20.â•‡ Relative size between the Landmark and the Trajector Relative size between Tr and Lm

Adessive

peal

Total

conventional same unconventional Total

193 â•⁄58 â•⁄49 300

140 â•⁄95 â•⁄65 300

333 153 114 600

relative size between the Landmark and Trajector was coded either as ‘conventional’ (Landmark > Trajector), ‘same’ (Landmark = Trajector) or ‘unconventional’ (Landmark < Trajector). The results indicate that in general, Landmarks tend to be indeed larger than Trajectors. Moreover, there is a difference between the adessive and the adposition peal ‘on’. The adessive is used when the Trajector is smaller than the Landmark and the adposition peal ‘on’ when the Trajector and the Landmark are of the same size or when the Trajector is bigger than the Landmark (Table 20). This difference is significant: χ²(2, N = 600) = 19.63, p < .001, φ = 0.18.

4.3

Summary of the variables

Many of the variables that are described and that have been put forward in the literature on variation were confirmed to have a significant effect on the use of the Estonian adessive and adposition peal ‘on’. Table 21 presents in a summary fashion all the variables argued to contribute to the alternation between the Estonian adessive and the adposition peal ‘on’; the p-values and effect sizes were obtained using the Chi-square test (see Gries, this volume). Out of the sixteen variables discussed, nine were highly significant, three marginally significant and four were not significant. The highly significant factors were the length of the Landmark phrase (in both words and syllables), the lexical complexity of Landmark, the verb lemma, word class of Landmark and Trajector, type, animacy and mobility of Landmark, and the relative size between Trajector and Landmark. The factors that were marginally significant were the syntactic function of the Landmark phrase, type of relation between Trajector and Landmark, and the number of Trajector; the factors that were not significant include the following: word order, animacy of Trajector, number of Landmark, and mobility of Trajector.

270 Jane Klavan Table 21. Variables that are argued to contribute to the alternation between the Estonian adessive and the adposition peal ‘on’ Variable name

Level for the adessive construction

Level for the adpositional construction

p-value

Effect size

Length of Lm phrase in words Length of Lm phrase in syllables Lexical complexity of Landmark Word order: Position of the locative phrase Word order: Nom_Loc/Loc_Nom Verb lemma Syntactic function of Lm phrase Word class of Lm Word class of Tr Type of relation between Tr & Lm Type of Lm Animacy of Lm Animacy of Tr Number of Lm Number of Tr Mobility of Lm Relative size Tr & Lm

2 or more words 4 or more syllables compound

1 word 1–3 syllables single lexeme

action verbs modifier

existence verbs

p < .001 p < .001 p < .001 p = .22 p = .19 p < .001 p = .02 p < .001 p < .001 p = .03 p < .001 p < .001 p = .51 p = .09 p = .04 p < .001 p < .001

– – 0.3 – – 0.2 0.1 0.1 0.2 0.1 0.1 0.1 – – 0.1 0.2 0.2

noun location

plural static Lm > Tr

pronoun pronoun abstract object animate

mobile Lm = Tr; Lm < Tr


One of the most surprising results among the less significant factors was the variable word order. Taking into consideration previous studies on grammatical variation, specifically the proposed end-weight principle (see Section 4.2.1), it was predicted that since the adposition peal ‘on’ adds an extra word to the locative phrase, thus making the whole locative phrase longer, it would prefer the final position within the clause. Instead, the locative phrases with the adessive occurred slightly more frequently in the final position. A possible explanation is the interaction between the length of the Landmark phrase and word order – longer Landmark phrases were used with the adessive in the dataset. It is clear, however, that this issue needs further research. Another morphosyntactic variable that seems to play at least a marginal role in the alternation between the two locative constructions is the syntactic function of the locative phrase. The first function is considerably more frequent, but if the locative phrase is used in the modifier function, it tends to be expressed by the adessive and not with the adposition. The results also show that there was a tendency for the adessive to occur when the Landmark was a location, and for the adposition peal ‘on’ to occur when it was an object. Unpredictably, the plurality of Trajector stood out – with plural Trajectors, the adessive construction was frequent. This interacts with the previous factor – type of Landmark – the adessive occurring with locations as Landmarks. Locations, in turn, tend to imply more than one Trajector. Out of the nine highly significant factors, five were morphosyntactic and four were semantic factors. Length of the Landmark phrase proved to be highly significant in a number of ways. First and foremost, the mean scores for the length of Landmark phrase both in words and syllables were different for the adessive and the adposition peal ‘on’. If the Landmark phrase is more than two words or more than four syllables long, the adessive is used; if the Landmark phrase is composed of only one word or is three or less syllables long, the adposition peal ‘on’ is used. The tendency to use the adposition with shorter Landmarks was also illustrated with the factor lexical complexity of the Landmark phrase – single lexemes were used with the adposition peal ‘on’ and compound words with the adessive. Unpredictably, there was a significant difference between the two constructions in the type of verb used – the adposition peal ‘on’ was frequently used with existence verbs. Another unpredictable significant factor was the word class of Trajector – the adposition peal ‘on’ was frequently used with pronominal Trajectors. From the semantic factors, type, animacy and mobility of Landmark, and the relative size between Trajector and Landmark were highly significant. Due to the fact that animate Landmarks with the adessive case express the possessor, the spatial support relation with animate Landmarks is expressed with the adposition peal ‘on’. The factor mobility of Landmark interacts with the type of Landmark – the adessive is used with static Landmarks, which in many cases are locations, and the adposition peal ‘on’ with mobile Landmarks, which in many cases are objects. Furthermore, the adessive is used when the Landmark is bigger than the Trajector; when the Landmark

272 Jane Klavan

and the Trajector are of the same size or when the Trajector is bigger than the Landmark, the adpositional construction was more frequent. Although the results of the monofactorial analysis indicate that a number of factors are significant in the alternation between the Estonian adessive and the adposition peal ‘on’, this way of analysing the data is not sufficient on its own. When speakers use either of these locative constructions, they probably do not consider the value of one factor only – in actual language use, all of the factors interact simultaneously and need to be analysed as such. Therefore, a multifactorial approach is necessary to determine which of the variables are more decisive and predictive for the choice of the construction. In the following section, I will present the results of a logistic regression analysis.

5. Multifactorial results. Logistic regression analysis Multiple logistic regression (Glynn 2007:â•›241–275; Baayen 2008:â•›195–208) is used to quantify the contribution of the factors presented above to the alternative use of the Estonian adessive and the adposition peal ‘on’. Being a confirmatory modelling technique, regression analysis gives probabilistic scores and calculates the explanatory power of the model (Glynn 2010:â•›257). I will not discuss here the details that lie behind this statistical technique. The interested reader is referred to elsewhere in this volume for a detailed overview of what the model does, what are its inner workings, its weaknesses and strengths; Glynn (2007, 2010), for example, gives an introduction to the mechanics of this technique. The results of the monofactorial analyses showed that many different factors were important in determining the difference between the adessive and the adpositional constructions. All of these factors were included in the regression modelling and several models were run with various combinations of these factors. Due to the concerns of multicollinearity, a number of factors could not be entered into the model simultaneously; this in turn increased the total number of models created. After comparing a range of models, the most significant and explanatory model was selected. The analysis was performed in R (version 2.10.1) and the model is presented below: Binominal Logistic Regression

Locative Phrase ~ Verb_Lemma + WO_LocNom + LM_LexComp + LM_WC + LM_Mobility

Deviance Residuals: Min

-2.1839

1Q

-1.0214

Median 0.2539

3Q

0.9923

Max

2.2529


Coefficients:

Estimate Std.Error

(Intercept)

z value Pr(>|z|)

-1.0388

0.3293

-3.155

Verb_Lemma Existence 1.4509

0.3347

4.334

Verb_Lemma Action Verb_Lemma Motion

Verb_Lemma Posture WO_LocNom Nom_Loc LM_LexComp Comp LM_WC Pron

LM_Mobility Mobile ---

0.3116 0.1471 0.3205 0.5132

-2.2415 2.0282 0.9781

0.2913 0.3485 0.3334 0.2061

1.070 0.422 0.962 2.490

0.3244

-6.911

0.1929

5.070

0.6405

3.167

0.00161 0.28471

**

1.46e-05 *** 0.67304 0.33626 0.01277

*

0.00154

**

4.83e-12 *** 3.97e-07 ***

Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 (Dispersion parameter for binomial family taken to be 1) Null deviance: 822.06 on 592 degrees of freedom

Residual deviance: 676.47 on 584 degrees of freedom AIC: 694.47

Number of Fisher Scoring iterations: 4 Summary of Model

Predictive Power of Model

D.f.: 8

C: 0.764

Model L.R .: 145.59 P: 0

Pseudo R²: 0.29

Somers’ Dxy: 0.527

Adding interactions and changing the order of the levels did not improve the model. The Nagelkerke R², a pseudo R² statistic for logistic regression, gives us an idea of the variability accounted for by the present model. As Glynn (2010:â•›259) points out, any figure above 0.3 is a sign of predictive power (for the model under discussion, the figure is just below 0.3–0.29). Another important score is the C-score (ROC), which is a scaled rank correlation between predicted and observed outcomes. Although not expressed in terms of probability of success, it can be interpreted as a rough indicator of such, where 1 represents prefect predictions and .5 pure chance. Although the score of C–0.764 does not indicate a predicatively strong model (.8 is typically taken as a predictive model), it does indicate that predictor variables explain a reasonable amount of the differences in use. The model was also checked for multicollinearity by calculating the variance inflation factors. Over dispersion does not appear to be a serious issue either. Although the indicators of the model’s explanatory power are all close to the lower boundary, it can be concluded that the model still bears some statistical significance and explanatory power and a cursory look at the results can be taken.

274 Jane Klavan

The model includes the factors type of verb (Verb_Lemma), word order (WO_ LocNom), lexical complexity of the Landmark phrase (LM_LexComplexity), word class of the Landmark (LM_WC) and mobility of the Landmark (LM_Mobility) and their different values or features – these predict the outcome of an example as either the adessive or the adpositional construction. In the list of the estimates of the coefficients, negative numbers predict the adessive construction and positive numbers the adpositional construction. Lexical complexity of the Landmark is the most important predictor for the adessive construction. This confirms the results of the monofactorial analysis, where we saw that when the Landmark word is a compound word, the adessive construction is used. Important predictors of the adpositional construction are the word class of the Landmark with the level of ‘pronoun’, verbs of existence, mobility of the Landmark with the level of ‘mobile’ and such a word order sequence where the locative phrase follows the Trajector. Glynn (2010:â•›260) points out that “[a]s a rule, any figure higher or lower than +/–1 is a relatively important predictor”. Since almost all of the significant levels or values are around +/–1, it can be concluded that in combination, the factors included in the model are fairly strong predictors. It seems, therefore, that although monofactorial results showed that for the present dataset a large number of factors were highly significant in determining the alternative use of the adessive and the adposition peal ‘on’ only five of these factors are predictive when we consider all of the factors together in combination, i.e. in a truly multifactorial situation what everyday language use no doubt is. One of the conclusions to be drawn from the fact that, in theory, the model could have more explanatory power is that there may be other factors not included in the present analysis that may play an even more significant role in determining the use of these two constructions. For instance, all of the discourse-functional variables, such as topic, register, preceding and subsequent mention of the adessive or adpositional construction, etc., and variables like idiolect and dialect are absent from the present analysis. Furthermore, the corpus data discussed in this study comes from a corpus of fiction, where the language used is that of edited written texts. Incorporating data from spoken or internet language (i.e. unedited texts) may provide very useful insights into the analysis and the model in general, which in turn would produce more coherent results that would predict the data more accurately. Nevertheless, both the monofactorial and multifactorial analyses of the present data also systematically indicate that there are significant differences between the Estonian adessive and the adposition peal ‘on’, as can be predicted if we proceed from the premises of Cognitive Grammar and Construction Grammar.


6. Conclusion The present chapter looked at the alternation between the Estonian adessive case and the adposition peal ‘on’ in the corpus of present-day written Estonian. Both the monofactorial and multifactorial analyses showed that the use of these two constructions is determined by a variety of morphosyntactic and semantic variables. More specifically, the multifactorial analyses of the data confirmed the statistical influence of the following factors: lexical complexity of Landmark, type of verb, word order, word class of Landmark, and the mobility of Landmark. Adessive tends to be used when the Landmark is lexically more complex, when it is static, and when the locative phrase precedes the Trajector. Adposition peal ‘on’ tends to be used together with verbs of existence and pronominal Landmarks, when the Landmark phrase is lexically simple, when the Landmark is animate and mobile, and when the locative phrase follows the Trajector. Even though the results of the corpus analysis confirmed the prediction that there are differences in the use of the adessive and the adposition peal ‘on’, we should be careful about drawing any far-reaching conclusions because the model obtained as a result of running the logistic regression analysis is only marginally powerful. This may indicate that there are other more important factors that better predict the alternation between the adessive and the adpositional construction which are absent from the present analysis (e.g. discourse-functional factors, idiolect, dialect). The corpus analysis results discussed in the present chapter do not facilitate a very good comparison with the results obtained by Klavan et al. (2011), since the latter included in their studies only semantic factors, thus excluding morphosyntactic ones. Nevertheless, as the logistic regression analysis showed, at least one of the semantic factors was an important predictor – mobility of Landmark; static Landmarks predict the adessive construction and mobile Landmarks the adpositional construction. Klavan et al. (2011) showed, in turn, that with locations as Landmarks the adessive is used, and with objects as Landmarks the adposition peal ‘on’ is used; since locations are predominantly static and objects mobile, there is converging evidence that the type of Landmark does influence the alternative use of the adessive and the adposition peal ‘on’. It is hoped that the chapter succeeded in demonstrating the necessity and utility of using a combination of methodologies in studying grammatical (and other) variation phenomena.

276 Jane Klavan

References Arnold, J. E., Wasow, T., Losongco, A., & Ginstrom, R. (2000). Heaviness vs. newness: The effects of complexity and information structure on constituent ordering. Language, 76(1), 28–55. Baayen, H. (2008). Analyzing linguistic data: A practical introduction to statistics using R. Cambridge: Cambridge University Press. DOI: 10.1017/CBO9780511801686 Balanced Corpus of Estonian (2008). Retrieved from Bartens, R. (1978). Synteettiset ja analyyttiset rakenteet lapin paikanilmauksissa [Suomalais-ugrilaisen Seuran toimituksia 166]. Helsinki: Suomalais-Ugrilainen Seura. Bresnan, J., Cueni, A., Nikitina, T., & Baayen, H. (2007). Predicting the dative alternation. In G. Bouma, I. Kraemer, & J. Zwarts (Eds.), Cognitive foundations of interpretation (pp. 69– 94). Amsterdam: Royal Netherlands Academy of Science. Bresnan, J., & Ford, M. (2010). Predicting syntax: Processing dative constructions in American and Australian varieties of English. Language, 86(1), 168–213. DOI: 10.1353/lan.0.0189 Cappelle, B. (2009). Contextual cues for particle placement: Multiplicity, motivation, modelling. In A. Bergs, & G. Diewald (Eds.), Context in construction grammar (pp. 145–191). Amsterdam & Philadelphia: John Benjamins. Carlson, L., & Van der Zee, E. (2005). Functional features in language and space: Insights from perception, categorization, and development. Oxford: Oxford University Press. Comrie, B. (1986). Markedness, grammar, people, and the world. In F. R. Eckman, E. A. Moravcsik, & J. R. Wirth (Eds.), Markedness (pp. 85–106). New York: Plenum. DOI: 10.1007/978-1-4757-5718-7_6 Cooper, W. E., & Ross, J. R. (1975). World order. In R. E. Grossman, J. L. San, & T. J. Vance (Eds.), Chicago linguistic society: Papers from the parasession on functionalism (pp. 63– 111). Chicago: Chicago Linguistic Society. Coventry, K. R., & Garrod, S. C. (2004). Saying, seeing, and acting: The psychological semantics of spatial prepositions. New York: Psychology Press. de Vega, M., Rodrigo, M. J., Ato, M., Dehn, D. M., & Barquero, B. (2002). How nouns and prepositions fit together: An exploration of the semantics of locative sentences. Discourse Processes, 34(2), 117–143. DOI: 10.1207/S15326950DP3402_1 Erelt, M., Kasik, R., Metslang, H., Rajandi, H., Ross, K., Saari, H., Tael, K., & Vare, S. (1993). Eesti keele grammatika II: Süntaks [The grammar of Estonian II: Syntax]. Tallinn: Eesti Teaduste Akadeemia Keele ja Kirjanduse Instituut. Erelt, M., Kasik, R., Metslang, H., Rajandi, H., Ross, K., Saari, H., Tael, K., & Vare, S. (1995). Eesti keele grammatika I: Morfoloogia [The grammar of Estonian I: Morphology]. Tallinn: Eesti Teaduste Akadeemia Eesti Keele Instituut. Erelt, M., Erelt, T., & Ross, K. (2007). Eesti keele käsiraamat [Handbook of Estonian]. Tallinn: Eesti Keele Sihtasutus. Feist, M., & Gentner, D. (2003). Factors involved in the use of in and on. In R. Alterman & D. Kirsh (Eds.), Proceedings of the twenty-fifth annual meeting of the Cognitive Science Society (pp. 390–395). Boston MA: Cognitive Science Society. Glynn, D. (2007). Mapping meaning: Toward a usage-based methodology in cognitive semantics. Unpublished PhD thesis, Katholieke Universiteit Leuven.


Glynn, D. (2010). Testing the hypothesis: Objectivity and verification in usage-based cognitive semantics. In D. Glynn, & K. Fischer (Eds.), Quantitative Cognitive Semantics. Corpus-driven approaches (pp. 239–270). Berlin: Mouton de Gruyter. DOI: 10.1515/9783110226423 Goldberg, A. (1995). Constructions: A construction grammar approach to argument structure. Chicago: University of Chicago Press. Goldberg, A. (2006). Constructions at work: The nature of generalization in language. Oxford: Oxford University Press. Greenberg, J. (1966). Language universals, with special reference to feature hierarchies. The Hague: Mouton de Gruyter. Gries, St. Th. (2003). Grammatical variation in English: A question of ‘structure vs. function’? In G. Rohdenburg, & B. Mondorf (Eds.), Determinants of grammatical variation in English (pp. 155–173). Berlin: Mouton de Gruyter. Hagège, C. (2010). Adpositions. Oxford: Oxford University Press. DOI: 10.1093/acprof:oso/9780199575008.001.0001 Hawkins, J. A. (1994). A performance theory of order and constituency. Cambridge: Cambridge University Press. Herskovits, A. (1986). Language and spatial cognition: An interdisciplinary study of the prepositions in English. Cambridge: Cambridge University Press. Klavan, J., Kesküla, K., & Ojava, L. (2011). Synonymy in grammar: the Estonian adessive case and the adposition peal ‘on’. In S. Kittilä, K. Västi, & J. Ylikoski (Eds.), Studies on case, animacy and semantic roles (pp. 113–134). Amsterdam: John Benjamins. Langacker, R. W. (1987). Foundations of Cognitive Grammar. Volume I: Theoretical prerequisites. Stanford: Stanford University Press. Langacker, R. W. (2008). Cognitive Grammar: A basic introduction. Oxford: Oxford University Press. DOI: 10.1093/acprof:oso/9780195331967.001.0001 Lestrade, S. (2010). The space of case. Unpublished PhD dissertation, Radboud University Nijmegen. Levin, B. (1993). English verb classes and alternations: A preliminary investigation. Chicago: University of Chicago Press. Luraghi, S. (1991). Paradigm size, possible syncretism, and the use of adpositions with cases in flective languages. In F. Plank (Ed.), Paradigms: The economy of inflection (pp. 57–74). Berlin: Mouton de Gruyter. Mondorf, B. (2003). Support for more-support. In G. Rohdenburg, & B. Mondorf (Eds.), Determinants of grammatical variation in English (pp. 251–304). Berlin: Mouton de Gruyter. DOI: 10.1515/9783110900019 Morphologically Disambiguated Corpus (2010). Retrieved from Ojutkangas, K. (2008). Mihin suomessa tarvitaan sisä-grammeja? Virittäjä, 112(3), 382–400. Palmeos, P. (1985). Eesti keele grammatika II: Kaassõna [The grammar of Estonian II: Adposition]. Tartu: TRÜ trükikoda. Rannat, R. (1991). Noomeni sünteetiliste ja analüütiliste vormide kasutus [The use of the synthetic and analytic forms of the noun]. Unpublished BA dissertation, University of Tartu. Rosenbach, A. (2003). Aspects of iconicity and economy in the choice between the s-genitive and the of-genitive in English. In G. Rohdenburgand, & B. Mondorf (Eds.), Determinants of grammatical variation in English (pp. 379–411). Berlin: Mouton de Gruyter.

278 Jane Klavan

Salm, S. (2010). Kaassõnade ‘sees’ ja ‘peal’ ning vastavate kohakäänete kasutust mõjutavad tegurid [The factors influencing the use of Estonian adpositions sees ‘in’ and peal ‘on’ and the corresponding locative cases]. Unpublished BA dissertation, University of Tartu. Szmrecsanyi, B. (2010). The English genitive alternation in a cognitive sociolinguistic perspective. In D. Geeraerts, G. Kristiansen, & Y. Peirsman (Eds.), Advances in cognitive sociolinguistics (pp. 141–166). Berlin & New York: Mouton de Gruyter. Talmy, L. (1983). How language structures space. In H. Pick, & L. P. Acredolo (Eds.), Spatial orientation: Theory, research and application (pp. 225–282). New York: Plenum Press. DOI: 10.1007/978-1-4615-9325-6_11 Vainik, E. (1995). Eesti keele väliskohakäänete semantika kognitiivse grammatika vaatenurgast [The semantics of Estonian external locative cases from the perspective of Cognitive Grammar]. Tallinn: Eesti Keele Instituut. Vandeloise, C. (1991). Spatial prepositions: A case study from French. Chicago: University of Chicago Press. Wasow, T. (1997). Remarks on grammatical weight. Language Variation and Change, 9(1), 81– 105. DOI: 10.1017/S0954394500001800 Wulff, S. (2003). A multifactorial corpus analysis of adjective order in English. International Journal of Corpus Linguistics, 8(2), 245–82. DOI: 10.1075/ijcl.8.2.04wul

A diachronic corpus-based multivariate analysis of “I think that” vs. “I think zero” Christopher Shank, Koen Plevoets, and Hubert Cuyckens Bangor University / University College Ghent / University of Leuven

This corpus-driven study seeks to explain the choice between the zero complement and the that complement constructions, when occurring with the mental state predicate think. Previous studies have identified a range of factors that are argued to explain the alternation patterns. Such studies have also proposed that there is a diachronic drift towards zero complementation. Based on a sample of 9,720 think tokens, from both spoken and written corpora, from between 1560–2012, we test the hypothesis of diachronic change and the effect of eleven proposed factors on the constructional alternation. Using logistic regression, we demonstrate that, contrary to previous studies, there is in fact a diachronic decrease in zero complementation. Moreover, the study also demonstrates the importance of understanding the interaction of the various factors that explain the near-synonymous relation, including, especially, between the spoken and written modes. Keywords: complementation, logistic regression, mental state verb, near-synonymy, that/zero alternation

1. Introduction This chapter uses a corpus-based approach in conjunction with logistic regression analysis to understand the near-synonymous motivation for the diachronically varied alternation of the complementizer zero and that constructions with the verb think. The analysis examines the periods from 1560 to 2010, in both written and spoken genres, as exemplified in the examples below. (1) I think that Powder is a vile bragger, he doth nothing but cracke. (CED, 1560–1673) (2) I think you can marry non but me; seinge we are sworne to be true. (CED, 1560–1673)

280 Christopher Shank, Koen Plevoets, and Hubert Cuyckens

In previous studies, it has been suggested that this alternation is being lost and that the zero form is generalizing to substitute the that form (Rissanen 1991; Thompson and Mulac 1991; Palander-Collin 1999). The current chapter seeks to test this hypothesis by means of a stepwise logistic regression analysis of (n = 5801) tokens of think, the most frequently used complement-taking verb of cognition, spanning the time period from 1560 to 2012. The literature has also put forward a number of motivating factors promoting the zero form. Logistic regression determines the importance of different factors by treating the alternation as a choice that it attempts to predict based on the behaviour of the factors. Our regression model tests whether the factors proposed in the literature do indeed predict the zero form, the potential impact that these individual factors may or do have upon each other when combined in terms of their overall predictive power from both a synchronic and diachronic perspective, and finally their ability to predict the zero form over time. Determining the interaction of time with each of the structural conditioning factors, this study adds an innovative diachronic perspective to existing research into zero/that alternation by testing the effect of time as a factor on the selection of the zero complementizer. Logistic regression represents an established technique in the attempts to understand near-synonymous constructions. Some recent applications of the technique to questions of constructional near-synonymy include Heylen (2005), Bresnan et al. (2006), Grondelaers et al. (2007; 2008), Divjak (2010), Glynn (2010), Speelman and Geeraerts (2010), Klavan (this volume), Levshina (this volume), Speelman (this volume), inter alia. This study extends the principle to diachronic research. Although the analysis is restricted to formal usage-features, it is believed that these formal characteristics of use are sufficient to adequately distinguish the two semantically extremely similar forms. Moreover, it is hoped that these differences can be charted over time. We start off with a review of the literature dealing with that/zero alternation in order to characterize the construction under investigation and to review the factors that have previously been said to condition the use of either that or zero complementation. In Section 3, our data and methodology are explained. After presenting our results in Section 4, we offer a conclusion in Section 5.

2. Review of the literature The increase in structural/clausal flexibility that emerged in English starting in the late ME and EModE periods had a profound impact on many facets of early English syntax especially in regards to the fixation of SVO word order, clause combining, and complementation patterns. One of more important shifts, especially in regards to grammaticalization research, has concerned the observed decrease in the frequency of the that-complementizer and a corresponding increase in the zero complementizer form

A diachronic corpus-based multivariate analysis 281

(Rissanen 1991; Hopper and Traugott 1993; Finegan and Biber 1995, and Palander-Â� Collin 1999). The most often cited study is that of Rissanen (1991) who used the Helsinki corpus to examine the development and use of the that/zero alternation in think, know, say, and tell constructions with object clauses in Late Middle and Early Modern English. His analysis revealed a steady increase in the deletion of that as an object clause link in think constructions from 14% in the years 1350 to 1420 up to nearly 70% by the period of 1640 to 1710. Other researchers such as Finegan and Biber (1995) have expanded upon Rissanen’s claims regarding the that/zero-alternation by analyzing a similar data set taken from the Archer Corpus to examine the period from 1650 to 1990. This type of analysis has illustrated the role variables such as genre plays in such alterations by demonstrating how more formal genres such as sermons, medical articles, and personal letters often retained the that-complementizer form. Initially, researchers used early corpus-based methodologies to document the diachronic increase in the zero complementizer in a number of different verbs (e.g. say, tell, think, know) from ME through PDE, then turned their attention to factors in both the matrix and complement clauses that might be motivating the observed and ongoing structural/clausal changes. With regards to matrix clause features, the subject has drawn attention from a number of authors. Thompson and Mulac (1991) utilized chi-square tests to demonstrate the impact that the higher relative frequency of a verb (e.g. think and guess) and the presence of I or you (versus other subject forms) as the subject of the matrix verb also facilitate the presence of the zero complementizer. Their findings were further complemented and built upon by Rissanen (1991) and Biber and Finegan (1995) who showed, via a simple proportional contrastive analysis, that the subject type (i.e. pronominal subjects), the person of the subject governing the object clause (especially 1st person), and again in the text type (especially with regards to informality), also contributed to a decline in the frequency of the that-clause. Other studies that have shown that pronouns, particularly ‘I’ or ‘you’, favour the use of zero include Elsness (1984), Tagliamonte and Smith (2005) and Torres Cacoullos and Walker (2009). Another matrix clause factor that has received attention in the literature is the absence of additional material in the matrix clause. It is believed that matrix clauses containing elements other than a subject and a (simplex) verb are more likely to be followed by that. Such elements may be adverbials (Thompson and Mulac 1991a; Torres Cacoullos and Walker 2009); negations or periphrastic forms in the verbal morphology of the matrix clause predicate (Thompson and Mulac 1991a; Torres Cacoullos and Walker 2009). The presence of intervening material between matrix and complement has been widely discussed as a factor favouring that (Finegan and Biber 1985; Rissanen 1991; Rohdenburg 1996; Tagliamonte and Smith 2005; Torres Cacoullos and Walker 2009). Adjacency of matrix and complement clause is believed to minimize syntactic and


cognitive complexity (Torres Cacoullos and Walker 2009). Besides the risk of ambiguity, which Rohdenburg (1996:â•›160) regards as a special type of cognitive complexity, the presence of intervening material has been related to a heavier cognitive processing load. In the words of Rohdenburg (1996:â•›161), “any elements capable of delaying the processing of the object clause and thus the overall sentence structure favour the use of an explicit signal of subordination”. Another factor which has received attention in the literature on zero/that alternation is the subject of the complement clause. It has been suggested that pronominal subjects as opposed to full NPs favour the use of zero (Elsness 1984; Finegan and Biber 1985; Rissanen 1991; Thompson and Mulac 1991a; Rohdenburg 1996; Tagliamonte and Smith 2005; Torres Cacoullos and Walker 2009). (3) Bill, I understand you have a special guest with you.

(COCA)

(4) Well, I’m not, because I understand that most of his girlfriends have either (COCA) been, you know, like the hooker or porn star types. The high discourse topicality of pronouns has been proposed as an explanatory principle (Thompson and Mulac 1991a:â•›248), as well as Rohdenburg’s (1996:â•›151) complexity principle, which entails that “in the case of more or less explicit grammatical options the more explicit one(s) will tend to be favoured in cognitively more complex environments”. While Elsness (1984) regards I and you as particularly conducive to zero complementation, Torres Cacoullos and Walker’s (2009:â•›28) multivariate study results in the following ordering of subjects from least to most favourable to that: it/ there < I < other pronoun < NP. Elsness (1984) adds that short NPs and NPs with definite or unique reference are more likely to select the zero variant than longer and indefinite NPs. In Kearns (2007a:â•›494), first and second person subjects (i.e. I, you but also we) are compared to third person subjects, but identical rates of zero and that are found for both data sets. Kearns (2007:â•›493; 2007b:â•›304) also examines the length of the complement clause subject as a possible factor, operationalizing it in terms of a three-way distinction between pronouns, short NPs (one or two words) and long NPs (three or more words). The study reveals significant differences, including one between short and long NPs. Finally, other factors that have been shown to be influential with regards to increasing the frequency of the zero form have included “appropriate light heavy weight distribution pattern in the matrix and complement clause”, an “anaphoric relationship or givenness of the complement clause” (summarized in Kaltenböck 2004:â•›52), coreferentiality of either tense or person between the matrix and complement clauses, and the absence of the harmony of polarity between the matrix and complement clauses (Torres Cacoullos and Walker 2009). A considerable amount of the corpus-based research into the loss of the complementizer with the matrix verb think, while informative and clearly important, has, however, been inherently limited in terms of actual


Table 1.â•‡ Zero and that in Early Modern English: Subject types (Risannen 1991:â•›281) Model 1 (1500–1570) Pronoun Say Tell Know Think Total

Model 3 (1640–1710)

other

Pronoun

other

zero

that

zero

that

zero

that

zero

that

37 â•⁄6 18 16 77

47 13 12 â•⁄7 79

â•⁄7 â•⁄2 â•⁄5 â•⁄6 20

33 â•⁄7 â•⁄5 â•⁄6 51

â•⁄80 â•⁄47 â•⁄22 â•⁄48 197

â•⁄8 25 13 â•⁄2 48

22 â•⁄9 â•⁄7 19 57

22 25 â•⁄4 â•⁄9 60

explanatory or predictive power due to what we believe are a number of underlying methodological issues. For example, the seminal and often cited work done by Rissanen (1991) and Finegan and Biber (1995) on that deletion utilized the Helsinki Corpus, which covers a period ranging from c730 to 1710 and also contains only 1.5 million words, a small corpus by today’s standards. The sample sizes that these authors worked with were often less than 30 tokens per period which had the unfortunate effect of severely limiting the generalizability of their results. An example of this type of limitation is presented below in Table 1. In addition to small sample sizes, a number of the individual factors which were deemed to be predictive of the zero form were developed by contrasting and comparing the simple percentages of occurrences of a feature or variable in the resulting data sets, as a high percentage of occurrence was assumed to predict the presence of the zero complementizer. This type of methodological approach may reveal general structural trends and patterns but it does not give any substantive insight into the actual significance a given factor is actually playing in positively or negatively influencing the that/zero deletion process. It also does not allow for any valid inferences to be made concerning the diachronic validity of a particular factor remaining a significant predictor over time. More recent studies have benefited from larger written and spoken corpora (e.g. the Cobuild, Brown and Santa Barbara corpora) which have allowed for larger samples to be extracted and analyzed. These improvements have also coincided with the incorporation by researchers of statistical techniques such as chi-square testing into their research design and methodologies. The results from these newer studies, however, are limited by the fact that a chi-square test only reveals if a relationship exists between two variables (i.e. the presence of a structurally predictive variable for the zero complementizer form and the presence of the zero complementizer form in a given sentence). It does not indicate which specific outcome or which diachronic direction is being predicted (the variable could actually have the opposite effect in that it could actually be predicting, contra expectations, the that complementizer form).


