Subido por antoine lonnet

(Oxford studies in theoretical linguistics 51) Alexiadou, Artemis Borer, Hagit Schäfer, Florian-The syntax of roots and the roots of syntax-Oxford University Press (2015)

Anuncio
The Syntax of Roots and the Roots of Syntax
OXFORD STUDIES IN THEORETICAL LINGUISTICS
general editors
David Adger and Hagit Borer, Queen Mary, University of London
advisory editors
Stephen Anderson, Yale University; Daniel Büring, University of California, Los Angeles; Nomi Erteschik-Shir,
Ben-Gurion University; Donka Farkas, University of California, Santa Cruz; Angelika Kratzer, University of
Massachusetts, Amherst; Andrew Nevins, University College London; Christopher Potts, Stanford University; Barry Schein, University of Southern California; Peter Svenonius, University of Tromsø; Moira Yip,
University College London
Recent titles
33 Events, Phrases, and Questions
by Robert Truswell
34 Dissolving Binding Theory
by Johan Rooryck and Guido Vanden Wyngaerd
35 The Logic of Pronominal Resumption
by Ash Asudeh
36 Modals and Conditionals
by Angelika Kratzer
37 The Theta System
Argument Structure at the Interface
edited by Martin Everaert, Marijana Marelj, and Tal Siloni
38 Sluicing
Cross-Linguistic Perspectives
edited by Jason Merchant and Andrew Simpson
39 Telicity, Change, and State
A Cross-Categorial View of Event Structure
edited by Violeta Demonte and Louise McNally
40 Ways of Structure Building
edited by Myriam Uribe-Etxebarria and Vidal Valmala
41 The Morphology and Phonology of Exponence
edited by Jochen Trommer
42 Count and Mass Across Languages
edited by Diane Massam
43 Genericity
edited by Alda Mari, Claire Beyssade, and Fabio Del Prete
44 Strategies of Quantification
edited by Kook-Hee Gil, Steve Harlow, and George Tsoulas
45 Nonverbal Predication
Copular Sentences at the Syntax-Semantics Interface
by Isabelle Roy
46 Diagnosing Syntax
edited by Lisa Lai-Shen Cheng and Norbert Corver
47 Pseudogapping and Ellipsis
by Kirsten Gengel
48 Syntax and its Limits
edited by Raffaella Folli, Christina Sevdali, and Robert Truswell
49 Phrase Structure and Argument Structure
A Case Study of the Syntax-Semantics Interface
by Terje Lohndal
50 Edges in Syntax
Scrambling and Cyclic Linearization
by Heejeong Ko
51 The Syntax of Roots and the Roots of Syntax
edited by Artemis Alexiadou, Hagit Borer, and Florian Schäfer
For a complete list of titles published and in preparation for the series, see pp 334–5.
The Syntax of Roots and
the Roots of Syntax
Edited by
ARTEMIS ALEXIADOU, HAGIT BORER,
AND FLORIAN SCHÄFER
1
3
Great Clarendon Street, Oxford, ox2 6dp,
United Kingdom
Oxford University Press is a department of the University of Oxford.
It furthers the University’s objective of excellence in research, scholarship,
and education by publishing worldwide. Oxford is a registered trade mark of
Oxford University Press in the UK and in certain other countries
© editorial matter and organization Artemis Alexiadou, Hagit Borer, and Florian Schäfer 2014;
© the chapters their several authors 2014
The moral rights of the authors have been asserted
First Edition published in 2014
Impression: 1
All rights reserved. No part of this publication may be reproduced, stored in
a retrieval system, or transmitted, in any form or by any means, without the
prior permission in writing of Oxford University Press, or as expressly permitted
by law, by licence or under terms agreed with the appropriate reprographics
rights organization. Enquiries concerning reproduction outside the scope of the
above should be sent to the Rights Department, Oxford University Press, at the
address above
You must not circulate this work in any other form
and you must impose this same condition on any acquirer
Published in the United States of America by Oxford University Press
198 Madison Avenue, New York, NY 10016, United States of America
British Library Cataloguing in Publication Data
Data available
Library of Congress Control Number: 2014934898
ISBN 978–0–19–966526–6 (hbk.)
978–0–19–966527–3 (pbk.)
Printed and bound by
CPI Group (UK) Ltd, Croydon, cr0 4yy
Links to third party websites are provided by Oxford in good faith and
for information only. Oxford disclaims any responsibility for the materials
contained in any third party website referenced in this work.
Contents
General preface
Notes on contributors
List of abbreviations
1. Introduction
Artemis Alexiadou, Hagit Borer, and Florian Schäfer
vii
viii
xi
1
2. From syntax to roots: A syntactic approach to root interpretation
Víctor Acedo-Matellán and Jaume Mateu
14
3. The roots of nominality, the nominality of roots
Paolo Acquaviva
33
4. Roots in transitivity alternations: Afto-/auto-reflexives
Artemis Alexiadou
57
5. Domains within words and their meanings: A case study
Elena Anagnostopoulou and Yota Samioti
81
6. The category of roots
Hagit Borer
112
7. On a low and a high diminutive: Evidence from Italian and Hebrew
Marijke De Belder, Noam Faust, and Nicola Lampitelli
149
8. The interaction of adjectival passive and Voice
Edit Doron
164
9. Roots and phases
Ángel J. Gallego
192
10. The ontology of roots and verbs
Lisa Levinson
208
11. Derivational affixes as roots: Phasal Spell-out meets English Stress Shift
Jean Lowenstamm
230
12. Building scalar changes
Malka Rappaport Hovav
259
vi
Contents
13. When roots license and when they respect semantico-syntactic
structure in verbs
Antje Roßdeutscher
282
References
Index
310
329
General preface
The theoretical focus of this series is on the interfaces between subcomponents of the
human grammatical system and the closely related area of the interfaces between the
different sub-disciplines of linguistics. The notion of ‘interface’ has become central
in grammatical theory (for instance, in Chomsky’s Minimalist Program) and in
linguistic practice: work on the interfaces between syntax and semantics, syntax and
morphology, phonology and phonetics, etc. has led to a deeper understanding of
particular linguistic phenomena and of the architecture of the linguistic component
of the mind/brain.
The series covers interfaces between core components of grammar, including syntax/
morphology, syntax/semantics, syntax/phonology, syntax/pragmatics, morphology/
phonology, phonology/phonetics, phonetics/speech processing, semantics/pragmatics,
and intonation/discourse structure, as well as issues in the way that the systems of
grammar involving these interface areas are acquired and deployed in use (including
language acquisition, language dysfunction, and language processing). It demonstrates,
we hope, that proper understandings of particular linguistic phenomena, languages,
language groups, or inter-language variations all require reference to interfaces.
The series is open to work by linguists of all theoretical persuasions and schools of
thought. A main requirement is that authors should write so as to be understood by
colleagues in related subfields of linguistics and by scholars in cognate disciplines.
The term ‘root’ is relatively familiar from morphology and from phonology, but
within the past 15 years, an increasing body of work has emerged which suggests that
there are empirical and conceptual advantages to assuming that the most basic
syntactic building block is neither a ‘word’ nor a Lexeme, but rather, a root, i.e. a
unit that is, in particular, devoid of a syntactic category. From the perspective of
such approaches, syntactic categories emerges as a result of the syntactic configuration, and are not, as more traditional approaches would have it, a property which
terminals bring with them into the syntax.
While there is an agreement, in such theoretical quarters, on what roots are not,
various scholars have pursued rather different solutions to the questions of what roots
are. Are they units of phonological representations, and if so, how delimited? Are
they units of meaning, and if so, how delimited? Do they have syntactic properties
(e.g. argument selection)? And finally, in the absence of category for roots, how do
syntactic constituents come to have a categorial label? This book provides an invaluable service in bringing together diverse answers to these questions, serving, simultaneously, as an introduction to the root-based approach, and as a tool to ‘rootists’
seeking to understand the diverse ramifications of the theoretical approach as a whole.
David Adger
Hagit Borer
Notes on contributors
Vı́ctor Acedo-Matellán is a postdoctoral researcher at Universidade do Minho, Portugal.
He received his Ph.D. in linguistics in 2010, at Universitat de Barcelona. His research interests
include issues in the syntax-lexicon interface and the syntax-morphology interface, and he has
worked on the argument and event structure of prefix and particle predicates. Among others,
he has published in Probus and in a volume within the series Syntax and Semantics by
Emerald.
Paolo Acquaviva is Senior Lecturer in Italian at University College Dublin. He is a graduate
of the University of Pisa and of the Scuola Normale Superiore of Pisa, where he obtained his
Ph.D. in 1993. His research centres on morphology and its interface with lexical semantics, in
particular on how linguistic categories shape the conceptualization of nouns. Lexical Plurals,
an extensive study into the varieties of non-canonical plurality, was published in 2008 by
Oxford University Press.
Artemis Alexiadou is Professor of Theoretical and English Linguistics at the Universität
Stuttgart. She received her Ph.D. in linguistics in 1994 from the University of Potsdam. Her
research interests lie in theoretical and comparative syntax, morphology, and most importantly in the interface between syntax, morphology, the lexicon, and interpretation. She has
published in journals, edited volumes, and conference proceedings.
Elena Anagnostopoulou obtained her Ph.D. in linguistics from the University of Salzburg
in 1994. After a post-doc at MIT (1997–1998), where she returned in 2007 as a Visiting
Associate Professor, she took a position at the University of Crete in 1998, where she is
currently Professor of Theoretical Linguistics. Her research interests lie in theoretical and
comparative syntax, with special focus on the interfaces between syntax, morphology, and the
lexicon, argument alternations, Case, Agreement, clitics and anaphora. She is the author of
The Syntax of Ditransitives: Evidence from Clitics (Mouton de Gruyter 2003), has co-edited
four volumes in theoretical linguistics, and has published in journals, edited volumes, and
conference proceedings.
Marijke De Belder is currently an FWO postdoc researcher at the KU Leuven campus
Brussel after having been a postdoc lecturer at Utrecht University, where she received her
Ph.D. in linguistics. Her research interests are morphosyntax and the syntax-lexicon interface.
More specifically, she has studied the syntax of roots, nominal inflection, vocabulary insertion,
derivational word-formation, and compounding. She published on Dutch nominal inflection
in the Journal of Comparative Germanic Linguistics and in Lingua.
Hagit Borer is Professor and Chair of Linguistics at Queen Mary, University of London. She
received her Ph.D. in linguistics at MIT, and has held professorial positions at the University
of Massachusetts at Amherst and at the University of Southern California. Her research
interests include syntax, morphosyntax, the syntax-semantics interface, and the acquisition
of syntax by children.
Notes on contributors
ix
Edit Doron is Professor of Linguistics at the Hebrew University of Jerusalem. She received a
Ph.D. in linguistics from the University of Texas at Austin in 1983. Her main research area is
the interface of semantics, morphology, and syntax, particularly in such languages as Hebrew,
Arabic, Aramaic, English, and French. She has published various articles on the following
topics: the Semitic verbal system, nominal predicates, adjectival passives, the subject-predicate
relation, resumptive pronouns, bare and mass nouns, ergativity, ellipsis, apposition, free
indirect discourse, the semantics of aspect and habituality, the semantics of voice, and
reference to kinds.
Noam Faust received his Ph.D. from Paris VII University in 2011. He is currently working at
the Hebrew University of Jerusalem. He is interested in phonological, morphological, and
morpho-syntactic structures, and how these can be detected through sound patterns. He has
published on these topics in the Semitic languages of Hebrew, Neo-Aramaic, and Tigre. He is
now conducting fieldwork on both Tigre and Nuer.
Ángel J. Gallego is a Professor Agregat at the Departament de Filologia Espanyola of the
Universitat Autònoma de Barcelona, where he defended his doctoral dissertation in 2007. He
is a member of the Centre de Lingüística Teòrica, a theoretical linguistic research group. His
main interests and publications concern the areas of syntax, comparative grammar, and
parametric variation (especially within Romance languages). He has published in journals
like Natural Language and Linguistic Theory, Probus, Theoretical Linguistics, or Linguistic
Analysis, and he is the author of the monograph Phase Theory (John Benjamins, 2010), and
has also acted as an editor in Phases. Developing the Framework (Mouton de Gruyter, 2012)
and El movimiento de constituyentes (with José M. Brucart, Visor, 2012).
Nicola Lampitelli is a lecturer at University of Tours (France). He received his Ph.D. in 2011
from University Paris 7. His research interests include the phonological form of morphemes,
the structures of words, and the phonology-syntax interface. He has published mainly on
Romance and Afroasiatic languages.
Lisa Levinson is an associate professor at Oakland University and received her Ph.D. from
NYU in 2007. She works on morphosemantics, trying to better understand what the atomic
units of compositional semantics are, and the extent to which those atomic units can be
mapped to atomic morphosyntactic constituents. She has recently published articles in the
journals Natural Language and Linguistic Theory and Syntax.
Jean Lowenstamm is Professor of Linguistics at Université Paris Diderot in Paris, France. He
received his Ph.D. from the University of Massachusetts, Amherst, in 1979. His research
interests include phonology, morphology, syntax, and interface issues. He has published on
those topics with special attention to Semitic, Germanic, and Romance languages. He is one of
the four editors of Brill’s Annual of Afroasiatic Languages and Linguistics.
Jaume Mateu is an associate professor of Catalan and current Director of the Center for
Theoretical Linguistics at Universitat Autònoma de Barcelona (UAB). He received his Ph.D.
in linguistics at UAB (2002). His research interests include the lexicon-syntax interface and
argument structure. Some of his recent publications are “Argument structure”, in A. Carnie
et al. (eds.), The Routledge Handbook of Syntax (2014); “Conflation and incorporation
x
Notes on contributors
processes in resultative constructions”, in V. Demonte & L. McNally (eds.), Telicity, Change,
and State: A Cross-Categorial View of Event Structure, Oxford: OUP (2012), 252–78; and “The
manner/result complementarity revisited: A syntactic approach” (joint work with Víctor
Acedo-Matellán), in M. C. Cuervo & Y. Roberge (eds.), The End of Argument Structure?
Vol. 38, Syntax & Semantics, Bingley: Emerald (2012), 209–28.
Malka Rappaport Hovav is Henya Sharef Chair in Humanities, head of the School of
Language Sciences and Professor of Linguistics at the Hebrew University of Jerusalem. She
received her Ph.D. in 1984 from MIT. Her research interests included lexical semantics,
argument structure, and morphology. She is author, along with Beth Levin, of Unaccusativity
(MIT Press, 1995) and Argument Realization (Cambridge University Press, 2005).
Antje Roßdeutscher is a senior researcher at the University of Stuttgart. She received her
Ph.D. in 1989. Her field of research is formal dynamic semantics (Discourse Representation
Theory) with special interest in lexical semantics, syntax-semantics interface, underspecified
DRT, and word-formation. She has published in journals such Theoretical Linguistics, Linguistics and Philosophy, and Linguistische Berichte.
Yota Samioti is a Ph.D. candidate in the Linguistics sector of the Philology department at the
University of Crete. She is currently working on the syntax and semantics interface regarding
adjectival participles in Greek.
Florian Schäfer is researcher at the the collaborative research centre SFB 732 ‘Incremental
Specification in Context’ at the Universität Stuttgart. He studied General and Theoretical
Linguistics at the University of Potsdam and completed his Ph.D. dissertation on the (anti-)
causative alternation in 2007 at the University of Stuttgart. His main research interests are
located in the theories of syntax, morphology, and lexical semantics and the interaction of
these modules of grammar.
List of abbreviations
{EX[V]}
Verbal Extended Projection
#(P)
Quantity (Phrase)
√&c
Roots and Category
a
adjectiviser
A(P)
Adjective (Phrase)
ACC
Accusative
ACT
Active Voice
AER
Agentive emphatic reflexives
aff
affixal realization
AG
Agree
ALL
Allative
APPL(P)
Applicative (Phrase)
AS-nominal
Argument Structure Nominal
Asp(P)
Aspect Phrase
AspQ
Aspect/Quantity
AUG
Augmentive
C
Categorial functor
C
Conceptual properties
C(AUS)
Causative template verb
C(P)
Complementizer (Phrase)
CCS
Categorial Complement Space
CAUS
Causative
CNT
Count
Compl
Complement
COS
Change of State
D(P)
Determiner (Phrase)
DAT
Dative
DEF
Definiteness
Deg(P)
Degree (Phrase)
xii
List of abbreviations
DIM
Diminutive
Div(P)
Division (Phrase)
DM
Distributed Morphology
DO
Direct Object
DRS
Discourse Representation Structure
DRT
Discourse Representation Theory
E(P)
Event Phrase
ExP
Extended Projection
F
Feminine
GB
Government and Binding
GEN
Genitive
I(NTNS)
Intensive template verb
I(P)
Inflectional (Phrase)
ILL
Illative
infer
inference
LC
Lexicalisation Constraint
Lex(P)
Lexical Phrase
LF
Lexical Function
LP
Lexical Phonology
LPM
Lexical Phonology and Morphology
M
Masculine
MCM
Multiple Contextualized Meaning
MH
Modern Hebrew
MID
Middle Voice
n
nominalizer
N(P)
Noun (Phrase)
NAct
Non-active
NEU
Neuter
NOM
Nominative
Num(P)
Number (Phrase)
P(P)
Prepositional (Phrase)
PART
Participle
PASS
Passive Voice
PF
Phonological Function
PIC
Phase Impenetrability Condition
PL
Plural
List of abbreviations
POSS
Possessive
PREF
Prefix
PST
Past
PTCP
participle phrase
Q(P)
Quantifier (Phrase)
QiTeL
Hebrew convention : Q, T, and L represent root consonants,
i, e vocalization of the unaffixed verb stem
Q-TL-L
Hebrew convention, root consonants
Q-TQ-T
Hebrew convention, root consonants
QiTuL
Hebrew convention, root consonats, plus vocalization pattern
R-nominal
Result nominal
REFL
Reflexive
Result(P)
Result (Phrase)
RO
Reference Object
Root(P)
Root (Phrase)
S(IMPL)
Simple template verb
SC
Small Clause
SG
Singular
Size(P)
Size (Phrase)
SPE
Sound Patterns of English
Spec
Specifier
T(P)
Tense (Phrase)
TRANS
Transitive
u
uninterpretable
v
verbalizer
V(P)
Verb (Phrase)
vC
Categorizing head v
vE
Eventivizing head v
VI
Vocabulary Insertion
Voice(P)
Voice (Phrase)
XSM
Exo-Skeletal Model
xiii
1
Introduction
ARTEMIS ALEXIADOU, HAGIT BORER,
AND FLORIAN SCHÄFER
1.1 Overview
The chapters in this volume are based on talks presented at two workshops entitled
Rootbound and Roots that were held in Los Angeles (February 2009) and in Stuttgart
(June 2009) respectively. These two workshops brought together scholars from
different schools of thought to discuss and debate the nature of roots and to
investigate, primarily, their interaction, or lack thereof, with syntactic structure. By
extension, and because views differ on what roots actually are, the chapters brought
together here comment not only on the syntax of roots, but also on their phonology,
semantics, and morpho-phonological role (or lack thereof), insofar as these turn
out to bear on their interaction with syntax.
Different perspectives notwithstanding, a number of important commonalities
have emerged which, in turn, highlight what are, in our view, the core issues of
concern to syntactic “root” scholars. In our introduction we offer a survey of these
issues.1
1.2 Roots and syntactic models
The relationship between syntactic structure and syntactic terminals has always been
at the core of important debates within generative grammar. Indeed, the very nature
of syntactic terminals is not a theoretically neutral issue. Are such terminals phonologically abstract or phonologically concrete? Do they correspond to features or to
1
Alexiadou and Schäfer’s research was supported by a DFG grant to project B6 of the collaborative
research center Incremental Specification in Context at the University of Stuttgart, which also financially
supported the organization of the Stuttgart Roots workshop. We would like to thank Patrick Lindert for
his assistance in finalizing the formatting of this volume. The title of this volume reflects the particular
focus around which we and our contributors came together—our interests in investigating the role that
roots play in syntactic representations, including event structure. Thanks to Paolo Acquaviva and Mark
Liberman for titular inspiration.
2
Artemis Alexiadou, Hagit Borer, and Florian Schäfer
actual fully listed items, possibly “words” or “lexemes”? Do such terminals have
syntactic properties that inform the structure they project, or more generally, do
structures project from terminals, or, alternatively, do the properties of terminals
derive from the structure that they are embedded within, and with the structure itself
constructed some other way? And finally, are there, in actuality, “terminals” in the
commonly understood sense altogether? Differently put, is there any theoretical
reason to assume that, e.g., an N head which is embedded within some nominal
extended projection actually contains something in it that emerges from some
vocabulary list, be it formal or substantive?
Within current approaches to syntactic structure which distinguish between
functional nodes (e.g., TP, DP) and non-functional nodes (e.g., NP, AP), an additional question emerges: Are terminals embedded within functional structure formally identical to terminals which are embedded within non-functional structure?
To wit, is the formal status of, e.g., the in the dog the same as that of dog, or are they
fundamentally distinct formal entities?
In the last decade or so, and parting way with the dominant approaches in the last
three decades of the 20th century, a body of research has emerged which seeks to
equate at least some syntactic terminals with roots, with the common understanding
that roots are distinct from “words” or “lexemes”, and at least potentially, more
minimal than either. Crucially, and as a common denominator to most of these
approaches, roots differ from both “words” and “lexemes” in that in and of themselves they do not have a syntactic category.
This said, the assumption that there exist syntactic terminals which are categoryless, roots, only goes part way towards accounting for the traditional properties
typically associated with “words” (or “lexemes”). Among such properties pivotal
ones that clearly interact with the syntactic structure involve selection (categorial as
well as thematic) and lexical semantics. To wit, traditional accounts (specific modes
of execution aside) have a particular verb, say kick, select a nominal complement and
assign two thematic roles to two arguments. However, if kick is not a verb, but rather
a category-less root, is it sensible to claim that it selects a complement of a particular
type, or that it assigns roles to event arguments? Equally important are issues which
concern phonological realization. While typically assumed to be syntactically inert
and potentially listed, it nonetheless remains the case that syntactic properties do
impact phonological realization (e.g., the realization of inflection, as well as the
realization of categorizing affixes). How to model the relevant interaction between
realization and syntax in the case of roots, however, remains at least prima facie
unclear. Again, an illustration may be helpful. Within traditional accounts, broke
emerges in past tense contexts through consultation with the lexical entry of the verb
break. But if break is a category-less root, rather than a verb, where could the
relevant information be stored?
Introduction
3
In the past ten years, and within the community of scholars who subscribe to the
view that category-less terminals are syntactically useful, distinct answers have
emerged to some of these questions. The purpose of this volume is to bring some
of these differing perspectives together, in the hope of elucidating what the empirical
consequences of the differing perspectives are, specifically on the way in which
different conceptions of roots bear on the construction of syntactic objects.
Within generative linguistics, the term root has been most dominantly used in the
context of word formation, where it is frequently identified with the notion stem (but
see Aronoff 1994 for some comments on the use of the term). Roots, as thus used,
are a minimal morpho-phonological base unit, where by base we refer specifically to
an intransitive core, which may then merge with affixes, themselves, in the relevant
sense, transitive. From that perspective, dog is a root (√DOG), but so is, for example,
struct (√STRUCT) as in instruct or construction. Within the historically prevailing
traditions in generative linguistics, however, dog is a licit syntactic terminal, whereas
struct is not. Specifically, and unlike dog, struct is neither a “word” nor a “lexeme”. It
is not clear what category it is, if any, and it cannot be meaningfully claimed to
exercise any selection, be it categorial or thematic. As such, then, it is not a licit
syntactic object. (We note as an aside that in Chomsky 1970, objects which are
category-less but which have selection properties are syntactically licit objects).
Suppose, however, we dispense with the assumption that listed terminals come
with a syntactic category. A number of important potential questions and consequences emerge immediately. The first set of questions concerns the formal nature of
roots. Harking back to an important debate within research on word formation and
lexical representatins, roots are fundamentally syntactico-semantic in nature (and
thus on a par with Lexemes, as in Beard 1995 and others), or, to the contrary, are
they fundamentally phonological in nature, as in Aronoff (1976)? And are they,
possibly, a conjunction of semantic and phonological properties, as might be suggested (albeit not for roots as such) in Allen (1978), Pesetsky (1982), and Kiparsky
(1982a, b, 1997)? Do roots have any syntactic properties (e.g., selection) which impact
their syntactic environment (e.g., Marantz 1997, harking back to Chomsky 1970) or,
are they possibly devoid of any syntactic properties altogether (e.g., Borer 2003), and
if the latter, how do selectional effects emerge?
And finally, are there actually veritable listed terminals which we may call roots,
and which merge, syntactically, as such? The latter claim has been challenged from
two rather different perspectives. Thus, in Ramchand (2008), neither roots nor other
listed (non-functional) terminals are syntactically present, and “words” are but the
realizations of complex structures in which terminals are featural in nature. For De
Belder and van Craenenbroeck (2011) on the other hand, terminals which correspond to roots do exist, but they are not populated by listed items. Rather, they are
null sets, whose existence is mandated by the properties of Merge (and specifically,
First Merge).
4
Artemis Alexiadou, Hagit Borer, and Florian Schäfer
An altogether different set of questions emerges once we consider the potential
interaction between roots as category-less items, and the properties of complex
words. To wit, if √STRUCT is a possible syntactic terminal, then instruct or structure
must be formed syntactically. Similarly, if √DOG is a syntactic terminal, dogs must
be formed syntactically. However, the formal nature of the syntactic operations that
can give rise to instruction or dogs are by no means agreed upon. One option would
be to assume that not only roots, but also affixes are syntactic terminals which merge
with the root. Such an approach would involve importing into the syntax the
configurational approaches to word formation otherwise advanced, within an autonomous morphological system, by Lieber (1980), Williams (1981), Selkirk (1982), and
more recently Ackema and Neeleman (2004) among others. On the other hand, one
may adopt a realizational approach, thereby allowing the formation of complex
words not through the presence of additional terminals, but through the spelling out
of syntactic distinctions on the tree, thus importing into the syntax approaches such
as those of Beard (1990, 1995) and Anderson (1992) (the latter for inflection only),
among others. Can these different approaches, once integrated into the syntax, be
shown to make different predictions, and can they be shown to overcome some
rather recalcitrant issues that have, traditionally, provided evidence for removing
word formation from the syntax altogether?
The issue here, we note, concerns not only the properties of roots, but also the
properties of affixes. Specifically, if one subscribes to the view that words have an
internal hierarchical structure, it must be the case that not only roots but also affixes
are syntactic terminals. But if so, what is the difference, if any, between affixes and
roots? The question doesn’t emerge, of course, within the realizational system, quite
simply because affixes as such do not exist. For a realizational (syntactic) approach,
however, the task is how to specify the presence of relevant syntactic properties that
condition particular realization (e.g., past tense), but which nonetheless do not
translate into structural complexity.
Finally, within the types of approaches under consideration here, roots are devoid
of category by assumption. However, under the plausible claim that in constituents
such as the dog, dog is at least in some relevant sense a noun, how does (the) dog,
coextensive with the category-less root √DOG, come to be a noun, and by extension,
how is categorization in general accomplished?
1.3 Specific issues and the structure of this volume
Importantly, this is not a book about word formation as such. Rather, it is about the
merits and the consequences, or lack thereof, of postulating category-less syntactic
terminals. For that reason, we did not attempt to integrate into this volume specific
debates on word formation which do not, as such, have syntactic ramifications,
nor have we included perspectives that postulate a fundamentally non-syntactic
Introduction
5
component of word formation. Discussions of the morpho-phonological properties
of roots or their lexical semantics were thus included insofar as they were couched
within the fundamental claim that the root is a valid syntactic object, either as a
terminal, or as a relevant unit of syntactic information.
Relative to the contributions in this volume, four main foci emerge from our brief
introduction. These foci do not, as such, divide the chapters of this volume into
groups, but rather, cut across them. More frequently than not, discussions of root
properties and their interaction with syntactic structure are closely interlaced,
resulting in a network of interconnections between the different chapters.
1.3.1 The meaning of roots in isolation and the selection of arguments
An important question that has been widely discussed in the recent literature is how
much meaning roots have in isolation, and to what extent that meaning informs
their syntactic merging possibilities. More specifically, we can identify the following
general approaches:
.
.
.
Quite independently of whether or not roots have meaning, as such, some
scholars subscribe to the view that roots do select arguments (e.g., Marantz
1997, 2000), thus informing some aspects of their syntactic context. Similarly for
Harley (2009b, c), although Harley suggests that root selection may, at times, be
mediated by formal structure.
Other authors, on the other hand, have argued explicitly that roots may have
meaning from which some aspects of their syntactic context may emerge. Thus
Rappaport Hovav and Levin (2010) make a strong case that root meaning
consists of a limited number of contrastive ontological properties (e.g., manner
vs. result). Embick (2004a), as well, claims that roots may have formal meaning
properties (specifically stative vs. eventive). Embick as well as Rappaport Hovav
and Levin further argue that meaning distinctions associated with roots inform
their syntactic merger possibilities, their potential categorization array, and the
availability of arguments. This general position is adopted by several of our
contributors, while others explicitly argue against it.
Finally, some scholars argue that not only do roots not have any meaning in
isolation, all (grammatical) meaning is associated with constituents larger than
roots (see Acquaviva 2008a and Borer 2013). For these scholars, the absence of
meaning correlates directly with the absence of arguments or any selection
properties. That position, as well, has been adopted by some of our contributors.
Let us consider some of our chapters from the perspective of this particular
debate. The claim that structure, rather than root ontology, determines interpretation is the point of departure for Acedo-Matellán and Mateu’s contribution, where,
in line with Hale and Keyser (1993, 2002) and Borer (2003, 2005), it is assumed that
argument structure is computed on the basis of syntactic configuration. Extending
6
Artemis Alexiadou, Hagit Borer, and Florian Schäfer
this approach, Acedo-Matellán and Mateu propose that the properties of roots are
contingent on their syntactic position. Appealing to a crucial distinction between the
conceptual and the syntactic properties of roots, they show that the conceptual
meaning of roots are opaque to the syntactic computation and hence must be
excluded from those aspects of the semantic interpretation that are built structurally.
As a consequence, ontologies of roots are grammatically spurious. Rather, what
might appear, intuitively, to be a grammatically active root meaning, such as result
or manner, is in fact an interpretation that is associated with a well-defined syntactic
structure. Even more specifically, they suggest that grammatical result interpretation
emerges whenever the root merges as the complement of a recursive P projection.
Grammatical manner interpretation, on the other hand, emerges whenever the root
is adjoined to v. Arguing directly against claims made by other contributors to this
volume (including Alexiadou, Anagnostopoulou and Samioti, Levinson, RappaportHovav, and Roßdeutscher), they strongly deny the claim that ontological classifications of roots can inform the linguistic derivation or place any conditions on it. To
the contrary, it is the structural position occupied by the root in the syntactic event/
argument structure which determines its properties, including those that appear
linked to meaning. As such, this conclusion is compatible with the Exo-Skeletal
approach, otherwise endorsed, in this volume, by Borer as well as by De Belder,
Faust, and Lampitelli.
A contrasting perspective on this same issue is offered in Alexiadou’s contribution, which examines the complex distribution of the prefix afto- “self ” in Greek
(and its Romance equivalent). She argues that the behavior of that prefix provides
important insights into the relevance of root ontology and sheds light on its
interaction with syntax in general and with the relevance of the manner/result
dichotomy in particular. It further sheds important light on the nature of the
Voice node, which, in Greek, hosts non-active morphology. Specifically, aftocombines neither with naturally reflexive predicates nor with mono-eventive predicates in general. If we assume that the properties of such predicates are contingent
on the presence of manner roots, and that manner roots merge as modifiers of v, to
give rise to a mono-eventive structure, and if we further assume that afto- indicates
the presence of a bi-eventive structure, then these effects can be explained. But if that
is the explanation, then it crucially depends on the claim that the ontology of the
root does translate, directly, into syntactic delimitations of its merger environment.
Ultimately endorsing similar conclusions, Doron investigates a particular subclass
of Hebrew adjectival passive participles formed in the causative template and shows
that their interpretation always includes an implicit external argument, even when
the external argument of the (active) verbal source is optional. As such, the behavior
of these passive participles parallels that of passive verbs in Hebrew, which also
obligatorily include implicit external arguments. The behavior of adjectival passive
participles in other morphological templates, by contrast, parallels the behavior of
Introduction
7
middle-voice verbs, which exclude external arguments. The conclusion drawn by
Doron is that the structure of adjectival passives must contain a Voice node, and that
the value of that Voice node is contingent on the voice values of the corresponding
verbs. Crucially, roots can be classified into various ontological types, and their
ontology correlates with the type of participles they build. Assuming the ontology
proposed, in essence, by Embick (2004a), she proposes that dynamic roots only give
rise to resultative participles, whereas roots that denote states may give rise to both
stative and resultative participles. With this distinction in place, Doron shows that
for the causative morphological template, the only available voice value is passive,
i.e., one which obligatorily introduces an (implicit) external argument. It thus
emerges that resultative participles in the causative templates must include a Voice
projection. It similarly emerges that the participial/adjectival structures proposed by
Kratzer (1994) and Embick (2004a), which do not include a Voice head, cannot
account for this correlation.
Levinson’s contribution, likewise, argues for the syntactic relevance of root ontology. In her contribution she explores the connections between the semantic properties of roots and morphosyntactic properties and argues that some correlations
between interpretation and morphosyntax can be derived from the semantic types
of the roots that form the lexical core of verbs. This idea in itself is not new, as for
example Rappaport Hovav and Levin (1998) argue for the existence of meaning
“constants” which determine aspects of a verb’s syntactic realization. However, in a
departure from Rappaport Hovav and Levin (1998), Levinson integrates this
approach into a Distributed Morphology approach to syntactic word formation.
The syntactic execution, as it turns out, gives rise to interesting predictions regarding
the interpretation of composition with roots. By putting forth an explicit formalization of verbal lexical decomposition, predictions concerning roots and composition
with them are shown to be borne out. In addition to contributing to our understanding of the ontology of roots, the chapter shows that apparent verb polysemy
frequently involves structural ambiguity which emerges in the context of root
polysemy. If on the right track, Levinson’s findings thus provide evidence that
roots are not semantically vacuous in isolation, contra Marantz (1996), Borer
(2005, 2013, this volume), Acquaviva (2008a, this volume), and Harley (2009a, b, c).
Rappaport Hovav’s chapter takes as its starting point, yet again, the claim that the
meaning components of verbal roots can inform grammatical structure. Specifically,
the chapter proposes a scalar analysis of verbs which highlights the structural
parallels between the semantics of change-of-state verbs and directed motion verbs
which is in turn supported by grammatical and interpretive properties shared by
both classes of verbs. In particular, scalar change verbs in both domains typically do
not encode a manner component, demonstrating what Rappaport Hovav and Levin
(2010) call manner/result complementarity. In turn, most verbs that do not lexicalize
scalar changes are shown to be manner verbs. The chapter further demonstrates that
8
Artemis Alexiadou, Hagit Borer, and Florian Schäfer
it is possible to isolate those components of the scalar semantics of event descriptions that are lexicalized in the root and those that are contributed by constituents in
the syntactic context of the root. Distinguishing those aspects of the event description which are lexically encoded from those which are not leads to a deeper understanding of the argument realization and interpretive properties of various classes of
verbs and lends further credence to the claim that it is possible to isolate and
explicate the grammatically relevant meaning components associated with a verbal
root. From Rappaport Hovav’s perspective, lexicalized meaning is a property of
roots, determined in the lexicon, and not structurally. This approach thus assumes,
as others do in this volume, an ontological classification of roots, which is grammatically relevant, a position explicitly denied in other contributions to this volume.
Finally, and yet again arguing for the significance of root meaning, Roßdeutscher
investigates the contribution of roots to the syntactic and semantic properties of
verbs. The leading question of her investigation is “How can the semantics of verbs
be constructed from their roots?” Roßdeutscher assumes that roots have a semantics
which is the source of argument structure and which determines whether they can
be selected by certain functional heads such as v (verbalizer), n (nominalizer) or
a (adjectivizer). For instance, eventive or “manner” roots like run are simple event
types which merge (directly) with v; the property root dry, on the other hand, creates
an argument for the bearer of the property dry to give rise to the de-adjectival verb to
dry. In contrast, entity-roots typically fill argument slots that are created by other
roots. E.g., the sortal root line in to underline a word satisfies one of the two
argument slots created by the preposition-like root under. (The other argument is
contributed by the direct object of the verb underline.) In German, where verb
formation of this kind is common, sortal roots fill argument slots of preposition-like
heads. Notwithstanding these results, Roßdeutscher notes that roots may enter word
formation operations that are incompatible with their ontology. In such cases, she
proposes, interpretation emerges as a result of the root being coerced into the
properties of the forming operations.
1.3.2 The syntax of roots
If roots lack a categorial specification, the way in which they come to be associated
with one, if indeed they do, is a pivotal question. Relative to this question, at least the
following approaches have emerged:
.
Syntactic categorization is achieved through the existence of specialized category labels, such as n, v, a which merge with an otherwise category-less roots, as
in Marantz (2000) and subsequent work. In such approaches, the root itself
never has a category. Rather, the category is associated with the node that
dominates it. As a consequence, any categorized constituent is at least binary
branching. Most contributions in our volume assume this type of categorization.
Introduction
.
.
9
Alternatively, syntactic categorization is an artifact of syntactic context. This
approach was first put forth in Chomsky (1970) for lexical entries in general and
regardless of their morphological complexity (and see also Marantz 1997). More
recently, it has been advanced, specifically for roots, by Borer (2003, 2005, 2013)
and subsequent work (cf. Alexiadou 2001, De Belder 2011b, among others).
Within that system, a root has a post-facto category as determined by a selecting
functor (e.g., the root becomes equivalent to V when selected by T, and
equivalent to N when selected by D) and categorized constituents need not be
binary branching.
An interesting twist relative to both approaches is that of Acquaviva (2009, this
volume), who accepts that roots are category-less, but nonetheless assumes that
the categorial status of roots, within the syntax, is by default nominal, unless
otherwise structured.
Chapters in our volume that touch specifically on these issues are those of Acquaviva, Borer, and De Belder, Faust and Lampitelli.
Acquaviva’s chapter investigates nouns as a primary lexical category. In his
contribution, he distinguishes individuation as discourse referent at the DP-level
from individuation as an abstract category, and argues that lexical nouns name the
latter, rather than the former. In turn, the granularity and the part-structure of the
denotation domain, including individual reference, emerge from the grammatical
structure occupying the middle field between the outer DP-level and the innermost
N- and root-level. Bringing forth empirical evidence to support his claim, Acquaviva
spells out the descriptive and explanatory advantages of his approach, giving rise to a
strong, falsifiable claim on what can and what cannot be a common noun in a
natural language. The pivotal role of nouns notwithstanding, Acquaviva nevertheless
argues that locating nominality (directly) on roots is over-simplistic and ultimately
wrong. Insofar as they differ from nouns, roots should not be stipulated to have the
semantic function of nouns; instead, their function is to differentially label the
syntactic construction that corresponds to a noun, and which interfaces the Conceptual/Intentional cognitive system as the name of a category concept.
Borer’s contribution focuses on the categorial properties of roots and proposes
that just like event structure, these emerge in the context of particular functional
structure and as a consequence of it. For Borer, functors, whether segments of
extended projections or derivational categorizers, are viewed as elements that partition the categorial space. Thus D and # (or Num) as well as the rest of the members
of the nominal extended projection project a nominal structure (say DP), and define
the domain of their complement as equivalent to N. A root residing at the bottom of
such an extended projection doesn’t need to merge with a category label (in turn
zero realized) nor undergo conversion. Rather, it becomes N-equivalent by virtue of
merging with a functor that defines its complement space as N. The model of
categorization outlined is contrasted with the model of categorization advanced by
10
Artemis Alexiadou, Hagit Borer, and Florian Schäfer
Chomsky (1970) and researchers working within the Distributed Morphology model.
Borer explicitly argues against linking the emergence of a category to zero-realized n,
v, and a, showing that if such zero-realized categorial heads are assumed across the
board, a number of very unfortunate formal consequences emerge, including the
failure to correlate morphosyntactic complexity with morpho-phonological complexity, and the need to postulate a host of syntactic locality restrictions which apply
across the board to zero-categorizers, but never to realized categorizers, and which
conspire to make the actual existence of such unrealized nodes virtually impossible
to detect (see, most recently, Embick 2010).
Syntactic properties of roots vs. those of (otherwise) categorized constituents are
likewise at the core of the contribution made by De Belder, Faust, and Lampitelli.
Looking at the properties of diminutives in Italian and Modern Hebrew, they argue
that there are two different positions for their merger. The first position involves the
functional domain of the noun, and hence takes as its input an already categorized
constituent. The second position involves merger with the root, and hence, from the
authors’ perspective, below the categorial head. The two positions differ with respect
to their productivity, with respect to the emergence of meaning compositionality,
and from the perspective of the word-formation strategy used. These differences, in
turn, can be accounted for by appealing to the distinct merge properties. Specifically,
the authors propose that the first categorial head demarcates a boundary between
two distinct domains. The domain below that head allows for non-compositional,
lexically listed meaning, whereas the domain that includes the categorial node hosts
functional projections whose meanings cannot be idiosyncratic.
1.3.3 The meaning of roots in context
However derived, if complex words are syntactically complex constituents, some
concern must be given to the emergence within such complex words of noncompositional meaning, i.e., meaning associated with a complex constituent
that cannot be computed from the meaning of its parts, and with transmission
with the meaning “gearbox” being a typical example. Syntactic approaches to noncompositionality do agree on the need to define a local syntactic domain within
which compositionality need not apply. What that domain might be, however, is not
agreed upon.
Some authors have argued that the domain of non-compositionality converges
with the domain of first categorization, see Arad (2003), Embick and Marantz
(2007a), and Embick (2010), as well as De Belder, Faust, and Lampitelli (this
volume). Alternatively, the domain of non-compositionality is defined by the presence of a functional bracket, as proposed in Borer (2013). And finally, other
domains have been proposed, e.g., by Alexiadou (2009) and by Harley (2009b),
each with its own distinct predictions.
In addition to the brief discussion already summarized in our discussion of De
Belder, Faust and Lampitelli (this volume), this question is at the heart of the
Introduction
11
contribution by Anagnostopoulou and Samioti. Anagnostopoulou and Samioti
investigate the conditions under which idiosyncratic meaning arises in word formation. More specifically, they investigate the hypothesis put forth in Arad (2003, 2005)
and Marantz (2001, 2007a) that idiosyncratic meaning is constrained by categorization. Their investigation is focused on Greek deverbal adjectives and adjectival
participles which provide extensive evidence for attachment below vs. above higher
heads (e.g., little v, Voice) of the adjectivizing/participial suffixes -tos and -menos,
respectively. Anagnostopoulou and Samioti propose, following Harley (2005), that
roots with fixed meaning fall into basic ontological types, naming events, things, or
states. As for roots that fail to have such specific fixed meaning, a categorial head
serves to provide an ontological classification, thereby giving rise to a fixed “meaning”, which is then retained throughout the derivation. Finally, the interpretation of
idioms provides support for the view that the head which delimits the domain for
idiomatic interpretations of adjectival participles and deverbal adjectives in Greek is
Voice, equivalent to the (little) v head which introduces the agent, as was proposed
originally in Marantz (1996, 1997). The conclusion then is that across the board, it is
the agent-licensing v that serves as a boundary for special meanings of both phrasal
idioms and complex words.
1.3.4 Phases and root phonology
In a hierarchical, syntactic approach to complex words, not only roots, but also
affixes must be listed. The common approach within root-based models is that while
both roots and affixes are listed, the lists are distinct and the listed items have
distinct properties. To illustrate, in (most) Distributed Morphology accounts, roots
are category-less but are listed with some phonology (see, e.g., Embick and Halle
2005; Embick 2010) while affixes are Vocabulary Items which can certainly be
associated with formal categorial properties, but which are, in turn, subject to late
insertion. Similar claims are made in Borer (2003, 2005, 2013), who likewise assumes
that roots are inherently linked with phonological information, but functional
vocabulary is subject to late realization. Different perspectives, are certainly possible.
Thus Harley (2009b) subscribes to the view that affixes are categorial and roots are
not, but nonetheless holds that both are subject to late insertion. In contrast, De
Belder (2011b) subscribes to the view that (derivational) affixes are in actuality roots,
and that both roots and affixes are devoid of any phonological properties and are
subject to late insertion.
As is well known, however, Vocabulary Items, and specifically affixes, come, in
English, in two varieties which are quite distinct from each other. Class 1 affixes
(+ boundary) which allow assimilation across an affix boundary, and which allow for
cyclical stress assignment, and class 2 affixes, (# boundary), which do not. It is not
clear, however, how any of the schematic pictures presented above can account for
this, and indeed, as Lowenstamm (this volume) points out, Distributed Morphology
12
Artemis Alexiadou, Hagit Borer, and Florian Schäfer
fails directly in offering no account for English Stress Shift, e.g., the emergence of
progressive cases such as átom, atómic, atomícity (but atómicness).
In turn, any discussion of affix types in English must take account of the
extremely influential Level Ordering Hypothesis (see primarily Lexical Phonology
and Morphology, Kiparsky 1982a, b, 1997). According to the Level Ordering Hypothesis, boundary types define two distinct domains of rule application which define not
only phonological rule application, but also semantic and syntactic characteristics.
Class 1 affixes define the inner domain—they are closer to the root/stem, they may
attach to non-words (and hence potentially non-categorized roots), and the combination may give rise to non-compositionality. In contrast, or so the claim goes, Class
2 affixes only attach to words (and hence by assumption to categorized constituents),
merge outside Class 1 affixes, and do not allow non-compositionality. If on the right
track, this picture comes very close to Chomsky’s notion of phase, insofar as phases
provide a natural juncture for the simultaneous realization of phonological, semantic, and syntactic properties. The question that we must ask, then, is whether the
formation of complex words can be usefully characterized by appealing to phases,
and if so, what the relevant phase is. Thus Embick (2010) explicitly proposes that
root categorization creates a phase domain, thereby defining a domain for the
application of phonological processes as well as for the possible emergence of noncompositional meaning (see the discussion above on the chapter by Anagnostopoulou and Samioti). A different phase-based approach is put forth in Borer (2013),
subscribing to the view that every instance of merge is, effectively, a phase, but
subject to extension through head remerger.
Two chapters in this volume specifically address the question of the interaction of
phases with roots. While one of them (Gallego) is primarily concerned with the
domain for argument structure determination, the other (Lowenstamm) takes on
directly the task of characterizing affixes vs. roots as based on Phase Theory.
In his exploration of the relationship between the properties of roots and the
theory of phases, Gallego asks whether all the properties of roots (by assumption
non-phase heads) can be derived from phase heads, by assuming grammatical
formatives such as category labels (n, v, etc.), φ-features (gender, number, person),
and structures that give rise to argument interpretation. That category and
φ-features contribute to the emergence of phases is a fairly standard assumption
in the current literature (see Chomsky 2007, 2008, Richards 2007), but the possibility
that argument structure is dependent on the presence of phase heads has not been
considered. Gallego explores this perspective following Gallego (2010, 2012) and
argues that it follows straightforwardly from Chomsky’s (2000 et seq.) Phase Theory.
Specifically, he proposes that argument structure is projected after the relevant
category-inducing morpheme merges with the root. In line with Chomsky’s (2007,
2008), he also assumes that phase heads provide non-phase heads with properties
through a process of feature inheritance, and that such inheritance is only forced in
Introduction
13
the case of unvalued φ-features. By assumption, now, only light verbs contain
unvalued φ-features, and the emerging prediction would be, rather contrary to
fact, that only roots that are dominated by light verbs could take arguments.
Addressing this apparent problem, Gallego proposes a distinction between “syntactic”
and “conceptual” arguments, linking the former to Chomsky’s unvalued φ-features,
and the latter to the conceptual content of roots. As an added bonus, Gallego notes
that as n never has unvalued φ-features; it never licenses syntactic arguments.
As noted already, Lowenstamm’s starting point is the observation that Distributed
Morphology accounts of vocabulary insertion have failed, altogether, to shed any
light on the nature of cyclic rule application, such that it gives rise to átom, atómic,
atomícity. A resolution as well as a conceptual simplification of the system are
available, Lowenstamm proposes, if one assumes the following: (a) stress-shifting
affixes are effectively transitive roots, rather than categorial labels; (b) the domains of
phonological rule application and spellout cannot be characterized usefully along the
lines of + vs. # boundaries. Rather, phases, or domains for the application of
phonological rules, should instead be defined on roots. Rules such as stress shift
would now apply to the most deeply embedded root, then reapply on the domain
defined by the next adjacent higher root, and so forth. The domain of the roots, now
to include cyclic “affixes”, would constitute Phase 1. The concomitant conclusion
would be that cyclic phonology is the hallmark of category-less roots, which,
together, constitute Phase 1, a domain in which categorially marked elements (e.g.,
non-cyclic affixes) are altogether excluded from that phase. As a consequence, affixes
no longer need to be divided into cyclic (i.e., + affixes) or non-cyclic (i.e., # affixes) as
such. Rather, our inventory would consist of roots and of affixes. The former would
constitute the innermost domain, the first phase, where merged elements would
include not only classical roots, but also, for example, √IC, √ATION. In that domain,
not only would phonological rule application be cyclical, but meaning could be noncompositional because of the absence of category labels. Beyond Phase 1, however,
affixes are categorial, rule application is never cyclic, and meaning, in the presence of
categorial labels, is always compositional.
2
From syntax to roots: A syntactic
approach to root interpretation
VÍCTOR ACEDO–MATELLÁN AND JAUME MATEU
2.1 Introduction
In recent years the important idea has been advanced that the interpretation of
arguments takes place on the basis of their syntactic position in event structure (see
Borer 2003, 2005) or, alternatively, argument structure configurations (see Hale and
Keyser 1993, 2002). For instance, to put it in Borer’s (2003:32) terms, “it’s not the case
that Agents project externally (universally), but rather, that nominal expressions
which project externally must be interpreted as Agents.” In this chapter we extend
this idea to the interpretation of roots: roots are structurally interpreted depending
on the position they occupy in the syntactic configuration. By drawing a crucial
distinction between the conceptual and the syntactic interpretations of roots, we will
take pains to show that roots, from the conceptual point of view, are opaque to the
syntactic computation and, hence, to the structural semantics of the linguistic expression. As a result, grammatically relevant ontologies of roots become spurious.*
Our chapter is organized as follows. In section 2.2 we present a neoconstructionist
view of argument structure. In section 2.3 we deal with the thematic interpretation of
roots with respect to the structures they appear in, focusing first on the so-called
Manner/Result complementarity and, second, on the syntactic properties of instrument-naming verbs. We provide overall conclusions in section 2.4.
2.2 Theoretical framework
We assume that argument structure is syntactically built: it is brought about by
the application of the operation Merge to primitive relational elements and
* We are very grateful to two anonymous reviewers for their constructive comments and suggestions.
This research has been sponsored by grants FFI2010-20634, FFI2011-23356 (Ministerio de Ciencia e
Innovación-FEDER), and 2009SGR1079 (Generalitat de Catalunya).
From syntax to roots: A syntactic approach to root interpretation
15
non-relational elements. The non-relational elements may be either roots or fullfledged DPs.
Relational elements alone may project structure. We propose the existence of two
such elements for the building of what has been called argument structure: v, an
event-encoding relational element, and p, an adpositional-like element. We, along
with Harley (2005) or Marantz (2005), assume that the semantic “flavors” v may
adopt arise structurally: i.e., for instance, a little v taking a DP specifier and a DP or
root complement is interpreted as DO, while one taking a DP specifier and a Small
Clause Result (cf. Hoekstra 1988) complement is interpreted as CAUSE. In a parallel
fashion, we follow Hale and Keyser’s (2002:ch. 7) claim that the so-called central and
terminal coincidence relations encoded in the element p can be read off a single
p-projection and a double p-projection, respectively.1 Accordingly, the semantic
flavors inherent to these relational argument structure elements can be derived
from configurational properties.2
(1)
Relational elements
v (eventive head) p (adpositional head)
Non-relational elements cannot project structure. Hence, roots cannot take complements or specifiers, and there does not exist any syntactic object like a RootP (but see
Marantz 1997 or Harley 2005, where roots are allowed to take complements; see Gallego,
this volume, for the idea that all elements manipulated by syntax are relational).3
The structures projected by relational elements plus their own intrinsic value (as an
event or as an adpositional relation) yield the structural semantics of the linguistic
expression (see Harley and Noyer 2000).4 Some relevant examples of syntactic argument structures that we will be dealing with in this chapter are the following ones:
1
According to Hale and Keyser (1993, 2002), a terminal coincidence relation (e.g., cf. to, out of, from,
etc.) involves a coincidence between one edge or terminus of the theme’s path and the place, while a
central coincidence relation (e.g. cf. with, at, on, etc.) involves a coincidence between the centre of the
theme and the centre of the place. For the correlation between terminal/central coincidence relation and
telicity/atelicity, respectively, see Mateu (2002).
2
As pointed out to us by Cedric Boeckx (personal communication), these two relational elements
could eventually be reduced to a single relational element capable of building predicates. The surface
distinction, then, between verbs and the categories representing p—mainly adpositions, but also adjectives:
see Mateu (2002), Amritavalli and Jayaseelan (2003), Amritavalli (2007), and Kayne (2009), for the claim
that adjectives can cross-linguistically be analyzed as resulting from the incorporation of a non-relational
element/a Noun to an adpositional marker—would be strictly morphological, and would depend again on
configurational factors: i.e., such an element, when merged with T, would surface as a verb, and it would
surface otherwise when not merged with T.
3
The unavailability of structures involving a RootP could be treated as the result of a crash at LF, rather
than the result of an intrinsic syntactic incapability of roots to project. See Gallego (this volume) for a
phase-theoretic account of the inability of roots to take (internal) arguments. See De Belder (2011b) and De
Belder and van Craenenbroeck (2011) for a syntactic, Merge-based account of the nature of roots, which,
according to these authors, are to be found exclusively as complements.
4
Under the view sketched in footnote 2, there not being any ontological difference between v and p, the
structural semantics would correspond solely to the semantic import of the configuration.
16
(2)
Víctor Acedo–Matellán and Jaume Mateu
a. Unergative creation/consumption event5
Sue danced: [vP [DP Sue] [v0 v √dance]]
Cf. Sue did a dance: [vP [DP Sue] [v0 v [DP a dance]]]
b. Transitive event of surface-contact6
He pushed the cart: [vP [DP He] [v0 v [pP [DP the cart] [p0 p √push]]]]
Cf. He gave the cart a push: [vP [DP He] [v0 v [pP [DP the cart] [p0 WITH [DP a
push]]]]]
c. Transitive event of change of state/location7
The strong winds cleared the sky: [vP [DP The strong winds] [v0 v [pP [DP the
sky] [p0 p [pP p √clear]]]]]
Sue shelved the books: [vP [DP Sue] [v0 v [pP [DP the books] [p0 p [pP p
√shelf]]]]]
Cf. Sue put the books on the shelf: [vP [DP Sue] [v0 v [pP [DP the books] [p0 p [pP
on [DP the shelf]]]]]]
d. Unaccusative event of change of state/location
The sky cleared: [vP v [pP [DP The sky] [p0 p [pP p √clear]]]]
Cf. He went to Paris: [vP v [pP [DP He] [p0 to [pP p (= AT) [DP Paris]]]]]
If a root is adjoined to the v head a complex event emerges, such as the following
ones (see McIntyre 2004, Embick 2004a, Mateu 2008, i.a.):
(3)
a. Complex creation event
Sue baked a cake:
[vP [DP Sue] [v0 [v √bake v] [DP a cake]]]
b. Complex transitive event of change of state/location
Sue hammered the metal flat:
[vP [DP Sue] [v0 [v √hammer v] [pP [DP the metal] [p0 p [pP p √flat]]]]]
Sue sneezed a napkin off the table:
[vP [DP Sue] [v0 [v √sneeze v] [pP [DP napkin] [p0 p [pP off [DP the table]]]]]]
5
See Volpe (2004), for the proposal that consumption verbs (e.g., eat, drink, smoke, etc.) are unergative
in Hale and Keyser’s (1993, 2002) sense. As is well-known, the latter claim that unergatives are transitive
verbs associated to creation processes (i.e., [do x]).
6
We follow Hale and Keyser’s (2002:44) proposal to treat this kind of predicate as featuring an abstract
preposition of central coincidence—that is, a single p projection. See also Mulder (1992:59), who claims
that push-verbs receive a Small Clause analysis. But see Harley (2005) for a different analysis.
7
Unlike Hale and Keyser (2002), we do not assign different syntactic argument structures to denominal
verbs like shelve and deadjectival verbs like clear: see our footnote 2, for the claim that adjective is not a
primitive element. The double p-structure, i.e., the one associated to terminal coincidence relation, can be
regarded as our version of Hoekstra’s (1988) Small Clause Result (cf. also Ramchand and Svenonius 2002, i.
a.).
From syntax to roots: A syntactic approach to root interpretation
17
c. Complex unaccusative event of change of state/location
The candle blew out:
[vP [v √blow v] [pP [DP The candle] [p0 [pP out X]]]]8
Sue danced into the room:
[vP [v √dance v] [pP [DP Sue] [p0 -to [pP in- [DP the room]]]]]
Importantly for our present purposes, non-relational elements are assigned a
particular interpretation depending on the position they occupy in the abovementioned structures, as either specifier, complement or adjunct of a v head, a single p
projection or a double p projection.9 For example, some relevant syntactic positions
of argument structure are interpreted as follows: spec-v is Originator, compl-v is
Incremental Theme, adjunct-v is Manner, spec-p is Figure, compl-single p is Central
Ground, and compl-double p is Terminal Ground (Result).
As for the nature of roots, they are constituted of two sets of properties: C and F.
F is a set of phonological properties. C is a set of conceptual properties readable only
at the C-I interface and unable, therefore, to determine the syntactic computation
in any way.10 The set of Cs of the roots contained in a linguistic expression in
combination with its structural semantics provides the semantic dimension of that
linguistic expression. In relation to the nature of C, Marantz’s (2001) distinction
between semantic properties and semantic features, shown in (4), becomes relevant:
(4)
“word (really, root) meanings don’t decompose; the semantic properties of
words (= roots) are different from the compositional/decompositional semantic features expressed through syntactic combination” (Marantz 2001:8).
8
The analysis of (3c) captures Svenonius’s (1996) proposal, assumed by Hale and Keyser (2002:229–
230), that bare particles like out can be analyzed as prepositions that incorporate a complement (i.e., the
Ground, represented by X in the example): such a proposal is coherent with maintaining the birelational
nature of p.
9
There are important restrictions on the distribution of DPs and roots in argument structure
configurations: while the former can be merged as specifiers but not as adjuncts to v or p, the latter can
be merged as adjuncts to v or p but not as specifiers. We take the reason for these restrictions to be
phonological and having to do with the different status of DPs and roots with respect to the transmission
of a phonological matrix to a phonologically empty head (v or p)—the operation referred to as conflation
(Hale and Keyser 2002); but see Section 3.1 for an important qualification. On the one hand, roots are
assumed to be copied into empty matrixes to be PF-licensed and this they can accomplish only if directly
merged with an empty head, either as complements or adjuncts. On the other hand, empty heads must be
provided a phonological matrix to be PF-licensed (unless a default Vocabulary Item is inserted therein),
and acquire one from the nearest possible non-relational element; however, if a DP is adjoined to the
empty head, becoming the nearest possible non-relational element, it cannot provide any phonological
matrix, since it is phonologically non-defective. See Acedo-Matellán 2010 for more related discussion.
10
An anonymous reviewer points out that for some roots the precise nature of C or F is not clear. For
example, what is the set C for the root -ceive in receive, deceive, conceive, etc.? And what is the set F for the
common root in think/thought? Since, for space reasons, we cannot provide a full answer to these
questions, we refer the reader to Borer (2005:351f.) for the interpretation of ceive-like roots as part of
idiomatic verbs and Acquaviva (2008a:15f.) for the proposal that roots name, rather than mean. On the
other hand, see Halle and Marantz (1993:129f.) for an analysis of root allomorphy of the think/thought kind
in terms of readjustment rules and Siddiqi (2009:27–63) for an analysis in terms of competition.
18
Víctor Acedo–Matellán and Jaume Mateu
Marantz’s semantic properties correspond to our set C, while his semantic features
are what we have termed the structural semantics of a linguistic expression. Following Marantz (1995) we take the meaning of a linguistic expression to be the sum of its
structural semantics and the semantic properties of the roots integrated in that
expression.
In this chapter we will be concerned with the set C; in particular, we would like to
show that C never tampers with syntax: semantic properties of roots never determine the syntactic computation and, hence, the structural semantics of the linguistic
expression.11 As a consequence, roots cannot be distributed into grammatically
relevant ontologies according to the type of C they encode, since there are no
grammatically relevant types of Cs. Quite on the contrary, roots receive a semantic
interpretation according to the syntactic position where they are merged (see above).
To show this we review some relevant cases where the structural semantics of the
linguistic expression is seemingly sensitive to properties of C, in particular to
whether C encodes a result, a manner or an instrument in the conceptual scene
evoked by the expression where it appears. By showing that even in these cases the
structural semantics is determined solely by the syntactic structure and the value of
the relational elements which project it, the claim is underpinned that C, the
conceptual dimension of roots, is opaque to the computation (see Borer 2003, 2005
and Åfarli 2007, i.a., for further discussion on so-called neoconstructionist
approaches).
2.3 Syntax determines how roots are thematically interpreted
In this section, we present two case studies that show that the lexical-semantic
classification/ontology of the root is not what predetermines the syntactic derivation.
Rather, we claim, it is the structural position the root occupies in the syntactic
argument structure which determines its thematic interpretation. First, we review
Mateu and Acedo-Matellán’s (2012) main arguments for a syntactic treatment of the
so-called Manner/Result complementarity (section 2.3.1) and, second, we analyze the
syntactic properties of instrument-naming verbs (section 2.3.2).
2.3.1 A syntactic approach to Manner/Result complementarity
First we deal with the so-called “manner/result complementarity” (see (5)) within
the syntactic model sketched out above, which lacks Rappaport Hovav and Levin’s
(1998, 2010) ontological categorization of roots and their deterministic integration
into non-syntactic event schemas.
11
See Embick (2000), however, for an argument, framed within a discussion on the syntax of deponent
verbs in Latin, that the choice of root can determine non-trivial syntactic effects. Interestingly, Embick’s
conclusion is that those syntactic effects cannot be derived from the semantic properties of the root [from
C], but from some formal feature “associated arbitrarily with certain Roots” (Embick 2000:1).
From syntax to roots: A syntactic approach to root interpretation
(5)
19
Manner/Result Complementarity: Manner and result meaning components are
in complementary distribution: a verb may lexicalize only ONE.
(Levin and Rappaport Hovav 2011)
Rappaport Hovav and Levin (2010) and Levin and Rappaport Hovav (2011) claim
that no verb encodes both manner and result: the manner in which something
comes to be in a state is unspecified for break-type verbs (e.g., break, fill, freeze,
melt, etc.), while the result is unspecified for wipe-type verbs (e.g., wipe, rub, scrub,
sweep, etc.). More generally, these authors claim that the manner/result complementarity is related to the lexicalization constraint in (6):
(6)
The Lexicalization Constraint: A root can only be associated with one position
in an event schema, as either an argument or a modifier.
Rappaport Hovav and Levin (2010) and Levin and Rappaport Hovav (2011) claim
that the root can be associated as a modifier in the event structure pattern of manner
verbs (see (7a)) or as an argument in the pattern of causative change of state
predicates (see (7b)). Given (6), it is predicted that the root in a single verb cannot
be associated to both modifier and argument positions.12
(7)
a. [ x ACT< ROOT> ]
b. [ x CAUSE [ y BECOME < ROOT> ]]
c. *[ [ x ACT< ROOT> ] CAUSE [y BECOME < ROOT>]] (* in a single verb)
According to the syntactic framework sketched out in section 2.2 above, our claim
is that the lexicalization constraint in (6) and its associated “Manner/Result complementarity” in (5) follow from how primitive elements of argument structure are
composed in the syntax (see Hale and Keyser 2002, Mateu 2002, and Acedo-Matellán
2010, i.a.). In particular, the descriptive observation in (5) can be accounted for in a
syntactic model where notions like Manner and Result become grammatically
relevant because they can be claimed to be configurationally read off the mere
syntactic argument structure: in particular, following previous syntactic treatments
of Talmy’s (2000) well-known typology of motion events (see Acedo-Matellán and
Mateu 2008 and Mateu 2008, i.a.), we argue that Manner is to be read off the
adjunction relation to v, whereas Result is to be read off the complement of the
double p-structure. Accordingly, the more general lexicalization constraint in (6) can
be shown to be derived from the syntactic fact that a single (monomorphemic) root
cannot act both as a v modifier and as a complement of a double p projection at the
same time.
12
See Rappaport Hovav and Levin (2010:26): “( . . . ) assuming that manner roots modify the predicate
ACT and result roots are arguments of BECOME, a root can modify ACT or be an argument of BECOME
in a given event schema. A root cannot modify both these predicates at once without violating the
lexicalization constraint.”
20
Víctor Acedo–Matellán and Jaume Mateu
Mateu and Acedo-Matellán’s (2012) neoconstructionist approach to the Manner/
Result complementarity contrasts with Levin and Rappaport Hovav’s lexicalist
approach in that Manner and Result are not meaning components of the root, but
interpretations derived from the position the root occupies in the syntactic structure.
From now on, we use capital letters to refer to Manner and Result in this sense. It
follows that, from our neoconstructionist perspective, expressions such as “Manner
root” or “Result root” are oxymoronic; at best, one could refer to “Manner constructions” and “Result constructions”, that is, constructions where the root is
adjoined to v and constructions where the root occupies the predicate position of
a Hoekstrian Small Clause Result, respectively. By contrast, we use “manner” and
“result”, in lowercase letters, to refer to the conceptual content of the root. In this
sense, we stick to Grimshaw’s (2005:85) claim that there are no constraints on how
complex the conceptual content of a root can be, unlike Rappaport Hovav and Levin
(2010:25), who claim that “[m]anner/result complementarity, however, involves the
root”. As we will see, a root may certainly involve manner and result simultaneously;
crucially, however, it may not be interpreted as Manner and Result simultaneously.
In our neoconstructionist approach, roots are not deterministically associated to
syntactic argument/event structures whereby lexicalist labels like “Manner verbs”
and “Result verbs” must be descriptively understood rather as “Manner constructions” and “Result constructions”. In particular, we think that Rappaport Hovav and
Levin’s (1998, 2010) claim that “result verbs” like break are less elastic (i.e., appear in
fewer syntactic argument structure constructions) than “manner verbs” like wipe
boils down to an E-linguistic fact, since counterexamples to their descriptive generalization can be found: for example, it is not the case that the root √break can only be
interpreted as Result (cf. the causative use in (8a) or the anticausative one in (8b))
since this root can also be structurally interpreted as Manner, as shown by the
examples in (8c) and (8d). Notice that (8c) and (8d) do not entail #The hammer head
broke nor #The boy broke, whereby off and into the room are not mere adjuncts but
are the Small Clause predicates. See McIntyre (2004) and Mateu (2008), i.a., for the
claim that the root is adjoined to a light verb in those cases that involve so-called
‘Manner conflation’. In (9) are the syntactic argument structures corresponding to
the examples in (8).
(8)
a.
b.
c.
d.
The
The
The
The
strong winds broke the glass.
glass broke.
hammer head broke off.
boy broke into the room.
(9)
a.
b.
c.
d.
[vP [DP The strong winds] [v0 v [pP [DP the glass] [p0 p [pP p √break]]]]]
[vP v [pP [DP The glass] [p0 p [pP p √break]]]]
[vP [v √break v] [pP [DP The hammer head] [p0 p [pP off X]]]]
[vP [v √break v] [pP [DP The boy] [p0 -to [pP in- [DP the room]]]]]
From syntax to roots: A syntactic approach to root interpretation
21
In our neoconstructionist approach, Rappaport Hovav and Levin’s (1998) observation that verbs such as break are less elastic than verbs such as wipe is to be
accounted for in terms of the compatibility between the structural semantics and the
conceptual content of the root. Thus, for example, one can agree that the root
√break is typically more used in constructions like (8a) and (8b), rather than in
constructions like (8c) and (8d). However, since usage factors like prototypicality
and frequency have to do with E-language, we consider that descriptive statements
like “the grammar of break is different from the grammar of wipe” are misleading
and should be rephrased as “the behavior of break is different from the behavior
of wipe”. See Rappaport Hovav, this volume, for a different perspective on verbs of
the break-type (encoding “scalar change”) and verbs of the wipe-type (encoding
“non-scalar change”) and the manner/result complementarity.
That said, we do acknowledge that there are some unquestionable cases of lack of
elasticity. We consider these cases to involve idioms in Borer’s (2005:25–29) sense. As
an example, consider the verb arrive, whose root is associated with a relational
element like p(ath). This explains its consistent use as an unaccusative verb: for
example, cf. Italian auxiliary selection in Gianni è arrivato “Gianni is arrived” (i.e.,
‘Gianni arrived’) vs. *Gianni ha arrivato ‘Gianni has arrived’.13
Turning back to the constraint in (6), it should not be regarded as an inescapable
stipulation on the formation of event structures (as in Rappaport Hovav and Levin’s
2010 lexical-semantic approach), but can be shown to be derived from the more
general formal fact that a root cannot be incorporated and conflated at the same time
(in a single verb). In particular, we assume that there are two ways of forming a verb:
i.e., via incorporation or via conflation (cf. Haugen’s 2009 revision of Hale and
Keyser’s 2002 distinction). We illustrate the difference with the case of the formation
of denominal verbs.14 In incorporation cases, the denominal verb (e.g., see (10a)) is
formed via copying the full matrix of the complement into the null verb. In
conflation cases, the denominal verb (e.g., see (11a)) is formed via direct adjunction
of a root to the null verb. In (10a) the root is structurally interpreted as Incremental
Theme (cf. (2a)), whereas in (11b) it is interpreted as Manner (cf. (3a)).15
13
In contrast, the following Italian example in (i), drawn from Sorace (2000:868; ex. (15c)), shows that a
root like It. √dur ‘last’ can be used in both unaccusative and unergative syntactic structures since, unlike
It. arrivare ‘arrive’ or It. venire ‘come’, It. durare does not involve an idiom in our sense: unlike arrivare or
venire, durare is not only compatible with an unaccusative use (cf. It. essere ‘be’-selection), but it can also
be allowed to be inserted in an unergative argument structure: [do √dur], whereby in this case auxiliary
avere ‘have’ is selected.
(i) Il
presidente {è/ha}
durato in carica due anni. (Italian)
the president IS/HAS lasted in post
two years
‘The president lasted in post for two years.’
14
It should be clear that we use the term denominal verb as a descriptive label: importantly, we adhere
to the view that roots do not bear a category (see Marantz 1995f., Borer 2005, among others). Accordingly,
so-called “denominal” verbs are in no way derived from a noun.
15
For further discussion on so-called Manner Conflation, see also Mateu (2002, 2008), McIntyre
(2004), Embick (2004a), Harley (2005), Zubizarreta and Oh (2007), Acedo-Matellán (2010), i.a.
22
Víctor Acedo–Matellán and Jaume Mateu
(10)
a. The boy smiled.
b. [vP [DP The boy] [v0 [v √smile] [√smile]]]
(11)
a. The boy smiled his thanks.
b. [vP [DP The boy] [v0 [v √smile v] [DP his thanks]]]
Rappaport Hovav and Levin (2010:footnote 3) point out that “for the purposes of
investigating manner/result complementarity, the specific type of predicate decomposition representation does not matter. The representations could be recast along
neo-Davidsonian lines ( . . . ) or as minimalist syntactic structures”. We disagree on
this point since the predictions of the semantic and syntactic approaches can be
quite different in an important way: for example, a brief comparison of KoontzGarboden and Beavers’s (2010) semantic approach with our syntactic one is
illustrative. As pointed out by these two semanticists, the Manner/Result complementarity in (5) cannot be said to hold as such when framed in truly semantic terms.
Koontz-Garboden and Beavers point out that conceived truth-conditionally, the
prediction is that there should be verbs encoding both manner and result, and
manner-of-death verbs can be claimed to fill in this gap. By using manner-ofdeath verbs like electrocute, drown or guillotine, Koontz-Garboden and Beavers
(2010) claim that Rappaport Hovav and Levin’s (2010) generalization with respect
to Manner/Result complementarity does not hold as such in semantic theory: the
former point out that its scope is narrower than the latter assume. However, in our
view, what Koontz-Garboden and Beavers (2010) show is not that the Manner/Result
complementarity in (5) is too strong; if anything, what they show is that (5) cannot
be formulated as such in purely semantic terms: Koontz-Garboden and Beavers
(2010:34) conclude that “we must admit the third and final logically possible class of
eventive roots, namely manner+result roots, contra RHL’s assumption that such
roots should not exist.”
As pointed out by Mateu and Acedo-Matellán (2012), Koontz-Garboden and
Beavers’s conclusion that a root can be claimed to conceptually express both manner
and result is compatible with our syntactic approach: we have nothing to say with
respect to which conceptual semantics a root element can encode.16 Our claim is that
when Manner and Result are understood in syntactic terms, there is a validity for
the descriptive constraint in (5). Consider for example the manner-of-death verb
guillotine in (12):
(12)
Joe guillotined Mary.
16
Cf. also Grimshaw’s (2005) important distinction between structural meaning and semantic content.
Following Hale and Keyser (1993f.), we assume that only the former can be syntacticized and then
constrained by well-known syntactic principles. In contrast, the complexity of conceptual content (i.e.,
Grimshaw’s 2005 semantic content) is not constrained by syntax. See also Borer (2005) for extensive
discussion on the need to sharply distinguish the meaning conveyed by grammatical structures from the
grammatically inert, conceptual content encapsulated in roots (in her terms, listemes).
From syntax to roots: A syntactic approach to root interpretation
(13)
23
a. [vP [DP Joe][v0 √guillotine [pP [DP Mary] [p0 p [pP p √guillotine]]]]]
b. [vP [DP Joe] [v0 [v √guillotine v] [DP Mary]]]
Our claim is that the syntactic argument structure corresponding to its use as a
telic causative predicate of change of state is the one in (13a), where the root is
structurally interpreted as Result: as pointed out above, Result is to be read off the
complement of the double p-structure. The root √guillotine is incorporated into
the null complex p en route to the null verb (see Hale and Keyser 2002 and Haugen
2009). The fact that its corresponding conceptual root encodes manner is not
structurally represented.
Similarly, the present neoconstructionist framework allows us to generate the
syntactic argument structure in (13b), where the root is now structurally interpreted
as Manner since it is adjoined to v. As noted above, in these cases the root is argued
to be compounded with the null verb via Conflation (see Embick 2004a, McIntyre
2004, Mateu 2008, Haugen 2009). (13b) will often be pragmatically ill-formed since
its structural interpretation would be “Joe created Mary by means of guillotining/
with a guillotine” (cf. (3a) and (11) above). However, as pointed out to us by an
anonymous reviewer, (13b) can be expected to be possible under the following
scenario: imagine a horror story in which zombies are created by chopping off
heads. Imagine a character named Joe who guillotines a woman named Elizabeth
in order to create a new creature: Mary. Given this scenario, the conflation structure
in (13b) would then be appropriate.17 Be this as it may, the important point for us
here is that (12) does obey the restriction in (5), since a root cannot be incorporated
(cf. (13a)) and conflated (cf. (13b)) at the same time. See also section 2.3.2 below, for
further discussion of other cases of instrument/manner verbs where (syntactic)
Manner conflation can be argued not to be involved.
One caveat is in order with respect to the Manner/Result complementarity in our
syntactic model. It is important to point out that such a complementarity only
emerges in cases where a monomorphemic verb, that is, a single root, is involved.
Accordingly, (5) does not hold for complex resultative constructions like John wiped
the table clean, where the verb encodes Manner and the Result component is
encoded by the adjective. Similarly, the out-prefixation construction exemplified in
(14) is a notable exception to Rappaport Hovav and Levin’s (2010) claim that
Manner/Result complementarity is manifested at the word level in English; in this
case the prefix encodes Result and the verb expresses Manner (see Marantz 2009:13):
(14)
John outswam/outdanced/outworked Mary.
17
Interestingly, as predicted by Talmy’s typology (2000), the Conflation pattern exemplified in (13b),
(3a), or (11b) can be found in English but not in Romance languages. See Mateu (2003, 2012) and AcedoMatellán (2010) for further discussion.
24
Víctor Acedo–Matellán and Jaume Mateu
Finally, as pointed out by Mateu and Acedo-Matellán (2012), another exception to
Rappaport Hovav and Levin’s (2010) abovementioned claim could be argued to be
the one discussed by Marantz (2001, 2005). According to him, the verb destroy and
related Latinate verbs (e.g., construct, instruct, restructure, obstruct, etc.) involve the
bimorphemic analysis depicted in (15): “√STROY is a manner root that incorporates
a particle, spelled out de-, that takes an ‘inner subject’ as the direct object of the
syntactically derived verb destroy” (Marantz 2001:21).18 According to him, the presence of the root √stroy, which, in (15), is intended to be structurally oriented
towards the external argument, would account for the ill-formedness of the anticausative variant of these verbs: e.g., cf. #The city destroyed / #The city constructed
/ #The boys instructed, etc. (see Alexiadou 2010 and Harley 2007 for further
discussion).
(15)
v
√STROY
city
deMarantz (2001:21)
2.3.2 Against a uniform treatment of instrument verbs
In this section we concentrate on the syntactic properties of so-called instrument
verbs and further purport to show that the ontological kind of a root does not
determine its thematic interpretation within the predicate in a grammatically relevant fashion, nor does it determine the place it has to be merged at within the
structure. Rather, the opposite holds: it is the root’s merging place that must be
claimed to determine its thematic interpretation. In particular we concentrate on
cases of denominal verbs like hammer, brush, or rake, where the incorporated root
refers to an object understood as the instrument used in the conceptual scene that
the whole predicate evokes. The discussion sets off from Harley’s (2005) proposal of
analysis for this class of verbs.
Harley (2005) proposes to derive the Aktionsart properties of a verb from a
combination of both the (un)boundedness of the root it incorporates and the
place in the syntactic configuration where that root is to be found. The proposal is
based on two tenets: on the one hand, the well-known fact that some properties of
the internal argument may determine the (a)telicity of the event (see Verkuyl 1972,
Dowty 1979, Tenny 1992, among others); on the other hand, a Halekeyserian
One caveat is in order regarding this quote. The combination of the prefix de- and the root √stroy is
not a case of incorporation in our sense: as illustrated in (10), incorporation of the root √smile into the
little v head does not involve affixation, but simply the copying of the phonological matrix of the former
into the latter.
18
From syntax to roots: A syntactic approach to root interpretation
25
l-syntactic approach to argument structure (Hale and Keyser 1993f.), where verbs
(and all predicates) decompose into (l-)syntactic structures according to their argument structure properties. Harley argues that, as overt objects may determine
VP telicity, so may roots merged as l-syntactic objects. As an example, the telicity
of unergative denominal verbs of birthing, like foal, whelp, or calve, and the atelicity of
unergative denominal verbs of bodily emission of fluids, like drool, sweat, or bleed,
depend, respectively, on the boundedness and unboundedness of the root they
incorporate (see (16), from Harley 2005:46, 47);19 this is possible because the root
of these verbs is directly merged as the complement of the light verb (see (17), from
Harley 2005:46, 48; the arrow is meant to express incorporation of the root into
the phonologically empty v head), whence its properties can determine the (a)telicity
of the event:20
(16)
a. The mare foaled {in two hours/#for two hours} (cf. bounded √foal)
b. The baby drooled {for two hours/#in two hours} (cf. unbounded √drool)
(17)
vP
v′
DP
v
√P
The mare/The baby
√FOAL/√DROOL
The analysis is successfully applied to other verbs, which, although assigned lsyntactic structures different from the one in (17), also seem to involve a homomorphism between the root and the event. However, when it comes to instrument
verbs like hammer, brush, or rake Harley observes that, in spite of the boundedness
of their roots, these verbs are not necessarily telic, as shown in (18) through (20)
(Harley 2005:60):
19
We note that birthing verbs like foal or whelp can be atelic if the predicate depicts an event in which
the mother produces several foals or whelps (cf. The bitch whelped for five minutes). This is a problem for
Harley’s (2005) analysis, since it is based on the allegedly inherent boundedness of the root. We think that
the structure of unergative predicates (see (17) and (2a)) is underspecified with respect to (a)telicity. In the
case of birthing verbs like whelp, the resulting predicate can be telic or atelic depending on whether the
root is understood as referring to a single whelp or several whelps, respectively. See footnote 24 for more
details on the non-grammatical character of the (a)telicity of these verbs. We thank an anonymous
reviewer for raising this point.
20
Actually, what is merged as the complement of v is the phrase projected by the root, √P. The
projecting ability of the roots is one of the points in which our analysis departs from that of Harley’s (see
section 2.2).
26
Víctor Acedo–Matellán and Jaume Mateu
(18)
John hammered the metal for/in five minutes.21
(19)
Sue brushed the dog for/in five minutes.
(20)
Jill raked the leaves for/in an hour.
In the face of this evidence Harley (2005:60) concludes that “[ . . . ] the source of
these denominal Roots cannot be within the argument structure of the vP, either as a
sister to v or in the Inner Subject or prepositional object positions of a Small Clause,
since elements originating in any of these positions do affect the telicity of their vPs.
Considering the thematic role of the incorporated nominal in these examples, this
makes sense: these incorporated nouns are neither Themes nor Location/Locatums,
but rather Instruments [emphasis added: VAM&JMF]. Instrumental phrases, in the
overt syntax, are adjuncts to vP, not arguments of it.”22
She then proposes that in these verbs the root is directly related to little v but in a
non-configurational way, that is, not holding a complement or specifier relationship
with it. This is shown in (21) (from Harley 2005:61):
(21)
vP
hammering
DP
Sue
v′
√P
v
√
(hit)
21
DP
the metal
An anonymous reviewer points out that a telic predicate headed by hammer is more difficult to
accept than a telic predicate headed by brush or rake, as shown in the examples of (18) through (20). We
agree with this reviewer that the difference in acceptability is due to world-knowledge reasons: while the
result of an event of brushing a dog or raking the leaves can be said to be salient or conventionalized, the
result of an event of hammering the metal cannot—although probably in the lexicon of smiths a telic
hammering of the metal is perfectly usual. See Kratzer (2000:4) for a similar remark on the telic instance of
push.
22
The Inner Subject and prepositional object positions of a Small Clause are positions in l-syntactic
configurations, whence an element may determine the aspectual interpretation of the event. Harley
proposes that the roots of deadjectival verbs of change of state (like clear or lengthen) and denominal
verbs of change of location (like saddle or butter) are merged, respectively, as predicate of a Small Clause
and prepositional object of a Small Clause-like configuration. The overt object, in both kinds of predicates,
originates as the subject of the Small Clause (it is, hence, an Inner Subject).
From syntax to roots: A syntactic approach to root interpretation
27
The same analysis, although not explicitly formulated, is proposed by Harley and
Haugen (2007:10), where it is stated that “English instrumental denominal verbs
always involve roots conflating directly with v, indicating manner [ . . . ]”. Haugen
(2009:254) also proposes, for the same verbs, that “the nominals [i.e., the roots:
VAM&JMF] are directly merged (or conflated) as adverbials directly into v.”
However, Harley’s (2005) conclusion seems to us to be too rash: the aspectual
ambiguity witnessed in the examples of (18) through (20) is not a sufficient condition
to infer that the roots in these predicates are not merged at some argumental
position. In addition, we strongly claim that, grammatically speaking, it is not
warranted that “these incorporated nouns [i.e., √hammer, for instance: VAM&JMF]
are neither Themes nor Location/Locatums, but rather Instruments” (Harley
2005:60). Hence, that observation cannot guide us in assigning them a place (an
adjunct position) in the structure. In particular, we claim that predicates such as
hammer in (18), can be analyzed as structurally ambiguous: on the one hand, the telic
hammer involves a structure hosting a preposition of terminal coincidence (in our
terms, a double p-projection). On the other, following Hale and Keyser’s
(2002:43–44) analysis of impact verbs, the atelic hammer (i.e., its normal use)
involves one hosting a preposition of central coincidence (in our terms, one
p-projection). The root √hammer, much as naming an instrument in the conceptual
scene evoked, is merged in one and the same argumental position: as a complement
of p in either case. This is represented in (22a) and (22b), respectively (arrows are,
again, the means to represent incorporation in our sense):23
(22) a. John hammered the metal (in five minutes).
vP
v′
DP
John
v
pP
DP
the metal
p′
pP
p
p
√hammer
23
We note that our analysis solves some phonological problems of Harley’s proposal in (21). In
particular, a source is provided for the phonological matrix of the verb, and, second, the problematic
empty root (the abstract HIT in (21)), copied for no (phonological) reason into v, disappears.
28
Víctor Acedo–Matellán and Jaume Mateu
b. John hammered the metal (for five minutes).
vP
DP
John
v′
v
pP
DP
the metal
p′
p
√hammer
Rough paraphrases of the structures in (22a) and (22b), respectively a transitive
event of change of state and a transitive activity, are the following ones (see Hale and
Keyser 2002:43–44):
(23)
a. John causes the metal to go INTO the state identified by √hammer
b. John provides the metal WITH properties identified by √hammer
The claim is, thus, that telic hammer-predicates are change-of-state predicates,
like break or open (in their most usual instantiations), while atelic hammer-predicates are atelic transitive predicates like push or shake (in their most usual
instantiations).24 There is evidence that telic hammer-predicates are in fact
24
In an exoskeletal framework such as the one adopted here, push can be construed as a telic changeof-state predicate (as in Sue pushed the button in seconds). See Kratzer (2000:4) or Borer (2005:128f.) for
similar remarks on this verb and its German counterpart. Importantly, there is evidence that the telicity of
these predicates is grammatically represented—contra Rappaport Hovav and Levin (1998: footnote 15),
who, following Brisson (1994), assume that telic instances of verbs like sweep are not representationally
different from their atelic counterparts. For instance, Kiparsky (1998:23, 24) shows that telic predicates
headed by the Finnish counterpart of shake require their objects to appear in accusative case—partitive
case automatically triggering an atelic reading:
(i) Ravist-i-n
mato-n.
shake-pst-1sg carpet-acc
‘I shook out the carpet.’
Thus, the telic uses of verbs like push, shake, iron, or sweep (see below) must be neatly distinguished from
other cases where telicity can be argued to depend solely on the properties of the root, and not on the
syntactic environment in which it is inserted. This is the case with telic and atelic unergative verbs like,
respectively, foal or drool—discussed by Harley (2005) and in this section—which inherit their eventive
(un)boundedness from the (un)boundedness of their roots, but which can be shown to be grammatically
indistinguishable from each other (for instance, it is interesting to point out that, in Italian, conceptually
telic figliare ‘foal’, unlike grammatically telic arrivare ‘arrive’ (and all grammatically telic intransitives),
selects the auxiliary avere ‘have’, and not essere ‘be’, in the perfect tenses).
From syntax to roots: A syntactic approach to root interpretation
29
change-of-state predicates. For instance, they, unlike atelic instances of hammer,
allow a restitutive reading of the adverb again, as shown in (24):25
(24)
a. John hammered the metal sheet in 5 minutes, but someone creased it. Sue
hammered it again, in 4 minutes.
b. #John hammered the metal sheet for 5 minutes. Sue hammered it again, for
4 minutes.
In (24a), a repetitive reading of the adverb (that is, one in which what is repeated
is the causing action) is precluded, since the subjects of each instance of hammer
have different references. The fact that a restitutive reading is accepted (being the
only one possible) suggests that the predicate encodes a final state over which again
takes scope. In the same line, hammer seems to be combinable with the prefix rewhich, according to Marantz (2005) encodes a restitutive reading and excludes a
repetitive one. This is shown in (25):26
(25)
The Damascus sword, for example, consisted of wrought-iron bars hammered
until thin, doubled back on themselves, and then rehammered to produce a
forged weld. (Encyclopaedia Britannica, “Forge Welding”: < http://www.
britannica.com/EBchecked/topic/639223/welding/7848/Forge-welding>)
The same contrast is observable in other telic change-of-state predicates (for
instance, headed by open, as in (26a)) versus atelic transitive predicates (for instance,
headed by push, as in (26b)):
(26)
a. Sue and John arrived at the strange temple. Sue opened the door, entered,
and closed the door. John waited outside and after five minutes {opened
the door again/reopened the door}.27
b. Sue pushed her car for a while and then took a break. #A gentle passer-by
stopped and {pushed it again/repushed it}.
Another proof that instrument verbs can be syntactically construed as telic is that
they admit depictive secondary predication (Rapoport 1993, Mateu 2002). As shown
in the examples of (27) and (28), telic predicates headed by instrument-naming verbs
such as brush or rake contrast with atelic predicates, which do not admit such
secondary predication (see (29)):28
(27)
25
Don’t brush the coat wet or you’ll ruin it.
See McCawley (1973), Fabricius-Hansen (1975), Dowty (1979), von Stechow (1996), among others.
But see Marantz (2009) for arguments that re- attaches to DPs.
27
The example is meant to be interpreted such that Sue and John had never opened the door until that
moment.
28
We are thankful to an anonymous reviewer for these examples.
26
30
Víctor Acedo–Matellán and Jaume Mateu
(28)
He raked the field dry.
(29)
??
He pushed the mare pregnant.
We emphasize that, given our exoskeletal assumptions, we are not making a
statement about brush, rake or push as verbs. Therefore, we are not claiming that
brush and rake admit secondary predication while push does not. Rather, the claim is
that telic predicates of change of state, such as (27) and (28), may host such
predication, while atelic predicates such as (29), may not.
More evidence comes from verbs naming the instrument with which someone is
killed, like guillotine or knife, which can easily head change-of-state predicates (see
Koontz-Garboden and Beavers 2010, for guillotine), as shown in (30) and (31):
(30)
It took me five minutes to guillotine Jim (with one slice). (Koontz-Garboden
and Beavers 2010: footnote 13)
(31)
[ . . . ] a leading Libyan dissident who was found knifed in his grocery shop in
west London. (found in Collins Wordbanks Online English corpus)
Similarly, verbs of household activities naming the instrument with which the
activity is carried out can be construed as (telic) changes of state, as illustrated below
by Catalan escombrar ‘sweep’, from escombra ‘broom’, and planxar ‘iron’, from
planxa ‘iron’:
(32)
He
escombrat la
cuina
en
have.1sg sweep.ptcp the kitchen in
‘I swept the kitchen out in five minutes.’
cinc
five
(33)
He
planxat
les camises en quinze
have.1sg iron.ptcp the shirts
in fifteen
‘I ironed the shirts in fifteen minutes.’
minuts.
minutes
minuts.
minutes
There is evidence, therefore, that the difference between the telic and the atelic
instances of instrument verbs is grammatical.29 Our claim is that the root is free to
be merged as the complement of an abstract preposition (see (22)): that is, an
argumental position, contra Harley (2005). In particular, when it is embedded within
a single p projection (encoding central coincidence; see section 2.2) it is interpreted
as a Central Ground; when it is embedded within a double p projection (encoding
terminal coincidence), it is interpreted as a Terminal Ground (i.e., Result). Crucially,
in neither case is it interpreted as an instrument (i.e., Manner), in spite of the fact
that it may represent an instrument in the conceptual scene. This is of course not to
say that these roots are not usually found in configurations akin to the one proposed
by Harley in (21), that is, merged in a non-argumental position. Following McIntyre
29
See footnote 24.
From syntax to roots: A syntactic approach to root interpretation
31
(2004) and Embick (2004a), among others, we claim that in a resultative construction such as the one in (34), the root √hammer is merged directly as an adjunct to v
via conflation, while another root occupies the place encoding resultant state. This is
represented in (34b):
(34)
a. John hammered the metal flat.
b.
vP
DP
John
v′
v
√hammer
pP
v
p′
DP
the metal
p
pP
p
√flat
In the structure of (34b) the root √hammer is adjoined to the eventive head v, and
is therefore identified with the grammatically-encoded event. In that sense, √hammer represents a Manner of the event.30
In conclusion, it has been argued that one and the same root may be inserted in
two different positions in the syntactic structure (not simultaneously): either in an
argumental position, e.g., as complement to a p head—which can either be taken as
complement by another p head or form a single p projection; see (22a) and (22b),
respectively—or in a non-argumental position, as an adjunct to the v head (see
(34b)). The root receives a different interpretation depending on the position where
it is freely merged: Terminal Ground in (22a), Central Ground in (22b) or Manner in
(34b). The fact that the root √hammer refers to the “instrument” with which the
action is carried out is orthogonal to its structurally imposed, grammatically relevant
interpretation.
2.4 Concluding remarks
What on an intuitive level seem to be intrinsic features of the root, such as Result,
Manner, etc., are in fact properties of the structure: e.g., Result is the interpretation
30
Of course, if √hammer is merged as complement to v it is understood as Incremental Theme: He has
been hammering all morning (i.e., “he has been doing hammering all morning”).
32
Víctor Acedo–Matellán and Jaume Mateu
of a root merged as the complement of a double-p projection and Manner is the
interpretation of a root adjoined to v. As a consequence, the ontological classification
of roots does not condition the linguistic derivation, as is assumed in the endoskeletal approach (Rappaport Hovav and Levin 1998, 2010, Rappaport Hovav, this
volume) and some work done within Distributed Morphology (Alexiadou, this
volume, Anagnostopoulou and Samioti, this volume, Doron, this volume, Levinson,
this volume, Roßdeutscher, this volume). On the contrary, it is the structural
position occupied by the root in the syntactic event/argument structure that
determines its interpretation, a result compatible with the exoskeletal approach
(Borer 2003, 2005; see also Acquaviva, this volume).
3
The roots of nominality,
the nominality of roots
PAOLO ACQUAVIVA
3.1 Introduction
A natural goal of a theory of grammar is to explain what speakers know when they
know lexical items. I will follow here recent syntactic approaches to lexical decomposition, in particular those that posit category-free roots in the make-up of lexical
categories (Marantz 2001, Arad 2003, 2005, Borer 2003, 2005) and pose the question
of how grammar expresses what we call nouns. To date, most work on lexical
decomposition has focused on verbs, with important extensions on deverbal nominalizations; by contrast, my aim here is to investigate nouns as a primary lexical
category—the roots of nominality.1
Baker (2003) addressed this same question, following syntactic assumptions that
do not involve discrete category-free roots. Baker’s work links up to a philosophical
tradition stemming from Frege (1884) and Geach (1962), which views the essence of
nominality in the ability to stand for S in a sentence of the form x is the same S as y.
There are two reasons for reconsidering the question. First, Baker’s approach views
nouns as sortal terms, that is, terms expressing a standard of sameness; but not all
nouns straightforwardly admit such an interpretation, and the very idea that nouns
lexically define a principle of identity (a way of being the same) as distinct from a
principle of application (what they are true of ) is not as obvious as it may appear
(see Barker 2010 for a critique). Second, defining nouns as sortal terms tends to blur
the distinction between a ‘nominal’ and a ‘noun’. In order to contribute to a theory
of lexical competence, we should focus on what makes nouns the kind of words they
are. This leads to the second aim: distinguishing various aspects of nominal inter1
The research reported here was supported by a fellowship funded by the Alexander von Humboldt
Stiftung, which I gratefully acknowledge. I would also like to thank Josef Bayer, Hagit Borer, Phoevos
Panagiotidis, and Carl Vogel for very useful discussion, as well as the organizers of the Stuttgart workshop
and the editors of this volume. Faults and omissions are my own responsibility.
34
Paolo Acquaviva
pretation and home in on nominality as an irreducible lexical property—the nominality of roots.
The argument has three stages. In the first part (sections 3.2–3.4), I will distinguish
individuation as defining a discourse referent at DP-level from individuation as
defining an abstract (kind-level) category of entities, and I will claim that lexical
nouns name such abstract categories; the granularity and the part-structure of the
denotation domain are specified by grammatical morphemes between the outer DPlevel and the innermost N- and root-level. The second part (3.5–3.8) justifies this
approach on empirical grounds, showing its descriptive and explanatory advantages
leading to falsifiable predictions on what can be and what cannot be a common noun
in a natural language. The last part (3.9–3.10), in contrast with widely held assumptions (in this volume, see especially the contributions by Alexiadou, Levinson,
Rappaport Hovav, Roßdeutscher) argues that roots do not encapsulate any aspect
of lexical semantic content, not even encyclopedic or lexical-categorial information
(contrast Gallego, this volume, and Borer 2005); in particular, the observations in
sections 3.9 and 3.10 suggest that locating nominality on roots is oversimplistic and
ultimately wrong. In so far as they differ from nouns, roots should not be stipulated
to have the semantic function of nouns; instead, their function is to differentially
label the syntactic construction that corresponds to a noun, and which interfaces the
Conceptual/Intentional cognitive system as the name of a category concept.
3.2 Two types of individuation
Both tree and tall are true of entities; but only the noun tree defines a type of entity.
As a necessary preliminary, let us distinguish individuation in this intuitive sense
from the individuation of an entity as a discourse referent. Discourse referents are
what we talk about: individuals in the domain of discourse, set up, tracked and
variously qualified by appropriate determiners and quantifiers, through deixis,
anaphora, and quantification (“StrongDP” in Zamparelli 2000, and < e>d in Borer
2005). The identification of discourse referents is ultimately anchored in the speakers’ spatiotemporal frame (Strawson 1959), and in this frame it is possible to make
identifying reference to them without the medium of a categorizing description (he,
that, the second from the left, another one). For this reason, anything can be a
discourse referent: the individuals referred to by pronouns or corresponding to the
assignment of value to variables, but also, without overt determiners, nominalized
properties which are subjects of individual-level predicates, as in big is beautiful, and
individuals denoted by names.
Nouns, more precisely common nouns, contribute to the individuation of discourse
referents indirectly, by a categorizing description: that tree, every book. To capture the
difference between tree and tall, we can follow a well-established tradition and call
“kinds” the categories of entities expressed by nouns (Carlson 1977, Krifka 1995,
Chierchia 1998, Zamparelli 2000), with the assumptions spelled out in (1):
The roots of nominality, the nominality of roots
(1)
35
• kinds are primary entities, not abstracted from individuals;
• nouns are names for kinds, and denote them rigidly;
• nouns are primarily entity-denoting names, not predicates.
The first assumption is taken from Krifka (1995) and Mueller-Reichau (2006). The
second generalizes to all kind-denoting nouns (not all nouns, as we will see) the
doctrine of direct reference of Putnam (1975) and Kripke (1972). The third adopts
Baker’s (2003) and Mueller-Reichau’s (2006) characterization of nouns (again, not
all) as kind-referring terms, but as names, and thus rigidly, and not as predicates. In
this respect, my suggestion is the mirror image of the “Nominal Description Theory”
of names, where names express “no substantive property but merely the property of
bearing that very name” (Bach 2002); or in the words of Elbourne (2008:197): “on
most occasions of use Alfred will mean ‘entity called Alfred and identical to a,’ where
a is an individual constant picking out an entity called Alfred.” Instead of starting
with nouns as predicates and characterizing names by the property of “being
named” and by a stipulation of identity with an unidentified referent, the alternative
I explore here starts with names and characterizes nouns as names for kinds; their
predicative use is derivative. This, I argue, is the ultimate source of the distinction
between the identifying function of nouns and the characterizing function of
adjectives and other predicates (cf. Barker and Dowty 1993).2
The hypothesis which emerges from the assumptions in (1) concerns the fundamental, irreducible nature of nouns as a lexical category. It does not concern “bare
nouns” in the sense of nouns embedded in a structure without determiners or
modifiers, but the constitutive properties of nouns as a lexical category. Clearly the
two issues are related, but the question “what is a noun” is prior to the question
“what is a bare noun”, that is, a noun in a particular syntactic context.
While the hypothesis presupposes that nouns are a linguistic category, it does not
presuppose that they are a primitive morphological or syntactic category. In fact, the
syntactic approach followed here (largely that of Borer 2005) decomposes lexical
categories into constructions made up of smaller syntactic pieces, at whose core lies a
category-free root. This leaves open the possibility that some fundamental properties
could make a construction a noun rather than an adjective or a verb, over and above
the choice of certain grammatical formatives. In what follows, I will pursue this line
and distinguish nominality as a grammar-internal characterization from a more
substantive characterization, in which what is morphosyntactically nominal also
encapsulates a certain cognitive value which no other lexical categories have, namely,
functioning as a name for a certain entity. Semantically, this is an < e>-typed kind;
cognitively, the kind-level entity named by (the construction interpreted as) a
2
Krifka (1995) calls “kinds” conventional kinds (e.g., gentleman) and “concepts” the larger class of
kind-level objects which include both conventional and non-conventional kinds (e.g., gentleman in a blue
tie). In discussing kind-level objects named by simplex nouns, I use “kind” only in the first sense, and
“concepts” for their psychological counterpart.
36
Paolo Acquaviva
common noun corresponds to a concept. Of course, it is not only nouns that have a
conceptual content; expressions corresponding to other lexical categories, and syntactic constructions of any complexity, nominal or otherwise, likewise have a
semantic value defining conceptual content. What singles out nouns, I hypothesize,
is naming one of the entity-types that make up a speaker’s conceptual ontology.
Nouns, then, individuate a category in a speaker’s ontology by naming it; this is
sharply different from the individuation of a discourse referent (be it object-level or
kind-level), which is accomplished by larger syntactic expressions and involves the
DP-level.
3.3 Concepts
That linguistic expressions are associated with some conceptual content is uncontroversial; the view that kinds are entities distinct from their instances, on the other
hand, and that they are mentally represented entities not reducible to abstractions
from instances, are significant theoretical choices. Through these choices, we can
single out the conceptual content of common nouns as the one which corresponds to
a concept identifying a category of entities. This establishes a connection with the
tradition of studies on category concepts in cognitive psychology and in cognitive
semantic approaches (cf. Rosch 1978, Jackendoff, Bloom and Wynn 1999, Lakoff
1987, Wierzbicka 1988). In particular, it is no accident that research on concept
representation and acquisition typically centers on sortals, variously defined as terms
or as concepts that provide standards of sameness and enable re-identification of
objects (Carey and Xu 1999, Xu 2007), going back to an old and important line of
research in metaphysics and in the philosophy of language (Frege 1884, Geach 1962,
Gupta 1980, Wiggins 2001). A key aspect of such conceptual categories is that their
definition depends less on the actual properties of the entities falling under them,
than on “canonical” or “typical” properties. This means that our understanding of a
natural kind like DOG or WATER rests on what is expected to be typical properties
of dogs and water, more than on actual biological or chemical properties; likewise,
artefacts like CHAIR or roles like PRESIDENT are defined, as concepts, more by the
functions associated with them than by properties of individual chairs or presidents.
Work in psychology (Bloom 1997, Gelman 2004) and in formal ontology (Simons
1987:111–127, Meirav 2003) significantly converge on this conclusion, the latter
emphasizing that reference to the constituent parts of actual object-level referents
cannot by itself account for our intuitions about what it means to be one and the
same entity; mode of constitution, function, purpose or other non-extensional
notions must be called upon.
Taking seriously this relative independence between an abstract concept and
its instances, I will envisage the primary denotation domain of common nouns as
a taxonomy of kinds, where each kind is an unanalyzable atom, including kinds
The roots of nominality, the nominality of roots
37
like WATER.3 This formal construct models the conceptual ontology of a speaker. It
is not clear to me whether all kinds in such a conceptual ontology should be part of a
taxonomy: the answer hinges on canonical names like Alfred, which may or may not
be related to the kind of entities they name, in this case the kind corresponding to
PERSON.4 Note that the subkind-of relation crucially differs from the part-of
relation. In fact, if wholeness, extension, overlap, and connection are all defined in
terms of the part-of relation, the atomicity of concept kinds derives from the single
assumption that no part-of relation is defined over their domain. At this basic level,
each concept refers to an atomic individual whose identity cannot depend on its
parts or instances, because no such parts are defined, and concepts are not defined
on the basis of their instances (contrast Chierchia 1998). Such kind-level individuals
are uniquely identified across all possible worlds.
As noted, the idea that kinds are primary and their instances are derived is not
new. Concepts are prior to object-level (particular) entities, not because they magically create objects, but because what counts as an object for a speaker depends on
that speaker’s conceptualization. Equating the kind-level entity referred to by a word
(say, the kind BEAR named by the noun bear) with the corresponding concept
(BEAR as an atomic member in the ontology represented in a speaker’s mind)
captures the intuition that the token-entities speakers can refer to are not simply
given, but depend on the type-entities into which speakers categorize their mental
representation of the world. This allows, I argue, for a predictive and empirically
falsifiable theory of nouns as linguistic terms.
One practical difference with related approaches is that canonical names, like
common nouns, identify kinds (rigidly), and only derivatively refer to particulars.
This might seem to obliterate an important distinction, namely that kinds have
descriptive content while names do not (Zamparelli 2000:174, following Carlson
1977). But on closer examination this distinction proves anything but clear-cut, in
both directions: canonical names, especially in different cultures, often have a
transparent descriptive content (and cf. Giorgione was so called because of his size,
which requires interpreting Giorgione as a name and -one as an evaluative suffix at
the same time); on the other hand, there are common nouns that identify a category
but express no descriptive predicate, for example acronyms like an FTE (“full-time
3
Boeckx (2010: 28–29) more generally views all the underived building blocks of syntax (corresponding, in his analysis, to lexical roots) as “conceptual addresses”. As will become clear, my view is instead that
concepts are named by categorized syntactic structures, not by roots.
4
Since I embrace the doctrine which treats all kind-referring nouns as proper names, I distinguish
those which have the syntax of names, like Alfred, by calling them “canonical names”. It seems that
canonical names, which effectively are kind-naming terms used as a discourse referent (as in Bear went to
visit Fox) should make up an unordered list, so that Fido is not related to dog by a taxonomic relation. We
can name entities whose ontological status is unclear to us (is Titania a woman? if so, is she a human
being?); and we routinely name scientific categories like Mus musculus that do not fit into the taxonomy of
common nouns (because categories of “folk biology”, like mouse or fish, often do not fit the scientific
taxonomy and so cannot be related to the kinds they name; see Dupré 1999).
38
Paolo Acquaviva
equivalent”), an administrative concept corresponding to one employee or student
as a unit for resource allocations. In fact, just about anything can take up the
function of a common noun, once it is plugged into the appropriate syntax: an
Einstein, no ifs and buts, German die Ichs ‘the egos’. But of course not every word
that can be so treated is a noun as a lexical entry. To clarify, let us distinguish three
aspects of nominality:
(2)
Three aspects of nominality
• the expression of an argument
• the expression of a discourse referent
• the expression of conceptual content
Every property can be nominalized in so far as it is treated as a non-predicable
individual: red-ness, looking tall, to be altruistic. Being nominal in this sense (that is,
in the sense of “argumental” as opposed to “predicative”) must be distinguished
from being an entity that reference is made to in the universe of discourse. In line
with most approaches, I will relate this aspect of nominality to the interpretive
function of determiners and quantifiers at the DP-level. What specifically concerns
nominality as a lexical property is the third aspect, on which I will focus. To
anticipate, naming a category of being in a speaker’s conceptual ontology is the
characteristic property of nouns, but not a definitional one. Nominality consists of a
cluster of properties; as a property of words, it is anchored in, but does not reduce to,
the naming of a category concept, represented as a kind-level entity in the speaker’s
ontology.
3.4 Three stages of nominal interpretation
A short comparison with previous analyses will illustrate how kind-naming nouns fit
into the syntax and semantics of DP structure. Aside from denying Longobardi’s
(1994) claim that only DP can express an argument (cf. Chierchia 1998, Baker 2003),
the proposal I am outlining integrates the approaches to DP structure of Zamparelli
(2000) and Borer (2005, this volume). Both associate D and DP-peripheral strong
quantifiers with the expression of a variable, which is assigned value in one of several
ways (by quantification, anaphora, ostension); on the other hand, DP-internal
projections express a description restricting the value of the variable. Zamparelli,
in particular, distinguishes three layers: an inner one expressing a kind, an intermediate one turning a kind into a predicate true of kind instances, and an outer one
expressing an individual variable. Borer (2005) likewise locates the expression of a
variable on D, and views the intermediate projections as specifying the unit granularity (“Division”) and the size (“Quantity”) of the reference domain over which the
variable will take value.
The roots of nominality, the nominality of roots
39
I follow this three-way division of DP, with the important modification that both
nouns and (canonical) names start off as expressions of type < e>, denoting categorizing atomic concepts. The second region adds information about part structure, and
turns the kind-names that will end up as common nouns into predicative expressions true of entities, or of relations between entities, as is plausibly the case for
relational nouns like sister, or between entities and events. (Canonical names like
Alfred or Mr. Bear are those that do not become predicates at this stage; correspondingly any adjective modifier does not restrict their reference.) The count/mass
difference arises in this second region, where the denotation domain is not a
taxonomy but a mereology, a semilattice organized by the part-of relation. It is the
grammatical context that turns water into an expression true of portions of matter,
and tree into an expression true of re-identifiable individuals with a stable granularity. Attributing this function to the grammatical context, rather than to intrinsic
and immutable properties of a “lexical entry”, has several advantages, as Borer (2005)
shows. We can understand why nouns that seem to be intrinsically mass or count
can often (not always) be coerced into an opposite interpretation, as in the “type/
portion” reading of three waters or the space-extended mass reading of there was dog
all over the road; we can generalize to all languages the structure necessary for
classifier languages, where the type of granularity is expressed by dedicated morphemes; we can associate the same function to determiners like the English indefinite a, and to a host of morphemes with packaging function, like the Classical Arabic
-at- in damʕ-at-un ‘a tear’ from the non-count plural damʕ-un ‘tears’. This does not
mean that nouns in isolation always neutralize the mass/count distinction, or that
the mass syntax of water is accidental. On the contrary, the conceptual content is
often precise enough to dictate the division expressed by grammar for normal uses of
a noun. But the two are not the same thing. The concept underlying what we
understand as a book will necessarily determine a way of being a whole, so that
knowing the meaning of book entails in all languages the ability to distinguish what
is a whole book from what is two or less than one. However, this concept can in
principle fit a count as well as a mass context (subject to language-particular
constraints): the coerced dog all over the road denotes matter, but to understand
that, we must know that this matter comes from individual bounded dogs.
The overview in (3) is based on Borer’s analysis of DP, in which functional head
positions correspond to open values with a category label, and elements occupying
them are ways to assign a range to the open value. In Borer’s model, the complement
of Div is the position hosting unanalyzable root listemes at the core of the whole DP.