In addition, the results presented in the synchronically-oriented studies are often used to make inferences about diachronic patterns and the effects of specific motivating factors; factors which were initially proposed or hypothesized by looking at small samples sizes (n < 30) and simply comparing the occurrence of a particular factor within a data set. It is with these types of methodological limitations and/or concerns in mind that we have designed our study. To address these aforementioned problems, our study utilizes an empirically motivated framework, a range of large diachronic corpora, and a statistical technique (regression analysis).

3. Data and methods of the current study Our analysis was based on tokens retrieved from the following written and spoken corpora (Tables 2 and 3). The Wordsmith concordance program was used first to identify the total number of inflected forms of think (i.e. think, thinks, thinking and thought) in both the written and spoken corpora from 1560–2012 per period. These results were then used to calculate the overall percentage of each inflected form relative to one another within the different periods. The percentages were then applied to the extracted subsets in order to ensure that the subsets would be proportionally similar in terms of inflected forms to the larger corpora from which they were taken. This two-step process resulted in Table 2.â•‡ Written corpora Sub-period of written English

Time span

Corpus

Number of words

Early Modern English (EModE)

1560–1710

â•⁄â•⁄2,848,314 Innsbruck Corpus of Letters CEECS I Corpus (1560–onward) CEECS II Corpus Corpus of English Dialogues (CED) Corpus of Early Modern English Texts (CMET) Lampeter Corpus (Early Modern English portion-up to 1710)

Late Modern English (LModE)

1710–1920

â•⁄15,413,159 Corpus of Late Modern English texts Extended Version (CLMETEV) Lampeter Corpus (Early Modern English portion (1710–onward)

Present-Day English (PDE)

1920–2009

The Time Corpus (Time) The Corpus of Contemporary American English – written component (COCA)

500,000,000


Table 3.â•‡ Spoken corpora Sub-period of spoken English

Time span

Early Modern English 1560–1710 (EModE)

Corpus

Number of words

Corpus of English Dialogues (CED) Old Bailey Corpus (OBC)

â•⁄â•⁄â•⁄â•‹980,320

Late Modern English (LModE)

1710–1913

Old Bailey Corpus (OBC)

113,253,011

Present-Day English (PDE)

1920–2012

The Corpus of Contemporary American English – spoken components (COCA) American National Corpus – spoken components (ANC) London-Lind Corpus (L-Lund) Alberta Corpus – 2010 component (Alberta)

133,180,448

Table 4.â•‡ Total number of tokens for think retrieved from the written and spoken corpora Date Written data

Total number of verbal forms

Date Spoken data

Total number of verbal forms

1560–1579 1580–1639 1640–1710 1710–1780 1780–1850 1850–1920 1920–1989 1990–2009 Total

(n = 100) (n = 638) (n = 1346) (n = 1440 ) (n = 1201) (n = 1297) (n = 280) (n = 317) (n = 6619)

1560–1579 1580–1639 1640–1710 1710–1780 1780–1850 1850–1913 1980–1993 1994–2012 Total

(n = 68) (n = 481) (n = 451) (n = 537 ) (n = 556) (n = 527) (n = 229) (n = 252) (n = 3101)

the following datasets for the verb think: (n = 3101 tokens from the written English corpora and n = 6619 tokens from the spoken English corpora). The full set (n = 9,720) of extracted sentences were analyzed and divided into those containing either a that-clause or a zero complementizer. The distribution of the remaining written (n = 2,217) and spoken (n = 3,584) sentences and resulting data sets are presented in Table 5 and Table 6. A comparison of the diachronic relative frequency patterns of the that versus zero forms per million words with the verb think indicates that the frequency of the zero form has remained relatively constant vis-à-vis the that-complementizer from 1560 to 2010 in both spoken and written genres. The zero form is clearly the more frequent


Table 5.â•‡ think in written corpora. Distribution of that-clauses and zero complementizer clauses from EModE to PDE in written corpora Period 1560–1579 1580–1639 1640–1710 1710–1780 1780–1850 1850–1920 1920–1989 1990–2009 Total

think-that

think-zero

think-that

think-zero

n

N

n

N

(n = 21) (n = 18) (n = 65) (n = 79) (n = 103) (n = 101) (n = 40) (n = 24) (n = 451)

214.00 â•⁄59.23 174.51 123.19 151.66 175.47 109.44 106.20

(n = 17) (n = 133) (n = 200) (n = 290) (n = 316) (n = 359) (n = 204) (n = 247) (n = 1766)

173.24 437.65 558.27 535.29 545.23 680.69 561.92 912.90

n: absolute frequency; N: normalized frequency per million

Table 6.â•‡ think in spoken corpora. Distribution of that-clauses and zero complementizer clauses from EModE to PDE in spoken English Period 1560–1579 1580–1639 1640–1710 1710–1780 1780–1850 1850–1913 1980–1993 1994–2012 Total

think-that

think-zero

think-that

think-zero

n

N

n

N

(n = 8) (n = 29) (n = 10) (n = 22) (n = 12) (n = 16) (n = 97) (n = 129) (n = 323)

â•⁄92.97 â•⁄86.37 â•⁄23.75 â•⁄45.64 â•⁄26.09 â•⁄47.50 449.18 471.64

(n = 28) (n = 116) (n = 212) (n = 412) (n = 439) (n = 418) (n = 857) (n = 779) (n = 3261)

â•⁄324.78 â•⁄345.48 â•⁄447.47 â•⁄854.10 â•⁄938.68 1305.45 3152.25 3139.33

n: absolute frequency; N: normalized frequency per million

form from 1560 to 2012 and this comports with all previous literature on think and claims regarding diachronic that/zero variation patterns. The (n = 2217) written and (n = 3584) spoken sentences containing either a that or zero complementizer clause were then coded for 26 features within three categories: corpus information, matrix clause features and complement clause features. The features included information such as the time period of the corpus (e.g. 1710–1780), the inflected form of the token and the full context in which it appeared. The matrix


1000.00 912.90

800.00 558.27

600.00 400.00 200.00

Written data – freq of the that –complemtizer

680.69 545.23

535.29

561.92

437.65 214.00 173.24

174.51

123.19

151.66

175.47

109.44

106.20

Written data – freq of zero complementizer

59.23

1990–2009

1920–1989

1580–1920

1780–1850

1710–1780

1640–1710

1580–1639

1560–1579

0.00

Figure 1.â•‡ Think in written data – that versus zero distribution per million words

3152.25

3139.33

Spoken data – freq of that –complementizer Spoken data – freq of zero complementizer

1305.45

449.18

471.64

1980–1993

47.50

1850–1913

26.09

1780–1850

1710–1780

1640–1710

1580–1639

447.47 92.97 324.78 345.38 86.37 45.64 23.75

938.68

1990–2012

854.10

1560–1579

3600.00 3200.00 2800.00 2400.00 2000.00 1600.00 1200.00 800.00 400.00 0.00

Figure 2.â•‡ Think in spoken data – that versus zero distribution per million words

and complement clauses of each extracted token were also coded for features such as person, tense, polarity, the length of the subject (pronoun / np-short 1–2 words / np-long 3+ words), and coreferentiality (or lack thereof). In addition, the presence (or absence) of additional elements within the matrix clause (elements between the subject and the matrix verb) was also noted along with intervening elements (between the matrix clause and the complementizer) and the location of the intervening elements (either pre/before or post/after the complementizer and before the complement clause subject). Once the coding process was completed, the data were submitted to the multiple logistic regression analysis modelled with the factors that have been claimed in the


Table 7.â•‡ Factors which favour the presence of the zero complementizer (summarized in Kaltenböck 2004:â•›52) 1. Matrix clause subjects are either I or you. (Elsness, 1984; Thompson and Mulac 1991; Tagliamonte and Smith 2005; Kearns 2007a) 2. The absence of extra elements in the matrix clause (viz. auxiliaries, indirect objects, adverbials) which reduce the ability of the matrix to function as an epistemic phrase by additional semantic content (cf. Thompson and Mulac 1991; Rohdenburg 1996) 3. The absence of intervening elements between the matrix and complement clause, making explicit boundary marking (disambiguation) with that unnecessary (Rissanen 1991; Tagliamonte and Smith 2005) 4. Pronominal subject of the complement clause, co-referential with the matrix clause subject (Elsness 1984; Torres Cacoullos and Walker 2009) 5. The length of the matrix clause subject (pronoun > np-short > np-long) (Thompson and Mulac 1991; Rissanen, 1991) 6. The length of the complement clause subject (it > pronoun > np-short > np-long) (Thompson and Mulac 1991; Rissanen, 1991; Rohdenburg 1996) 7. Coreferentiality of tense between the matrix and complement clauses (Torres Cacoullos and Walker 2009) 8. Coreferentiality of polarity between the matrix and complement clauses (Torres Cacoullos and Walker 2009)

literature (see Section 2) to favour the presence of the zero complementizer. Table 7 lists the factors that were included in our analysis.1 The statistical technique for our analysis is stepwise logistic regression analysis (which was run with the stepAIC-function in the R library MASS).2 The stepwise selection procedure was both-ways. The maximal model contained all main effects

1. In fact, we ran multiple analyses which also incorporated factors such as a pronominal subject (versus NP) in the matrix clause (Elsness 1984), a pronominal subject (versus NP) in the complement clause (Thompson and Mulac 1991; Rissanen 1991), and a 1st or 2nd person as the complement clause subject (Thompson and Mulac 1991). All factors cited in Section 3 were first analyzed in separate models and then also in various small subsets of factors, in order to get a better understanding of their relative potential in explaining the that/zero complementizer. These exploratory steps eventually led to the factors in Table 7. We would like to thank Dylan Glynn for his helpful comments in this respect. 2. The logical extension for our many predictors would be to fit a mixed-effects model. However, none of the factors in our data is a random effect, so we opted for a straightforward logistic regression, as explained in Speelman (this volume). The stepwise selection procedure then serves to identify the important predictors. The general outline of this methodology was suggested to us by Stefan Th. Gries, for which we express our gratitude.


plus the two-way interactions with mode and period (together with the interaction between period and mode itself).3 The resulting model after stepwise selection contained 9 main effects and 8 interactions which predict the zero form; see Table 8 for the model summary.4 The model fits well with the predicted variation (C-statistic, or the area under the ROC) just above the threshold of 80%. The rather modest explained variation (as expressed in Nagelkerke’s pseudo-R²) of 27% is just below the baseline of 30% (see Speelman, this vol., for how to interpret the model diagnostics of the C-statistic or Nagelkerke’s R²).5 The next section discusses the coefficients.

4. Discussion of the results Because of the rather complex structure of our regression model (with 9 main effects and 8 interactions), we interpret the results in Table 8 with the visual aid of so-called effect plots (obtained with the R library effects; see Gries 2013:â•›303). Furthermore, we divide the discussion into the main effects, the interactions with mode and the interactions with period. In 4.1, we present the nine main effects of period, mode, matrix-internal elements, intervening elements between the matrix and complement clause, complement clause subject length, I or you as the matrix clause subject, harmony of tense, subject coreferentiality and harmony of polarity, which predict the use of the zero form. In 4.2 we discuss the four statistically significant interactions with mode, viz. interactions with intervening elements between the matrix and complement, the absence of intervening elements between the matrix and complement clauses, subject coreferentiality, and the length of the complement clause subject. In 4.3, we finally offer the diachronic picture of the conditioning factors for zero use, i.e.

3. Mode is the distinction between written and spoken language. In the results of the statistical analysis, however, its label has been changed to TYPE. 4. The ANOVA type III tests revealed that the main effect of coreferentiality of tense (CC.T. co.ref) and the two interactions of mode with matrix-internal elements (mat.int:TYPE) and with the length of the complement clause subject (CC.length:TYPE) are “border-significant” (i.e. slightly higher than 0.05). This is reflected in Table 8 in the fact that some levels of these factors are not significant or are also border-significant. As the (stepwise) selection of the factors was based on Akaike’s Information Criterion and not on Likelihood-Ratio testing, the discussion of the results in Section 4 will nevertheless cover all effects. 5. The model was also checked for multicollinearity and although TYPE and period revealed variance inflation factors of 5.4 and 5.2, respectively, the rest of the predictors received scores of approximately 4 or under. Given the complexity of the model, we feel the assumption of orthogonality is met. See Speelman (this volume) for an explanation of the question of multicollinearity in logistic regression.

Estimate Error z value (Intercept) 7.14770 0.85237 8.386 mat.int: mat int -1.12456 0.14464 -7.775 CC.length: Pronoun -2.79835 0.86310 -3.242 CC.length: NP Short -3.99093 0.86946 -4.590 CC.length: NP Long -4.79996 0.95717 -5.015 interv: interv -2.55597 0.55110 -4.638 I.or.U: non -0.89387 0.09886 -9.042 Mode: written -2.26338 0.46837 -4.832 CC.Pol.co.ref: non -1.35818 0.39459 -3.442 CC.Pcoref: coref 1.08250 0.28079 3.855 CC.T.co.ref: non -0.15270 0.08981 -1.700 Period -0.54998 0.11807 -4.658 interv: interv * Mode: written -1.38356 0.30941 -4.472 mat.int: mat int * Mode: written -0.33652 0.18580 -1.811 Mode: written * CC.Pcoref: coref -1.38028 0.34021 -4.057 Mode: written * Period 0.31019 0.05057 6.134 CC.Pol.co.ref: non * Period 0.29591 0.06973 4.243 CC. length Pro * Period 0.33643 0.12078 2.786 CC. length NP short * Period 0.44723 0.12154 3.680 CC. length NP long * Period 0.48836 0.13219 3.694 interv: interv: * Period 0.17660 0.08285 2.132 CC.length: Pro * Mode: written 0.78566 0.37388 2.101 CC.length: NP Short * Mode: written 0.38367 0.36568 1.049 CC.length: NP Long * Mode: written 0.16318 0.40981 0.398 ------------------------------Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Coefficients:

Table 8.â•‡ Stepwise regression analysis; 9 main effects and 8 two-way interactions Pr(>|z|) < 2e-16 7.55e-15 0.001186 4.43e-06 5.31e-07 3.52e-06 < 2e-16 1.35e-06 0.000577 0.000116 0.089100 3.19e-06 7.77e-06 0.070107 4.97e-05 8.59e-10 2.20e-05 0.005343 0.000234 0.000220 0.033043 0.035610 0.294098 0.690489 *** *** ** *** *** *** *** *** *** *** . *** *** . *** *** *** ** *** *** * *


A diachronic corpus-based multivariate analysis 291 Table 8. (continued) Deviance Residuals:

Min -3.1176

1Q 0.2271

Median 0.3528

3Q 0.4883

Max 2.8074

Null deviance: 4557.8 on 5800 degrees of freedom Residual deviance: 3607.8 on 5777 degrees of freedom AIC: 3655.8 Model L.R.: 949.99 d.f.: 23

C: 0.803 R2: 0.278

Legend: The following coding abbreviations have been utilized in the regression analysis: mat.int = absence of matrix internal elements, CC.length = length of complement clause subject, interv = lack of intervening elements between matrix and complement clauses, I. or U. = Subject pronoun I or you, CC.Pol.co.ref = polarity between the complement and matrix clauses, and CC.Pcoref= coreferentiality of person between the complement and matrix clauses.


the interactions with period. It shows that there are significant changes across time in the extent to which mode, the absence of intervening elements between the matrix and complement clauses, length of the complement clause subject, and harmony of polarity predict the use of zero.

4.1

Main effects

The first step in our process involved determining whether the factors proposed in the literature indeed predict the zero complementizer form. The results from this initial procedure are presented in Figures 3–6. As it gives the overall picture of the diachronic change, we begin by examining the main effect of time or ‘period’ as a predictor of the zero form.

Main effect: Period

Figure 3a shows that there is a steady loss of the zero form from the earliest period of 1560 to the most recent period of 2012. The decrease is gradual but statistically significant. This finding is most noteworthy as it runs counter to previous claims that there is a diachronic increase in zero complementation.

Main effect: Mode

1.0

Predicted probability of SubC = ’zero’


The analysis of the effect of mode (i.e. spoken versus written language) upon the presence of the zero form reveals that the zero form occurs significantly more often in the spoken rather than the written data (Figure 3b). This is consistent with claims in the literature that zero complementation is especially common in spoken language. This confirms findings from previous studies (cf. Finegan and Biber 1985; Rissanen 1991).

0.8 0.6 0.4 0.2 0.0 1

2

3

4 5 Period

a. Period

6

7

8

1.0 0.8 0.6 0.4 0.2 0.0 Spoken

b. Mode (or Type)

Figure 3.â•‡ Main effects – structural factors predicting the zero form

Type

Written


Although the difference between the main effects of spoken versus written data is slight, it will become apparent in 4.2, where we discuss the interactions of the other factors with it, that the effects of conditioning factors may differ quite vastly depending on the mode.

Main effect: Absence of matrix internal elements

In Figure 4a, the analysis reveals that the absence of elements within the matrix clause is a good predictor for the zero form; when the matrix clause does not contain any material in addition to the subject and the verb, the probability of the zero form is significantly greater compared to matrix clauses that do contain extra material.

Main effect: Absence of intervening elements

In Figure 4b, we see that the absence of intervening material between the matrix and complement clause is an even stronger predictor for the zero form. When intervening material is present, the zero complementizer rate drops to just above 60%. Thus, when there are elements separating the matrix clause from the complement clause, there is almost a 40% chance that that will be realized. This adds support to the validity complexity principle proposed in the literature (cf. Rohdenburg 1996).

Main effect: Length of the complement clause subject

1.0



In Figure 5a, we turn our attention to the complement clause subject, examining the role that the weight of the subject (i.e. it, pronoun, NP-short and NP-long) plays in predicting the presence of the zero form. The results in this plot confirm the order that Torres Cacoullos and Walker (2009) previously arrived at; the subject pronoun it is more likely to occur with the zero form than other pronouns, which are still better

0.8 0.6 0.4 0.2 0.0 abs

mat.int

a. Matrix internal elements

Written

1.0 0.8 0.6 0.4 0.2 0.0 abs

interv

b. Intervening elements


interv

1.0




0.8 0.6 0.4 0.2 0.0 it

pro

np-short

np-long

1.0 0.8 0.6 0.4 0.2 0.0 I or U

a. Length of Complement

non

I.or.U

CC.length

b. Subject I or you


predictors of zero than short noun phrases and long noun phrases. This shows that the shorter and referentially lighter complement clause subjects are better predictors of the zero form.

Main effect: I or you

Turning our attention back to the main clause subject, we now examine the effect of the subject pronouns I or you as a predictor (Figure 5b). As with complement clause length, our results confirm previous literature (see Elsness 1984; Thompson and Mulac 1991; Tagliamonte and Smith 2005; Kearns 2007; Torres Cacoullos and Walker 2009) that the 1st and 2nd person is indeed a better predictor than other subject forms (i.e. 3rd person and plural forms) for the presence of the zero form.

Main effect: Cotemporality

In Figure 6a, we examine the effect of cotemporality, i.e. whether a construction in which the verbs of the matrix and complement clauses have the same tense is more likely to be used with the zero complementizer. In the plot above, there is not much difference between cotemporality and non-cotemporality, and Table 8 indeed revealed this effect to be border-significant. Nevertheless, the coefficient for non-cotemporality was negative, demonstrating that cotemporality tends to favour the zero form.

Main effect: Coreferentiality of person between matrix and complement clause

Figure 6b presents the effect of another type of ‘harmony’ between the matrix and complement clause, viz. coreferentiality between the respective subjects. Structures with coreferential subjects are slightly more likely to get a zero form than those with non-coreferential subjects.


0.6 0.4 0.2 0.0

a. Cotemporality

1.0

1.0

0.8

0.8

0.6 0.4 0.2 0.0

b. Coreferentiality of Person

Predicted probability of SubC=’zero’

0.8



1.0

0.6 0.4 0.2 0.0

c. Harmony of polarity


Main effect: Harmony of polarity between the matrix and complement clauses

The main effect of harmony of polarity between the matrix and complement clauses (Figure 6c) is that it is the disharmonious patterns that significantly predict the zero form. When there is harmony of polarity between the matrix and complement clauses, the zero form is slightly less likely to occur than with a mixed, i.e. disharmonious, pattern. This refines Torres Cacoullos and Walker’s (2009) finding that harmony of polarity is not a significant conditioning factor for the zero form. Now that we have completed our discussion of the main effects, we look at the interaction with the spoken or written mode.

4.2

Mode

In this section, we see that mode has an impact on the strength of other factors. Recall that although there was a significant difference in the main effect between spoken and written language, the difference was not that great. This section, however, reveals that some factors may be better predictors for the zero form in one mode as opposed to the other mode.

Mode: Absence of intervening elements

Figure 7a allows us to compare the conditioning effect of intervening elements between matrix clauses and the complement clause in the spoken and written modes. Recall that absence of intervening elements was a very good predictor overall. The interaction confirms this earlier finding; in both panels we observe a dramatic difference in complementizer use between the presence and absence of material. A notable difference, however, resides in the extent to which the presence of intervening material in the written mode predicts the zero form. When there is intervening material in the written mode, we are much less likely to get the zero form than in the spoken mode, so much so that the explicit complementizer that in fact becomes more likely; the zero rate drops to below 0.4. It may be that writers are more led by the complexity


abc TYPE : spoken

0.8 0.6 0.4 0.2

TYPE : spoken

1.0 Predicted probability of SubC=’zero’


1.0

abc

interv TYPE : written

mat int TYPE : written

0.8 0.6 0.4 0.2 0.0

0.0 abc

abc

interv

a. Intervening elements

mat int mat.int

interv

b. Internal elements

Figure 7.â•‡ Mode and structural factors predicting the zero form

principle than speakers and feel the need to insert that to make clause boundaries clearer when intervening material risks impairing clarity.

Mode: Absence of matrix internal elements

Figure 7b presents the results for the effect of elements between the matrix clause subject and verb in the spoken and written modes respectively. We see that the absence of matrix-internal elements is a good predictive factor for both spoken and written data. The steepness of the plot lines shows that the difference in predictive power between the presence and absence of matrix internal elements is comparable for both modes; however, in the written mode, less zero is used overall. This is indicated by the lower points for both the spoken and written levels.

Mode: Coreferentiality of person between the matrix and complement clause

The analysis of the effect of the coreferentiality of person between the matrix and complement clause subjects reveals an interesting difference with regard to mode (Figure 8a). Coreferentiality of person leads to higher levels of zero in the spoken data. In the written data, confidence intervals show that the difference between coreferentiality and non-coreferentiality is not significant.

Mode: Length of the complement clause subject

The final factor that we examine in this section is the effect that the length of the complement clause subject has as a predictor of the zero form relative to mode. As before, the plot in Figure 8b shows that within the spoken data the following cline exists: it > pro > np-short > np-long. A comparison between the two modes shows that short and long NPs tend more strongly towards that in the written mode than in the


non TYPE : spoken

it

0.8

0.6

0.4

0.2

0.0 non

coref

TYPE : spoken

1.0



1.0

coref TYPE : written

0.8

0.6

0.4

0.2

0.0 it

pro np-short np-long

CC.Pcoref

a. Coreferentiality of person

pro np-short np-long TYPE : written

CC. length

b. Length of the complement subject

Figure 8.â•‡ Mode and structural factors predicting the zero form

spoken mode. Overall, the length of complement clause subject has a stronger effect on written data than on spoken data. Again, the complexity principle, i.e. the need to mark off clause boundaries, may motivate writers’ choice of the that-complementizer as opposed to the zero form. In addition, the concern with clarity fostered by standardization and prescriptivism may also play a role. We now turn to the final stage of our analysis and look at the effect of the structural factors across the eight time periods (i.e. 1560–2012). Thus, in the following sections, we discuss the interactions with period which came out as significant.

4.3

Period

The interaction effects with period were significant with the following factors: mode, absence of intervening elements, complement clause length, and cotemporality of polarity between the matrix and complement clauses. This final step in the analysis offers a diachronic perspective; it shows whether the import of a given factor becomes stronger or weaker over time.

Period: Mode

Figure 9a shows the effect of mode over time. In the earliest periods, the zero form was far more prevalent in the spoken data relative to the written data but over time, as the zero form has gone down in the spoken mode and increased in the written mode, in PDE the two modes are at the same predictive level. As Figure 9a shows, the endpoints in PDE for both modes are almost identical which suggests that nowadays mode, in and of itself, is no longer a good or a significant predictor of the zero form with these verbs anymore.


1 2 TYPE : spoken

3 4 5 6

7 8

0.8

0.6

0.4

0.2

0.0

1 2

TYPE : written

1 2 3 4 5 6 7

8

interv : abs

1.0



1.0

7 8

0.8

0.6

0.4

0.2

0.0

1 2 3 4 5 6 7

Period

a. Mode (or Type)

3 4 5 6 interv : interv

8

Period

b. Intervening elements

Figure 9.â•‡ Period and structural factors predicting the zero form

Period: Absence of intervening elements

The analysis of the diachronic effect of the absence of intervening elements between the matrix and complement clauses (Figure 9b) gives a result which confirms what has been argued in the literature on that/zero variation, namely that the absence of intervening elements is a strong predictor of the zero form. The results show that this trend is decreasing over time; however, it still remains quite robust relative to the presence of intervening elements. The values in the right panel suggest that intervening elements predict the explicit that-complementizer throughout all periods, although the effect gets weaker, but this finding is somewhat less robust, as shown by the larger confidence intervals.

Period: Length of the complement clause subject

In Figure 10a, the analysis of the effect of the length of the complement clause subject over time shows a clear division between it and other pronouns versus NPs in that the former two have been and still remain the stronger predictors of the zero form while the latter (i.e. NPs) are actually increasing in their own respective predictive abilities of the zero form. Still, they have yet to reach the level of it or other pronouns. Furthermore, an examination of the start and endpoint for it and other pronouns shows that they are higher compared to NPs at any stage of their development and that it and other pronouns remain the stronger predictive factors in PDE.

Period: Harmony of polarity between the matrix and complement clauses

The final significant effect over time to be discussed in this section is the interaction between harmony of polarity and period. Figure 10b shows that when there is harmony of polarity, there is a distinct tendency towards more that over time; harmony



3 4 5 6

7 8

CC.length : pro

1 2

0.8 0.6 0.4 0.2 1.0

CC.length : np-short

CC.length : np-long

0.8 0.6 0.4 0.2 0.0

1 2 3 4 5 6 7

8

1.0

1.0

0.0


1 2 CC.length : it

CC.Pol.co.ref : coref

7 8

0.8

0.6

0.4

0.2

0.0

1 2 3 4 5 6 7

Period

a. Length of the complement subject

3 4 5 6

CC.Pol.co.ref : non

8 Period

b. Harmony of polarity

Figure 10.â•‡ Period and structural factors predicting the zero form

of polarity used to be a stronger predictor of the zero form than it is now. This trend is the opposite in the non-harmonious data; here, the level of zero use has actually increased over time.

5. Discussion This study has shown that, contrary to claims and speculation in the literature to the effect that there has been an overall diachronic tendency towards more zero complementizer use at the expense of that-complementation, the most frequent complement-taking mental verb in Present-Day English think, in fact, exhibits a diachronic decrease in zero complementation and a concomitant increase in that use. This trend can be observed both as a main effect when the data for this verb is aggregated and the interactions via mode (i.e. spoken and written data) are explored. The rigorous methodological approach developed and utilized in this study, and the attention given to ensuring sufficiently large and representative sample sizes from each period, has also highlighted the fundamental problems seen in previous work on this topic which have often relied heavily upon descriptive statistical processes. As evidenced by our initial presentation of findings in Section 3, a reliance on descriptive statistics (often presented in the literature in conjunction with a chi-square test) can unintentionally obscure important interactions between factors or variables and/or not reveal the stability or robustness of diachronic trend-lines or patterns. From a descriptive perspective, it would appear that the zero form for think is robust or is at least remaining consistent over time and thus one could reasonable infer that the factors which have been proposed to facilitate the zero form are either


equally predictive or also remain significant over time. It is only when a methodology such as the one used in this study is applied that the true role of the various factors becomes apparent along with the diachronic robustness of the predicted or expected trends and/or patterns vis-à-vis a dependent variable such as the presence of the zero complementizer. In addition to invalidating the long-standing assumption that this particular complement-taking verb has diachronically developed towards higher levels of zero complementation, this study also highlights the need to differentiate between individual factors when examining complementation patterns. It became apparent, firstly, that the extent to which the factors mentioned in the literature actually predict zero use may differ considerably in terms of predictive power, as was revealed by the analysis of interactions between various factors and mode (i.e. spoken versus written data) or period (i.e. time). The effect of intervening elements of material between the matrix and complement clause was a case in point. As we observed in Figure 4, a strong predictor overall, when intervening material was present the zero complementizer rate dropped to just above 60%. Its predictive power, however, is revealed to be much stronger in the written mode than in the spoken mode; when intervening material is present in the written mode, that is favoured. This effect was also shown to be quite robust and stable over time (i.e. period) with a slight decrease seen in PDE (Figure 9). The apparent significance, however, of the length of the complement clause subject was not limited to just one mode. This factor was significant in both the written and spoken data sets. Our results therefore suggest that the following cline may be present in terms of predicting the zero form: it > Pro > NP-short > NP-long zero ---------------------------------that This cline, however, when examined diachronically, also revealed a small but consistent decrease for both the it and pro subject forms over time towards PDE. This general decline was seen with a great majority of the factors tested in this study. Finally, the analysis of mode revealed that the zero form occurs overall significantly more often in the spoken rather than the written data. Yet, once again, by diachronically examining the effect of mode over time (i.e. period) we see that in PDE this previously significant finding now essentially disappears. In the most recent period, the trend lines indicate that both modes are equally predictive of the presence of the zero form and that mode has lost any real significance as a predictive factor. We believe that these results have shown the strength of the methodology and approach. Having established the structural factors that lie behind the choice of the complementizer for think and having demonstrated that this alternation is not merely a diachronic phenomenon, two important steps should now be taken. Firstly, the data set needs to be extended to include a larger set of mental state verb types. This may


reveal additional differences in the way that /zero alternation has evolved with similar verbs of cognition, as well as shedding more light on how the effect of a conditioning factor may differ from verb to verb. Secondly, the addition of a range of semantic and pragmatic features to the analysis may offer more direct insights into the differences between the two forms. Although the manual annotation involved may necessitate smaller samples, extending the analysis in this manner will further explain the role of the alternation.

6. Conclusion In this chapter we have tried to demonstrate the advantages of utilizing a rigorous empirically oriented framework and statistical analysis when exploring the diachronic development of clause combining and complement clause patterns with the verb think. Our approach has allowed us to test a range of motivating factors for actual diachronic significance, to calculate the cumulative effects of these factors/variables against each other and over time, and to determine what factors, if any, retain any predictive power for the presence of the zero complementizer form. In addition, our methodology has permitted us to identify the importance that the absence of intervening elements of material between the matrix and complement clauses, the length of the complement clause subject and the effects of mode and period play in predicting the presence of the zero form Finally, our use of a regression analysis has allowed us to assess the effect of factors, or lack thereof, across mode and over time. Our findings indicate that, contrary to expectations and predictions, there is a steady statistically significant loss of the zero form from the earliest period of 1560 to the most recent period of 2012. It is based on these findings and our overall results that we believe the approach demonstrated in this chapter has real potential for helping us to understand the role or roles that structural factors play in distinguishing near-synonymous forms in some aspects of diachronic language change.

References Aijmer, K. (1997). I think – an English modal particle. In T. Swan, & O. Jansen Westvik (Eds.), Modality in Germanic languages: Historical and comparative perspectives (pp. 1–47). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110889932.1 ANC = American National Corpus (2002). Linguistic Data Consortium, University of Pennsylvania. Bolinger, D. (1972). That’s that. (Janua linguarum. Series Minor, 155). The Hague: Mouton de Gruyter.


Bresnan, J., Cueni, A., Nikitina, T., & Baayen, H. (2007). Predicting the dative alternation. In G. Bouma, I. Kraemer, & J. Zwarts (Eds.), Cognitive foundations of interpretation (pp. 69– 94). Amsterdam: Royal Netherlands Academy of Science. Brinton, L. J., & Traugott, E. C. (2005). Lexicalization and language change. (Research Surveys in Linguistics.) Cambridge: Cambridge University Press. DOI: 10.1017/CBO9780511615962 BROWN = Francis, W. N., & Kucera, H. (1979). The Brown Corpus. Department of Linguistics, Brown University. Bybee, J. (2002). Cognitive processes in grammaticalization. In M. Tomasello (Ed.), The new psychology of language Vol. II (pp. 145–167). New Jersey: Lawrence Erlbaum CEECS I & II = Corpus of Early English Correspondence Sampler (CEECS) . CEMET = Corpus of Early Modern English texts (Extended version). See De Smet (2005). CLMETEV = Corpus of Late Modern English texts (Extended version). See De Smet (2005). COCA = Davies, M. (2008). The Corpus of Contemporary American English (COCA): 1990– present. Retrieved from . De Smet, H. (2005). A Corpus of Late Modern English. ICAME Journal, 29, 69–82. Divjak, D. (2010). Structuring the lexicon: A clustered model for near-synonymy. Berlin & New York: Mouton de Gruyter Elsness, J. (1984). That or zero: A look at the choice of the object clause connective in a corpus of American English. English Studies, 65, 519–533. DOI: 10.1080/00138388408598357 Finnegan, E., & Biber, D. (1995). That and zero complementizers in Late Modern English: Exploring Archer from 1650–1990. In B. Aarts, & C. Meyer (Eds.), The verb in contemporary English: Theory and description (pp. 241–257). Cambridge: Cambridge University Press. Glynn, D. (2010). Testing the hypothesis: Objectivity and verification in usage-based Cognitive Semantics. In D. Glynn, & K. Fischer (Eds.), Quantitative Cognitive Semantics: Corpus-driven approaches (pp. 239–270). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110226423 Gries, S. Th. (2013). Statistics for linguistics with R: A practical introduction. 2nd revised edition. Berlin & Boston: Mouton de Gruyter. DOI: 10.1515/9783110307474 Grondelaers S., Speelman, D., & Geeraerts, D. (2008). National variation in the use of er “there”. Regional and diachronic constraints on cognitive explanations. In G. Kristiansen, & R. Dirven (Eds.), Cognitive Sociolinguistics: Language variation, cultural models, social systems (pp. 153–204). Berlin & New York: Mouton de Gruyter. Grondelaers, St., Geeraerts, D., & Speelman, D. (2007). A case for cognitive corpus linguistics. In M. Gonzalez-Marquez, I. Mittelberg, S. Coulson, & M. Spivey (Eds.), Methods in Cognitive Linguistics (pp. 49–169). Amsterdam & Philadelphia: John Benjamins. Heylen, K. (2005). A quantitative corpus study of German word order variation. In S. Kepser, & M. Reis (Eds.), Linguistic evidence: Empirical, theoretical and computational perspectives (pp. 241–264). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110197549.241 Kaltenböck, G. (2004). That or no that – that is the question: On subordinator suppression in extraposed subject clauses. Vienna English Working Papers, 13, 49–68. Kearns, K. (2007a). Epistemic verbs and zero complementizer. English Language and Linguistics, 11, 475–505. DOI: 10.1017/S1360674307002353 Kearns, K. (2007b). Regional variation in the syntactic distribution of null finite complementizer. Language Variation and Change, 19, 295–336. DOI: 10.1017/S0954394507000117 LAMPETER = Lampeter Corpus of Early Modern English Tracts (1641–1732). Retrieved from .


Palander-Collin, M. (1999). Grammaticalization and social embedding: I THINK and METHINKS in Middle and Early Modern English. Helsinki: Tome LV. Quirk, R., Greenbaum, S., Leech, G., & Svartvik, J. (1972). A grammar of contemporary English. London: Longman. Rissanen, M. (1991). On the history of that/zero as clause object links in English. In K. Aijmer, & B. Altenberg (Eds.), English corpus linguistics: Studies in honor of Jan Svartvik (pp. 272– 289). London & New York: Longman. Rohdenburg, G. (1996). Cognitive complexity and increased grammatical explicitness in English. Cognitive Linguistics, 7, 149–182. DOI: 10.1515/cogl.1996.7.2.149 Speelman, D., & Geeraerts, D. (2010). Causes for causatives: The case of Dutch ‘doen’ and ‘laten’. In T. Sanders, & E. Sweetser (Eds.), Causal categories in discourse and cognition (pp. 173– 204). Berlin & New York: Mouton de Gruyter. Tagliamonte, S., & Smith, J. (2005). No momentary fancy! The zero ‘complementizer’ in English dialects. English Language and Linguistics, 9, 289–309. DOI: 10.1017/S1360674305001644 Thompson, S. A., & Mulac, A. (1991). A quantitative perspective on the grammaticalization of epistemic parentheticals in English. In E. Traugott, & B. Heine (Eds.), Approaches to grammaticalization (pp. 313–339). Amsterdam & Philadelphia: John Benjamins. TIME = Davies, M. (2007). TIME Magazine Corpus (1920s–2000s). Retrieved from . Torres Cacoullos, R. & Walker, J. A. (2009). On the persistence of grammar in discourse formulas: A variationist study of that. Linguistics, 47, 1–43. DOI: 10.1515/LING.2009.001 Traugott, E. C., & Dasher, R. B. (2002). Regularity in semantic change. Cambridge: Cambridge University Press. Traugott, E. C., & Konig, E. (1991). The semantics-pragmatics of grammaticalization revisited. In E. C. Traugott, & B. Heine (Eds.), Approaches to grammaticalization, Volume 1. Focus on theoretical and methodological issues (pp. 189–218). Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/tsl.19.1

Section 2

Statistical techniques

Techniques and tools Corpus methods and statistics for semantics Dylan Glynn


The use of corpora in semantic research is a rapidly developing method. However, the range of quantitative techniques employed in the field can make it difficult for the non-specialist to keep abreast with the methodological development. This chapter serves as an introduction to the use of corpus methods in Cognitive Semantic research and as an overview of the relevant statistical techniques and software needed for performing them. The discussion and description are intended for researches in semantics that are interested in adopting quantitative corpus-driven methods. The discussion argues that there are fundamentally two corpus-driven approaches to meaning, one based on observable formal patterns (collocation analysis) and another based on patterns of annotated usage-features of use (feature analysis). The discussion then introduces and explains each of the statistical techniques currently used in the field. Examples of the use of each technique are listed and a summary of the software packages available in R for performing the techniques is included. Keywords: collocation analysis, corpus linguistics, semantics, statistics, usage-feature analysis (behavioural profile)

1. Introduction This chapter offers an explanation of the corpus methods represented in the book and a brief overview of the various statistical techniques employed. It is designed as a resource for those less familiar with the field, but also as a reference for those already working with corpus-driven methods in Cognitive Semantics. Specifically, corpus-Â� driven Cognitive Semantics is understood as the work beginning with Dirven et al. (1982), Schmid (1993, 2000), Geeraerts et al. (1994, 1999) and Gries (1999, 2003), and currently represented in the edited volumes of Gries and Stefanowitsch (2006), Stefanowitsch and Gries (2006), Lewandowska-Tomaszczyk and Dziwirek (2009), Glynn and Fischer (2010), Geeraerts et al. (2010), Divjak and Gries (2012), Gries and

308 Dylan Glynn

Divjak (2012), Pütz et al. (2012), Reif et al. (2013), Glynn and Sjölin (2014), and in the monographs Hilpert (2008, 2012), Divjak (2010a), Gilquin (2010), Dziwirek and Lewandowska-Tomaszczyk (2011), Hoffmann (2011), and Glynn (forthc.). In this chapter, corpus-driven Cognitive Semantics is argued to divide into two methodologies, or analytical approaches, based either on the formal analysis of collocations or the semantic analysis of features. This proposed distinction is described in Section 2. Following this, Section 3 describes the quantitative techniques used in such research. It lists and explains the techniques and offers examples of how they are used, giving detailed references on where the application of each technique is explained in the literature.

2. Collocations and features: Two approaches to corpora A common misconception amongst cognitive linguists is that corpus-driven research, and indeed, the quantitative analysis of corpus data, does not involve any close analysis of actual examples. This is not necessarily true at all. Within Cognitive and Functional Linguistics, broadly speaking, there is a wide range of approaches to corpus data, from simply counting the number of occurrences of a given form in a given context to the development of complex computational models trained on enormous text banks. For corpus-driven research in semantics, where the ‘meaning’ of a given linguistic form is in question, it is possible to broadly identify at least two approaches. All the studies in the first section of this book fall into one of these two categories. The first of these is based on formal, and therefore, observable, patterns. We can term this approach ‘collocation analysis’. Secondly, the corpus analysis can be based on patterns of annotated features, which we term ‘feature analysis’. In the former, the analysis seeks to identify formal patterns so as to interpret them as indices of meaning structure and in the latter, the analysis seeks to directly identify semantico-pragmatic patterns through close manual annotation. Although the approaches can be combined (cf. Stefanowitsch and Gries 2008), they tend to be used separately and possess distinct strengths and weaknesses. The first ‘type’ of corpus-driven research, collocation analysis, is more established and is typical of mainstream Corpus Linguistics. Collocation studies identify the co-occurrence of linguistic forms in a given sample of naturally occurring language. Firth’s (1957:â•›179) now famous phrase, “you shall know a word by the company it keeps”, is a succinct way of capturing the aim of this approach. When extended to other parts of language, such as syntactic patterns or indeed text types and genres, the large-scale study of collocation is a powerful tool for making generalisations about language use. Cognitive and Functional Linguistics are particularly concerned with why a given form is used and so it follows that in order to answer research questions

Techniques and tools 309

of this nature, inferences as to the semantic, functional, or conceptual motivation for the collocation must be made in post hoc interpretation. Despite this subjective step in the use of collocation analysis in Cognitive-Functional Linguistic research, the analytical approach has important advantages. To the extent that one can retrieve forms automatically, one can consider extremely large samples, making studies (relatively) representative of a given language or part of language. Secondly, forms are objectively identifiable, making this step largely independent of subjective analysis. However, this statement warrants qualification. Even if a form is objectively identifiable, linguists are typically interested in only certain uses of a given form and, often, these specific uses cannot be retrieved automatically. In such situations, the decision as to which occurrences are representative of the category is typically a question for debate (cf. Perek, this volume, 61–86). Moreover, collocation studies rely on some measurement of association. Raw frequency of co-occurrence can be misleading because if one of the forms is extremely frequent, then relatively high co-occurrence may just be a result of the overall high frequency of that form. The problem of how to determine the degree of association, or ‘attraction’, is fundamental. Common ways of measuring the degree of association for lexical co-occurrence are the mutual information (MI) score, the z-score (standard score), the t-score and the log-likelihood. Many Corpus Linguistics programs, both on-line and stand-alone, automatically generate some of these scores. Collostructional analysis is one alternative to such measures. Developed by Stefanowitsch and Gries (2003, 2005) and Gries and Stefanowitsch (2004a, 2004b) and described in Hilpert (this volume, 391–404), it is a suite of methods that use the Chi-squared or Fisher exact test to compute degree of association. These techniques allow the researcher to consider the co-occurrence, not just of lexemes, but also of syntactic patterns. Collostructional analysis has proven popular in Cognitive Linguistics. One of the newest advances in the use of collocation is the application of Word Space modelling to semantic research questions within computational linguistics. The principle is to extend the analysis of collocation beyond one or two words or even syntactic patterns, to whole lines, paragraphs and even entire texts. Such approaches give rich collocation-based behavioural profiles of a given linguistic form. The implications for such analytical techniques in semantics are only now being realised. This methodology is not represented in the volume. Peirsman et al. (2010) and Sagi et al. (2001) are examples of the application of these methods to research in semantic relations. The number of studies employing a collocation approach, even restricted to Cognitive Linguistics, is enormous. A small sample of recent studies includes Newman and Rice (2004, 2006), Deignan (2005), Delorge (2009), Pęzik (2009), Van Bogaert (2010), Colleman (2010) and Zeschel (2010). Applications of collostructional analysis include Wulff (2006), Wulff et al. (2007), Hilpert (2008, 2009) and Gilquin (2010). In general terms, it is possible to identify a second quantitative approach in corpus-driven Cognitive Semantics, one that focuses on the manual analysis of

310 Dylan Glynn

usage-features. Although less traditional in the mainstream of Corpus Linguistics, the general principle has a long tradition in Cognitive Linguistics (Dirven et al. 1982; Rudzka-Ostyn 1989, 1995; Fillmore and Atkins 1992; Geeraerts et al. 1994) and, more recently, is gaining currency in Functional Linguistics (Fischer 2000; Scheibman 2002; Kärkkäinen 2003; Pichler 2013). The principle of combining the results of this usage-feature analysis with multivariate statistics begins with Geeraerts et al. (1999) and Gries (2003). It is termed the behavioural-profile approach by Gries and Divjak (2009) and Divjak and Gries (2009) and multifactorial usage-feature analysis by Glynn (2009, 2010b).1 The principle is simple: for a large sample of a given linguistic phenomenon, various formal, semantic, and/or social ‘linguistic features’ (or ‘ID tags’ in the terminology of Gries and Divjak 2009) are identified and ascribed to each occurrence. It is worth noting that the method per se has also been independently developed in social psychology and computational linguistics. In the former, it is termed the analysis of components (cf. Scherer 2005; Fontaine et al. 2013) and in the latter, sentiment analysis (Wiebe et al. 2005; Verdonik et al. 2007; Daille et al. 2011; Balahur and Montoyo 2012; Read and Carroll 2012; Taboada and Carretero 2012). The approach consists of the repeated application of what is essentially a ‘traditional’ linguistic analysis to hundreds, or even thousands, of naturally occurring examples. This procedure results in a quantified usage-profile of the linguistic phenomenon in question. Usage-feature analysis is employed, with varying degrees of statistical sophistication, to examine phenomena of all kinds, from syntactic variation and semantics (Heylen 2005a; Bresnan et al. 2007; Speelman et al. 2009), to discourse studies and conversation analysis (Scheibman 2002; Kärkkäinen 2003; Flores Salgado 2011; De Cock 2014a, 2014b), and even gesture research (Zlatev and Andrén 2009; Morgenstern et al. 2011). The limitations of the approach are twofold. Firstly, the detailed manual analysis is as subjective as any traditional linguistic analysis and is open to the same vagaries, theoretical biases and human error. Secondly, the manual analysis, or annotation, of examples is meticulous and laborious. This, combined with the simple practical reality of limited resources, means that samples are relatively small. The resulting sample

1. Since Multifactorial Usage-Feature Analysis (Behavioural Profile Approach) is less known in corpus circles, a selection of current examples of its use include: Geeraerts et al. (1999), Gries (1999, 2003, 2006, 2010), Szmrecsanyi (2003, 2010), Wulff (2003, 2009), Heylen (2005a), Divjak (2006, 2010a, 2010b), Divjak and Gries (2006, 2009), Bresnan et al. (2007), Grondelaers et al. (2007, 2008), Glynn (2009, 2010a, 2010b, 2014a, 2014b, forthc.), Janda and Solovyev (2009), Speelman et al. (2009), Speelman and Geeraerts (2010), Krawczak and Glynn (2011, in press), Krawczak and Kokorniak (2012), Levshina (2012), Levshina et al. (2013a, 2013b), Krawczak (2014a, 2014b), and Deshors (2014). Doctoral dissertations focusing on developing the method include Gries (2000), Grondelaers (2000), Heylen (2005a), Glynn (2007), Arppe (2008), Robinson (2010b), Deshors (2011), Levshina (2011), Barnabé (2012), Klavan (2012), and Diehl (2014).


Table 1.â•‡ Observational differences in collocation and feature analysis of corpora Stage 1: Analysis of data Stage 2: Interpretation of analysis

Collocation

Feature

objective subjective

subjective objective

size makes it more difficult to be sure of representativity and harder to obtain statistically significant results. The advantages of the approach are also twofold. Firstly, the method allows the operationalisation and quantification of traditional linguistic analyses. This is no trivial matter because it permits hypothesis testing and produces falsifiable results for research questions not easily approached using traditional corpus methods (c.f. Geeraerts 2010; Glynn 2010b; Stefanowitsch 2010). Secondly, an important strength lies in the possibility of treating the results obtained through the usage-feature analysis with multivariate statistics. This is especially important for non-modular theories of linguistics, such as Cognitive Linguistics, because multivariate statistics permits an analysis to handle the complexity of the interaction of the different dimensions of language structure simultaneously (such as lexis, syntax, phonology, society, etc.), creating a multidimensional and socio-conceptually realistic profile of the use of a linguistic form or the role of a linguistic function. Geeraerts (2011) compared the two corpus approaches, underlining that both are subjective, but are at different stages in their application. Table 1 summarises Geeraerts’ point about subjectivity. Juxtaposing the two analytical approaches like this is, of course, a simplification. At the first stage of analysis, collocation studies are often not entirely objective because of questions such as what constitutes a ‘form’. Firstly, forms are polysemous and only certain uses may be relevant for a given study. In such a situation, manual selection is often the only solution. Secondly, the forms themselves are typically composite and so formal variation itself can cause category issues. In other words, is a given formal variant an example of the form in question or is it a ‘different’ form? Again, in such situations, subjective categorisation enters the analysis. Turning to feature analysis, the subjective first step is not always particularly subjective. Often, feature analysis is largely based on observable phenomena. For example, grammatical features can be crucial to usage-feature analysis and are annotated automatically, or if done manually, are done so objectively. At the stage of interpretation the same objective-subjective blurring occurs. For collocation analysis, as Desagulier (this volume, 145–178) shows, statistical analysis can help add a degree of objectivity to interpreting the collocation patterns observed. A similar caveat is needed for the usage-feature method. Although multivariate statistics may help us to objectively distinguish semantico-pragmatic patterns from non-patterns, we still must decide if those patterns answer the research question at hand, which is an inherently subjective step.

312 Dylan Glynn

3. Statistical techniques and tools Often one of the most confusing issues in the application of quantitative techniques to linguistic research is the myriad of different techniques available. This section is primarily intended for the reader who has some experience with quantitative methods, presenting an overview of the techniques relevant to corpus linguistic research. For the reader who has little experience in quantitative techniques, the overview will be technical, but it is hoped, still informative. It is important to understand that statistics is a rapidly growing science with constant new advances as well as many uncertainties and conflicts. Perhaps more importantly, we must also remember that statistical techniques are only analytical tools. No statistical technique will identify a linguistic fact or explain any linguistic structure. Nevertheless, statistical tools can be used by linguists to help look for language structure – assuming one knows where to look. They can also be used to confirm the probability that the results of an analysis are not a chance occurrence. Statistics can help linguists struggle with what they have been doing for centuries, describe and explain language, but they are only tools in that endeavour. Just as there is sometimes a misconception that statistics can answer linguistic questions, there exists a misconception that quantitative corpus-driven research is devoid of ‘real’ linguistic analysis. Nothing is further from the truth. Corpus-driven linguists deal with real language and in large quantities. The ‘numbers’ presented in corpus-driven research are not the analysis; they are a quantitative summary of the analysis, which must, in turn, be interpreted. Corpus-driven linguists, for the most part, deal with language in a relatively close and fine-grained way; they just deal with large quantities of it. One of the aims of this book is to showcase and explain the use of a small set of statistical techniques that can be helpful for traditionally trained linguists in their research. The aim is not to teach statistics or the computer programs for performing statistical analyses, but simply to introduce some of the possibilities. In this section, we begin with a short description of the computer applications available for performing statistics, and then briefly consider a fundamental theoretical question for the statistical sciences – type of data. This question is essential to understand before one can decide which statistical techniques are appropriate in a given situation. This is followed by a systematic summary of the techniques currently used in the field, examples of their use, as well as examples of texts that explain how they are used. The description ends with a detailed list of the different commands and packages for performing these statistical techniques in the programming suite R.

Statistical software

There are many computer applications, commercial and otherwise, that enable the researcher to perform statistical analysis. In this volume, the statistical program that


is used by most authors is R. This program is, in fact, a powerful programming suite with enormous potential. The explanatory chapters all use R and the reader is taken step-by-step through the necessary “code”, or command lines, needed to perform the analyses. No attempt is made to demonstrate the full functionality of the program, merely to offer a working knowledge of how to perform specific analyses. This volume focuses upon R for three reasons. Firstly, it is a free and cross-platform program. Secondly, since it is open source, as soon as new statistical techniques develop, new software modules are written and uploaded for the public. Thirdly, the programme is one of the two most commonly used programs for statistics in the social sciences (there are, of course, many more, especially devoted to specific techniques). The other most frequently used program in the social sciences is SPSS. Like R, it is also an extremely powerful tool, as widely used, but also includes a graphic user interface (unlike R). Since R is equally powerful, arguably more up to date, entirely free and used by the majority of authors in the book, the only negative is its command-line interface. However, in the following chapters, the command-line is given simple step-by-step instructions and, it is hoped, will not pose too many problems for the beginner. It is true that the command-line may seem daunting at first, but if the steps are followed line-by-line, the only difference with ‘button-for-button’ (as in a graphic interface application) is one of familiarity. Other important application suites include SAS, Statistica, and Stata, which are all powerful and versatile. SAS is command-line, like R and in some ways, R can be seen as the open-source version of SAS. It is arguably the most complete statistical programming suite, but is rarely used in the social sciences. Statistica and Stata are comparable to SPSS. They too have graphic user interfaces, are relatively user friendly and, just like SPSS, are costly. Statistica is restricted to the Windows operating systems, but has a relatively large and helpful online community. Stata is cross-platform, but is probably less common than Statistica. It is not really possible to say which suite is the best, since certain techniques are extremely well covered in one suite and not the other. Due to its being open source, R is surely the suite with the most options and also the quickest to respond to developments within the domain of statistics, but, of course, that does not mean its implementation of those techniques is the best. If the reader is familiar with any of these other programs, the descriptions of the statistical techniques in the book, as well as their interpretation and application, will still be useful. Lastly, it should be noted that a graphic user interface is under development for R. This is not drawn upon because its development is not yet complete and the commands/R sessions described in this book are sufficiently straightforward that readers who are not familiar with statistics or command-line will not have problems following.

Types of data

Before choosing a statistical technique, one must first know what ‘type’ of data one is dealing with. This is because different types of data require different statistical

314 Dylan Glynn

techniques. The most basic distinction is between what is called continuous data and categorical data. The former typically come from measurements and therefore make a continuum, for example 1.0, 1.1, 1.2 … 1.8, 1.9, 2.0. This kind of data is probably the most common and comes from diverse sources such as age, time, height, dosage, temperature, response times, and, arguably, grammatical judgements. Continuous data are typical in psychology and psycholinguistics. The second kind of data is categorical, also called ‘discrete data’, ‘tabular data’ or ‘count data’. It is this kind of data, as corpus linguists, with which we are most often concerned. Such data include, for example, the frequency of occurrence of a linguistic form, the number of times it occurs in a given tense, or in a given register. In these examples, the data are said to be nominal because each of the occurrences is independent from the other. However, categorical data can also be ordered. This is the case when, for example, the categories follow a natural sequence or ranking, such as young, middle-aged, and old or when a sentence is short, medium or long in length. Ordered categorical data share properties of both nominal categorical and continuous data. Grammatical judgements, on a scale of 1 to 7, for example, could be argued to be continuous or ordered categorical. Technically, it is ordered because a respondent cannot enter 3.5, for example, but is forced to make a discrete choice upon what is, in reality, a continuous scale of acceptability. However, if we assume that no respondent would perceive differences to the degree of 3.5, then we can treat the scale as a true measurement, and therefore, continuous. Table 2.â•‡ Types of data in statistics Data type

Example of data

Continuous 1, 1.1, 1.2 … 1.8, 1.9, 2 1, 2, 3, 4, 5, 6, 7 Ordered short, medium, long cold, warm, hot Nominal apples, peaches, pears y’all, you lot, youse

Description

Example of use

Sequential (ordered) but non-discrete / continuous Sequential (ordered) but discrete / non-continuous Independent and discrete categories

Response times in Psycholinguistics Different periods in diachronic linguistics Different lexemes in Corpus Linguistics

Although there is occasionally debate on the issue, most statistical techniques are designed for one of the kinds of data. For example, least squares estimation and linear regression are used for continuous data and maximum likelihood and logistic regression for categorical, just as principle components analysis is used for continuous data and correspondence analysis for categorical. Table 2 summarises the differences.

Statistical techniques for corpus linguistics

Statistics is an immense science – there are countless tests and corrections for those tests. There are even more exploratory techniques with various algorithms that each technique can employ and different ways for representing results of those exploratory techniques. Confirmatory analysis has again as many different techniques, but this


time, seemingly endless sets of diagnostics to check the validity of the results. It must be stressed that the techniques presented here only scratch the surface of what is possible, but also of what problems exist. We begin with significance tests and association measures. Although not statistical techniques per se, they are tools that are important to the field. We then cover exploratory methods, the results of which cannot be used to make claims about structure beyond the sample. In other words, what is found with these techniques may be restricted to the corpus or the extract of the corpus being examined. These exploratory techniques do not test hypotheses or make predictions about the population (real language). The description then turns to confirmatory techniques, which are more complex in their application but which make predictive claims and can test hypotheses in terms of statistical significance or the probability that observed structures exist in real language beyond the sample.

Sample, significance and independence

Establishing that the occurrence of something in a given sample is more or less common than would be expected by chance or that two sets of data are more different than would be expected by chance are basic steps in inductive research. Pearson’s Chisquared test and Fisher’s Exact test are omnipresent in research based on samples of categorical data. Gries (this volume) explains these tests and shows how to apply them in R. Other tests useful for corpus data include the exact binomial test, McNemar’s paired Chi-squared test, and the proportions test. These are used for investigating relations in frequency tables. An excellent explanation of these tests and their commands in R can be found in Dalgaard (2008: Ch. 8) and Baayen (2008:â•›Section 4.1.1). See also Gries (2009b:â•›125–127, 158–176; 2013:â•›165–172), Everitt and Hothorn (2010: Ch. 3), and Adler (2010:â•›360–367).

Collocation and association measures

Within Cognitive Linguistics, collostructional analysis has proven to be one of the most important methods for investigating collocations. Developed by Stefanowitsch and Gries (2003, 2005) and Gries and Stefanowitsch (2004a, 2004b), the principle can be combined with a range of association measures for determining the degree of collocational ‘strength’ (the measure is typically calculated with a p-value obtained from a Fisher exact test, log-transformed). These calculations are not yet implemented in most corpus annotation or concordance software. However, Stefan Gries has developed R scripts (semi-automated sets of commands) for performing the tests.2 Hilpert (this volume, 391–404) explains three varieties of collostructional analysis: collexeme analysis, distinctive collexeme analysis, and covarying collexeme analysis. 2. For more information, contact Stefan Th. Gries. His contact details can be found on his website: http://www.linguistics.ucsb.edu/faculty/stgries/.

316 Dylan Glynn

Examples of use include Wulff (2003), Hilpert (2008), Stefanowitsch and Gries (2008), Colleman (2009), and Gilquin (2010). The aim of quantifying degree of association between two forms in terms of frequency is not unique to the collostructional suite. Corpus Linguistics has developed an array of calculations to determine relative degree of association, especially between individual words. The most common are the mutual information (MI), the z-score, the t-score, and the log-likelihood. There is important variation in the results obtained from using any one test over another. Evert (2009) offers a detailed discussion on the matter; see also Wiechmann (2008), Wulff (2010) and Desagulier (this volume, 145–178). The z-score and the t-score are both explained with the R-code in Johnson (2008: Ch. 3) and Dalgaard (2008: Ch. 5). The freely available Ngram Statistics Package extracts sequences from a corpus and calculates a range of association measures. All these scores are used extensively in collocation-based corpus linguistics.

Cluster analysis

Cluster analysis is a diverse family of techniques, which, as the name suggests, cluster data. K-means clustering is used when one knows how many clusters there should be in advance; the technique ‘sorts’ the data accordingly. More common in semantic research is hierarchical clustering, which is used as an exploratory technique for the identification of clusters in the data. Importantly, by identifying clusters, it also sorts the data into the clusters it has ‘discovered’. The technique begins with a set of features and then uses them to group the features of a given variable (for instance, a list of senses, concepts, words, or constructions). It represents the results in a dendrogram, a kind of plot that depicts groups in an intuitively transparent way as dependencies clustered in branches. Cluster analysis is an excellent technique for determining which forms are similar to each other and which are different. It is explained by Divjak and Fieller (this volume, 405–442). Other explanations using R code include Crawley (2007:â•›742–744), Baayen (2008:â•›138–148), Johnson (2008: Ch. 6), and Ledolter (2013: Ch. 15). Härdle and Simar (2007: Ch. 11), Izenman (2008: Ch. 12), Drenan (2009: Ch. 25), Everitt and Hothorn (2010:â•›18), Afifi et al. (2011: Ch. 16) and Marden (2011: Ch. 12) represent detailed, yet approachable, explanations without R code. Everitt et al. (2011) is surely the most comprehensive work devoted to the technique, and although quite technical, is a systematic and excellent reference for using cluster analysis. The book provides no explanations for performing the analysis, but does give information on which software packages are available for many of the analyses it describes. Examples of use in Cognitive Linguistics include Schulze (1991), Chaffin (1992), Myers (1994), Sandra and Rice (1995), Ravid and Hanauer (1998), Rice et al. (1999), Gries (2006), Divjak (2006, 2010а), Divjak and Gries (2006), Gries and Hilpert (2008), Valenzuela Manzanares and Rojo López (2008), Janda and Solovyev (2009), Louwerse and Van Peer (2009), Robinson (this volume, 87–116), Glynn (2010a, 2014a, 2014b,


this volume, 117–144), Levshina (2012), Szmrecsanyi (2013), and Krawczak and Glynn (in press).

Correspondence analysis

Correspondence analysis is an exploratory technique that helps identify associations in the data, such as patterns in the combinations of linguistic features. The technique is designed for dealing with complex interactions where it is not known a priori which dimension, be that syntax, semantics, pragmatic, or social context, that structures the behaviour of the data. For instance, it can help find which semantic features typically occur with a set of grammatical forms or constructions, but also how these two dimensions interact relative to social variation. It visualises these associations in biplots, which, although arguably difficult to interpret, represent rich depictions of complex structures. Glynn (this volume, 443–486) explains the application and interpretation of two varieties of correspondence analysis: binary correspondence and multiple correspondence analysis. There exist several comprehensive books devoted to the technique: Benzécri (1980, 1992), Murtagh (2005), Greenacre (2007 [1993], 2010), and Le Roux and Rouanet (2010). Amongst these, Greenacre (2007) is probably the standard book of reference. Useful introductions include Le Roux and Rouanet (2004: Chs. 2 and 5), Everitt (2005: Ch. 5), Härdle and Simar (2007: Ch. 13), Baayen (2008: Ch. 5), Izenman (2008: Ch. 17), and Husson et al. (2011: Chs. 2 and 3). The last of these, Husson et al. (2011), is particularly clear and includes some of the most recent developments. Examples of use include Arppe (2006), Glynn (2007, 2009, 2010a, 2010b, 2014a, 2014b, this volume, 117–144), Szelid and Geeraerts (2008), Plevoets et al. (2008), Glynn and Sjölin (2011), Krawczak and Glynn (2011), Barnabé (2012), Krawczak and Kokorniak (2012), Nordmark and Glynn (2013), Levshina et al. (2013b), Desagulier (this volume, 145–178; in press), Delorge et al. (this volume, 39–60), Fabiszak et al. (this volume, 223–252), and Krawczak (2014a, 2014b; in press).

Multidimensional scaling

This technique is similar to correspondence analysis in its functionality and output. It identifies correlations between levels (features) in frequency tables. Explanation in R can be found in Rencher (2002: Ch. 15, Section 1), Everitt (2005: Ch. 5), Baayen (2008:â•›136–138), Drenan (2010: Ch. 23), Maindonald and Braun (2010 [2003]:â•›383– 384), and Everitt and Hothorn (2009: Ch. 17; 2011:â•›121–127). A new volume, which is one of the most comprehesive applied works on the technique to date and one that includes explanation in R, is Borg et al. (2013). Adler (2010:â•›525, 541ff., 564) lists the wide range of functions in R for applying multidimensional scaling, but without examples of use. Härdle and Simar (2007: Ch. 15) and Izenman (2008:â•›13) offer more detailed explanations of how the technique functions. See Le Roux and Rouanet (2004:â•›12– 14) and Cadoret et al. (2011) for comparison between multidimensional scaling and

318 Dylan Glynn

correspondence analysis. Borg and Groenen (2005) is a complete description, containing both mathematical theory and details of application and interpretation. Cox and Cox (2001) is equally detailed, though more concerned with mathematical theory. Nevertheless, the work includes helpful chapters on biplots and correspondence analysis. Examples of its use within the field include Bybee and Eddington (2006), Clancy (2006), Croft and Poole (2008), Szmrecsanyi (2010), Hilpert (2012), Heylen and Ruette (2013), and Ruette et al. (in press, forthc.). Although not a corpus study, Berthele (2010) is another recent example.

Configural frequency analysis

This is a simple and powerful technique, yet surprisingly uncommon outside the German linguistic tradition. It can be seen as a simplified log-linear analysis (see below) or as multiple Chi-squared tests; indeed, it functions by creating log-linear combinations of factors to predict cell frequencies typically based on Chi-squared tests. The technique offers possibilities for significance testing in multivariate models where no clear response variable exists, by identifying which correlations in a multiway frequency table are significant. The main limitation for the application of this technique is sample size. For a given analysis, all cells must have at least one occurrence and a minimum of 20% should have more than 5 occurrences. An excellent explanation, though with no R code, can be found in Tabachnick and Fidell (2007: Ch. 16). Gries (2009b:â•›240–252) offers a clear explanation of how to implement it, but note that this is omitted from the newest version of his book (Gries 2003). Von Eye (2002) is a textbook devoted to the subject and von Eye et al. (2010) represents the state-of-the-art. Hierarchical configuration frequency analysis has been used by Stefanowitsch and Gries (2005, 2008), Wulff et al. (2007), Hilpert (2009, 2012), Jing-Schmidt and Gries (2009), Schmidtke-Â� Bode (2009), Berez and Gries (2010), Hoffmann (2011), and Kööts et al. (2012).