Following De Belder (2008) and Alexiadou (2011), I will assume instead an intermediate projection between Div and the root, to encode information like gender and
noun class, which when present are noun-inherent but not root-inherent. The
function of providing a root-external position for lexeme-defining derivational
morphology is the same as that ascribed to the nominalizer [n] in Kihm (2005)
40
Paolo Acquaviva
and Acquaviva (2008a, b), in line with the Distributed-Morphological assumption
that roots are categorized by heads like [n] or [v] (see Embick and Marantz 2008). In
Borer’s system, however, since the whole structure determines lexical categorization,
there is no need for specially posited categorizing heads. It seems better, then, to view
this inner head as a nominalizer that ascribes the resulting nominal structure to a
paradigmatic class, without necessarily selecting a value for Division, that is, without
fixing the noun’s countability or its number value (cf. the head called “Class” in Picallo
2005 or “Sort” in Déchaine et al. 2012). De Belder (2008) shows that this determinant
of the noun’s lexical semantics qualifies its reference as having a part structure induced
by extension, typically in space (which explains her choice of the label “Size”; cf. De
Belder, Faust, and Lampitelli, this volume, for a refinement of this analysis, with two
loci for evaluative affixes, above and below the nominalizing head). Evaluative affixes
with a derivational function, like the Dutch diminutive -je in vilt-je ‘felt piece’, or the
Russian diminutive -ok in tsvet-ók ‘flower’, seem likely candidates as realizations of
this inner Division position, since they fix a gender value and at the same time
establish a count interpretation which cannot be coerced. Importantly, nouns thus
formed can be pluralized, which shows that this inner-level determination can
combine with an outer-level division, where pluralization is expressed.
In sum, the assignment of an index for discourse-level reference, the division of
the reference domain, and the conceptual categorization correspond to distinct
regions in the DP structure:
(3)
DP
discourse-level reference
Det
division
QuantityP
Quantity
DivisionP
Div
categorization
SortP
Sort
root
Division extends over two projections, the inner one expressing “lexical” properties
which, while grammatical in nature, depend on the identity of the lexeme defined by
the overall structure. For example, the Italian masculine noun frutto expresses a
‘piece of fruit’, but the corradical frutta, feminine, expresses ‘fruit’ as a non-denumerable mass term, with a corresponding shift in conceptualization (for instance,
there are appearance or dimensional properties which apply to a piece of fruit and
The roots of nominality, the nominality of roots
41
not to fruit in general, and conversely). As this example shows, it is not the root
alone which identifies a concept, but a structure comprising at least the inner “Sort”
head. In fact, I will argue that lexical roots do not by themselves identify a concept;
even when no other element appears, as in the English cat, what expresses a noun’s
lexical semantics is always a syntactic construction, never a root alone (although the
root in this structure might suffice to identify a concept). Below we will also consider
cases where regular pluralization triggers a shift on conceptualization, as in brain
‘brain organ’ brains ‘organs’ or ‘brain matter, intelligence’. When this happens, it
seems plausible that the plural marking on Div marks a valuation of both Div and
Size; in other words, the lexicalized reading of the regular inflectional plural can be
represented as involving an appropriate marking on the innermost functional head.
Turning now to the three functions associated with these syntactic regions, note
that the distinction between discourse referents and concepts is not one between two
different types of entities, but between different ways to individuate; that is, to make
identifying reference to something. As noted, anything can be a discourse referent,
including a kind:
(4)
The Dodoi is extinct. Iti used to live in Madagascar.
It would be misleading, therefore, to correlate directly the transition from the N-level
to the D-level with the map from kinds to instances. Such a map is certainly needed,
but not as the meaning of number, division, or of determiners in general. The
observed cases of discourse referents not based on lexicalized nouns (your oh’s, the
why’s, das Ich) constitute further evidence that discourse referents are not necessarily
derived by instantiation from a kind with descriptive content. More generally, there
is actually no need to associate a concept, in the sense clarified earlier, with all lexical
nouns. Consider a nominalization like destruction:
(5)
a. Kim’s destruction of the vase
b. [NP -tion [EP Kim [AspQ the vase [L-D destroy ]]]]
(L-D = lexical domain; EP = event; AspQ = Aspect/quantity)
(Borer 2003)
(5b) shows the syntactic structure assigned by Borer (2003) to the transparent
deverbal reading (cf. Borer, this volume, especially her (5), and references cited
there). All we know about the lexical semantics of this noun corresponds to the
lexical semantics of the verb, plus a nominalizing affix -tion which caps the structure
and turns it into an argument. In so far as the interpretation of destruction as an
event noun is perfectly transparent and preserves all aspectual and argumentstructural information of the verb to destroy, there is no reason to assume a concept
DESTRUCTION in addition to the lexical content of the verb. More strongly, if we
associated a concept with anything that can take the syntactic role of a noun, we
should posit concepts for all properties, because any property can be nominalized
and used as an argument in an atomic predication, as in red is beautiful or being tall
42
Paolo Acquaviva
is a property; not only that, but as we have noted there is simply no limit to what can
play the role of a noun in a syntactic structure, as in your oh’s, or non-linguistic
expressions like whatever can take the place of *** in an that kind of ***. Such a
Platonistic regress, which would treat as a concept just anything that can fill the
syntactic place of a noun, would deprive the notion of concept of all psychological
substance, and ultimately of all theoretical usefulness.
In sum, there is no mutual implication between discourse referents, concepts, and
common nouns. Not all DPs contain a lexical noun, and not all nouns name
concepts; in particular, not transparent nominalizations. In their transparent deverbal reading, nouns like destruction do not name a kind, and so do not identify a
category concept. They are morphosyntactically nouns (because of the grammarinternal properties of the morpheme realized as -tion), but their conceptual content
consists of that of the corresponding verbs, nominalized in the first of the three
senses distinguished in (2) above.5 In other words, when destruction encapsulates a
VP, its conceptual content includes information about a type of event, event
structure, and argument structure, all of which is different from what I have called
a category concept; the notion of ‘destruction’ is derivative from that of ‘to destroy’,
and does not identify a type of entity. By contrast, not only dog, but also forms
ending in -tion like audition or circumlocution name entity types, and so can be
taken to name category concepts. This highlights the fact that, in the current
analysis, there is no fixed size for the map between a piece of syntactic structure
and a concept: a category concept may be identified by a root alone, like cat (when
qualified as a noun by its insertion in a nominal syntactic environment; see below);
or by a complex consisting of root-like elements, like pro-fession; or by a morphological complex involving a Sort head, like the Russian molot-ók ‘hammer’ or, it is
claimed, brains as ‘intelligence’, where pluralization realizes not only Div but Sort as
well (this marks an important difference with approaches that associate either
argumental information or just bare “encyclopedic” content systematically to the
root; cf. the opening remarks in Gallego, this volume). Correspondingly, there are
nouns that involve a lexical root but do not name a concept, like destruction in its
transparently denominal reading but also, as I will claim below, nouns like contents.
The underlying intuition is that concepts are not “the” meaning of a fixed-sized
chunk of syntactic structure, but grammar-external elements with cognitive reality,
related to grammar-constructed formal representations, in the case of lexical nouns, by
a naming relation (see Acquaviva and Panagiotidis, 2012, for arguments that the
content encapsulated in a syntactic structure does not make up, but rather constrains,
5
A reviewer queries the status of the morpheme -tion in (5b) in relation to the structure in (3). The
countability and number properties of destruction when used in this sense derive from its nature as
“Argument Structure Noun”, which in this framework go back to the presence of a full-fledged VP. By
itself, -tion simply turns the VP into something that is morphologically a noun and can be inserted as such
in a DP; for concreteness, I identify it with Sort (Alexiadou 2011 simply calls it [n], in a discussion of Greek
and English deverbal nominalizations in -m and -ing).
The roots of nominality, the nominality of roots
43
the meaning of a lexical concept). Since the semantic content of functional heads is
not generally compatible with this naming function, the portion of DP structure
which can name a concept is limited to the root and its immediate environment.
3.5 Morphology and part structure
Viewing the semantics of nouns as based primarily on category concepts, with
syntactic structure mediating the expression of discourse referents (kind- or
object-level), has a number of advantages. The most immediate one is a better
understanding of how morphology affects noun lexical semantics.
The previous section distinguished categorization and division, and related them
to specific regions in DP. By reference to these two basic functions, we can capture
the semantic relatedness of many morphological phenomena; in addition, we can
predict that morphemes that realize these functions must also spell out the innermost heads in DP, here corresponding to Sort and Div: gender, number, classifiers,
evaluatives, but not case or definiteness, for example.
Division, which specifies the part structure of the reference domain, is the function of classifier-like affixes that in many languages produce nouns that are necessarily count, usually with a concrete reading:
(6)
a. ħṭeb
‘fire wood’
ħTeb-a
‘piece of fire wood’
Moroccan Arabic (Harrell 1962:80)
b. glao
‘rain’
glav-enn
‘raindrop’
Breton (Trépos 1968:67)
As is well known, evaluative affixation (typically diminutives) often has the same
function:
(7)
a. molót’
‘to grind’
molotók
‘hammer’
Russian
b. pane
Brot
‘bread’
panino
Brötchen
‘bread roll’
Italian
German
c. meabhair
‘mind’
meabhróg
‘thoughtful girl’
Irish
Derivation from number words illustrates the function of division with particular
clarity:
(8)
a. sieben
‘seven’
der Sieben-er
German
‘the number seven’ (bus)
b. des’jat
‘ten’
des’jat-ka
Russian
‘number ten’ (bus, playing card)
44
Paolo Acquaviva
Whatever number words like ‘seven’ are true of (see Wiese 2003 for an innovative
approach), they do not denote entities extended in time or space. Suffixation with
diminutive-singulative, or with the “individuating” -er of Germanic, derives count
nouns whose reference ranges over sets of concrete unit-objects, identified but not
described by a numeral.
Lieber (2004:37) qualifies the output of suffixation by -ee, -er, -ist, -ent as concrete
and related to a function or activity. My suggestion is that these and similar
“individualizing” affixes primarily express individual entities with a stable granularity derived from the base. The nouns they derive are systematically count nouns,
although nothing makes this necessary in Lieber’s account; and they are not necessarily extended in space as the label “material” suggests, witness eventive readings
like all-nighter. The eventive reading is prominent in the Italian nouns derived with
the suffix -ATA (a shorthand for -ata and other past participle endings), alongside
other packaging readings like “quantity contained in N”; cf. Acquaviva (2005):
(9)
a. nuotare
‘to swim’
nuotata
‘a swim’
b. stupido
‘stupid’
stupidata
‘a stupid act’
c. cucchiaio
‘a spoon’
cucchiaiata
‘a spoonful’
Italian
Although a concrete reading typically accompanies this packaging function, it is
by no means necessary. De Belder (2008) notices that a choice between two Dutch
suffixes determines whether the count reading imposed on the mass base ‘felt’ refers
to concrete objects or to subkinds:
(10)
vilt
‘felt’
sg, mass
no [DIV]
vilt-en
‘felt types’
pl, count
[DIV]
vilt-je(-s) ‘felt piece(s)’ sg/pl, count [DIV]+[Size]
atomless
kind-based granularity
object-based granularity
Once the function of division is clear, we can detect it in many more cases than
would appear at first sight. Bittner (2002), for instance, observes that all masculine
derivational suffixes in modern German derive count nouns. We can also venture
wide-ranging predictions: a nominal affix with a clear singulative value, like the
Breton -enn or the Slavonic diminutives, should never create nouns interpreted as
nominalized properties (like redness or love). The reason is that suffixes with
dividing function form nouns with a complex reference domain, articulated into
identifiable parts (single events, unit objects, fragments, and complexes of them);
and a nominalized property does not refer to a domain partitioned into units.
We also expect division-affecting morphology to determine count readings more
frequently than mass. For Borer (2005), a count interpretation corresponds to a
partition based on fixed units, while mass arises where there is no division, or it fails
The roots of nominality, the nominality of roots
45
to define stable units. It is no accident, then, that dividing morphemes are mostly,
perhaps overwhelmingly, count nouns. Lieber (2004:40), who notices this fact, cites
the English “compounding stem” -ware as one of the few counterexamples (footware, glass-ware). Revealingly, the only general “massifier” involves pluralization,
which Borer shows to be a pure divider: cf. brainbrains, fundfunds, in the noncount reading ‘intelligence/brain matter’ and ‘funding’ (cf. Acquaviva 2008b for
extensive discussion). But the dividing function of plural is needed anyway, even
for count nouns, in so far as plurality is an inflectional category (in English and
similar languages). So, even in those cases, the mass reading lacks a dedicated mass
morpheme, according to expectations.
Since plural is a divider, and division typically (not always) affects a reference
domain extended in space or time, the choice of number feature can distinguish
between a reading as abstract kind or as a space- or time-extended entity. A littlestudied example is provided by mass plurals like waters (cf. Ware 1979, Tsoulas 2006,
Acquaviva 2008b, Alexiadou 2011).
(11)
water
kind-level reading or indefinite description for quantities of water
the water applicable to contextually given quantities
the waters large quantities only (*some waters, *the formula of waters)
The granularity of the reference domain can be a central component in the content
of a concept. Small wonder, then, that division-expressing morphology can also shift
the conceptual categorization itself. This is often accomplished by choice of gender,
word class, or by irregular inflectional morphemes, which in the present account
involve a marking on the head Sort, like in penny pennies (coins) pence (value)
(cf. (20–22) below).
Plural can also select an instance reading without a categorizing concept. This is
the analysis I propose, following Wierzbicka (1988), for nouns like contents or riches
(see Acquaviva 2008b:104–105, 116–118 for discussion).
(12)
a. contents, beginnings, belongings, furnishings
X-[pl]: undescribed particulars named by a contingent property X
b. depths, deeps, heights, riches
X-[pl]: concrete extensions of the property x, where X is not the kind they
instantiate
Nouns like table or even furniture name a category of beings, and entities that are
tables or furniture instantiate this kind. Entities that are contents, beginnings, or
riches, by contrast, are undescribed and uncategorized. Being contained, taking place
at the start of a process, or making up wealth, are contingent facts about entities and
do not categorize them, not even on the basis of their purpose and function (as is
arguably the case of furniture). Two falsifiable predictions arise: (i) such nouns
should never allow kind-level predication, as in # depths/contents/heights are
46
Paolo Acquaviva
widespread; and (ii) such nouns should not allow semantic incorporation; we can
have ‘to poetry-read’ (Farkas and de Swart 2003), but not something like ‘to contents-carry’, in any language.6
3.6 Concept-naming noun stems in word formation
The most straightforward empirical justification for basing the semantics of nouns
on concepts comes from the value of noun stems in word formation. As is well
known, nouns that are part of a lexical predicate systematically display an interpretation typically characterized as “non-referential”. The most usual illustration
comes from compounds, where a non-head noun cannot refer to specific entities: no
individual dog in dog food, no particular taxi in taxi driver. An object can be a head
rest before any head has rested on it, and, as Borer (2005) points out, stone requires a
count reading in stone throwing but a mass one in stone carving. This is what
characterizes kind-level reference, as in the object-independent value of Coke in
Coke bottle, which identifies a bottle type regardless of whether the bottle-instances
ever have contained or will contain Coke. If kind-level reference is basic and objectlevel reference is syntactically mediated, then this sort of “non-referential” value is
precisely what we expect from nouns embedded as bare nominal stems in a larger
lexical predicate.
But the point extends well beyond this familiar observation. Since concepts are
more than an abstraction from their instances, the view that nouns denote concepts
provides a single revealing account for three distinct phenomena.
The first is the Canonical Use Constraint of Kiparsky (1997): if a verb V is derived
from a noun N, the semantic contribution of N to V is not entirely derivable from
the objects in the extension of N. For example, shelve and saddle mean something
like ‘to put in a shelf-like structure for storing’ and ‘to place a weighty implement on
a living being for a purpose’, not just ‘to put in a shelf ’ and ‘to put a saddle on’, as
shown by the oddness of the examples in (13) (see Borer, this volume, for critical
discussion).
(13)
a. # he had absent-mindedly shelved the screws and forgot where he’d put
them
b. # she saddled the floor (for a moment)
A related phenomenon was brought to light in connection with denominal verbs
(Clark and Clark 1979), but Aronoff (1980) pointed out that it is much more general.
6
It is not clear to me whether the property X in X-[pl] is necessarily expressed by an existing predicate
in the language, as the English examples suggest. If so, this is another empirical prediction. H. Borer and
H. Harley (p.c.) have both pointed out to me that rapids (in a river) and shorts (as a garment) clearly
identify a category concept, even though they are morphologically parallel to contents, and keep a
synchronic semantic connection with the adjectives rapid and short. The conclusion must be that this
morphological type can, but does not have to, create nouns which do not identify a kind.
The roots of nominality, the nominality of roots
47
The example in (14a) is not a contradiction because to nurse does not mean ‘to be a
nurse and behave accordingly’, but something like ‘to do what a nurse does’: the
meaning of the verb is not based on the set of entities nurse is true of (not even
intensionally constructed), but on a concept that defines being a nurse on the
basis of prototypical activities and functions. This component is what good modifies
in (14b).
(14) a. he nurses well (but he’s not a nurse)
b. he’s a good nurse (but he’s not a nurse)
‘he does well what nurses do’
Thirdly, consider “Compositional Idioms” like plastic flower or kosher bacon (Katz
and Pitt 2000). No entity that satisfies the description flower or bacon can ever be,
respectively, made of plastic or not made of pork. Yet these phrases are very different
from round square, and we know immediately what they must mean. Instead of
positing hidden syntactic structure to mediate the sense ‘materially constituted of ’,
as Katz and Pitt (2000) do, we can read plastic flower as ‘individual in plastic related
to the concept flower’, rather than instantiating it. As in (14b), the modifier qualifies
a description of object-level entities, and the noun identifies a concept, without the
former being an instance of the latter. Far from being an ad-hoc device, the relation
to a concept is independently needed to describe the semantics of root compounds
like dog bed, and I suggest it also underlies derivational suffixations like tenner ‘entity
related to the number ten’ (cf. (8) above).
3.7 Semantic filtering in derivational morphology
It is important to stress that although concepts belong to the conceptual/intentional
cognitive component, and not to grammar per se, naming concepts is a semantic
property of nouns as linguistic terms. We might expect this semantic fact to have
repercussions elsewhere in the grammar, and this is precisely what seems to happen
in a puzzling constraint in English derivational morphology. Aronoff (1976) noted
the following generalization: denominal adjectives like ornamental have the structure X-ment-al, where X is not a verb (more precisely, not a verb form):
(15)
a.
b.
*XV -ment-al
ornament
regiment
employment
agreement
shipment
X cannot be a verb (form)
*ornaV
ornamental
regimental
*regiV
employV
agreeV
shipV
(Aronoff 1976:54)
*employmental
*agreemental
*shipmental
Thus stated, the generalization is arbitrary. It is also non-absolute: Aronoff noticed
the exceptions govern-ment-al and develop-ment-al, to which we can add judgement-al. One could conceivably view the base X as a category-free root, in the sense
48
Paolo Acquaviva
of Distributed Morphology; (15) would then restrict -al affixation to root nominalizations X-ment, where X must be a root and not a verb. But, apart from not
explaining the exceptions, this would incorrectly rule in cases like agree-ment-al, if
the verb agree has the structure √agree-Øv, where agree spells out the root.
The exceptions hold the key to a more revealing and predictive account. Governmental is acceptable in so far as it means ‘relative to government’, not ‘relative to the
act of governing’. In pointing this out, Aronoff also notes that -al otherwise has no
problems attaching to affixes that turn verbs into action nominalizations, witness
organ-iz-ation-al. In fact, the nominal suffixes -ation, -ence, -or actually select -al as
adjectival suffix (organizational, reverential, professorial). No such preferential relation relates -ment and -al, which does not mean that X-ment-al is impossible but
only that, aside from -ation, -ence, -or, the choice of -al is a property of individual
words and is not driven by morphology. But there is a distinct type of selection
involving -al, semantic and not morphological, restricting the bases which -al can
attach to (‘inwards’), and this is where concepts make the right distinction. Semantically, -al suffixation derives a property-referring adjective from an entity-referring
noun. But not all nouns are entity-referring in the required sense. Depart-ment-al
must mean ‘relative to a department’, where depart-ment is morphologically bimorphemic but semantically atomic: it cannot be interpreted as ‘the act of departing’.
Likewise, govern-ment-al means ‘relative to government’ where this is an institution,
an entity (concrete or abstract), not the property nominalization from ‘to govern’. It
can hardly be accidental that judge-ment-al means ‘having a censuring attitude’, not
‘relative to judging’; and develop-ment-al seems better paraphrased as ‘relative to the
developing phase in life’ rather than ‘relative to developing’ in general. This
approach is empirically revealing: we can add argumental to the list of exceptions,
where argu-ment describes an entity (a logical category) and not the property
nominalization form ‘to argue’. In sum, -al can be suffixed to X-ment- if this base
describes an entity, in a sense that excludes abstract property nominalizations: this is
the sense in which concepts define entities. The base must define a category of being,
concrete or abstract, in such a way that its semantics does not reduce to the content
of a verb treated as an argument, like destruction in (5). In sum:
(16) a. -al is selected by certain nominalizing affixes: -ation, -ence, -or (organizational, reverential, professorial), but not -ment
(morphological selection, outwards)
b. otherwise, -al attaches to a N expressing a concept (≠ transparent nominalization)
(semantic restriction, inwards)
This explanation is predictive and falsifiable, because it identifies an implication
between a certain interpretation of X-ment and the possibility of X-ment-al. If the
latter is morphologically admissible, the former must denote a concept, that is, it
The roots of nominality, the nominality of roots
49
must define an entity over and above the content of a property nominalization
(the implication does not go in the other direction). Usage can introduce new forms
XV -ment-al; but in such cases XV -ment should always mean more than ‘the act/fact
of XV -ing’. And this prediction, right or wrong, involves reference to concepts,
defined as detailed in section 3.1 in such a way as to exclude transparent deverbal
nominalizations.
3.8 Impossible nouns
The approach being defended here envisages two basic values for lexical nouns: a
nominalized property based on a predicate, or a kind-level entity corresponding to a
concept. The first was exemplified by transparent deverbal nominalizations like
destruction, agreement, and contents, and although much more needs to be said
about them, I will leave them aside.7 The second comprehends proper names and
common nouns, which refer rigidly to kind-level concepts and are mainly distinguished by how the concept is related to a discourse referent. Since the concepts
named by nouns are elements of the speaker’s conceptual ontology, they must be
cognitively relevant and definable across possible worlds. The following descriptions
arguably fail to characterize concepts meeting these requirements, and so we predict
that no natural language lexicalizes them as the content of a simplex common noun:
(17)
a. number of planets
b. round square
c. not-blue thing
no world-invariant concept
contradictory property
non-possession of a property
No name can refer across all possible worlds to the category of being ‘the number of
planets’, because this can only be fully defined by reference to the planets in a set of
worlds. More generally, natural languages do not seem to name entities which are
definable only in a given world and situation, like instants in a particular event or
stages in the existence of a particular entity. We have names for ‘war’ as an event
type, and arguably for particular events like World War II (as distinct from the
definite description the second world war), but not for ‘the first day of war’. Likewise,
if 1,000 travelling persons correspond to 2,000 passengers in a year, we do not
identify the passenger-events by name, as opposed to the persons. We can name
instead the category ‘passenger’, as a stage type defined independently of particular
events in a world (along with so-called phase sortals like boy or caterpillar). Since
formal languages have no problem identifying such particulars by means of logical
constants, this limitation points to a distinctive property of natural language.
‘Round square’, on the other hand, is a legitimate way to describe an entity in a
world-independent manner; but, being a contradictory property, it would lack any
There is probably an additional variety, which I will not discuss: “dummy sortals” like thing or entity
(Griffin 1977:60, Pelletier 1979:11).
7
50
Paolo Acquaviva
cognitive content as a name for a category of entities. Being ‘not-blue’ is also a wellformed predicate, and it is true of a non-null domain across worlds; but as a
maximally uninformative description, not even restricted to colored things or to
existing things (unlike ‘non-blue’), it seems exceedingly unlikely, if not formally
impossible, as a category in the conceptual ontology of any community of language
users.
Examples such as (17) illustrate that not all predicates can correspond to a rigidly
identifiable kind, especially when it must be cognitively relevant as a category of
conceptual ontology. While these remarks are far from a fully worked-out account,
they show that an approach along these lines can lead to an explanatory theory of
possible nouns in natural language.8
3.9 Concepts and the content of roots
If the analysis proposed here is correct, there is a linguistically relevant distinction
between nouns that denote a type of entity and nouns that merely nominalize a
predicate. The notion of category concept as elucidated in section 3.1 is an attempt to
capture this distinction. Two major consequences follow. First, we cannot define
nouns as expressions that denote concepts, since some do and some do not. Second,
if a word like argument is made up of the same morphemes in the nominalization
reading (‘arguing event’) as in the concept-denoting reading (‘logical category’), then
it is not the choice of morphemes that selects the interpretation. Concepts are not
encapsulated in morphemes, at least not as a general rule. Where, then, are they
linguistically encapsulated?
The question here is not the largely terminological one whether the two senses of
argument, transmission, or constituency are two nouns or one. The question is what
formal object we are talking about when we say that a noun names or does not name
a concept; part of the general question of what makes a noun a noun. The preceding
sections have provided some partial answers. I have argued that being an argument
and identifying a discourse referent are key nominal properties, but they are not
properties of nouns as lexical items; and that naming a concept is a lexical property
that only nouns have, but not all. We can now add that nouns which merely
nominalize a predicate must have a different underlying structure from those that
denote concepts, as illustrated by Borer’s (2003) syntactic analysis of destruction in
(5), which I follow. However, a difference in syntax can explain why the two readings
differ, but it does not tell us where the concept reading comes from—and so fails to
identify the source of a fundamental aspect of nominality as a lexical property. In a
framework where nouns are built on a category-neutral root, the answer centers on
the status of roots. To see why, suppose the two readings of argument have the
8
A fuller treatment should derive the requirements that Gupta (1980) formalized as modal constancy
and modal separation: what is one N at world w cannot be two distinct Ns at world w0 .
The roots of nominality, the nominality of roots
51
structures outlined in (18). As in (5), Event P and Aspect-Quantity P identify a verbal
structure, while the NP level of (5) is replaced by nP, and the root is simply labelled
ROOT, in keeping with (3):
(18) a. [nP -ment [EP [AspQ [ROOT argue ]]]]
b. [nP -ment
[ROOT argue ]]
Plausibly, it is the content of the verbal heads that determines the verbal interpretation
of the predicate in (18a). But (18b) contains a subset of the heads in (18a), and yet it
denotes an entity-defining concept. Neither -ment nor the root, by themselves, can
trigger this interpretation, since they appear in (18a) too. Nor does grammar automatically determine a concept reading for any root embedded under [n], if only
because not all roots appear as nouns. Otherwise, grammar should allow us to embed,
for instance, the root of long in a nP, to name some entity-type distinct from the
nominalization ‘length’. This is not what happens; concepts are not forced into
existence like that. Notice that we are not forced to assume that argument has a
monomorphemic structure in its reading as a result noun which denotes an event (but
is not an argument-structure noun); that may well be the correct analysis for cases like
proposition verbalized as to proposition, as Borer argues (this volume, see her (69)).
In sum: concepts are what distinguish entity-denoting nouns from predicate
nominalizations; this contributes to our understanding of nominality as a lexical
property; when we ask what in a word “means” a concept, we face fundamental
questions about the content of roots:
.
.
do roots determine (encapsulate, select . . . ) nominality?
do roots determine lexical semantics—in particular, the interpretation as a concept?
A positive answer to the first question conflicts with the assumption of category-free
roots, yet most studies that embrace this assumption also posit that the information
contained in roots determine a lexical category, directly via class diacritics like
“declension II” or indirectly through lexical meanings like ‘growing process’ or
‘cutting event’ (see Acquaviva 2009 for critical discussion). An influential variant
of the latter position holds that roots are typed as modifiers of events or entities, or
as “names” for entities, states, relations, or manners (cf. Marantz 2001). A simplistic
approach where all lexical roots are so typed, and where types correlate one-to-one
with syntactic categories, would just reinstate nouns, verbs, adjectives, and prepositions as morphosyntactic primitives. Besides, the noun–verb distinction cannot be
anchored so easily in an ontological distinction between events and entities. Not only
are events also entities; many nouns are true of events, like war, party, or event (as
well as adjectives like protracted, presumably).9 It would seem more plausible to view
9
In connection with event-denoting nouns, Carlson (1998:39) notes that “principles that apply to verbs
and sentences might well not carry over to these expressions, a point often left aside in the literature on
events.”
52
Paolo Acquaviva
this sort of typing as a consequence of the meaning inherent in a root: some roots
might have very specific content and are compatible with a single lexical category,
while others may be semantically underspecified and syntactically flexible. This
seems to be the most widespread interpretation of category-free roots, worked out
in some detail in Arad (2003, 2005).
The problem with root meanings is that they are so elusive. On the basis of
Modern Hebrew, Arad (2005) argued for semantically contentful but underspecified
roots, identifying broad categories like mental activity, relation, motion, positional
state, manner, or emission (p. 104). Empirical arguments may well force us to
conclude that Hebrew and Semitic roots are semantically contentful and categorytyped (cf. especially Doron, this volume).10 But the behavior of roots in Semitic does
not readily generalize across languages (as Arad 2005 points out).
Apart from the major issue of cross-linguistic generalization, characterizing the
elusive content of roots is problematic because it must do without a great deal of
idiosyncratic, word-specific information. Basilico (2008:735–736), for instance,
clearly shows that selectional restrictions are not encapsulated in a root, since they
change according to whether a verb appears alone or in construction with a particle:
(19)
a. the author wrote a story/#an idea
b. the author wrote up an idea
Of course, being in construction with a particle also affects transitivity (Basilico
2008:742), as in the poet is writing/#the poet is writing up; but while argument
structure is plausibly a function of the syntactic context, selectional restrictions
would seem to be an idiosyncratic lexical property, associated with a listeme and
not given by grammar. Examples like (20) tell us that, whatever this listeme is, the
root alone does not express it. Indeed Basilico associates this interpretation with the
Root+particle or Root+[v] complex, not with the root, although he follows Arad
(2003) in thinking that roots are contentful but compatible with a range of interpretations, fixed by the element they combine with (p. 741). Notice that if selectional
properties are stripped away from the content of a root, serious theoretical consequences follow. The claim by Hale and Keyser (2005:19), for instance, that the
“conflation” between V and its object N in verbs like dance reduces to semantic
selection, becomes problematic if their V is a category-free root and does not encode
selectional properties. Along with selectional properties, the putative content of roots
can lack other fine idiosyncratic characterizations. Carlson (1998:43) argues that the
verbs in they kissed and she kissed her are “two separate lexical items”, because their
encyclopedic content varies subtly, only the second sentence being true in a situation
10
Arad (2005:102–103) notes that natural kinds and kinship terms tend to have unrelated roots in
Hebrew, while “Hebrew frequently and productively creates names for places, institutions, instruments
and artefacts from highly underspecified roots.”
The roots of nominality, the nominality of roots
53
where A kissed B’s hand. Whatever it is that encodes this difference, it is not the
invariant root.
The reflection on root semantics has privileged the verbal domain. Category
concepts bring these suggestions into focus for the domain of nominal lexical
semantics. In a nutshell, there is robust empirical evidence that these concepts are
sometimes identified by roots alone and sometimes by roots plus something else. It
is easy to find examples where the same invariant base is associated with different,
sometimes sharply different concepts in different grammatical environments:
(20)
same root, distinct environments, distinct concepts: class
a. class 3 4 m-ti ‘tree’ mi-ti ‘trees’
Swahili (Polomé 1967:95, 103)
class 7 8 ki-ti ‘chair’ vi-ti ‘chairs’
b. class 11/4 9/10 u-siku ‘night’ siku ‘day’
(21)
same root, distinct environments, distinct concepts: gender and number
a. mālus ‘apple tree’ (fem) mālum ‘apple’ (neut)
Latin
b. membro masc ‘member’,
pl. membra fem. ‘limbs’ membri masc. ‘members’
Italian
The Italian plurals in (21b) illustrate the widespread phenomenon of plural doublets
identifying distinct “senses” where this vague notion is sharpened by reference to
concepts as entity-defining category (cf. Acquaviva 2008a). Inflectional doublets can
interact with derivational morphology, as in the Russian network illustrated in (22):
(22)
singular
tsvet ‘color’
tsvet-ók ‘flower’
plural
tsvetá ‘colors’
tsvetý ‘flowers’
Russian
‘Flower’ is a basic-level concept, but the noun tsvetók is derived in the singular by
suffixation of the singulative (diminutive) -ok to the base tsvet-. This same base,
morphologically an unanalyzable root, means ‘color’. In the plural, the two readings
are distinguished not by suffixation, but through different inflectional endings
(a common occurrence in Russian and similar inflecting languages). This description
simplifies the synchronic outcome of language change, as it leaves aside the use of
tsvet in the meaning of ‘flower, blossom’ (as well as the regular plural tsvetkí from
tsvetók; cf. also the verb tsvestí ‘to blossom’). The point is that, synchronically, the
same root expresses two concepts of sharply different entities.
Note that invoking root homonymy for every such case would mean giving up on
the idea of roots as meaningful entities distinct from words. Consider the two
readings of collection as transparent nominalization (as in the frequent collection of
mushrooms by visitors) and as so-called “result nominal” (as in come see my stamp
collection). These readings are related, in fact the label “result nominal” presupposes
54
Paolo Acquaviva
that the object-denoting one arises from the same predicate underlying to collect. If
we assume that they share a root, we can express the difference between this pair and
clear homonyms like the readings of plant as ‘vegetable’ and ‘industrial works’. But
sharing the same root evidently does not guarantee a key lexical semantic property of
nouns, namely, their ability to identify a category of entities.
In sum, if roots have any content by themselves, in many cases it is so vague and
coarse-grained that it cannot define category concepts. This result reinforces the
view that roots are assigned interpretation only in a context (Arad 2003, 2005,
Marantz 2001). Why view them as meaningful in themselves, then? As the phonological side of Semitic roots provides a template that by itself is unpronounceable, it
is coherent to think that the semantic side is also uninterpretable on its own. If so,
the “underspecified” meaning of roots that are not obviously nominal or verbal,
often invoked but never elucidated (Arad 2005 is the only exception I know of ), is so
elusive because it is an epiphenomenon; the reification of semantic relations between
meaningful larger units. The alternative, explored in Acquaviva (2009), is to view
roots as syncategorematic morphemes, acting as differential indexes for semantically
interpreted constructions that embed them; essentially, like cran- in cranberry.
Notice that this predicts the existence of semantically empty roots, naturally
accommodating two types of widespread phenomena that are otherwise problematic: roots shared by words that are historically related but synchronically conceptually unrelated, like re-ceive, con-ceive, or per-ceive (cf. Aronoff 1976); and the
use of the same roots as open-class and closed-class words, commonplace in
classifier languages (see Contini-Morava and Tobin 1999, Aikhenvald 2003:353–
367). The latter phenomenon finds a parallel in lexical verbs that double up as
grammatical elements, like the German werden ‘to become’ or the Italian venire ‘to
come’ used as passive auxiliaries, or the Irish caith ‘to cast, spend’ which in the future
has an idiomatic deontic reading (caithfidh mé a rá ‘I must say’).
The point may be strengthened: if category-neutral roots do not determine an
interpretation specified for a logical type, they are not intelligible. As noted by
Panagiotidis (2009), this would explain why roots must be categorized at all (Embick
and Marantz 2008), relating in a principled way grammatical and semantic underspecification.
3.10 The content of roots
Joining together the threads of this discussion, our approach leads to specific
conclusions about roots as theoretical objects.
Roots do not stand in a one-to-one relation with concepts. Semantically atomic
nouns need not also be morphologically atomic. This sums up the evidence considered in the last section, and generalizes to Borer’s (this volume) observation that
morphologically complex nouns like audition or proposition “have exactly the cluster
The roots of nominality, the nominality of roots
55
of properties we find with underived nouns.” In particular, they allow verbal conversion (to proposition), unlike other conceptually atomic complex nouns like
collection–*to collection. Reference to concepts clarifies the intuition that nouns
like audition or proposition are semantically lexicalized, even though they have the
same internal structure as transparent nominalizations like imposition (cf. *to
imposition). More precisely, all nouns with structure [root + nominal affix] which
allow verbal conversion (like proposition) must identify a concept as a whole
complex and not by their root alone, although the opposite does not hold.
It is not grammar that determines which complex noun identifies a concept as a
whole, which is plausible given the gradience and idiolectal variability of some
judgments (cf. especially Lowenstamm, this volume). Concepts, as understood
here, affect grammar as the semantic values of nouns (where nouns are not a
fixed-sized syntactic construction), but they have a cognitive reality outside of the
generative language component. Given a non-grammatically determined pairing of
concepts with nouns, the language faculty deterministically fixes how the linguistic
pieces making up a noun can be combined and interpreted. This means that there is
ultimately one and the same reason why to proposition and argumental are acceptable but *to imposition and *agreemental are not: as nouns, proposition and
argument are the minimal encapsulations of a concept, imposition and agreement
are not.
If roots do not correlate with concepts, they do not correlate with the category
‘noun’ either. When collection is a transparent nominalization, the root -lect- underlies a verbal structure. As a concept name, collection incorporates no such verbal
structure: the shared root is neither nominally nor verbally typed but truly categoryfree, unless we stipulate as many homophonous root alternants as the lexical
categories they can instantiate. As noted, this makes it very problematic to assume
that any given root is categorized as a modifier of one basic ontological type like
event, (stative) property, or entity (contrast Roßdeutscher, this volume, and the
literature cited there, in particular Marantz 2001).
Of course, not all roots are so clearly category-free. In many cases, and to different
degrees of systematicity according to language, roots do determine lexical category.
But this is a strength of the present account, not a weakness. Concepts allow us to
define a central aspect of lexical nominality anchoring it outside grammar, in the
function of linguistic terms as concept names. The advantage of this perspective is
that we do not have to view roots as the source of all idiosyncratic information in a
word. But if roots are not defined in terms of a content they often lack, nothing bars
them from having a privileged or even unique association with a certain lexical
category or a certain type of content. In a coherently separationist perspective, roots
spell out those non-categorial nodes around which grammar builds “words”.
Depending on language-particular constraints and historical accident, the choice
of root correlates more or less directly with the choice of lexical category and of
56
Paolo Acquaviva
lexeme-specific content (such as naming a concept). This perspective captures
distinctions that would be obliterated by a simplistic contextual definition of lexical
categories. Consider cost and price, both monomorphemic, both possible as noun
or verb:
(23) a.
in ’85, the house cost very little
the cost was prohibitive
Lola cost/costed the project
b. *in ’85, the house priced very little
the price was reasonable
the agent priced the house reasonably
Although both roots can appear as verbs, price lacks the stative interpretation (cf.
Harley 2005). In this reading (and for many speakers only in this reading), cost
exhibits the lexeme-specific behavior of a strong verb, which denominal verbs
systematically lack. Correspondingly, the stative interpretation is the one that does
not rely on a nominal concept. In sum, price is primarily nominal. There is no need
to posit a grammatical marking on the root or an invisible nominalizing morpheme
(cf. Borer, this volume, for the inadequacies of this type of solution): simply, price
names a concept, as a historical fact about the English vocabulary. Grammar does
the rest: the root may appear in a nominal or in a verbal syntactic frame, but the
interpretation will include the corresponding concept.
3.11 Conclusion
The manifold facets of verb lexical semantics offer many perspectives for lexical
decomposition. Nouns, by contrast, appear uninteresting: they are true of their
referents, perhaps with some complications for event-dependent and relational
nouns. I have explored an alternative perspective, which does not take noun
referents for granted but instead focuses on the conceptualization of entity-types
as the distinctive property of nouns. This perspective on noun lexical semantics is
descriptively useful and leads to a number of predictions; above all, it allows a
characterization of nominality as a primitive lexical property. This leads to the
question whether concepts are the content of lexical roots. The facts strongly suggest
this is not the case. The evidence from nouns, then, is that roots do not encapsulate
core lexical information, but are best seen as morphological atoms, as opposed to
atomic mappings of sound and meaning.
4
Roots in transitivity alternations:
Afto-/auto-reflexives
ARTEMIS ALEXIADOU
4.1 Introduction
Cross-linguistically, there are three semantic/conceptual classes of reflexive verbs. In
a language with a so-called two-form reflexive system such as Dutch, these are
distinctly marked1. As is well known, Dutch has two reflexive pronouns, the simple
reflexive pronoun zich and the complex reflexive pronoun zichzelf. Dutch makes a
morpho-syntactic distinction between three classes of reflexive verbs. Crucially, this
morpho-syntactic distinction reflects a semantic or conceptual difference (see Kemmer (1993) and references therein). The following verb-classes can be identified:
Inherently reflexive (or inherently reciprocal) verbs: these verbs do not allow
replacing the reflexive pronoun with a referential DP. Furthermore, only the simple
reflexive pronoun is allowed in languages like Dutch:2
(1)
Jan
schaamt zich/*zichzelf/*Marie.
John shames refl/refl-self/Mary
‘John is ashamed.’
Naturally reflexive (or naturally reciprocal) verbs: the reflexive pronoun can be
replaced by a referential DP. In out-of-the-blue contexts, the simple reflexive is
strongly preferred. (The complex reflexive becomes acceptable under focus.)
1
I am indebted to Elena Anagnostopoulou, Terje Lohndal, Florian Schäfer, Giorgos Spathas, Jim
Wood, and Katerina Zombolou for insightful comments and discussions. My research has been supported
by a DFG grant to project B6 Underspecification in Voice Systems and the Syntax–Morphology Interface
within the Collaborative Research Center (SFB) 732 Incremental Specification in Context at the Universität
Stuttgart.
2
Interestingly, the class of inherent/absolute reflexive verbs in English is rather limited, and for this
English also uses the reflexive pronoun, see Brame (1983), and recently Siemund (2010), who suggests that
such forms are lexicalized with the verb, cf. Geniušiene_ (1987).
58
(2)
Artemis Alexiadou
a. Jan
waste
zich/??zichzelf/Marie.
John washed refl/refl.self/Mary
‘John washed (Mary).’
b. Jan scheerde zich/??zichzelf/Peter.
John shaved refl/refl.self/Peter
‘John shaved (Peter).’
Naturally reflexive verbs come from a number of semantic subclasses which all
represent events that carry “. . . inherent in their meaning [ . . . ] the lack of expectation
that the two semantic roles they make reference to will refer to distinct entities . . .”
(Kemmer 1993:58). So-called “grooming verbs” such as ‘shave’, ‘wash’ or ‘dress’ form
one main subgroup of naturally reflexive verbs; a further group would be verbs of
movement. Naturally reciprocal verbs involve, for example, verbs of social (‘meet’) or
affectionate (‘kiss’) events but also verbs of antagonistic events (‘fight’).
The third class are naturally disjoint verbs (called other directed verbs in Koenig
and Vezzosi 2004): a referential DP can replace the reflexive pronoun. The complex
reflexive is strongly preferred. We can say that these verbs express events that carry
the expectation that the two semantic roles they make reference to will refer to distinct
entities (e.g., ‘hate’, ‘accuse’, ‘kill’, . . . ).
(3)
Zij haat ??zich/zichzelf/Peter.
She hates refl/refl-self/Peter
‘John hates himself/Peter.’
Other languages, such as Greek, use the same non-active (NAct) morphology that
appears in the passive—and with certain anti-causative predicates—to mark the first
two classes, see (4b) and (5) (see Tsimpli 1989, Embick 1998, Alexiadou and Anagnostopoulou 2004, Zombolou 2004, and references therein).
(4)
(5)
a. O Janis
diavase
to
vivlio.
the John-nom read-ACT the book-acc
‘John read the book.’
active
b. To vivlio diavastike apo to Jani.
the book read-NAct by the John
‘The book was read by John.’
passive
a. I
Maria
drepete.
the Maria-nom
shame-NAct-3sg
‘Maria is ashamed of herself.’
inherent reflexive
b. I Maria
htenistike.
the Mary-nom
combed-NAct-3sg
‘Mary combs (herself ).’
naturally reflexive
Roots in transitivity alternations: Afto-/auto-reflexives
c. I Maria
htenise
the Mary-nom combed-ACT-3sg
‘Mary combed Peter.’
59
ton Petro.
the Peter-acc
There is no transitive counterpart of (5a) with active morphology, while there is one
of (5a) as in Dutch and English, (5c).
Of special interest here is the counterpart of (3). In Greek, as in English and
Dutch, a transitive variant is possible containing the complex reflexive DP ton eafto
tu, (6a), Iatridou (1988), Anagnostopoulou and Everaert (1999), Spathas (2010).
However, an intransitive variant is also possible via non-active morphology, which
carries reflexive interpretation but only when prefixed with the element afto- ‘self ’.
In this case, the complex form cannot occur (6b):
(6)
a. O
Janis
katigorise ton eafto tu/to Petro.
the John-nom accused the self his/the Peter
‘John accused himself/Peter.’
b. O
Janis
afto katigorith-ik-e
*ton eafto tu.
the John-nom self accused-NAct-3sg the self his
‘John accused himself.’
In (6b), it is the combination of the element afto and the non-active morphology
which is responsible for the reflexive interpretation; without afto the result is a
passive, and without the non-active the afto-prefixed form is ungrammatical.
A similar pattern is found across Romance involving auto-prefixation, a learned
loan from Greek. In French, Italian, and Spanish, transitive/other directed predicates
can be reflexivized by auto, and in this case they necessarily require the presence of
si/se (Italian data in (7) from Castella 2010):
(7)
a. Gianni incorona se stesso.
Gianni crowns himself
Italian
b. Gianni si è autoincoronato.
Gianni refl is self-crowned
‘Gianni crowned himself.’
The data in (6b) and (7b) raise several questions concerning the status of the afto-/
auto- prefix, and the status of non-active morphology and reflexive clitics respectively. Several authors have assumed that it is the presence of afto/auto that marks
the predicate as reflexive, and as a result non-active morphology and reflexive clitics
can no longer be seen as bringing about a reflexive interpretation (see, e.g., Embick
1998, Labelle 2008, and Marelj and Reuland 2012).
In this chapter, I will first be concerned with establishing the factual background
on afto-/auto-prefixation. As will be shown, these prefixes are sensitive to the type of
predicate involved, i.e., they must attach to an other-directed predicate, which,
60
Artemis Alexiadou
however, must be an agentive accomplishment. I will show that afto-/auto- can only
appear on verbs that can bear non-active morphology and importantly in the absence of
the prefix receive a passive only interpretation (Spathas, Alexiadou, and Schäfer to
appear). This supports an analysis of afto-/auto- according to which its role is distinct
from that of non-active morphology and that of reflexive clitics. When the two
co-occur, non-active morphology functions as a marker of detransitivization, in
agreement with what the literature claims the function of this morphology is.
The question I will then turn to is the syntax of afto-prefixed predicates. Three
main approaches have been proposed in the literature, which all treat afto- as a
reflexivizer: a) Tsimpli (1989) argues that afto- reflexives are passives that take a
reflexive interpretation via the presence of the reflexive clitic afto-. b) Embick (1998,
2004b) argues that afto- predicates are unaccusative, and afto attaches to the root
and reflexivizes it by means of compound formation. c) Labelle (2008) argues that
auto- predicates combine with a special kind of Voice head, labeled middle Voice
(see Doron 2003, Alexiadou and Doron 2012), and the prefix attaches to the verb in
the lexicon. This Voice head does not introduce an external argument, but can
introduce the former internal argument of the root, as an external argument.
While I will partially agree with Tsimpli and Embick that afto- takes a passive/
unaccusative syntax as its input, I will argue that afto- is not a reflexivizer. Instead, it
is an actor-oriented intensifier, which in combination with a passive syntax asserts
that its associate (the internal argument) is the only agent in all relevant sub-events
(Spathas, Alexiadou and Schäfer to appear). Crucially, afto- functions like an
agentive intensifier (see Howell 2012, Gast and Siemund 2006). The main point is
that we need to distinguish between reflexivizers and triggers of reflexive interpretation. Reflexivizers are functions that turn relations into reflexive predicates. Triggers
of reflexive interpretation, which I argue afto- is, bring about a reflexive interpretation only when combined with a Voice structure that lacks an external argument.
This is a reflexivization strategy only available in languages that have a semantically
weak, underspecified, valency reducing morphology, such as Greek, see Gast and
Siemund (2006). As we saw above, in Greek, passives, anti-causatives and reflexives
all surface with non-active morphology.
The behavior of this prefix provides important insights into the relevance of root
ontological classification (manner vs. result) and its interaction with syntax, as well
as on the nature of Voice, which hosts non-active morphology in Greek. Since afto-/
auto- does not combine with naturally reflexive predicates, built on the basis of
manner roots, and mono-eventive predicates in general, this suggests that the
encyclopedic classification of roots as manner vs. result plays an important role
with respect to the type of structures these may occur in (see e.g., the contributions
of Doron, Rappaport Hovav, Anagnostopoulou and Samioti. Levinson, and
Roßdetuscher in this volume, contra Acedo-Matellán and Mateu). In addition, this
analysis has implications for the type of Voice available in languages like Greek. The
Roots in transitivity alternations: Afto-/auto-reflexives
61
facts discussed here provide support for the view advanced in Alexiadou and Doron
(2012) that the type of Voice head involved is actually underdetermined/unspecified
for the passive vs. reflexive distinction, i.e., it is a middle Voice head (Alexiadou and
Schäfer 2014, Doron 2003, this volume, Alexiadou, Anagnostopoulou and Schäfer, to
appear). This Voice is interpreted as reflexive in the context of naturally reflexive
predicates, and as passive in the context of naturally disjoint predicates, in the
absence of overt material to yield a different interpretation. Predicates that only
yield an anti-causative interpretation in the presence of non-active morphology are
ruled out as input to afto-prefixation, even if they involve result roots.
The chapter is structured as follows: in section 4.2 I describe the distribution of
afto-/auto-. In section 4.3, I turn to the analysis of afto-, and compare it to previous
analyses. In section 4.4, I conclude my discussion.
4.2 The distribution of afto-/auto-prefixation
4.2.1 Lexical categories that combine with afto-/autoIn Greek and Romance, afto-/auto- can attach to nouns, adjectives, and verbs, see,
e.g., Zombolou (2004), and Arquiola (2003) for Spanish, Dugas (1992) for French.
Greek examples are given in (8):
(8)
a. Verbs:
b. Nouns:
c. Adjectives:
afto-katastrefome
afto-kritiki
afto-didaktos
self-destroy-NAct
self-criticism
self-educated
While in (8a) the verb bears non-active morphology, the complex relationship
between afto- and the non-active morphology is clarified by a class of verbs which
have afto-, but no non-active morphology. Examples of this are given in (10), see also
Zombolou (2004) and Embick (1998). Similar examples are reported in Dugas (1992)
for French, see also Mutz (2004):
(9)
Active afto-verbs
Afto-Form
aftoktono
aftomolo
aftoshedhiazo
Translation
Transitive Form
commit suicide *ktono
change sides
*molo
improvise
shedhiazo = ‘design’
The difference between (8a) and (9) is that whereas (8a) has an active transitive
counterpart without afto, with aftoktono there is no corresponding verb ktono. In the
case of aftoshedhiazo, there is a form shedhiazo, but the former is not the reflexive of
the latter. Thus, I conclude that the forms in (9) do not represent the productive
62
Artemis Alexiadou
pattern of afto-prefixation in Greek, which is related to reflexivization and is
illustrated by examples such as (8a) and (6).3
Note that in English, self-prefixation applies mainly to adjectives and nominals,
and similar observations hold for German (see Chomsky 1970, Edmondson and
Plank 1978, Grimshaw 1982, König 2011):
(10)
a. self-destructive, self-absorbed, self-conscious, self-educated
b. self-pity, self-interest, self-study, self-destruction
There are some verbal forms, generally considered to be back formations, as can be
seen by the fact that verbal form contains the nominal ‘stem’ in, e.g., (11a):
(11)
a. This game will self-destruct in 10 minutes.
b. You need to self-educate before you can progress in your career.
In this chapter, I will focus on afto-/auto- prefixation on verbs, hence the English
data will not be further discussed.4
4.2.2 Verb class restrictions
4.2.2.1 A transitivity restriction? At first sight it seems that afto- prefixation is only
possible with transitive predicates. Further Greek examples are given in (12), from
Zombolou (2004:171):
3
The prefix has been argued to have two further functions, see Dugas (1992), and Mutz (2004): a)
auto- can function as an anti-causative marker, which indicates an action that takes place without the
agent, and b) auto- can function as a focalizing/excluding marker, which does not modify the argument
structure of the element it attaches to, but it acts only on the referential level by indicating that the
participants of the action referred to are not the prototypical ones, but some less expected ones. In
Romance, according to Mutz (2004:357) and Dugas (1992), these two other uses are primarily found in the
nominal domain. The distribution of the Greek data suggests that maybe it is the focalizing marker that is
at work.
4
Chomsky (1970) pointed out that in this is a self-inflicted wound, there is no antecedent to license self,
suggesting that the prefix is not subject to the same conditions that license reflexive pronouns. Moreover,
he observed that self-educated cannot have the passive as its input, as passives are incompatible with
reflexivization. A further point is that in several cases the interpretation is not obligatorily a reflexive one.
For instance, self-study is not study of oneself, but rather refers to something to be studied alone. Moreover,
in several cases a transparent reflexive verbal structure cannot be transformed into a self-compound, e.g.,
*Hamlet is self-shooting/self-killing/self-amusing/self-frightening (from Edmondson and Plank 1978).
An important point made by Edmondson and Plank (1978), and taken up in König (2011), is that in
English these compounds involve the intensifier/emphatic element self-, and not the complex reflexive
pronoun (*himself-educated). This is particularly clear in German, where intensifiers differ in form from
reflexives. Thus, these authors concluded that intensification proper rather than reflexivization is involved
in self-prefixation:
(i)
a. Selbstbildnis, Selbstkritik
self-blindness self-criticism
b. *Sichbildnis, *michselbstkritik
sefl-blindness myself-criticism (Edmondson and Plank 1978:394)
Roots in transitivity alternations: Afto-/auto-reflexives
(12)
a. afto-eskorizome
self-exile-NAct-1sg
‘exile myself ’
b. afto-anerume
self-refute-NAct-1sg
‘refute myself ’
c. afto-diorizome
self-appoint-NAct-1sg
‘appoint myself ’
d. afto-diafimizome
self-adverstize-NAct-1sg
‘advertize myself ’
e. afto-katigorume
self-blame-NAct-1sg
‘blame myself ’
f. afto-sikofandume
self-defame-NAct-1sg
‘defame myself ’
g. afto-katadikazome
self-sentence-NAct-1sg
‘sentence myself ’
h. afto-penevome
self-praise-NAct-1sg
‘praise myself ’
63
The list in (12) suggests that the predicates that undergo afto- prefixation are
transitive (or take small clause complements, e.g., appoint) and become reflexive
via afto-/auto-. This conclusion is supported by the fact that afto-prefixation is not
allowed with monadic predicates in Greek, neither unergatives (12a), nor unaccusative (13b), even if they surface with non-active.
(13)
a. *afto-gelieme
self-laugh
*afto-perpatieme
self-walk
b. *afto-erhome
self-come
*afto-petheno
self-die
The above suggests only ordinary transitive predicates undergo regular afto-prefixation (i.e., predicates that are naturally disjoint). As already mentioned, in Greek,
outside the context of afto-prefixation, these verbs can be construed in a reflexive
interpretation only via the presence of an anaphor DP:
(14)
O
Janis
katigori
the John-nom accuses
‘John accuses himself.’
ton eafto tu.
the self his
When they appear in the non-active morphology, and without afto-, the only
available interpretation is a passive one:
(15)
O
Janis
katigorite.
The John-nom accuse-NAct-3sg
‘John is accused.’
64
Artemis Alexiadou
To achieve a reflexive interpretation, crucially afto- must attach to the verb and nonactive morphology must surface:
(16)
O
Janis
aftokatigorite.
The John-nom self-accuse-NAct-3sg
‘John accuses himself.’
4.2.2.2 Afto- as a reflexivizer Inherently reflexive verbs do not permit aftoprefixation either (17c). Importantly, however, naturally reflexive verbs disallow
afto-prefixation as well, although these verbs can in principle appear in a transitive
construal (18a vs. 18b):
(17)
a. O
Janis
drepete.
the John-nom shame-NAct-3sg
‘John is ashamed of himself. ’
b. *O Janis drepi ti Maria.
The John shames the Mary-acc
c. *O Janis afto-drepete.
John self-shames-NAct-3sg
(18)
a. O Janis
pleni ti Maria.
the-John washes the-Mary-acc
‘John washes Mary.’
b. O Janis
plenete/*afto-plenete.5
the-John-Nom wash-NAct-3sg/self-wash-NAct-3sg
The above restriction has been pointed out in Embick (1998), who observed that in
Greek, only Kemmer’s (1993) ‘body action’/grooming verbs, e.g., shave, comb, etc.,
are interpreted as reflexive via non-active morphology. These verbs cannot combine
with afto-.
A similar restriction is discussed for Romance.6 As is well known, in these
languages, naturally reflexive predicates are formed on the basis of the reflexive
5
(19b) is arguably ambiguous between a passive, which is made explicit in the presence of a by-phrase,
and a reflexive reading, Tsimpli (1989). Theophanopoulou-Kontou (1980) claims that a passive reading is
not available.
(i)
O Janis plenete
apo ti mitera tu.
the John washes-NAct by the mother his
‘John is washed by his mother.’
6
Grimshaw (1982:109) notes that in English self- can only be added to dyadic adjectives, and it is in
complementary distribution with the second argument of the adjective:
(i)
a. *self-tall
b. Bill is self-confident/*Bill is self-confident of his abilities.
c. Bill is confident of his abilities.
Roots in transitivity alternations: Afto-/auto-reflexives
65
clitic se/si (19a). Predicates that are other-directed can form reflexives via attaching
the prefix auto-, but, as mentioned in section 4.1, the presence of se is required in this
case. For instance, Labelle (2008) points out that in French, predicates like autoanalyser ‘self-analyse’ (also autofinancer ‘self-finance’, autoproclamer ‘self-proclaim’,
autocélébrer ‘self-celebrate’, autodétruire ‘self-destruct’) are reflexive via auto. In
spite of this, these verbs require the reflexive clitic, as shown by the ungrammaticality of (19c), (see also Dugas 1992). See also Mendoza (2010) and Arquiola (2003)
for Spanish, and Castella (2010) for Italian:
(19)
a. Jean se lave.
Jean refl wash-3sg
‘John washes himself.’
b. Jean s’ autoanalyse.
Jean refl self-analyze-3sg
‘John self-analyzes himself.’
c. *Jean autoanalyse.
As in Greek, auto- cannot attach to inherent/natural reflexives. This is illustrated in
(20) with a Spanish example from Mendoza (2010). See also Castella (2010) for
Italian, and Labelle (2008), Dugas (1992) for French:
(20)
*Pedro se auto-lava.
‘Pedro washes himself ’
This led Embick and Labelle to formulate the following generalization:
(21) a. Afto-/auto- prefixation reflexivizes roots/predicates.
b. Non-active voice morphology/se clitics do not reflexivize
verbs, but appear on verbs which are reflexivized by other
means; i.e., by virtue of being ‘inherently reflexive’, or by
virtue of afto/auto.
However, the distribution of this prefix is sensitive to very subtle factors: it is
sensitive to the passivizabilty of the predicate and to its aspectual classifiction: it
has to be an accomplishment predicate.
4.2.2.3 Two restrictions: Aktionsart + passivizability First, note that the associate of
afto- cannot be the goal argument of a di-transitive predicate, while this is possible in
Romance, (22b) from Arquiola (2003):
(22) a. *afto-doristike ena aftokinito.
self-donated a car
b. autorregalar-se un libro
Spanish
‘to give oneself a book as a present’
66
Artemis Alexiadou
This relates to the fact that reflexivization of a dative/benefactor argument is
generally out in Greek, see Papangeli (2004), but possible in Romance. Note,
however, that the associate of afto- can be the theme argument of a ditransitive
predicate (22c):
c. afto-sistinome sti Maria
self-introduce.NAct.1sg to-the Maria
‘I introduce myself to Mary.’
The above leads to the conclusion that afto- prefixation is only allowed with
predicates that are transitive and importantly can form a passive. This explains
why we have the distibution in (22), as theme arguments of ditransitives, but not goal
arguments, can passivize (Anagnostopoulou 2003).
Further support for this comes from the observation made in Spathas, Alexiadou
and Schäfer (to appear) that afto- prefixation is out with
a. transitive predicates that can form anti-causatives but not passives (23, Alexiadou
and Anagnostopoulou 2004),
b. verbs that cannot form a passive due to a gap in the morphological paradigm
(e.g., break in 24, Alexiadou and Anagnostopoulou 2004), and
c. deponent verbs (25), i.e., verbs with non-active morphology (Embick 1998) and
an animate subject, which, however, cannot passivize,
suggesting that it is the ability to form a passive and not the presence of non-active
morphology that matters.7
(23)
a. O Janis
ekapse to spiti.
the John-nom burnt the-house-acc
‘John burnt the house.’
b. *To spiti kaike
apo ton Jani.
the house burnt.NAct by the John
c. *afto-kaika.
self-burn.NAct-1sg
(24)
a. *afto-spastika.
self-break-NAct-1sg
b. *spastike
apo to Jani.
break-NAct-3sg by the John-acc
7
Katerina Zombolou (p.c.) points out that several deponent verbs such as aminome ‘defend-NAct’
permit afto-prefixation: afto-aminome ‘self-defend-NAct’. It is not clear to me how to account for cases
like this at the moment. It could very well be that these involve back formations from the self-N form, as is
the case with English self-destruct.
Roots in transitivity alternations: Afto-/auto-reflexives
67
(25) a. *afto-metahirizome.
self-use-NAct-1sg
b. metahirizome
use-NAct-1sg
The behavior of the class of psych-verbs further supports this generalization. As
(26–28) show, subject experiencer and dative experiencer predicates do not allow
afto-/auto-prefixation. While dative experiencer predicates are considered unaccusatives (Anagnostopoulou 1999, Belletti and Rizzi 1988) and hence are expected not
to undergo afto-/auto- prefixation, subject experiencer predicates are other-directed
predicates, can form a passive, but still do not tolerate afto-prefixation.8 Arquiola
(2003), Castella (2010), and Dugas (1992) report similar facts for Spanish, Italian, and
French respectively:
(26)
a. O Janis agapai ton eafto tu/ti Maria.
the John loves himself/the Mary
‘John loves himself/Mary.’
subject experiencer
b. *O Janis aftoagapiete.
the John self-loves-NAct
(27)
a. O Janis ftei ston eafto tu.
the John bothers the self his
‘John bothers himself.’
b. *O Janis afto-ftegete.
the John self-bother-NAct-3sg
(28)
*autoamarse ‘to love oneself ’, *autoodiarse ‘to hate oneself ’
Arquiola (2003:500)
Object experiencer predicates can undergo afto- prefixation, but only on their
agentive reading. As Alexiadou and Ioardachioaia (2012) point out, annoy-type
object experiencer predicates can be analyzed as a sub-type of the causative alternation. Their intransitive variants as well as their transitive variants can be interpreted either as agentive or as causative. Evidence for this comes from the fact that
an agentive by-phrase or a causer PP are licit in the case of the intranstive variant
that bears non-active morphology (29b). An agent or a causer can surface as the
subject of the transitive variant as well (29a). On the other hand, interest-type
predicates are unambiguously stative (Anagnostopoulou 1999); their external argument is not an agent/cause but rather a target/subject matter, see Doron (2011). This
can be seen from the fact that their intransitive variant can be modified by a PP that
is distinct from that introducing agents/causers (29b vs. 30b) (Castella 2010):
8
But cf. Landau (2010).
68
(29)
Artemis Alexiadou
a. Ta nea/o Janis
enohlisan ti Maria.
the news/the John annoyed the Mary.acc
‘The news/the John annoyed Mary.’
b. I Maria enohlithike
apo to Jani/me ta nea.
the Mary annoyed.NAct by the John/with the news
‘Mary got annoyed by John/from the news.’
c. I Maria
aftoenohlithike
*me ta nea.
the Mary self-annoyed-NAct-3sg with the news
‘Mary annoyed herself *by the news.’
(30)
a. Ta fita/i Maria
endiaferun to Jani.
the plants/the Mary-nom interest the John
‘Plants/Maria interest John.’
b. O Janis
endiaferthike ja ta fita/ja ti Maria/*apo ti Maria.
the John interested-NAct for the plants/for the Mary/by the Mary
‘John was interested in plants/in Mary/*by Mary.’
c. *O Janis afto-endiaferete.
the John self-interest-NAct-3sg
(29c) and (30c) show that only annoy-type predicates can undergo afto-prefixation.
Finally, note that agentive/causative object experiencer verbs that have intransitive
variants that surface with active morphology and cannot form a passive disallow
afto-prefixation (31d):
(31)
a. Ta nea/ta pedia
thimosan to Jani.
the news/the children angered the John-acc
‘The news/the children angered John.’
b. O Janis thimose
me ta nea/me ta pedia.
the John angered-3sg with the news/with the children
‘John got angry with the news/with the children.’
c. *O Janis thimothike
the John angered-NAct
apo ti Maria.
by the Mary
d. *afto-thimono/-me.
self-anger-ACT-1sg/NAct-1sg
As we saw, afto-prefixation is out with stative predicates. It is also out with
achievement predicates (32b), even if they can form a passive.9
(32)
a. O Janis vrike ta gialia tu.
the John found the glasses his
‘John found his glasses.’
9
Thanks to Florian Schäfer for raising this issue.
Roots in transitivity alternations: Afto-/auto-reflexives
69
b. Ta gialia vrethikan.
the glasses found-NAct-3pl
‘The glasses were found.’
c. *Ta gialia aftovrethikan.
the glasses self-found
As the external argument of achievement predicates is an experiencer and not
an agent, (32a–c) support the generalization that afto-/auto-prefixation takes place
only if the external argument of a transitive predicate can receive an agentive
interpretation.
Note further that non-core transitive predicates, in the sense of Levin (1999),
which are standardly classifed as activities, also do not tolerate afto- prefixation (33),
even if they can form a passive:
(33) *afto-skupistike
self-swept
*afto-magireftike
self-cooked
Non-core transtive verbs can form a passive, e.g., to fagito magireftike apo ti Maria
‘the meal cooked-NAct by Mary’. Thus, the problem seems to be related to the fact
that these predicates are non-core transitive, i.e., they have a mono-eventive structure, and not a bi-eventive, accomplishment structure.
On the other hand, the verbs that can undergo afto- prefixation are core transitive
predicates in the sense of Levin (1999). As can be seen in (34), like other core
transitive predicates (35), and unlike non-core transitive ones (34b), verbs that can
have the afto-prefix do not allow unspecified objects.
(34) a. *O Janis espase.
the John-nom broke-3sg
intended: ‘John broke.’
b. O Janis skupise.
the John-nom swept-3sg
‘John swept.’
(35)
*O Janis eksorise.
the John-nom exiled-3sg
intended: ‘John exiled.’
I thus conclude that there is sufficient evidence for the generalization that predicates
that can undergo afto-prefixation obey two conditions: i) they have to be agentive/
causative accomplishments, and ii) they can undergo passivization. This supports
Arquiola’s (2003) observation that if the external argument of the predicate is not an
agent or a cause, auto-prefixation is out. In other words, ‘being a dyadic predicate is
a necessary condition, although not a sufficient one. In addition to that, the external
70
Artemis Alexiadou
argument must be an agent or must have the possibility of receiving an agentive
interpretation’ (Arquiola 2003:500).
4.2.2.4 Afto- is not a reflexivizer The overall behavior of afto- suggests that it is
very similar to agentive emphatic reflexives in English, which are also only
compatible with accomplishments (Howell 2012), see (36):
(36)
a. John built the house himself.
b. #John forgot the keys himself.
c. #John loves Mary himself.
accomplishment
achievement
state
As Hole (2002) also notes, non-agentive predicates are barred from interacting with
agentive intensifiers. There are two more sets of data that further support this
comparison, from Spathas, Alexiadou, and Schäfer (to appear). First, note that
agentive emphatic reflexives/intensifiers (AERs) indicate direct involvement of
their associate (the external argument), (37).
(37)
John built the house (full/partialLY/half ) himSELF.
‘John built the house (fully/partially/half ) without help.’
Afto-predicates, like AERs, can be modified by degree expressions that measure the
level of involvement of the external argument (prosodic prominence, indicated with
capital letters, is necessary for the relevant reading). The transitive variant doesn’t
have the relevant reading, (39).
(38)
O
Janis sxedhon AFTO-eksoristike.
the John almost
self-exiled.NAct
‘John exiled himself almost without help.’
(39)
O Janis sxedhon eksorise ton EAFTO tu.
the John almost exiled the self
his
6¼‘John exiled himself almost without help.’
Importantly, however, it can be shown that afto- is not a reflexivizer: focused
reflexivizers under focus-sensitive operators, e.g., negation, induce two distinct
inferences ((40), Spathas 2010). The Greek reflexive anaphor only allows Infer1,
(41). Focused afto- behaves like neither; it only gives rise to Infer2, (42).
(40)
John didn’t exile himSELF.
a. Infer1: John exiled someone other than John.
b. Infer2: Someone other than John exiled John.
Roots in transitivity alternations: Afto-/auto-reflexives
(41)
O Janis den eksorise ton EAFTO tu.
the John not exile.NAct the self
his
a. Infer1: John exiled someone other than John.
b. # Infer2: Someone other than John exiled John.
(42)
O Janis den AFTO-eksoristike.
the John not self-exile.NAct
a. #Infer1: John exiled someone other than John.
b. Infer2: Someone other than John exiled John.
71
If afto- is indeed not a reflexivizer, but rather a trigger of reflexive interpretation, then we
would expect to find cases in which a) afto is impossible, and b) it is only the presence
of non-active morphology that leads to a reflexive interpertation, contra Embick.
Indeed, there is a class of predicates that receive a reflexive interpretation in the
absence of afto-prefixation in Greek. Papangeli (2004) gives a list of predicates that
receive a reflexive interpretation only via the presence of non-active morphology:
(43)
klinome ‘close myself ’, anigome ‘open myself ’, klidonome ‘lock myself ’
htipieme ‘hit myself, psych interpretation’, hapakonome ‘fill myself with pills’,
etc.
Note that the predicates involved are core transitive verbs. A similar observation is
made in Alexiadou and Anagnostopoulou (2004), who point out that several verbs
that undergo the causative alternation in Greek when they occur in the anticausative variant with animate arguments bear the non-active form only and have
a reflexive (and often a psych) interpretation (45b). See also Zombolou (2004) who
observed that not all transitive predicates can have a reflexive interpretation in their
intransitive form; only those that are causative can:
(44)
a. O Janis
lerose
to trapezomandilo.
the John-nom dirtied-ACT the tablecloth-acc
‘John dirtied the tablecloth.’
b. To trapezomandilo lerose/lerothike
apo mono tu.
the tablecloth
dirtied-ACT/-NAct by itself
‘The tablecloth got dirty.’
(45)
a. O Janis
lerose
ti Maria.
the John-nom dirtied-ACT the Mary-acc
‘John dirtied the Mary.’
b. I Maria
lerothike/*lerose.
the Mary-nom dirtied-NAct/ACT
‘Mary dirtied herself. ’
72
Artemis Alexiadou
What is now the difference between these verbs and the ones in (8)? Certain of
these predicates cannot form a passive, thus they don’t obey the second restriction
associated with afto-prefixation. In addition, the most natural interpretation of
the predicates in (43) and (45) in the presence of non-active morphology and
an animate subject is either a subject experiencer, or predicates expressing a change
in body posture, i.e., actually a natural reflexive interpretation. Thus afto- is disallowed.
To conclude, the discussion in this section supports the view expressed in Manney
(2000), according to which the predicates that require afto to express reflexivity are
those that denote actions that the subject does not normally perform upon himself.
For Manney, these verbs use the prefix to emphasize that it is indeed the subject who
carries out the action upon himself.
In the next section, I turn to an analysis of afto-/auto- predicates.
4.3 Towards an analysis of afto- prefixation
4.3.1 A distributed morphology perspective
In section 4.2, I concluded that afto-/auto- attaches to predicates that are otherdirected accomplishments and can form a passive. Building on Manney (2000), Hole
(2002), and Howell (2012) the semantic contribution of the prefix is to emphasize
that the agent of the event is no other than the predicate’s internal argument.
In this section, I introduce the building blocks that are relevant to compute the
syntax and semantic contribution of afto-. I assume that several functional heads are
involved in the encoding of verbal meaning. The verbal predicate itself is decomposed into two layers: i) a v head, which is a verbalizer head bringing about event
implications, and ii) a root (see Marantz 2005, Alexiadou, Anagnostopoulou and
Schäfer 2006 (AAS)). I further assume that roots fall into two main classes depending on their encyclopedic semantics: manner vs. result (following Rappaport Hovav
and Levin 2010, see also Anagnostopoulou and Samioti, Rappaport Hovav, and
Roßdeutscher, this volume). According to Rapapport Hovav and Levin, verbs built
on the basis of manner roots specify as part of their meaning a manner of carrying
out an action (e.g., run, laugh), while verbs built on the basis of result roots specify
the coming about of a result state (e.g destroy, open). This categorization influences
a) the way roots merge into the structure, and b) the number of arguments they can
contain. Following Embick (2004b), there are two ways according to which roots
merge with verbalizers, either as modifers of v or as complements of v:
(46) a. modifiers of v, direct Merge
v
e.g. hammer
√
v
b. complements of v
v
e.g. flatten
v
√
Roots in transitivity alternations: Afto-/auto-reflexives
73
The type of merge is sensitive to the manner vs. result/state classification of roots.
Manner roots merge as modifiers of v, state roots merge as complements of v. The
former give rise to activity, non-core transitive, predicates, while the latter to
accomplishment, core-transitive, predicates.
(47) is thus the structural representation of an unaccusative/anti-causative predicate, e.g., the door opened. The root expresses a result state.
(47)
[vP v [ Root/ResultP ]]
The external argument is severed from the verbal predicate and is introduced by a
functional verbal head (Voice in Kratzer 1996, Marantz 1997, AAS 2006, Bruening
2012), as in (48).
(48)
[Voice [v [ ResultP ]]]
I propose that afto-attaches to Voice; see Hole (2002) for German selbst. It selects
a Voice structure that lacks an overt external argument (see Bruening 2012), hence
NActive morphology (Embick 1998, AAS 2006), see (49). It contributes the information that its associate is the only agent in every sub-event e, the event predicated
over. Unlike adverbial intensifiers, the associate of afto- is the internal argument.
(6b) can be paraphrased as follows: “There was an event e of someone accusing John
and John was the agent of every sub-event of e.”, see Spathas, Alexiadou and Schäfer
(to appear) for details:
(49)
[afto [VoiceP +Ag [vP [ResultP]]]]
What can afto-prefixation tell us about Greek Voice (alternations)? I think it
provides support for the view that non-active Voice in Greek is underdetermined/
underspecified. It is actually middle Voice (Doron 2003, this volume, Alexiadou and
Doron 2012, Alexiadou and Schäfer 2014, Spathas, Alexiadou and Schäfer to appear).
Greek, unlike Hebrew and English/German, does not have a dedicated passive Voice
head. The default interpretation of middle Voice is reflexive with naturally reflexive
predicates and passive with naturally disjoint predicates. In the presence of overt
lexical material that specifies its implicit argument it gives rise to non-ambiguous
interpretations: a passive, when an agentive by-phrase is present ((50a) Alexiadou
and Doron 2012, Alexiadou and Schäfer 2014), with naturally reflexive predicates,
and a reflexive in the case of naturally disjoint predicates prefixed with afto- (50b):
(50)
a. O Janis
plithike
apo ti Maria.
the John-nom washed-NAct-3sg by the Maria
‘John was washed by Mary.’
b. O Janis
afto-eksoristike.
the John-nom self-exiled.NAct-3sg
‘John became self-exiled.’
74
Artemis Alexiadou
The above helps us understand why afto- prefixation is out with naturally reflexive
predicates. Middle Voice receives its default interpretation in this context, hence
overt material is ruled out as superfluous. In addition, wash is not an accomplishment predicate, it is an activity predicate. Hence, afto-prefixation might be out for
this reason. It should be pointed out, however, that many speakers of Greek never
identify the non-active form of wash with a passive reading. If this is the case, then
there is an additional reason why afto-prefixation is expected to be out. But see
Zombolou (2004), Castella (2010), and Labelle (2008) for a different line of argumentation.
Afto-prefixation works only with accomplishment predicates that can form a
passive: thus accomplishment verbs that don’t allow the prefix, but surface with
non-active morphology in intransitive variants can safely by classified as anticausative only (e.g., Greek kill, *afto-skotothike ‘self-killed-NAct’/*skotothike apo
to Petro ‘killed-NAct-S3g by Peter’).
We saw that in Romance it has to be se/si that appear together with auto-. From
the perspective of Greek, the predicates that undergo auto-prefixation should be
seen as reflexive passives, raising thus the question whether middle Voice is also
available in Romance. As has been noted in the literature (e.g., Belletti 1982 and
many others), reflexive passives in Romance can only be built from verbs that can
assign accusative Case, and hence are out with unergative and unaccusative predicates. Standardly, these cases are analyzed as follows: se bears the agent thematic
role and accusative Case, while the DP argument saturates the internal thematic role
and bears nominative Case. For this reason, reflexive passives cannot co-occur with
agentive PPs, although they contain an implicit external argument.
Why do English and German not have such prefixes? As Gast and Siemund
(2006) note, actor-oriented intensifiers can function as reflexive markers in contexts
in which the main verb has undergone some operation of diathetic change, and
typically, they are used in combination with a middle marker as expressions of
reflexivity. This is exactly the case in Greek (and also in Russian, and in Otomí, see
Gast and Siemund 2006). However, English and German, unlike Greek, do not have
underspecified valency-reducing morphology, i.e., the passive is always a passive,
and the reflexive is sufficiently distinct. This in turn supports Alexiadou and Doron’s
(2012) analysis of the Greek non-active morphology as middle Voice, a head underspecified for a passive vs. reflexive interpretation, see also Spathas, Alexiadou and
Schäfer (to appear).10
In the next section, I turn to a comparison with previous accounts.
10
What about English (and German) self (see note 4)? The prefix is productively used to form synthetic
compounds in English (Conti 2007), and it has actually been identified as an intensifier (König 2011). See
Bruening (2014) for a recent discussion self-prefixation in the context of adjectival passives.
Roots in transitivity alternations: Afto-/auto-reflexives
75
4.3.2 A comparison with previous accounts
As far as I know, none of the previous analyses discuss the verb class restrictions
established here. The analysis presented here shares some of the intuitions expressed
in Theophanopoulou-Kontou (1980) and Tsimpli (1989) who argued for a passive
analysis of afto-predicates, though it differs greatly in details and implementation.
Theophanopoulou- Kontou proposes that afto- is similar to the agentive by phrase,
the difference being that it incorporates into the verbal lexeme. Tsimpli (1989) argues
that afto- reflexives are passives that take a reflexive interpretation via the presence
of the reflexive clitic afto- under I°, which is co-indexed with the suffix -te. The
passive suffix absorbs the external thematic role and the co-indexation between afto
and the passive suffix leads to the reflexive interpretation. I argued here against the
status of afto- as a reflexivizer.
The analysis differs from Zombolou (2004) who argues that afto- is a manner
adverb corresponding to the phrase apo monos tu ‘by oneself ’, which offers, in her
view, a uniform account of self-prefixation across bases, (51–52):
(51)
a. O Janis eksoristike
apo monos tu.
the John exile-NAct by alone his
‘Kostas exiled himself by himself. ’
b. *O Janis afto-eksoristike apo monos tu.
the John self-exiled-NAct by alone his
(52)
a. aftodinamos = dinatos apo monos tu
self-powered powered by oneself
b. afto-kratoras = kratai apo monos tu
self-ruler ‘emperor’ rules by oneself
c. aftoptis = vlepi apo monos tu
eyewitness see by oneself
Assuming, as in Schäfer (2007), that the by itself phrase has an interpretation along
the lines of ‘without outside force’, I note that this is also possible with naturally
reflexive verbs, which, however, do not allow afto-prefixation. This is not accounted
for under Zombolou’s analysis:
(53)
O Janis den plenete
apo monos tu.
the John neg wash-NAct-3sg by himself
‘John does not wash by himself.’
In addition, note that the compounds in (52a–c) are neither productive nor compositional in meaning.
The analysis here also differs from that offered in Rivero (1992) who takes afto to
be a case of noun incorporation and explains the presence of non-active Voice as
an instance of Case absorption. While also here non-active morphology signals
76
Artemis Alexiadou
de-transitivization, there is no argument incorporation. In a sense, Rivero’s analysis
is close to the LF-incorporation analysis of the Greek reflexive pronoun eaftos that
has been proposed in Anagnostopoulou and Everaert (1999). In (54), the predicate is
reflexive-marked since one of its arguments is a SELF anaphor. This means that two
arguments of the predicate should be coindexed to satisfy Reflexivity Condition A of
Reinhart and Reuland (1993). This is not the case, however, since the two elements
that are coindexed in (54) are, structurally, arguments of different predicates:
(54)
[O Petros]j agapai [ton eaftoi tuj]i.
the Petros(N) loves the self(A) his(G)
‘Petros loves himself. ’
To deal with this issue, these authors proposed that another type of complex
predicate formation applies, one in which the noun and the verb are composed.
This is possible because eaftos is a defective noun, and such nouns can be assumed to
be licensed by abstract incorporation, see (55):
(55)
[o Petros]i eaftoj-agapai [ton tj tui]i
As a result, the verb–noun complex counts as one syntactic and semantic predicate.
Again, however, Rivero’s account predicts that all predicates that can take the
complex anaphor as a complement should allow afto- prefixation, but this is clearly
not the case, e.g., wash and the psych predicates discussed in section 4.2.
Embick (2004b) considers two options, the first one being that afto is the realization of an external argument that has cliticized onto the verb. The analysis presented
here shares the intuition that afto is somehow related to the external argument, but
as an agentive intensifier. The second option that actually Embick argues for is that
afto- is in effect adverbial. As Rivero (1992) argued, Greek shows what is classified as
‘adverb incorporation’ more generally. Embick suggests that afto-prefixation is a
type of compounding, and crucially involves attachment to a Root. However, under
Kratzer’s view, according to which the external argument is severed from the root,
crucially Embick’s analysis, according to which afto attaches to the root, cannot be
adopted (a point made by Spathas, Alexiadou and Schäfer to appear and in Ahn
2012; see Hole 2002).
Finally, this analysis also differs from the ones presented in Labelle (2008),
Arquiola (2003), and Castella (2010), see also Marelj and Reuland (2012), who
claim that reflexivity must be visible at the level where auto-prefixation applies,
i.e., “early” in the derivation.
I will concentrate on Labelle’s (2008:844f.) account here. Labelle claims that the prefix
auto- derives monadic verbs from dyadic verbs in the lexicon. By prefixing a twoargument verb with auto- we obtain a different verb selecting only one argument.
Schematically, this is given in (56) below. According to this analysis, afto-/auto- is a
Roots in transitivity alternations: Afto-/auto-reflexives
77
reflexivizer and not a trigger of reflexive interpretation in the sense defined in the
introduction:
(56)
analyzer = λxλyλe [analyze(e, x) and Agent(e, y)]
→autoanalzser = λxλe [analyze(e, x) and Agent(e, x)]
She further notes:
When an auto- prefixed verb is introduced in syntax, it selects an Agent coreferential with the
object. This selectional restriction is as in (56), where the prefix auto introduces the Agent in
the verb’s lexical entry. Importantly, for this analysis to work, the verb’s external argument
cannot be severed from the root. The verb is then treated as an exception to Kratzer’s
generalization that external arguments do not appear in the lexical representation of verbs.
Adding se is redundant. It does not add information to what the verb lexically contains. The
Agent contributed by se is non-distinct from the Agent role already present in the lexical entry
of the verb. When the DP subject is introduced in Spec,Voice, it is interpreted as being both
the Agent and the Theme of the event, as required by the meaning of the verb.
Labelle stresses that se is obligatory, because the meaning of the verb requires the
Agent to be the same entity as the Theme. Without se, Active Voice would introduce
a distinct variable for the subject, and we would end up with two distinct Agents,
violating the principle that a thematic role can only be assigned once. The sentence
would be semantically incoherent. In this sense, we may say that se is a grammatical
means of classifying a predicate as reflexive: it classifies the predicate as reflexive by
marking it as being one in which two arguments, the external argument and an
internal argument, are coreferential. This analysis cannot be adopted, if we follow
Kratzer’s view that external arguments are severed from the root.
Note also that the above would suggest that afto-prefixed predicates should
behave like unergatives. In fact Tsimpli (1989, 2006) took such predicates to be
unergatives. Tsimpli notes that a derived subject in Greek cannot control into
purpose clauses, as shown in the passive example in (57a). In contrast, subjects of
naturally reflexive predicates and afto- reflexives can (57b–c). This suggested to
Tsimpli that the subjects in (57b–c) cannot be analyzed as ‘deep’ objects:11
(57) a. *O Janis dolofonithike ja na gini iroas.
the John murdered-NAct-3sg for subjunct become hero
‘John was murdered to become a hero.’
11
Note here that (57c) is problematic for Tsimpli’s analysis, according to which afto-reflexives are
actually passives, as the derived subjects of passives do not allow control into purpose clauses (57a). She
thus argues in this case event control is involved. It is not problematic for the present account, since aftocombines with middle Voice, which is very similar to passive Voice, the crucial difference between the two
being the absence of disjoint reference effects with middle Voice, Spathas, Alexiadou, and Schäfer (to
appear).
78
Artemis Alexiadou
b. I Maria htenistike ja na vgi ekso.
the Mary combed-NAct-3sg for subjunct go out
‘Mary combed herself to go out.’
c. O Janis aftodioristike diefthindis ja na elenchi tin eteria.
the John self-appointed-NAct-3sg directer fo subjunct control the company
‘John appointed himself as director to control the company.’
However, applying unaccusativity diagnostics to such predicates, points to a different behavior. First of all, natural reflexives can form adjectival participles, a test
which, according to Markantonatou (1992), distinguishes unaccusative from unergative predicates in Greek.
(58)
a. pesmeno filo
fallen leaf
b. *tregmenos anthropos
run man
c. plimeno pedi
washed child
d. aftoeksorismenos politikos
self-exiled
politician
Second, Alexiadou and Anagnostopoulou (1999) showed that in Greek, possessor
sub-extraction is possible from the postverbal subject of an unaccusative verb, but
not from that of an unergative (cf. Anagnostopoulou 1999). Applying this test to
natural and afto- reflexives again suggests that they pattern with unaccusative/
passive predicates:
(59)
a. Tinos diavases to vivlio?
whose read-2sg the book-acc
‘Whose book did you read?’
b. Tinos efige to aftokinito?
whose left-3sg the car
‘Whose car left?’
c. *Tinos etrekse to aftokinito?
whose ran-3sg the car
intended: ‘Whose car ran?’
(60)
a. Tinos plithikan
ta pedia?
whose washed-NAct-3pl the children
‘Whose children got washed?’
b. Pias horas
aftoeksoristikan
i politiki?
which country self-exiled-NAct-3pl the politicians
‘The politicians of which country got self-exiled?’
Roots in transitivity alternations: Afto-/auto-reflexives
79
4.4 Conclusions
In this chapter, I investigated the properties of the verb classes that take afto(/auto-). I first showed that it is sensitive to the type of predicate involved, i.e., it
must be an other-directed agentive/causative predicate that can crucially build a
passive form with non-active morphology, whose presence is an indication that
de-transitivization has taken place. I then proposed an analysis, according to which
afto-/auto- attaches to a passive structure, actually to a non-active structure involving middle Voice, which is interpreted as a default passive in the case of naturally
disjoint predicates.
This analysis suggests that roots do have some meaning in isolation, and that this
influences the morpho-syntactic combinations they can enter in important ways, i.e.,
result roots enter as complements of verbalizers and give rise to bi-eventive structures, manner roots enter as modifiers of v, a view shared by other contributions in
this volume as well. Importantly, however, the analysis shows that it is the structural
environments roots are found in that is responsible for the restrictions on aftoprefixation. If, crucially, causative/accomplishment interpretations arise in that roots
can enter the derivation in very specific ways, depending on their ontological
classification (manner vs. result), and that external arguments are severed from
the root, then afto- triggers a reflexive interpretation only when it combines with a
non-active structure of a causative predicate.
According to my analysis, we need to draw a distinction between elements that are
truly reflexivizers and elements that simply trigger a reflexive interpretation. The
latter are present only in languages such as Greek, where non-active morphology is
underspecified. Embick’s and Labelle’s generalization is partially correct: non-active
morphology is a signal of de-transitivization, it signals the presence of a special
Voice, namely middle Voice, underspecified for the passive vs. reflexive distinction
(Alexiadou and Doron 2012, Alexiadou and Schäfer 2014, Spathas, Alexiadou, and
Schäfer to appear, Alexiadou, Anagnostopoulou and Schäfer to appear). The default
interpretation of middle Voice is reflexive with naturally reflexive predicates and
passive with naturally disjoint predicates.
Finally, note that there is another prefix in Greek, namely the reciprocal prefix
alilo ‘each other’, which can combine with prototypical reciprocal predicates (Kemmer 1993) and is subject to two restrictions: the predicate must surface with nonactive morphology, and its sole argument must be plural:
(61)
Ta pedia
alilokitunde.
the children each-other look-NAct-3pl
‘The children look at each other.’
Labelle (2008) offers a similar description for French: predicates like entreregarder
‘look at one another’ (also entraccuser ‘accuse one another’, entraider ‘help one
80
Artemis Alexiadou
another’, entredéchirer ‘tear one another to pieces’, entredétruire ‘destroy one
another’) receive a reciprocal interpretation via the prefix. In addition, they all
require a plural subject and the presence of the reflexive clitic, as shown by the
ungrammaticality of the b example below.
(62)
a. Les participants s’ entreregardèrent.
the participants refl entre-look-at-PST-3P
‘The participants looked at one another.’
b. *Les participants entreregardèrent.
In Greek (and Romance) reciprocal prefixation is out with naturally reciprocal
predicates. As is well known, these languages use non-active morphology and the
reflexive clitic se to mark naturally reciprocal predicates, respectively:
(63)
a. *Ta pedia alilosinandithikan.
the children each-other met-NAct-3pl
b. Ta pedia
sinandithikan.
the children met-NAct-3pl
‘The children met each other.’
This prefix is not sensitive to the same Aktionsart restrictions that characterize afto-:
for instance, as (64) shows, alilo- prefixation is allowed with stative subject experiencer predicates:
(64)
Ta pedia
alilomisiunde.
the children each-other-hate-NAct-3pl
‘The children hate each other.’
It is not sensitive to the presence of a passive structure, i.e., it is possible with clear
anti-causatives:
(65)
Ta pedia
aliloskotothikan.
the children each-other-kill-NAct-3pl
‘The children kill each other.’
Obviously the issue awaits further research, but in principle, one would need to say
that the reciprocal prefix combines with a Voice structure that lacks an overt
external argument (hence non-active morphology is present). See Bruening (2004)
for some discussion on verbal reciprocals.
5
Domains within words and their
meanings: A case study
ELENA ANAGNOSTOPOULOU AND YOTA SAMIOTI
5.1 Introduction
The goal of this chapter is to investigate the conditions under which idiosyncratic
meaning arises in word formation.1 More specifically, we investigate the hypothesis
put forth in Arad (2003, 2005) and Marantz (2001, 2007) that idiosyncratic meaning
derives from locality conditions on the interpretation of roots. This hypothesis is
stated in (1):
(1)
The Marantz/Arad Hypothesis
Roots are assigned an interpretation in the context of the first category assigning
head/phase head merged with them, which is then fixed throughout the derivation
We test (1) in the domain of Greek verbal adjectives and adjectival participles which
provide extensive evidence for attachment below vs. above higher heads (i.e., little v,
Voice) of the adjectivizing/participial suffixes -tos and -menos, respectively. Our
investigation leads to the following results:
(i) Affixes serving as verbalizers and the little v head contributing eventiveness
cannot be collapsed. Greek verbal adjectives are shown to be semantically and
syntactically small (i.e., they do not have event implications, a semantic
property correlating with the fact that they do not support adverbs and
other modifiers), and yet they may contain verbalizers inside the adjectivizing
1
Different versions of this chapter have been presented at the 32nd GLOW Colloquium (University of
Nantes, April 15–18 2009), the Roots Workshop (University of Stuttgart, June 2009) and the Workshop on
Linguistic Interfaces (University of Ulster, December 2–4 2010). We would like to thank the audiences for
questions, suggestions and discussions, especially Asaf Bachrach, Orin Percus, Peter Svenonius, Gertjan
Postma, Danny Fox, Alec Marantz, David Embick, Andrew Koontz Garboden, Hagit Borer, Lisa Travis,
Christina Sevdali, Raffi Folli, and Heidi Harley. Special thanks to Winfried Lechner, Artemis Alexiadou,
and Terje Lohndal for their comments.
82
Elena Anagnostopoulou and Yota Samioti
suffix -tos. This leads us to propose that there are two ways of interpreting
the Arad/Marantz hypothesis, depending on whether we take (morphological) verbalizers or (abstract) eventivizing heads to qualify as the relevant
phase heads in (1).
(ii) If we look at the interpretation of an unspecified root, as suggested by Arad
(2003, 2005) on the basis of Multiple Contextualized Meaning (MCM) in
Hebrew, then Greek provides some support for the view that specified
meaning is indeed determined at the point of categorization. Even though
many Greek roots have a rather fixed meaning, unlike Hebrew, there are also
roots whose meaning is unspecified, similarly to Hebrew. Interestingly, the
presence of a verbalizing head in verbal adjectives formed on the basis of the
latter type of roots serves as a context determining their interpretation
(similarly to prefixes with Latinate roots in English and Ancient Greek
roots in Greek), as expected under Arad’s hypothesis. More specifically, we
propose that roots fall into basic ontological types naming events, things, and
states following Harley (2005). Roots must be classifiable in terms of this
basic ontology (see also Alexiadou, this volume, Levinson, this volume,
Rappaport Hovav, this volume and Roßdeutscher, this volume; contra
Acedo-Matellán and Mateu, this volume, who argue against this). If they
are not, a categorizing head serves to classify them (cf. Borer this volume for
a discussion of categorization of roots). The classification itself makes them
acquire what is sensed as fixed “meaning”, which is then retained throughout
the derivation.
(iii) If we look at idiomatic interpretations, as suggested by Marantz on the basis
of idiomatic interpretations of adjectival passive participles vs. compositional
interpretations of verbal passive participles in English, then (1) is not supported by the Greek data. Verbal adjectives involving affix attachment below
v (whether v is taken to be a verbalizer or an eventivizer) as well as participles
showing affix attachment above v may both have idiomatic meanings, and
there is no necessary correlation between the meaning of verbs and the
meaning of participles which include a verbal head, contra Marantz (2001,
2007b). On the other hand, there is strong evidence from various directions
that the head which delimits the domain for idiomatic interpretations of
adjectival participles and verbal adjectives in Greek is Voice, the little v head
introducing the agent, as was proposed in earlier work by Marantz (1996,
1997). We conclude that the node that syntactically projects an agent serves
as a boundary for special meanings of idioms at the word level, similarly to
what has been claimed for idioms at the phrasal level.
The chapter is organized as follows. In section 5.2, we provide the necessary
background on inner vs. outer affixation focusing on how this distinction has been
related to the question of idiosyncratic interpretation. In section 5.3, we present an
Domains within words and their meanings: A case study
83
outline of the properties of Greek -menos participles and -tos verbal adjectives,
investigating their architecture in connection to their syntax, semantics and morphology. We present evidence that morphological verbalizers should be dissociated
from the abstract heads contributing eventiveness. This leads us to conclude that
the categorizing head v (vC), which verbalizes an acategorial root, cannot be
collapsed with the eventivizing v head (vE), which functions as the BECOME/
RESULT/FIENT operator identified in recent decomposition analyses. Finally, in
section 5.4 we turn to the question of domains for idioms. We argue that neither
the verbalizing head vC nor the eventivizing head vE serve as boundaries for
idioms; the relevant domain is defined by Voice, the head projecting an agent.
5.2 Background: Two domains for word formation, inner
and outer affixation
As is well known, there are two types of word formation, one forming words
showing irregularities, such as paradigmatic gaps, non-predictable meaning, irregular forms, and one for morphologically productive, semantically transparent and
morpho-phonologically predictable forms. In many theories, this difference has been
linked to the hypothesis that there are two places for word formation: lexicon vs.
syntax (e.g., Wasow 1977, Dubinsky and Simango 1996, Horvath and Siloni 2008),
derivation vs. inflection (e.g., Anderson 1982, Perlmutter 1988, Spencer 1991), level I
vs. level II (e.g., Siegel 1979, Kiparsky 1982b, Mohanan 1986). Adjectival/stative vs.
verbal/eventive passive participles in languages like English have been widely
assumed in the literature to present a paradigmatic case exemplifying the double
nature of word formation. Stative/adjectival participles may show special morphology, while eventive/passive participles always show regular morphology:
(2)
a. The shaven man
b. The man was shaved by John.
Moreover, stative participle formation is associated with idiosyncrasy in meaning,
unlike verbal participle formation:
(3)
a. The hung jury
(#Someone hung the jury.)
b. *The jury was being hung.
In English and other languages the stative vs. eventive participle distinction correlates with a split between derivation and inflection (Wasow 1977, Marantz 2001,
2007a, Horvath and Siloni 2008). Stative participles are adjectives and passive ones
are verbs, as shown by very vs. very much-modification in (4) and (5). In English,
very modifies exclusively adjectives and very much exclusively verbs, as indicated by
84
Elena Anagnostopoulou and Yota Samioti
the contrasts between (4b) vs. (4c) and (5b) vs. (5c), respectively. The fact that the
stative participle respected is modified by very in (4a) presents evidence that it
qualifies as an adjective. On the other hand, passive respected in (5a) is modified
by very much qualifying as a verb.
(4)
a. Your family was very respected.
b. John is very fond of your family.
c. *John is very much fond of your family.
Adjectival
(5)
a. Your family was very much respected by the neighbors. Verbal
b. John very much respects your family.
c. *John very respects your family.
Wasow (1977) argued that the properties distinguishing adjectival from verbal
participles may receive a principled explanation if the two participle types are
formed in two different components of the grammar (an approach widely adopted
since then;2 see Horvath and Siloni 2008 for a recent analysis along these lines). On
this view, adjectival passives are formed in the lexicon, and they show idiosyncratic
forms and meanings due to the fact that words in the lexicon have special listed
properties. On the other hand, verbal passives are formed in syntax, the locus of
regularity, productivity, and compositionality, which in turn explains their transparency in form and meaning.
For syntactic approaches to morphology such as Distributed Morphology (DM)
or Borer (2005), properties traditionally attributed to lexical listing (e.g., categorization, argument structure) are reduced to mechanisms of the computational system.
In such models, there can be no well-defined distinction between lexical and
syntactic word formation. This raises the question of how the distinction between
regular and irregular word formation can be derived. In order to capture this
distinction, Marantz (2001, 2007a) and Arad (2003, 2005) propose to reconstruct
the ‘two places to build words” approach within the syntax, as ‘two domains for
word formation’ (see also Embick 2010). More specifically, they propose that there
are two domains for word formation delimited by a category-defining head, as
shown in (6). Attachment of x directly to the root, as in (6a), leads to irregular
word formation, while attachment above the category defining heads (little v, a, n),
as in (6b), leads to regular word formation:
2
But cf. Strong Lexicalist approaches such as Levin and Rappaport’s (1986), who argue that all
participles are formed in the lexicon, adjectival participles are formed by verbal ones via a category
changing rule, and the properties distinguishing adjectival from verbal participles derive from their
difference in category. See Dubinsky and Simango (1996) for evidence against this view and Marantz
(2007a) for discussion of Wasow’s analysis, as compared to Levin and Rappaport’s.
Domains within words and their meanings: A case study
(6)
a.
x
Root
b.
85
x
x
n, v, a
Root
n, v, a
On this view, there is a split between “inner” and “outer” morphology, as defined
in (7):
(7)
“Inner morphology” attaches to roots or complex constituents below the first
little x (x={v,n,a}) node head (phase head) above the root. All morphology
above the first x node is “outer morphology” including all “category changing”
derivational morphology. (Marantz 2007a: 5; Marantz’s (3), (2))
The main properties associated with inner vs. outer affixation are summarized in
Table 5.1:
Table 5.1. Inner vs. outer affixation
Inner Affixation
Outer Affixation
Regularity
Potential special form and
special meaning
Predictable form and
predictable meaning
Selection
Attaches inside morphology
determining lexical category
May attach outside morphology
determining lexical category
According to Marantz and Arad, phasehood provides the key towards understanding
why there is an inner vs. outer split in word-formation. Assuming that category
defining heads are phase heads in the sense of Chomsky (2000, 2001), they define
cyclic domains leading to the semantic interpretation and phonological spell-out of
the chunk of syntactic structure which includes the root or a more complex
constituent plus the category determining head. Once such phase heads are merged,
the structure is shipped off to PF and LF for pronunciation/interpretation and from
that point on, spell-out and interpretation cannot be altered. Moreover, further
affixation cannot reach into properties of the Root. The empirical generalization
this proposal is based on will be called here “the (non-)compositionality generalization” in (8):
(8)
The (non-)compositionality generalization
When affixes attach directly to the root, idiosyncratic meanings may arise.
When affixes attach outside category defining heads, the result is a meaning
predictable from the meaning of the stem.
Arad (2003, 2005) presents extensive evidence from Hebrew denominal verbs supporting (8) and the “inner” vs. “outer” architecture.
86
Elena Anagnostopoulou and Yota Samioti
5.3 Greek -menos participles and -tos verbal adjectives:
Outer and inner architecture
Greek has a rich set of adjectives/participles based on verbs, ending in -tos and
-menos.3 As has been argued for in Anagnostopoulou (2003) and Alexiadou and
Anagnostopoulou (A&A 2008), -menos participles denote states resulting from prior
events and are formed by outer cycle attachment (above vP or above VoiceP; see
Doron, this volume for an extensive discussion of Voice in Hebrew participles). On
the other hand, -tos verbal adjectives lack event implications, a property taken by
Anagnostopoulou (2003) and A&A (2008) to provide evidence for inner-cycle
attachment. If this picture is correct, then -menos and -tos forms present an ideal
empirical domain for testing the validity of the Marantz/Arad hypothesis (1), and the
non-compositionality generalization in (8). Assuming (1) and (8), -menos participles
are expected to have a meaning predictable from the meaning of the corresponding
verbs, while -tos forms are expected to be highly idiosyncratic.
However, as will be shown in this section, a closer investigation of the morphology
of -tos forms in connection to their syntax and semantics reveals that this simple
picture needs to be revised. It turns out that the architecture of -tos verbal adjectives
is considerably more complex than described by Anagnostopoulou (2003) and A&A
(2008), for two main reasons.
First, it is actually not easy to decide on the basis of morphology whether -tos
verbal adjectives involve inner or outer cycle attachment: sometimes -tos attaches
directly to the root and sometimes outside a verbalizer, with absolutely no difference
in the interpretation and syntax of the two cases. Importantly, -tos verbal adjectives
uniformly lack event implications and the concomitant syntactic differences associated with them, regardless of whether -tos attaches directly to the root or to the
root+verbalizer complex.
Second, -tos attachment directly to the root provides some evidence for Arad’s
hypothesis based on Multiple Contextualized Meaning (MCM). If the meaning of
the root is completely unspecified, the presence of a verbalizer fixes it first, and then
-tos attaches to the Root+verbalizer complex. But idiomaticity is a different issue, and
the domain for idiomatic interpretation of verbal adjectives/participles turns out to be
bigger than the domain defined by the first categorizing or eventivizing head: idiomatic meanings may arise at any point in the extended vP domain, provided that the
properties associated with an agent are absent.
The discussion is organized as follows. In section 5.3.1, we summarize from the
previous literature the main evidence in favor of postulating a more complex
3
Nakas 1978, Setatos 1984, Lascaratou 1991, Lascaratou and Philippaki-Warburton 1984, AnastasiadiSimeonidi 1994, Markantonatou et al. 1996, Kordoni 2002, Anagnostopoulou 2003, Alexiadou and Anagnostopoulou 2008, Samioti 2008, in progress.
Domains within words and their meanings: A case study
87
architecture for -menos participles than for -tos forms. In section 5.3.2, we concentrate on the properties of -tos verbal adjectives which lead to a revision of the
Anagnostopoulou (2003)/A&A (2008) picture in two respects: (a) There are classes
of -tos verbal adjectives which express negation and possibility/ability (the latter
implicating an agent). These involve attachment above vP/VoiceP, similarly to
-menos participles. (b) A class of -tos verbal adjectives involves attachment below
the little v head contributing eventiveness but not necessarily below verbalizers. It
thus follows that morphological verbalizers cannot be equated with abstract eventivizing (little v) heads. In section 5.3.3, we address the generalizations governing (i)
the distribution of -menos vs. -tos in participles/verbal adjectives and (ii) the presence vs. absence of verbalizers in Greek verbal adjectives. We argue that the
morphological piece -menos is always associated with the Perfect of Result (Kratzer
1994, 2001), i.e., it is present with participles expressing that the subject is in the
target or result state of a prior event. On the other hand, -tos is present in a
heterogeneous set of cases all having in common that they uniformly lack event
implications. We furthermore propose that verbalizers in Greek verbal adjectives are
required on roots that do not fall under the basic ontological type “event”, following
Harley’s (2005) root typology. Having revised and refined the Anagnostopoulou
(2003)/A&A (2008) picture of the -menos vs. -tos distinction, we finally proceed to
the question of idiomaticity. In section 5.4, we test (1)/(8) by comparing the interpretations of the several types of -tos verbal adjectives to those of -menos participles
and of the corresponding verbs. We conclude that agentivity and not eventiveness is
the key property blocking idiomatic readings from participles, in accordance with
Marantz (1996, 1997) and contra Marantz (2001, 2007a).
5.3.1 The simple A (2003)/A&A (2008) picture: -menos
is an outer- and -tos an inner-affix
Next to adjectives, Greek has two adjectival constructions: the participle in -menos
and what traditional grammars call the “verbal adjective” in -tos:
(9)
anig-men-os
open-men-masc.sg.nom
anix-t-os
‘opened’/‘open’
open-t-masc.sg.nom
They have a similar function to adjectives, i.e., they appear in attributive and
predicative positions:4
(10)
a. to
anih-t-o
the-neut.sg.nom open-t-neut.sg.nom
‘the open window’
parathiro
window
4
Passives are synthetic in Greek and therefore there is never an ambiguity between verbal and
adjectival passives in Greek, unlike English.
88
Elena Anagnostopoulou and Yota Samioti
b. to
anig-men-o
parathiro
the-neut.sg.nom open-men-neut.sg.nom window
‘the opened window’
(11)
a. to parathiro ine anih-t-o
the window is
open-t-neut.sg.nom
‘The window is open’
b. to parathiro ine anig-men-o
the window is
open-men-neut.sg.nom
‘The window is opened’
Participles and verbal adjectives show a number of semantic and syntactic differences discussed in Markantonatou et al. (1996), Georgala (2001), Kordoni (2002),
Anagnostopoulou (2003), A&A (2008), among others. The most fundamental one is
that -menos participles denote a state resulting from a prior event, while -tos verbal
adjectives lack event implications. They denote what has been called “a characteristic
state” by Markantonatou et al. (1996). Consider the examples in (12):
(12)
a. I patates
ine tiganis-men-es.
the potatoes are fry-men-fem.pl.nom
‘The potatoes are fried.’
b. I patates
ine tigani-t-es.
the potatoes are fry-t-fem.pl.nom
‘The potatoes are fried.’
(12a) conveys the meaning that the potatoes are fried as a result of a frying event:
they have been fried, e.g., a minute ago, and are now ready to be eaten. On the other
hand, (12b) simply expresses the fact that the potatoes are cooked in a particular way
(“characteristic state” interpretation): they are fried (rather than, e.g., cooked). A
context bringing out this difference between the two adjectival forms is (13), based on
Embick (2004a: 357). As pointed out by Embick, the complement of verbs of creation
cannot be a state resulting from a prior event because this would be a contradiction.
And as can be seen in (13), -tos forms expressing characteristic states are licit in such
a context, while -menos participles are not, due to their eventive subcomponent:
(13)
a. I porta
chtistike
anix-t-i/*anig-men-i.
the door-Nom built-NAct,3sg open/opened
‘The door was built open/*opened.’
b. Magirepsa to kotopoulo vras-t-o/*vras-men-o.
cooked-I the chicken boiled/boiled
‘I cooked the chicken boiled.’
Domains within words and their meanings: A case study
89
c. Eftiaksa tis patates
tigani-t-es/*tiganis-men-es.
made-I the potatoes fried/fried
‘I made the potatoes fried.’
Crucially, the difference between the two forms with respect to event implications is
associated with a number of syntactic differences. -menos participles can be modified
by manner adverbs (14a) and can license by-phrases and instrument PPs (15a), while
-tos verbal adjectives cannot (14b, 15b):
(14)
a. Afto to vivlio ine kala gra-men-o.
this the book is
well written
‘This book is well-written.’
b. *Afto to kimeno ine kala grap-t-o.
this the text
is well written
(15)
a. O tixos ine xtis-men-os me mistri/apo ton ergati.
the wall is
built
with trowel/by the worker
‘The wall is built with a trowel/by the worker.’
b. *O tixos ine xtis-t-os me mistri/apo ton ergati.
the wall is built
with trowel/by the worker
Anagnostopoulou (2003), following Kratzer (2001), furthermore points out that
-menos participles can denote target and resultant states (Parsons 1990: 234–235).
The former describe states that are in principle reversible; the latter introduce states
that hold forever after the event that brings them about. Target state participles in
(16) are compatible with the adverbial akoma ‘still’, while resultant state participles
in (17) are incompatible with it:
(16)
Ta
the
pedhia
children
(17)
*To theorima
the theorem
ine
are
akoma
still
krimena.
hidden
ine akoma apodedigmeno.
is
still
proven
Target and resultant state -menos participles are not only semantically but also
syntactically distinct. Target state -menos participles cannot license agent and
instrument PPs and agentive adverbials. As shown by (18), by-phrases and instrument phrases are incompatible with akoma ‘still’:
(18)
a. Ta lastixa ine (*akoma) fuskomena
the tires
are (still)
inflated
‘The tires are still inflated by Mary.’
apo tin
by
the
Maria.