Linear discriminant analysis

Discriminant analysis is a classification technique that functions in a similar way to logistic regression and classification tree analysis (see below). However, linear discriminant analysis requires normally distributed data and continuous predictor variables, two conditions that are rarely met in Corpus Linguistics.3 Venables and Ripley (2002:â•›331–338), Crawley (2007:â•›744–747), Baayen (2008: 154–160), Adler (2010:â•›440–444) and Maindonald and Braun (2010:â•›385–391) offer explanations appropriate for the intermediate user. Everitt (2005:â•›Ch 7), Härdle and Simar (2007: Ch. 12), Tabachnick and Fidell (2007: Ch. 9), Izenman (2008: Ch. 8) and Afifi et al. (2011: Ch. 11) offer more substantial descriptions of discriminant analysis, 3. Cf. Stevens (2001), Arppe (2008:â•›164), Baayen (2008:â•›154), Heylen et al. (2008), and Hoffman (2011:â•›95) for discussion on the problems associated with the implementation of discriminant analysis. See also Divjak (2010a:â•›138) who defends its use.


but offer no explanation for performing the analysis in R. Given the criteria are met, the method is a powerful classification technique and has been used by Gries (2003), Wulff (2003), and Divjak (2010a) in the field.

Classification tree analysis

An alternative to linear discriminant analysis is a data mining technique designed for categorical data called classification tree analysis. It is closely related to another technique termed regression tree analysis, which is used for continuous data. Together they are referred to as CART (or classification and regression tree analysis). The classification tree analysis technique employs an algorithm called recursive partitioning. For a given binary response variable (a vs. b), the algorithm begins with this alternation and asks which of the predictors (the other variables in the model) is best at predicting the choice between the two alternatives in the response variable. The algorithm continues this process for each of the two branches until all the predictor variables are ‘used up’. This re-occurring branching gives us a ‘tree’ that shows how the different variables predict the outcome, a vs. b. Classification tree analysis is explained and presented with R code in Crawley (2007: Ch. 21), Baayen (2008:â•›148–154), and Adler (2010:â•›406–117, 446–452). Other substantial descriptions include Venables and Ripley (2002: Ch. 9), Everitt and Hothorn (2010: Ch. 9), Maindonald and Braun (2010: Ch. 11), and Marden (2011: Ch. 11). The method has enjoyed some popularity in Cognitive Linguistic research, being both straightforward to apply and to interpret. Within the field, examples of its use include Klavan et al. (2011), Robinson (2012; this volume, 87–116), and Levshina et al. (this volume, 205–222). Bootstrapping regression trees and, what is termed, the random forests technique, represent an important avenue for the development of these techniques. Bootstrapping is a widely used technique that randomises the data in order to test explanatory strength and, thus, to ascertain confidence scores for the observed data through comparison with the randomised version of the data. The application of such techniques to classification tree analysis is opening up a new set of statistical alternatives to logistic regression analysis (see below). See Everitt and Hothorn (2010:â•›170–173), Strobl et al. (2009a), Adler (2010:â•›414–417), and Maindonald and Braun (2010:â•›369–372) for a description. Such techniques have yet to be applied in the field.

Regression analysis

In its various forms, regression analysis is one of the most widely used and powerful techniques in statistics. The importance of regression techniques lies in their ability to ‘predict outcomes’. The outcome is the term used to refer to a linguistic choice or a linguistic variant. This can be any kind of linguistic phenomenon, from lexemes, gestures, grammatical constructions and phonological patterns to the meanings of words, pragmatic functions, even gender, period, sociolect or dialect. The principle

320 Dylan Glynn

of how a regression analysis works is simple. The regression analysis takes our linguistic analysis of the data and builds a model that attempts to predict the behaviour of whatever phenomenon we are interested in explaining. If the model can predict which linguistic phenomenon (choice or variant, for example) is used, based on the linguistic analysis, then we can say that the analysis is accurate and, at least adequate, in distinguishing the phenomena under consideration. The linguistic choice or variety is understood as the response variable, which is ‘predicted’ by the independent variables, or the factors and features of the linguistic analysis. The model provides a great deal of information about how the linguistic analysis predicts the behaviour of the response variable but three pieces of information are crucial. Firstly, it tells us which of the linguistic factors and features are statistically significant in predicting the outcome. Secondly, it tells us the effect size of those features and factors; in other words, the relative importance of that factor or feature in predicting the outcome. Lastly, it tells us how accurately a combination of all the significant factors and features distinguish between the linguistic phenomena (the forms, uses or varieties being investigated). The following sections summarise several types of regression that are designed for categorical outcomes. This family of regression techniques are typically referred to as logistic regression. The standard references for logistic regression modelling include Agresti (2013 [1990, 2002], also 2007) and Hosmer and Lemeshow (2013 [1989, 2000]). Harrel (2001, also 2012) and Faraway (2006, also 2002) are also widely used reference books for the technique. Two other useful references include Hilbe (2009) and Menard (2010, also 2002). Once the basics have been mastered, and perhaps even before then, these books should be consulted. Especially useful is Thompson (2009), an unpublished and freely downloadable book that accompanies, step-by-step, Agresti’s work, with the R code needed to perform most of what his books cover. A note of caution is needed for the reader with little experience in statistics. None of the aforementioned books are designed for novice users, but they need to be consulted before regression analysis is used in research. Actually performing regression analysis is not particularly difficult. The complexity of confirmatory modelling lies not in applying the techniques (fitting the models), but in knowing which of the many algorithms and options one should use for the data and also applying and understanding the diagnostics of the model. Since confirmatory modelling tests hypotheses, one runs the risk of what is termed a Type I Error. This is statistics parlance, more or less, for demonstrating something to be true, when it is not. Before one reports findings obtained with regression modelling, one should always have the results thoroughly checked by a statistician.


Binary logistic regression

Currently, the most common regression analysis for categorical data is binary logistic regression. This technique takes one or more ‘predictor’ or ‘explanatory’ variables and attempts to predict the outcome of a binary response variable, such as the use of one sense or near-synonym over another (start vs. begin, for instance). The regression analysis ‘models’ the data, permitting it to indicate which features, or ‘levels’, are most important in distinguishing the binary outcome. It also indicates the statistical significance of each of these predictions. Finally, scores for the overall success of the model in predicting the outcome can be obtained. As one of the most widely employed techniques in categorical statistics, there exists a diverse range of tutorials and textbooks devoted to it. Specifically designed for linguists, Speelman (this volume, 487–533) offers a concise introduction to applying the technique, so too does Baayen (2008: Ch. 6), Dalgaard (2008: Ch. 2008), Johnson (2008: Ch. 5), and Gries (2009b:â•›291–306; 2013: Ch. 5). Speelman and the latter two explanations include R code. Crawley (2005: Ch. 16) also includes lucid explanations of much of the R code needed. More general explanations, which remain accessible to the relative beginner, include Chatterjee and Hadi (2006: Ch. 12), Faraway (2006: Chs. 2–4), Gelman and Hill (2007: Ch. 5), Sheather (2009: Ch. 8), Everitt and Hothorn (2010: Ch. 7), Maindonald and Braun (2010: Ch. 8), Azen and Walker (2011: Chs. 8, 9), and Field et al. (2012: Ch. 8). As mentioned above, the ‘standard’ references for the technique include Harrell (2001), Faraway (2006), Hilbe (2009), Menard (2010: Chs. 8, 9), Agresti (2013: Chs. 4–7; 2007: Chs. 4, 5), and Hoshmer and Lemshow (2013). The technique is widely used in sociolinguistics and has a well-established tradition in Cognitive Linguistics. A few examples of use include Szmrecsanyi (2003, 2006), Heylen (2005b), Grondelaers et al. (2007, 2008), Speelman et al. (2009), Divjak (2010a), Glynn (2010b, this volume, 117–144), Robinson (2010a, 2010b, this volume, 87–116), Speelman and Geeraerts (2010), Deshors (2011, 2014), Levshina (2011), and Deshors and Gries (this volume, 179–204).

Loglinear analysis

Multiway frequency analysis or loglinear analysis is a technique not yet widely used in the field. Unlike binary logistic regression, loglinear analysis is not limited to determining the difference between a maximum of two possibilities. Therefore, it can be used to predict the behaviour of several senses, lexemes, or constructions. The technique is similar to configural frequency analysis, described above. Where configural frequency analysis examines configurations of sets of cells in a multiway frequency table, log-linear analysis looks at the interaction of variables that make up the multiway frequency table. Another way to think of loglinear analysis is to think of it as a logistic regression analysis without a response variable (start vs. begin, for instance).

322 Dylan Glynn

Instead of this response variable, one attempts to predict the actual frequencies for each variable with the minimal number of factors. Gries (this volume) offers a brief introduction to the technique, where it is termed “Poisson regression”. Adler (2010:â•›394–395, 444) offers a very short explanation, but suggests a range of functions in R that can be used for fitting loglinear models (Adler 2010:â•›227, 425, 437–438, 543, 557–558, 569). Thompson’s (2009: Chs. 8, 9) R manual for Agresti (2002) has two detailed chapters devoted to the technique. Short explanations include Oakes (1998: Ch. 5), Agresti (2007: Ch. 7; 2013: Chs. 9, 10), Faraway (2006:â•›61–67, 93–95), Dalgaard (2008: Ch. 15), Gries (2009b:â•›240–248; 2013:â•›324– 327), Tarling (2009: Ch. 7), Braun (2010:â•›258–266), Afifi et al. (2011: Ch. 17), Azen and Walker (2011: Ch. 7), Smith (2011: Ch. 4), Field et al. (2012: Ch. 18), and Ledolter (2013: Ch. 7). Von Eye and Mun (2013) is a new volume devoted to the technique and includes practical explanations in R. However, the book is relatively theoretical and may prove challenging for learners. For users of SPSS, Tabachnick and Fidell (2007: Ch. 16) present a thorough explanation. Kroonenberg (2008) is an approachable, non-technical, volume devoted to the topic, and Christensen (1997) is older and more technical, but comprehensive. Finally, Hilbe (2011) offers a less orthodox discussion, contextualising loglinear modelling as a means for identifying multivariate dependencies. With an example-based discussion, the author reveals how the approach ties in with other techniques. Within the field of Cognitive Linguistics, Krawczak and Glynn (in press) and Glynn (forthc.) are examples of its use.

Multinomial logistic regression

This extension of binary logistic regression (explained above) is also called polychotomous logistic regression, or polytomous logistic regression. The principle is the same as for binary logistic regression, save that there are multiple nominal outcomes. The technique, however, still requires a base line for the model, that is, an outcome that serves as the point of reference for the ‘other’ outcomes (start vs. begin, set off and commence, for example). Arguably the most approachable descriptions to date are Hilbe (2009: Ch. 10), Orme and Combs-Orme (2009: Ch. 3), and Ledolter (2013: Ch. 11), but see also Agresti (2007: Ch. 6). Arppe (2008) represents a detailed study on possible alternatives to this technique. For SPSS users, Tarling (2009: Ch. 6) and Azen and Walker (2011: Ch. 10) include a step-by-step example-based explanation. For Stata uses, Long and Freese (2006) is clear; its explanations are also useful independent of the statistical package used. The application of multinomial logistic regression is not straightforward and the technique has not yet enjoyed wide use in the field. However, as quantitative approaches to semantics continue, its application is likely to be an important contribution. Arppe (2008), Nordmark and Glynn (2013), Krawczak (2014a, 2014b, in press), and Glynn (forthc.) represent examples of its application in Cognitive Linguistics.


Ordinal logistic regression

Also referred to as ordered multinomial logit regression or proportional odds regression, the technique is a special case of logistic regression where the response is multiple and ordered, such as ‘short’, ‘medium’, ‘long’ or ‘young’, ‘older’, and ‘oldest’. At least three ways of modelling ordinal regression exist; the most common is called the proportional method. The principle is straightforward. Rather than a binary response, one has a series of response variables. For example, for an ordered list of choices A, B, C or D, one attempts to predict the outcome of A versus B, C, or D, then in turn A or B versus C and D, and finally A or B or C versus D. If these response variables A, B, C, and D are ordered, this can be interpreted as determining what factors predict that ordering. The most accessible explanations of such modelling can be found in Baayen (2008: Ch. 6), Hilbe (2009: Ch. 9), Orme and Combs-Orme (2009), and Tarling (2009: Ch. 8). O’Connell (2006) is a user-friendly textbook devoted to the technique, but intended for users of SPSS. Long and Freese (2006) is comparable for users of Stata. Agresti (2013:â•›86–98) offers a description of some of the basic issues and tests involved with ordered categories, and Agresti (2007: Ch. 6) offers a more detailed description, though somewhat theoretical. In terms of theory, Agresti (2010) represents a comprehensive work of reference. Johnson and Albert (1999) is a detailed and somewhat technical book devoted to the subject. This is a good reference, but has little explanation on application and only includes a software guide for program MATLAB.

Mixed-effects logistic regression

Sometimes also called multilevel modelling or hierarchical modelling, this technique is similar to ‘normal’ logistic regression, except that the model accounts for both ‘fixed’ effects (that is, the predictors in the model) and ‘random’ effects (or factors we know a priori are ‘noise’ in the model). For example, if one is looking at examples from a small set of sources, such as a set of authors in a diachronic corpus or speakers in discourse analysis, one does not want the individual traits of those authors or speakers influencing the outcome of the analysis. These unwanted effects are treated as ‘random’ in the model. Put simply, mixed-effects regression analysis accounts for those ‘unwanted’ factors, and ‘neutralises’ their effects, preventing them from skewing results. The principle can be applied to any form of regression, including the ordinal and multinomial regression explained above. Speelman (this volume, 487–533) offers a succinct explanation. An older, but thorough, description can be found in Edwards (2000: Ch. 4). Gellman and Hill (2006) offer an extremely detailed, yet approachable, book on the matter. Crawley (2007: Ch. 19), Baayen (2008: Ch. 7), Maindonald (2008: Ch. 10), Sheather (2009: Ch. 10), and Tarling (2009: Ch. 9) give clear introductions to the method, as does Johnson (2008:â•›255–260). See also Frawley (2007: Ch. 19), who gives one of the clearest explanations on how to distinguish random variables from fixed

324 Dylan Glynn

variables, and Maindonald and Braun (2010: Ch. 10), who offer a thorough description of the interpretation of the output in R. Finally, Hox (2010) is a work devoted to the technique. It is broad in its coverage, with a theoretical orientation, but it remains approachable for the faux-débutant, serving as an excellent book of reference. Mixed models are beginning to become more common in the Cognitive Linguistic literature; examples include Bresnan et al. (2007), Divjak (2010b), Klavan (2012), Levshina et al. (2013a; this volume, 205–222); Krawczak and Glynn (in press), and Glynn (2014a). Table 3 summarises the different techniques described here. Although the table systematically covers the techniques for categorical data, it does not include any techniques for continuous data. Moreover, it does not include many of the recent advances and variants, such as random forest classification or hierarchical configural frequency analysis. Tabachnick and Fidell (2007:â•›29–31) offer an excellent breakdown of many of the multivariate techniques available; so too does Baayen (2008:â•›Appendix B). Tummers et al. (2005), Heylen et al. (2008) and Gilquin and Gries (2009) offer extensive discussions on the quantitative state-of-the-art in Cognitive Linguistics. Just as the number of different statistical techniques can be overwhelming for someone first learning, so too can the number of packages and commands available for performing them in R. Packages are modules that expand R’s functionality and the commands are the computer prompts to make them operate. One of R’s most important strengths is the fact it is a vibrant community, with countless active internet fora and just as many people writing packages to refine and advance the application of every imaginable statistical technique. The downside to this, of course, is that a simple search request on the Internet can result in in an overload of information and options. In response to this problem, Table 4 represents a concise reference list for the functions and packages in R for performing the multivariate techniques described above. It is far from complete, being designed as a quick reference for the intermediate user who wishes to get started on a method with which he or she is not yet familiar. Also included are references for tutorials and textbooks on the functions and packages. A complete list would be impossible since many of the techniques have a number of packages devoted, or partially devoted, to them and other techniques have many variants. Moreover, it must be remembered that for the confirmatory techniques, there also exist large numbers of diagnostic and visualisation options, most of which are performed with the use of other more general or more specific packages and functions. Certain books can be recommended for the reader who wishes to go back and investigate the basics that this volume skips, and also for the reader who wishes to delve deeper into the kinds of methods presented here. Baayen’s (2008) Analyzing Linguistic Data is an excellent place to start. Another highly recommended guide for starting statistical analysis using R in Linguistics is Dalgaard’s (2008)’s Introducing Statistics with R. If used in combination with Baayen (2008), one should be able to move on

Techniques and tools 325 Table 3. Quantitative techniques and their usage in corpus-based Cognitive Linguistics Technique

Type

Object collocation

Example of application

Explanation

T-score, Z-score, MI score

measure

strength

identifying multi-word patterns – Biber (2009) identifying constructional variants – Wong (2009)

Evert (2009), Biber & Jones (2010)

Chi-squared test, univariate Fisher’s exact Test

probability / independence

synonymy, constructional – Wulff (2006) polysemy, lexical – Robinson (2010)

Dalgaard (2008), Everitt & Hothorn (2009), Gries (this volume)

Collostructional analysis

univariate

collocation strength

synonymy, constructional – Hilpert (2008) synonymy, constructional – Gilquin (2010)

Stefanowitsch & Gries (2003), Gries & Stefanowitsch (2004a), Hilpert (this volume)

Hierarchical cluster analysis

multivariate associations btw. objects of single variable

polysemy, lexical – Gries (2006) synonymy, lexical – Divjak (2010a)

Baayen (2008), Everitt et al. (2011), Divjak & Fieller (this volume)

Multidimensional multivariate associations btw. objects of scaling multiple variables

synonymy, morphological – Croft & Poole (2008) relations between variants Berthele (2010)

Baayen (2008), Izenman (2008), Everitt & Hothorn (2010)

Correspondence analysis

synonymy, concepts – Szelid & Geeraerts (2008) polysemy, constructional – Glynn (2009)

Le Roux & Rouanet (2010), Husson et al. (2011), Glynn (this volume)

Configural multivariate associations btw. objects of frequency analysis multiple variables

polysemy, constructional – Hilpert (2009) synonymy, constructional – Hoffmann (2011)

von Eye (2002), Tabachnick & Fidell (2007), Gries (2009b)

Discriminant analysis

synonymy, constructional – Gries (2003) synonymy, lexical – Divjak (2010a)

Tabachnick & Fidell (2007), Baayen (2008), Maindonald & Braun (2010)

Classification tree multivariate identify factors that lead to analysis an outcome / prediction

polysemy, lexical – Robinson (2012a) synonymy, lexical – Levshina et al. (this vol.)

Venables & Ripley (2002), Everitt & Hothorn (2010), Maindonald & Braun (2010)

Loglinear analysis multivariate predict correlation multiple response variables

synonymy, constructional – Krawczak & Glynn (in press) Kroonenberg (2008), Hilbe (2011), polysemy, lexical – Glynn (forthc.) Smith (2011)


multivariate predict outcome binary response variable

synonymy, constructional – Szmrecsanyi (2003) synonymy, lexical – Speelman & Geeraerts (2010)


multivariate predict outcome ranked response variable

synonymy. lexico–constructional – Klavan (2012) Baayen (2008), Tarling (2009), synonymy, constructional – Glynn & Krawczak (forthc.) Orme & Combs-Orme (2009)

multivariate associations btw. objects of multiple variables

multivariate identify factors that lead to an outcome / prediction

Orme & Combs-Orme (2009), Everitt & Hothorn (2010), Speelman (this volume)

Multinomial multivariate predict outcome logistic regression multiple response variables

synonymy, lexical Arppe (2008) synonymy, lexical – Krawczak (2014a)

Long & Freese (2006), Tarling (2009), Orme & Combs-Orme (2009)

Mixed-effects multivariate predict outcome logistic regression include random variables

synonymy, constructional – Divjak (2010b) synonymy, constructional – Levshina et al. (this vol.)

Baayen (2008), Maindonald & Braun (2010), Smith (2011)

326 Dylan Glynn

Table 4.â•‡ Functions and packages for categorical multivariate statistics in R Technique

Function

Package

R code tutorial

Hierarchical cluster analysis

hclust

stats*

Crawley (2007:â•›738ff.); Zhao (2013)

agnes

cluster

Kaufman & Rousseeuw (2005); Maechler (2013)

pvclust

pvclust

Suzuki & Hidetoshi (2006); Suzuki (2013)

kmeans

stats*

Crawley (2007:â•›742ff.); Zhao (2013)

clara

cluster

Kaufman & Rousseeuw (2005); Maechler (2013)

pamk

fpc

Hennig (2013)

corresp

MASS*

Venables & Ripley (2002:â•›326ff.); Ripley (2013)

ca

ca

Greenacre (2007); Neandić & Greenacre (2007)

anacor

anacor

de Leeuw & Mair (2009a, 2013a)

MASS*

Venables & Ripley (2002:â•›329f.); Ripley (2013)

K-means cluster analysis Binary correspondence analysis

Multiple correspond- mca mjca ence analysis Multidimensional scaling

ca

Greenacre (2007); Neandić & Greenacre (2007)

MCA

FactoMineR

Lê et al. (2008); Husson et al. (2013)

cmdscale

stats*

Baayen (2008:â•›136ff.); Johnson (2008:â•›208ff.)

sammon

MASS*

Maindonald & Baun (2010:â•›284f.)

smacofSym

smacof

de Leeuw & Mair (2009b, 2013b)

cfa

Funke et al. (2007); von Eye & Mair (2008)

Configural frequency cfa hcfa analysis Linear discriminant Analysis

Classification tree analysis / Random forest classification

cfa

Gries (2010:â•›248ff.); von Eye et al. (2010:â•›265ff.)

cfa2

cfa24

No tutorials available, cf. Schönbrodt (2013)

lda

MASS*

Baayen (2008:â•›167ff.); Maindonald & Braun (2010:â•›385ff.)

discrim

ade4

Chessel et al. (2004); Chessel & Dufour (2013)

rda

klaR

Roever et al. (2013)

rpart

rpart

Zhao (2012:â•›32ff.); Therneau et al. (2013)

tree

tree

Venables & Ripley (2002:â•›266)

ctree

party

Zhao (2013:â•›29ff.)

cforest

party

Strobl et al. (2009a, 2009b)

randomForest randomForest Maindonald & Braun (2010:â•›351ff.);

Liaw & Wiener (2002) Loglinear analysis

glm

MASS*

Maindonald & Braun (2010:â•›258ff.); Baguley (2012)

loglm

MASS*

Thompson (2009:â•›142ff.); Baguley (2012)

quasipois

aod

Lesnoff & Lancelot (2013)

4. The package cfa2 is not currently in the CRAN repository for R but can be found in the RForge repository. This repository is typically used for packages still under development. A simple command listed on the Rforge site for the package will install a package as effortlessly as installation using the ‘normal’ method in R.


Table 4.â•‡ (continued) Technique

Function

Package

R code tutorial


glm

MASS*

Baayen (2008:â•›195ff.); Everitt & Hothorn (2010:â•›122ff.)

lrm

rms†

Harrell (2001:â•›257ff.; 2012:â•›221ff.); Baayen (2008:â•›195ff.)

MCMClogit

MCMCpack

Martin et al. (2010)

polr

MASS*

Faraway (2006:â•›117ff.); Maindonald & Braun (2010:â•›270ff.)

lrm

rms†

Baayen (2008); Johnson (2008)

clm

ordinal

Christensen (2012)

multinom

nnet

Faraway (2006:â•›106); Thompson (2009:â•›118)

polytomous

polytomous

Arppe (2014)

mlogit

mlogit

Field et al. (2012:â•›325); Croissant (2013)

lmer

lme4

Baayen (2008:â•›278ff.); Bates (forthc.:â•›1ff.)

glmmPQL5

MASS*

Johnson (2008:â•›255ff.); Thompson (2009:â•›179ff.)

MCMCglmm

Hadfield (2010)


Multinomial logistic regression Multilevel logistic regression (mixed effects) *

MCMCglmm

Recommended package for the base installation. This means it comes ‘pre-installed’.

†

In older textbooks and tutorials, this package is called Design. The package rms is simply a new version. The command line to use the package is unchanged and so older descriptions remain helpful.

from what is covered in this volume on all three fronts – developing knowledge of R, the basic statistical principles and tests, as well as advanced statistical analysis. Gries’ (2009a) Quantitative Corpus Linguistics with R is another book to consider. Although an excellent book, it is designed more for corpus linguistics per se than multivariate analysis. More in line with the focus of this volume is Gries’ (2009b) Statistics for Linguistics Using R. It covers the basics thoroughly and introduces some multivariate statistical techniques. A new edition, Gries (2013), expands the chapter on precisely the techniques covered in this volume Johnson’s (2008) Quantitative Methods in Linguistics is good for a debutant level statistics textbook using R – it explains both the command line and statistics lucidly and concisely. However, it ‘orders’ the different statistical techniques relative to different subfields of linguistics. This could be misleading for the novice and not particularly logical for the reader with some knowledge in the field, since most of the techniques are not at all restricted to the subfield Johnson ascribes to them. However, the expla-

5. The function glmmPQL uses a so-called penalised quasi-likelihood, which has lost favour in the research community (Crawley 2007:â•›655). Although the functions lmer and MCMCglmm are more up to date in this regard, glmmPQL in the MASS package still works perfectly well, especially when learning since some of the command line is closer to other regression functions a learner may have already mastered.

328 Dylan Glynn

nations of the techniques are clear, especially concerning the issues that lie between the very basic and more advanced study, such as understanding data distribution and samples. Slightly more advanced books, though approachable, are Everitt and Hothorn (2010; 2011). These volumes are excellent textbooks for researchers with an introductory knowledge in statistics and/or with R, but who wish to adopt multivariate techniques – veritable handbooks. Although the examples are not linguistic, they are clear and well chosen. The statistical techniques covered are all explained through the use of examples. The demonstration of the R code is systematic and complete. Finally, Keen (2010) offers a thorough coverage of the graphic possibilities in R. Appropriate for novice and expert alike, the book is practically orientated with detailed examples of the R code.

References Adler, J. (2010). R in a nutshell: A desktop quick reference. Sebastopol: O’Reilly Media. Afifi, A., May S., & Clark, V. A. (2011). Practical multivariate analysis (5th ed.). London: Chapman & Hall. Agresti, A. (2007). An introduction to categorical data analysis (2nd ed.). Hoboken: John Wiley. DOI: 10.1002/0470114754 Agresti, A. (2010). Analysis of ordinal categorical data (2nd ed.). Hoboken: John Wiley. DOI: 10.1002/9780470594001 Agresti, A. (2013) [1990, 2002]. Categorical data analysis (3rd ed.). New York: John Wiley. Arppe, A. (2006). Frequency considerations in morphology: Finnish verbs differ, too. SKY Journal of Linguistics, 19, 175–189. Arppe, A. (2008). Univariate, bivariate and multivariate methods in corpus-based lexicography – A study of synonymy. Unpublished PhD dissertation, University of Helsinki. Azen, R., & Walker, C. (2011). Categorical data analysis for the behavioral and social sciences. New York & Hove: Routledge. Baayen, R. H. (2008). Analyzing linguistic data: A practical introduction to statistics using R. Cambridge: Cambridge University Press. DOI: 10.1017/CBO9780511801686 Baguley, T. (2012). Loglinear models. Online Supplement 5 to Serious stats: A guide to advanced statistics for the behavioral sciences. Basingstoke: Palgrave. Available at: http://www. palgrave.com/psychology/baguley/students/supplements.html. Balahur, A., & Montoyo, A. (2012). Semantic approaches to fine and coarse-grained feature-based opinion mining. In H. Horacek, E. Métais, R. Muñoz, & M. Wolska (Eds.), Natural language processing and information systems (pp. 142–153). Berlin: Springer. Barnabé, A. (2012). Le schème du chemin en grammaire et sémantique anglaises. Unpublished PhD dissertation, Université Bordeaux 3. Bates, D. (Forthcoming). lme4: Mixed-effects modeling with R. Heidelberg & New York: Springer. Preprints available at: http://lme4.r-forge.r-project.org/lMMwR/lrgprt.pdf. Benzécri, J.-P. (1980). Pratique de l’analyse des donnees. Paris: Dunod. Benzécri, J.-P. (1992). Correspondence analysis handbook. New York: Dekker.


Berthele, R. (2010). Investigations into the folk’s mental models of linguistic varieties. In D. Geeraerts, G. Kristiansen, & Y. Peirsman (Eds.), Advances in cognitive sociolinguistics (pp. 265–290). Berlin & New York: Mouton de Gruyter. Biber, D. (2009). A corpus-driven approach to formulaic language in English: Multi-word patterns in speech and writing. International Journal of Corpus Linguistics, 14, 275–311. DOI: 10.1075/ijcl.14.3.08bib Biber, D., & Jones, J. (2009). Quantitative methods in Corpus Linguistics. In A. Lüdeling, & M. Kytö (Eds.), Corpus Linguistics: An international handbook. Vol. 2. (pp. 1287–1304). Berlin & New York: Mouton de Gruyter. Borg, I., Groenen, & Mair, P. (2013). Applied multidimensional scaling. Heidleberg & New York: Springer. DOI: 10.1007/978-3-642-31848-1 Borg, I., & Groenen, P. (2005). Modern multidimensional scaling (2nd ed.). Heidelberg & New York: Springer. Bresnan, J., Cueni, A., Nikitina, T., & Baayen, H. (2007). Predicting the dative. In G. Bouma, I. Krämer, & J. Zwarts (Eds.), Cognitive foundations of interpretation alternation (pp. 69– 94). Amsterdam: Royal Netherlands Academy of Arts and Sciences. Bybee, J., & Eddington, D. (2006). A usage-based approach to Spanish verbs of ‘becoming’. Language, 82, 323–355. DOI: 10.1353/lan.2006.0081 Cadoret, M., Lê, S., & Pagès, J. (2011). Multidimensional scaling versus multiple correspondence analysis when analyzing categorization data. In B. Fichet, D. Piccolo, R. Verde, & M. Vichi (Eds.), Classification and multivariate analysis for complex data structures (pp. 301–308). Heidleberg & New York: Springer. DOI: 10.1007/978-3-642-13312-1_31 Chaffin, R. (1992). The concept of a semantic relation. In A. Lehrer, & E. Kittay (Eds.), Frames, fields, and contrasts: New essays in semantic and lexical organisation (pp. 253–288). London: Lawrence Erlbaum. Chatterjee, S., & Hadi, A. (2006). Regression analysis by example. London: John Wiley. DOI: 10.1002/0470055464 Chessel, D., & Dufour, A.-B. (2013). Analysis of ecological data: Exploratory and Euclidean methods in environmental sciences. Available at: http://cran.r-project.org/web/packages/ ade4/ade4.pdf. Chessel, D., Dufour A.-B, & Thioulouse, Y. (2004) The ade4 package – I: One-table methods. R News, 4, 5–10. Christensen, R. (1997). Log-linear models and logistic regression (2nd ed.). Heidleberg & New York: Springer. Christensen, R. (2012). A tutorial on fitting cumulative link models with the ordinal package. Available at: http://cran.r-project.org/web/packages/ordinal/vignettes/clm_intro.pdf. Clancy, S. (2006). The topology of Slavic case: Semantic maps and multidimensional scaling. Glossos, 7, 1–28. Colleman, T. (2009). The semantic range of the Dutch double object construction. A collostructional perspective. Constructions and Frames, 1, 190–221. DOI: 10.1075/cf.1.2.02col Colleman, T. (2010). Beyond the dative alternation: The semantics of the Dutch aan-Dative. In D. Glynn, & K. Fischer (Eds.), Quantitative Cognitive Semantics: Corpus-driven approaches (pp. 271–304). Berlin & New York: Mouton de Gruyter. Cox, T., & Cox, M. (2001). Multidimensional scaling (2nd ed.). Boca Raton: Chapman & Hall. Crawley, M. (2005). Statistics: An introduction using R. Southern Gate & Hoboken: John Wiley. DOI: 10.1002/9781119941750

330 Dylan Glynn

Crawley, M. (2007). The R book. Chichester: John Wiley. DOI: 10.1002/9780470515075 Croft, W., & Poole, K. (2008). Inferring universals from grammatical variation: Multidimensional scaling for typological analysis. Theoretical Linguistics, 34, 1–37. DOI: 10.1515/THLI.2008.001 Croissant, Y. (2013). Estimation of multinomial logit models in R: The mlogit packages. Available at: cran.r-project.org/web/packages/mlogit/mlogit.pdf. Daille, B., Dubreil, E. Monceaux, L., & Vernier, M. (2011). Annotating opinion–evaluation of blogs: The Blogoscopy corpus. Language Resources and Evaluation, 45, 409–437. DOI: 10.1007/s10579-011-9154-z Dalgaard, P. (2008). Introductory statistics with R (2nd ed.). Dordrecht: Springer. DOI: 10.1007/978-0-387-79054-1 De Cock, B. (2014a). A discourse-functional analysis of speech participant profiling in spoken Spanish. Amsterdam & Philadelphia: John Benjamins. De Cock, B. (2014b). The discursive effects of Spanish impersonals uno and se. In D. Glynn, & M. Sjölin (Eds.), Subjectivity and epistemicity: Corpus, discourse, and literary approaches to stance (pp. 103–120). Lund: Lund University Press. De Leeuw, J., & Mair, P. (2009a). Simple and canonical correspondence analysis using the R package anacor. Journal of Statistical Software, 31, 1–18. De Leeuw, J., & Mair, P. (2009b). Multidimensional scaling using majorization: The R package smacof. Journal of Statistical Software, 31, 1–30. De Leeuw, J., & Mair, P. (2013a). anacor: Simple and canonical correspondence analysis. Available at: cran.r-project.org/web/packages/anacor/anacor.pdf. De Leeuw, J., & Mair, M. (2013b). SMACOF for multidimensional scaling. Available at: http:// cran.r-project.org/web/packages/smacof/smacof.pdf. Deignan, A. (2005). Metaphor and Corpus Linguistics. Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/celcr.6 Delorge, M. (2009). A diachronic corpus study of the constructional behaviours of reception verbs in Dutch. In B. Lewandowska-Tomaszczyk, & K. Dziwirek (Eds.), Studies in Cognitive Corpus Linguistics (pp. 249–272). Frankfurt/Main: Peter Lang. Desagulier, G. (In press). Le statut de la fréquence dans les Grammaires de Constructions: ‘simple comme bonjour’? Langages. Desagulier, G. (Submitted). Quite new methods for a rather old issue: Exploring and visualizing collocation data from the BNC with correspondence analysis. Deshors, S. (2011). A multifactorial study of the uses of may and can in French-English interlanguage. Unpublished PhD dissertation, University of Sussex. Deshors, S. (2014). Identifying different types of non-native co-occurrence patterns: A corpus-based approach. In D. Glynn, & M. Sjölin (Eds.), Subjectivity and epistemicity: Corpus, discourse, and literary approaches to stance (pp. 387–412). Lund: Lund University Press. Diehl, H. (2014). On modal meaning in the uses of quite, rather, pretty and fairly as degree modifiers in British English. Unpublished PhD dissertation, Lund University. Dirven, R., Goossens, L., Putseys, Y., & Vorlat, E. (1982). The scene of linguistic action and its perspectivization by speak, talk, say, and tell. Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/pb.iii.6 Divjak, D. (2006). Ways of intending: A corpus-based Cognitive Linguistic approach to near-synonyms in Russian. In St. Th. Gries, & A. Stefanowitsch (Eds.), Corpora in Cognitive Linguistics: Corpus-based approaches to syntax and lexis (pp. 19–56). Berlin & New York: Mouton de Gruyter.


Divjak, D. (2010a). Structuring the lexicon: A clustered model for near-synonymy. Berlin & New York: Mouton de Gruyter. Divjak, D. (2010b). Corpus-based evidence for an idiosyncratic aspect-modality relation in Russian. In D. Glynn, & K. Fischer (Eds.), Quantitative Cognitive Semantics: Corpus-driven approaches (pp. 305–331). Berlin & New York: Mouton de Gruyter. Divjak, D., & Gries, St. Th. (2006). Ways of trying in Russian: Clustering behavioral profiles. Corpus Linguistics and Linguistic Theory, 2, 23–60. DOI: 10.1515/CLLT.2006.002 Divjak, D., & Gries, St. Th. (2009). Corpus-based Cognitive Semantics: A contrastive study of phrasal verbs in English and Russian. In B. Lewandowska-Tomaszczyk, & K. Dziwirek (Eds.), Studies in Cognitive Corpus Linguistics (pp. 273–296). Frankfurt/Main: Peter Lang. Divjak, D., & Gries, St. Th. (Eds.). (2012). Frequency effects in language learning and processing. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110274059 Drenan, R. (2009). Statistics for archaeologists: A common sense approach (2nd ed.). Heidelberg & New York: Springer. Dziwirek, K., & Lewandowska-Tomaszczyk, B. (2011). Complex emotions and grammatical mismatches: A contrastive corpus-based study. Berlin & New York: Mouton de Gruyter. Edwards, D. (2000). Introduction to graphical modelling (2nd ed.). Heidelberg: Springer. DOI: 10.1007/978-1-4612-0493-0 Everitt, B. S. (2005). An R and S-PLUS companion to multivariate analysis. London: Springer. Everitt, B. S., & Hothorn, I. (2010). A handbook of statistical analyses using R (2nd ed.). Boca Raton: Taylor & Francis. DOI: 10.1201/9781420079340 Everitt, B. S., & Hothorn, I. (2011). An introduction to applied multivariate analysis with R. Munich: Springer. DOI: 10.1007/978-1-4419-9650-3 Everitt, B. S., Landau, S., Leese, M., & Stahl, D. (2011). Cluster analysis (5th ed.). Chichester: John Wiley. DOI: 10.1002/9780470977811 Evert, S. (2009). Corpora and collocations. In A. Lüdeling, & M. Kytö (Eds.), Corpus Linguistics: An international handbook (pp. 1212–1249). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110213881.2.1212 Faraway, J. (2002). Practical regression and anova using R. Available at: cran.r-project.org/doc/ contrib/Faraway-PRA.pdf. Faraway, J. (2006). Extending the linear model with R: Generalized linear, mixed effects and nonparametric regression models. London: Taylor & Francis. Field, A., Miles, J., & Field, Z. (2012). Discovering statistics using R. London & Thousand Oaks: Sage. Fillmore, C., & Atkins, B. (1992). Toward a frame-based lexicon: The semantics of risk and its neighbours. In A. Lehrer, & E. Kittay (Eds.), Frames, fields, and contrasts: New essays in semantic and lexical organisation (pp. 75–102). London: Lawrence Erlbaum. Firth, J. R. (1957). A synopsis of linguistic theory 1930–1955. In J. R. Firth (Ed.), Studies in linguistic analysis (pp. 1–32). Oxford: Basil Blackwell. Fischer, K. (2000). From Cognitive Semantics to Lexical Pragmatics: The functional polysemy of discourse particles. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110828641 Flores Salgado, E. (2011). The pragmatics of requests and apologies: Developmental patterns in Mexican students. Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/pbns.212 Fontaine, J., Scherer, K., & Soriano, C. (Eds.). (2013). Components of emotional meaning: A sourcebook. Oxford: Oxford University Press. DOI: 10.1093/acprof:oso/9780199592746.001.0001 Funke, S., Mair, P., & von Eye, A. (2007). cfa: R package for the analysis of configuration frequencies. Available at: http://cran. R-project.org.

332 Dylan Glynn

Geeraerts, D. (2010). The doctor and the semantician. In D. Glynn, & K. Fischer (Eds.), Quantitative Cognitive Semantics: Corpus-driven approaches (pp. 63–78). Berlin & New York: Mouton de Gruyter. Geeraerts, D. (2011). Entrenchment, conventionalization, and empirical method. Presented at the 44th Meeting of the Societas Linguistica Europaea, Logroño. Geeraerts, D., Grondelaers, S., & Bakema, P. (1994). The structure of lexical variation: Meaning, naming, and context. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110873061 Geeraerts, D., Grondelaers, S., & Speelman, D. (1999). Convergentie en Divergentie in de Nederlandse Woordenschat. Amsterdam: Meertens Instituut. Geeraerts, D., Kristiansen, G., & Peirsman, Y. (Eds.). (2010). Advances in cognitive sociolinguistics. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110226461 Gelman, A., & Hill, J. (2007). Data analysis using regression and multilevel/hierarchical models. Cambridge: Cambridge University Press. Gilquin, G. (2010). Corpus, cognition and causative constructions. Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/scl.39 Glynn, D. (2007). Mapping meaning: Toward a usage-based methodology in Cognitive Semantics. Unpublished PhD dissertation, University of Leuven. Glynn, D. (2009). Polysemy, syntax, and variation: A usage-based method for Cognitive Semantics. In V. Evans, & S. Pourcel (Eds.), New directions in Cognitive Linguistics (pp. 77– 106). Amsterdam & Philadelphia: John Benjamins. Glynn, D. (2010a). Synonymy, lexical fields, and grammatical constructions: A study in usage-based Cognitive Semantics. In H.-J. Schmid, & S. Handl (Eds.), Cognitive foundations of linguistic usage-patterns: Empirical studies (pp. 89–118). Berlin & New York: Mouton de Gruyter. Glynn, D. (2010b). Testing the hypothesis: Objectivity and verification in usage-based Cognitive Semantics. In D. Glynn, & K. Fischer (Eds.), Quantitative Cognitive Semantics: Corpus-driven approaches (pp. 239–270). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110226423 Glynn, D. (2014a). The conceptual profile of the lexeme home: A multifactorial diachronic analysis. In J. E. Díaz-Vera (Ed.), Metaphor and metonymy across time and cultures (pp. 265– 293). Berlin & New York: Mouton de Gruyter. Glynn, D. (2014b). The social nature of anger: Multivariate corpus evidence for context effects upon conceptual structure. In I. Novakova, P. Blumenthal, & D. Siepmann (Eds.), Emotions in discourse (pp. 69–82). Frankfurt/Main: Peter Lang. Glynn, D. (Forthcoming). Mapping meaning: Corpus methods for Cognitive Semantics. Cambridge: Cambridge University Press. Glynn, D., & Sjölin, M. (2011). Cognitive Linguistic methods for literature: A usage-based approach to metanarrative and metalepsis. In A. Kwiatkowska (Ed.), Texts and minds: Papers in cognitive poetics and rhetoric (pp. 85–102). Frankfurt/Main: Peter Lang. Glynn, D., & Krawczak, K. (Forthcoming). Social cognition, Cognitive Grammar and corpora: A multifactorial approach to epistemic modality. Cognitive Linguistics. Glynn, D., & Fischer, D. (Eds.). (2010). Quantitative Cognitive Semantics: Corpus-driven approaches. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110226423 Glynn, D., & Sjölin, M. (Eds.). (2014). Subjectivity and epistemicity: Corpus, discourse, and literary approaches to stance. Lund: Lund University Press. Greenacre, M. (2007) [1993]. Correspondence analysis in practice (2nd ed.). London: Chapman & Hall.


Greenacre, M. (2010). Biplots in practice. Bilbao: Fundación BBVA. Gries, St. Th. (1999). Particle movement: A cognitive and functional approach. Cognitive Linguistics, 10, 105–145. DOI: 10.1515/cogl.1999.005 Gries, St. Th. (2000). Towards multifactorial analyses of syntactic variation: The case of particle placement. Doctoral dissertation, University of Hamburg. Gries, St. Th. (2003). Multifactorial analysis in Corpus Linguistics: A study of particle placement. London: Continuum Press. Gries, St. Th. (2006). Corpus-based methods and Cognitive Semantics: The many senses of to run. In St. Th. Gries, & A. Stefanowitsch (Eds.), Corpora in Cognitive Linguistics: Corpus-based approaches to syntax and lexis (pp. 57–99). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110197709 Gries, St. Th. (2009a). Quantitative Corpus Linguistics with R: A practical introduction. London: Routledge. DOI: 10.1515/9783110216042 Gries, St. Th. (2009b). Statistics for Linguistics with R: A practical introduction (1st ed.). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110216042 Gries, St. Th. (2010). Behavioral profiles: A fine-grained and quantitative approach in corpus based lexical semantics. The Mental Lexicon, 5, 323–346. DOI: 10.1075/ml.5.3.04gri Gries, St. Th. (2013). Statistics for linguistics with R: A practical introduction (2nd ed.). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110307474 Gries, St. Th., & Divjak, D. (2009). Behavioral profiles: A corpus-based approach to cognitive semantic analysis. In V. Evans, & S. Pourcel (Eds.), New directions in Cognitive Linguistics (pp. 57–75). Amsterdam & Philadelphia: John Benjamins. Gries, St. Th., & Hilpert, M. (2008). The identification of stages in diachronic data: Variability-based neighbor clustering. Corpora, 3, 59–81. DOI: 10.3366/E1749503208000075 Gries, St. Th., & Stefanowitsch, A. (2004a). Extending collostructional analysis: A corpus-based perspective on ‘alternations’. International Journal of Corpus Linguistics, 9, 97–129. DOI: 10.1075/ijcl.9.1.06gri Gries, St. Th., & Stefanowitsch, A. (2004b). Co-varying collexemes in the into-causative. In M. Achard, & S. Kemmer (Eds.), Language, culture, and mind (pp. 225–36). Stanford: CSLI. Gries, St. Th., & Divjak, D. (Eds.). (2012). Frequency effects in language representation. Berlin & New York: Mouton de Gruyter. Gries, St. Th., & Stefanowitsch, A. (Eds.). (2006). Corpora in Cognitive Linguistics: Corpus-based approaches to syntax and lexis. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110197709 Grondelaers, S. (2000). De distributie van niet-anaforisch er buiten de eerste zinsplaats: SocioÂ� lexicologische, functionele en psycholinguïstische aspecten van er’s status als presentatief signaal. Doctoral dissertation, University of Leuven. Grondelaers S., Geeraerts, D., & Speelman, D. (2007). A case for a cognitive Corpus Linguistics. In M. Gonzalez-Marquez, I. Mittleberg, S. Coulson, & M. Spivey (Eds.), Methods in Cognitive Linguistics (pp. 149–169). Amsterdam & Philadelphia: John Benjamins. Grondelaers S., Speelman, D., & Geeraerts, D. (2008). National variation in the use of er “there”: Regional and diachronic constraints on cognitive explanations. In G. Kristiansen, & R. Dirven (Eds.), Cognitive Sociolinguistics: Language variation, cultural models, social systems (pp. 153–204). Berlin & New York: Mouton de Gruyter. Hadfield, J. (2010). MCMC methods for multi-response generalized linear mixed models: The MCMCglmm R package. Journal of Statistical Software, 33, 1–22.

334 Dylan Glynn

Härdle, W., & Simar, L. (2007). Applied multivariate statistical analysis. Heidelberg & New York: Springer. Harrell, F. (2001). Regression modeling strategies: With Applications to linear models, logistic regression, and survival analysis. Heidelberg & New York: Springer. Harrell, F. (2012). Regression modeling strategies. Unpublished manuscript, available at: www. biostat.mc.vanderbilt.edu/wiki/pub/Main/RmS/rms.pdf. Hennig, C. (2013). Flexible procedures for clustering. Available at: http://cran.r-project.org/ web/packages/fpc/fpc.pdf. Heylen, K. (2005a). A quantitative corpus study of German word order variation. In St. Kepser, & M. Reis (Eds.), Linguistic evidence: Empirical, theoretical and computational perspectives (pp. 241–264). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110197549.241 Heylen, K. (2005b). Zur Abfolge (pro)nominaler Satzglieder im Deutschen: Eine korpusbasierte Analyse der relativen Abfolge von nominalem Subjekt und pronominalem Objekt im Mittelfeld, 264. Doctoral dissertation, University of Leuven. Heylen, K., & Ruette, T. (2013). Degrees of semantic control in measuring aggregated lexical distances. In L. Borin, A. Saxena, A., & T. Rama (Eds.), Approaches to measuring linguistic differences (pp. 353–374). Berlin & New York: Mouton de Gruyter. Heylen, K., Tummers, J., & Geeraerts, D. (2008). Methodological issues in corpus-based Cognitive Linguistics. In G. Kristiansen, & R. Dirven (Eds.), Cognitive Sociolinguistics: Language variation, cultural models, social systems (pp. 91–128). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110199154.2.91 Hilbe, J. (2009). Logistic regression models. London: Chapman & Hall. Hilbe, J. (2011) [2007]. Negative binomial regression (2nd ed.). Cambridge: Cambridge University Press. Hilpert, M. (2008). Germanic future constructions: A usage-based approach to language change. Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/cal.7 Hilpert, M. (2009). The German mit-predicative construction. Constructions and Frames, 1, 29–55. DOI: 10.1075/cf.1.1.03hil Hilpert, M. (2012). Constructional change in English: Developments in allomorphy, word formation, and syntax. Cambridge: Cambridge University. Hoffmann, Th. (2011). Preposition placement in English: A usage-based approach. Cambridge: Cambridge University Press. DOI: 10.1017/CBO9780511933868 Hosmer, D., & Lemeshow, S. (2013) [1989, 2000]. Applied logistic regression. Hoboken: John Wiley. DOI: 10.1002/9781118548387 Hox, J. (2010). Multilevel analysis: Techniques and applications (2nd ed.). Hove & New York: Routledge. Husson, F. Josse, J., Lê, S., & Mazet, J. (2013). Multivariate exploratory data analysis and data mining with R. Available at: http://cran.r-project.org/web/packages/FactoMineR/ FactoMineR.pdf. Husson, F., Lê, S., & Pagès, J. (2011). Exploratory multivariate analysis by example using R. London: Chapman & Hall. Izenman, A. (2008). Modern multivariate statistical techniques: Regression, classification and manifold learning. Heidelberg & New York: Springer. DOI: 10.1007/978-0-387-78189-1 Janda, L., & Solovyev, V. (2009). What constructional profiles reveal about synonymy: A case study of the Russian words for sadness and happiness. Cognitive Linguistics, 20, 367–393. DOI: 10.1515/COGL.2009.018 Johnson, K. (2008). Quantitative methods in linguistics. Oxford: Blackwell.


Johnson, V., & Albert, J. (1999). Ordinal data modeling. Heidelberg & New York: Springer. Kärkkäinen, E. (2003). Epistemic stance in English conversation: A description of its interactional functions, with a focus on I think. Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/pbns.115 Kaufman, L., & Rousseeuw, P. (2005) [1990]. Finding groups in data: An introduction to cluster analysis. Hoboken: John Wiley. Keen, K. (2010). Graphics for statistics and data analysis with R. Boca Raton: CRC Press. Klavan, J. (2012). Evidence in linguistics: Corpus-linguistic and experimental methods for studying grammatical synonymy. Doctoral Dissertation, University of Tartu. Klavan, J, Kesküla K., & Ojava, L. (2011). Synonymy in grammar: The Estonian adessive case and the adposition peal ‘on’. In S. Kittilä, K. Västi, & J. Ylikoski (Eds.), Studies on case, animacy and semantic roles (pp. 1–19). Amsterdam & Philadelphia: John Benjamins. Krawczak, K. (2014a). Shame and its near-synonyms in English: A multivariate corpus-driven approach to social emotions. In I. Novakova, P. Blumenthal, & D. Siepmann (Eds.), Emotions in discourse (pp. 84–94). Frankfurt/Main: Peter Lang. Krawczak, K. (2014b). Epistemic stance predicates in English: A quantitative corpus-driven study of subjectivity. In D. Glynn, & M. Sjölin (Eds.), Subjectivity and epistemicity: Corpus, discourse, and literary approaches to stance (pp. 355–386). Lund: Lund University Press. Krawczak, K. (In press). Corpus evidence for the cross-cultural structure of social emotions: Shame, embarrassment, and guilt in English and Polish. Poznań Studies in Contemporary Linguistics. Krawczak, K., & Glynn, D. (2011). Context and cognition: A corpus-driven approach to parenthetical uses of mental predicates. In K. Kosecki, & J. Badio (Eds.), Cognitive processes in language (pp. 87–99). Frankfurt/Main: Peter Lang. Krawczak, K., & Kokorniak, I. (2012). Corpus-driven quantitative approach to the construal of Polish ‘think’. Poznań Studies in Contemporary Linguistics, 48, 439–472. DOI: 10.1515/psicl-2012-0021 Krawczak, K., & Glynn, D. (In press). Operationalising construal: Of/about prepositional profiling for cognitive and communicative predicates. In C. M. Bretones Callejas (Ed.), Construals in language and thought: What shapes what? Amsterdam: John Benjamins. Kroonenberg, P. (2008). Applied multiway data analysis. New York: John Wiley. DOI: 10.1002/9780470238004 Lê, S., Josse, J., & Husson, F. (2008). FactoMineR: An R package for multivariate analysis. Journal of Statistical Software, 25, 1–18. Le Roux, B., & Rouanet, H. (2004). Geometric data analysis: From correspondence analysis to structured data analysis. Dordrecht: Kluwer. Le Roux, B., & Rouanet, H. (2010). Multiple correspondence analysis. London & Thousand Oaks: Sage. Ledolter, J. (2013). Data mining and business analytics with R. Hoboken: John Wiley. DOI: 10.1002/9781118596289 Lesnoff, M., & Lancelot, R. (2013). Analysis of overdispersed data. Available at: http:// cran.r-project.org/web/packages/aod/aod.pdf. Levshina, N. (2011). A usage-based study of Dutch causative constructions. Doctoral dissertation, University of Leuven. Levshina, N. (2012). Comparing constructicons: A usage-based analysis of the causative construction with doen in Netherlandic and Belgian Dutch. Constructions and Frames, 4, 76–101. DOI: 10.1075/cf.4.1.04lev

336 Dylan Glynn

Levshina, N., Geeraerts, D., & Speelman, D. (2013a). Towards a 3D-grammar: Interaction of linguistic and extralinguistic factors in the use of Dutch causative constructions. Journal of Pragmatics, 52, 34–48. DOI: 10.1016/j.pragma.2012.12.013 Levshina, N., Geeraerts, D., & Speelman, D. (2013b). Mapping constructional spaces: A contrastive analysis of English and Dutch analytic causatives. Linguistics, 51, 825–854. DOI: 10.1515/ling-2013-0028 Lewandowska-Tomaszczyk, B., & Dziwirek, K. (Eds.). (2009). Studies in Cognitive Corpus Linguistics. Frankfurt/Main: Peter Lang. Liaw, A., & Wiener, M. (2002). Classification and regression by randomForest. R News, 2, 18–22. Long, J. S., & Freese, J. (2006) [2001]. Regression models for categorical dependent variables using Stata. College Station: Stata Press. Louwerse, M., & Van Peer, W. (2009). How cognitive is cognitive poetics? The interaction between symbolic and embodied cognition. In G. Brône, & J. Vandaele (Eds.), Cognitive poetics goals, gains and gaps (pp. 423–444). Berlin & New York: Mouton de Gruyter. Maechler, M. (2013). Cluster analysis extended. Available at: http://cran.r-project.org/web/ packages/cluster/cluster.pdf. Maindonald, J. (2008). Using R for data analysis and graphics: Introduction, code and commentary. Available at: http://www.maths.anu.edu.au/˜johnm/r/usingR.pdf. Maindonald, J., & Braun, J. (2010) [2003]. Data analysis and graphics using R (3rd ed.). Cambridge: Cambridge University Press. Marden, J. (2011). Multivariate statistical analysis: Old school. Department of Statistics, University of Illinois at Urbana-Champaign. Available at: istics.net/pdfs/multivariate.pdf. Martin, A. D., Quinn, K. M., & Park, J. H. (2010). Markov chain Monte Carlo (MCMC) package. Available at: http://mcmcpack.wustl.edu/. Menard, S. (2002). Applied logistic regression analysis (2nd ed.). London & Thousand Oaks: Sage. Menard, S. (2010). Logistic regression: From introductory to advanced concepts and applications. London & Los Angeles: Sage. Morgenstern, A., Blondel, M., Caët, S., & Boutet, D. (2011). Hearing children’s use of pointing gestures: From pre-linguistic buds to the blossoming of communication skills. Presentation at SALC III, Copenhagen. Murtagh, F. (2005). Correspondence analysis and data coding with R and Java. London: Chapman & Hall. DOI: 10.1201/9781420034943 Myers, D. (1994). Testing for prototypicality: The Chinese morpheme gong. Cognitive Linguistics, 5, 261–280. DOI: 10.1515/cogl.1994.5.3.261 Neandić, O., & Greenacre, M. (2007). Correspondence analysis in R, with two- and three-dimensional graphics: The ca Package. Journal of Statistical Software, 20, 1–13. Newman, J., & Rice, S. (2004). Patterns of usage for English sit, stand, and lie: A cognitively-inspired exploration in corpus linguistics. Cognitive Linguistics, 15, 351–396. DOI: 10.1515/cogl.2004.013 Newman, J., & Rice, S. (2006). Transitivity schemas of English eat and drink in the BNC. In St. Th. Gries, & A. Stefanowitsch (Eds.), Corpora in Cognitive Linguistics: Corpus-based approaches to syntax and lexis. (pp. 225–260). Berlin & New York: Mouton de Gruyter. Nordmark, H., & Glynn, D. (2013). anxiety between mind and society: A corpus-driven cross-cultural study of conceptual metaphors. Explorations in English Language and Linguistics, 1, 107–130.


O’Connell, A. (2006). Logistic regression models for ordinal response variables. London & Thousand Oaks: Sage. Oakes, M. (1998). Statistics for Corpus Linguistics. Edinburgh: Edinburgh University Press. Orme, J., & Combs-Orme, T. (2009). Multiple regression with discrete dependent variables. Oxford: Oxford University Press. DOI: 10.1093/acprof:oso/9780195329452.001.0001 Peirsman, Y. Heylen, K., & Geeraerts, D. (2010). Applying word space models to sociolinguistics: Religion names before and after 9/11. In D. Geeraerts, G. Kristiansen, & Y. Peirsman (Eds.), Advances in Cognitive Sociolinguistics (pp. 111–139). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110226461 Pęzik, P. (2009). Extraction of multiword expressions for corpus-based discourse analysis. In B. Lewandowska-Tomaszczyk, & K. Dziwirek (Eds.), Studies in Cognitive Corpus Linguistics (pp. 249–272). Frankfurt/Main: Peter Lang. Pichler, H. (2013). The structure of discourse-pragmatic variation. Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/silv.13 Plevoets, K., Speelman, D., & Geeraerts, D. (2008). The distribution of T/V pronouns in Netherlandic and Belgian Dutch. In K. Schneider, & A. Baron (Eds.), Variational pragmatics: Regional varieties in pluricentric languages (pp. 181–209). Amsterdam & Philadelphia: John Benjamins. Pütz, M, Robinson, J. A., & Reif, M. (Eds.) (2012). Cognitive Sociolinguistics: Social and cultural variation in cognition and language use. (Special edition of Annual Review of Cognitive Linguistics, 10.) Ravid, D., & Hanauer, D. (1998). A prototype theory of rhyme: Evidence from Hebrew. Cognitive Linguistics, 9, 79–106. DOI: 10.1515/cogl.1998.9.1.79 Read, J., & Carroll, J. (2012). Annotating expressions of Appraisal in English. Language Resources and Evaluation, 46, 421–447. DOI: 10.1007/s10579-010-9135-7 Reif, M., Robinson, J. A., & Pütz, M. (Eds.). (2013). Variation in language and language use: Linguistic, socio-cultural and cognitive perspectives. Frankfurt/Main: Peter Lang. Rencher, A. (2002). Methods of multivariate analysis (2nd ed.). New York: John Wiley. DOI: 10.1002/0471271357 Rice, S., Sandra, D., & Vanrespaille, M. (1999). Prepositional semantics and the fragile link between space and yime. In M. Hiraga, C. Sinha, & S. Wilcox (Eds.), Cultural, typology and psycholinguistic issues in Cognitive Linguistics (pp. 107–127). Amsterdam & Philadelphia: John Benjamins. Ripley, B. (2013). Support functions and datasets for Venables and Ripley’s MASS. Available at: http://cran.r-project.org/web/packages/MASS/MASS.pdf. Robinson, J. A. (2010a). Awesome insights into semantic variation. In D. Geeraerts, G. Kristiansen, & Y. Piersman (Eds.), Advances in Cognitive Sociolinguistics (pp. 85–109). Berlin & New York: Mouton de Gruyter. Robinson, J. A. (2010b). Semantic variation and change in present-day English. Doctoral dissertation, University of Sheffield. Robinson, J. A. (2012). A sociolinguistic perspective on semantic change. In K. Allan, & J. A. Robinson (Eds.), Current methods in Historical Semantics (pp. 191–231). Berlin & New York: Mouton de Gruyter. Roever, C., Raabe, N., Luebke, K., Ligges, U., Szepannek, G., & Zentgraf, M. (2013). Classification and visualization. Unpublished manuscript available at: http://cran.r-project.org/ web/packages/klaR/klaR.pdf.

338 Dylan Glynn

Rudzka-Ostyn, B. (1989). Prototypes, schemas, and cross-category correspondences: The case of ask. In D. Geeraerts (Ed.), Prospects and problems of prototype theory (pp. 613–661). Berlin & New York: Mouton de Gruyter. Rudzka-Ostyn, B. (1995). Metaphor, schema, invariance: The case of verbs of answering. In L. Goossens, P. Pauwels, B. Rudzka-Ostyn, A.-M. Simon-Vandenbergen, & J. Vanparys (Eds.), By word of mouth: Metaphor, metonymy, and linguistic action from a cognitive perspective (pp. 205–244). Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/pbns.33 Ruette, T., Ehret, K., & Szmrecsanyi, B. (In press). Frequency effects in lexical sociolectometry are insubstantial. In H. Behrens, & S. Pfänder (Eds.), Again on frequency effects in language. Berlin & New York: Mouton de Gruyter. Ruette, T., Geeraerts, D., Peirsman, Y., & Speelman, D. (Forthcoming). Semantic weighting mechanisms in scalable lexical sociolectometry. In B. Szmrecsanyi, & B. Waelchli (Eds.), Aggregating dialectology and typology: Linguistic variation in text and speech, within and across languages. Berlin & New York: Mouton de Gruyter. Sagi, E., Kaufmann, S., & Clark, B. (2011). Tracing semantic change with latent semantic analysis. In K. Allan, & J. Robinson (Eds.), Current methods in Historical Semantics (pp. 161– 183). Berlin & New York: Mouton de Gruyter. Sandra, D., & Rice, S. (1995). Network analyses of prepositional meaning: Mirroring whose mind – the linguist’s or the language user’s? Cognitive Linguistics, 6, 89–130. DOI: 10.1515/cogl.1995.6.1.89 Scheibman, J. (2002). Point of view and grammar: Structural patterns of subjectivity in American English conversation. Amsterdam & Philadelphia: John Benjamins. DOI: 10.1075/sidag.11 Scherer, K. (2005). What are emotions? And how can they be measured? Social Science Information, 44, 693–727. DOI: 10.1177/0539018405058216 Schmid, H.-J. (1993). Cottage and co., idea, start vs. begin: Die kategorisierung als grundprinzip einer differenzierten bedeutungsbeschreibung. Tübingen: Max Niemeyer. DOI: 10.1515/9783111355771 Schmid, H.-J. (2000). English abstract nouns as conceptual shells: From corpus to cognition. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110808704 Schmidtke-Bode, K. (2009). Going-to-V and gonna-V in child language: A quantitative approach to constructional development. Cognitive Linguistics, 20, 509–553. DOI: 10.1515/COGL.2009.023 Schönbrodt, F., Collins, L., & Stemmler, M. (2013). cfa2: Configuration frequency analysis with a design matrix. Available at: http://www.rforge.net/cfa2/. Schulze, R. (1991). Getting round to (a)round: Towards the description and analysis of a ‘spatial’ predicate. In G. Rauh (Ed.), Approaches to prepositions (pp. 253–74). Tubingen: Günter Narr. Sheather, S. (2009). A modern approach to regression with R. New York: Springer. DOI: 10.1007/978-0-387-09608-7 Smith, R. (2011). Multilevel modeling of social problems: A causal perspective. Heidelberg: Springer. DOI: 10.1007/978-90-481-9855-9 Speelman, D., & Geeraerts, D. (2010). Causes for causatives: The case of Dutch ‘doen’ and ‘laten’. In T. Sanders, & E. Sweetser (Eds.), Causal categories in discourse and cognition (pp. 173– 204). Berlin & New York: Mouton de Gruyter.