Mary
b. Ta lastixa ine (*akoma) fuskomena me tin tromba.
the tires
are (still)
inflated
with the pump
‘The tires are still inflated with the pump’
90
Elena Anagnostopoulou and Yota Samioti
Moreover, note that there are two types of manner adverbials, i.e., (a) manner
adverbs that are result-oriented in that they modify the visible result of an event such
as ‘sloppily’, ‘well’ and (b) manner adverbs that modify the initiator of the action
such as ‘carefully’, ‘deliberately’ (agent-oriented). As shown below, agent-oriented
modifiers are incompatible with akoma (19), while adverbs denoting the visible
result (result-oriented) are compatible with it (20):
(19)
To thisavrofilakio itan (*akoma)
the safe
was (still)
‘The safe was still cautiously opened.’
prosektika
cautiously
(20)
Ta malia mu ine (akoma) atsala
the hair my is
still
sloppily
‘My hair is still sloppily combed.’
anigmeno.
opened
xtenismena.
combed
On the basis of these considerations, A&A (2008) proposed that Greek has three
types of participles with three different structures, depending on the height of
attachment of the participle morpheme. According to A&A (2008), a layer Asp
(=stativizer) attaches to the root in -tos participles, to vP in -menos target state
participles and to VoiceP in -menos resultant state participles:
(21) [AspP Asp X]
(where X= Root, vP or VoiceP)
-tos verbal adjectives lack event implications (they are licensed after verbs of creation
in (13), they do not permit result-oriented modification) and agentivity (no agentoriented modification, no by-phrases and instruments). A&A (2008) take this to
mean that they involve root-attachment:
(22)
root attachment of Asp
ASP
√ANIG
open
Asp
-t-
-menos targets state participles which include the implication of an event (they are
not licensed after verbs of creation in (13), they license result oriented modifiers) but
lack agentivity (no agent-oriented modification, no by-phrases and instruments)
involve v attachment (v is taken by A&A 2008 to be the eventivizing head):
(23)
v attachment of Asp
Asp
v
√ANIG
open
men
Domains within words and their meanings: A case study
91
Finally, -menos resultant state participles which include both implication of an event
and agentivity (agent-oriented modification, licensing of by-phrases and instruments) involve Voice attachment (Voice is taken to introduce the external argument):
(24)
Voice attachment of Asp
Asp
Voice
men
v
√ANIG
open
v
Important for present purposes is the proposal that -tos verbal adjectives instantiate “inner-cycle” attachment while -menos participles “outer-cycle” attachment:
(25)
root-cycle
-tos
root
outer-cycle attachment
-menos functional head
… root…
The distinction in (25) leads us to expect that -tos forms will show properties of inner
affixation and -menos forms properties of outer affixation from the perspective of
hypotheses (1)/(8). Prima facie evidence in favor of the “inner” vs. “outer” division in
(25) comes from the observation that there is a striking difference in the productivity
of -menos as opposed to -tos participles. While all verbs of the appropriate semantic
type, i.e., all telic, many atelic, and even some (coerced?) statives can form -menos
participles, there are many gaps in the formation of -tos participles. Some examples
illustrating this are listed in (26):5
(26)
a.
b.
c.
d.
e.
vizag-menos *vizax-tos
rimag-menos *rimax-tos
doule-menos *doulef-tos
louz-menos
*lous-tos
pirag-menos *pirax-tos
‘nursed’
‘ruined’
‘worked out’
‘shampoo bathed’
‘hurt, tampered with’
5
Note that out of 2,722 verbs listed in the online Anastasiadi-Simeonidi reverse index (< http://www.
Komvos.edu.gr/dictionaries/dictOnLine/DictOnLineRev.htm>) that have been checked so far, 688 verbs
form both -menos and -tos participles, 1,866 form only -menos participles and 168 verbs only -tos
participles. These numbers need to be checked more carefully.
92
Elena Anagnostopoulou and Yota Samioti
5.3.2 A more complex picture: Different types of -tos participles
A closer look into the properties of -tos verbal adjectives reveals that the Anagnostopoulou (2003)/A&A (2008) classification of participles in (25) needs to be refined.
As it turns out, there are some clear cases of -tos involving outer cycle attachment
and some more controversial ones:
(I) A clear case of outer cycle attachment is discussed in Samioti (2008, in
progress). As pointed out by Markantonatou et al. (1996), there is also a class of
-tos participles denoting ability/possibility rather than characteristic state:
(27)
a. Afti i dikaiologia ine pistef-t-i.
this the excuse
is
believe-t-fem.sg.nom
‘This excuse can be believed/is believable.’
b. Afto to asteri ine ora-t-o.
this the star is see-t-neut.sg.nom
‘This star can be seen/is visible.’
Other participles of this type are katortho-t-os ‘achievable’, bore-t-os ‘doable’, epitefkt-os ‘doable’, antilip-t-os ‘perceivable’, aisthi-t-os ‘perceivable’, ap-t-os ‘touchable’,
thea-t-os ‘visible’, ia-t-os ‘curable’, fori-t-os ‘transportable’, noi-t-os ‘conceivable,
thinkable’, katanoi-t-os ‘understandable’, anek-t-os ‘tolerable’, ipofer-t-os ‘tolerable’
etc. Samioti (2008, in progress) argues that -tos in ability participles attaches above v.
Evidence for this is drawn from the fact that they license agent (28a) and instrument
PPs (28b) and agent-oriented adverbs of the type found in English middles (29):
(28)
a. I istoria tou ine pistef-t-i apo olous.
the story his is believable by everyone
‘His story can be believed by everyone.’
b. To vouno
ine ora-t-o me kialia.
the mountain is visible with binoculars
‘The mountain is visible with binoculars.’
(29)
To mathima ine efkola katanoi-t-o.
the lesson
is
easily understandable
‘The lesson can be easily understood.’
Samioti argues that ability -tos participles pattern syntactically with Greek dispositional middles as these have been described by Lekakou (2005). According to
Lekakou, the (dispositional) middle Voice head in Greek is a variant of the passive
Voice, unlike English and other languages. An implicit external argument is present
and, therefore, by-phrases and instruments are licit (see Lekakou 2005 for detailed
argumentation). Adopting this analysis for ability/possibility participles, Samioti
proposes that -tos attaches to the (middle) VoiceP (see Samioti 2008, in progress,
for details).
Domains within words and their meanings: A case study
93
(II) A related, though less straightforward, case of high attachment is illustrated in
(30). Greek has a productive process yielding ability/possibility -tos participles which
consist of the verbal root and the adjectival prefix aksio- ‘worth-’:
(30) a.
b.
c.
d.
aksi-agapi-tos
aksi-o-thavmas-tos
aksi-o-meleti-tos
aksi-o-katafroni-tos
worth loving
worth admiring
worth studying
worth despising
Semantically, such formations clearly implicate an agent (a generic/impersonal agent
corresponding to English one), pointing to the presence of Voice. Syntactically, they
differ from the participles in (27)–(29) in that they do not license adverbs and
instruments, and agentive PPs are marginally tolerated when they express universal
quantification:
(31)
a. To arthro ine aksiomeletito (*prosektika).
the article is
worth-studying carefully
‘The article is worth studying carefully.’
b. O planitis ine aksiomeletitos (*me to tileskopio).
the planet is worth-studying with the telescope
‘The planet is worth studying with the telescope.’
c. To arthro ine aksiomeletito (?apo olus/ ?apo ton kathena
the article is
worth-studying by all/by the everyone/
?apo ton opjondipote).
by the anyone
‘The article is worth being studied by all/everyone/by anyone.’
Morphologically, these participles involve compounding, as evidenced by the fact
that they systematically employ the compound marker -o- indicated in (30) (see Ralli
2001, 2006 for discussion of -o- in Greek).6 We tentatively suggest that they indeed
contain Voice (perhaps the type of “middle” Voice also contained in ability/possibility
participles), while the restrictions in (31) reduce to restrictions on compounding.7
6
Some examples of compounds displaying this marker are the following:
(i) a.
b.
c.
d.
7
xart-o-peksia
trapez-o-ipallilos
maxer-o-piruno
Afstr-o-Ungaria
‘card playing’
‘bank employee’
‘knife and fork’
‘Austria-Hungary’
The issue of compounding deserves further investigation. Samioti (in progress) compares the
properties of aksio- compounds to -tos participles involving compounding with the manner adverbs
efkola- ‘easily’ and dhiskola ‘with difficulty’. These are formations like efkol-o-diavas-tos lit. ‘easy-read’ -tos
‘something that can be easily read’ and diskol-o-diavas-tos lit. ‘difficult-read’ -tos ‘something hard to read’.
Samioti points out the following: (i) All three types of compounds disallow instrument PPs. (ii) In efkoloand dhiskolo-compounds, the adverbs efkola- ‘easily’ and dhiskola ‘with difficulty’ are manner adverbs, a
fact which can potentially explain why further manner adverbs cannot be licensed with them. Note,
94
Elena Anagnostopoulou and Yota Samioti
(III) Negated participles instantiate a third case where, arguably and more controversially, -tos participles present an instance of outer cycle attachment. Negated
participles obligatorily surface with -tos in Greek, as shown by the fact that the
negated -menos forms in (35) are ungrammatical. Note that, even in cases when there
is no corresponding non-negated -tos form available, such as (33) and (34), the
presence of -tos instead of -menos is triggered by negation:
(32)
a. gra-men-os
written
b. a-graf-t-os
un-written
(grap-t-os)
(33)
a. pli-men-os
washed
b. a-pli-t-os
un-washed
(*pli-tos)
(34)
a. diavas-menos
read
b. a-diavas-t-os
un-read
(*diavas-tos)
(35)
a. *a-diavas-menos
un-read
b. *a-pli-men-os
un-washed
At first sight, negated participles behave as if they involve attachment to the root.
They do not imply a prior event, and this correlates with the lack of the key syntactic
properties associated with a vP/VoiceP architecture (see Anagnostopoulou 2003
following Kratzer 1994, and A&A 2008 who adopt a root-attachment analysis
based on the facts in (36)–(38)). More specifically, they resist manner modification,
as illustrated in (36b), and they do not license by-phrases and instruments, as shown
in (37b) and (38b), respectively. In all respects they differ from their non-negated
-menos counterparts in (36a), (37a), and (38a):
(36)
a. I
Maria ine prosektika xtenis-men-i.
the Maria is
carefully
combed
‘Mary is carefully combed.’
manner adverbs
though, that this explanation does not immediately carry over to aksio-compounds, because aksio- ‘worth’
is not a manner but rather an evaluative adverb. (iii) Interestingly, aksio-compounds differ from efkoloand dhiskolo- compounds with respect to the type of agentive PP they license. While aksio-compounds
marginally license PPs introduced by the preposition apo ‘from’, which generally introduces agents in
Greek (see, e.g., Alexiadou and Anagnostopoulou 2004, Alexiadou, Anagnostopoulou, and Schäfer 2006
for discussion and references), efkolo- and dhiskolo- compounds license ‘agentive’ PPs introduced by the
preposition gia ‘for’, as in (i) (we would rather call these PPs “relevance PPs”, as they express the person
for which something is easy or difficult to do):
(i) Afto to fagito ine efkol-o-magirefto gia oles tis nikokires
This the food is easily-cook-to
for all the housewives
‘This food is easy to cook for all housewives’
A similar fact has been noted for some (though not all) dispositional middles in English, namely the
existence of well-formed examples like No Latin text translated easily for Bill (compare though to *These
books don’t sell for the average shopkeeper); see among others Stroik (1992, 1995, 1999) and Ackema and
Schoorlemmer (1995) for discussion.
Domains within words and their meanings: A case study
95
b. *I Maria ine aprosekta
a-xtensi-t-i.
the Maria is sloppily/not carefully un-combed
intended: ‘Mary is sloppily uncombed.’
(37) a. To vivlio ine gra-men-o apo tin Maria.
the book is
written
by the Mary
‘The book is written by Mary.’
by-phrases
b. *To vivlio ine a-graf-t-o apo tin Maria.
the book is unwritten by the Mary
intended: ‘The book is unwritten by Mary.’
(38)
a. To vivlio ine gra-men-o me grafomixani.
the book is written
with typewriter
‘The book is written with a typewriter.’
instruments
b. *To vivlio ine a-graf-t-o me grafomixani.
the book is unwritten with typewriter
intended: ‘The book is unwritten with a typewriter.’
From the point of view of their productivity, morphology, and interpretation,
however, they behave as if they involve outer cycle attachment (cf. Embick
2004a:359, fn. 6 for English negated un-participles). First, negated -tos participles
are fully productive, similarly to -menos partciples and unlike -tos ones. Second, in
cases of morphologically conditioned stem allomorphy, negated -tos participles do
not show the same allomorphy as the corresponding non-negated -tos form:
(39)
a. a-graf-tos
un-written
b. grap-tos
written
c. *a-grap-tos
un-written
Third, they clearly form the negation of -menos participles, in cases where there is no
corresponding -tos form (see (33) and (34) above). And fourth, when all three forms
exist, negated participles negate -menos and not -tos participles. For example, a-graft-os in (32) above means ‘something that has not been yet written’ and not ‘something that is not in a written form’ (see Setatos 1984, Anastasiadi-Simeonidi 1994).
But if negated participles involve the negation of the -menos forms, then this leads to
the postulation of an “outer-cycle” architecture for them (contra Anagnostopoulou
2003 and A&A 2008). Under such an analysis, the reason why -tos and not -menos
surfaces on negated participles might be a morpho-phonological one, related to
stress: the suffix -ménos requires stress while the negation prefix a- triggers obligatory recession of stress, a conflict resolved by the insertion of -tos.
(IV) The final case we would like to discuss is the most interesting one from the
perspective of hypotheses (1)/(8) discussed in this chapter. Note first that Greek
productively employs verbalizing suffixes which have been analyzed as root verbal-
96
Elena Anagnostopoulou and Yota Samioti
izers by Alexiadou (2001, 2009; see Giannakidou and Merchant 1999, Ralli 2001 for
discussion of these elements):
(40) Root-verbalizing elements
Greek: -iz, - on-, -en/an, -ev,- -az, -a Alexiadou (2001, 2009):
(41)
a. aspr-iz-o,
whiten
kathar-iz-o
cleaned
b. pag-on-o,
freeze
ler-on-o
dirty
c. diaplat-en-o, arost-en-o
widen,
become sick
d. sten-ev-o,
tighten
berd-ev-o
confuse
e. diav-az-o,
read
f. pul-a-o,
sell
xal-a-o
destroy
mir-az-o
split, share
As expected by the “outer-analysis” of -menos participles in (25), verbalizers (or their
allomorphs as in the (b), (c), (d), and (f) examples) are systematically present in
-menos participles, providing morphological evidence for outer cycle attachment:
(42) a. aspr-iz-menos,
whitened,
kathar-iz-menos
cleaned
b. pag-o-menos,
frozen,
ler-o-menos
dirtied
c. diaplat-i-menos,
widened,
arost-i-menos
sickened
d. sten-e-menos,
tightened,
berd-e-menos
confused
e. diav-az-menos,
read,
mir-az-menos
split, shared
f. pul-i-menos,
sold,
xal-az-menos
destroyed
Root verbalizers are often disallowed in -tos participles, a fact which was taken by
A&A (2008) as evidence for root attachment on -tos in (25):8
(43) a. *aspr-is-tos,
whitened,
*kathar-is-tos
cleaned
b. *pag-o-tos,
frozen,
*ler-o-tos
dirtied
c. *diaplat-i-tos,
widened,
*arost-i-tos
sickened
8
For some of these examples (the ones based on adjectival roots) there is an alternative well-formed
formation based on the root + the adjectival ending: *kathar-is-tos (cleaned) vs. kathar-os (clean),
*aspr-is-tos (whitened) vs. aspr-os (white), *sten-ef-tos (tightened) vs. sten-os (tight). A&A 2008 analyze
this as an instance of blocking. See below for discussion.
Domains within words and their meanings: A case study
d. *sten-ef-tos,
tightened,
*berd-ef-tos
confused
e. *diav-as-tos,
read,
*mir-as-tos
split, shared
97
But, quite unexpectedly, many Greek characteristic state -tos participles that do
include such verbalizing elements:
(44)
a.
b.
c.
d.
e.
f.
g.
h.
i.
axn-is-tos
koudoun-is-tos
vathoul-o-tos
vid-o-tos
if-an-tos
magir-ef-tos
fit-ef-tos
angali-as-tos
evodi-as-tos
‘steaming hot’ axn-iz-o
‘steam’
‘ringing’
koudoun-iz-o ‘ring (a bell)’
‘hollow’
vathoul-on-o ‘hollow out’
‘screwed’
vid-on-o
‘screw’
‘woven’
if-en-o
‘weave’
‘cooked’
magir-ev-o
‘cook’
‘planted’
fit-ev-o
‘plant’
‘embraced’
angali-az-o
‘embrace’
‘fragrant’
evodi-az-o
‘be fragrant’
Despite the presence of verbalizers, the -tos participles in (44) do not have event
implications (they denote characteristic states), and they do not license manner
modification, agent PPs or instruments:
(45) a. *To fagito
the food
ine kala/prosektika
is well/carefully
magir-ef-t-o.
cooked
b. *To fagito
the food
ine magir-ef-t-o
is cooked
apo tin Maria.
by the Mary
c. *Ta fita
the plants
ine fit-ef-t-a
are planted
me diaforetika ergalia.
with different instruments
The characteristic state participles of the type illustrated in (44) are important to
our discussion, as they show that the abstract little v heads described in the decomposition literature, i.e., the semi-functional heads introducing eventive interpretations (and licensing result-oriented manner modification) must be dissociated from
morphological verbalizers. This poses an immediate problem concerning the proper
understanding of (1) and (8): the question is raised what counts as a phase head in
-menos and -tos participles, the verbalizers in (40), which are also present in -tos
verbal adjectives like (44), or the abstract eventive little v heads licensing modifiers
and PPs, which are consistently absent from characteristic state -tos verbal adjectives, even when they include a morphological verbalizer?
In the next sections, we will argue that morphological verbalizers indeed serve as
contexts for meaning assignment to unspecified roots, in accordance with (1)/(8), as
these generalizations have been interpreted by Arad (2003, 2005) in her work on
Hebrew. But this is a separate issue from the question of idiomatic interpretations,
98
Elena Anagnostopoulou and Yota Samioti
where the domain for unpredictable meaning of participles is considerably larger,
and neither verbalizers nor eventivizers have a special status, but rather Voice.
5.3.3 Revisions and generalizations
The discussion in the previous sections has led us to revise the typology of Greek
verbal adjectives in two respects:
(a) First, -tos does not always attach below little v. There is one clear case where
-tos attaches above little v (ability/possibility -tos, Samioti 2008, in progress)
and two more cases where -tos possibly attaches above little v (aksio-compounding, negated -tos participles). It is an open question at this point whether the
different types of -tos verbal adjectives involve a single morphological piece
-tos inserted in stative environments lacking event implications, with the three
different interpretations resulting from the specific environments of insertion
(a case of polysemy) or whether Greek has three different –tos’s, i.e., ability
-tos, negation -tos, characteristic state -tos (a case of homophony). It is interesting though that Greek is not unique in showing the same morphology for
stativized expressions and for expressions denoting ability/possibility, i.e., the
counterparts of -able adjectives in English. Chichewa (Dubinsky and Simango
1996: 759, ex. (17)) and Malagasy (Travis 2005) are similar in this respect.
(b) Second, the morphological decomposition of -tos characteristic state verbal
adjectives leads to the identification of verbalizing heads in a number of cases,
as was shown by the examples (44). Crucially, as shown by (45), these
morphological verbalizers do not have the syntactico-semantic properties of
the abstract little v head identified and discussed in the decomposition
literature (Embick 2004a; Alexiadou, Anagnostopoulou and Schäfer 2006;
Marantz 2001, 2007a; Harley 2012 and related literature), i.e., they do not
contribute eventiveness nor do they license modifiers and argumental PPs.
Alexiadou (2009) reaches exactly the same conclusion for result nominals in
Greek, which may contain a verbalizer which, however, does not contribute
eventiveness and does not license arguments.
In view of these revisions, what are then the generalizations governing (i) the
distribution of -menos vs. -tos in participles/verbal adjectives and (ii) the presence
vs. absence of verbalizers in Greek verbal adjectives?
Starting from the first issue, the affix -menos signifies the “Perfect of Result”, i.e., it
is present whenever the participle denotes a target or resultant state resulting from a
prior event (Kratzer 1994; 2001). For -tos it is unclear whether it presents a case of
polysemy or homophony. Under the first option, there is a single morphological
piece -tos inserted whenever there is no prior event implied, maybe as an elsewhere
form in the absence of a Perfect of Result operator. Alternatively, there are three
different -tos morphemes, namely ability -tos, characteristic state -tos, and the
Domains within words and their meanings: A case study
99
negated Perfect of Result -tos inserted for morphophonological reasons. We conclude that verbal adjectives with -tos involve either inner- or outer-cycle attachment,
contra A&A (2008), namely:
(i) Ability/possibility -tos forms clearly show outer-cycle attachment, and they
contain a (middle) VoiceP.
(ii) -tos forms involving aksi-o- compounding (e.g., aksi-o-meletitos ‘worthstudying’) possibly also contain a VoiceP and hence are formed by outercycle attachment.
(iii) When a target/resultant state participle is negated there are semantic and
morphological reasons for positing -tos attachment above vP or VoiceP,
though the syntactic evidence is less clear.
(iv) Characteristic state verbal adjectives do not have event implications and lack
all syntactic properties associated with a little v head providing evidence for
inner-cycle attachment. They pose, however, a puzzle. While -tos attaches
directly to the root in a number of cases, it attaches to the Root+verbalizer in
another set of cases. This leads to the second question that needs to be
addressed, namely what is the explanation for the presence vs. absence of
morphological verbalizers in -tos characteristic state participles. In answering
this question, we would like to pursue hypothesis (46):
(46)
Selection hypothesis:
-tos selects expressions naming events
With Harley (2005) we assume that roots fall into basic ontological types naming
events, things and states. We can then formulate the following generalizations
concerning the absence or presence of verbalizers inside -tos:
Generalization I: -tos directly attaches to roots which can be characterized as
Rootevent in terms of Harley’s ontology, as shown in (47). These forms involve
roots with a specified meaning (they express particular types of events) yielding
verbs (as in the first column of 47) by combining with the verbal inflectional ending
either directly (as in 47b,d,e,f,g) or via the formative -n- (in 47a,c).9
(47)
a.
b.
c.
d.
e.
f.
g.
ftiax-n-o ‘make’ ftiax-tos ‘made’
lin-o
‘loosen’ li-tos
‘loose’
per-n-o
‘take’
par-tos ‘taken’
klev-o
‘steal’
klef-tos ‘stolen’
din-o
‘give’
do-tos
‘given’
plek-o
‘knit’
plek-tos ‘knitted’
klin-o
‘close’
klis-tos ‘closed’
9
There are a number of further observations one can make here. First, the absence of -n- in the verbal
adjectives as in (47a) could be viewed as evidence that -tos directly attaches to the root and not to the verb,
if it can be ensured that the reason for the absence of -n- is not morpho-phonological. Second, in (47c,e,g)
-tos attaches to the perfective stem (marked by stem allomorphy in (47c,e) and by the presence of -s- in
(47g)), a fact that could, in principle, receive either a semantic or a phonological explanation. The question
raised by such examples is to what extent it is legitimate to talk in these cases about -tos attachment to
roots rather than stems. We have to leave these questions open for now.
100
Elena Anagnostopoulou and Yota Samioti
In these cases, the selection requirement of -tos stated in (46) is satisfied because -tos
directly combines with roots naming events (and see fn. 9 on the question of whether
root or stem is the appropriate notion).
Generalization II: -tos does not combine with Rootthing. It combines with Rootthing +
verbalizer.10 In a number of cases illustrated in (48) roots have a specified meaning
naming a thing and can be combined directly with nominal inflection forming a
noun (first column in 48). -tos cannot directly combine with such roots. As shown
by the third column in (48), they first become verbal by combining with a verbalizer
(cf. the second column in (48) where such roots are turned into verbs by a verbalizer
plus the verbal inflection), and then -tos attaches to the root + verbalizer complex:
(48)
a.
b.
c.
d.
e.
f.
afr-os
axn-os
vid-a
koudoun-i
koumb-i
skep-i
‘foamN’
‘steamN’
‘screwN’
‘ringN’
‘buttonN’
‘roofN’
afr-iz-o
axn-iz-o
vid-on-o
koudoun-iz-o
koumb-on-o
skep-az-o
‘foamV’
‘steamV’
‘screwV’
‘ringV’
‘buttonV’
‘coverV’
afr-is-tos
axn-is-tos
vid-o-tos
koudoun-is-tos
koumb-o-tos
skep-as-tos
‘foaming’
‘steaming hot’
‘screwed’
‘ringing’
‘buttoned’
‘covered’
Note that in some cases, such as (49), the denominal verb formed by the Rootthing +
verbalizer is deviant (second column in 49, i.e., such expressions cannot name events
and therefore they cannot be verbs), while the corresponding -tos participle is
perfect:
(49)
kamban-a ‘bellN’ ??kamban-iz-o ‘bellV’
kamban-is-tos ‘sounding like a bell’
We propose that the presence of a verbalizer in formations of this kind is necessary
to satisfy the selection requirements of -tos in (46).
An issue that needs to be discussed at this point concerns the criteria by which the
verbal and adjectival forms in the second and third columns of (48), (49) above are
indeed root-derived (based on Rootthing), as opposed to being derived from zerorelated nouns (e.g., based on Root plus a little n with a zero exponent). A way to
decide this is to apply Kiparsky’s (1982c) criteria. Kiparsky argues that root-based
formations do not entail the existence of the corresponding nouns while nounderived ones do entail the existence of the corresponding nouns (see Arad 2003, 2005
for extensive discussion of Hebrew based on Kiparsky). For many of the forms in
(48), (49) it can indeed be demonstrated that they are root derived. For example,
vidono and vidotos in (48c) do not entail the existence of a screw but rather they refer
to the type of movement required for putting two pieces together. In a similar
manner, koumbono and koumbotos in (48e) do not entail the existence of a button
(they refer to closing something by making use of a particular device), koudounizo
and koudounistos (48d) /kambanistos (49) do not entail the existence of a ring/bell
10
See Alexiadou (2009) for discussion of such formations.
Domains within words and their meanings: A case study
101
(they mean sounding like a ring/bell), skepazo and skepastos (48f ) do not entail a
roof (they mean cover and being covered) and axnizo and axnistos (48b) do not
entail steam (they mean hot and cosy like anything that is steaming hot).11
Generalization III: As shown in (50), -tos does not combine with Rootstate +
verbalizer because there is an adjective blocking the -tos form (as pointed out by
A&A 2008).
(50)
aspr-iz-o ‘whiten’
mavr-iz-o ‘blacken’
kitrin-iz-o ‘yellow’
prasin-iz-o ‘green’
kathar-iz-o ‘clean’
megal-on-o ‘grow’
aspr-iz-men-os
aspr-os/i/o
‘white’ *is-tos
mavr-iz-men-os mavr-os/i/o ‘black’ *is-tos
kitrin-iz-men-os kitrin-os/i/o ‘yellow’ *is-tos
prasin-iz-men-os prasin-os/i/o ‘green’ *is-tos
kathar-iz-men-os kathar-os/i/o ‘clean’ *is-tos
megal-o-men-os megalos
‘big’
*o-tos
In principle, however, the formation Rootstate + verbalizer + tos is possible. And
indeed, such forms do exist, but they have specialized uses. For example, the -tos
form in (51) is used only for food:
(51)
kokin-iz-o ‘redden’
kokin-os/i/o
‘red’
kokin-is-tos ‘with a red sauce’
Overall, we have checked many -tos participles containing the verbalizers -iz-, -az-,
-ar-, -on-, -ev-, and it turns out that the majority of them are as in (48)/(49), i.e., they
are based on Rootthing+ verbalizer. These have corresponding nouns of the form
Rootthing+ nominal inflection. A few -tos participles with a verbalizer are based on
Rootstate (adjective or adverb/preposition), as in (51) and (52) below:
(52)
a.
b.
c.
d.
stogil-ev-tos round-ef-tos
thab-o-tos
misty-o-tos
xor-is-tos
without-is-tos
antam-o-tos together-o-tos
‘round/rounded’
‘misty/blurred’
‘separate’
‘together’
There is a final set of cases, to be discussed immediately, which leads to a particular
interpretation of the hypothesis (1)/(8).
Generalization IV: verbalizers turn undefined roots into Constructevent and then
-tos attaches to the Rootundefined+ verbalizer. A group of -tos participles is based on
a Root with no clear meaning (call it Rootundefined) which combines with a verbalizer.
This type of root has unspecified meaning in the sense that (a) one couldn’t assign
an exact meaning to it and (b) there is no corresponding noun or adjective or verb of
the form Rootundefined + Inflection. Two different subcases fall under this category.
(i) A number of undefined roots represent sounds or movements or shapes (they
are often formed by reduplication). A verbalizer must necessarily attach to them
11
Of course, one would have to check all such formations available in the Greek lexicon in order to
draw firm conclusions concerning this issue. For example, by Kiparsky’s criteria the verbal and adjectival
forms in (48a) seem to qualify as being based on a zero-derived noun.
102
Elena Anagnostopoulou and Yota Samioti
before they become verbs (surfacing with verbal inflectional endings), and then they
enter further derivation (becoming adjectives, as in (53), or nouns):
(53)
a.
b.
c.
d.
e.
f.
g.
h.
kakar-is-tos
tsitsir-is-tos
trekl-is-tos
tourtour-is-tos
gourl-o-tos
koxl-as-tos
xarxal-ef-tos
paspat-ef-tos
‘cackling’
‘sizzling/frizzling’
‘staggering’
‘shivering/shuddering’
‘goggling’
‘bubbling’
‘rummaging’
‘fiddling’
(ii) There is a residue of roots which seem completely and totally undefined before
a verbalizer attaches to them, making them of type “event”. -tos suffixation follows:
(54)
a.
b.
c.
d.
e.
kt-is-tos
sk-is-tos
str-o-tos
lig-is-tos
sik-o-tos
‘built’
‘slit’
‘smooth, regular’
‘bent’
‘raised’
The two types of roots described above, especially the ones in (54), seem to us to be
very close to what Arad (2003, 2005) describes for Hebrew. One couldn’t exactly tell
what they mean. According to Arad (2003, 2005), there are two types of languages: In
Hebrew-type languages (also Georgian, Russian) roots may be assigned numerous
interpretations in different morphophonological environments (MCM). This correlates with the fact that (i) roots are semantically underspecified and (ii) the inventory
of roots in the languages in question is small. On the other hand, in English-type
languages most roots are assigned meaning in one environment only (exception to
this are Latinate bound roots like √fer, √cieve etc.; see fn 12 below). This correlates
with the fact that (i) they are semantically specified and (ii) the inventory of roots is
large. Greek seems to be an English-type language. A large number of Greek roots
are semantically specified.12 However, there are exceptions to this in both English
(the Latinate roots mentioned above; see also fn. 12) and Greek (see Alexiadou and
Anagnostopoulou 2011, 2013 for discussion of the Greek counterparts of words based
on Proto-Indo-European roots, where prefixes drawn from the Ancient Greek
prepositional inventory fix the meaning of unspecified roots). The examples in
(53), (54) constitute a further class of exceptions. In the cases listed in (53) and (54)
lack of clear/stable meaning correlates with the inability to classify the roots according
to Harley’s ontology. Adding a verbalizer to these roots on the one hand makes them
12
As in English, exception to this are roots with a Proto-Indo-European origin like √fer. These form
multiple verbs with very different interpretations depending on the prefix they combine with: dia-fer-o
‘differ’, pro-fer-o ‘pronounce’, ana-fer-o ‘report’, pros-fer-o ‘offer’, ek-fer-o ‘formulate’.
Domains within words and their meanings: A case study
103
classifiable in terms of Harley’s basic ontology and at the same time it fixes their
meaning, which is then retained throughout the derivation. We believe that this
double function of verbalizers in the case at hand provides a key towards understanding why categorizers serve as contexts for meaning assignment to unspecified
roots. Roots must be classifiable in terms of a basic ontology naming events, things,
and states. If they are not, as in (53), (54) or in the cases of the Hebrew type MCM
Arad discusses, a categorizing head serves to classify them. The classification itself
makes them acquire what is sensed as fixed “meaning”, which is then retained
throughout the derivation. We conclude that the Greek facts in (53) and (54) support
an interpretation of (1)/(8) along the lines of Arad (2003, 2005), where the crucial
head is the categorizing head (in the cases we discuss vC) which makes roots
classifiable in terms of Harley’s (2005) basic ontology.
5.3.4 Summary: How many participles?
We have arrived at the following picture of Greek adjectival participles and verbal
adjectives w.r.t. the question of outer vs. inner cycle architecture:
(55)
Greek Participles and Verbal Adjectives
1) Two types of -menos participles:
a) target state -menos participles
b) resultant state -menos participles
outer attachment
outer attachment
2) Four types of -tos participles:
a) ability -tos participles,
b) aksio- compounding -tos participles
c) negation -tos participles
d) characteristic state -tos participles:
(56)
root-cycle = tos + Rootevent
-tos
root
outer attachment
outer attachment (possibly)
outer attachment (possibly)
direct attachment to Rootevent
attachment to Rootthing+ verbalizer
attachment to Rootstate+ verbalizer
attachment to Rootundefined+ verbalizer
outer-cycle attachment = ability-tos
-tos
Voice
VoiceP
… root …
Does not count as outer-cycle = tos +[Rootthing/state/undefined + verbalizer]
-tos
v
root
104
Elena Anagnostopoulou and Yota Samioti
There is one question raised by these participles which will only be mentioned
briefly and will not be answered here, concerning the polysemy vs. homophony issue
mentioned earlier in the chapter. In principle, this question can be asked for both
-menos participles and -tos participles. For -menos participles the question is the
following. Are there two different -menos suffixes one yielding target states and one
yielding resultant states which have different selection requirements and different
syntax (target state -menos necessarily attaching below Voice) or is there one -menos
yielding resultant states or target states depending on the context of insertion? In
view of the close similarity in meaning between resultant and target state participles,
we think that the second option is definitely more plausible: there is only one -menos
suffix yielding states resulting from events; the resultant vs. target state difference
must derive from the context, e.g., the idiosyncratic properties of the root, the size of
the complement (if the complement is VoiceP then the state is necessarily a resultant
state), plus some further properties of the structure. For example, the presence of
adverbs has been noted by Kratzer to provide the component necessary for a target
state to verbal roots otherwise yielding resultant states ((57) is Kratzer’s example
(21)):
(57) a. * Meine Haare waren
my hairs
were
‘My hair was still cut.’
immer noch
still
geschnitten.
cut
b. Meine Haare waren immer noch schlampig geschnitten.
my hairs
were still
sloppily
cut
‘My hair was still cut sloppily.’
For -tos the issue is more involved since it is not clear that -tos has a uniform
semantics in characteristic state participles, negated participles and ability/possibility
participles. Moreover, the cross-linguistic evidence we reviewed does not provide
straightforward support in favor of the one over the other alternative hypothesis. On
the one hand, English distinguishes -ed (in negated and simple state adjectival
participles) from -able (for ability/possibility adjectives) supporting the homophony
hypothesis. On the other hand, Chichewa and Malagasy use the same element for
both types of adjectives, like Greek, providing crosslinguistic support in favor of the
polysemy alternative (the unifying property of all three kinds being the absence of a
prior event).
Turning to the issue of idiosyncracy discussed in this chapter, characteristic state
-tos verbal adjectives are particularly relevant to the hypothesis (1)/(8) for two
reasons:
a) First, they present evidence that the presence of a verbalizer does not, in itself,
contribute eventiveness and the syntactic properties associated with it.
Domains within words and their meanings: A case study
105
b) Second, they present evidence that the presence of a verbalizer fixes the
meaning of unspecified roots which cannot be classified according to Harley’s
typology.13
We are now in a position to turn to the question of idiomaticity. We do so in the
next section, and we refer the reader to Anagnostopoulou and Samioti (2013) for a
more complete discussion.
5.4 A bigger domain for Idioms
5.4.1 Neither vC nor vE are boundaries for idioms
We assume a working definition of idioms along the lines of Svenonius (2005:1):
“. . . listed syntactic structures which [ . . . ] have unpredictable meanings in the way
words do, but consist of more than one ‘piece’.” In multi-word idioms, “piece” is the
phonological word. In single-word idioms “piece” is the head (Roots, affixes). Here
we are looking at single word idioms. In view of the dissociation between verbalizing
heads (categorizing heads, vC) and little v (the head contributing eventiveness,
licensing arguments etc., vE) introduced in the previous discussion, the Marantz/
Arad hypothesis as a hypothesis about idioms can, in principle, be tested in two
ways:
a) Taking literally the proposal that once a root is categorized it is assigned a
range of meanings fixed for the rest of the derivation, what should be tested is
whether the presence of a verbalizer in -tos participles yields such an effect. In
other words, are there significant differences between -tos verbal adjectives
with a verbalizer and -tos adjectives without a verbalizer with respect to
idiomaticity, in comparison to the corresponding verbs?
b) Assuming, alternatively, that it is the presence of eventive v which defines a
cyclic domain for interpretation (phase), it should be tested whether the
presence of an eventive little v fixes meaning in a way that affixes attached
above it always lead to predictable interpretations (outer affixation). In other
words, are there significant differences between -tos and -menos forms and the
corresponding verbs with respect to idiomaticity?
The answer to the first question is negative. The internal composition of -tos
verbal adjectives denoting characteristic states does not correlate in any way with
(non-)idiomaticity. Characteristic state verbal adjectives of both types (simplex
without a verbalizer or complex with a verbalizer) can have idiomatic readings
lacking from the corresponding verbs. This is illustrated in (58) and (59) below:
13
The cases where aspectual prefixes determine the meaning of verbal (i.e., event denoting) roots, as in
English with words based on Latinate roots, in Russian (Arad 2003: 775, 2005) and in Greek (mentioned in
footnote 12) should also be taken into account and might lead to a partial modification of this statement.
106
Elena Anagnostopoulou and Yota Samioti
Characteristic state -tos verbal adjectives showing direct attachment of -tos to Rootevent:
Verb
Participle
(58) a. sfing-o
tighten
sfix-tos
tight
Idiomatic interpretation
of participle only
‘careful with money’
b. ftin-o
spit
ftis-tos
lit. spitted ‘spitting image’
c. klin-o
close
klis-tos
lit. closed
‘introverted’
Characteristic state -tos verbal adjectives showing attachment of -tos to Root+
verbalizer:
(59)
Verb
Participle
Idiomatic interpretation
of participle only
‘close friend’
a. kol-a-o
glue-1sg
kol-i-tos
lit. glued
b. xtip-a-o
bang, hit, whip
xtip-i-tos
lit. whipped
‘striking’
c. xon-ev-o
digest
xon-ef-tos
no lit. meaning
‘inside the wall’
d. karf-on-o
nail
karf-o-tos
no lit. meaning
‘very fast/direct’
The answer to the second question is again negative. -menos participles may have
idiosyncratic meanings, just like the -tos verbal adjectives in (58) and (59):
(60)
striv-o
twist
stri-menos geros
‘crotchety old man’
strif-to tsigaro
lit. ‘twisted (rolled) cigarette’
Strikingly, the -menos participle in example (60) only has the idiomatic reading
when modifying a [+human] noun, while the verb and the -tos participle can only
have the literal meaning. But this is the reverse of what is expected if (1)/(8) are
understood as applying to idioms and if the relevant phase head is taken to be vE
(the abstract eventivizing head). What would be expected under such an interpretation of (1)/(8) is that the characteristic state -tos participle has the idiomatic reading
and the -menos participle the compositional meaning. It is furthermore expected
that the idiomatic reading of the -menos participle depends on the idiomatic interpretation of the corresponding verb, i.e., exactly the opposite of what we see in (60). As
we also discuss in detail in Anagnostopoulou and Samioti (2013), there is extensive
evidence that there is no necessary correlation between the idiomatic meaning of
Domains within words and their meanings: A case study
107
verbs and the idiomatic meaning of the corresponding -menos participles. We
distinguish between the following three cases:
i.
Idiomatic verb
kathar-iz-o
lit.
‘clean’
idiom. ‘kill’
(61)
Non-idiomatic participle
kathar-iz-menos
only lit. ‘cleaned’
a. Ton katharisan xtes
to vrady.
him cleaned-3sg yesterday the evening
‘They killed him yesterday evening.’
b. ?Aftos o anthropos ine katharismenos.
this the man
is cleaned
‘This man is cleaned.’
ii.
Non-idiomatic verb
trav-a-o
only lit. ‘pull’14
Idiomatic participle
trav-ig-menos
lit. pulled
idiom. ‘far fetched’
Verb
(62) a. O Janis travik-s-e tin porta.
the Janis pulled
the door
‘Janis pulled the door.’
b. *O Janis travik-s-e to epixirima.
the Janis pulled
the argument
intended: ‘Janis pulled the argument.’
Participle
(63) a. ?I porta ine trav-ig-meni.
the door is
pulled
‘Janis is pulled.’ lit. interpretation
b. To epixirima
ine
trav-ig-meno.
the argument
is
pulled
‘The argument is far fetched.’ only idiom. interpretation
14
This verb can have an idiomatic reading as a phrasal idiom (i) either when it combines with the
prefix para (meaning exaggeration) or when combined with clitic-doubled object to skini ‘the rope’:
(i) a. O Janis to paratravikse.
the Janis it para-pulled
‘John went too far.’
b. O Janis to travikse to skini.
the Janis it pulled the rope
‘John went too far.’
Crucially, for present purposes, one does not ‘pull the argument’ or ‘pull the story’; compare (62b) to (63b).
108
iii.
(64)
Elena Anagnostopoulou and Yota Samioti
Idiomatic verb: one meaning Idiomatic participle: another
xon-ev-o
lit. ‘digest’
idiom. ‘like’
xon-e-menos
lit. ‘digested’
idiom. ‘understood’
Verb—Participle literal
(65) a. O Janis
xonep-s-e to fagito.
the Janis digested
the food
‘Janis digested the food.’
b. To fagito ine xone-meno.
the food is digested
‘The food is digested.’
Verb idiomatic—Participle idiomatic
(66) a. O Janis
den xonevi ta mathimatika.
the Janis not digests the math
‘Janis dislikes mathematics.’
b. Ta mathimatika den ine xone-mena.
the mathematics not are digested
‘Math is not understood.’
In conclusion, neither vC nor vE constitute boundaries for idiomatic meanings of
participles in Greek, contrary to what one might expect on the basis of (1)/(8).
5.4.2 Agentive Voice is a boundary for idioms
While idiomatic interpretations are not blocked by vC or vE, agentive features
systematically destroy non-compositional interpretations of -tos and -menos forms.
Specifically:
(a) -tos verbal adjectives denoting ability/possibility which implicate an agent,
never have idiomatic readings. All participles belonging to this class have exactly the
same meanings as the corresponding verbs, and their interpretations are always
predictable. Some of these forms are listed below:
(67)
katortho-t-os ‘achievable’, bore-t-os ‘able’ ‘doable’, epitefk-t-os ‘doable’, antilipt-os ‘perceivable’, aisthi-t-os ‘perceivable’, ap-t-os ‘touchable’, thea-t-os ‘visible’, ia-t-os ‘curable’, fori-t-os ‘transportable’, noi-t-os ‘conceivable, thinkable’,
katanoi-t-os ‘understandable’, anek-t-os ‘tolerable’, ipofer-t-os ‘tolerable’, etc.
(b) Whenever the adjectival prefix aksio- (‘worth’) combines with a idiomatic
participle yielding a modal ability/possibility interpretation which implicates an
implicit agent, the non-compositional meaning is lost, as shown in (68):
Domains within words and their meanings: A case study
(68)
109
a. trav-ig-menos aksi-o-travix-tos
lit. ‘pulled’
only lit.: ‘worth pulling’
idiom. ‘far fetched’
b. stri-menos aksi-o-strif-tos
lit. ‘twisted’ only lit.: ‘worth twisting’
idiom. ‘crotchety’
(c) Agentive adverbs like ‘deliberately’ and agent-oriented manner adverbs like
‘carefully’ systematically block idiomatic interpretations (see Alexiadou, Anagnostopoulou, and Schäfer 2006 among many others for discussion of these properties in
connection to agentive Voice):
(69)
a. trav-ig-menos
prosektika/skopima travigmenos
lit. ‘pulled’
only lit.: ‘carefully/deliberately pulled’
idiom ‘far fetched’
b. stri-menos
lit. ‘twisted’
idiom. ‘crotchety’
prosektika/skopima strimenos
only lit.: ‘carefully/deliberately twisted’
(d) The same effect is triggered by agentive and instrument PPs in (70):
(70)
stri-menos jeros
idiom. ‘crotchety man’
BUT
stri-menos apo kapion/me kati only lit: ‘twisted by someone/with something’
The effects of agentivity illustrated above are consistent with Marantz’s (1996, 1997)
generalization that the syntactic head that projects an agent defines a locality domain
for special meanings:
(71)
boundary for domain of special meaning (Marantz’s 1997 (6))
agent
v
head projecting agent
Predictions: (Marantz’s 1997 (7)):
(72)
a. No idioms with fixed agents
(root in agent position, context for special meaning within the VP)
b. No eventive-passive idioms, but possible non-eventive stative idioms
c. No idioms with causative morpheme and lower agentive verb, but
possible idioms with causative and lower non-agentive verb
110
Elena Anagnostopoulou and Yota Samioti
In Marantz’s (1997) system, the difference between verbal and adjectival participles
with respect to idiomaticity, i.e., prediction (72b), is linked to agentivity; in Marantz
(2001, 2007a), on the other hand, it is linked to the verbal/eventive vs. adjectival/
stative nature of English participles. While the two hypotheses are difficult to tell
apart on the basis of English participles which collapse agentivity with eventiveness
and verb-hood, and lack of agentivity with stativity and adjective-hood, the Greek
data discussed in this chapter show that agentivity and not eventiveness or verbhood is what is crucial. Throughout the chapter and in this section, we have only
looked at adjectival/stative forms with or without verbalizers, event implications and
agentivity. We saw that idiomatic meanings freely occur with -menos participles and
-tos verbal adjectives as long as the relevant morphological constructs do not
implicate an agent. But when agentivity features are present, compositional interpretations are enforced on both -tos and -menos forms preserving the literal meanings of the corresponding verbs. Assuming, furthermore, that agentivity is linked to
the head Voice, this leads to the conclusion that Voice is a boundary for special
meanings within the word.
5.5 Concluding remarks
In this chapter, we discussed the issue of idiosyncratic meaning on the basis of a case
study, namely Greek -menos participles and -tos verbal adjectives denoting characteristic state, negation and ability/possibility. We investigated the Marantz/Arad
hypothesis in (1):
(1)
The Marantz/Arad Hypothesis
Roots are assigned an interpretation in the context of the first category assigning head/phase head merged with them, which is then fixed throughout the
derivation.
and addressed two questions:
.
First, which head counts as a phase head for meaning assignment, the abstract
eventivizing head vE or the verbal categorizing head vC? We provided evidence
in favor of dissociating the two heads based on Greek characteristic state -tos
verbal adjectives which lack event implications but can include verbalizing
heads. We offered some evidence that verbalizing heads lacking event implications indeed serve as contexts for meaning assignment to roots that have
completely unspecified meaning and are therefore not classifiable in terms of
a basic typology “thing”, “event”, “state” (Harley 2005). This is in line with what
Arad (2003, 2005) proposes for Hebrew.
Domains within words and their meanings: A case study
.
111
The second question concerns idiomaticity within words, what the domain is
for special meaning in single word idioms. We argued that this domain is larger
than the first phase head, whether we take the relevant head to be vE or vC. The
boundary for special meanings is Voice, i.e., the head that syntactically projects
an agent, as proposed by Marantz (1996, 1997) and contra Marantz (2001, 2007a).
After we wrote the first draft of this chapter, it came to our attention that Marantz
(2013) revises his 2001/2007 position with respect to idioms and adopts a view very
close to the one argued for in the present chapter. Marantz (2013) distinguishes
between two meaning domains: (a) What he calls “the domain for contextual
allosemy.” This is determined immediately, by the first phase head, a contentful
categorizing head, vE in our terms. Marantz argues that the vC head we identify in
our work on Greek participles (i.e., the verbalizer included in characteristic state -tos
participles) should be viewed as a semantically null head, akin to do in English dosupport, which, being semantically empty, does not block the (meaning assignment)
relationship between the phase head above it (taken to be little a, the adjective
forming head) and the root below it. (b) The domain for idioms. This must be
separated from contextual allosemy, and it has as a boundary the Voice head that
introduces the external argument, in accordance with Marantz (1996, 1997). We
discuss Marantz (2013) in detail in Anagnostopoulou and Samioti (2013; the present
chapter was written before Anagnostopoulou and Samioti 2013), where we focus on
the domains for idioms as opposed to domains for allosemy.
6
The category of roots
HAGIT BORER
6.1 Introduction
Consider, as our starting point, the paradigm in (1):
(1)
a. formable; sortable; faceable; coastable; primable; palatable
b. formal; sortal; facial; coastal; primal; palatal
In (1), the Content of, e.g., coastal is related to the nominal Content of coast, but
coastable is Content-related to the verbal Content of coast. Nor is this a coincidence.
Rather, whenever compositional, -able derivatives are related to the verbal Content,
while compositional -al derivatives are related to the nominal Content. But if coast is
a root, i.e., √coast, then where could this categorial correspondence possibly come
from? Specifically, if the representation for (1a–b) is as in (2), the nominal vs. verbal
source Content correlation would be altogether obscured:1
(2)
a. √COAST – able
b. √COAST – al
Within classical accounts of English word formation, it is typically assumed that,
e.g., -al indeed only merges with nouns, and that coast, in the context of coastal, is
somehow nominalized. This is either because coast is listed as an N to begin with, or
because it is derived from some other listed instantiation of coast, (e.g., the listed
coastV). In the latter case, the derivation must involve a silent N affix, effectively a
phonologically null variant of -ation or -age. Similarly, -able only merges with verbs,
and coast, in the context of coastable, is somehow verbalized, either because it is
listed as such (and thus nominal coast is derived from it), or is rendered verbal by
means of a silent V affix merging with a listed N entry. That type of account, we
note, certainly captures the Content-relatedness of coastV and coastable vs. coastN
1
And where Content is perceived as substantive, conceptual meaning, roughly Frege’s Sinn, and
throughout notated with italic capitals (e.g., BALL).
The category of roots
113
and coastal. However, as is immediately evident, it does so at the cost of postulating
multiple null categorizers with distinct, and otherwise difficult to prove, formal
properties. Typically, we find that categorizing affixes project a specific category
and select a specific category, and cases of phonologically realized homophony,
within the derivational domain, are rather rare. In contrast, phonologically unrealized categorizers (henceforth zero-categorizers) are ambiguous (at the very least)
between N and V, and select (at the very least) either V or N. But why should that
be so?2
The conceptual objection above, as it turns out, is only one of numerous empirical
and theoretical problems otherwise facing the claim that English has zero categorizers, a matter I return to in sections 6.3 and 6.4 below. Before doing so, however, let
us consider an alternative which dispenses with zero categorizers, and which has
independent empirical as well as conceptual advantages.
Note now that categorial selection, as typically conceived, is fundamentally a
matching system, and as such, has a redundancy built into it by definition. The
categorial statements that would be required, for, e.g., coastal (or governmental, for
that matter) as well as for will coast or coasted (or will crystalize/crystalized), would
be as in (3)–(4) and consist of first stating the categorial selection properties of -al
and T (the latter a segment of a verbal Extended Projection), and second, of stating
independently the categorial properties of their potential complements, however
derived (note that by assumption government is already N and crystalize is already V
due to prior operations):
(3)
a. -al projects A
-al must merge with N
b. will projects T, T a segment of {Ex[V]}
{Ex[V]} must merge with V
(4)
√COAST → N (somehow)
√COAST → V (somehow)
[N[govern]ment]
[V[crystal]ize]
(See fn. 3 for a note on the notations here and directly below.)
The alternative to such redundancy is to assume that the categorial properties of
roots emerge in the context of a particular structure, and as a result of that structure.
More concretely, suppose we conceive of the functional vocabulary as a set of
operators which effectively divide the syntactic space, with each formal functor
defining not only its own inherent space, but also that of its complement. By way
of illustration from the visual domain, consider the collection of black markings in
(5a). We can conceive of that collection of dots as projecting the delineated space
defined by (an extension of ) its outside boundaries and as such identical to (5b). In
2
English -ing may project V, N, and A, but only merges with V. Similarly, in English (and universally)
some affixes typically are ambiguous between A and N, but merge with an identical base (e.g., AmericanN,A;
IsraeliN,A; MarxistN,A; vs., e.g., JapaneseA,*N). English -al is, to the best of my knowledge, the only case of
true ‘homophony’ for phonologically realized affixes, merging with N to project A, and merging with V to
project N.
114
Hagit Borer
turn, (5a) can also be said to define a complement space in the most concrete sense of
complement, i.e., that which is delineated within the projecting structure and which
is not black. The delineated white space under consideration crucially has no
properties beyond those defined by the black markings, however otherwise determined. Specifically, that inner space is not a BALL, as such, although an object that
does have the inherent properties of a BALL, such as the BALL in (5c), could, of
course, fit into that space. What could not, however, fit into that space is anything
that has inherently defined properties which are incompatible with those of BALL—
i.e., that space cannot accommodate a NON-BALL, one instantiation of which is,
presumably, the CUBE in (5d).
(5)
a.
b.
c.
d.
Insofar as we can refer to (5c) as a BALL, we can now think about the space
defined by the markings in (5a) as BALL-equivalent. The shape in (5d), however, is
neither BALL nor BALL-equivalent. Crucially, the BALL-equivalent space in (5a) is
not identical to the space defined by the outer shape in (5a), in turn identical to (5b).
Neither, note, could fit into that BALL-equivalent space. By similar logic, while the
complement space (5a) is BALL-equivalent, the complement space defined by (6a) is
TRIANGLE-equivalent, insofar as it would fit the shape in (6b), but not, e.g., the
shapes in (5a–d):
(6) a.
b.
We can now extend this picture to syntactic category labels. Consider, for instance,
the verbal Extended Projection (ExP) which, by assumption, defines its categorial
complement space (CCS) as V-equivalent. We can now view the verbal Extended
Projection as a set ({Ex [V]}) consisting of all functional nodes which define their CCS
as V-equivalent. T, to exemplify, defines its CCS as V-equivalent, and is thus a
segment of the verbal Extended Projection (ExP-segment), or, more specifically,
a member of {Ex [V]}.3 Now, insofar as the CCS of {Ex[V]} is V-equivalent, it
3
A note is in order here about notation. {Ex[V]} stands for the set of (non-categorial) nodes that define
V as their Categorial Complement Space, or, in more familiar terms, the verbal Extended Projection. The
segments of {Ex[V} (its ExP-segments) are, e.g., T, Asp, Modal, etc. Similarly, of course, {Ex[N]} or
{Ex[A]}. On the other hand Categorizers, to spell out as -ation or -ate, are not ExP-segments, but are
(partially) formally equivalent to an Extended Projection as a whole. Such categorizers, or C-functors, are
notated as, e.g., CN[V], where C stands for a categorial functor, N for the categorial value it projects, and
[V] for the categorial value of its CCS. By assumption, C is a semantically bleached functor, unless
otherwise specified, and thus, e.g., CN[V] to spell out as /πation, al, ance/, but ERN[V] to spell out /πer/ or
ABLEV[A], resulting not only in the syntactic structures corresponding, respectively, to CN[V] or CV[A], but
also whatever semantic value is associated with ER or ABLE respectively. Finally, /πxyz/ is a shorthand for
phonological realization. Actual phonological representations are not attempted.
The category of roots
115
will exclude anything which is not V-equivalent, but would allow, equally happily,
both items which already are V (e.g., crystalize) or items which fail to have any
categorial properties altogether, i.e., category-less roots. Similar rationale applies
to the functor to be realized as -alA: it projects A, and defines an N-equivalent
CCS (CA[N]). The statements in (3)–(4) could now be dispensed with, and replaced
with the statements in (7), in turn formally equivalent, within the terminology used
here, to (8):
(7)
a. -al projects A
The (Categorial)
Complement Space
of -al is N-equivalent.
b. will projects T, T an ExP-segment of {Ex[V]}
The (Categorial) Complement Space of {Ex[V]},
is V-equivalent.
(8)
b. T ∈ {Ex[V]}
a. CA[N]
WILLT → /πwill/
CA[N]→ /πal, ic, ous/
Returning to √COAST and √FACE, we may now say that when they merge with
what spells out as /πal/ or /πthe/ respectively, they are N-equivalent, but when they
merge with /πable/ or with /πwill/, they are V-equivalent. Importantly, roots are not
assigned a category as such, nor does a categorial conversion operation of any sort
take place. Rather, they become N-equivalent or V-equivalent because the categorial
space has been divided by a functor, and the ‘space’ into which these roots have
been ‘poured’ so to speak, defines an N- or a V-equivalent space respectively.
Importantly, with the exception of the inherent categorial properties of C-functors,
the system, as a whole, has no inherent categorial properties, but only properties
of equivalence. Thus while we can think of the label projected by e.g., CN[V]-/π-ation/
as N, the category of its CCS is not in actuality V, but rather equivalent to V,
and hence C=V.
Illustrations of categorial space division in distinct contexts are in (9)–(10). Note
that across the board, roots, by assumption inherently category-less, and derived
forms, which by assumption are already categorized by an otherwise existing
C-functor (e.g., ship or ize), merge in identical structures:
(9)
a.
b.
c.
d.
e.
f.
[T
[T
[A
[#/Q
[CL
[A
willT
pstT
ABLEA[V]
MUCH#
DIVCL
CA[N]]
[C=Vπ√coast]
[C=Vπ√coast]
[C=Vπ√coast]
[C=Nπ√coast]
[C=Nπ√coast]
[C=Nπ√coast]
...
...
...
...
...
...
→ will coast
→ coasted
→ coastable
→ (too) much coast
→ coasts
→ coastal
Finally, in Borer (2005a–b, 2013), I argue that ExP-segments consist of pairs in which semantic functions
(e.g., WILL, MUCH, PST) bind an empty variable head. For ease of exposition, and as the matter is by and
large orthogonal for the topic under consideration here, a more traditional execution is opted for in the
text.
116
(10)
Hagit Borer
a.
b.
c.
d.
e.
f.
[T
[T
[A
[#/Q
[CL
[A
willT
pstT
ABLEA[V]
MUCH#
DIVCL
CA[N]]
[C=V
[C=V
[C=V
[C=N
[C=N
[C=N
crystalizeV] . . .
crystalizeV] . . .
crystalizeV] . . .
governmentN] . . .
governmentN] . . .
governmentN] . . .
→
→
→
→
→
→
will crystalize
crystalized
crystalizable
(too) much government
governments
governmental
A point of clarification is in order before proceeding. In Borer (2013) I argue in
some detail that roots are (pure) phonological indices. The notation π√xyz, when
used, is in reference, specifically, to that conception of roots. √XYZ, absent the π
superscript, is in reference to a more general notion of roots, e.g., as in Distributed
Morphology, where opinions on the phonological properties of roots vary. The
significance of phonological faithfulness as preserved through this view of roots
may not be at the core of this particular article, but the central ways in which it
informs the author’s views should become evident enough.
Some consequences of the representations in (9)–(10) are worth pointing out.
Note first that all roots are category-equivalent once they have merged with any
functor. Therefore, although roots have no categorial properties on their own, in
most, if not all structural contexts they do acquire categorial equivalence. As a result,
there is no need to introduce into our syntactic vocabulary category-less terminals.4
A second consequence of significance is that both [C=N π√coast] and [C=V π√coast]
represent a single, non-branching terminal, and crucially, the emergence of a
categorial context for roots has been accomplished without increasing the derivational complexity of these forms relative to some bare root ‘source’ (i.e., making
the categorized form more complex than the root), and without deriving any one
categorial instantiation of the root from another. Similarly, note that constituting a
CCS (e.g., of -al or -ation) does not add any complexity to government or
crystalize, although these are, already, complex forms, representing a previous
operation which merged e.g., the root π√govern with /πment/, the latter a realization
of CN[V].
6.2 Contextual categorization: Evidence and competing accounts
6.2.1 Chomsky (1970)
The model of categorization outlined directly above is very simple and straightforward. In this and the next few sections, I will undertake to compare it to other
categorization models so as to highlight its advantages, as well as to provide further
evidence for it.
4
Roots may merge with roots, to give rise, specifically, to compounds. See Borer (2013) for some
discussion of categorization in those contexts. See also section 6.3 below.
The category of roots
117
It is appropriate to start this discussion by highlighting the difference between the
contextual categorization adopted here and the contextual categorization model put
forth in Chomsky (1970) (and see also Ouhalla 1991, Picallo 1991, and Marantz 1997).
Quite similarly to the account directly above, in Chomsky’s (1970) a category-less
item, possibly a root, is inserted into the syntactic structure, and it is the syntactic
structure that determines its category. Thus direct, by assumption a category-neutral
item, when inserted in a verbal context, gives rise to a verb and the subsequent
‘verbal’ instantiation of the root properties (e.g., obligatory subject, possible direct
object). On the other hand, in a nominal context, direct would give rise to a noun
and the subsequent ‘nominal’ instantiation of its properties (e.g., optional arguments, of-insertion). The account is conceptually akin to the one put forth here,
insofar as in both it is the structural context that is responsible for categorial
properties, and not the inherent, presumably listed properties of a terminal. That
the accounts are nonetheless vastly different, empirically as well as conceptually,
becomes evident when we consider the specific relationship between syntax and
phonological realization and the structural role (or lack thereof ) of functors. Thus
for Chomsky, the nominal realization of, e.g., direct could be /πdirection/ as in (11):
(11)
NP
N
direct
form
form
→
→
→
VP
(of NP)
V
/πdirection/
/π formation/
/π form/
direct
form
NP
→
→
/πdirect/
/πform/
Fundamentally, Chomsky’s approach sets aside morphological complexity as a
non-syntactic issue, and the fact that /πformation/ as well as /πdirection/ are bimorphemic while /π form/ and /πdirect/ are mono-morphemic (prefixing aside) is by
assumption syntactically irrelevant. Nor is it relevant (indeed, even acknowledged)
that insofar as /π form/ is at least morpho-phonologically a subpart of /πformation/,
that subpart corresponds to the verbal, rather than the nominal Content of /πform/.
The notion of what is a basic, non-categorial unit, possibly root, in some sense, is thus
fundamentally divorced from phonological realization or from derivational history.
While it appears rather clear that the notion of lexical item, or possibly root in
Chomsky (1970) must carry some phonological content, morpho-phonological considerations, as such, play no role in its determining its grammatical properties.5
5
To wit, if roots, or listed items, have no phonological content altogether, there is little to ensure that
the verbal instantiation of some root would be phonologically similar, in any way, to its nominal
instantiation. Little, in other words, to exclude the existence of a category-neutral entry that would be
realized as, e.g., recline in a verbal context, but as sofa in the nominal one.
118
Hagit Borer
The perspective advanced here likewise subscribes to the view that roots do not
have a category. In fact, it subscribes, even more explicitly, to the view that roots are
altogether devoid of any syntactic properties. However insofar as Categorial morphology is fundamentally hierarchical, I subscribe to the view that it is per force
syntactic. Equally importantly, I assume that structural complexity tallies with
morpho-phonological complexity and entails phonological faithfulness. As a consequence, /πdirection/ and /πformation/ cannot possibly be phonological instantiations of roots, because they are not mono-morphemic, and rather, represent a
complex hierarchical structure, at the very minimum, that of a (derived) R-nominal
(in the sense of Grimshaw 1990 and Borer 2013), as in (12):
(12)
CN[V]
CN[V]
/π ation/
[C=Vπ √direct]
[C=Vπ √form]
/π direct/
/π form/
Of course, it follows equally directly in both Chomsky’s system and in ours that
/πdirection/ or /π formation/ cannot instantiate a verbal context. However, the
rationale for why that is so is distinct. In Chomsky’s system, /π formation/ is a
phonological spell-out associated specifically with a nominal context, making its
non-occurrence in verbal contexts a straightforward matter. Equally straightforwardly, in the system presented here, /π formation/ is the spell-out associated with
(12), and presupposes the existence of a distinct phonological representation
embedded within it, i.e. /π form/, spelling out the V-equivalent constituent [C=V
π
√form]. Matters, however, do become less straightforward when we consider the
fact that π√form has (at least) two possible nominal instantiations: formation which
is morpho-phonologically complex, and formN, which is not, and where the latter,
but not the former, is homophonous with the verbal instantiation of the same root.
The ramifications of this matter play a considerable role in the discussion in the
remainder of this chapter and in the argument for the contextual notion of categorization outlined here.
6.2.2 Distributed Morphology
Assuming category-less roots and a strict separation between lexical listing and
syntactic projection, Marantz (2000 and subsequent work) and the bulk of work
within Distributed Morphology propose that categories project as separate nodes (n,
v, a—so-called little n, little v, little a). Such categorial nodes are part of the
functional inventory of the language, and are subject to Vocabulary Insertion (VI),
by assumption a post-syntactic procedure which associates a phonological form with
functional terminals. For, e.g., n, such phonological form could be -ation, or -er, or a
The category of roots
119
phonologically null suffix. Importantly, these categorial nodes may also accomplish
additional work. Thus, it won’t do to postulate just a (little a), because clearly the a
that would spell out as /πable/ is distinct from the a that would spell out as /πal/.
Effectively, then, categorial labels are functors, complete with both a category and
with some additional syntactico-semantic properties, and at least at times, with a
unique spell-out. Crucially, roots as such in this system, even once embedded within
the structure, never have a category. To see that this is the case, consider first the
structure in (13). The phonological output for this structure may be either (13i) or
(13ii), which, for Distributed Morphology, would have an identical structure, and
would only differ from each other insofar as for n, VI would insert /ation/ in one
case, but /πØ/ in the other.6 Beyond this, note that in (13), and regardless of the
phonological instantiation of n, √FORM is a (permanently) category-less syntactic
terminal, as the category label is associated exclusively with the functional item n
and with the structure which it projects:
(13)
[n n √FORM]
i. /π form/
ii. /π formation/
Consider now a, and specifically when instantiated as /πal/. It could merge with
the structure in (14), giving rise to (14i) and (14ii):
(14) [a a [n n √FORM]]
i. /π formal/
ii. /π formational/
But presumably, a-/πal/ can also merge with the root √FORM directly, giving rise
to the structure in (15):
(15) [a a √FORM] ]
/π formal/
In (14), a does not merge with a root, but rather with an already existing categorial
structure which is a projection of n, and with n realized as /πØ/. In (15), on the other
hand, it merges with the root. Either way, the root itself is not labeled. Rather, what is
labeled is always only the structure dominating it.
The category-less status of roots within syntactic structures is a rather important
matter for several reasons. We note, first, that contrary to common perception, e.g.,
little n does not assign a category to √FORM, regardless of whether it is null or not.
In (13ii), the constituent n never corresponds to /πform/. Its minimal instantiation is
/πation/ (if phonologically separable) and its maximal instantiation is /πformation/.
Within /πformation/, the part that spells out as /π form/, and which is a distinct
constituent, remains category-less. A very similar logic holds for (13i) in which the
6
-ation and ØN, note, are not in competition, contra the competition model of insertion developed in
Halle and Marantz (1993). Tacitly, then, such analyses introduce a formal distinction between C-functors
and ExP-segments, where competition does hold.
120
Hagit Borer
structure is identical, and where the Ø-nature of the affix is syntactically irrelevant—
what is n is never the root itself, but rather its categorial sister or the branching
constituent that dominates it. Strikingly, then, this is a radically distinct perspective
on the instantiation of roots in syntactic structures than that put forth in Chomsky
(1970), where ‘roots’ as such are not syntactic objects, but where every ‘root,’ in any
syntactic context, is inevitably fully categorized and where category-less terminals
are impossible. Rather, the structures in (13)–(15) amount to the introduction of a
novel syntactic primitive, with syntactic properties yet to be determined. For example, if indeed category-less roots are possible syntactic terminals, do they project? Do
we expect a Rootmax distinct from Rootmin and the sub-phrasal syntax that such
projecting structures might give rise to? Alternatively if roots cannot project and
must, effectively, instantaneously merge with a category label (i.e., they are definitionally Rootmin/max), why should that be?7
No less important, in this context, is the status of the merger of a root with any
ExP-segment. Within Distributed Morphology, this doesn’t seem to happen. Rather,
a root must merge with a categorizer before merging with any functional node (in
the conventional sense). But why should that be? Why can a root merge directly with
v but never with T or D? Are these conditions on the distribution of roots? Are these
conditions on the distribution of functors? If these are conditions on roots, e.g., the
need for roots to instantaneously merge with a category label, what does such a
requirement follow from? If, on the other hand, these are conditions on functors,
e.g., T selects v, D selects n and so on, and then n or v merge with the root, don’t
such conditions amount, effectively, to a surrogate categorization of the root by T or
D, rendering the presence of an additional n or v categorizer superfluous?
When compared with (13)–(15), the account developed here likewise postulates an
N-projecting C-functor in conjunction with π√form, to be spelled out as /πation/.
The structure of formation, in the account developed here, is nonetheless different
from (13) exactly insofar as π√form, as such, is not a syntactic object. The C-functor
eventually to be realized as /πation/, rather, is an instance of CN[V], which is to say
that its CCS is V-equivalent and thus the structure of formation would (minimally)
be on a par with (9f ), where /π form/ instantiates a V-equivalent constituent. In this
respect, we note, CN[V]-/πation/ is no different from {Ex[V]}, the set of nodes that
define the verbal Extended Projection. T, as a segment of {Ex[V]}, defines a
V-equivalent CCS, and thus any root merging with it is V-equivalent, as already
illustrated in the structures in (9a–b).
7
The first approach is pursued by Harley (2009b), who indeed develops a full sub-phrasal syntax for
roots. In turn, however, at least one major problem with assuming that roots are permanently uncategorized syntactic objects is the fact that roots never appear to spell out in the absence of a category, a matter
already noted rather importantly in Marantz (1996). But if uncategorized roots are licit syntactic objects, it
is not clear why they shouldn’t be able to spell out as such.
The category of roots
121
Considerably more crucial, however, are the differing claims concerning the
structure of what spells out as /π form/ in (13). In the contextual categorization
account outlined here, /π form/ is always mono-morphemic, insofar as it spells out
a single, non-branching terminal in all its instantiations. If it is a ‘noun’, it is because
it is embedded in a larger structure that defines it as N-equivalent (e.g., 9d–e). If it is
a ‘verb’, it is because it is embedded in a larger structure that defines a V-equivalent
categorial space (e.g., 9a–c). Thus while /π formation/ spells out a bi-morphemic
structure consisting of a V-equivalent constituent and a C-functor, /π form/ is a
single terminal, crucially with categorial properties that are determined solely within
its larger syntactic context.
The reader may now wonder whether it is, in fact, the case that the non-branching
nature of /πform/ follows from contextual categorizing. After all, what is to prevent
the derivation in (12) from proceeding exactly as it does, with the nominal functor
having a V-CCS (i.e., CN[V]) but with CN[V] itself being phonologically unrealized,
i.e., /πØ/? In other words, what, specifically, would block a zero realization of CN[V]
alongside /πation/, a zero realization of CA[N] alongside /πal/, and a zero realization
of CV[A/N] alongside /πize/, making formN and formV bi-morphemic in such contexts? Certainly not anything that emerges from contextual categorization as such!
Indeed, zero realizations of C-functors are in principle compatible with contextual
categorizing, although, as we note, they are not inevitable. The converse, however, is
not so. Absent contextual categorizing (and absent non-combinatorial re-bracketing
such as conversion), zero realizations of C-functors emerge as the only way to derive
the categorial properties of otherwise categorially unmarked roots. Ipso facto, should
it turn out that zero realizations of C-functors can—and should—be dispensed with,
the existence of contextual categorizing becomes inevitable.
The following two sections are devoted to a detailed argumentation against the
existence of zero realization for English C-functors. Concretely, I will argue that
there are, in English, no zero realizations which are categorially equivalent to -ation
or -ize, or in more traditional terms, no categorial zero affixes. In the absence of zero
realizations for C-functors, the structure in (13) could not be assumed to underlie
what is, eventually, pronounced as (the) form, and insofar as form does have an
instantiation which is N-equivalent, that instantiation would have to be contextually
determined. In turn, and insofar as contextual categorization is inevitable for deriving such nominal instantiations of form, we must doubt the utility of structures such
as those in (13) altogether, or, more generally, the assumption that roots, within
structures, are (permanent) category-less terminals, and that categorization is always
achieved through the presence of a branching structure and a dedicated categorial
node.
122
Hagit Borer
6.3 Against English Zero C-functors: Part I
6.3.1 Some general considerations
Before I turn to empirical issues concerning the existence (or lack thereof) of zero
instantiations for C-functors, it might be worthwhile highlighting the fact that any
model which relies on productive zero realization per force entails a distinction
between morpho-phonological complexity and syntactic complexity.8 The failure of
correlation works in both directions. As the tree in (13) already showed, an equally
complex syntactic tree may give rise to two outputs, one morpho-phonologically
complex (formation) and the other morpho-phonologically simple (formN). Similar
disassociation occurs in the other direction, where a single realization, regardless of
its morpho-phonological complexity, may correspond to varying degrees of morphosyntactic complexity. To the extent that all framed constituents in (16) could
spell out as /πform/, /πform/ may correspond to a single syntactic node, as in (15) or
(16c,d), or to the merger of two syntactic nodes, as in (13) or in (16a,b,e). For that
matter, it is not clear that it cannot, in principle, correspond to three or four, as in
(16h,i). Nor is the problem limited to morpho-phonologically simple constituents.
Formal may be bi-morphemic, as in (16c) or tri-morphemic, as in (16e), etc. In turn,
once morpho-phonological complexity is divorced from (morpho-)syntactic complexity, the result is not only trees of distinct complexity for a simplex morphophonological representation, or varying degrees of morpho-phonological complexity
for a single syntactic tree, but also the wholesale elimination of any empirically
sound and non-circular methods of correlating morphological analysis with structural complexity:9
8
The claim here is specifically about zero realizations of C-functors, i.e., functors which project a
category and which define a CCS, and where, as we shall see, the proliferation of presumed zero
realizations is particularly troubling. While a typological universal along such lines might emerge
from a deeper scrutiny, the author is certainly not in a position to assert its existence. Nor is the
claim made here assumed to carry over, necessarily, to argument structure changing operations, such as,
e.g., break.CAUSEØ, assuming the validity of such representations, or to the English middle construction
(books sell well).
Importantly, the claim explicitly does not carry over to any realizational properties that might be
associated with ExP-segments (e.g., T, D). See Borer (2013) for a detailed motivation of the relevant
syntactic, semantic, and phonological distinction between C-functors and ExP-segments, and for the
explicit endorsement of the view of ‘inflection’ as amorphous, in the sense of Anderson (1992).
9
Specific illustrations of such failures, both within Distributive Morphology and within Kiparsky’s
(1982a, 1997) system will be pointed out as we proceed.
The category of roots
(16)
a. [t PST
[v vØ
√FORM]
]
b. [d DEF
[n nØ
√FORM]
]
c. [v vize
[a aal
√FORM
d. [d DEF
[n nation √FORM
e. [v vize
[a aal
[n nØ √FORM
]
] ]
f. [d DEF
[n nation [v vØ √FORM
]
] ]
g. [d DEF
[n ning
[v vØ √FORM
]
] ]
h. [a aal
[n nØ
[v vØ √FORM
]
] ]
i. [v vØ
[ n nØ
[v vØ √FORM
]
] ]
]
123
]
] ]
(Subscript for realization)
Having noted the conceptual difficulties for zero categorizers, let us turn our
attention to some rather serious empirical problems. In the remainder of this
section, I review some problems for zero categorizers as postulated (differently) in
Lexical Phonology and Morphology and in Distributed Morphology. The specific
explicit arguments put forth by Kiparsky (1982a, 1997) in favor of zero categorizers in
English are tackled in section 6.4.
6.3.2 Zero categorizers: The problem of distributional restrictions
Consider now the phonologically unmarked noun–verb alternations in (17). These
include cases in which one might want to consider the ‘basic’ concept an ‘object’ as
well as cases in which one might want to consider the ‘basic’ concept an ‘action’:
(17)
the
the
the
the
the
the
the
the
the
the
the
salute
form
chair
floor
lamp
dance
kiss
run
walk
feed
show
to
to
to
to
to
to
to
to
to
to
to
salute
form
chair
floor
lamp
dance
kiss
run
walk
feed
show
Certainly, and as (17) illustrates, unmarked noun–verb alternations in English are
extremely productive. Typically, the assumption is that one of these categorial
instantiations is the basic one, and that the other one is derived from it by zero-
124
Hagit Borer
affixation (see, for instance, Lexical Phonology and Morphology). To illustrate, the
noun walk is derived from the verb walk, having the representation in (18a), while
the verb chair is derived from the noun chair, having the structure in (18b):
(18) a. [Vwalk ] [N[Vwalk ] ØN ]
b. [Nchair] [V[Nchair] ØV ]
The representations in (18) involve a number of important assumptions. First,
they presuppose the existence of a listed basic lexical item, possibly a Lexeme in the
sense of Beard (1995), with a categorial label from which other categorically marked
forms may be derived. Second, they assume the existence of (at least) two null
categorial suffixes—a nominal one merging with verbs and a verbal one merging
with nouns. Finally, they subscribe to the view that, e.g., /πwalk/ may correspond to
both a mono-morphemic and a bi-morphemic structure.
Within Distributed Morphology, and coupled with the assumption that roots are
devoid of category, the relevant representations would, rather, be as in (19a–b), with
the Vocabulary Items corresponding to n and v realized as /πØ/ in both cases:
(19) a. [n[√WALK ] nØ ]
b. [n[√CHAIR] nØ]
[v[√WALK] vØ ]
[v[√CHAIR] vØ]
The representations in (19) likewise involve a number of important assumptions.
One of these assumptions is shared across both accounts—both assume the existence, in English, of zero equivalents of /πize/ and /πation/. In other respects,
however, the underlying assumptions are very different. First, (19) presupposes no
basic, listed, categorial forms. Rather, whatever listing exists consists of category-less
roots. Second, there is no derivational relationship between the nominal and the
verbal instantiations of roots. Rather, both are derived directly from the root. And
finally, as a result of the previous two assumptions, the verbal and the nominal
instantiations are equally complex, and specifically, the realization /πwalk/ always
corresponds to (at least) bi-morphemic structure regardless of its nominal or verbal
instantiation.
Finally, (20) would be the structure assigned to (17) in the contextual categorization system endorsed here (with D standing in for any ExP-segment of {Ex[N]} and
with T standing for any ExP-segment of {Ex[V]}):
(20) a. [ D [C=N
b. [ D [C=N
π
π
√walk ]]
√chair]]
[
[
T
T
[C=V
[C=V
π
π
√walk ]]
√chair]]
The representations in (20), like those in (19), deny the existence of a basic
categorially marked form, as well as any direct derivational relationship between
the verbal and the nominal instantiations. As in (19), both variants are related
directly to a root. Like the representations in (19), then, they do not postulate
differing levels of morphological complexity for the verbal and nominal instantia-
The category of roots
125
tions. However, (20) differs from (19) in two important ways. First, in (20), neither
verbal nor nominal instantiations involve the merger of an additional head, be it
nominal or verbal, eventually to remain phonologically unrealized. Secondly, and
relatedly, all forms are non-branching terminals and thus mono-morphemic syntactically as well as morpho-phonologically.
The table in (21) might be useful in comparing the assumptions across these three
approaches:
(21)
LPMt DMt XSMt
one categorial instantiation basic, the other derived from it
both forms ‘derived’ from a category-less root
both forms are mono-morphemic
both forms are bi-morphemic (at least)
zero categorizers
contextual categorizing
yes
no
no
no
yes
no
no
yes
no
yes
yes
no
no
yes
yes
no
no
yes
t
LPM: Lexical Phonology and Morphology; DM: Distributed Morphology;
XSM: Exo-Skeletal Model
With this exposition in mind, consider now the ungrammatical cases in (22b) and
(23b):
(22)
a. a salutation
an arrival
a neighborhood
a writer
the kindness
the ability
(23) a. *a crystalize
*an instantiate
*an acidify
*an encase
*a fatten
*an enlighten
b. *to salutation
*to arrival
*to neighborhood
*to writer
*to kindness
*to ability
b. to crystalize
to instantiate
to acidify
to encase
to fatten
to enlighten
From the sample of cases in (22)–(23), as compared to those in (17), a tentative
generalization emerges: phonologically unmarked categorial noun–verb alternations
in English appear impossible for morpho-phonological derivatives.
A further consideration of the facts, however, results in the conclusion that the
restriction cannot involve derivational complexity as such. The primary compounds
in (24) are indeed excluded as verbs, but not so the primary compounds in (25),
which are, presumably, derivationally complex as well:
126
(24)
Hagit Borer
a. math teacher
b. *to math teacher
mass destruction
*to mass destruction
law enforcement
*to law enforcement
fellow traveler
*to fellow traveler
piano recital
*to piano recital
word formation
*to word formation
(25) a. wardrobe
blackboard
chicken wire
wallpaper
grandstand
network
b. to
to
to
to
to
to
wardrobe
blackboard
chicken wire
wallpaper
grandstand
network
In turn, note that the compounds in (24) are themselves headed by a nominal
derivative, while the compounds in (25) are headed by a form that is itself (morphophonologically) mono-morphemic, and hence plausibly co-extensive with a root.
Differently put, the compounds in (25) are headed by forms otherwise belonging to
the (17) group. The generalization that thus emerges does not concern the derivational complexity of the forms in (22)–(23) as such. Rather, what appears to make a
difference is the presence vs. absence of an actual overt categorizer such as -ationN,
-izeV, -enV-, -hoodN, etc. We may state the generalization as (26):10
(26)
Categorially marked forms in English may not undergo a (phonologically)
unmarked noun–verb alternation.
Corroboration is available in a corpus study of some 1,200 cases of unmarked noun–
verb alternations, in which Clark and Clark (1979) found exactly 6 cases in which the
alternation involved an overt C-functor. The six forms are in (27). With the exception of blockade, they have been uniformly rejected by native speakers I have
consulted.
(27) a. to tourist; to launderette; to laundress; to lover; to allowance; to blockade
A slightly higher number was found among instrumentals, where out of 127
instances, 11 were derivatives, but nonetheless allowed a verbal occurrence, including
to computer, to glider and to elevator. We note that these, too, are of questionable
felicity, and that even if more acceptable, remain a very small portion of the
vocabulary, when compared to the overwhelming number of mono-morphemic
forms in English that are attested, simultaneously, as nouns and as verbs.
Consider now how this generalization can be handled. In LPM, and harking back
to the representations in (18), it is extremely difficult to see how the ungrammatical
10
The presence of pairs such as a portion/to portion; a condition/to condition and others appears to cast
doubt on the generalization in (26). I take up this matter directly in section 6.4.4.
The category of roots
127
cases in (22)–(24) can be ruled out. To exemplify, it is very hard to see why (18a–b)
are licit, but not so (29a–c) or (31c–d). Nor is it obvious why the forms in (25) are
licit, presumably with the structures in (31a–b) but not so the forms in (24),
effectively with the identical structure, as in (31c–d):
(28)
a. [V [N acidi] ify]
b. [N [V arrive ] al]
c. [N [A absurd] ity]
(30)
a.
b.
c.
d.
[N[N
[N[A
[N[N
[N[N
wall]
black]
math]
mass]
paper]
board]
teacher]
destruction]
(29) a. *[N[V [N acidi] ify] ØN ]
b. *[V[N [V arrive ] al]ØV ]
c. *[V[N [A absurd ] ity]ØV ]
(31)
a
b.
c.
d.
[V[N[N
[V[N[A
*[V[N[N
*[V[N[N
wall]
black]
math]
mass]
paper]ØV]
board]ØV]
teacher]ØV]
destruction]ØV]
Within LPM, at least some affixal combinations are excluded by level mismatches.
Specifically, Kiparsky (1982a, 1997) assumes that the ØN merging with verbs is a Level
I affix. Thus, if -enV is a Level II affix, as seems plausible, the absence of *the fatten
follows from a level mismatch. Similarly, possibly, for *the encase, under the
assumption that the categorial enV- prefix is Level II. Level mismatch, however,
cannot rule out any of the cases involving a putative ØV, by assumption a Level II
affix, when merging with derived nouns, or cases in which a putative ØN is merging
with clear Level I verbal affixes such as -ify or -ate. Even more problematically, there
is no obvious reason why ØV, always a Level II affix, can merge with the compounds
in (25), but not to the compounds in (24).
The Distributed Morphology representations in (19) fare no better. Crucially, for
DM, as for LPM, the syntax which underlies /π acidify/ and the verbal instantiation of
/π form/ is identical, with the difference between them reducing to the choice of
Vocabulary Items for v. Similarly, the syntax underlying /π formation/ and /π form/N
is identical, with the difference between them reducing to the choice of Vocabulary
Items for n. It is thus not clear what is to exclude the relevant representations in
column B of (32), structurally identical, for all intents and purposes, to those in (19)
or to, e.g., crystallization or government(al)ize in column C:11,12
11
As already noted in footnote 6, in DM both form and formation are derived by n merger to the root,
making competition, as such, irrelevant for the VI of categorizers. As a result, we cannot appeal to
competition with to form to exclude to formation.
12
A sole account of primary compounds within DM is available in Harley (2009b), where the structure
proposed is as in (i), and where, presumably, a verbal structure underlying to wallpaper would involve the
merger of a zero-realized v node, either as in (iia) or as in (iib). It remains entirely unclear, however, why
wallpaper can merge with vØ while math teacher or mass destruction cannot:
i.
[nØ [[√WALL]√PAPER ]] → [√PAPER -nØ [[√WALL]√PAPER] ]
ii.
a. [√PAPER -nØ [[√WALL]√PAPER] ] → [vØ [√PAPER -nØ [[√WALL]√PAPER]]] →
[√PAPER -nØ-vØ [√PAPER-nØ [[√WALL]√PAPER]]
b. [vØ [[√WALL]√PAPER ]] → [√PAPER -vØ [[√WALL]√PAPER]]
128
(32)
Hagit Borer
A
a. [v[√CRYSTAL] vize]
B
*[n[v[√CRYSTAL] vize]nØ]
C
[n[v[√CRYSTAL]
vize] nation]
b. [n[√GOVERN] nment] *[v[n[√GOVERN] nment]vØ]
*[v[a[n[√GOVERN] nment] aal] vØ] [v[a[n[√GOVERN]
nment] aal] vize]
But suppose we assume that there are no zero-realized C-functors, thereby
excluding both (18) and (19), and opting for the contextual categorization representations in (20). An interesting contrast now emerges between, e.g., /π formation/ and
/π form/N. In both cases, the root π√form is rendered category-equivalent contextually. In both cases, the C-equivalent root π√form spells out identically (stress
notwithstanding) as /π form/. Within /π formation/, however, /π form/ is rendered
V-equivalent by the CN[V] to be realized as /πation/. In contrast, in /π form/ it is
N-equivalent, as defined by (some) nominal ExP-segment (e.g., D):13
(33) a.
CN[V]
CN[V] [C=Vπ√form]
[C=Vπ√form]
/π form/
b.
D
D
[C=Nπ√form]
/π form/
CN[V]
/πation/
With these configurations in place, accounting for (22)–(24) becomes entirely
straightforward. Consider first the items in (17). Crucially, no C-functors are
present. Rather, these are cases very much like (33b), with a root embedded directly
under some ExP-segment and thus rendered categorially-equivalent to the Categorial Complement Space of the relevant Extended Projection. A, the, or three, by
assumption ExP-segments of {Ex[N]}, would thus render roots such as π√salute,
π
√jump, or π√floor N-equivalent. Will, infinitival to, or PST, all ExP-segments of
{Ex[V]}, will all equally successfully render roots such as π√dance, π√floor, and
π
√lamp V-equivalent, thereby deriving the alternation. Such an alternation is possible for the forms in (17), however, precisely because roots have no inherent
category and can thus in principle be rendered C-equivalent to any CCS. This,
however, is not the case for the forms in (22)–(23), which are all categorical already.
Salutation already is an N, by virtue of being headed by CN[V]ation; encase already is a
V, headed by CV[N/A]-en etc.. In these cases, the roots π√salute and π√case having
already merged with a C-functor, have already been rendered C-equivalent, and are
now embedded within a larger, categorially specified constituent. But if phonologi13
And see Borer (2013) on the adjunction in (33a).
The category of roots
129
cally unmarked C-functors are not attested in English, and zero-marked noun–verb
alternations are never mediated by affixation, then it follows that, e.g., /πacidify/
cannot be an instantiation of a noun, quite simply because it is already V, by virtue
of /πify/, itself the realization of a CV[N] functor. If now embedded under, e.g., D, it
would become N-equivalent, and a clash would emerge resulting in ungrammaticality. As such, the ungrammaticality of /πacidify/ as N is exactly on a par with its
ungrammaticality if embedded, e.g., under /πity/, by assumption an instantiation of
CN[A] (offensive categorial clash enboxed and highlighted):
(34) a. ∗[ D
[c=n [c=n π√acid] Cv[n]/ify]
b. ∗[ Cn[a]/ity
∗the acidify
]
[c=a[c=n π√acid] Cv[n]/ify]
]
∗acidificity/∗acidifity
Acidify could, of course, merge with some instance of CN[V]. However, in the
absence of phonologically null instantiations for such functors, the presence of such
a functor would be phonologically detectible. The result is, of course, licit, giving rise
to acidification as well as to acidifying, both with a straightforward structure.
A similar account is readily available for the compounds in (24), once we consider
the fact that it is the head of the compound, rather than the compound as a whole,
that is categorially crucial. Because, e.g., π√stand is a root devoid of category, it is
equally comfortable in a V-equivalent and an N-equivalent context, giving rise to
both a nominal and a verbal licit instantiation of grandstand. Because teacher and
destruction are nouns headed by (phonologically overt) CN functors, spelling out as
/πer/ and /π ation/ respectively, embedding mass destruction or math teacher within a
V-equivalent context would result in a clash and in ungrammaticality, as (35)
illustrates:14
π√paper]];
(35) a. [D [c=n
[wall]
[T [c=v [wall] π√paper]]
π
∗
b. [D [c=n ERn[v] [c=v [math] √teach]]; [T [c=v ERn[v] [c=v[math] π√teach]]
Concluding this particular argument, note that the challenge here for proponents
of zero categorizers is to explain why it is that such categorizers systematically fail to
merge with overt affixes. There are, most certainly, restrictions on the co-occurrence
of affixes that must be captured within any approach and which, presumably, could
be stated within any of the approaches so far outlined (see especially Fabb 1988 and
Plag 1999, as well as discussion in Borer 2013). However, the challenge here is to
explain why such a generalization should hold for all zero categorizers, regardless of
their syntactic properties, insofar as neither zero N nor zero V categorizers can
merge with any overt affixes. Within Distributed Morphology, this is extremely
14
Although by and large orthogonal to the argument here, care was nonetheless taken to avoid
examples of verbal compounds which could be argued to emerge from the back-formation of synthetic
compounds (i.e., to hand shake but *math teach).
130
Hagit Borer
puzzling, given the fact that the phonological realization of Vocabulary Items should
be altogether orthogonal to their syntactic and hierarchical properties. From that
perspective, the fact that n and v display identical distributional restrictions exactly
when the Vocabulary Item they correspond to is null is difficult to explain. For
Lexical Phonology and Morphology, matters are further complicated by the fact that
the affixes not only belong to different categories, but also to different levels, and are
therefore, by assumption, distinct not only syntactically, but also morphologically,
semantically and phonologically. Why, then, should they display the same distributional restrictions? Needless to say, none of these complications emerge if one adopts
a contextual approach to the categorization of roots, coupled with the absence of
zero realizations for C-functors.
6.3.3 Zero categorizers: The problem of deverbal nominals
Grimshaw (1990) observes that Ø-deverbal nominals in English are largely excluded
as AS-nominals (Argument Structure Nominals, Grimshaw’s Complex Event Nominals). The generalization does have its exceptions, and thus the nouns in (36) can be
used as AS-nominals, although they are homophonous with their verbal instantiation. Nonetheless, the existence of such exceptions pales in comparison with the
overwhelming validity of the claim for a huge, productive class of nouns with
invariant verbal correlates, which are systematically barred as AS-nominals (see
(37) for a small subset of the relevant nouns, and (38) for the illustration that they
cannot be AS-nominals):15
(36)
change (and exchange), release, use, murder, discharge, endeavor, abuse, access,
consent, resolve.
(37) (an) admit; (an) arrest; bite(s); (a) break; cause(s); (a) chase; climb(s); (a)
descent; export(s); (a) fall; (a) float; (a) follow-up; (a) frown; (much) hate; (the)
hold; (the) import; (a) jump; (a) kill; (some) kiss; (a) laugh; (the) lick; (a) lie
down; lift(s); (a) look; (little) love; (a) make; (some) mock; (a) move; (every)
raid; (a) ride; (a) rock; (a) roll; run(s); (a) scream; (a) sit-in; (a) smile; (a)
smoke; (a) stand; (a) take; (a) talk; (a) think; (a) touch; (a) turn; (a) twist; (a)
view; (a) walk; (a) whisper; . . . (and see throughout for many more examples)
15
(36) is from Newmeyer (2009) who lists 11 such (valid) cases but claims the existence of ‘literally
dozens’, although he does acknowledge that “the majority—perhaps the large majority—of AS-nominals
are morphologically complex”, attributing it to historical reasons. His overall conclusion is that the form of
AS-nominals or any of their properties can be neither predicted nor correlated with the verbal source, and
hence must be listed.
Importantly, focusing on the cases in (36) leaves without any explanation whatsoever the picture in (37).
Here we have a productive, systematic, pervasive phenomenon of ‘deverbal’ underived nouns all of which
absolutely do not allow for AS-nominal syntax, in spite of the fact that many of them denote a (simple)
event and have a Content that is extremely event-friendly. Listing the nouns in (36) as AS-nominals may
indeed help explain their properties but it will be doing so at the expense of any hope of capturing the
generalization which underlies the properties of the massive class of nouns in (37). For comments on how
to accommodate the exceptional morpho-phonological nature of the cases in (36), see Borer, (2013).
The category of roots
(38) *the
*the
*the
*the
*the
*the
131
walk of the dog for three hours
dance of the fairy for a whole evening
(gradual) fall of the trees for two hours/in two minutes
salute of the officers by the subordinates
frown of the director for seven minutes
raid of the town in order to retrieve the prisoners
We note, in augmenting the solidity of Grimshaw’s original observation, that the
absence of AS-nominal instantiation for the nouns in (37) is in actuality extremely
surprising, given the fact that most of these nouns maintain an extremely close
Content relationship with their corresponding verbal instantiation, and the fact that
most of them can, indeed sometimes must, denote a (simple) event (cf. 39)—all
factors which would seem to encourage an AS-nominal formation, but the latter is,
nonetheless, almost without exceptions barred. For the sake of completeness, we
note that AS-nominals based on the same root are readily available with overt
nominal affixation:
(39)
the walk/dance/kiss/salute/touch/view/smoke/scream/roll lasted several hours
the arrest/bite/fall/raid/talk/kill/sit-in/turn/smile took place at 5 a.m.
(40)
the
the
the
the
the
(41)
the importation of goods from China in order to bypass ecological regulations
the salutation of the officers by their subordinates
walking of the dog for three hours
dancing of the fairy for a whole evening
(gradual) falling of the trees for two hours (multiple events)
saluting of the officers by the subordinates
viewing of the results by the visiting committee
In Borer (1993, 1999, 2013) I argue in some detail that AS-nominals must include
verbal structure. A schematic structure of an AS-nominal is in (42) with AspQ and E
event-related nodes, by assumption ExP-segments of {Ex[V]} (immaterial structural
details omitted):
(42)
Cn[v]
Cn[v]
E
[c=v π√transmit] Cn[v] subj
E
Aspq
DPq
obj Aspq
C=V
[c=vπ√transmit]
132
Hagit Borer
Crucially, none of the relevant properties of AS-nominals follow from the specific
spell-out properties of the circled C-functor in (42), quite regardless of whether it
spells out as /πation/ (transmission) or as /π al/ (transmittal) or /π ance/ (transmittance), or, for that matter, as /πing/ (transmitting). Rather, the properties of
AS-nominals derive from the fact that an event complex, itself consisting of
ExP-segments of a verbal Extended Projection, is embedded under some CN[V].16
Given the orthogonal nature of the realization of CN[V], we note, there is little reason
from the structure in (42), as such, to exclude a zero realization of CN[V]. Had there
been such a possible realization, little would need to change about the structure.
Differently put, should it turn out to be the case that CN[V] in (42) must correspond
to an overt phonological instantiation, the reason for that requirement could not be
reduced to the structure of (42). What is patently clear, however, is that unless zero
categorizers are excluded in English, the systematic exclusion of the nouns in (37) as
AS-nominals becomes impossible.
And indeed, excluding the nouns in (37) as well as countless others like them as
AS-nominals is a result that cannot be derived in accounts which subscribe to zero
categorizers, quite independently of the structure in (42). Neither Kiparsky (1982a,
1997) nor Marantz (1997, 2010) or Harley (2009a,b) can capture this overwhelming
generalization. Nor can it be captured by Chomsky’s (1970) original execution.
Consider first the account put forth by Kiparsky (1982a, 1997). In Kiparsky’s
system, deverbal ØN merges at Level I, but so do, at the very least, nominalizers
such as -ationN, -alN, and -ance/enceN.17 This fact in and of itself, then, cannot be
somehow used to derive the absence of the AS-nominal reading of nominals zeroderived from verbs. On the other hand, denominal ØV merges at Level II. If all cases
in (37) were cases of verbs derived from nouns, and, crucially, if one maintains the
generalization that AS-nominals must be derived from verbs, then the exclusion of
zero AS-nominals could follow. Specifically, in such cases, verbalizing of a listed
noun would take place at Level II, with the consequence that subsequent nominalizing could only be available with Level II nominalizers, e.g., -ing, but not with ØN,
and nor, for that matter, with -ation, -al or -ance/ence:
(43) [Npass]
listed form
→ [V[Npass] ØV] —*→
Level II
affix
[N[V [Npass] ØV] ØN]
*Level I
affix
Of course, the immediate problem for such an account would be the fact that the
forms in (37) do not bar ‘zero’ nominalization (if such exists) across the board—they
only bar it in the context of AS-nominals. If, indeed, all of these are derived verbs,
16
There are, in fact, semantic and syntactic differences between CN[V], a bleached syntactic operator
that may spell out, at the very least, as /πation, ence, ment, al/, and INGN[V], which is a semantic as well as
a syntactic operator. These differences, however, are by and large orthogonal to our purpose here. The
reader is referred to Borer (2013) for a detailed discussion.
17
The status of -ment is unclear insofar as it displays mixed level diagnostics.
The category of roots
133
then across the board, one would have to assume that the basic form is a nominal
one, and that e.g., walk, hate or think are all cases in which the verbal instantiation
derives from the nominal one. It is hard to see, in fact, that in a system such as this
English could afford too many underived verbs, with the possible exception, of
course, of endeavor, consent, and other members of the list in (36), which would
thus stand as the sole class of underived verbs in English. To this rather implausible
consequence, we should add the fact that at least some of the forms under consideration do allow nominalization with {-ation} (e.g., salute, form, import, export),
and hence must already be verbs at Level I, raising anew the question of why, at least
in these cases, ØN suffixation couldn’t give rise to AS-nominals. Finally, we note that
the claim that all unmarked noun–verb alternations in English are cases of verbs
derived from nouns comes extremely close to arguing that there is, indeed, a root,
but, contrary to the claim put forth here or in Distributed Morphology, it is nominal
in nature (and on this claim, see specifically Hale and Keyser, 1993, as well as
Acquaviva, 2008a, this volume).
Turning now to Chomsky (1970) (as well as Marantz 1997), we note that as by
assumption in these executions the morpho-phonological complexity of root instantiations in distinct categorial contexts is syntactically irrelevant, whether a root spells
out as form or as formation in a nominal context cannot be expected to give rise to
any syntactically meaningful results, making any statement of the actual difference
impossible.
Within executions which involve the merger of roots with categorial labels and
which do allow such categorial labels to be phonologically null, excluding form as an
AS-nominal but including formation, or, for that matter, excluding run but including running, cannot be accomplished. Ignoring potentially intermediate functional
nodes or matters of execution otherwise orthogonal to the main issue under discussion, (44) is, as far as I can see, an exhaustive list of all possible root–categorizer
combinations that may yield a derived nominal (and where aff is a theory neutral
reference to a phonologically overt affixal realization):
(44) a. Nouns derived from Verbs b. Nouns derived from Roots
[N n [V v [√ROOT]]]
[N n [√ROOT]]
vØ
nØ
nØ
naff
vØ
naff
nØ
vaff
naff
vaff
None of these representations predicts syntactic differences based on the phonological realization of affixes as overt or null. If one subscribes to the view that
AS-nominals must be derived from verbs, then the structures in (44b) would,
presumably, be excluded as AS-nominals. This, however, would only yield the
exclusion of the nouns in (37) as AS-nominals with the added assumption that n
134
Hagit Borer
may only spell out as Ø when it merges with roots but not when it merges with v,
whether vØ or vaff or, for that matter, that nØ may not merge with any affix, thereby
ruling out all representations in (44a) and reducing the problem to the one already
noted in subsection 6.3.2. Either way, it is clear that the only potential solution would
involve some special statement about phonologically null categorial nodes, contingent not on their syntax, but on their phonology, thereby setting them apart from
phonologically overt categorial labels and raising the obvious question concerning
their usefulness and their theoretical cost. We add, in this context, that proposing
that zero categorizers may only merge with roots makes it that much more pivotal to
justify their existence, insofar as, by assumption, they could only merge where
contextual categorization would otherwise be available.
Consider now a contextual categorization system, coupled with the claim that
English doesn’t have zero categorizers. If AS-nominals must include a verbal structure, as in (42), then the exclusion of the nouns in (37) as AS-nominals follows
straightforwardly from the fact that they do not contain a verbal projection. Rather,
they are roots in a nominalizing context, with no additional structural complexity.
They are, in other words, truly mono-morphemic, and whatever relationship they
hold to their verbal counterparts is mediated through the existence of a common
root and not through a direct derivational relationship.
A verbal constituent, in turn, can only emerge from a root in V-equivalent
contexts. Such contexts involve either verbal ExP-segments, such as E and AspQ,
or a C-functor with a verbal CCS. Considering again the structure in (42), the root
embedded under E and AspQ is per force V-equivalent. It further follows that such a
V-equivalent constituent must be nominalized in order for a derived AS-nominal to
emerge. In the absence of Ø realizations for such nominalizers, such nominalizations
must be overt, directly leading, again, to the conclusion that the forms in (37) cannot
possibly be AS-nominals. The unfolding of such a derivation is in (45):
(45)
a.
[
π
√form]
b.
[E [C=V[π√form] [ASP [C=V π√form] [C=V π√form]]]
c. [N [π√form]-CN[V]
/πing/
/πation/
*/πØ/
[E [C=Vπ√form] [ASP [C=V π√form] [C=V π√form]]]]
→ /π forming/
→ /π formation/
→ */πform/
6.3.4 Zero categorizers: The selection problem
The English default past tense marking, -ed, is entirely systematically the only one
available for overtly derived verbs (see also Embick 2010). There are no overt verbal
derivatives that take either -Ø past tense (plus potential stem allomorph), or -t/-d,
the other Vocabulary Items available in Halle and Marantz (1993) for past tense
marking. Put differently, ‘irregular’ past tense marking in English is only available
for morpho-phonologically mono-morphemic forms.
The category of roots
135
The very same situation holds for English plural marking. When compared with
past tense, the variety of plural forms in English is in actuality quite a bit bigger,
allowing for a (putative) -Ø coupled with a stem allomorph (men, women, feet, geese,
fish, sheep, etc.); -en, likewise associated with potential stem allomorphs (oxen,
children, etc.), and the default -s, but also a variety of originally Latin or Greek
plural forms such as -i (foci, loci), -a (data, phenomena), and others. Nonetheless, the
only plural form associated with overtly derived nouns, including those derived with
suffixes of a Latino-Greek origin such as -logy, is the default -s. In other words, just
as in the case of the past tense, ‘irregular’ plural marking in English is only available
for morpho-phonologically mono-morphemic forms.
From the point of view of root-based systems, this is in actuality a true bonanza,
insofar as it provides direct evidence for the existence of roots. Specifically, the
default marking on all derived forms follows directly from the claim, hardly controversial, that all instances of selection are local, coupled with the assumption that
all unpredictable phonological information is associated specifically with roots. The
reason derived forms are restricted to a default instantiation, then, is that the
relationship between the root and the inflectional marking is not local, a fact
illustrated in (46) for past tense and in (47) for plural (the representations deliberately sidestep the question of whether inflection is morphemic):18
(46)
[T-PST[V[π√instant]CV[N]-ate]-ed]
[T-PST[V[π√acid ] CV[N]-ify]-ed]
[T-PST[V[π√harmon] CV[N]-ize]-ed]
[T-PST[V[π√fat] CV[A]-en]-ed]
[T-PST[V CV[A/N]en-[π√case]]-ed]
(47)
[CNT-(PL)[N[π√form]CN[V]-ation]-s]
[CNT-(PL)[N[ π√defer] CN[V]-ment]-s]
[CNT-(PL)[N[π√sister] CN[N]-hood]-s]
[CNT-(PL)[N[π√refer] CN[V]-al]-s]
[CNT-(PL)[N[π√able] CN[A]-ity]-s]
[CNT-(PL)[N[π√weak] CN[A]-ness]-s]
[CNT-(PL)[N[π√read]CN[V]-ing]-s]
[CNT-(PL)[N[π√fellow] CN[N]-ship]-s]
[CNT-(PL)[N[π√write] CN[V]-er]-s]
The problem which now emerges for models subscribing to the existence of zero
categorizers is immediately evident. Consider the specific set of assumptions made in
Distributed Morphology, recalling, specifically, that by definition any categorized
18
At least some of the C-functors in (46)–(47) have a semantic function as well as a syntactic one, and
hence, in actuality, ERN[V], SHIPN[N] etc. The matter is set aside here for expositional reasons. See Borer
(2013) for a fuller discussion.
136
Hagit Borer
stem is complex, at the very least bi-morphemic, regardless of whether the node that
it merges with is phonologically overt or null—see (19a–b), repeated here as (48):
(48)
a. [n[√WALK] nØ] [v[√WALK] vØ]
b. [n[√CHAIR] nØ] [v[√CHAIR] vØ]
In turn, DM does assume, entirely correctly, in my view, that markers such as past
tense merge with verbs, and not with roots, and that presumably markers such as
plurality merge with nouns, and not with roots. But if every verb and every noun are
complex, then it follows that the relationship between the root and the past tense
marker, or between the root and the plural marker, is no more local in the case of
run than it is in the case of instantiate, and no more local in the case of foot than it is
in the case of formation or fellowship. For DM, specifically, the relevant representations are thus as in (49) and (50) respectively, where any statement concerning the
local selection of non-default inflection is impossible to make:19
(49)
a. [[√INSTANT]-v]-pst]];
↓ ↓
VI:
-ate -ed
[[√ACID] -v] -pst]];
↓
↓
-ify -ed
b. [[√RUN] -v]-pst]]; [[√BEND ]-v] -pst]]];
↓ ↓
↓ ↓
VI:
Ø Ø
Ø -t
(50)
a. [[√FORM]-n]-pl]];
↓ ↓
VI:
-ation -s
[[√DEFER]-n]-pl]];
↓ ↓
-ment -s
b. [[√GOOSE]-n]-pl]]; [[√OX]-n]-pl]];
↓ ↓
↓ ↓
VI:
Ø Ø
Ø -en
[[√FAT] -v]-pst]]…
↓ ↓
-en -ed
[[√WALK]-v] -pst]]
↓ ↓
Ø -ed
[[√SISTER]-n]-pl]]…
↓ ↓
-hood -s
[[√BOOK]-n]-pl]]
↓ ↓
Ø -s
The failure of local selection in such cases is noted and discussed in Embick (2003,
2010). By way of attempting to solve the problem, Embick suggests that locality is
sensitive to phases, rather than strict locality. By assumption, the first phase is that
which involves categorization, and so, for, e.g., run, the first phase is [[√RUN ]-v].
The irregular past tense selected by run, in turn, is adjacent to this first phase,
making it local in the relevant sense. We note that the solution does predict that, e.g.,
verbalize would have a regular past tense, insofar as there is, by assumption, a phase
19
Lexical Phonology and Morphology fares somewhat better on this front, but still faces problems, a
matter to which I return in section 6.4.5.
The category of roots
137
(boundary) between -al and -ize. The solution nonetheless fails insofar as it fails to
distinguish, within its own terms, between [[√RUN ] -vØ]-pst] or [[√GOOSE ]-nØ]
-pl]] on the one hand, and [[√ACID ] -v-ify]-pst] or [[√BREV ]-nity] -pl]] on the
other hand. Both brevity and acidify, by assumptions internal to DM, define first
phase, and there is little to distinguish them structurally from run or goose. And yet,
irregular inflection is excluded. In fact, insofar as the selection of -ify, -ate, en-,
-ment, -ance, -al, and so forth is root-conditioned, they must merge with roots, and
the output defines first phase, where irregular inflection, by this logic, cannot be
excluded.20
In contrast, the selection properties thus far outlined present no problem whatsoever to a system of contextual categorization. We already noted that in XSM, unlike
DM, morpho-phonological constituent structure goes hand in hand with morphosyntactic constituent structure, and specifically, that morpho-phonological terminals
are not branching syntactic constituents. While in Distributed Morphology both
formV and formN are syntactically complex and neither PST nor PL can be adjacent
to the root, for contextual categorization, both are terminals categorized without the
addition of syntactic complexity, as in (51):
(51) a. [CNT-(PL) [C=N π√run ] ] [T-PST [C=V
b. [CNT-(PL) [C=N π√foot] ] [T-PST [C=V
π
√run ] ]
√foot] ]
π
By the same logic, derivatives with overt categorizers are complex (see (46) and
(47), with some examples repeated below):
(52) [T-PST[V[π√liqui]CV[N]-ify]-ed]
[T-PST[V[π√fat] CV[A]-en]-ed]
[CNT-(PL)[N[ π√defer] CN[V]-ment]-s]
Given these representations, we are in an excellent position to state the distribution of irregular inflection locally and as encoded directly with roots. In (51), the
relationship between the roots and PST or PL is local, while it is not in (52), where a
categorizing C-functor dominates the root. The formal advantage, we note, derives
directly from the fact that given contextual categorization, a root may be categorized,
or, more accurately, rendered category-equivalent, without increasing its hierarchical complexity. Not so in a system that assumes zero categorizers, where, for
principled formal reasons, a root is never adjacent to an inflection marker (a segment
20
Insofar as the derivations in (49b) and (50b) involve, in DM, the realization of pst as Ø following a Ø
realization of v and n, the system as a whole cannot exclude consecutive zero-affixation, nor can it
maintain that zero-realized morphemes may only merge with the root. Neither one of these assumptions,
then, can be appealed to, to exclude the relevant cases in (44a) or to delimit the structural possibilities in
(16). The choice for proponents of zero categorizers in resolving this matter is thus rather limited, and
must be based, fundamentally, on treating zero affixation, syntactically every bit as real as overt affixation,
as nonetheless transparent to root selection, and as having this property in parallel in both its n and v
instantiations.