Speelman, D., Tummers, J., & Geeraerts, D. (2009). Lexical patterning in a Construction Grammar: The effect of lexical co-occurrence patterns on the inflectional variation in Dutch attributive adjectives. Constructions and Frames, 1, 87–118. DOI: 10.1075/cf.1.1.05spe Stefanowitsch, A. (2010). Empirical Cognitive Semantics: Some thoughts. In D. Glynn, & K. Fischer (Eds.), Quantitative Cognitive Semantics: Corpus-driven approaches (pp. 355– 380). Berlin & New York: Mouton de Gruyter. Stefanowitsch, A., & St. Th. Gries. (2003). Collostructions: Investigating the interaction of words and constructions. International Journal of Corpus Linguistics, 8, 209–243. DOI: 10.1075/ijcl.8.2.03ste Stefanowitsch, A., & St. Th. Gries. (2005). Covarying collexemes. Corpus Linguistics and Linguistic Theory, 1, 1–43. DOI: 10.1515/cllt.2005.1.1.1 Stefanowitsch, A., & St. Th. Gries. (2008). Register and constructional meaning: A collostructional case study. In G. Kristiansen, & R. Dirven (Eds.), Cognitive Sociolinguistics: Language variation, cultural models, social systems (pp. 129–152). Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110199154.2.129 Stefanowitsch, A., & Gries, St. Th. (Eds.). (2006). Corpus-based approaches to metaphor and metonymy. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110199895 Stevens, J. 2001. Applied multivariate statistics for the social sciences (4th ed.). Mahwah: Lawrence Erlbaum. Strobl, C., Hothorn, T., & Zeileis, A. (2009a). Party on! A new, conditional variable importance measure for random forests available in the party package. The R Journal, 1, 14–17. Strobl, C., Malley, J., & Gerhard T. (2009b). An introduction to recursive partitioning: Rationale, application, and characteristics of classification and regression trees, bagging, and random forests. Psychological Methods, 14, 323–348. DOI: 10.1037/a0016973 Suzuki, R. (2013). Hierarchical clustering with p-values via multiscale bootstrap resampling. Available at: http://cran.r-project.org/web/packages/pvclust/pvclust.pdf. Suzuki, R., & Hidetoshi, S. (2006). Pvclust: an R package for assessing the uncertainty in hierarchical clustering. Bioinformatics, 22, 1540–1542. DOI: 10.1093/bioinformatics/btl117 Szelid, V, & Geeraerts, D. (2008). Usage-based dialectology: Emotion concepts in the Southern Csango dialect. Annual Review of Cognitive Linguistics, 6, 23–49. DOI: 10.1075/arcl.6.03sze Szmrecsanyi, B. (2003). Be going to versus will/shall: Does syntax matter? Journal of English Linguistics, 31, 295–323. DOI: 10.1177/0075424203257830 Szmrecsanyi, B. (2006). Morphosyntactic persistence in spoken English: A corpus study at the intersection of Variationist Sociolinguistics, Psycholinguistics, and Discourse Analysis. Berlin & New York: Mouton de Gruyter. DOI: 10.1515/9783110197808 Szmrecsanyi, B. (2010). The English genitive alternation in a cognitive sociolinguistic perspective. In D. Geeraerts, G. Kristiansen, & Y. Peirsman (Eds.), Advances in Cognitive Sociolinguistics (pp. 141–166). Berlin & New York: Mouton de Gruyter. Szmrecsanyi, B. (2013). Grammatical variation in British English dialects. Cambridge: Cambridge University Press. Tabachnick, B., & Fidell, L. (2007). Using multivariate statistics (5th ed.). London: Pearson. Taboada, M., & Carretero, M. (2012). Contrastive analyses of evaluation in text: Key issues in the design of an annotation system for attitude applicable to consumer reviews in English and Spanish. Linguistics and the Human Sciences, 6, 275–295. Tarling, R. (2009). Statistical modelling for social researchers: Principles and practice. London & New York: Routledge.

340 Dylan Glynn

Therneau, T., Atkinson, E., & Foundation, M. (2013). An introduction to recursive partitioning using the RPART routines. Available at: http://cran.r-project.org/web/packages/rpart/ vignettes/longintro.pdf. Thompson, L. (2009). S-PLUS (and R) manual to accompany Agresti’s categorical data analysis (2002). Available at: home.comcast.net/~lthompson221/Splusdiscrete2.pdf. Tummers, J., Heylen, K., & Geeraerts, D. (2005). Usage-based approaches in Cognitive Linguistics: A technical state of the art. Corpus Linguistics and Linguistic Theory, 1, 225–261. DOI: 10.1515/cllt.2005.1.2.225 Valenzuela Manzanares, J., & Rojo López, A. M. (2008). What can language learners tell us about constructions? In S. De Knop, & T. De Rycker (Eds.), Cognitive approaches to pedagogical grammar? A volume in honour of René Dirven (pp. 197–230). Berlin & New York: Mouton de Gruyter. Van Bogaert, J. (2010). A constructional taxonomy of I think and related expressions: Accounting for the variability of complement-taking mental predicates. English Language and Linguistics, 14, 399–428. DOI: 10.1017/S1360674310000134 Venables, W., & Ripley, B. (2002). Modern applied statistics with S (4th ed.). Heidelberg: Springer. DOI: 10.1007/978-0-387-21706-2 Verdonik, D., Rojc, M., & Stabej, M. (2007). Annotating discourse markers in spontaneous speech corpora on an example for the Slovenian language. Language Resources and Evaluation, 41, 147–180. DOI: 10.1007/s10579-007-9035-7 von Eye, A. (2002). Configural frequency analysis: Methods, models, and applications. Mahwah: Erlbaum. von Eye, A., & Mair, P. (2008) A functional approach to configural frequency analysis. Austrian Journal of Statistics, 37, 161–173. von Eye, A, Mair, P., & Mun, E.-Y. (2010). Advances in configural frequency analysis. London: Guilford Press. von Eye, A, & Mun, E.-Y. (2013). Log-linear modeling: Concepts, interpretation, and application. Hoboken: John Wiley. Wiebe, J., Wilson, T., & Cardie, C. (2005). Annotating expressions of opinions and emotions in language. Language Resources and Evaluation, 39, 165–210. DOI: 10.1007/s10579-005-7880-9 Wiechmann, D. (2008). On the computation of collostruction strength: Testing measures of association as expressions of lexical bias. Corpus Linguistics and Linguistic Theory, 4, 253– 290. DOI: 10.1515/CLLT.2008.011 Wong, M. (2009). Gei constructions in Mandarin Chinese and bei constructions in Cantonese: A corpus-driven contrastive study. International Journal of Corpus Linguistics, 14, 60–80. DOI: 10.1075/ijcl.14.1.04won Wulff, S. (2003). A multifactorial corpus analysis of adjective order in English. International Journal of Corpus Linguistics, 8, 245–82. DOI: 10.1075/ijcl.8.2.04wul Wulff, S. (2006). Go-V vs. go-and-V in English: A case of constructional synonymy? In St. Th. Gries, & A. Stefanowitsch (Eds.), Corpora in Cognitive Linguistics: Corpus-based approaches to syntax and lexis (pp. 101–126). Berlin & New York: Mouton de Gruyter. Wulff, S. (2009). Rethinking idiomaticity: A usage-based approach. London: Continuum. Wulff, S. (2010). Marrying cognitive-linguistic theory and corpus-based methods: On the compositionality of English V NP-idioms. In D. Glynn, & K. Fischer (Eds.), Quantitative Cognitive Semantics: Corpus-driven approaches (pp. 223–238). Berlin & New York: Mouton de Gruyter.


Wulff, S., Stefanowitsch, A., & Gries, St. Th. (2007). Brutal Brits and persuasive Americans: Variety-specific meaning construction in the into-causative. In G. Radden, Köpcke, K.-M., Berg, Th., & Siemund, P. (Eds.), Aspects of meaning construction (pp. 265–281). Amsterdam & Philadelphia: John Benjamins. Zeschel, A. (2010). Exemplars and analogy: Semantic extension in constructional networks. In D. Glynn, & K. Fischer (Eds.), Quantitative Cognitive Semantics: Corpus-driven approaches (pp. 201–221). Berlin & New York: Mouton de Gruyter. Zhao, Y. (2013). R and data mining: Examples and case studies. Unpublished manuscript. Available at: http://www.rdatamining.com. Zlatev, J., & Andrén, M. (2009). Stages and transitions in children’s semiotic development. In J. Zlatev, M. Andrén, C. Lundmark, & M. Johansson Falck (Eds.), Studies in language and cognition (pp. 380–401). Newcastle: Cambridge Scholars.

Statistics in R First steps Joost van de Weijer and Dylan Glynn Lund University and University of Paris VIII

The R Project for Statistical Computing is one of the most comprehensive and widely used software options for statistical analysis. Moreover, it is open source, freely available and entirely cross-platform. It is for these reasons that the following chapters all employ R to demonstrate the application and interpretation of statistics. Like the commercially available software SAS, but unlike three other widely used suites (SPSS, Stata, and Statistica), R is principally used in command line. The need to work with commands rather than a graphical user interface can be a challenge for novice users, especially when combined with the task of learning statistics. However, commands given in a step-by-step fashion is arguably simpler than a graphic interface, which can overwhelm the novice user with options. This chapter is an introduction to R focusing on how to import data and make sure those data are in the correct format for analysis. Knowledge of each of these steps is assumed in the following chapters. Keywords: dataframe, cross-tabulation, formatting data, importing data

1. Installing R R is freely available from www.r-project.org. On the site, one will find the CRAN server link where the various versions (Linux, MacOSX and Windows) can be downloaded. Follow the instructions to install R onto a personal computer.

2. Commands Once R has been installed on the computer, you can launch it by double-clicking the program’s icon. A window entitled “R Console” should open. In this window, you will see some introductory text, followed by an empty line that starts with a “>” sign. This

344 Joost van de Weijer and Dylan Glynn

sign is called the prompt. Unlike many other programs, R does not have a graphical interface with drop-down menus or dialogue windows to do an analysis. Instead, you need to tell the program what to do by typing commands directly in the console window. Numerous examples of commands are given throughout this book. Commands can be simple or complex. An example of a command is given below. This command tells R to read a text file and thus load the data into its working memory. The actual command below will not work since the location of the file is not specified, but it serves to explain the principle behind using command line. mydata = read.table("dataframe.txt")

For many users new to R, working with commands is difficult at first. Commands have a very strict structure (called syntax). If they are not entered entirely correctly, they do not work. The source of the error can be small (a missing comma, for instance) and the resulting error message is usually not very helpful to a beginner trying to determine what the problem actually is. Nevertheless, writing and memorising commands gets easier through practice. A helpful strategy is to keep and maintain a personal collection of command examples that you have used before and which are useful for the kinds of analysis you typically carry out. Let us examine the above command line by identifying its constituent elements. The central element is the word read.table. This is one of many functions that are built in into R and it is used for importing external data. More examples of the read.table function are discussed/explained below in Section 4. A function, such as read.table, is a way of telling the program to do something, in this case load the data, which is in the format of a table, into R. Note that read.table is followed by an opening parenthesis and, a bit further on at the end, a closing parenthesis. The text within the parentheses is the name of a data file, enclosed in quotation marks. The action that is specified by the function is often performed ‘on something’; here, the data file called dataframe.txt which is stored on the computer. This part of the command is called the argument . Then notice the part to the left of the read.table command in the example above. The result of the action performed by the command is saved internally in R under the name mydata. Practically, this means that the content of the external data file is copied, and saved in R under a new name. This name is more or less arbitrarily chosen. There are some restrictions on the choice of names, but almost anything will do, as long as it starts with a letter and does not contain any special characters. The name could as well have been olddata, sunday2112, pilotstudy, or xxx. It is up to you to choose the name, choose something short and easy to remember. Data or results that are saved in R are usually referred to as objects. Once an object has been created, it will be available until the program is quit or until it is manually deleted from R. An object can be a single number, a series of numbers, a complete data file, a graph, the output of an analysis, and so on. These are all stored in the so-called

Statistics in R 345

workspace. In order to see a listing of what is in the workspace, type the command ls(). In order to delete or remove and object type remove() with the name of the object between the parentheses.

3. The data file Creating a data file that will be loaded into R is an important step in the analysis and one that often leads to confusion when first learning how to use R. The idea is to take the data, from whatever their source, and put them in a plain text file. This step needs extreme care since text files can include hidden formatting and other information that will prevent the data from being loaded. Typically, data are held in a spreadsheet file (such as those produced by MS Office’s Excel, Open Office’s Calc or iWork’s Numbers) or a database (such as those produced by MS Office Word’s Access, Open Office’s Base, and Apple’s FileMaker). When just beginning, the tabular layout of a spreadsheet is arguably the easiest. This is because when data are displayed in a spreadsheet, one can easily see whether all cells are complete, whether the columns are aligned properly and how many cases there are. Once one is confident that the data are all clear and there are no empty cells and so forth, one can copy and paste the data directly into a plain text file. It is important that the file itself does not contain any ‘formatting’. If the file contains formatting or invisible mark-up, R may not be able to read the file. There are many text-editing programs that can contain hidden formatting (such as WordPad for Windows and TextEdit for MacOSX). However, other text editors automatically ‘strip’ any formatting, hidden or otherwise (such as NotePad in Windows and TextWrangler in MacOSX). If the data are contained in a database, one needs to export the data to a plain text file. This option is also available in the spreadsheet programs. Exporting data helps to eliminate the problem of hidden formatting and mark up. If the data contain diacritics or non-roman characters, this option is the safest one because it is usually possible to specify the encoding used in the text file (unicode is the preferred option here). The data in the text file, whether taken from a database or a spreadsheet, normally exists in one of two formats: a flat or ‘raw’ dataframe form, where each row represents a case and each column represents a variable, and a cross-tabulation or a contingency table where this information is summarised numerically. An example of the flat or ‘raw’ dataframe format is illustrated in Table 1. In most spreadsheets and databases, the data can be saved as a tab or comma delimited textfile, which means that there is a tab or a comma between each column. On screen, the columns in the textfile may not look perfectly straight, but this does not necessarily mean there is a problem – remember, there is no formatting. R reads any type of column delimitation. In the commands, it is assumed the data are tab-delimited. The flat dataframe layout of the data normally results from the


Table 1.â•‡ Example of a flat dataframe verb tense person figurativity run past 1 literal run past 1 metaphor jog future 2 metaphor run past 2 literal literal skip past 2 jog future 1 literal

Table 2.â•‡ Example of a cross-tabulation jog run skip

future 2 0 0

past 0 3 1

pers.1 1 2 0

pers.2 1 1 1

literal 1 2 1

metaphor 1 1 0

manual analysis of examples in a database or spreadsheet. In this format, the numbers of occurrences are not indicated, but instead each occurrence is listed in a large ‘flat’ file. This format of the data is the preferred format for multivariate statistics. One of the most common problems faced when starting statistical analysis in R is that the annotation or concordance tool used to obtain the data exports the results in a numerical table (described below). Typically, such tables have omitted much of the important information and cannot be used in most multivariate analyses. It is important to make sure that whatever corpus tool being used the data can be exported to the flat dataframe format. The second typical format for data is a numerical tabulation. We refer to this as a cross-tabulation or a contingency table. In contrast to the raw dataframe, this format is a numerical summary of the data and is typically a result of manually counting occurrences or the results of questionnaires. However, the format can also be generated by a number of spreadsheet, database, annotation and concordance programs. An example of a cross-tabulation is shown in Table 2. Using the first column in Table 1 as the row names, the data in Table 2 are the equivalent to those above. It is important to note that this table would look considerably different if we were to take a different column in Table 1 and use it to create the row names. This data format is sometimes also referred to as a frequency table, contingency table, xtab, or pivot table.

Statistics in R 347

4. Importing the data into R 4.1

Importing data from a flat dataframe

The two formats of data are loaded into R differently. We will begin with the raw datafile. The most common command for importing data into R is read.table(), which was introduced in Section 1. Here we present this command again, but now with two additions. mydata = read.table("dataframe.txt", header=TRUE, sep="\t")

In this example, there are three arguments to the command rather than just one as in the earlier example. The arguments have been separated by commas. The first argument, "dataframe.txt" is the filename. The second argument, header=TRUE, indicates that the first line of the data file contains the variable names, as is the case in Table 1. If this argument had been omitted, then the fields in the first line of the data file would have been interpreted as values rather than as names, and the columns would have been labelled automatically instead (V1, V2, etc.). The third argument, sep="\t", indicates that the columns are separated by tabs. Had the columns been separated by commas instead, then the third argument would have been sep=",". The object that has been created with the read.table() command above is called a dataframe, and it has been named mydata. However, these commands will still not work – we have to add one further piece of information. We need to tell R where to find the data file. There are four possibilities: mydata = read.table(file.choose(), header=TRUE, sep="\t") mydata = read.table("clipboard", header=TRUE, sep="\t") mydata = read.table("users/linguist/data/dataframe.txt", header=TRUE, sep="\t") setwd("/Users/linguist/data") mydata = read.table("dataframe.txt", header=TRUE, sep="\t")

In the first alternative, the name of the file has been replaced by file.choose(). This argument causes a dialogue window to open from which the data file can be chosen. With the second alternative, it is assumed that the data have been copied to the clipboard and R takes the data from there. This is similar to the copy and paste function common in applications such as MS Word and Excel. Depending of the operating system and the version of R being used, the second option often results in problems. The third option tells R to go to a specific location and open the file. That location can be on the hard drive of a personal computer, on a server, or even on the Internet. The location is indicated as a path, where slashes “/” represent folders. This location


or path needs to be between inverted commas. In MacOSX and Linux, the command given above indicates that the data file is in a folder called ‘data’, which is in a folder called ‘linguist’, which in turn is in a folder called ‘Users’, which is on the boot drive of the system. In Windows, one needs to add the name of the disk or volume typically indicated by a lowercase letter followed by a colon, “c:” being the default label for a boot drive. Therefore, for Windows, the equivalent command line looks like this: mydata = read.table ("c:users/linguist/data/dataframe.txt", header=TRUE, sep="\t")

The fourth alternative, finally, consists of two commands. The first command sets the so-called working directory, which is the path to the folder in which the datafile is stored. The second command imports the data into a dataframe object. If you do not know what the current working directory is, you can find out by typing the command getwd(). The object that is being created with read.table() is called mydata. It is important to check that the dataframe created from a data file has been properly imported into R. It is not uncommon that something goes astray during the process of importing, and not always does this result in an error message in R. Therefore, next we provide examples of commands that can help make sure that the newly imported data is in order. The simplest way of seeing what the data looks like in R is by typing the name of the object. We called our data object mydata. mydata

Typing the name of the object, here mydata, will display the contents of the entire object. However, if the object is a flat dataframe, this will result in the entire dataset being displayed, which can be extremely cumbersome. In the case of a flat dataframe, a better option is to look at the first few rows, using the command head(): head(mydata)

This command will display the column names of the dataframe (these are the names that were in the first row of the data file; the headers in a spreadsheet and the cell labels in a database), followed by the first six rows. There is a corresponding command, tail(), that displays the last six rows of the dataframe. A second command that is useful in this regard is summary(). This command generates a numerical summary of the dataframe. This is also useful for spotting spelling mistakes and so forth in the analysis. Remember that R treats lowercase and uppercase letters as distinct items and does not ignore invisible characters such as a space or a tab. It is rare that a flat dataframe is without any mistakes. After importing data, one must typically return to the database or spreadsheet and correct many small typographic errors.

Statistics in R 349

summary(mydata)

A third command that provides information about a dataframe is the command str(). This command offers information about the structure of a dataframe. When applied to the data from Table 1, we receive the following output: str(mydata) 'data.frame' : $ verb : 2 2 1 2 3 1 $ tense : 2 2 1 2 2 1 $ person : $ figurativity: 1 2 2 1 1 1

6 obs. of 4 variables: Factor w/ 3 levels "jog","run","skip": Factor w/ 2 levels "future","past": int 1 1 2 2 2 1 Factor w/ 2 levels "literal","metaphor":

The first line of the output shows that mydata is a dataframe with six observations (rows) and four variables (columns). The next four lines show the four variables and three types of information about them. First of all, they show their names. Second they show what type of variables they are. Here, three of the four variables are labelled as Factor, which is the usual type for categorical variables. Additionally, the str() command shows how many levels these variables have, that is, the number of features that these variables possess. The variables tense and figurativity have two levels, while verb has three. The fourth variable, person, is labelled as int, which means that the values of that variable are whole numbers. The numbers 1 and 2 are labels for first and second person, respectively. This has revealed an important error. R has assumed that this variable is numerical because the labels used in the analysis were numbers. One solution is to add a letter to the label in the spreadsheet or database, the other is to change it in R. This is further explained in Section 5.3.

4.2

Importing cross-tabulations

If the data are in a cross-tabulated format, as exemplified in Table 2, loading the data is a little different. There are two ways to load the data in this case. Firstly, one can repeat the command used above in Section 4.1 to load the data from a flat dataframe, but add the argument row.names = 1. mydata = read.table(file.choose(), header=TRUE, sep="\t", row.names=1)

The read.table command is not designed for data in this format, so R does not always treat the data as it should (for example, the str() and summary() commands do not work). Nevertheless, for most purposes, loading the data in this way does not


Table 3.â•‡ Layout for read.ftable() command verb

future 2 0 0

jog run skip

past 0 3 1

1stPers 1 2 0

2ndPers 1 1 1

literal 1 2 1

metaphor 1 1 0

pose any problems. If one wishes to see that the cross-tabulation has been correctly loaded, enter the object name: mydata

This brings up the cross tabulation for inspection. A second, more orthodox, way to load a cross-tabulation is to use the command read.ftable(). Notice an “f ” has been added to the function read.table. mydata = read.ftable(file.choose(), sep="\t")

This command tells R that the data are in the cross-tabulated format. Note that the header=TRUE argument has been removed. The read.ftable command is sensitive to the actual layout of the data, especially to how the row and column names are placed in the text file. For this command to work, the name of the first row must be located on a previous line, independent from the column names. One must also remember to add a ‘blank’ first column beneath the first row-name. Table 3 exemplifies this layout. Both commands, read.table and read.ftable(), expect a blank line (a return carriage) at the end of the text file, beneath the table.

4.2.1 Transpose a contingency table Sometimes it is useful to transpose a cross-tabulation. Transposition means that the entire data object is rotated by 90 degrees. That is, the rows become columns and the columns become rows. For a raw dataframe, this is rarely useful, but for cross-tabulations, since the data are summarised as a given variable and its levels are relative to another variable (or variables and levels), inverting the table means that a statistical technique may actually examine the data from a different perspective. Transposition can be done in R using the command t(), as in the following example: mydata2 = t(mydata)

The object mydata2 is now an inverted or transposed version of the of the original cross-tabulation.

Statistics in R 351

5. Making changes to a dataframe in R It is not uncommon that information in a dataframe needs to be changed or that new information needs to be added. A possible way of doing this is to make changes directly in the data file and to re-load the data, but for various reasons it might be desirable to keep the data file unchanged, and to do the modifications to the dataframe in R instead.

5.1

Creating objects

To begin, the principle of creating objects in R needs to be explained. Two synonymous signs are used for creating object, “=” and “ New Document from the R menu. A new window will open in which one can add commands, exactly as they are typed in the console window. An example is shown below: mydata=read.table("datafile.txt") str(mydata) summary(mydata)

Commands can be run directly from the script window. Click once on the command that one wishes to run, then type command-Enter (Macintosh) or Control-r (Windows). The command and its output are then automatically displayed in the Console window. There are several advantages of working with scripts. One advantage is that one builds up a collection of scripts with commands that are often used. A second advantage is that scripts can contain personal comments. Comments can be anything from the date that the script was created, to the purpose of the script or the explanation of a complex command. The following is an example of a very short script, which includes the example command that we saw above, with a comment added giving information on the date the script was created: # Example R-script, created December 2011 mydata=read.table("datafile.txt")

Statistics in R 361

The first line starting with the hash-sign (“#”) is the comment. The second line is the command. Comments have no effect if run, they are there only for additional information or explanation and are ignored by R. A third advantage of working with scripts is that script files allow one to write very long commands on multiple lines. Commands to make plots, typically, become very long. Writing them on separate lines (as was shown above in Section 6) usually makes the structure of the command more transparent, and makes it easier, if there were something wrong with the command, to spot the error. Finally, R offers the possibility of syntax colouring, which also can help with seeing the structure in R commands. Syntax colouring means that different parts of the script are shown in different colours. A specific colour is reserved for a comment, another colour for a command keyword, a third colour for numbers, and so on. For these reasons, script files can be a great help in learning to work with R. Script files are normally saved on the computer with the extension .R.

9. Extending functionality with packages When R is installed for the first time it contains many basic functions for doing analyses and making plots. These base functions can be complemented with other functions that are geared towards more special-purpose analyses, such as the ones described in this book. Many so-called packages exist that contain functions used within a specific discipline, or for producing specific types of plots, or doing specific types of analysis. An incredibly rich collection of packages for performing all kinds of statistical analyses exists, and this collection is constantly being improved upon and added to. The packages are small and download quickly. To use a package, it must first be installed, that is, it must be downloaded from the R-website and saved to the computer. One way of installing a package is with the install.packages() command. The following example shows how to install the package called ca, for doing correspondence analysis described in Glynn (this volume): install.packages("ca")

This will firstly produce a prompt asking the user to choose a server (CRAN mirror) and a list of different options will appear. Once a server has been selected, the package can be installed. Another way to install packages is to the use the menu options. Under MacOSX, it will be found under the menu “Packages & Data > Package Installer” and under Windows “Packages > Install packages(s)”. Under MacOSX, choose the server and then simply type the first letters of the name of the package and it should appear. Then


select “install”. It is recommended that the option in the radio button “install dependencies” be checked. Under Windows, the procedure is the same except that one must scroll through a long list to locate the desired package but install dependencies is set to default and does not be worried about. An extremely common point of confusion for new users is the difference between installing and loading packages / libraries. The packages contain the libraries that R needs to perform its tasks. One only needs to install a package once on a computer, but the library associated with it, must be loaded each time R is booted. Loading a package can be done with the library() command: library(ca)

Once loaded, a package is available until R is quit. Remember that it must be re-loaded with the library() command each time. Error messages that result from unloaded packages can be one of the most frustrating experiences for novice users. A good way of preventing this from happening is to add the library() command to the R-script.

10. Going further This short introduction is designed to help new users get started with R. There are many things to discover in R, and it will probably take some time to get a good grasp of the kinds of commands, packages and graphs that are useful for the kind of analysis for which one wants to use R. To conclude, we offer three pieces of advice that we found helpful on the way of becoming experienced R users. First, as mentioned above in Section 7, there is a built-in help function that shows information about the syntax of a command, some examples, and often links to other related commands. A call for help on a command is obtained with help() or with ?(), placing the command in parentheses. Second, the Internet is a good place to search for help. There are numerous sites with blogs, tutorials, and user fora. Here one can find R code, pose questions, and see graphs. For the less experienced users, we can recommend the site Quick-R, which contains many clear examples. Furthermore, the R-website also offers a manual. Finally, the number of books on statistics using R within many different disciplines grows steadily. Books for linguistic analysis are Baayen (2008), Dalgaard (2008), Johnson (2008) and Gries (2009, 2012). Focusing on graphics, Keen (2010) and Mittal (2011) are accessible to beginners and are relatively complete. Other introductory books include Crawley (2007), Everitt & Hothorn (2009), Maindonald & Braun (2010) and Adler (2010).

Statistics in R 363

References Adler, J. (2010). R in a nutshell: A desktop quick reference. Sebastopol: O’Reilly Media. Baayen, H. (2008). Analyzing linguistic data: A practical introduction to statistics using R. Cambridge: Cambridge University Press. DOI: 10.1017/CBO9780511801686 Crawley, M. (2007). The R book. Chichester: John Wiley. DOI: 10.1002/9780470515075 Dalgaard, P. (2008). Introductory statistics with R (2nd ed.). Dordrecht: Srpinger. DOI: 10.1007/978-0-387-79054-1 Everitt, B. S., & Hothorn, I. (2009). A handbook of statistical analyses using R (2nd ed.). Boca Raton: Taylor & Francis. Gries, St. Th. (2009). Quantitative corpus linguistics with R: A practical introduction. London: Routeledge. DOI: 10.1515/9783110216042 Gries, St. Th. (2012). Statistics for linguistics with R: A practical introduction. Berlin & New York: Mouton de Gruyter. Keen, K. (2010). Graphics for statistics and data analysis with R. Boca Raton: CRC Press. Johnson, K. (2008). Quantitative methods in linguistics. Oxford: Blackwell. Maindonald, J., & Braun, J. (2010). Data analysis and graphics using R (3rd ed.). Cambridge: Cambridge University Press. Mittal, H. V. (2011). R graphs cookbook. Birmingham: Packt.

Appendix: The tablebind-script This script can be copied or carefully entered into R and saved as a function, explained in Section 7. The script and the data file can be downloaded from http://dx.doi.org/10.1075/hcp.43. 13wei.additional. tablebind=function(df.flat) { if(ncol(df.flat)|z|) (Intercept) 0.35333 0.32786 1.078 0.281 ageC -0.08591 0.01950 -4.406 1.06e-05 *** […]

Logistic regression 517

This analysis is identical to the previous one in every detail except (a) that the name of the age predictor has changed and (b) that the intercept now refers to the fitted logit for a person with a centred age of zero (i.e. with an age of 51.6). We see that according to the model, a person of an age of 51.6 has an average logit of 0.35 (which corresponds to odds of 1.4 to 1 and a proportion of about 59%), which, according to the significance test, is not significantly different from a logit of zero (i.e. from odds of 1 to 1 and a proportion of 50%). Going back to the model with regular age as predictor and moving to the far more important estimate which is the slope age, we see that according to the model for each year that is added to the age, 0.09 is subtracted from the average logit (and conversely for each year that is subtracted from the age, 0.09 is added to the average logit). The odds are requested as follows: > exp(coef(fig1Mid.glm)) (Intercept) age 119.8829912 0.9176727

At age zero, the odds are 120 (i.e. 120 to 1) and with each additional year these odds are multiplied by 0.91. The 95% confidence intervals for the estimates are requested as follows: > confint(fig1Mid.glm) Waiting for profiling to be done… 2.5 % 97.5 % (Intercept) 2.8377877 7.32063059 age -0.1296267 -0.05201228

Therefore, we are 95% confident that the correct slope for age, on the logit scale, is between –0.13 and –0.05. The 95% confidence intervals for the odds are requested as follows: > exp(confint(fig1Mid.glm)) Waiting for profiling to be done… 2.5 % 97.5 % (Intercept) 17.0779426 1511.1565962 age 0.8784232 0.9493172

We see now very clearly that there is quite a lot of uncertainty regarding the intercept (the confidence interval for the intercept covers probabilities from 94% up to 99.9%) but that does not concern us here, since we zoom in on the effect of age. We see that we are 95% confident that the odds ratio between consecutive years is between 88% and 95%.

518 Dirk Speelman

# figure 6 cdplot(variant ~ age, data=fig1Mid, ylevel=2:1) rug(jitter(fig1Mid$age), col=“white”) # figure 7 plot(Predict(fig1Mid.lrm, age, fun=plogis))

Old

1.0

0.8

Young

Variant

0.6

0.4

0.2

20

40

60 Age

80

0.8

0.6

0.4

0.2

0.0 20

40

Age

60

80

Figures 6 and 7.â•‡ Observed (left) and fitted (right) effect of age on variant choice

Figures 6 and 7, which were generated with the code above, enables us to see what all this means on the proportions of success scale. The left plot has nothing to do with the regression model. It is a conditional density plot of our data. A conditional density plot is related to a spinogram [spineplot()] in the way a density plot relates to a histogram. A conditional density plot depicts the observed relative proportion of the two variants for different age groups, but instead of selecting a discrete set of age groups, it applies a smoothing technique that enables us to see the gradual transition of this proportion along the age axis. At the bottom of the plot, we have also added the individual items (with their correct ages, but with some horizontal jitter added so that the items do not hide each other too much) so that we can see how much information is behind the different parts of the conditional density plot. The right plot is generated with plot(Predict(fig1Mid.lrm, age, fun=plogis)). This instruction plots the effect of age on the proportions of success scale and also plots 95% confidence margins around the fitted averages (the argument fun=plogis causes the plot to be on the proportions scale instead of on the logit scale). We see that the model rather nicely mirrors the pattern that is observed in the conditional density plot.