138
Hagit Borer
of an Extended Projection, in the terminology developed here, or functional node, as
conventionally perceived).21
6.3.5 Interim summary: Root-based systems
The account put forth here shares with Distributed Morphology the claim that roots
may merge with potentially affixal category labels, C-functors in the present account.
The two accounts differ, however, in their approach to the categorization of roots. As
noted already, roots, in and of themselves, not only do not have a category inherently
in DM, but remain a category-less terminal in the structure. As such, then, no
categorial generalizations can ever be applied to nodes that are co-extensive with
roots, and all categorial operations by definition must involve a complex structure.
Because, by assumption, the terminal occupied by the root is never categorized, zerorealized categorizers become inevitable, for otherwise, there is quite simply no way of
associating e.g., √WALK, with a verbal structure, eventually to spell out (in the
context of PST) as /πwalked/.
Zero-realized categorizers, however, come with a heavy cost, as just illustrated.
Conceptually, they result in an across-the-board failure to correlate morpho-phonological complexity with syntactic complexity, thereby depriving us of our primary
window into the internal structure of complex words. In the absence of strict
mapping between morpho-phonological complexity and structure, what could
bear on the existence, or lack thereof, of structures such as (53a) when compared
with structures such as (53b), and what, if anything, would be the consequences of
structural proliferation?
(53)
a. [v v [nn [v v [nn [√FORM]]]]] (form)
[√FORM]]
(form)
b. [v v
[a a [v v [ √FORM]]] (formal)
[a a
[ √FORM]] (formal)
The matter is particularly tricky within DM because of the assumption (see
Marantz 2000, Arad 2003, Marvin 2002) that non-compositionality is delimited by
the domain of (first) categorization. Formal as in (53b) is predicted to be noncompositional, but formal as in (53a) must be compositional. However, in the
absence of any independent evidence for the complexity of structure postulated,
the claim runs the danger of being circular.
Empirical problems abound as well, and we noted several, all requiring treating zerorealized categorizers as distinct, across the board, from realized categorizers, and with
identical distinctions holding for both verbal and nominal instantiations, in and of
itself already a surprising and disturbing result. Unlike overt categorizers, they may not
attach to already derived forms and appear licit only when merging directly with the
root; unlike overt categorizers, they cannot give rise to AS-nominals, and unlike overt
21
Note that at least in principle, both N and V derived from the same root could show irregular
inflection. The situation, however, appears not to be attested, e.g., run–ran, but runs; foot–feet but footed.
See Borer (2013) for some discussion.
The category of roots
139
categorizers, they do not create an opaque domain for selection. Altogether, then,
and across their categorial instances, zero affixes very much behave like they are
simply not there. What evidence, then, can be brought forth to bolster their
existence, and how compelling is that evidence?
A detailed argument for the existence of zero categorizers is, as it turns out, put
forth in Kiparsky (1982a, 1997) within a specific level-ordering based hypothesis, that
of Lexical Phonology and Morphology. Some problems for that approach were
already pointed out in this section, insofar as these were common to all systems
which do not incorporate contextual categorization. No attempt to do away with
zero categorizers in English could possibly be complete, however, without challenging the substantial argumentation in their favor mounted by Kiparsky (1982a, 1997).
A fuller critique of Kiparsky’s arguments is undertaken in Borer (2013). A summary
of that critique is outlined below.
6.4 Against English zero categorizers, part II: Kiparsky (1982a, 1997)
Kiparsky, as already noted, argues for (at least) two zero categorial affixes in English,
one which derives nouns from verbs and which is a Level I, and another one which
derives verbs from nouns, and which is a Level II. Insofar as there are distinct zero V
and zero N categorizers, Kiparsky’s account tallies with that of Distributed Morphology. However, in contrast with DM as well as with the account put forth here,
he assumes a directional derivational relationship - Ø-derived nouns and Ø-derived
verbs are not derived independently from roots, but from forms that are already
categorial. In turn, the major force of his argument derives from the distinct properties of ØN and ØV. We note, specifically, that in a system that assumes no zero
categorizers in English and which rather subscribes to the view that both verbal and
nominal instantiations of unmarked noun–verb alternations are categorized contextually, such differences, if present, are unexpected and in principle problematic,
unless they can be explained by appealing either to general differences between
nouns and verbs, or to the properties of the functor which merges with the root.
For Level Ordering-based Morphology systems, of which Lexical Phonology and
Morphology is the most articulated execution, the most fundamental claim is that
(English) affixes can be divided into (at least) two distinct pools, with distinct
semantic, morphological and phonological properties associated with each, and
with one of these pools, Level I affixes, ordered before (=inside of ) the second,
Level II affixes. Some of the crucial properties of each pool are listed in (54). Sample
affixes of each level are in (55):22
22
The explanatory value of the classification has, in turn, come under considerable criticism in the past
two decades, beginning, specifically, with Fabb (1988) and continuing with Plag (1999). See Borer (2013) for
some additional discussion.
Forms in parentheses are in reference to affixes whose Level status is either mixed or in dispute.
140
Hagit Borer
(54) LEVEL I (+ boundary) affixes
Merge with words and non-words
May or may not be productive
Output may or may not be compositional
May or may not involve irregular inflection
Affect (some) phonological rules
assimilation
stress shift
Tend to be Latinate
Tend to merge with Latinate bases
(55)
CV
CN[V]
CA
CN[A]
(-ize), -ify, -ate
-ation, -al, -anc(y)/enc(y), (-ment)
-ous, (-able), -al
-ity
LEVEL II (# boundary) affixes
Merge only with words
Productive
Output compositional only
Regular inflection only
Do not affect (same) phonological rules
no assimilation
no stress shift
Tend to be Germanic
-en, en-, be-, (-ize)
-ing, (-ment)
-en/ed, -er, -ful, -less (-able)
-ness
N–N, A–N compounds
6.4.1 Two denominal verbs in English?
Consider now the paradigm in (56)–(57), and examining specifically the relationship
between the verbal and nominal instantiations of hammer, paint, tape, and lacquer:
(56)
a. I hammered the nail in (with my sandal).
b. I painted the wall (with lacquer).
(57)
a. I taped the picture (*/#with pushpins).
b. I lacquered the wall (*/#with paint).
c. Screw the fixture on the wall (*/#with nails).
By way of explaining the contrast, Kiparsky (1997) notes that the Content of paint
appears to be considerably more flexible than that of lacquer, and hence it is possible
to paint with something other than paint. Similarly, the Content of hammer is
considerably more flexible than that of tape, and hence it is possible to hammer
with things which are not a hammer. The anomaly of the instrumental specification
in (57), it would appear, emerges from the fact that taping can really only be done
with tape, and lacquering can only be done with lacquer. But why should that be so?
The reason, Kiparsky argues, is that the verbs tape and lacquer are derived from the
corresponding nouns with a Ø-verbal suffix which is a Level II suffix. Level II affixes
merge with words and are compositional. On the other hand, hammer and paint, as
nouns, are derived from the corresponding verb at Level I. Level I affixation may be
non-compositional. Specifically, the Content of the verb hammer is not actually USE
A HAMMER, but rather is entirely compatible with a broader Content, say DELIVER REPEATED BLOWS WITH A HEAVY OBJECT. The Content of the
verb paint clearly is not actually COVER WITH PAINT, but rather COAT WITH
DECORATIVE LIQUID FINISH. In turn, because tape and lacquer are derived
The category of roots
141
from nouns at Level II, they are, so to speak, accountable to the Content of the
original nouns TAPE and LACQUER, resulting in the relevant restrictions.
As Harley and Haugen (2007) show, however, the judgments in (57) appear to
result from a certain misclassification of the canonical Content of tape, lacquer, and
screw, respectively. Thus consider the examples in (58)–(59):
(58)
a. Lola taped the poster to the wall with Band-Aids/mailing labels.
b. Screw the fixture on the wall with nails—
OK, provided that the nails are twisted to affix the fixture (Harley and
Haugen 2007)
(59)
Bento boxes of the week: lacquered with bitter persimmon juice (Google
search)
Thus while the infelicity of (57a) may, indeed, emerge from the fact that the verb
tape cannot mean AFFIX in general, neither does hammer mean HIT in general.
Rather, hammer means something like DELIVER REPEATED BLOWS WITH A
HEAVY OBJECT, and tape means something like AFFIX WITH STICKY STRIPS.
Similarly, lacquer is not strictly COVER WITH LACQUER nor is it COVER WITH
DECORATIVE FINISH, but rather something like CREATE A GLOSSY HARD
FINISH. Once this is granted, it emerges that there is quite simply no evidence
from the paradign in (56)–(57) for two distinct Ø-categorizers operating at two
different levels.23
6.4.2 Productivity?
[V[N]ØV], a Level II operation, is very productive, Kiparsky claims. Not so
[N[V]ØN], which is a Level I operation. We note, however, and with (37) in mind,
that the non-productivity of [N[V]ØN] is highly questionable. By way of some
specific illustrations, note that the following are all possible, and some are clearly a
recent addition (the Boston Globe ran a headline with an embed three weeks after the
beginning of the first Gulf War):
(60)
a. an àdmit (a newly admitted student)
b. an admìt (a record of a positive admission decision or a newly admitted
student)
c. give it a think (note the absence of blocking from thought)
d. a responsive read (proof-reading process, from a magazine editorial board)
e. there are people on that list who deserve a listen
f. a scheduled (court) hear
g. an embed (a journalist ‘embedded’ in a US military unit)
23
And see Anagnostopoulou and Samioti (this volume) for a discussion of this paradigm from a
different angle.
142
Hagit Borer
6.4.3 Stress shift?
V→N alternations may give rise to stress shift, as in (61a), Kiparsky observes, but not
so the N→V alternations in (61b):
(61)
progréssV→ prógrèssN
a. permítV→ pérmitN;
b. pátternN → *pattérnV
admítV→ ádmitN
If Level II affixation is involved in N→V, and if stress shift is applicable only to the
output of Level I affixation, the absence of stress shift in the change from N→V vs. its
possibility in V→N cases would be explained.
The claim, however, is only compelling if we take it for granted that permit and
pattern are listed with one categorial instantiation, and the second one is derived
from it. If we assume, instead, that both nominal and verbal instantiations are
derived from a root (or a compound root), e.g., as in (62), stress shift, as such, is
non-existent, and what we have in its stead is distinct stress patterns in English,
rather systematically, for compound verbs and compound nouns, coupled with the
relatively plausible claim that prefixed forms, such as those in (61a), are at least
phonologically compound-like in nature. We note in this context that the Compound Stress Rule in English would systematically give rise to initial stress, although
by assumption compounding is Level II, making the assignment of initial stress to
nouns a poor test for Level I affixation altogether.
(62)
a. {Ex[V]} [C=Vπ√gress] [pro] → [C=V[pro][C=Vπ√gress] [pro]] → /πprogress/
b. {Ex[N]} [C=Nπ√gress] [pro] → [C=N[pro][C=Nπ√gress] [pro]] → /πprógress/
Compound Stress Rule¿
Further support for a directional treatment of the stress alternation emerges,
Kiparsky claims, from the existence of triplets such as those in (64). What is notable
about these triplets is the emergence of a second instantiation of the verb, but with
the nominal stress pattern:
(63)
compóundV
permítV
cómpoundN
pérmitN;
cómpoundV
pérmitV
Such triplets, Kiparsky argues, must have the analysis in (64). The original verb,
permít, merged with ØN at Level I with stress shift resulting. The second occurrence
of the verb, as pérmitV, on the other hand, clearly takes the nominal form as its
input, and being a Level II formation, does not allow for the change of stress, giving
rise to pérmitV:
(64)
a. [compóund]V → [[cómpound]V
ØN]NØV]V(Level II)
ØN]N(Level
I)
→[[cómpound]V
b. [permít]V → [[pérmit]V ØN]N(Level I) →[[pérmit]V ØN]NØV]V(Level II)V
The category of roots
143
Note now that intuitively, it is of course entirely obvious that the emergence of
pérmitV is a result of the existence of pérmitN, with its altered Content. The question,
however, is whether the conclusions drawn by Kiparsky (1982a) are inevitable.
Consider, specifically, the possibility that the nominal instantiation of pérmitN.is
derived as in (62b), and as such, not from the verbal instantiation permít, but rather
directly from the root. Suppose now that as related to its altered Content, /πpérmit/
has become relisted as a separate root (roughly, OFFICIAL SANCTION). Alongside
the derivation from root-particle compounding depicted in (62b), then, the language
has now acquired the roots π√pérmit (and π√cómpound), phonological representations inclusive of penultimate stress, which, we expect, would now proceed to occur
as both V- and N-equivalent, but where the existence of the root-derived nouns in
(65b) is obscured by the existence of the nouns /πpérmit/ and /πcómpound/ as derived
through (62b), and as of yet, with an identical Content:
(65)
a. {Ex[V]} [C=Vπ√pérmit] → /πpérmit/; {Ex[V]} [C=Vπ√cómpound] → /πcompound/
b. {Ex[N]} [C=Nπ√pérmit] → /πpérmit/; {Ex[N]} [C=Nπ√cómpound] → /πcómpound/
6.4.4 An argument from ordering
We noted briefly, in footnote 10, the existence of cases such as (66):
(66) a. a portion, the condition, rations, a motion, the air condition, positions,
a proposition, the audition, questions, sanctions
b. to portion, to condition, to ration, to motion, to air condition, to position, to
proposition, to audition, to question, to sanction
The cases, at least prima facie, appear to contradict the claim made in section 6.3.2,
according to which unmarked noun–verb alternations in English are only attested with
underived forms, by assumption roots, and in fact, Kiparsky (1982a) explicitly suggests
that such cases lend support for postulating a ØV affix at level II, where it follows the
affixation of -ation, the latter a clear Level I affix. Using a similar rational, ØV merger
with compounds, as in (25) (repeated here as (67)) argues for ØV ordering after
compounding, itself a Level II operation, and hence for ØV as a Level II affix:
(67) a. wardrobe
blackboard
wallpaper
grandstand
network
b. to wardrobe
to blackboard
to wallpaper
to grandstand
to network
A closer look at these cases reveals, rather surprisingly, that the only -ation-ending
forms which undergo unmarked noun–verb alternation are forms which in their
144
Hagit Borer
nominal instantiation do not have a coherent verbal source from which they could
possibly be (synchronically) derived. These and similar cases could be divided into
two groups. For cases in (68a), the subtraction of -ation gives rise to an otherwise
unattested root verb. In (68b) are cases where the subtraction of -ation leaves behind
an English verb, but with a Content that is clearly divorced from that of the complex
form in (66):24
(68)
a. (port-), cond-, rat-, motb. to pose, to propose, to audit, to quest, (to port)
That the forms in (66) which do appear to have a ‘source’ are not in any way
whatsoever related to that putative source can be seen from the complete impossibility of using them as AS-nominals with the Content of the source verb. Thus there
is no way the underlined nouns in (70) could possibly be the AS-nominals expressing the events in (69). In fact, the nominals in (70) are altogether ungrammatical as
AS-nominals, barring aspectual modifiers, argumental by-phrases, and purpose
clauses:
(69)
a. The model posed in front of the camera for several hours (in order to give
herself a chance to deliver the correct photograph).
b. Mary proposed a solution in two minutes (in order to please her boss).
c. The authorities audited my tax records for several weeks (in order to
establish my guilt).
d. John quested for love for years (in order to become happy).
(70)
a. ?the position of the model in front of the camera (*for several hours) (*in
order to . . . )
b. *Mary’s proposition of a solution (in two minutes) (in order to . . . )
c. *the proposition of a solution by Mary
d. *The authorities’ audition of my tax records (for several weeks) (in order
to . . . )
e. *the audition of my tax records
f. *the question for love (for years) by adult males
?= with the intended reading as events associated with (69a)
The ungrammaticality of the AS-nominals in (70) is quite noteworthy, because the
relevant forms here behave very differently from other cases of non-compositional
24
To the best of my ability to ascertain, and with the possible exception of to inconvenience, unmarked
V–N alternations are not attested with the nominal realizations /πal; ance–ence/, both of them Level I (e.g.,
*to recital; *to nuisance). Nor are there any with /πity/, similarly Level I. -age is sporadically attested with
bandage and voyage and possibly also mortgage and baggage. Interestingly, the most common instance
involves -ure, with, at the very least, to puncture, to gesture, to picture, to culture, to conjecture, and to
structure. In all these cases, the verbal source is extremely difficult to spot, and in fact, cases in
contemporary English with a transparent verbal source for -ure are altogether rare, suggesting that its
affixal instantiation may very well be a historical matter.
The category of roots
145
derived nominals, of which there are many. Transmission or contraction (as in birthrelated spasm), to exemplify, certainly come with Content no longer computable
from the Content of either transmit or contract. Nonetheless, the form has retained
its ability to express a fully compositional event, as (71) illustrates:
(71)
a. Mary’s transmission of the documents (in two minutes) (in order to . . . )
b. the slow contraction of the rubber saddle for three hours
It thus emerges that the forms in (66) with the putative sources in (68b) have gone
beyond acquiring a non-compositional Content alongside their compositional one,
and rather proceeded to lose their ‘internal verb’ altogether. Put differently, they are
no longer complex in any meaningful sense. Unsurprisingly, and given the absence
of a discernible verb, the cases in (68a) do not give rise to AS-nominals either, as (72)
illustrates:
(72)
a. *the motion of the model in front of the camera (*for several hours) (*in
order to . . . )
b. *Mary’s condition of the furniture (in two minutes) (in order to . . . )
c. *the ration of the food by the government
Rather, the forms in (66) have the exact same cluster of properties typically found
with category-less roots. They exhibit an unmarked alternation with verbal forms,
conditioned solely by Extended Projections, they display no internal complexity and
may not function as AS-nominals, and finally, they require an overt C-functor in
order for an AS-nominal to emerge. In view of this, there is little reason to believe
that /πation/ as it occurs on the forms in (66) is an instance of CN[V], or that these
forms are complex and derived. I will assume, rather, that their historical emergence
notwithstanding, for present day English, they have been re-listed as roots, thereby
accounting for the fact that their synchronic properties are, indeed, those that we
fully expect from roots.
The status of the forms in (66) in present day English notwithstanding, the
strongest argument against allowing the merger of a ØV with a derived noun
comes from the observation that the N–V alternation in English is extremely
productive, and that practically any mono-morphemic noun in English has a
homophonous verbal correlate. As a result, proponents of Level II ØV would surely
want to postulate it as an extremely productive affixation rule. And yet the cases in
(66) are quite rare and clearly not productive. Nor is there a clear account, for
proponents of Level II ØV, for the fact that ØV is only available to derivatives that
have clearly lost all residue of their compositional Content. Why should ØV merge
with question, but not with formation? Why should it merge with section, but not
with admission, or admittance, for that matter? In short, there is simply no possible
146
Hagit Borer
explanation, as based on the Level II status of ØV, for the complete impossibility, in
English, of the forms in (73):25
(73) *to construction, *to availability, *to performance
Turning to the compounds in (67), we note that analyzing the verbal variant as
involving ØV suffers from exactly the same problem. I noted already in discussing
this matter that alongside the grammaticality of the verbal instantiations in (67), we
have the contrastively ungrammatical cases in (24), repeated here as (74):26
(74)
a. math teacher
b. *to math teacher
mass destruction
*to mass destruction
law enforcement
*to law enforcement
fellow traveler
*to fellow traveler
piano recital
*to piano recital
word formation
*to word formation
In Kiparsky’s system, there is simply no reason for this contrast. Compounds are a
Level II process, as is the affixation of ØV, and compound formation is oblivious to
the derivational history of its head. There is no difference, compound-wise, between
chicken wire and word formation. That it is exactly the compounds with a nominally
marked head that block verbalization, but not others, follows from a contextual
categorization, but cannot be derived by appealing to the properties of a Level II ØV.27
6.5 Conclusion
This chapter was devoted to the development of a categorization system in which
syntactic category is exclusively a prerogative of functors. Pivotal, here, was the
perception of categorial labels as corresponding to divisions within the syntactic
space. Our C-functors thus define an outer spatial domain of a particular type (e.g.,
N, V), as well as a complement space of a particular type (likewise N, V), a function
25
Note that insofar as we can happily come up with transformationalization, preventing ØV merger
with compositional derived nouns by blocking appears rather dubious.
26
There is a direct argument here against Lowenstamm’s view (this volume) of (effectively Level I)
categorizers as roots, and with, e.g., acceptance having the structure in (i) (and see also De Belder, 2011b):
i. [n[[√ACCEPT] [√ANCE]] nØ]
Specifically, while wallpaper would clearly allow the structure in (iia), acceptance, or for that matter
piano recital, would be barred in that very same (putative) structure:
ii.
a. [v[[√WALL] [√PAPER]] vØ]
b. *[v[[√ACCEPT] [√ANCE]] vØ]
c. *[v[[√PIANO] [[√RECITE][√AL]]] vØ]
27
In conjunction with the claim, in Kiparsky (1982a), that irregular inflection is always Level I, an
interesting prediction emerges, within LPM, relative to the derivational properties of V–N pairs such as
shit or spit, with their irregular past tense realization. For a detailed discussion of these issues, see Borer
(2013).
The category of roots
147
we notated as CX[Y]. A slightly distinct but nonetheless contextual categorization
emerged in the contexts of Extended Projections, viewed as a set of (consecutive)
nodes with a shared CCS, and where the formal equivalent of a C-functor is not a
single ExP-segment (e.g., D, T), but rather the set as a whole, defining an outer
spatial domain of a particular type (e.g., what is typically referred to as DP) and
defining its complement space as a particular type (say N).28
Crucially, within the system developed here, syntactic categories are never inherently associated with non-functors, e.g., roots, an assumption shared with Distributed Morphology. Unlike the categorization system developed in DM, however,
contextual categorization effectively amounts to the denial of the existence, in any
syntactic tree, of uncategorized domains. Roots, then, may not have an inherent
category, but within any syntactic structure they will per force be co-extensive with a
categorized domain.
Of some significance, now, is the fact that in a system of contextual categorization,
zero-realized functors are certainly not excluded in principle, and whether or not
they do exist is fundamentally an empirical, rather than a theoretical issue. Not so,
however, for (hierarchy-based) systems that do not subscribe to contextual categorization, be they root based, as is DM, or lexeme-based, such as LPM. To wit, if
contextual categorization is absent, and assuming √BREAK to be a root, zero
functors are essential to derive both [N break] and [V break]. Similarly, absent
contextual categorization, the lexeme break is either N or V. Either way, a zero
functor is essential to derive its second categorial instantiation. By contrast, in a
system that subscribes to contextual categorization, zero functors are not necessary,
quite simply because √BREAK will emerge as N in the context of, e.g., D, and as V in
the context of T without any need for C-functors.
To the extent that zero categorizers are essential in any hierarchical system that
doesn’t subscribe to contextual categorization, compelling arguments against the
existence of such zero categorizers in English per force favors contextual categorization. Much of this chapter was devoted to showing that regardless of whether noncategorial roots are assumed (DM) or not (LPM), assuming zero categorizers in
English results in substantial empirical and formal costs, and few, if any, theoretical
or empirical advantages. The conclusion, then, is inevitable. If the structure of
complex words is to be integrated into a hierarchical system, categorization must
be contextual.
Beyond contextual categorization, in itself a theoretically significant issue, a
consequence of much greater significance emerges as well. Crucially, in the absence
of zero categorizers and with contextual categorization, a considerably higher level of
correspondence can be maintained between morpho-phonological complexity and
28
And see Borer (2013) for a fuller perspective on the categorial properties of Extended Projections.
148
Hagit Borer
hierarchical structure, altogether an extremely encouraging result. Morpho-phonological cues, much more so than any other linguistic factor, remain the first and
foremost markers of relatedness, not only within the phonological domain, but
within the syntactic and semantic domain as well. By Occam’s razor, any model
that takes the structural significance of such cues seriously should, then, be favored.
7
On a low and a high diminutive:
Evidence from Italian and Hebrew
MARIJKE DE BELDER, NOAM FAUST,
AND NICOLA LAMPITELLI
7.1 Introduction
Cross-linguistically, diminutives display mixed behavior with respect to meaning. To
illustrate, take the Italian examples below. Example (1) shows a diminutive with a
compositional meaning; examples (2)–(4), on the other hand, have non-compositional meanings.1
(1)
nas-ino
nose.dim
‘small nose’
(2)
pan-ino
bread.dim
‘sandwich’
(3)
cas-ino2
house.dim
‘brothel’
(4)
telefon-ino
telephone.dim
‘cell phone’
[Italian]
1
We take compositionality and non-compositionality to be the equivalent of “semantic predictability”
and “semantic unpredictability”, respectively. See Bertinetto (1995) for a slightly different use of these
terms in morphology.
2
Cf. Pianigiani (1926).
150
Marijke De Belder, Noam Faust, and Nicola Lampitelli
Whereas the combination n+dim in (1) simply refers to a smaller version of what the
noun denotes, in (2)–(4) the diminutive morpheme derives a new denotation. As a
result, (1) cannot be combined with an augmentative morpheme, because a concept
cannot be small and big at the same time, as can be seen in (5). The diminutives in
(2)–(4), in contrast, freely combine with an augmentative. This is illustrated in (6).
(5)
*nas-in-one
nose-dim-aug
(6)
pan-in-one
bread-dim-aug
‘big sandwich’
[Italian]
The data above argue for two types of diminutivization: compositional and noncompositional diminutives.
This chapter argues that the distinction between the two diminutives can be
captured in terms of syntactic structure. More specifically, we propose two different
positions for diminutives. The first position we refer to as SizeP. It is situated in the
functional domain of the noun, between the categorial head n° and the projection
which hosts number marking (see De Belder 2011a for a detailed discussion of this
head, see further Acquaviva this volume). Because it realizes functional material it is
characterized by full productivity and compositionality. The second one directly
merges with the root and realizes a lexical position below the category head,
henceforth LexP.3 Because of its low position, a diminutive in Lex° is oblivious to
the nature of the category head above it. As such, it is not restricted to nouns: a
verbal head may select the same configuration of LexP+√. In addition, because it
does not realize a functional head, it may show lexical gaps and non-compositional
meaning.
Summing up, we propose that diminutives can realize both functional and
derivational heads cross-linguistically. Both positions are shown in (7).4
3
A reviewer raises an alternative, according to which the low dim is not a head, but rather a modifier
adjoined to the root, as in Wiltschko and Steriopolo (2007). However, note that -in imposes its formal
features to the base: the item pane ‘bread’ has the theme vowel -e, whereas its diminutive panino
‘sandwich’ has -o (*pano, *panine). In Italian, final vowels are the exponents of gender and number
features. In addition, low -in can change the gender of the base (as German -chen): donna woman-F-sg
‘woman’ vs. donnina woman-dim-F-sg and donnino woman-dim-M-sg. These facts point to the head
status of such an item. Therefore lexical material such as -in- in panino ‘sandwich’ can be treated as a head
root which projects √P, as in Lowenstamm’s (this volume) account of English suffixes. An alternative is to
assume a low position, which we label Lexo. The differences between the two approaches are beyond the
scope of this chapter.
4
Our proposal relies on two previous studies: De Belder (2008) proposes the projection SizeP to
introduce diminutive inflection on nouns, whereas Lampitelli (2010) proposes that Italian diminutives are
introduced by a projection between nP and √.
On a low and a high diminutive: Evidence from Italian and Hebrew
(7)
151
DivP
Div⬘
Div°
SizeP
Size⬘
Size°
nP
n⬘
n°
LexP
Lex⬘
Lex°
√
Data from Romance, Semitic, Germanic and Slavic languages will provide support
for our proposal.
Our analysis is consistent with the syntactic approach to word-formation proposed by Embick and Noyer (2007), Harley and Noyer (1999) and Marantz (1997,
2001) among the most relevant ones. More precisely, we assume that the first
category-assigning head demarcates a frontier between two different structural
domains. The domain below that head is reserved for lexical meaning and thus
non-compositionality, whereas the one above hosts functional projections, whose
meanings cannot be idiosyncratic (see Anagnostopoulou and Samioti this volume
for an opposing view).
The chapter is structured as follows. Section 7.2 discusses predictions which follow
from the proposal. We first show that a language can formally distinguish between
the two diminutives (}7.2.1). In section 7.2.2, it emerges that both positions can be
filled simultaneously. Section 7.2.3 shows that the lower diminutive combines with
uncategorized material and that it does not determine the category of this combination, which can therefore be the base of either a noun or a verb. Finally, we discuss
the prediction that some languages may have a derivational diminutive without
having an inflectional one or vice versa (}7.2.4). After having discussed these predictions, we address some theoretical consequences of the proposal in section 7.3. The
final section concludes and sums up.
7.2 Predictions on the behavior of Lex° and Size°
As mentioned above, this chapter explores the hypothesis that diminutives may spell
out two different syntactic positions, which we refer to as Lex° and Size°. If our
152
Marijke De Belder, Noam Faust, and Nicola Lampitelli
hypothesis is correct, then the following four situations must be attested in some, if
not all languages:
(8)
Predictions
a. Different morphological strategies of diminutivization may correspond to
the two different positions;
b. Both positions may be filled simultaneously;
c. The derivational diminutive is oblivious as to the category head that selects
it; and
d. The two types of diminutives may exist independently of each other.
In what follows, Modern Hebrew (henceforth MH) will be shown to exemplify (8a).
(8b) will be confirmed by data from various other languages. (8c) states the prediction that one and the same low diminutive formation is not necessarily restricted to a
single category: the same diminutive may yield nouns, verbs, adjectives, and adverbs.
Hebrew and Italian will again provide the relevant data. Finally, the prediction in
(8d) will be confirmed by languages that lack one of the two positions.
7.2.1 Prediction #1: different morphological strategies of diminutivization
may correspond to the two different positions
The first logical consequence of our hypothesis concerns how Lex° and Size° are
realized cross-linguistically. As we saw in the introduction for Italian -in, the same
morpheme can be used to realize both positions. Still, by assuming two positions or
domains, it is predicted that the form of the diminutive may reflect its position, i.e.
that different behavior may be detected in the different domains with respect to
form, too.
This is indeed the case in Modern Hebrew (henceforth MH), which shows two
different paths to diminutivization. MH uses either a concatenated morpheme -on or
a reduplicated and discontinuous morpheme QTaLTVL, wherein the V is sometimes
(though not always) the vowel of the corresponding noun.5
(9)
Diminutivization in MH
a.
b.
c.
d.
e.
f.
Noun
xazir
bacal
xatul
kélev
géver
xamor
‘pig’
‘onion’
‘cat’
‘dog’
‘man’
‘donkey’
Templatic Dim
xazarzir
‘piglet’
bcalcal
‘shallot’
xataltul
‘kitten’
klavlav
‘puppy’
gvarvar
‘macho’
*xamarmor
Concatenative Dim
xazir-on ‘small pig’
bcal-on
‘small onion’
xatul-on ‘small cat’
kalb-on
‘small dog’
gavr-on
‘small man’
xamor-on ‘small donkey’
5
Surface mismatches between V in QTaLTVL and a corresponding noun concern only segholates (see
9d and 9e). It has been shown on independent grounds in Faust (2011) that these nouns involve an
underlying /a/ between the two last consonants, which would explain the vowel in their respective
templatic diminutive.
On a low and a high diminutive: Evidence from Italian and Hebrew
153
The interpretation of concatenative diminutives is strictly compositional. In
addition, it is a fully productive morphological process. In the terms of the present
study, this means that it is spelled out at the Size° position.
Conversely, templatic diminutives must be the product of Lex°, because of the
following two properties. Firstly, their interpretation is not semantically predictable.
They do not denote a smaller counterpart of the basic noun. For example, xazarzir is
not simply a “small pig”. Secondly, they are not morphologically productive.6 The
non-concatenative diminutive applies only to a closed group of roots, as is shown by
the illicitness of *xamarmor.7
If so, in MH we find a case for two positions for diminutives on both semantic and
morphological grounds: two different morphological strategies exist that correspond
to two distinct semantic effects. The first strategy, the templatic one, involves a low
diminutive and may result in non-compositionality; the second strategy—concatenation—points to a high diminutive, and is always compositional.8
7.2.2 Prediction #2: Both positions may be filled simultaneously
If two positions exist for diminutives, one predicts that both positions could be filled
simultaneously. In other words, one expects to find a compositional diminutive on
top of a non-compositional one. This is indeed borne out in many languages. In the
introduction it was already pointed out that the lower diminutive may co-occur with
an augmentative in Italian. The example is repeated in (10). In the same vein, the
lower diminutive can co-occur with a higher, compositional diminutive in this
language:
(10)
pan-in-one
bread-dimlex-augsize
‘big sandwich’
(11)
pan-in-etto
bread-dimlex-dimsize
‘small sandwich’
[Italian]
Similar data are found in other languages:
(12)
stół
‘table’
[Polish]
6
By “not productive” we mean that no new words are formed with this strategy. In other words, it is
impossible to predict whether an item will be eligible for this strategy. This situation stands in sharp
contrast to the productive diminutive strategy of -on suffixation, which in principle is possible for any
noun in the language. For a discussion of productivity with emphasis on MH, see Bolozki (1999).
7
Unlike Hebrew, Italian -in does both jobs with the same root: casino ‘brothel’, casina ‘small house’ (cf.
casa ‘house’).
8
See Faust (2011) and Faust and Lampitelli (2012) for further discussion of the construction of
templates by templatic morphemes situated low in the structure.
154
Marijke De Belder, Noam Faust, and Nicola Lampitelli
(13)
stoł-ek
table-dimlex
‘chair’
(14)
stół-ecz-ek
chair-dimlex-dimsize
‘small chair’
(15)
bolso
‘bag’
(16)
bols-illo
bag-dimlex
‘pocket’
(17)
bols-ill-ito
bag-dimlex-dimsize
‘small pocket’
(18)
kalb
‘dog’
(19)
klayb
‘puppy’
(20)
klayb-un
puppy-dimsize
‘small/cute puppy’
(21)
xazir
‘pig’
(22)
xazarzir
pig-dimlex
‘piglet’
(23)
xazarzir-on
piglet-dimsize
‘small piglet’
[Spanish]
[Tunisian Ar.]
[MH]
The Polish and Spanish examples display a sequence of diminutive suffixes, of which
the inner one gives rise to non-compositional meaning. In Tunisian Arabic and MH,
non-concatenative morphology is used to derive a possibly non-compositional
diminutive, and a strictly compositional diminutive is realized as a suffix. In conclusion, two diminutives may co-occur. If one assumes two different positions for
diminutives, this is unsurprising.
On a low and a high diminutive: Evidence from Italian and Hebrew
155
We conclude the discussion of this prediction by giving the structure for the MH
double-diminutive xazarziron ‘small piglet’:
(24)
Structure of MH double diminutive
DivP
Div⬘
xazarzir-on
‘small piglet’
SizeP
Div°
Size⬘
Size°
-on
nP
n⬘
n°
LexP
Lex⬘
Lex° √xzr
CVCaCCVlexC
7.2.3 Prediction #3: The derivational diminutive is oblivious to the category
head that selects it
LexP, as we have been using it here, is an elaboration on the root, in that it both
precedes the category-assigning head and has no category of its own. Therefore, in
order to prove the existence of such a position, we have to show that the combination of a root and a diminutive morpheme may serve as the base of more than one
category. In other words, neither the root nor its combination with the diminutive
affix are category-specific.9 This section suggests that this is the case of reduplicated
diminutives of the type Q-TL-L and Q-TQ-T in MH.
Before we can approach these diminutives, we must examine the general case of
QiTeL verbs and nouns. MH has several verb types. The most productive type is
called QiTeL (wherein by convention {Q,T,L} represent root consonants and {i,e} the
vocalization of the unaffixed verb stem). The majority of the verbs in this group have
related nominal forms (also called action nouns) of the shape QiTuL:
9
Unlike the affixes which are discussed in Borer (this volume), the diminutive does not force a specific
categorial interpretation on the root.
156
(25)
Marijke De Belder, Noam Faust, and Nicola Lampitelli
QiTeL, QiTuL in MH
a.
b.
c.
d.
Verb
xipes
šitef
nipec
kilef
Action noun
xipus
‘search’
šituf
‘sharing’
nipuc
‘shattering’
kiluf
‘peeling’
We would like to propose that although QiTeL and QiTuL are certainly related, it
is wrong to posit a relation of derivation in either direction. We make this claim for
three empirical reasons. First, QiTuL nouns may have an idiosyncratic meaning,
only vaguely related to that of the verb (26a,b). In fact, there may even be no existing
verbal base (26c,d). Finally, a sub-group of QiTeL verbs has {o,e} vocalization (26e,f ).
Their corresponding nominal melody is still {i,u}, not *{o,u}. Thus, even the form of
the QiTuL noun is not influenced by that of the verb.
Not all the verbs in (26) are vocalized alike. The fact that they nevertheless express
the same verbal pattern can be established by the [e] vocalization of the prefix m- of
the participle, which uniquely characterizes the pattern under discussion.
(26)
QiTuL is not derived from QiTeL
a.
b.
c.
d.
e.
f.
Verb
miten
cimek
roken
pocec
Participle
me-maten
me-camek
me-roken
me-focec
Action
‘moderate’ mitun
‘shrink’
cimuk
sikuy
biyuv
‘empty’
rikun
‘explode’ picuc
noun
‘moderation/(economic) depression’
‘shrinking/raisin’
‘chance’10
‘gutter’
‘emptying’ (*rokun)
‘explosion’ (*pocuc)
We take the facts in (26) as sufficient evidence against a derivation of QiTuL directly
from QiTeL (or vice versa). That said, it is clear that both forms are related through a
common base. The relation is thus best described as follows: first the template is
specified; the category is assigned only subsequently.
This derivation is modeled in (27). First, the root is combined with an action
head ι (Doron 2003) that inserts a template Q_T_L. The formal effect of ι is to
syllabify the root by creating vocalic positions within it. However, this syllabified
root is not yet vocalized; its vocalization will be determined only by the category
head subsequently merged with it.
10
We take it that form reflects structure: even though these nouns do not have corresponding verbs, we
assume that they have to share a structure with QiTuL nouns that do.
On a low and a high diminutive: Evidence from Italian and Hebrew
(27)
157
QiTeL and QiTuL are both derived from Q_T_L
a. cimek ‘to shrink’
b. cimuk ‘shrinking/ raisin’
v → cimek
v
ι → c_m_k
ι
√cmk
{i,e} Q-T-L
n → cimuk
n
ι → c_m_k
ι
√cmk
{i,u} Q-T-L
In (27) we use Doron’s diacritic ι as the morpheme that the Q_T_L template realizes.
Notice, however, that the position is the same position that we have labeled LexP,
namely between the root and the category-assigning head.
Having established that QiTeL and QiTuL are not derived from one another, we
may return to diminutives. Another sub-group of QiTeL verbs is QiTLeL, i.e. verbs
with a third reduplicated radical. Such verbs, which often carry a pluractional
diminutive meaning (diminutives and pluractionals often coincide in verbal morphology11), are shown in (28). They may be related to a basic QaTaL-type verb
(28a,c), or exist alongside an equi-radical QiTeL verb (28d,e). Other QiTLeL verbs
may have no equivalent (28f,g). Regardless of origin, all the forms in (28) have both
the verbal and the nominal versions.
(28) QiTLeL verbs
related item
QiTLeL
verb
a. caxak ‘to laugh’
cixkek
b. laxaš ‘to whisper’
lixšeš
c. kafac ‘to jump’
kifcec
d. kiven ‘to aim/to direct’ kivnen
e. išer
‘permit (auth.)’
išrer
f. fikšeš
g. širbev
“diminutive”
action noun
cixkuk
‘giggle’
lixšuš
‘whisper quietly’
kifcuc
‘jump around’
kivnun
‘fine-tune’
išrur
‘allow bureaucratically’
fikšuš
‘commit a small error’
širbuv
‘carelessly insert, stick out’
We have chosen verbs of the type QiTLeL because some of these verbs have
non-diminutive equivalents in QaTaL, and are thus contrastable to those. But, as
discussed at length in Greenberg (2010), QiTeL contains many other diminutive
(/pluractional) verbs of the fully reduplicated sub-group QiTQeT:
11
Literature on pluractionals is vast since Cusic (1981). Tovena (2010, 2011) addresses some related
issues, namely on Italian verbal diminutivization from a semantic point of view.
158
(29)
Marijke De Belder, Noam Faust, and Nicola Lampitelli
Diminutives in QiTQeT
a.
b.
c.
d.
QiTQeT
milmel
nimnem
rixreax
pitpet
‘mutter’
‘doze’
‘sniff ’
‘babble’
Related item
mila ‘word’
nam ‘sleep’
reax ‘odor’
Now consider the tree representation for a QiTLeL diminutive. There is now one
other morpheme to realize before we apply the category head: the diminutive
morpheme. This morpheme is called [dim1] in (30); together with ι, it assigns the
root with the special subpattern of QiTeL, namely QiTLeL (we name it [dim1] in
order to distinguish it from the diminutive realized as QiTQeT). In this example, we
regroup both ι and [dim1] under the same label, LexP.
(30)
cixkek and cixkuk ‘giggle’
a. cixkek ‘to giggle’
v
v
b. cixkuk ‘a giggle’
→ cixkek
LexP
ι,[dim1]
→ c_xk_k
√cxk
{i,e} Q_TL_L
n
n
→ cixkux
LexP
ι,[dim1]
→ c_xk_k
√cxk
{i,u} Q_ TL_L
In (30) we have what our proposal predicted: a diminutive devoid of category,
which can serve as the basis for items of more than one category (the same
structures, with a different diminutive morpheme [dim2], will yield the QiTQeT–
QiTQuT pairs in 29).
Recall that MH also had a concatenative diminutive -on, which we suggested was
exclusively SizeP material. The noun cixkuk, but not the verb cixkek, should be able
to appear before on. This is indeed the case: cixkukon is presented in (31); the suffix
-on is never found on any verbal form.
(31)
cixkukon ‘a small giggle’
SizeP → cixkukon
Size
-on
nP → cixkuk
n
LexP → Q_TL_L
ι, [dim]
√cxk
On a low and a high diminutive: Evidence from Italian and Hebrew
159
Modern Hebrew, if so, exemplifies the lack of category in LexP. This, of course,
should be demonstrable in other languages too.
Consider the examples from Italian in (32). The root √fischi in Italian has the
approximate denotation of ‘whistle’. Two nouns may be derived from it: one with a
simple class marker o (32a) and the other with the same class marker preceded by an
exponent ett, already familiar for its diminutive meaning (32b). Both nouns have
parallel verbs (32c,d). In the case of fischio-fischiare, it is hard to tell whether there is
a derivational relation or not between the two. Nothing, in truth, indicates that there
is one, and the relation might very well be one of a shared base, as we saw above for
MH. The case of fischietto-fischiettare is clearer: the verb cannot be derived from the
noun, because the semantic relation between the two is non-compositional.
(32)
LexP in Italian is category-less in Italian
a. fischi-o
c. fischi-are
whistle.sg.m
whistle-infinitive
‘whistle (the action)’
‘to whistle’
b. fischi-ett-o
whistle-DIM-sg.m
‘whistle (the object)’
d. fischi-ett-are
whistle-DIM-infinitive
‘to emit short whistles repeatedly
(not necessarily with a fischietto)’
Just like in the MH case, there is no relation of derivation between the noun fischietto
and the verb fischiettare; rather, both are derived from a common base. We propose
that this common base is a complex structure root+LexP.
The relevant tree diagrams are in (33). In both fischi-ett-o and fischi-ett-are, the
exponent -ett- occupies the low position. It merges with the root. As we saw for MH,
this position has no category and is oblivious to the category head that will be
assigned to it. If it is n we derive fischi-ett-o, but if it is v we derive fischi-ett-are.
Neither is derived from the other.
(33) fischi-ett-o vs. fischi-ett-are
a. fischi-ett-o ‘whistle (object)’
n→
n
fischi-ett-o
LexP
-ett
b. fischi-ett-are ‘whistle (pluract.)’
√fischi
v → fischi-ett-are
v
LexP
-ett
√fischi
To conclude this section, we have seen that low diminutives (templatic in MH)
occupy a position close to the root, which we call LexP.12 This position is lower than
12
Other accounts place diacritics on roots (see Embick and Halle 2005) or let L-nodes select for types
of roots (Acquaviva 2008a).
160
Marijke De Belder, Noam Faust, and Nicola Lampitelli
the category-assigning head, and thus not marked for category. However, as we saw
above (e.g. MH -on), the possibility of yielding more than one category is not shared
by all diminutive markers. The existence of both types of diminutive markers is
expected if diminutives can indeed occupy two positions: a pre-categorial, derivational one and a post-categorial, functional one.
7.2.4 Prediction #4: The two types of diminutives may exist independently
of each other
Under the assumption that languages only select a subset of the features of Universal
Grammar (Iatridou 1990), a Size° head (i.e. a head that hosts a productive diminutive) may not exist in some languages. That said, if the positions SizeP and LexP exist
independently of one another, such languages may still allow for low diminutives.
This situation is found in English, for example. It is well known that English only
contains some sporadic diminutives, such as piglet, gosling and catkin (Marchand
1960). Similar examples can be found in French and Egyptian Arabic. Both are
languages without a productive diminutive, but some diminutive markers can be
found in derivations. This is illustrated in (34)–(35) for French and in (36)–(37) for
Egyptian Arabic, in which the diminutive is realized by means of non-concatenative
morphology.
(34)
fill-ette
girl-dimlex
‘young girl’
(35)
livr-et
book- dimlex
‘small notebook’
(36)
bint
‘girl’
(37)
bannuuta
girl-dimsize
‘young girl’ (template: QaTTuuL)
[French]
[Egyptian Ar.]
We can conclude that the structurally low diminutive can exist in a given language,
independently from the structurally high one. The question immediately arises
whether the opposite holds as well. Do some languages have only high diminutives,
without having a low one? It is hard to provide this question with an answer from a
methodological point of view: it is (nearly) impossible to prove that a given lexicon
contains not a single example of a non-compositional diminutive. However, if such a
case may exist, it might be found in a Creole language, in which a diminutive
morphology has developed only recently. In Mauritian Creole, for example, there is a
On a low and a high diminutive: Evidence from Italian and Hebrew
161
strategy for the formation of diminutives which originates in the adjective ‘small’.
The adjective is illustrated in (38), the newly formed diminutive in (39).
(38)
tipti sez
small chair
‘small chair’
(39)
ti-sez
chair.dim
‘small chair’
It may well be the case that such languages with limited, recent morphology
show the first reflexes of a high diminutive, while no derivational diminutive exists
as of yet.13
Summing up, the low and high diminutives may exist independently of one
another. It is clear that the low diminutive occurs in languages in which no
productive high diminutive exists. The other side of the coin is harder to prove.
We suggested that the first occurrences of the high diminutive in Creole languages
may not have a low counterpart.
7.3 Further issues
Most attempts at the formalization of diminutives have assumed a single locus for all
diminutive morphology, which corresponds to what we have been calling the high
diminutive (See, e.g., Bachrach and Wagner 2007, Ott 2011a, Stump 1993). Wiltschko
and Steriopolo (2007) is an exception to that trend. In that paper, diminutives are
also classified according to the level they attach to, which is either the root or a
category head. As in the present study, the fact that a diminutive marker is shared by
more than one category serves as proof for a low level of attachment. The authors
further show that this state of affairs has an effect on the order of affixes, with some
diminutives appearing below derivational morphology. The proposal and analysis of
the present chapter should be viewed as further confirmation for this two-level
distinction.
In addition to this distinction, Wiltchko and Steriopolo submit that a diminutive
can be either a head or a modifier. If a diminutive marker imposes gender or
conjugation class features on its base, it is a head; if these features of the base remain
unaltered, the diminutive is an adjunct. This distinction is incompatible with our
findings, because it does not coincide with the compositionality condition. Dutch
high diminutives, for example, are completely compositional and change the gender
of the noun they attach to, as shown below. The noun ster ‘star’ in (40) has common
13
That said, since a low diminutive is only optionally non-compositional, it is hard to prove on the
basis of semantics alone that a language like Mauritian Creole doesn’t have such low diminutives.
162
Marijke De Belder, Noam Faust, and Nicola Lampitelli
gender. This can be concluded from the fact that it selects the common definite
article de. However, when it is diminutivized, it selects the neuter article het, as
illustrated in (41).
(40)
de ster
the star
‘the star’
(41)
het ster-etje
the star-dimsize
‘the small star’
Italian compositional diminutives completely overwrite the inflectional class
information (i.e., theme vowel) of the base. For example, the diminutive of nav-e
‘boat’ is nav-in-a, not *nav-in-e.14 Pending further evidence, we do not see a
justification for the head-modifier distinction.
We would like to point out a specific aspect of our proposal regarding the relation
between syntactic information and exponence. As we have seen for Italian, -in- may
be the realization of either SizeP (nasino ‘small nose’) or LexP (panino ‘sandwich (lit.
small bread)’). We would like to use this last example to clarify the status of LexP
and show two different advantages it may have.
The first advantage is in modeling grammatical change. Consider how the noncompositional panino noun comes into being. First, we have a compositional
diminutive noun (40a), with -in- realizing Size°. When this combination becomes
grammaticalized, -in- loses its diminutive meaning. It is no longer interpreted as
realizing the head Size°, and comes to occupy Lex°, below nP (40b):
(42)
Compositional vs. non-compositional pan-in-o
b. Grammaticalized
a. Pre-grammaticalized
pan-in-o ‘sandwich’
pan-in-o ‘small bread’
SizeP → pan-in-o
nP → pan-in-o
[dim]
-in-
n
n
n
√pan
LexP
-in-
√pan
Crucially, the item in (40b) is still morphologically decomposable: this is only
possible if a low position such as Lex° is available for -in- to move into.
It may be concluded from the relative independence of exponent and position
that the former is a sign similar to the root. Like a root, it does not have a category
or a single fixed position. Lowenstamm (this volume) accords such morphemes
14
Italian diminutives can marginally change the gender of the basic noun, i.e. donna ‘woman.F-sg’ and
donnone ‘ugly woman.M-sg’.
On a low and a high diminutive: Evidence from Italian and Hebrew
163
the status of a bound root (see also De Belder 2011b). Our analysis conforms to
such a view.
The second advantage of LexP is that it is not exclusively reserved for diminutives.
Indeed, as we have said above, LexP is more generally a position for root augmentation: other morphemes may also be represented as hosted by LexP. We have argued
for this to be the case in MH, where an action head ι was also included under the
label LexP. We conclude this section of the chapter with a similar example from
Italian:
(43)
Derivational LexP
a. bors-a
handbag-f.sg.
‘handbag’
[Italian]
b. bors-eggi-o
handbag-Lex°-m.sg.
‘the act of mugging’
c. bors-eggi-are
handbag-Lex°-infinitive
‘to mug (not just to steal a borsa)’
Both the noun borseggio and the verb borseggiare (41c) are attested. We propose that
morphological material (in the example above: -eggi-) which is not the expression of
any type of diminutivization, is best analyzed as being in a position Lex°, merging
directly with the root below the categorial head. We leave for future work the
question of interplay between the different exponents that may occupy the Lex°
position.
7.4 Conclusion
On the basis of Italian data, we proposed that two different projections (Lex° and
Size°) should be postulated to host what has always been called “diminutive”
morphology.
We then showed that the four predictions which followed from our proposal are
all correct, using data drawn mainly—but not only—from Modern Hebrew and
Italian. We illustrated that the low and high positions may correspond to different
morphological strategies, that both positions can be filled simultaneously, that the
low diminutive is oblivious to the category with which it merges and that a language
may have one position, but not the other.
After a short discussion of further implications of the proposal, we concluded that
LexP is an independent projection which is merged with the root below the
categorial node.
8
The interaction of adjectival
passive and Voice
EDIT DORON
8.1 Introduction
Passive participles are derived in Hebrew by the three different verbal templates of
the language. The present study shows that there is a difference in syntactic structure
between adjectival passive participles derived by the causative template and those
derived by the other two templates. The difference stems from the different values of
the non-active voice available in the three templates. As a result, the adjectival
passive participle of the causative template has a richer structure than that of the
other templates. In particular it projects the external argument, which the other
templates do not.1
Verb, noun, and adjective stems in Semitic languages are derived from consonantal roots by different intercalations, called templates, of consonantal patterns,
vowel sequences and affixes. While there are many templates that derive nouns from
roots, the number of verbal templates, called binyanim in Hebrew grammatical
tradition, is extremely limited. In Modern Hebrew, setting aside voice variation,
each verb stem is derived by one of exactly three templates. These templates, also
found in Classical Hebrew and the other ancient Semitic languages (Akkadian,
Aramaic, Arabic, Ge’ez), are traditionally known as (a) the simple template, (b) the
intensive template, and (c) the causative template. Each verbal template derives, in
addition to finite verbs, two participles: an active participle and a passive participle.
The present study is concerned with the passive participle.
1
I am grateful to the audiences and organizers of the following workshops: the Root Workshop held at
the Hebrew University in December 2008, the Rootbound Workshop held at the University of Southern
California in February 2009, and the Root Workshop held at Universität Stuttgart in June 2009. I am
indebted to Hagit Borer, Anita Mittwoch, Malka Rappaport Hovav, and two anonymous reviewers for
their feedback. This work was supported by ISF grant 1157/10 to Edit Doron and Nora Boneh.
The interaction of adjectival passive and Voice
165
The passive participles of the three templates are introduced in section 8.2 of the
chapter. Section 8.3 shows that passive participles of all three templates can be
categorized as adjectives in addition to their categorization as participial forms of
verbs. Section 8.4 discusses the difference between the interpretation of adjectival
passives of the simple and intensive templates and the richer interpretation of
adjectival passives of the causative template, the latter also including the verb’s
external argument. Section 8.5 reviews the verbal template system and the syntactic
derivation of verbs. Section 8.6 discusses the derivation of adjectival passives within
this system, demonstrating that the minimal structure which derives adjectival
passives in the simple and intensive templates is a middle-voice structure not
including the verb’s external argument, whereas the minimal structure which derives
adjectival passives in the causative template is a passive-voice structure obligatorily
including the verb’s external argument. This difference is shown to be determined by
the difference in the values which non-active voice can assume for the different
verbal templates. Section 8.7 concludes the chapter.
8.2 The three passive participle templates of Hebrew
Example (1) below shows the three passive participles derived from the root √sdr
‘order’, and example (2) illustrates their adjectival use as noun modifiers. The
subscript found with each passive participle (S/I/C) denotes the corresponding
template (simple/intensive/causative) which derives it from the root:
(1)
Adjectival passive participles derived from the root √sdr ‘order’
a. simple
sadurS
‘ordered (in a configuration)’
b. intensive
mesudarI ‘arranged (tidy, orderly)’
‘settled (regulated, pre-arranged)’
c. causative musdarC
(2)
a. zug sadurS
b. xéder mesudarI
pair ordered
room arranged
‘an ordered pair’
‘a tidy room’
c. heskem
musdarC.
agreement settled
‘a pre-arranged agreement’
The categorization of passive participles as adjectives is illustrated by their cooccurrence with the pronominal copula pron (Doron 1983, 1986), which is not
found with verbs:
(3)
šaršéret
ha-mazon b-a-téva
hi
sduraS/mesudéretI/musdéretC.
(the-)chain (of ) the-food in-the-nature pron ordered/orderly/regulated
‘The food chain in nature is ordered/orderly/regulated.’
An additional example is shown below of the three passive participles derived from
the root √kp ‘bend’ in (4). Their adjectival distribution as noun-modifiers is illustrated in the examples in (5), found on the Internet.