4.2

Output for multiple logistic regression

For the discussion of multiple logistic regression in R, we introduce a new non-fictitious data set. The object noemen contains 446 instances of the noemen/heten alternation pattern that were found in the spontaneous speech section of the Belgian part of the Spoken Dutch Corpus (Oostdijk 2000). In informal Belgian Dutch, the verb noemen is sometimes used with the meaning ‘to be called’, as in (1) and (2). The other variant, heten, as in (3) and (4), is the variant that most speakers in Belgium would call ‘the correct variant’ (and is the only one that is used in The Netherlands; the verb noemen is very well known in The Netherlands, but is only used with the meaning ‘to call’, as in (5), which, in Belgium too, is the most dominant use of noemen). The analysis we discuss here only contains cases such as (1) to (4). Cases such as (5), in which noemen has a very different meaning, were excluded from the analysis. 1. Ik noem Dirk. (I’m called Dirk.) 2. Dat noemt heteroscedasticiteit. (That’s called heteroscedasticity.) 3. Ik heet Dirk. (I’m called Dirk.) 4. Dat heet heteroscedasticiteit. (That’s called heteroscedasticity.) 5. Ik noem dat onzin. (I call this nonsense.) We use head() to show the structure of the data. > head(noemen) variant conver.type generation region occup.type 1 heten face.to.face from.1980 LI neutral 2 heten face.to.face from.1980 AB higher.edu 3 heten face.to.face from.1980 AB higher.edu 4 heten face.to.face from.1980 AB higher.edu 5 heten face.to.face before.1980 OV higher.edu 6 heten face.to.face before.1980 LI higher.edu

The variable variant is the response variable, with levels heten and noemen (noemen will be the success level in the analyses). The predictor conver.type (conversation type) has levels face.to.face and telephone (face.to.face will be the reference level in the analyses). The predictor generation is a rough classification of the speakers in two age groups: the reference level is before.1980 (born before 1980) and the other level is from.1980 (born in 1980 or later). All recordings in the corpus are roughly from the same period (around 2000), so what we try to find with the predictor generation is an apparent time effect. Ideally, we would have liked to have used age as a numerical predictor, but we restrict our analysis to a simpler categorical predictor.18 The predictor region represents the region of birth 18. Although in general it is better to use the ‘richer information’ in a numerical variable age, rather than the ‘simplified information’ in a categorical variable generation, we refrained

520 Dirk Speelman

of the speakers. Values are AB (the central provinces Antwerp and Brabant), LI (the province Limburg in the east of the country), WV (the province West-Flanders in the west of the country), and OV (the province East-Flanders in between Antwerp and West-Flanders). The reference value is OV. Finally, the variable occ.type (occupation type) is a coarse-grained classification of the occupation types of the speakers into the categories higher.edu (profession requires higher education), no.higher.edu (profession does not require higher education), and neutral (professions for which the link to higher education is either variable or unknown; obviously this is a ‘rest category’). The level higher.edu is the reference value. We start by running a model with all predictors (but without interactions), and we call the resulting model noemen.glm.1 (glm output) and noemen.lrm.1 (lrm output). The less important or redundant parts in the output have been suppressed. > noemen.glm.1 summary(noemen.glm.1) […] Null deviance: 618.06 on 445 degrees of freedom Residual deviance: 506.24 on 438 degrees of freedom AIC: 522.24 […] > noemen.dd options(datadist=“noemen.dd”) > noemen.lrm.1 noemen.lrm.1 […]

Obs 446 heten 228 noemen 218

Model Likelihood Ratio Test LR chi2 111.82 d.f. 7 Pr(> chi2) |Z|) 0.0026 noemen.glm.2 summary(noemen.glm.2) […] Null deviance: 618.06 on 445 degrees of freedom Residual deviance: 482.34 on 435 degrees of freedom AIC: 504.34 […] > noemen.lrm.2 noemen.lrm.2 […] Model Likelihood Ratio Test Obs 446 LR chi2 135.73 heten 228 d.f. 10 noemen 218 Pr |Z|) Intercept -0.8518 0.2018 -4.22 + + + + + > + +

predict.info plot(allEffects(noemen.glm.2))


1 2 3 4 5 6 7 8

Group.1 OV WV AB LI OV WV AB LI

Group.2 before.1980 before.1980 before.1980 before.1980 from.1980 from.1980 from.1980 from.1980

x.mean 0.66666667 0.45833333 0.31395349 0.06250000 0.59459459 0.67647059 0.77777778 0.59459459

x.sd 0.18788276 0.18866238 0.19044277 0.04043570 0.16932037 0.09336409 0.09800288 0.10651400

x.n 63.00000000 96.00000000 86.00000000 48.00000000 37.00000000 34.00000000 45.00000000 37.00000000

Incidentally, the code we just used can easily be modified to generate the list of predicted probabilities for all categories in the data (i.e. all possible combinations of values of the predictors that occur in the data). We do this below, but we suppress most of the output. There are 43 lines in the output because 5 of the 48 theoretically possible combinations do not occur in the data. > aggregate(predict(noemen.glm.2, type=“response”), + list(noemen$region, noemen$generation, + noemen$conver.type, noemen$occup.type), + predict.info) Grp.1 Group.2 Group.3 Group.4 x.mean x.sd x.n 1 OV before.1980 face2face higher.edu 0.43974421 0.00 23.00 2 WV before.1980 face2face higher.edu 0.24715753 0.00 35.00 3 AB before.1980 face2face higher.edu 0.15073898 0.00 45.00 4 LI before.1980 face2face higher.edu 0.02497667 0.00 18.00 […] 42 AB from.1980 telephone neutral 0.89284457 0.00 8.00 43 LI from.1980 telephone neutral 0.74984350 0.00 6.00

5. Model diagnostics In linear regression, which relies heavily on the assumption of normal errors, a lot of model diagnostics exist that involve inspecting the residuals in the model and looking for indications that the model assumptions are violated. Compared to that, the model assumptions in logistic regression are much less demanding, but still there is a clear need to test if nothing is wrong in the way the model fits the data. The assumption of a binomial distribution of the response implies certain expectations about the amount of variability we can expect in the data given a specific model. If we detect an amount of variability that is not in line with these expectations – a phenomenon which is called underdispersion when we detect less variation than expected, and overdispersion if we detect more variation than we expect (overdispersion is, obviously, the more common problem) – then this undermines the assumption. A first quick check we can do to test for this is to compare the residual deviance in the

528 Dirk Speelman

model to the residual degrees of freedom in the model; should the residual deviance be much larger than the residual degrees of freedom, then that would be an indication of overdispersion. In noemen.lrm.2, this comparison gives no indication of overdispersion (Residual deviance: 482.34 on 435 degrees of freedom). Second, we can perform the so-called Hosmer-Lemeshow-Cessie goodness of fit test. A significant result in that test would indicate a significant lack of fit. We do not obtain a significant lack of fit here (p = 0.0755884), but the rather low p-value does indicate that the fit is far from good. > residuals(noemen.lrm.2, type=”gof”) Sum of squared errors Expected value|H0 80.6027138 81.3454554 Z P -1.7768775 0.0755884

SD 0.4180039

If any of the above techniques signals a problem, the most likely reason is that important predictors are missing from the model (or that the model is for some other reason too simplistic to provide an acceptable fit). Another thing that should be tested routinely is whether there are any correlation patterns among the predictors that are so outspoken that this makes the model less reliable (this is the so-called problem of multicollinearity). We can test this with the command vif() which we run on the lrm model without interactions. Values higher than four should be considered problematic. If they occur, one of the problematic variables should be removed from the model and the model should be rerun and tested again.21 > vif(noemen.lrm.1) conver.type=telephone 1.178568 region=AB 1.689968 occup.type=no.higher.edu 1.386220

generation=from.1980 1.674683 region=LI 1.626993 occup.type=neutral 1.654916

region=WV 1.724891

Finally, it can also happen that individual items in your data set have too big an influence on your model. Candidate problematic cases can be detected as shown below, together with the estimate they could affect the most (the numbers are row numbers in the data set). Inspection of these cases involves manual inspection as well as running a model with these cases removed and testing whether indeed the models differ

21. An alternative solution would be to use a dimension reduction technique to remove the redundancy from the predictors.


substantially. The decision to permanently remove certain items from the analysis should only be a last resort (and obviously should be well documented). > which.influence(noemen.lrm.2) $Intercept [1] “265” “267” “268” “270” “271” $generation [1] “151” $region [1] “238” “239” “450” $`generation * region` [1] “188” “238” “239” “450”

6. Variable selection Logistic regression is a powerful and versatile tool, but it is not a very easy technique to use. One of the more difficult aspects of its use is variable selection. Variable selection, sometimes also called model selection, is the process of deciding which of the candidate predictors or interactions between predictors to include in the (final) model. From a ‘mechanical hypothesis testing perspective’, it might seem that this is not a process, but simply is a ‘one step procedure’ that boils down to including all candidate predictors to the model and running the model. However, in just about any non-trivial study, the task is more complicated than that and issues of the fit, interpretability and replicability of the model arise. On the one hand, the comparison of models often helps us acquire a deeper understanding of the relation between the model(s) and the data. On the other hand, we must be careful that the very process of multiple model comparison does not trick us into ‘data fishing’. In sum, variable selection is the sometimes-tricky task of striking a balance between on the one hand misreading or oversimplifying the patterns in the data and on the other hand overfitting to the ‘noise’ in the data. An interesting reference regarding this topic is Harell (2001).

7. Which conditions should my data set meet? It has been mentioned before that logistic regression analysis can handle data from corpus studies as well as from experimental designs. In the case of corpus studies, most of the time the best approach to collecting your data is to retrieve all occurrences of the phenomenon under scrutiny from your corpus and include all these occurrences in your data set (of course, you do exclude all spurious hits and all cases with contexts in which variation is not possible). In general, you do not have to try to balance your data set. Firstly, logistic regression can handle a lot of correlation patterns among the

530 Dirk Speelman

predictors. Secondly, in logistic regression, variable levels within one variable do not have to be balanced, neither for the response variable nor for the predictors. The only downside of unbalanced data is that unbalanced data often require larger data sets. This brings us to the following point. How many data points do I need for my study? This crucially depends on the number of regressors you consider for inclusion in the model. Several rules of thumb exist. Here I only mention one general rule of thumb, which is the one I believe is most important. The rule starts from the overall number of occurrences in the data set for the least frequent response level. For instance, in the noemen data set this would be 218, the overall frequency of the variant noemen (the other variant, heten, has a frequency of 228). The rule of thumb then says that the maximum number of regressors one should consider for inclusion in the model must not be more than this frequency divided by 20. So, in the noemen data set, the maximum number of regressors must not be more than 10 or 11 (see Harrell (2001:â•›60–61) and the studies that are cited there, for a more detailed and nuanced discussion of sample size issues). The biggest risk related to the inclusion of too many regressors is overfitting. Overfitting is fitting a model that follows the patterns in the data so closely that it picks up idiosyncratic as well as generalizable patterns. An overfitted model will not be reproducible in replication studies. On the one hand, it will typically exaggerate the size of effect that does exist, and on the other hand, it will also detect pseudo-effects that are actually no more than accidental idiosyncratic irregularities in the data set. Next to rules of thumb such as the one just mentioned, there are other, more advanced, techniques that help reduce the risk of overfitting (see e.g. Harell 2001 on cross-validation and shrinkage).

8. Beyond the limits of traditional binomial logistic regression In a sense, it is inappropriate to include a section called ‘Beyond the limits of traditional binomial logistic regression’ in an introductory text that has barely scratched the surface of what can be done within traditional binomial logistic regression. Moreover, just like any other popular modern technique in statistics, logistic regression is constantly being enhanced, and therefore the boundaries of what is ‘traditional binomial logistic regression’ are unclear ones and they are probably shifting. Still, there is a need to position what is discussed in this text in a wider context and point to some techniques that extend binomial logistic regression or can be seen as alternatives to it. The first thing that springs to mind is multinomial logistic regression, which is a further elaboration of logistic regression that supports categorical response variables with more than two levels, and proportional odds models (also called ordinal logistic regression), which support ordinal response variables.


A second extension is the incorporation into generalized linear models (and therefore also in logistic regression models) of mixed model analysis. In mixed model terminology, all predictors that were discussed in this text are so-called fixed effect predictors. Mixed models allow for the combination in one model of both fixed effect predictors and so-called random effects (also called random factors), which are identifiable sources of random variation. Random effects are useful for many contexts, but let me give just one example. Suppose that in the data set for some Corpus Linguistics study you have 1000 items that were produced by 200 different speakers. In other words, many of these speakers have produced several items in your data. Obviously this fact can introduce certain patterns in the data (a type of interdependence between certain observations) because behaviour by the same speaker is likely to be more similar than behaviour by different speakers. Ignoring this fact when analysing the data may lead to biased results and incorrect conclusions. On the other hand, one does not want to treat speaker as a fixed effect because (a) it would have 200 levels which would make it both a terribly ‘dominant’ and a terribly uninterpretable predictor, and (b) fitting fixed effect slopes for individual speakers would not be interesting and would not lead to replicable studies or general, reproducible results (in other words, the model would overfit to irrelevant idiosyncratic patterns in the data). This is an example where ‘speaker’ would be a good candidate for a random factor. Contrary to the levels of a fixed effect predictor, the levels of a random factor that occur in a data set are assumed to be a random selection from a typically much larger set of levels in the population. In other words, if we replicate a study, the levels of the fixed effect will be the same as in the original study. The levels of a random factor, however, will typically be another random sample from the large set of levels in the population. In the case we just discussed, it will typically be other speakers. Mixed models then are designed to account for the variation caused by random factors without giving them the status of fixed effects. Yet another avenue is the exploration of non-linear relations between numerical or ordinal predictors and logits. The assumption of linearity is hardly ever obvious. Often it is an acceptable simplification of reality, often it is not. Within traditional binomial regression, we can use polynomials and transformed predictors. Beyond it, we can explore splines for logistic models and generalized additive logistic models. Finally, it should be mentioned that beyond generalized linear models (either with or without random effects) there are many other statistical tools that can be used for classification (i.e. for predicting class membership of items on the basis of the properties of the items) and that the types of analyses discussed in this text can be seen as classification tasks with the classes being the levels of the response variable. Examples of such techniques are classification trees and forests (a.o. conditional inference trees and random forests), support vector machines, memory based learning, Bayesian networks, and many others.

532 Dirk Speelman

9. Further reading A gentle, yet broader introduction to logistic regression can be found in Pampel (2000). Some more advanced textbooks on logistic regression are Hosmer and Lemeshow (2000), and Hilbe (2009). More information on the VARBRUL tradition of logistic regression in Sociolinguistics can be found in Paolillo (2002) and Tagliamonte (2006), amongst others.

References Arnold, J., Wasow, Th., Losongco, A., & Ginstrom, R. (2000). Heaviness vs. newness: The effects of complexity and information structure on constituent ordering. Language, 76, 28–55. Berkson, J. (1944). Application of the logistic function to bio-assay. Journal of the American Statistical Association, 39, 357–365. Cedergren, H., & Sankoff, D. (1974). Variable rules: Performance as a statistical reflection of competence. Language, 50, 33–56. DOI: 10.2307/412441 Cox, D. R. (1969). The analysis of binary data. London: Chapman and Hall. Fox, J. (2003). Effect displays in R for generalised linear models. Journal of Statistical Software, 8(15), 1–27. Retrieved from . Grondelaers, S., Speelman, D., & Geeraerts, D. (2002). Regressing on er. Statistical analysis of texts and language variation. In A. Morin, & P. Sébillot (Eds.), 6èmes journées internationales d’analyse statistique des données textuelles (pp. 335–346). Rennes: Institut National de Recherche en Informatique et en Automatique. Harrell, F. E. (2001). Regression modeling strategies: With applications to linear models, logistic regression, and survival analysis. Berlin: Springer. Hilbe, J. M. (2009). Logistic regression models. London: Chapman & Hall/CRC Press. Hosmer, D., & Lemeshow, S. (2000). Applied logistic regression (2nd ed.). New York: Wiley. DOI: 10.1002/0471722146 Johnson, D. E. (2008). Getting off the GoldVarb standard: Introducing Rbrul for mixed-effects variable rule analysis. Language and Linguistics Compass, 3, 359–83. DOI: 10.1111/j.1749-818X.2008.00108.x Keune, K., Ernestus, M., van Hout, R., & Baayen, H. (2005). Social, geographical, and register variation in Dutch: From written mogelijk to spoken mok. Corpus Linguistics and Linguistic Theory, 1, 183–223. DOI: 10.1515/cllt.2005.1.2.183 Nelder, J., & Wedderburn, R. (1972). Generalized linear models. Journal of the Royal Statistical Society: Series A, 135, 370–384. DOI: 10.2307/2344614 Oostdijk, N. (2000). The spoken Dutch corpus: Overview and first evaluation. In S. Markantontou, S. Piperidis, & G. Stainhauoer (Eds.), Proceedings of the second international conference on language resources and evaluation (pp. 887–893). Athens: Institute for Language and Speech Processing. Pampel, F. C. (2000). Logistic regression: A primer. Thousand Oaks, CA: Sage.


Paolillo, J. (2002). Analyzing linguistic variation: Statistical models and methods. Stanford: CSLI. Sankoff, D. (1988). Variable rules. In U. Ammon, N. Dittmar, & K. J. Mathheier (Eds.), Berlin sociolinguistics: An international handbook of the science of language and society, Vol. 2. (pp. 984–997). Berlin & New York: Walter de Gruyter. Sankoff, D., Tagliamonte, S., & Smith, E. (2005). Goldvarb X: A variable rule application for Macintosh and Windows. Department of Linguistics, University of Toronto. Tagliamonte, S. A. (2006). Analysing sociolinguistic variation. Cambridge: Cambridge University Press. DOI: 10.1017/CBO9780511801624 Williams, R. S. (1994). A statistical analysis of English double object alternation. Issues in Applied Linguistics, 5, 37–58. Wilson, E. B., & Worcester, J. (1943). The determination of L. D. 50 and its sampling error in bio-assay. Proceedings of the National Academy of Sciences, 29, 257–262. DOI: 10.1073/pnas.29.8.257

Name index

A Adler, Jeffreyâ•… 315, 317, 318, 319, 322, 362 Afifi, Abdelmonemâ•… 316, 318, 322 Agresti, Alanâ•… 320, 321, 322, 323, 482 Aijmer, Karinâ•… 184, 185, 281 Albert, Jamesâ•… 323 Aldenderfer, Markâ•… 99 Allan, Keithâ•… 88 Allerton, Davidâ•… 147, 148, 149, 150, 174 Altenberg, Bengtâ•… 149 Alviar, Joséâ•… 406 Amberber, Mengistuâ•… 224 Andrén, Matsâ•… 310 Apresjan, Juriâ•… 10 Arnold, Jenniferâ•… 259, 488 Arppe, Anttiâ•… 14, 122, 187, 310, 317, 318, 322, 325, 327, 482 Athanasiadou, Angelikiâ•… 147, 149, 152, 174 Atkins, Berylâ•… 22, 310 Azen, Raziaâ•… 321, 322 B Baayen, Haraldâ•… 136, 137, 218, 272, 315, 316, 317, 318, 319, 321, 323, 324, 325, 326, 327, 362, 382, 405, 407, 435, 460, 482 Backhaus, Klausâ•… 435 Baguley, Thomâ•… 326 Balahur, Alexandraâ•… 310 Barðal, Jóhannaâ•… 41 Barnabé, Aurélieâ•… 310, 317 Barrio, Gregarioâ•… 106 Bartens, Raijaâ•… 254, 265, 266 Barthélemy, Jean-Pierreâ•… 18 Bartmiński, Jerzyâ•… 12 Bartning Ingeâ•… 180 Bates, Douglasâ•… 327

Bates, Elizabethâ•… 187, 188 Beitel, Dinaraâ•… 95 Bellavia, Elenaâ•… 17 Benki, Joséâ•… 101 Benzécri, Jean-Paulâ•… 148, 158, 317, 446, 447, 482 Beretta, Alanâ•… 87 Berkson, Josephâ•… 487 Berthele, Raphaelâ•… 318, 325 Biber, Douglasâ•… 155, 281, 282, 283, 292, 325 Blank, Andreasâ•… 87 Blashfield, Rogerâ•… 99 Boas, Hansâ•… 83 Boers, Frankâ•… 22 Bolinger, Dwightâ•… 148, 281 Bondarko, Aleksandrâ•… 12, 13 Borg, Ingwerâ•… 317, 318 Bortz, Jürgenâ•… 376 Boscarino, Josephâ•… 106 Bouma, Gerlofâ•… 208 Brace, Nicolaâ•… 102 Braun, Johnâ•… 136, 317, 318, 319, 321, 322, 324, 325, 326, 327, 362 Breckenridge, Jamesâ•… 95 Bresnan, Joanâ•… 101, 206, 218, 254, 259, 264, 266, 267, 280, 310, 324 Broccias, Cristianoâ•… 63, 64, 69 Brooks, Patriciaâ•… 366 Brugman, Claudiaâ•… 19, 20, 88 Bucholtz, Maryâ•… 88 Bybee, Joanâ•… 15, 190, 206, 318, 367, 397 Byloo, Pieterâ•… 182, 191 C Cadoret, Marineâ•… 317 Cappelle, Bertâ•… 255, 260, 262 Carlson, Lauraâ•… 265 Carretero, Martaâ•… 310 Carroll, Johnâ•… 310

Casad, Eugeneâ•… 22 Cedergren, Henriettaâ•… 488 Ceulemans, Evaâ•… 216 Chaffin, Rogerâ•… 25, 316 Chambers, Jackâ•… 88 Chatterjee, Sampritâ•… 136, 137, 321 Chaturvedi, Anilâ•… 106 Chessel, Danielâ•… 326 Chomsky, Noamâ•… 123 Christensen, Runeâ•… 322, 327 Church, Kennethâ•… 147 Cienki, Alanâ•… 21 Clancy, Stevenâ•… 318 Clatworthy, Janeâ•… 99 Coates, Jenniferâ•… 181, 182, 190, 191 Coleman, Lindaâ•… 20 Colleman, Timothyâ•… 40, 41, 42, 48, 49, 309, 316 Collins, Peterâ•… 181, 183, 184, 185, 191, 201 Combs-Orme, Terriâ•… 322, 323, 325 Comrie, Bernardâ•… 226, 254, 256, 265, 266 Cooper, Marthaâ•… 93, 432 Cooper, Williamâ•… 259 Coșeriu, Eugenioâ•… 10 Coulmas, Florianâ•… 88 Coventry, Kennyâ•… 265 Cox, Davidâ•… 487 Cox, Michaelâ•… 318 Cox, Trevorâ•… 318 Crawley, Michaelâ•… 316, 318, 319, 321, 323, 326, 327, 362, 382, 383 Croft, Williamâ•… 19, 21, 62, 70, 71, 83, 318, 325 Croissant, Yvesâ•… 327 Cruse, Alanâ•… 10, 145, 153 Culioli, Antoineâ•… 12 Cuyckens, Hubertâ•… 17, 19, 20, 21, 88, 118

536 Corpus Methods for Semantics

D D’Andrade, Royâ•… 225 Dąbrowska, Ewaâ•… 20, 22, 89, 228 Daille, Béatriceâ•… 310 Dalgaard, Peterâ•… 315, 316, 321, 322, 324, 325, 362 Danielewiczowa, Magdalenaâ•… 225 Davies, Markâ•… 146, 154, 393 De Clerck, Bernardâ•… 40, 41, 46, 49, 52 De Cock, Barbaraâ•… 310 De Haan, Ferdinandâ•… 182 De Leeuw, Janâ•… 326, 349, 350, 474, 475, 476 De Schutter, Gertâ•… 48 de Stadler, Leonâ•… 22 De Sutter, Gertâ•… 206 de Vega, Manuelâ•… 266, 268 Deane, Paulâ•… 17, 19, 20 Degand, Liesbethâ•… 207 Deignan, Aliceâ•… 309 Delaere, Isabelleâ•… 482 Delbeque, Nicoleâ•… 22 Delorge, Martineâ•… 44, 46, 51, 309 DeMaris, Alfredâ•… 136 Depraetere, Ilseâ•… 190 Desagulier, Guillaumeâ•… 317 Deshors, Sandraâ•… 190, 201, 310, 321 Dewell, Robertâ•… 17, 229 Dickey, Stephanâ•… 227, 228, 229, 230, 232, 233, 234 Diehl, Hanneleâ•… 310 Divjak, Dagmarâ•… 27, 89, 95, 128, 131, 157, 163, 174, 180, 186, 188, 192, 195, 226, 235, 236, 307, 308, 310, 321, 324, 325, 435, 481 Dixon, Robertâ•… 64 Dodge, Yadolahâ•… 136, 244 Downing, Angelaâ•… 171 Drenan, Robertâ•… 316, 317 Dubois, Danièleâ•… 19, 23 Duda, Richardâ•… 99 Dufour, Anne-Béatriceâ•… 326, 477 Dunbar, Georgeâ•… 19, 87 Duyck, Wouterâ•… 42 Dziwirek, Katarzynaâ•… 307, 308

E Eckert, Penelopeâ•… 88 Eddington, Davidâ•… 206, 318 Edmonds, Philipâ•… 153 Edwards, Davidâ•… 323 Einspruch, Ericâ•… 100 Elsness, Johanâ•… 281, 282, 288, 294 Erades, Pieterâ•… 278 Erelt, Matiâ•… 255, 256, 257 Evans, Vyvyanâ•… 17, 19, 120 Everitt, Brianâ•… 93, 94, 99, 162, 315, 316, 317, 318, 319, 321, 325, 327, 328, 362, 435 Evert, Stefanâ•… 147, 316, 325 Eyrich, Christophâ•… 22 F Fabiszak, Małgorzataâ•… 126, 247 Faraway, Julianâ•… 136, 137, 320, 321, 322, 327, 382 Fauconnier, Gillesâ•… 13 Feist, Micheleâ•… 265, 266 Fellbaum, Christianeâ•… 73 Fidell, Lindaâ•… 101, 318, 322, 324, 325 Field, Andyâ•… 105, 321, 322, 327 Fillmore, Charlesâ•… 18, 19, 20, 21, 22, 25, 27, 83 Firth, Johnâ•… 146, 308 Fischer, Kerstinâ•… 205, 236, 307, 310 Fleischman, Suzanneâ•… 190 Flores Salgado, Elizabethâ•… 310 Fontaine, Johnnyâ•… 310 Ford, Marilyn â•… 254, 259, 264, 266, 267 Fortescue, Michaelâ•… 223, 224 Fought, Carmenâ•… 88 Fox, Johnâ•… 525 Freese, Jeremyâ•… 322, 323, 325 Fuchs, Catherineâ•… 26 Funke, Stefanâ•… 326 G Gabrielatos, Costasâ•… 182, 184, 185 Garrod, Simonâ•… 265 Gass, Susanâ•… 188 Geeraerts, Dirkâ•… 10, 15, 16, 17, 18, 19, 20, 21, 22, 23, 26, 27, 40,

87, 88, 118, 119, 120, 121, 122, 124, 206, 207, 208, 209, 215, 217, 219, 280, 307, 310, 311, 317, 321, 325, 488 Geiger, Richardâ•… 22 Gelman, Andrewâ•… 321 Gentner, Dedreâ•… 265, 266 Gifi, Albertâ•… 450 Gilbert, Ericâ•… 174 Gilquin, Gaëtanelleâ•… 17, 122, 157, 165, 308, 309, 316, 324, 325 Givón, Talmyâ•… 12, 13 Glynn, Dylanâ•… 16, 18, 21, 27, 46, 61, 72, 73, 74, 83, 117, 119, 120, 132, 134, 135, 136, 205, 223, 224, 226, 234, 236, 238, 253, 254, 272, 273, 274, 280, 288, 307, 208, 310, 311, 316, 317, 321, 322, 324, 325, 482, 506 Goddard, Cliffâ•… 224, 225 Goldberg, Adeleâ•… 12, 20, 42, 49, 61, 63, 66, 67, 83, 151, 152, 156, 173, 174, 367, 392 Gower, Johnâ•… 414, 418, 420, 421, 422, 437, 438, 482 Green, Paulâ•… 106 Greenacre, Michaelâ•… 46, 134, 148, 158, 239, 317, 326, 445, 446, 447, 448, 449, 450, 466, 467, 467, 470, 471, 473, 482 Greenberg, Josephâ•… 267 Gries, Stefanâ•… 18, 22, 25, 26, 27, 62, 66, 67, 68, 70, 81, 87, 94, 95, 117–128, 130, 131, 135, 137, 141, 156, 157, 160, 162, 163, 165, 174, 180, 186, 192, 193, 194, 195, 201, 206, 216, 217, 226, 236, 254, 288, 289, 307, 308, 309, 310, 315, 316, 318, 319, 231, 322, 324, 325, 326, 327, 372, 378, 382, 387, 391, 392, 393, 395, 397, 405, 407, 435 Grochowska, Alinaâ•… 230 Groenen, Patrickâ•… 318 Grondelaers, Stefanâ•… 10, 206, 236, 280, 310, 321, 488 H Hadfield, Jarrodâ•… 327 Hadi, Aliâ•… 136, 137, 321 Hagège, Claudeâ•… 254, 256, 265, 266

Name index 537

Halliday, Michaelâ•… 12 Hanauer, Davidâ•… 316 Hanks, Patrickâ•… 87, 147 Härdle, Wolfgangâ•… 316, 317, 318 Harnad, Stevanâ•… 405 Harrell, Frankâ•… 211, 321, 327, 505, 530 Haspelmath, Martinâ•… 40 Hawkins, Barbaraâ•… 19 Hawkins, Johnâ•… 259 Hebda, Annaâ•… 224, 247 Hennig, Christianâ•… 326, 432 Hermerén, Larsâ•… 182, 185, 197 Herskovits, Annetteâ•… 20, 265, 268 Heylen, Krisâ•… 27, 106, 205, 206, 280, 310, 321, 324 Hidetoshi, Shimodairaâ•… 326 Hilbe, Josephâ•… 320, 321, 322, 323, 325, 532 Hill, Jenniferâ•… 321, 323 Hilpert, Martinâ•… 27, 61, 156, 157, 175, 308, 309, 316, 325, 403 Hirst, Graemeâ•… 153 Hoffmann, Thomasâ•… 308, 318, 325 Hommerberg, Charlotteâ•… 382, 283 Hopper, Paulâ•… 7, 281 Hosmer, Davidâ•… 101, 105, 212, 320, 528, 532 Hothorn, Torstenâ•… 214, 315, 316, 317, 319, 321, 325, 327, 328, 362 Hox, Joopâ•… 324 Huddleston, Rodneyâ•… 182, 190, 201 Husson, Francoisâ•… 317, 325, 326, 448, 482 Hyltenstam, Kennethâ•… 180 I Ito, Takehikoâ•… 187, 188 Izenman, Alanâ•… 316, 317, 318, 325 J Jackendoff, Rayâ•… 224 Janda, Lauraâ•… 19, 20, 21, 22, 27, 118, 230, 310 Janssen, Theoâ•… 40 Järvikivi, Juhaniâ•… 122 Johnson, Danielâ•… 488

Johnson, Keithâ•… 316, 321, 326, 327, 362, 382, 405, 407, 435 Johnson, Markâ•… 224 Johnson, Richardâ•… 408 Johnson, Valenâ•… 323 Jones, Jamesâ•… 325 Jones, Stevenâ•… 11 Jones, Valâ•… 99 Junker, Marie-Odileâ•… 224 K Kallel, Amelâ•… 101 Kaltenböck, Guntherâ•… 282, 288, 393, 402 Kastovsky, Dieterâ•… 10 Kaufman, Leonardâ•… 326, 428 Kay, Paulâ•… 20, 21, 83, 152 Kearns, Katherineâ•… 282, 288, 294 Keen, Kevinâ•… 328, 362 Kemmer, Suzanneâ•… 206, 207, 209, 210 Kennedy, Graemeâ•… 149, 155 Keune, Karenâ•… 488 Kilborn, Kerryâ•… 187, 188 Kilgarriff, Adamâ•… 16, 74, 119, 147, 156 Kittay, Evaâ•… 19, 21, 23 Klavan, Janeâ•… 26, 258, 275, 280, 310, 319, 324, 325 Kleiber, Georgeâ•… 19, 118 Klein, Deborahâ•… 87 Kleinbaum, Davidâ•… 101 Klinge, Alexâ•… 182, 185 Kokorniak, Iwonaâ•… 247, 310, 317, 482 Konat, Barbaraâ•… 224, 247 Kowalczyk, Teresaâ•… 482 Krawczak, Karolinaâ•… 27, 224, 247, 310, 317, 322, 324, 325, 482 Kreitzer, Anatolâ•… 17, 20 Krishnamurthy, Rameshâ•… 88 Kristiansen, Gitteâ•… 88 Kroonenberg, Pieterâ•… 322, 325 Kruisinga, Etskoâ•… 378 Kudrnáčová, Naděždaâ•… 124 Kustova, Galinyâ•… 225 L Labov, Williamâ•… 88 Lakoff, Georgeâ•… 10, 12, 14, 17, 18, 19, 20, 22, 25, 119, 120, 224

Lancelot, Renaudâ•… 326 Langacker, Ronaldâ•… 7, 25, 49, 145, 151, 219, 227, 234, 254, 268 Le Roux, Brigitteâ•… 317, 325, 448, 482 Lê, Sébastienâ•… 326, 477 Leacock, Claudiaâ•… 19, 88 Ledolter, Johannesâ•… 316, 322 Leech, Geoffreyâ•… 154, 181, 182, 201 Legendre, Pierreâ•… 414, 418 Lehrer, Adrienneâ•… 19, 21, 22, 23, 26, 88, 120 Lehrer, Keithâ•… 26 Lemeshow, Stanleyâ•… 101, 105, 212, 320, 528, 532 Lemmens, Maartenâ•… 21, 22 Lesnoff, Matthieuâ•… 326 Lestrade, Sanderâ•… 254, 256, 265, 266 Levin, Bethâ•… 63, 263 Levshina, Nataliaâ•… 26, 216, 217, 310, 317, 319, 321, 324, 325 Levy, Josephâ•… 132 Lewandowska-Tomaszczyk, Barbaraâ•… 22, 88, 307, 308 Liamkina, Olgaâ•… 17 Liaw, Andyâ•… 326 Lienert, Gustavâ•… 376 Lindner, Susanâ•… 19, 20, 232 Lipka, Leonardâ•… 10 Locke, Philipâ•… 171 Loewenthal, Judithâ•… 210 Lohmann, Arneâ•… 375 Long, Scottâ•… 323, 325 Lorenz, Gunterâ•… 149, 156 Louwerse, Maxâ•… 316 Luraghi, Silviaâ•… 254, 265, 266 Lutzeier, Peterâ•… 10 Lyons, Johnâ•… 10 M MacLaury, Robertâ•… 19, 23 MacWhinney, Brianâ•… 180, 187, 188 Maechler, M.â•… 326 Magidson, Jayâ•… 106 Maindonald, Johnâ•… 136, 317, 318, 319, 312, 323, 324, 325, 326, 327, 362 Mair, Patrickâ•… 326, 449, 450, 474, 475, 476


Malchukov, Andrejâ•… 40 Manning, Chrisâ•… 147 Marden, Joaquimâ•… 147 Marques de Sá, Joaquimâ•… 136 Medin, Douglasâ•… 206 Meex, Birgittaâ•… 17 Mel’čuk, Igorâ•… 10 Melis, Ludoâ•… 20 Menard, Scottâ•… 212, 320, 321 Mervis, Carolynâ•… 206, 216 Milligan, Glennâ•… 93, 432 Milroy, Lesleyâ•… 88 Mittal, Hrishiâ•… 362 Moisl, Hermannâ•… 94, 99 Mojena, Richardâ•… 98 Mondorf, Brittaâ•… 254, 259, 260 Montoyo, Andrésâ•… 310 Morgan, Pamelaâ•… 20 Morgenstern, Aliyahâ•… 310 Moss, Marcâ•… 104 Mulac, Anthonyâ•… 280, 281, 282, 288, 294 Müller, Henrikâ•… 182, 185 Mun, Eun-Youngâ•… 322 Murphy, Gregoryâ•… 87 Murphy, Lynnâ•… 10, 11 Murtagh, Fionnâ•… 317, 482 Myers, Danâ•… 25, 316 N Neandić, Olegâ•… 326 Neff, Joanneâ•… 184, 185 Nelder, Johnâ•… 487, 505 Nerlich, Brigitteâ•… 19, 88 Nevalainen, Terttuâ•… 149, 174 Newman, Johnâ•… 20, 44, 309 Nicholls, Dianeâ•… 88 Nordmark, Henrikâ•… 317, 322 Norušis, Marijaâ•… 102 Norvig, Peterâ•… 19 Nuyts, Janâ•… 190, 191 O O’Connell, Annâ•… 323 Oakes, Michaelâ•… 322 Ojutkangas, Kristaâ•… 254, 265, 266, 268 Oksanen, Jariâ•… 449, 477 Oostdijk, Nellekeâ•… 519 Orme, Johnâ•… 322, 323, 325 Otani, Naokiâ•… 186, 193, 194, 195

P Pagès, Jérômeâ•… 146 Palander-Collin, Minnaâ•… 280, 281 Palmeos, Paulineâ•… 256, 257, 266 Palmer, Frankâ•… 182, 190, 191 Palmer, Garyâ•… 225 Pampel, Fredâ•… 532 Paolillo, Johnâ•… 532 Paprotté, Wolfâ•… 21, 23 Paradis, Caritaâ•… 147, 148, 149, 150, 152, 153, 162, 164, 173, 174 Pasich-Piasecka, Agnieszkaâ•… 230 Pawłowska, Reginaâ•… 225 Pedersen, Tedâ•… 156 Peng, Chao-Yingâ•… 104, 105 Pęzik, Piotrâ•… 309 Pichler, Heikeâ•… 309 Piernikarski, Cezaryâ•… 229 Pinker, Stevenâ•… 63, 64 Plevoets, Koenâ•… 317, 443, 482 Poole, Keithâ•… 318, 325 Przybylska, Renataâ•… 228, 230, 232 Pütz, Martinâ•… 19, 23, 88, 308 Q Quine, Willardâ•… 153 Quirk, Randolphâ•… 148, 150 R Radden, Günterâ•… 19, 182, 229 Rakova, Marinaâ•… 19, 88 Rannat, Rutaâ•… 254 Rao, Vithalaâ•… 106 Rastier, Françoisâ•… 10 Rauh, Gisaâ•… 23 Ravid, Doritâ•… 316 Ravin, Yaelâ•… 19, 88 Read, Jonathonâ•… 310 Reed, Susanâ•… 190 Reif, Monikaâ•… 88, 308 Rencher, Alvinâ•… 317, 482 Rice, Sallyâ•… 18, 20, 21, 26, 119, 309, 316 Ripley, Brianâ•… 318, 319, 325, 326, 454 Rissanen, Mattiâ•… 149, 174, 280, 281, 282, 283, 288, 292

Robinson, Justynaâ•… 88, 89, 90, 98, 106, 111, 310, 316, 319, 321, 325 Roever, Christianâ•… 326 Rohdenburg, Günterâ•… 281, 282, 288, 293 Romney, Kimballâ•… 444, 482 Rosch, Eleanorâ•… 119, 206, 216 Rosenbach, Anetteâ•… 255 Ross, Johnâ•… 259 Rouanet, Henryâ•… 317, 325, 448, 482 Rousseeuw, Peterâ•… 326, 428, 432 Rudzka-Ostyn, Brygidaâ•… 19, 20, 21, 22, 23, 118, 228, 310 Ruette, Tomâ•… 318 S Saffran, Jennyâ•… 367 Sagi, Eyalâ•… 309 Salkie, Raphaelâ•… 185, 193, 195 Salm, Siretâ•… 258 Saltini, Annaâ•… 106 Sanders, Joséâ•… 22 Sandra, Dominiekâ•… 20, 26, 119, 316 Sankoff, Davidâ•… 488 Sarmento, Simoneâ•… 182, 184, 185 Schaffer, Margueriteâ•… 206 Scheibman, Joanneâ•… 310, 367 Scherer, Klausâ•… 310 Schermer-Vermeer, Everdinaâ•… 51 Schmid, Hans-Jörgâ•… 15, 21, 22, 26, 106, 122, 173, 318, 366, 397, 481 Schmidtke-Bode, Karstenâ•… 318, 481 Schneider, Walterâ•… 13 Schulze, Rainerâ•… 20, 25, 316 SchuÌ‹tze, Hinrichâ•… 147 Schwarz, Monikaâ•… 19, 23 Sethuraman, Nityaâ•… 367 Shaw, Davidâ•… 406 Sheather, Simonâ•… 321, 323 Sheskin, Davidâ•… 382 Shiffrin, Richardâ•… 13 Simar, Léopoldâ•… 316, 317, 318 Simon-Vandenbergen, AnneMarieâ•…149 Sinclair, Johnâ•… 155

Name index 539

Śmiech, Witoldâ•… 227, 228, 232 Smith, Ericâ•… 488 Smith, Jenniferâ•… 281, 282, 288, 294 Smith, Robertâ•… 322, 325 Solovyev, Valeryâ•… 27, 310, 316 Speelman, Dirkâ•… 27, 101, 136, 206, 207, 208, 209, 215, 217, 219, 223, 236, 243, 280, 288, 289, 310, 321, 325, 355, 382, 482 Spooren, Wilbertâ•… 22 Steckel, Joelâ•… 106 Stefanowitsch, Anatolâ•… 13, 18, 26, 62, 66, 67, 68, 70, 81, 131, 147, 152, 156, 157, 160, 162, 165, 307, 309, 311, 315, 316, 316, 325, 391, 392, 393, 397 Stepanov, Juriâ•… 12 Stevens, Jamesâ•… 318 Stoffel, Cornelisâ•… 146, 150, 155 Storjohann, Petraâ•… 153 Storms, Gerritâ•… 216 Strobl, Carolinâ•… 319, 326 Stukker, Ninneâ•… 207, 209 Suzuki, Ryotaâ•… 326 Sweetser, Eveâ•… 87 Szelid, Veronikaâ•… 317, 325, 482 Szmrecsanyi, Benediktâ•… 136, 254, 255, 310, 317, 318, 321, 325 Szwedek, Aleksanderâ•… 224 T Tabachnick, Barbaraâ•… 101, 318, 322, 324, 325 Tabakowska, Elżbietaâ•… 228, 229, 233, 243 Taboada, Maiteâ•… 310 Tagliamonte, Saliâ•… 281, 282, 288, 294, 488, 532

Talmy, Leonardâ•… 12, 25, 265, 268 Tan, Pang-Ningâ•… 94, 99 Tarling, Rogerâ•… 322, 323, 325 Taylor, Johnâ•… 17, 19, 20, 22, 23, 88 Therneau, Terryâ•… 326 Thompson, Lauraâ•… 320, 322, 326, 327 Thompson, Susanâ•… 280, 281, 282, 288, 294 Tomasello, Michaelâ•… 15, 218, 366 Torres Cacoullos, Renaâ•… 281, 282, 283, 288, 293, 294, 295 Tottie, Gunnelâ•… 382, 383 Traugott, Elisabethâ•… 146, 281 Tribushinina, Elenaâ•… 122 Tryon, Robertâ•… 406 Tsohatzidis, Savasâ•… 19, 23 Tuggy, Davidâ•… 10, 19 Tummers, Joséâ•… 9, 101, 205, 324 Turner, Markâ•… 13 Tyler, Andreaâ•… 17, 19, 120 V Vainik, Eneâ•… 254 Van Belle, Williamâ•… 40, 48 Van Bogaert, Julieâ•… 309 Van der Leek, Frederikeâ•… 63, 64 Van der Zee, Emileâ•… 265 Van Langendonck, Willyâ•… 40, 48 Van Peer, Willieâ•… 316 Vandeloise, Claudeâ•… 17, 19, 20, 265, 268 Vanhove, Martineâ•… 88 Vanparys, Janâ•… 19 Venables, Williamâ•… 318, 319, 325, 326, 454 Vendler, Zenoâ•… 191, 225, 236

Verdonik, Darinkaâ•… 310 Verhagen, Arieâ•… 48, 205, 207, 209, 210, 215 Verschueren, Jefâ•… 10, 19 Victorri, Bernardâ•… 10, 26 von Eye, Alexanderâ•… 318, 322, 325, 326 Vorkachev, Sergeyâ•… 12 Vorlat, Emmaâ•… 22 W Walker, Cindyâ•… 321, 322 Walker, Jamesâ•… 281, 282, 288, 293, 294 Ward, Joeâ•… 163 Wasow, Thomasâ•… 259, 260, 262 Wedderburn, Robertâ•… 487, 505 Weller, Susanâ•… 444, 482 Wichern, Deanâ•… 408 Wiebe, Janyceâ•… 310 Wiechmann, Danielâ•… 147, 316 Wiener, Matthewâ•… 326 Wierzbicka, Annaâ•… 10, 12, 19, 224, 225 Williams, Robertâ•… 488 Wilson, Dianaâ•… 367 Wilson, Edwinâ•… 487 Wong, Mayâ•… 325 Worcester, Janeâ•… 487 Wulff, Stefanieâ•… 27, 201, 254, 259, 309, 310, 316, 318, 319, 403 Z Zawada, Brittaâ•… 19, 88, 325 Zelinsky-Wibbelt, Corneliaâ•… 19 Zeschel, Arneâ•… 309 Zhao, Yanchangâ•… 326 Zlatev, Jordanâ•… 16, 19, 119, 310

Subject index

A Access (software)â•… 345 ade4â•… 326, 452, 477 Adessive caseâ•… 253–275 adpositional constructionâ•… 253–275 agglomeration methodâ•… see linkage algorithm Agglomerative algorithmsâ•… see linkage algorithms agnesâ•… 326, 423, 425, 428–430, 433, 439, 440 AICâ•… see Akaike’s information criterion Akaike’s information criterionâ•… 136, 211, 212, 213, 218, 508, 510 amapâ•…439 anacorâ•… 326, 449, 452, 474– 476, 479 anovaâ•… 385–387, 511, 512 aodâ•…326 apeâ•… 480, 481 as.distâ•… 438, 439 as.numericâ•… 509, 513 average linkageâ•… see linkage algorithms B Base (software)â•… 345 behavioural-profile approachâ•… see multifactorial usagefeature analysis binary correspondence analysisâ•… see correspondence analysis binary logistic regressionâ•… see logistic regression binomial logistic regressionâ•… see binary logistic regression, under regression analysis binomial testâ•… 147, 165, 315 biplotâ•… 317, 318, 447, 448, 454, 455, 458–460, 464, 465, 472, 478, 479, 780

bootstrap resamplingâ•… 99, 138, 163, 164, 319, 434, 480, 481 bootstrap validationâ•… see bootstrap resampling BP approachâ•… see multifactorial usage-feature analysis Burt matrixâ•… 134, 239, 449, 450, 470, 471, 474 C c-score, goodness-of-fit, underâ•… see c-statistic caâ•… 326, 361, 362, 447, 452, 454, 466, 467, 470, 472, 474, 480 Calc (software)â•… 345 canonical correspondence analysisâ•… see correspondence analysis carâ•…387 categorical dataâ•… 94, 226, 314, 315, 319–321, 324, 326, 349, 352, 355, 358, 366, 367, 369, 371, 409–411, 413, 414, 416, 418, 421, 438, 443, 446, 478, 489, 490, 495, 502, 503, 505, 519, 520 categorical variablesâ•… see categorical data causative constructionâ•… 12, 21, 67, 165, 205–219 cbindâ•… 354, 357 cfaâ•…326 cfa2â•…326 cforestâ•…326 Chi-square automatic interaction detection (CHAID)â•… 105, 106 Chi-squared testâ•… 132, 138, 147, 151, 156, 169, 259, 266, 309, 315, 318, 325, 370–381, 445, 446, 462, 463, 511 chisq.testâ•… 370, 373, 374, 376, 378, 281, 462

city block distanceâ•… see Manhattan, under distance measures claraâ•… 326, 428 clmâ•…327 classification and regression tree analysis (CART) classification tree analysisâ•… 89, 99, 105–110, 213–218, 318–319, 325, 326, 531 random forest analysisâ•… 319, 324, 326, 531 regression tree analysisâ•… 213, 319 Clustan (software)â•… 93, 94 clusterâ•…437–440 cluster analysis fuzzy cluster analysisâ•… 428 hierarchical agglomerative cluster analysisâ•… 91–101, 126, 228, 129, 132, 148, 157–159, 162, 163, 164, 192–195, 201, 226, 227, 234, 238, 245–247, 316, 217, 325, 326, 405–441, 478–482 k-means cluster analysisâ•… 316, 326, 422, 427, 428, 430 k-medoid cluster analysisâ•… 428 cluster.statsâ•…432 clustering algorithmsâ•… see linkage algorithms cmdscaleâ•… 326, 433, 440 coefficients (logistic regression)â•… see estimated coefficients of predictors Cognitive Linguisticsâ•… 7–10, 13, 16 –18, 21, 22, 25, 26, 187, 206, 309, 310, 311, 405, 482 Cognitive Semanticsâ•… 7, 17, 18, 19, 21, 22, 24–26, 28, 117–120, 158, 173, 205


collexeme analysisâ•… see collostructional analysis collinearityâ•… see multicollinearity collostructional analysis collexeme analysisâ•… 62–63, 65–67, 70, 72–73, 75–76, 82, 84, 147, 154, 156–164, 168, 393–397 covarying-collexeme analysisâ•… 315, 392, 400–403 distinctive collexeme analysisâ•… 147–148, 154, 157, 158, 162, 165–172, 201, 315, 392, 397–400 colnamesâ•… 351, 352, 353 complementationâ•… 229, 244, 247, 280–300, 382–387 complete linkageâ•… see linkage algorithms conative constructionâ•… 62–67, 81–84 confidence intervalsâ•… 99, 296, 298, 475, 501, 517 configural frequency analysisâ•… 135, 136, 138, 321, 324, 325, 236 Construction Grammarâ•… 42, 52, 63, 66, 151, 152, 173, 254, 392 contingency tableâ•… 66, 134, 148, 158, 345, 346, 350, 356–358, 446, 447, 453, 457, 463, 464, 469, 477, 481, 512 continuous dataâ•… 314, 318, 319, 324, 352, 358, 365, 366, 409, 411, 413, 416, 418–420, 433, 436, 477, 489, 490, 505, 519 correspondence analysis binary correspondence analysisâ•… 132–134, 240, 242, 326, 446, 448, 449, 451, 454–457, 461–467, 475–477, 479, 481 canonical correspondence analysisâ•… 449, 461, 474 detrended correspondence analysisâ•… 449, 477 multiple correspondence analysisâ•… 134, 226, 239, 317, 326, 382, 446, 448–450, 454, 457–460, 470–474, 477, 480, 481

corres.fncâ•… 461, 462, 465,

466, 469 correspâ•… 326, 454–456 count dataâ•… see categorical data covarying-collexeme analysisâ•… see collostructional analysis Cox and Snell’s pseudo R2 statisticâ•… see goodness-of-fit cross-tabulationâ•… see contingency table ctreeâ•…326 D

datadistâ•… 510, 520 dataframeâ•… 344–354, 356–358, 365, 453 decision tree analysisâ•… see classification tree analysis, under classification and regression tree analysis degree modifiersâ•… 146–164, 173–174 degrees of freedomâ•… 44, 49, 106–108, 137, 169, 195, 381, 509, 511, 528 dendrogramâ•… 92, 93, 95–99, 128–131, 148, 158, 163, 164, 193–195, 316, 422–427, 478–481 Designâ•…327 detrended correspondence analysisâ•… see correspondence analysis deviance residualsâ•… see residual deviance dianaâ•…428 dichotomous logistic regressionâ•… see binary logistic regression, under regression analysis discrete dataâ•… see categorical data dissimilarity matrixâ•… see distance measures distâ•… 423, 436–438, 440, 480, 481 distance measuresâ•… 92, 128, 130, 163, 193 Canberraâ•… 130, 131, 163, 192 Euclideanâ•… 128–130, 163, 417–420, 426, 427, 437, 438, 446, 480

Mahalanobisâ•… 417, 420, 421, 437, 438 Manhattan (City-block)â•… 131, 417–420, 433, 436–438 phi square (mean square)â•… 92, 94 distinctive collexeme analysisâ•… see collostructional analysis divisive algorithmsâ•… 426–428 drop1â•… 522, 525 Dunning’s log-likelihood ratioâ•… 161 Dxyâ•… see Somers’ Dxy under goodness-of-fit DynGraph (software)â•… 169, 477 E estimated coefficients (of predictors / log odds ratios)â•… 101–103, 136–139, 211–213, 215, 218, 244, 273, 274, 289, 292– 299, 384–387, 507, 519–527 estimates (logistic regression)â•… see estimated coefficients of predictors Euclidean distanceâ•… see distance measures Excel (software)â•… 345, 347, 453, 457, 463, 464 F

FactoMineRâ•… 169, 326, 452,

458, 477, 482

factorâ•… 352, 355 fannyâ•…428

FileMaker (software)â•… 345 Fisher-Yates exact testâ•… 54, 66, 68, 147, 156, 165, 244, 267, 309, 315, 325, 371, 395 fisher.testâ•…371 fpcâ•… 326, 432 frequency dataâ•… see categorical data frequency tableâ•… see contingency table Functional Linguisticsâ•… 8–13, 308, 310 furthest neighbour linkageâ•… see complete linkage, under linkage algorithms fuzzy cluster analysisâ•… see cluster anlaysis

Subject index 543

Fuzzy Set Theoryâ•… 26, 27, 119, 150, 153, 181, generalized linear modelsâ•… see regression analysis G

glmâ•… 326, 327, 383, 384, 505–507,

509, 510–513, 515–517, 520, 524

glmmPQLâ•…327

Goldvarb X (software)â•… 488 Goodman-Kruskal’s Gammaâ•… see goodness-of-fit goodness-of-fit c-statisticâ•… 104, 136, 137, 212, 243, 244, 273, 289, 515 Cox and Snell’s pseudo (generalised) R2 statisticâ•… 104, 105, 515 Goodman-Kruskal’s Gammaâ•… 212 Hosmer-Lemeshow-Cessie testâ•… 104, 105, 212, 528 log likelihood ratio test (–2 log likelihood)â•… 104, 105, 137, 212, 509, 511, 521, 522 McFadden’s pseudo (generalised) R2 statisticâ•… 138, 213 Nagelkerke’s pseudo (generalised) and adjusted R2 statisticâ•… 105, 137–140, 196, 243, 291, 511, 515 Somers’ Dxyâ•… 212, 273, 514 Gower universal similarity coefficientâ•… 418, 420–422, 437, 438 H

hcfaâ•…326 hclustâ•… 129, 130, 163, 326, 423,

431, 433, 440, 441, 479–481

headâ•… 348, 505, 506, 513, 514, 519

hierarchical agglomerative cluster analysis (HAC)â•… see cluster analysis hierarchical modellingâ•… see mixed-effects logistic regression, under regression analysis homalsâ•… 452, 476, 477

homogeneity analysisâ•… see indicator multiple correspondence analysis, under correspondence analysis Hosmer-Lemeshow-Cessie testâ•… see goodness-of-fit I indicator multiple correspondence analysisâ•… see correspondence analysis inertiaâ•… 46, 53, 133, 134, 170, 239, 240, 446, 447, 449–451, 461, 465, 467–474, 476 interceptâ•… 215, 218, 384, 491, 498–500, 503, 504, 507–509, 511, 521, 522 J Jacard coefficientâ•… 414, 416, 436, 437 joint multiple correspondence analysis (search for joint)â•… see correspondence analysis K k-means cluster analysisâ•… see cluster analysis klaRâ•…326 kmeansâ•…326 L

languageRâ•… 452, 460–462, 465,

466, 469, 479

ldaâ•…326 levelsâ•… 355, 506

linear discriminant analysisâ•… 216, 318, 319, 326 linear regressionâ•… see regression analysis linkage algorithms average linkageâ•… 94, 424–426, 439, 480 complete linkage (furthest neighbour)â•… 94, 424–426, 439 single linkage (nearest neighbour)â•… 94, 129, 194, 423, 424

Wardâ•… 92, 94, 130, 163, 246, 428, 431–434, 440, 480–481 listâ•… 369, 373–375, 378, 381, 526, 527 lme4â•…327 lmerâ•… 218, 327 log likelihood ratio testâ•… see goodness-of-fit log likelihood test (–2 log likelihood)â•… see log likelihood ratio test under goodness-of-fit log oddsâ•… 221, 215, 494 log odds ratiosâ•… see estimated coefficients of predictors logistic regressionâ•… see regression analysis logitsâ•… 493–495, 497, 498, 531 loglinear analysisâ•… 174, 321, 325, 326, 475 loglmâ•…326 lrmâ•… 327, 505, 510, 511, 515, 516, 518, 520, 521, 524, 528 M Mahalanobis distanceâ•… see distance measures Manhattan distanceâ•… see distance measures Marascuilo procedureâ•… 378, 379, 380 MASSâ•… 288, 326, 327, 454, 456– 458, 479, 480 matching coefficientâ•… 414–417, 437 MATLAB (software)â•… 323 matrixâ•… 369, 373–375 MCAâ•…326 mcaâ•… 326, 457, 458 McFadden’s pseudo R2 statisticâ•… see goodness-of-fit MCMCglmmâ•…327 MCMClogitâ•…327 MCMCpackâ•…327 McNemar’s Paired Chi-squared test mergeâ•…354 MI Scoreâ•… see mutual information score


mixed effectsâ•… see mixed-effects logistic regression under regression analysis mixed modelâ•… see mixed-effects logistic regression under regression analysis mixed-effects logistic regressionâ•… see regression analysis mjcaâ•… 326, 470–474, 480 mlogitâ•…327 monaâ•…428 mosaic plotâ•… 525, 526 multicollinearityâ•… 136, 138, 139, 243, 272, 273, 289, 528 multidimensional scalingâ•… 317, 325, 326, 406, 440, 468 multifactorial usage-feature analysis (also behaviouralprofile approach)â•… 118–123, 124, 125, 135, 161, 180, 186, 192, 195, 200, 201, 224, 226, 236, 309–311 multilevel logistic regressionâ•… see mixed-effects logistic regression, under regression analysis multilevel modellingâ•… see mixed-effects logistic regression, under regression analysis multinomâ•…327 multinomial logistic regressionâ•… see regression analysis multiple correspondence analysisâ•… see correspondence analysis multiple logistic regressionâ•… see logistic regression, under regression analysis multiway frequency analysisâ•… see loglinear analysis mutual information scoreâ•… 147, 149, 151, 156, 309, 316, 325 N Nagelkerke’s adjusted pseudo R2â•… see R2 score Nagelkerke’s pseudo R2 statisticâ•… see goodness-of-fit nearest neighbour distanceâ•… see single linkage algorithm, under linkage algorithms

njâ•… 480, 481 nnetâ•…327

Nominal dataâ•… 226, 314, 322, 369, 410, 414 NotePad (software)â•… 345 Numbers (software)â•… 345 Numerical variablesâ•… see continuous data O odds ratiosâ•… 53, 54, 78, 82, 101–103, 211, 215, 267, 371, 372, 493–495, 498, 503, 514, 516, 517, 521, 522 operationalisationâ•… 9, 13–15, 22, 27, 73, 118, 119, 122, 123, 135, 139, 208, 215–217, 258, 259, 262, 282, 311, 407, 408, 444 ordered dataâ•… see ordinal data ordered multinomial logit regressionâ•… see ordinal logistic regression, under regression analysis ordinalâ•…327 ordinal dataâ•… 100, 314, 323, 409–412, 418, 530, 531 ordinal dataâ•… 100, 314, 323, 409, 410, 411, 418, 530, 531 ordinal logistic regressionâ•… see regression analysis overdispersionâ•… 508, 527, 528 P

pamâ•…428 pamctdpâ•… 452, 477, 478 pamkâ•…326 parâ•…358 partyâ•… 326, 477

Pearson residualsâ•… 44, 46, 52, 58, 59, 132, 138, 238, 239, 259, 263, 370, 373, 374, 377, 462, 463 periphrastic causativesâ•… 165, 207 PhyloClustâ•…481 phylogenetic tree plotâ•… 480, 481 plotâ•… 359, 439, 440, 454, 455, 457, 458, 461, 467, 473, 481, 518, 522, 525, 526 Poisson regressionâ•… see regression analysis polrâ•…327

polychotomous logistic regressionâ•… see multinomial logistic regression, under regression analysis polytomousâ•…327 polytomous logistic regressionâ•… see multinomial logistic regression, under regression analysis predictâ•… 509, 512, 513, 526, 527 predictor estimates / coefficients (logistic regression)â•… see estimated coefficients of predictors prettyâ•… 145, 6, 148, 149, 152, 157, 159–175 prop.tableâ•… 380, 381 proportional odds logistic regressionâ•… see ordinal logistic regression, under logistic regression proportional odds regressionâ•… see ordinal logistic regression, under regression analysis proportionsâ•… 376, 492, 493, 494, 495, 496, 497, 498, 500, 503, 512, 518, 522, 523 proportions testâ•… 315 prototype effectsâ•… 9, 13–16, 18, 25, 26, 62, 88, 117–128, 136, 138, 139, 141, 206, 216, 217, 219 Prototype Set Theoryâ•… see prototype effects pvclustâ•… 163, 164, 326, 431, 433, 434, 440, 441, 479–481 pvrectâ•… 434, 440 Q

quasipoisâ•…326

R R (software)â•… 312, 313, 324–328, 343–345, 362–364, 368–369, 395, 437, 444, 482, 505 radial network analysisâ•… 18, 24, 26, 27, 118–120, 123 randomForestâ•…326 rbindâ•…354 rdaâ•…326 read.tableâ•… 344, 347, 348 receiver operating characteristic curveâ•… see c-statistic

Subject index 545

regression analysis binary logistic regressionâ•… 136–141, 193, 195–200, 210–213, 321, 322, 325, 327, 243–245, 272–274, 289–299, 382, 488–530 generalized linear modelsâ•… 383, 505, 508, 531 linear regressionâ•… 314, 382, 383, 489–494, 496–499, 501, 515 mixed-effects logistic regressionâ•… 288, 323, 325 multinomial logistical regressionâ•… 135, 187, 322, 323, 325, 327, 382, 488, 530 ordinal logistic regressionâ•… 323, 325, 327, 530, 531 Poisson regressionâ•… 322, 382–387, regression tree analysisâ•… see classification and regression trees analysis relevelâ•… 355, 506 residual devianceâ•… 507, 508, 511, 520, 527, 528 residualsâ•… 491, 492, 501, 507, 515 see also residual deviance and Pearson residuals rmâ•…351 rmsâ•… 137, 327, 505 ROCâ•… see c-statistic under goodness-of-fit rpartâ•…214, 326

S SAS (software)â•… 313 scaleâ•…420 semantic classâ•… 62, 63, 66, 70, 73, 75, 147, 158, 161, 175, 209, 215, 218 sentiment analysisâ•… 310 silhouetteâ•…440 silhouette plotâ•… 433 silhouette validation techniqueâ•… 139, 194, 432, 433, 434 similarity measureâ•… see distance measures single linkageâ•… see linkage algorithms smacofâ•…326 smacofSymâ•…326 sociolinguisticsâ•… 117, 158, 321, 406, 488, 532 Somers’ Dxyâ•… see goodnessof-fit SPSS (software)â•… 93, 94, 100– 102, 105, 106, 111, 313, 322, 323 Stata (software)â•… 313, 322, 323 Statistica (software)â•… 313 statistical significanceâ•… 75, 78, 98, 104, 107, 127, 135, 138, 244, 273, 315, 321, 444, 448, 475 statsâ•… 326, 436, 437 strâ•…349 subsetâ•…354 summaryâ•… 348, 349, 355, 360

T tâ•… 350, 440, 441, 461, 481 t-scoreâ•… 147, 309, 316, 325 tableâ•… 356, 357, 513 tabular dataâ•… see categorical data TextEdit (software)â•… 345 TextWrangler (software)â•… 345 titleâ•… 461, 469 treeâ•…326 U underdispersionâ•… 387, 527 usage-feature analysisâ•… see multifactorial usagefeature analysis V VARBRUL (software)â•… 488, 532 variance inflation factorâ•… 136, 137, 244, 273, 289, 528 veganâ•… 449, 452, 477 vifâ•…528 vifâ•… see variance inflation factor ward algorithm, see linkage algorithms W WordGen (software)â•… 42 WordPad (software)â•… 345 Wordsmith (software)â•… 284 Z z-scoreâ•… 93, 309, 316, 325 X χ2 testâ•… see Chi-squared test

Corpus Methods for Semantics. Quantitative studies in polysemy and synonymy

Quantitative Methods in Cognitive Semantics: Corpus-Driven Approaches (Cognitive Linguistic Research)

Quantitative Methods in Cognitive Semantics: Corpus-Driven Approaches (Cognitive Linguistic Research)

Quantitative Methods in Linguistics

Quantitative Methods In Linguistics

Quantitative Methods For Investment Analysis

Quantitative Methods For Investment Analysis

Corpus, Cognition and Causative Constructions (Studies in Corpus Linguistics)

Quantitative Methods in Project Management

Lexis in Contrast: Corpus-based Approaches (Studies in Corpus Linguistics)

Quantitative Methods in Reservoir Engineering

Quantitative methods in derivatives pricing

Lexis in Contrast: Corpus-based Approaches (Studies in Corpus Linguistics)

Collocations in a Learner Corpus (Studies in Corpus Linguistics)

Corpus Approaches to Grammaticalization in English (Studies in Corpus Linguistics)

Belief and Synonymy

Semantics and Cognition (Current Studies in Linguistics)

Small Corpus Studies and Elt: Theory and Practice (Studies in Corpus Linguistics)

Boolean Semantics for Natural Language (Studies in Linguistics and Philosophy)

Semantics and Cognition (Current Studies in Linguistics)

Keyness in Texts (Studies in Corpus Linguistics)

Perspectives on Corpus Linguistics (Studies in Corpus Linguistics)

Quantitative Methods for Decision Makers (4th Edition)

Quantitative Social Research Methods

Quantitative Methods for Business (4th Edition)

Quantitative Methods for Business, 5th Edition

Quantitative Social Research Methods

Words and Phrases: Corpus Studies of Lexical Semantics (Language in Society)

Semantics in Acquisition (Studies in Theoretical Psycholinguistics)

Corpus Linguistics in Literary Analysis: Jane Austen and her Contemporaries (Studies in Corpus and Discourse)

Corpus Methods for Semantics. Quantitative studies in polysemy and synonymy

Quantitative Methods in Cognitive Semantics: Corpus-Driven Approaches (Cognitive Linguistic Research)

Quantitative Methods in Cognitive Semantics: Corpus-Driven Approaches (Cognitive Linguistic Research)

Quantitative Methods in Linguistics

Quantitative Methods In Linguistics

Quantitative Methods For Investment Analysis

Quantitative Methods For Investment Analysis

Corpus, Cognition and Causative Constructions (Studies in Corpus Linguistics)

Quantitative Methods in Project Management

Lexis in Contrast: Corpus-based Approaches (Studies in Corpus Linguistics)

Quantitative Methods in Reservoir Engineering

Quantitative methods in derivatives pricing

Lexis in Contrast: Corpus-based Approaches (Studies in Corpus Linguistics)

Collocations in a Learner Corpus (Studies in Corpus Linguistics)

Corpus Approaches to Grammaticalization in English (Studies in Corpus Linguistics)

Belief and Synonymy

Semantics and Cognition (Current Studies in Linguistics)

Small Corpus Studies and Elt: Theory and Practice (Studies in Corpus Linguistics)

Boolean Semantics for Natural Language (Studies in Linguistics and Philosophy)

Semantics and Cognition (Current Studies in Linguistics)

Keyness in Texts (Studies in Corpus Linguistics)

Perspectives on Corpus Linguistics (Studies in Corpus Linguistics)

Quantitative Methods for Decision Makers (4th Edition)

Quantitative Social Research Methods

Quantitative Methods for Business (4th Edition)

Quantitative Methods for Business, 5th Edition

Quantitative Social Research Methods

Words and Phrases: Corpus Studies of Lexical Semantics (Language in Society)

Semantics in Acquisition (Studies in Theoretical Psycholinguistics)

Corpus Linguistics in Literary Analysis: Jane Austen and her Contemporaries (Studies in Corpus and Discourse)

Recommend Documents