166
Edit Doron
derived from the root √kp ‘bend’
‘stooped, subordinate’
‘bent’
‘subordinated’
(4)
Adjectival passive participles
a. simple
kapupS
b. intensive
mekupapI
c. causative
mukpapC
(5)
a. gab
kapupS hu
lo
gzera
mi-šam-áyim.
back stooped pron not decree from-sky-dual
‘A stooped back is not decreed by God.’
b. gib’ol mekupapI dmuy-bérek
stalk bent
like-knee
‘a knee-shaped bent stalk’
c. ha-yéda’
hayom hu
be-’ópen
totáli mukpapC
the-knowledge today
pron in-manner total subordinated
l-a-nihul.
to-the-administration
‘Knowledge nowadays is totally subordinated to administration.’
A puzzling contrast in interpretation is found between the simple/intensive
adjectival passives in the examples above and the causative adjectival passives. The
simple and intensive participles denote a state s holding of an individual x, which
may or may not result from a previous event e, whereas the causative participle
includes both quantification over an event e which brings about the state s, and over
a participant y of e. y is understood as different from x, and thus the causative
participle denotes a relation between two individuals. In the examples above, the
state in (a) and (b) is not necessarily the result of an event of change, and, even if it
is, this event does not necessarily involve an additional participant. In (c), the state is
the result of an event including an additional participant.
We thus find in Hebrew a classification of adjectival passive participles which
includes a class additional to the two classes recognized by Embick 2004a. Embick
distinguished among adjectival passives in English between stative and resultative
participles. Stative participles describe basic states and are directly derived from the
root by an adjectival aspectual head Asp, without an intermediary verbalizing node.
Resultative participles denote states that are results of the events which bring them
about, and include a V node. An adjectival passive in English thus has one of two
structures: [Asp √Root] or [Asp [V √Root]], corresponding to the stative and
resultative readings respectively. For example, the adjectival passive closed can be
constructed in two different ways, depending on whether it denotes a basic state, or
the result state of an event of closing. In neither case is Voice, the verbal head which
introduces external arguments, included in the derivation. Therefore in neither
structure is the external argument part of the interpretation of the adjectival
participle (but see Anagnostopoulou 2003, Meltzer 2006, 2011 for a different view).
The interaction of adjectival passive and Voice
167
In Hebrew, stative and resultative readings are found with each of the simple and
intensive adjectival passives. But the situation is different for causative adjectival
passives, though they too have two readings. One is the stative reading, similarly to
the other adjectival passives. The other is the resultative reading; but in the case of a
causative participle, its structure is different from the resultative structure of the
other adjectival passives, in that it obligatorily includes the external argument. In this
chapter I would like to offer an account for this phenomenon. In particular, I start by
showing in the next section that this property of causative participles is not due to
their simply being participial forms of passive verbs. There exist causative adjectival
passives, as already demonstrated above, yet their resultative structure includes the
external argument.
8.3 Adjectival vs. verbal passive participles
In Hebrew adjectival passive participles can be distinguished from verbal passive
participles (Doron 1999). This contrast is discussed in the present section. In
subsequent sections, verbal passive participles will be set aside, since the puzzle
that constitutes the topic of the chapter relates to adjectival passive participles. The
aim of this section is to show that the puzzling behavior of causative passive
participles is not an indication that they are only categorized as verbal.
Verbal passive participles form part of the verbal system of Hebrew. In general, to
each active transitive verb there corresponds a non-active verb, but non-active does
not necessarily mean passive. Non-active variants are often middle-voice verbs (as
elaborated in section 8.5 of the chapter). In principle, there are two non-active verbal
templates, passive (pass) and middle (mid), corresponding to each active (act)
verbal template. Some active verbs have two corresponding non-active verbs, one
in the passive template and the other in the middle template. Intensive-template
verbs (intns) usually have both a passive and a middle corresponding verb. But
other active verbs only correspond to a single non-active verb. Simple-template
verbs (simpl) have a corresponding middle-voice verb but no passive verb. Causative-template verbs (caus), on the other hand, do not have a corresponding middlevoice verb, only a passive verb. This fact about causative verbs will turn out to be
crucial for the solution to the puzzle formulated above.
We begin with examples in the simple template, where there is no corresponding
passive verb, and in particular no participial forms of passive verbs. It follows that
the simple template passive participle is adjectival only (A), not verbal (V). The only
non-active verbal participle of such verbs is the middle-voice participle (simpl-midpart):
(6)
a. ha-šá’ar sagurS
/ nisgar
the-gate closed (A)/ close-simpl-mid-part (V)
‘The gate is closed/is closing.’
168
Edit Doron
b. ha-xalon
šaburS
/ nišbar.
the-window
broken (A) / break-simpl-mid-part (V)
‘The window is broken/is breaking.’
In the intensive and causative templates, each adjectival passive participle related to
a transitive active verb is typically also categorized as a verbal passive participle,
intns-pass-part and caus-pass-part respectively:
(7)
a. ha-mismak
mešuxzarI.
the-document reconstructed (A)/reconstruct-intns-pass-part (V)
‘The document is reconstructed (A)/is being reconstructed.’
b. ha-kise
muzazC.
the-chair moved (A) / move-caus-pass-part (V)
‘The chair is moved (A)/is being moved.’
The distinction between verbal and adjectival passive participles holds for intensive
and causative participles as well, though verbal and adjectival passives are homophonous. This can be demonstrated by inspecting environments which distinguish
adjectival from verbal participles (Doron 1999). One is the future tense copula yihye
‘will be’, which takes A but not V complements. The contrast is illustrated in (8) for
the simple template examples where the adjectival and verbal participles have
different forms. Only the adjectival participle appears in this environment. A verbal
participle cannot appear there (rather, the verb would have to be temporally
inflected):
(8)
be-ša’a xameš ha-šá’ar yihye
sagurS /
*nisgar.
at-hour five
the-gate will-be closed (A)/ *close-simpl-mid-part (V)
‘At five the gate will be closed (A)/will be closing.’
Similarly, the verb nir’e ‘seem’ and the aspectual verb notar ‘remain’ take A but not
V complements:
(9)
/ * nisgar.
ha-šá’ar nir’e / notar
sagurS
the-gate seems / remains closed (A) / *close-simpl-mid-part (V)
‘The gate seems/remains closed/*closing.’
In the context of of kbar ‘already’ + temporal interval, adjectival and verbal participles are interpreted differently; an adjectival participle is interpreted as resultative,
whereas a verbal participle is interpreted as progressive:
(10)
sagurS
’éser daqot.
a. ha-šá’ar kbar
the-gate already closed (A) ten
minutes
‘The gate has been closed for ten minutes now.’
The interaction of adjectival passive and Voice
169
b. ha-šá’ar kbar
nisgar
’éser daqot.
the-gate already close-simpl-mid-part ten
minutes
‘The gate has been in the process of closing for ten minutes now.’
These contexts also distinguish between homophonous verbal and adjectival
participles like the ones in (7) above. Since these participles can appear in adjectival
environments such as the complement of ‘remain’, ‘seem’ and the future tense
copula yihye, and receive a resultative interpretation in the context of kbar ‘already’
+ temporal interval, in addition to the progressive interpretation, it follows that
intensive and causative passive participles are adjectival in addition to being passive
verbs:
(11) Intensive passive participle
notar/nir’e/yihye
mešuxzarI.
a. ha-mismak
the-document remains/seems/will-be reconstructed (A)
‘The document remains/seems/will be reconstructed (A).’
b. ha-mismak mešuxzarI
the-document reconstructed (A)/reconstruct-intns-pass-part (V)
kbar šbu’-áyim.
already week-dual
‘The document has been reconstructed (A)/(V) for two weeks now.’
(12) Causative passive participle
a. ha-kise notar/nir’e/yihye
muzazC.
the chair remains/seems/will-be moved (A)
‘The chair remains/seems/will be moved (i.e., to a different location).’
b. ha-kise muzazC
kbar éser daqot.
the-chair moved (A)/move-caus-pass-part (V) already ten minutes
‘The chair has been moved (A)/(V) for ten minutes now.’
This section has shown that causative passive participles, as well as the other
participles, can be categorized as adjectives. In subsequent sections, verbal passive
participles will be set aside, and we will only be interested in adjectival passive
participles.
8.4 The interpretation of passive participles
8.4.1 Stative passive participles
Having set aside verbal passive participles, we must also set aside adjectival passive
participles which are basic adjectives. Though morphologically constructed with
participial exponents, these basic adjectives are not understood as the result of a
dynamic event, even when they correspond to dynamic verbs. Following Embick
170
Edit Doron
(2004a), I call these stative adjectival passives. Hebrew examples are given below in
all three templates. Note that these participles mostly correspond to English nonparticipial adjectives, and that many denote individual-level properties. The examples listed in (13) are all derived from roots which also derive verbs:
(13)
a. S stative passive participles
hadurS ‘elegant’, ka’urS ‘ugly’, xatubS ‘slender’, qabua’S ‘permanent’, xašubS
‘important’, šatuaxS ‘flat’, ’akurS ‘muddy’, ’alubS ‘miserable’, ’acubS ‘sad’,
barurS ‘clear’, atumS ‘opaque’, qalušS ‘sparse’, spur-imS ‘few-pl’, sabukS
‘entangled’, sadurS ‘ordered’
b. I stative passive participles
mehudarI ‘elegant’, meku’arI ‘ugly’, me’uxarI ‘late’, meluklakI ‘dirty’, mesurbalI ‘clumsy’, mesugalI ‘able’, mecuyanI ‘excellent’, meruxaq ‘distant’, mexuspasI ‘rugged’, metumtamI ‘imbecile’, meruvaxI ‘spacious’, mešuneI
‘strange’, meyuxadI ‘special’
c. C stative passive participles
muqdamC ‘early’, muladC ‘innate’, mupšatC ‘abstract’, mubhaqC ‘distinct’,
muplagC ‘extreme’, mupra’C ‘deranged’, mubxarC ‘select’, muxlatC ‘total’,
mukanC ‘ready’, mušlamC ‘perfect’, mušxatC ‘corrupt’, murkavC ‘complicated’, mukšarC ‘talented’
One indication for viewing participles such as the ones in (13) as basic adjectives
not including any verbal component in their derivation is that they are incompatible
with adverbs that modify events and can be used as modifiers of resultative participles, such as biqpida ‘carefully’, berišul ‘carelessly’, bexipazon ‘hastily’. This type of
test was suggested by Kratzer (1994).
(14) a. *te’un
barurS biqpida
argument clear
carefully
* ‘a carefully clear argument’
b. * béged
meku’arS berišul
garment ugly
carelessly
* ‘a carelessly ugly garment’
c. * pirxax mupra’S bexipazon
brat
deranged hastily
* ‘a hastily deranged brat’
Second, each of these participles is compatible with the claim that it denotes a
state that has not been brought about by an event, but is a lifetime property. (15)
illustrates this point:
The interaction of adjectival passive and Voice
(15)
171
a. ezor
ha-yá’ar haya tamid sabukS
be-’ecim.
(the-)area (of ) the-forest was always entangled with-trees
‘The forest area has always been entangled with trees.’
b. ha-síax
haze camax meruxaqI me-ha-gader.
the-shrub this grew distant
from-the-fence
‘This shrub grew at a distance from the fence.’
c. xataltul nolad mukanC l-a-xayim.
kitten (is) born ready
for-the-life
‘A kitten is born ready for life.’
Moreover, many stative participles are derived from roots which do not derive
matching verbs in the same template, if they derive verbs at all. Some roots only
derive nouns and adjectives in addition to participles, and some uniquely derive
participles and their nominalizations and thus are clearly not constructed from
verbs:
(16)
Participle
a. gašumS ‘rainy’
maluaxS ‘salty’
garumS ‘bony’
xasudS ‘pious’
xarucS ‘diligent’
Equi-rooted lexical items
(equi-templatic verbs do not exist)
géšem ‘rain’ N
mélax ‘salt’ N
gérem ‘bone’ N
xésed ‘charity, grace’ N
xaricut ‘diligence’ N
b. memušqapI ‘bespectacled’
mezuqanI ‘bearded’
menumarI ‘spotted’
me’ušarI ‘happy’
memušma’I ‘obedient’
mišqap-áyim ‘glasses-dual’ N
zaqan ‘beard’ N
namer ‘leopard’ N
’óšer ‘happiness’ N
šama’ ‘hear-simpl-act’ V
c. mušlagC ‘snowy’
mugazC ‘carbonated’
mudlaqC ‘infected’
mubtalC ‘unemployed’
mucaqC ‘solid’
šéleg ‘snow’ N
gaz ‘gas’ N
daléqet ‘infection’ N
batel ‘idle’ A
yacaq ‘cast-simpl-act’ V
Finally, the choice of template for many of these participles is arbitrary. There are
many pairs of participles, some shown in (13), consisting of the same root with two
different participial templates but no discernible meaning difference, such as ka’urS/
meku’arI both meaning ‘ugly’, rabua’S/meruba’I both meaning ‘square’, xacupS/
mexucapI both meaning ‘impertinent’, hagunS/mehuganI both meaning ‘decent’,
meturapI/mutrapC both meaning ‘nutty’, etc.
This section discussed stative participles, which are basic adjectives not involving
any verbal structure. These participles are irrelevant to the puzzle distinguishing
172
Edit Doron
causative from other participles, since the puzzle concerns participles which include
an eventive component as part of their interpretation.
8.4.2 Resultative passive participles
Unlike the participles discussed in the previous section, there are participles whose
meaning involves the existence of an event that brought about the state denoted by
the participle. Such participles are typically compatible with the adverbs biqpida
‘carefully’ and berišul ‘carelessly’ which modify events, indicating that their structure
includes V. Following Embick 2004a, I call these resultative participles.
Resultative participles are the ones relevant to the puzzle formulated in section 8.2:
whereas the external argument is not part of the structure of the simple/intensive
participles, it is in the case of causative participles. This contrast can only be found in
the resultative interpretation of passive participles, where they denote states typically
resulting from an event.2
8.4.2.1 Simple and intensive resultatives
intensive participles only:
(17)
We start by considering simple and
a. S resultative passive participles
katubS ‘written’, qašurS ‘tied’, šaburS ‘broken’, raxucS ‘washed’, apuyS
‘baked’, qanuyS ‘bought’, makurS ‘sold’, baduqS ‘verified’, daxuyS ‘postponed’, paturS ‘solved’, akulS ‘eaten’, la’usS ‘chewed’, laquaxS ‘taken’, maxucS ‘smashed’, manua’S ‘prevented’, matuaxS ‘tense’, rapuyS ‘loose’, kapupS
‘stooped’, atumS ‘sealed’, parucS ‘breached’, ‘arukS ‘organized’, sagurS
‘closed’, patuaxS ‘open(ed)’
2
The distinction between stative and resultative interpretations is not detectable for adjectival passive
participles corresponding to active verbs which are themselves stative, i.e., where the verb does not denote
a dynamic event to begin with. In this case, the event bringing about the result state is itself a state and is
not distinguished from the result state. I will therefore leave aside in the text adjectival passive participles
related to stative verbs. Such participles can be derived in all three templates, e.g., xasumS ‘blocked’,
mekutarI ‘surrounded’, mustarC ‘hidden’, in parallel to their corresponding active verbs:
(i) a. ha-sela’ xosem
et ha-knisa.
the-rock blocks-simpl-act acc the entrance
et ha-’ir.
b. ha-gba’ot mekatrot
the-hills surround-intns-act acc the city
c. ha-‘ec mastir
et ha-knisa.
the-tree hides-caus-act acc the entrance
Other examples of causative template passive participles derived from stative verbs are mugdar ‘defined’,
mugbal ‘limited’, mubla’ ‘concealed’, mu’arax ‘appreciated’, muncax ‘eternalized’, mugan ‘protected’,
muban ‘understood’, mubxan ‘distinct’, mutar ‘permitted’, mukar ‘familiar’, mu’ar ‘lighted’. These adjectives are relational, and thus include an additional argument. Therefore, even if viewed as resultative, these
causative participles fall under the generalization described in the chapter: causative passive participles
include an argument in addition to the holder of the state.
The interaction of adjectival passive and Voice
173
b. I resultative passive participles
mesudarI ‘arranged’, mesuraqI ‘combed’, megulaxI ‘shaven’, mexumamI
‘heated’, mequrarI ‘cooled’, mebušalI ‘cooked’, mešulamI ‘paid’, meyuda’I
‘informed’, me’umanI ‘trained’, meyucarI ‘manufactured’, meturgamI
‘translated’, metuknanI ‘planned’, mekupapI ‘bent’, metuqanI ‘corrected’,
meluksanI ‘angled’, mexudadI ‘sharpened’, mesubakI ‘complicated’, mepuzarI ‘dispersed’, mequšarI ‘connected’, me’uqalI ‘curved’, mesulsalI ‘curled’,
mequba’I ‘fixed’, mexusanI ‘immune’
Some of these participles also have a stative reading in addition to the resultative
one. But even the resultative reading, which entails a dynamic event, does not entail
the existence of an additional participant within this event, unlike the case of passive
verbs (Kratzer 2000). In each of the examples (18)–(22) below, the state is not
claimed to have been brought about by an additional participant. In (18a), the boy
could have combed himself, whereas the passive verb in (18b) entails that someone
else combed him. In (19a), the air could have cooled on its own, as is explicitly stated
in the attested sentence in (19c), but this is not possible with the passive verb in (19b),
which entails that an external participant cooled the air. The same thing is shown for
warming, registering, washing and dressing in (20)–(22).3
(18)
3
a. ha-yéled
the boy (is)
mesuraqI.
combed
A reviewer reports finding the following sentences acceptable:
(i) a. ha-yeled suraq
’al-yedey ’acmo.
the-boy comb-intns-pass by
himself
‘The boy was combed by himself.’
b. ha-avir qurar
’al-yedey ’acmo.
the-air cool- intns-pass by
himself
‘The air was cooled by itself.’
I find these sentences ungrammatical, though they might slightly improve if himself is focused (cf. Doron
and Rappaport Hovav 2009), as in the following example:
(ii) ?? ha-nibxéret muyna
lo ’al-yedey ha-me’amen ’ela ’al-yedey ’acma.
the team.fem sort-intns-pass not by
the-coach but by
herself
‘The team was sorted out not by the coach but by itself.’
In general, according to Baker, Johnson, and Roberts (1989), there is an implicit argument in the
morphology of passive verbs, which is interpreted as a generic or impersonal pronoun. Thus in (iiia), an
impersonal ‘they’ is said to have sorted the team. This is in contrast with (iiib), where the team is sorted out,
maybe through a process of classification internal to the team:
(iii) a. ha-nibxéret muyna.
the team sort-intns-pass
‘The team was sorted out (V).’
b. ha-nibxéret memuyenetI.
the team (is) sorted
‘The team is sorted out (A).’
174
Edit Doron
b. — /→ ha-yéled suraq.
the boy comb-intns-pass
‘The boy was combed.’
(19)
a. avir mequrarI
air cooled
‘cooled air’
b. — /→ ha-avir qurar.
the air cool-intns-pass
‘The air was cooled.’
c. beyn
še ha-’ananim
nepuxim o daqim, tamid medubar be-gušim
whether that the-clouds (are) puffy
or slim, always it-is-dealt with-blocks
šel avir mequrarI. avir yaxol le-hitqarer
be-’et še-hu mitromem
of air cooled.
air can to-cool-intns-mid when that-it rises
‘Whether the clouds are puffy or slim, one is always dealing with blocks of
cooled air. Air may cool wile rising.’
(20)
a. avir mexumamI
air warmed ‘warmed air’
b. — /→ ha-avir xumam.
the air warm-intns-pass
‘The air was warmed.’
c. cinor ze no’ad
le-haknis l-a-manóa’
avir conen yoter
hose this is-intended to-bring.in to-the-engine air cool more
me-xazit
ha-rékeb ve-lo avir mexumamI
from-(the-)front (of) the-car and-not air warmed
me-ezor
taha-manóa’.
from-(the-)area (of ) (the-) compartment (of ) the-engine
‘This hose is intended to bring into the engine cooler air from the front of
the car and not warmed air from the area of the engine compartment.’
d. be-imun ’im ximum ata
mexumamI ve kašir le-metixot
in training with warm-up you (are) warmed
and fit for-stretches
ve- tnu’ot
xadot.
and-movements sharp
‘In training with a warm-up, one is warmed up and fit for stretches and
sharp movements.’
(21)
l-a-ši’ur.
ha-talmid
rašumS
the student (is) registered for-the-class
(Doesn’t mean he was registered by someone else)
The interaction of adjectival passive and Voice
(22)
175
ha-yeladim
rexucimS ve-lebušimS be-pijámot.
the children (are) washed and dressed in pyjamas
(Doesn’t mean they were washed or dressed by someone else.)
The attested examples in (23)–(26) below, collected from the Internet, are examples where the events which bring about the states denoted by the passive participles
typically do not involve an additional argument:
(23)
’aruk ‘organized’
li-qlot
et ha-mictarfim
ha méšeq
ha-yisra’eli eyno ’arukS
the economy the-Israeli is-not organized to-absorb acc the-joiners
ha-xadašim le-kóax
ha-avoda b-a-šanim ha-ba’ot.
the-new
to-(the-)force (of) the-work in-the-years the-next
‘The Israeli economy is not organized to absorb the new people joining the
work force in the next years.’
(24)
paruc ‘breached’
txum
ha-hipnóza b-a-árec
parucS laxalutin
(the-)area (of) the-hypnosis in-the-country (is) breached completely
b-o.
ve-ap-exad lo mamaš mepaqé’ax ’al ha-’osqim
and no.one not really supervises on the-practicing at-it
‘The field of hypnosis in the country is completely breached and no one really
supervises its practitioners.’
(25)
katub ‘written’
katubS b-a-kokabim: taxazit
šbu’it
written in-the-stars: horoscope weekly
‘Written in the Stars:
Weekly Horoscope’
(26)
qašur ‘tied’
ha-’ubar
qašurS be-xével ha-tabur.
the fetus (is) tied
with the umbilical cord
‘The fetus is tied with the umbilical cord.’
This section has argued that the verb’s external argument is not part of the
interpretation of resultative passive participles in the simple/intensive templates.
8.4.2.2 Causative resultatives
the causative template:
We now consider resultative passive participles in
(27) C resultative passive participles
musdarC ‘arranged’, muzazC ‘moved’, muqpaC ‘frozen’, mumasC ‘melted’,
murtaxC ‘boiled’, musaqC ‘heated’, mupalC ‘thrown down’, mucabC ‘posted’,
mudaxC ‘deposed’, mu’ataqC ‘moved, copied’, mušbatC ‘put on strike’, mucatC
176
Edit Doron
‘ignited’, mušxazC ‘sharpened’, mukxadC ‘extinct’, mušmadC ‘exterminated’,
mukeC ‘beaten’, mupacC ‘scattered’, musatC ‘incited’, mušxatC ‘ravaged’, muxlatC ‘decided’, mukanC ‘prepared’, mušlamC ‘completed’, murkavC ‘combined’,
mukšarC ‘trained’
Here too we find some ambiguous participles, for example mukšarC ‘trained’,
which is also listed among stative participles in section 8.4.1 with the meaning
‘talented’. Focusing on resultative interpretations, causative participles are distinguished from simple and intensive participles in that they obligatorily include an
additional participant in the causing event. in (28a) below, where modification with
biqpida ‘carefully’ selects the resultative reading of the causative participle, it is
understood that the athletes were trained by others rather than having trained on
their own. This is not so for the intensive participle in (28b), where the athletes can
be understood as having carefully trained on their own:
(28)
a. sporta’im mukšarC-im biqpida
athletes trained-pl carefully
‘carefully trained athletes’
b. sporta’im me’umanI-im biqpida
athletes trained-pl
carefully
‘carefully trained athletes’
In each of the examples (29)–(32) below, the causative participle involves an
additional participant other than the holder of the state. In (29), the causative
dressed, unlike the simple dressed, entails that the boy did not dress on his own.
The same holds for the causative separated, silenced, posted in (30)–(32), contrasting
with the corresponding simple/intensive participles, which do not entail an additional participant. In this respect causative resultative participles are like the verbs
illustrated in the last section, but they are demonstrably not verbs.
(29)
a. mulbašC vs. labušS
dressed
dressed
b. ha-yéled
mulbašC → Someone else dressed him
the boy (is) dressed
c. ha-yéled
labušS
the boy (is) dressed
(30)
→ He could have dressed himself
a. mupradC vs. parudS
separated
separated
b. zug
mupradC → Someone/something separated them
couple separated
‘a separated couple’
The interaction of adjectival passive and Voice
177
c. zug
parudS
→ They might have separated of their own accord
couple separated
‘a separated couple’
(31) muštaqC vs. mešutaqI
silenced
paralyzed
(32)
a. mucabC vs. memuneI
posted
appointed
b. ha-mélek
mucabC be-roš-ha-caba
the-king (is) posted at-(the-)head(of)the-army → Someone else posted him
c. ha-mélek
memuneI ’al ha-caba
the-king (is) appointed over the-army → He probably appointed himself
Surprisingly, this is also true for causative participles derived from verbs (many of
which are de-adjectival), where the external argument is optional. These are verbs
where both the causative and the anticausative have the same form, i.e., these verbs
do not necessarily have an external argument in the active voice. Yet the passive
participle entails such an argument:
(33)
a. ha-se’ar hilbin.
the-hair whitened-caus-act
‘The hair whitened.’ (possibly of itself )
b. se’ar mulbanC
hair whitened
‘whitened hair’ → someone/something whitened it (not by itself )
(34)
a. ha-xalab hexmic.
the-milk turn-sour-caus-act
‘The milk turned sour.’ (possibly of itself )
b. xalab muxmacC
milk soured
‘soured milk’ → someone/something turned it sour (not by itself )
(35)
a. ha-lexay-áyim
he’edimu
l-o.
the-cheek-dual redden-caus-act to-him
‘His cheeks reddened.’ (possibly of themselves)
b. lexay-áyim mu’adamotC
cheek-dual reddened
‘reddened cheeks’ → someone/something made them red
(not by themselves)
178
(36)
Edit Doron
a. ha-šqedim gadlu
l-o.
the-tonsils grow-simpl-act to-him
‘His tonsils grew.’ (possibly of themselves)
b. šqedim mugdalimC
tonsils enlarged
‘enlarged tonsils’ → someone/something made them exceed the size
they would reach by themselves4
This section has shown that the verb’s external argument is part of the interpretation of resultative passive participles in the causative template.5
8.5 The root-template system
8.5.1 The architecture of the verbal system
The asymmetry between simple/intensive participles and causative participles
derives from the architecture of the verbal template system of Hebrew. This system
expresses two dimensions of meaning (Doron 2003, 2008). One is agency—the
thematic role of the external argument. The marked templates (the causative and
the intensive) express the thematic role of the verb’s external argument: cause and
agent respectively. The simple template functions as default and is neutral as to the
external argument’s role.6
The agency dimension of the template system is illustrated by the table below,
where the different active templates are shown (in boldface) intertwined with the
(originally Greek) root √spg ‘sponge’.
4
There is group of adjectives like mugdalC ‘enlarged’: muqtanC ‘reduced’, mugbarC ‘increased’, mu’acC
‘accelerated’, murxabC ‘extended’, mu’arakC ‘elongated’, etc., which mean ‘above-the-norm’. These adjectives denote the target state of an event with an external cause responsible for this above-the-norm result.
The event is not necessarily a dynamic event of change over time, but of change from the expected to the
actual value of a given dimension. It is reminiscent of the “derived stative” interpretation of adjectival
passives discussed in Koontz-Garboden 2010, whereby events develop not in the temporal dimension
but in a spatial dimension. Here the dimension of change is spatial in that it spans different specimens
of a kind.
5
There are resultative participles predicated of the active subject (Blau 1954, Doron 1999): muclaxC
‘successful’, mupnamC ‘introverted’, murvaxC ‘has-gained’, mupsadC ‘has-lost’, mupla’C ‘mysterious’,
qasumC ‘enchanted’, mumarC ‘convert’. Also masurS ‘devoted’, mexušabI ‘calculated’, mesukanI ‘dangerous’. I leave aside these examples of voice reversal.
6
Many roots derive singleton verbs, verbs which do not contrast with an equi-rooted verb in another
template or with an equi-rooted noun/adjective. There is no contrast associated with such roots, and no
meaning emerges in these cases; the template is sometimes arbitrary, and often dictated by phonological
considerations. But though not every verb in the causative template is causative, it is nevertheless the case
that in every alternating pair of equi-rooted verbs, it is the causative-template verb which is the causative
counterpart of the simple-template verb, and this is never reversed. Equally, every intensive-template verb
alternating with an equi-rooted simple-template verb is agentive, and this is never reversed. Thus, the
verbal template system expresses meaning where there is contrast.
The interaction of adjectival passive and Voice
(37)
179
√spg ‘sponge’
Simple
Intensive
Causative
Active Voice sapag ‘absorb’ sipeg ‘dab, swab, dry hispig ‘impregnate,
by sponging up’
make absorb’
The following sentences serve to demonstrate the agentive nature of the subject of
the intensive verb in (38a) versus the causative nature of the subject of the causative
verb in (38c). (38b) is ungrammatical since abstract entities, such as his good
education, are not agentive; (38c) is grammatical since abstract entities can be causes.
(38)
a. aba sipeg
be-matlit et micx-o ha-meyuza’ šel ha-yeled.
father dab-intns with-cloth acc front-his the-sweaty of the-boy
‘Father dried up the boy’s sweaty forehead with a cloth.’
ha-tob sipeg
et micx-o ha-meyuza’ šel ha-yeled.
b. *xinuk-o
education-his the good dab-intns acc front-his the-sweaty of the-boy
‘His good education dried up the boy’s sweaty forehead.’
ha-tob hispig
et ha-yeled be-’arakim.
c. xinuk-o
education-his the good impregnate-caus acc the-boy with-values
‘His good education impregnated the boy with values.’
An additional example is constructed with the root √šlt ‘control’. Verbs in
Modern Hebrew are not exclusively derived from bare roots, but are sometimes
derived from nouns or adjectives (Arad 2003). An example is the intensive verb
within the table (39) below, which, though ultimately derived from the root √šlt
‘control’ like the simple and causative verbs in the same table, is not derived like the
latter from the bare root, but rather indirectly, after the root has been categorized as
a noun. The intensive verb is thus in fact derived from the noun šélet ‘sign post’:
(39)
√šlt ‘control’
Simple
Intensive
Causative
Active Voice šalat ‘control’ šilet ‘fit with sign posts’ hišlit ‘impose’
The agentive nature of the subject of the intensive verb is illustrated by (40b), in
contrast with the non-agentive nature of the subject of the simple verb in (40a)
(which is a stative verb, hence non-agentive) and of the causative verb in (40c):
(40)
a. xóser séder šalat
b-a-rexobot.
lack order control-simpl in-the-streets
‘Disorder ruled the streets.’
180
Edit Doron
b. * xóser séder šilet
et
ha-rexobot.
disorder
fit-with-sign-posts-intns acc the-streets
‘Disorder fitted the streets with sign posts.’
c. xóser séder hišlit
páxad b-a-rexobot.
lack
order impose-caus fear
in-the-streets
‘Disorder imposed fear in the streets.’
The agency dimension can be schematically summarized as in (41). The external
argument of each verb is not the argument of the root but of a little-v head (Marantz
1984, Kratzer 1994, Borer 2003, Alexiadou, Anagnostopoulou, and Schäfer 2006)
which assigns a thematic role depending on its syntactic environment: Agent in
the environment of the agency head ι (normally realized as the morphological
exponent intns, the intensive template), Cause in the environment of the agency
head γ (normally morphologically realized as the caus template), and undetermined
elsewhere:
(41)
v
agency-head
environment
a. v/ _ ι
b. v/ _ γ
c. v
denotation of v
λyλe[Agent(e,y)]
λyλe[Cause(e,y)]
λyλe[Agent/Caus (e,y)]
default exponent
of agency-head
intns
caus
simpl
Beyond the three active templates, the rest of the verbal system expresses alternations of voice (diathesis). To each active template there correspond in principle
two non-active templates: passive and middle. As already mentioned in section 8.3
above, some of the active templates only have a single non-active corresponding
template, which accounts for the fact that the total number of Hebrew templates is
only seven. Whereas intensive-template verbs have in general both a passive and a
middle corresponding verb, simple-template verbs have a corresponding middle but
no passive, and causative-template verbs—a passive but no middle. The non-active
templates are illustrated in (42) and (43) below with the roots √spg ‘sponge’ and √šlt
‘control’ by respectively expanding the tables in (38) and (39) along the voice
dimension. Note that there are also lexical gaps, e.g., there are no attested examples
of an intensive passive verb derived from the root √spg ‘sponge’, for no apparent
good reason.7
7
The meaning of the unattested intensive passive verb is expressed by the causative passive verb; the
following is an attested example. Even in the active voice, the causative verb is sometimes used instead of
the intensive verb, which is considered literary.
(i)
kétem
ha-soler ’al ha-kbiš
huspag
be-xómer
meyuxadI.
(the-)stain (of ) the-fuel
on the road dab-caus-pass with-substance special
‘The fuel stain on the road was dabbed with a special substance.’
181
The interaction of adjectival passive and Voice
(42)
√spg ‘sponge’
Voice
Simple
Active
sapag ‘absorb’ sipeg ‘dab, swab,
sponge up’
Intensive
Causative
hispig ‘impregnate,
absorb’
make
Passive —
—
huspag ‘be impregnated, be
made to absorb’
Middle nispag ‘be
absorbed’
histapeg ‘dab
oneself ’
—
(43) √šlt ‘control, govern’
Voice
Simple
Active
šalat ‘control’ šilet ‘fit with sign posts’
Intensive
Causative
hišlit ‘impose’
Passive —
šulat ‘be fitted with signposts’ hušlat ‘be imposed’
Middle nišlat ‘be
controlled’
hištalet ‘impose oneself ’
—
All verbs derived by the non-active templates are intransitive. But while the
external argument can be totally obliterated in the derivation of the middle verb, it
always implicitly participates in the derivation of the passive verb, and may be
expressed explicitly as an al-yedey ‘by’ phrase. In middle voice verbs, the external
argument is often totally missing, but it may be implicit, or identified with the
internal argument. This optionality gives rise to a variety of interpretations for the
middle voice, which have been discussed in the typological literature (Klaiman 1991,
Kemmer 1993). These interpretations are the ones found for the Hebrew middle
templates (and the Greek middle verbs) as well: anticausative, reflexive/reciprocal,
medio-passive, dispositional, behavioral (Alexiadou and Doron 2012, Alexiadou this
volume).8
Voice can be syntactically expressed as a feature [±Voice] of the head V
which verbalizes the root: V[+Voice] is either active or non-active, whereas V[ Voice]
8
It has been claimed that the template system also marks aspectuality, since in some cases, middle
verbs are the inchoative (punctual) counterpart of unbounded (atelic) active verbs, e.g., hityašev ‘sit-down.
mid’ vs. yašav ‘sit.act’ (Arad 2005, Schwarzwald 2008). I do not believe the system marks aspectuality, as
the aspectual contrast is reversed in other examples, where it is the active verb which is punctual, and the
middle verb which is atelic, e.g., halak ‘leave.act’ (also ‘walk’) vs. hithalek ‘walk-around.mid’, xala ‘fall-ill.
act’ vs. hitxala ‘pretend-to-be-ill.mid’, yabaš ‘turn-dry.act’ vs. hityabeš ‘be-in-the process-of-drying.mid’
(also ‘turn-dry’), exer ‘arrive-late.act’ (also ‘be-late’) vs. hit’axer ‘be-late.mid’; or both are atelic e.g., samak
‘rely.act’ and histamek ‘base-oneself.mid’; or both puctual, e.g., ’acar ‘stop.act’ and ne’ecar ‘stop.mid’.
Thus, aspectual contrasts vary in their direction, if at all present, and are reducible to contrasts in
agentivity and in general to the thematic distinctions expressed by the template system.
182
Edit Doron
is non-active. Thus passive verbs are classified with active verbs as [+Voice] and with
middle verbs as non-active, ie [ Active]:
(44)
+ Voice
+ Active
Active
– Active
Passive
– Active
+ Voice
Passive
– Voice
Middle
There are only three possible combinations of [±Active ±Voice], since [+Active]
entails [+Voice]. These three combinations are encoded by the three-valued voice
dimension of the template sytem. The exponent of V[+Act+Voice] is act, the active
templates, the exponent of V[ Act Voice] is mid, the middle templates, and the
exponent of V[ Act+Voice] is pass, the passive templates.9
The implicit external argument in the interpretation of resultative causative participles, in contrast to the lack of such an argument in simple/intensive participles, is
reducible to the nature of the verbalizing head of the different templates. In the case
of the verbalizing head of the causative template (the agency head γ), non-active voice
is realized as [ Active +Voice], a passive-voice head which always requires v, i.e.,
which always introduces an (implicit) external argument. In the case of the verbalizing head V of the other templates, non-active voice can be realized as [ Active
Voice], a middle-voice head which does not requires v, i.e., which does not
introduce an external argument. This will be elaborated in section 8.6 below.
8.5.2 Sample verbal derivations
8.5.2.1 Simple and intensive templates I assume the Distributed Morphology
approach to word structure (following Halle and Marantz 1993, Hale and Keyser
1993, Marantz 1997 and subsequent work). I am also making the assumption that
roots denote properties or relations of objects and eventualities (including events
and states). In particular, this means that roots have arguments. For example, of the
two arguments of the transitive verb daxap ‘push’ in (45) below, one is the argument
of the root √dxp (the “internal argument”), and the other (the “external argument”)
is a Cause argument introduced by the [+Active] voice. Similarly for the verb ’aqap
‘overtake’ in (46), where the external argument is an Agent.
The verbs in (45)–(47) are transitive verbs with no corresponding adjectival
passives. These are verbs which do not entail the existence of a target state (Parsons
1990, Kratzer 2000). Rather, the root relates an object x to a dynamic event e, with no
specified target state. As argued in Doron 1999, a target state is a necessary condition
9
In previous work, I have used the notation μ for V[
notation π for V[ Act+Voice], the passive-voice head.
Act Voice],
the middle-voice head, and the
The interaction of adjectival passive and Voice
183
for the derivation of adjectival passives in Hebrew. Thus these roots do not derive
adjectival passives.
(45)
Simple Template
z daxap
et-x
z push-simpl acc-x
‘z pushed x’
∗ daxup ‘pushed’
S
v
v
z
λz λe [Cause(e,z)]
λe [push(e, x) & Cause(e,z)]
λz λe [push(e, x) & Cause(e,z)]
v10
V
V
[+Active]
λe [push(e,x)]
R
x
(46)
λe [push(e,x)]
√dxp
Simple Template
z ’aqap
et-x
z overtake-simpl acc-x
‘z overtook x’
v
λz λe [Agent(e,z)]
∗ ’aqupS ‘overtaken’
λe [overtake(e,x) & Agent(e,z)]
v
z
v
λx λe [push(e,x)]
λz λe [overtake(e, x) & Agent(e,z)]
V
V
[+Active]
λe [overtake(e,x)]
R
x
λe [overtake(e,x)]
√‘qp
λx λe [overtake(e,x)]
As argued in Doron 2003, intensive verbs are verbs where the root is modified by
the agency head ι. Roots modified by ι only co-occur with v that introduces Agents
rather than Causes. Modified roots not including a target state do not derive an
adjectival passive, similarly to unmodified roots. Thus, the passive participle mexuzarI ‘courted’ corresponding to the verb xizer ‘court’ in (47) below only has a verbal
passive interpretation (i.e., is only interpreted progressively) but no adjectival interpretation.
10
Following Kratzer, functional heads do not combine with their complements by the usual mode of
function application, but by a different mode she calls “identification”. For example, identification takes
place in (45) in the subtree where v and V are combined, by applying λPλzλe[v(e,z) & P(e)] to the
denotation of V.
184
Edit Doron
(47)
Intensive Template
z xizer
axare-x
z courted-intns after-x
‘z courted x’
v
mexuzari ‘courted-intns-pass-part’ (V)
∗ mexuzari ‘courted’ (A)
λe [court(e,x) & Agent(e,z)]
v
z
λz λe [court(e,x) & Agent(e,z)]
v
λz λe [Agent(e,z)]
V
λe [court(e,x)]
V
[+Active]
ι
λe [court(e,x)]
ι
x
ι
λx λe [court(e,x)]
√xzr
Verbs with a corresponding adjectival passive, such as the verbs in (48)–(51)
below, are those verbs which have a target state as part of the interpretation of the
root (or the root modifed by ι). There are two cases: one where the root is dynamic,
yet, unlike the dynamic roots above, includes a target state, as in (48)–(49). These
verbs will have a resultative but not a stative adjectival passive. Roots that are stative
often allow the construction of stative in addition to resultative participles, as in
(50)–(51).
In examples (48) and (49), the dynamic event resulting in the target state is part of
the root’s interpretation.11 In such cases, the adjectival passive is resultative only:
(48)
Simple Template
z makar
et-x
z sell-simpl acc-x
‘z sold x’
makurS ‘sold’
v
v
z
λz λe [Agent(e,z)]
λeλs [Target(e,s) & sell(e, x) & Agent(e,z)]
v
λzλeλs [Target(e,s) & sell(e, x) & Agent(e,z)]
V
V
[+Active]
λeλs [Target(e,s) & sell(e, x)]
R
x
λeλs [Target(e,s) & sell(e,x)]
√mkr λxλeλs [Target(e,s) & sell(e,x)]
11
We thus agree with the view of Levinson (this volume), Roßdeutscher (this volume) according to
which roots may lexicalize an event together with its result component.
The interaction of adjectival passive and Voice
(49) Intensive Template
z cilem
et-x
z photograph-intns acc-x
‘z photographed x’
v
meculami ‘photographed’
λeλs [Target(e,s) & photograph(e,x) & Agent(e,z)]
v
z
185
λzλeλs [Target(e,s) & photograph(e,x) & Agent(e,z)]
λz λe [Agent(e,z)] v
V
λeλs [Target(e,s) & photograph(e,x)]
ι
V
[+Active]
λeλs [Target(e,s) & photograph(e,x)]
ι
x
λxλeλs [Target(e,s) & photograph(e,x)]
ι12
√clm λxλeλs [Target(e,s) & photograph(e,x)]
In contrast, in examples (50)–(51), the dynamic event resulting in the target state is
not part of the interpretation of the root/noun from which the verb is derived, but
part of the verbalizing head V. In such cases, the adjectival passive does not
necessarily entail the event, and in many examples has both a resultative and a
stative interpretation.
(50)
Simple Template
z patax
et-x
z open-simpl acc-x
‘z opened x’
patuaxS ‘open(ed)’
v
v
z
λz λe [Cause(e,z)]
λeλs [Target(e,s)]
λeλs [Target(e,s) & open(s, x) & Cause(e,z)]
λzλeλs [Target(e,s) & open(s, x) & Cause(e,z)]
v
V
V
[+Active]
λeλs [Target(e,s) & open(s, x)]
R
x
λs [open(s,x)]
√ptx λxλs [open(s,x)]
12
In order not to complicate the verbal structures, the interpretation λe[Action (e)] of the agency head
ι is left out in subsequent structures. The effect of ι on the structure is here reduced to its co-occurrence
with v that introduces Agents rather than Causes.
186
Edit Doron
(51) Intensive Template
z šilet
et-x
z fit-with-signposts-intns acc-x
‘z fitted x with signposts’
v
z
mešulati ‘has (been fitted with) signposts’
λeλs [Target(e,s) & have(s, ∩λu[signpost(u)], x) & Agent(e,z)]
v
λzλeλs [Target(e,s) & have(s, ∩λu[signpost(u)], x) & Agent(e,z)]
λz λe [Agent(e,z)] v
λeλs [Target(e,s) & have(s, ∩λu[signpost(u)], x)]
V
λs [have(s, ∩λu[signpost(u)], x)]
ι
λeλs [Target(e,s)]
V
[+Active]
ι
x
λxλs [have(s, ∩λu[signpost(u)], x)]
ι
λPλxλs [have(s, ∩P, x)]
λxλs [have(s, ∩λu[signpost(u)], x)]
A
N
A13
N
λu[signpost(u)]
√šlt
λxλyλs [control(s,y,x)]
8.5.2.2 Causative templates In causative verbs, the verbalizing head which
combines with the root and its arguments is γ, typically realized by the
exponent caus. As with the simple and intensive templates, if the root does not
include a target state, no adjectival passive is derived. For example, the passive
participle munšamC ‘respirated’, derived from the dynamic root √nšm ‘breathe’,
only has a verbal passive interpretation (is only interpreted progressively) and
does not have an adjectival interpretation. Dynamic roots which include a target
state, like √qny ‘buy’, derive a resultative adjectival passive, in this case muqneC
‘imparted’, parallel to the simple resultative qanuyS ‘bought’. An example with a
stative root is shown in (52).
13
As shown in Doron (2003), the intensive and causative agency heads ι and γ have canonical
interpretations when combined with nominalized roots, i.e., when functioning as “outer” rather than
“inner” morphology in the terms of Anagnostopoulou and Samioti (this volume). Intensive template verbs
are canonically interpreted as verbs of putting, i.e., an action resulting in x having N, and thus they include
the property A: “having N” derived from N. The symbol ∩ denotes the nominalization type shift mapping
a property denoted by N to its individual (kind) correlate (see Chierchia 1984). In the denotation of N,
“signpost” is shorthand for “instrument of control” derived by applying the nominalizing head N
(interpreted as the “instrument” type-shift) to the denotation R of the root λxnλe[control(e,xn)]. I assume
“instrument” to be defined as λRλu∃xn∃e [instrument(e, u) & R(e,xn)].
The interaction of adjectival passive and Voice
(52) Causative Template
z hišlit
et-y be-x
z impose-caus acc-y at-x
‘z imposed y on x’
mušlatc ‘imposed’
v
λz λe [Cause(e,z)]
λeλs [Target(e,s) & control(s,y,x) & Cause(e,z)]
v
z
187
λzλeλs [Target(e,s) & control(s,y,x) & Cause(e,z)]
γ
v
λeλs [Target(e,s) & control(s,y,x)]
λeλs [Target(e,s)]
γ
[+Active]
R
λs [control(s,y,x)]
y
R
λxλs [control(s,y,x)]
√šlt λxλyλs [control(s,y,x)]
at-x
For some causative verbs (many of which are de-adjectival), the external argument is optional. These are verbs where both the causative and the anticausative
verbs, e.g., (53) and (54) respectively, have the same form.
(53) Causative Template
z hilbin
et-x
z white-caus acc-x
‘z whitened x’
mulbanc ‘whitened’
v
v
z
λz λe [Cause(e,z)]
λeλs[Target(e,s) & white(s,x) & Cause(e,z)]
λzλeλs [Target(e,s) & white(s,x) & Cause(e,z)]
γ
v
λeλs [Target(e,s) & white(s,x)]
λeλs [Target(e,s)]
γ
[+Active]
R
x
λs [white(s,x)]
A
√lbn
A
(54)
Causative Template
x hilbin
x white-caus
‘x whitened’
λxλs [white(s,x)]
mulbanc ‘whitened’
γ
λe λs [Target(e,s)]
λxλs [white(s,x)]
γ
[+Active]
x
λeλs [Target(e,s) & white(s,x)]
R
λs [white(s,x)]
A
A
λxλs [white(s,x)]
√lbn
λxλs [white(s,x)]
188
Edit Doron
8.6 The derivation of adjectival passives
The adjectival passives corresponding to the simple/intensive verbs in (48)–(51)
above are shown below in (55)–(58). The passive participle structure is categorized
as A. Semantically, A is interpreted as an aspectual operator existentially binding the
event argument if there is one (Kratzer 2000). Dynamic roots only derive resultative
participles, whereas roots denoting states may derive both stative and resultative
participles.14 Resultative participles are derived from the root by the minimal nonactive structure. Since both simple and intensive templates have a [ Voice Active]
template, this will be the structure derived. The corresponding structure with a
[+Voice Active] template would have an external argument in addition and would
not be the minimal structure. Thus, simple and intensive participles are derived
without the merge of v, i.e., without the external argument found in the verbal
derivation.
(55)
Resultative participle only
x
makurs
‘x(is) sold’
A λs∃e [Target(e,s) & sell(s,x)]
A
λRλs∃e [R(e)(s)]
V
λeλs [Target(e,s) & sell(s,x)]
V
[–Voice]
R
x
(56)
√rxc λxλeλs [Target(e,s) & sell(s,x)]
Resultative participle only
x meculami
‘x is photographed’
A
λRλs∃e [R(e)(s)] A
λs∃e [Target(e,s) & photograph(e,x)]
V
λeλs [Target(e,s) & photograph(e,x)]
ι
V
[–Voice]
λeλs [Target(e,s) & photograph(e,x)]
ι
xι
ι
14
λeλs [Target(e,s) & sell(s,x)]
λxλeλs [Target(e,s) & photograph(e,x)]
√clm
λxλeλs [Target(e,s) & photograph
(e,x)]
A state denotation of the root is not a sufficient condition for the existence of a stative passive
participle. There is no such participle corresponding to (59) below, as a matter of lexical gap, nor to (60),
where it is blocked by the basic adjective laban ‘white’. Moreover, since stative participles are directly
derived from the root without the mediation of an agency-head, the contrast in participial exponent
between S/I/C participles is often arbitrary, as was shown in section 4.1.
The interaction of adjectival passive and Voice
(57)
189
a. Stative participle
patuaxs
open
λxλs [open(s,x)]
A
λxλs [open(s,x)]
√ptx
A
b. Resultative participle
x
patuaxs
‘x (is) opened’
A
V
A
λRλs∃e [R(e)(s)]
V
[–Voice]
λeλs [Target(e,s)]
λs∃e [Target(e,s) & open(s,x)]
λeλs [Target(e,s) & open(s,x)]
R
λs [open(s,x)]
x
λxλs [open(s,x)]
√ptx
(58) a. Stative participle
mešulati
has sign posts
λxλs [have(s, ∩λu[signpost(u)], x)]
A
λPλxλs [have(s, ∩P, x)]
N
A
N
λu[signpost(u)]
√šlt
b. Resultative participle
x mešulati
‘x is fitted with sign posts’
A
λRλs∃e [R(e)(s)] A
λs∃e [Target(e,s) & have(s, ∩λu[signpost(u)], x)]
λeλs [Target(e,s) & have(s, ∩λu[signpost(u)], x)]
V
λs [have(s, ∩λu[signpost(u)], x)]
ι
V
λeλs [Target(e,s)]
[–Voice]
λxλs [have(s, ∩λu[signpost(u)], x)]
ι
x
ι
λPλx∃s [have(s, ∩P, x)]
λxλs [have(s, ∩λu[signpost(u)], x)]
A
N
A
N
λu[signpost(u)]
√šlt
190
Edit Doron
Turning to causative verbs, the verbalizing head γ is inherently [+Voice], since there
is no middle voice in the causative template. As [+Voice Active] requires v (i.e., the
external argument is part of the structure of the passive voice), the external argument is obligatorily part of the derivation of the causative participle, as seen in (59):
(59)
Resultative participle
y
mušlatc be-x
‘y(is) imposed on x’
A
λRλs∃e∃z [R(z)(e)(s)]
λs∃e∃z [Target(e,s) & control(s,y,x) & Cause(e,z)]
v
A
v
λzλe [Cause(e,z)]
λzλeλs [Target(e,s) & control(s,y,x) & Cause(e,z)]
γ
λeλs [Target(e,s) & control(s,y,x)]
λeλs [Target(e,s)]
R
γ
[+Voice –Active]
y
λs [control(s,y,x)]
R
be-x
λyλs [control(s,y,x)]
√šlt
λxλyλs [control(s,y,x)]
Moreover, since [+Voice Active] requires v, then the external argument is part of
the derivation of the causative participle, whether or not it is part of the corresponding [+Voice+Active] verb. We thus account for the contrast between the active verb
in (53)–(54) above, where the external argument is optional, and the resultative
adjectival passive in (60) below, where it is obligatory:
(60) Resultative participle
x
mulbanc
‘x (is) whitened’ (by an external cause)
A
λs∃e∃z [Target(e,s) & white(s,x) & Cause(e,z)]
λRλs∃e∃z [R(z)(e)(s)] A
λzλe [Cause(e,z)]
v
v
λzλeλs [Target(e,s) & white(s,x) & Cause(e,z)]
γ
λeλs [Target(e,s) & white(s,x)]
γ
λeλs [Target(e,s)]
[+Voice –Active]
R
λxλeλs [Target(e,s) & white(s,x)]
A
x
A
λxλs [white(s,x)]
√lbn
λxλs [white(s,x)]
8.7 Conclusion
Resultative passive participles derived by the causative template, unlike those derived
by the simple or intensive template, always include in their interpretation existential
The interaction of adjectival passive and Voice
191
quantification over the external argument of the corresponding active verb (even if
the external argument is optional in the representation of the verb). This correlates
with an additional distinction between the causative template and the other
templates: whereas simple and intensive verbs have middle-voice forms, the only
non-active form of causative-template verbs is the passive, a voice which obligatorily
introduces an (implicit) external argument. The correlation between these two
properties indicates that resultative participles are derived from their roots in a
verbal structure which also includes the specification of voice. The simpler structures
proposed by Kratzer (1994) and Embick (2004a), which do not include a Voice head,
cannot account for this correlation.
9
Roots and phases
ÁNGEL J. GALLEGO
9.1 Introduction
A long-standing assumption within GB-based projectionist approaches is that lexical
items (e.g., destroy and destruction) sharing the same root (e.g., √(DE)STROY)
participate in a process of argument structure inheritance whereby arguments of
the verb are kept in nominal counterparts as part of the lexical, root-bound,
information.* This fruitful line of research stemming from Chomsky (1970) has
been developed in some form or another by Alexiadou (2001, 2009), Borer (2005,
2009), Grimshaw (1990), Harley (1995, 2004, 2009a, 2009b), Marantz (1997 et seq.),
Picallo (1991), Ramchand (2008), and many others. Whereas all the variants of this
approach share the idea that roots are category-neutral elements that encode noncompositional (i.e., atomic) and encyclopedic information, they differ with respect to
whether these units can encode other types of information. Here two main views can
be distinguished, depending on how empty (or bare) roots are:
(1) √ROOT typology
√ROOT
partially bare
totally bare
category
argument
structure
semantic
type
conceptual
content
no
yes
yes
yes
no
no
yes
yes
no
no
no
yes
* The material in this chapter was presented at the conference Roots: Word formation from the
perspective of “core lexical elements”, held at Universität Stuttgart, Stuttgart (Germany), whose audience
I would like to thank for questions and discussion. For comments to previous versions of this chapter I
would like to thank two anonymous reviewers. Special thanks go to Cedric Boeckx, Ignacio Bosque, Noam
Chomsky, Jaume Mateu, and Juan Uriagereka for discussing the issues raised in this chapter with me. I am
Roots and phases
193
As (1) shows, some authors assume that roots still preserve some argument-taking
properties (they can take internal arguments) and have a semantic type ([±bound],
[±eventive], etc.) associated. This is the view held by Alexiadou (2001, 2009), Harley
(2005, 2009a, 2009b), and Marantz (1997). According to other authors, roots are
totally argument-free elements (both external and internal arguments being introduced by functional heads), but still preserve an inherent semantic denotation, as in
Marantz (2001, 2003, 2005). These two views, which fall under the “partially bare”
label in (1), are illustrated in (2), where the root √(DE)STROY takes an internal
argument (as in (2a)), or does not take arguments at all (as in (2b)):
(2)
vP
a.
v
John
v
vP
b.
v
John
v
√DESTROYP
√DESTROY
DP
the city
v
√STROY
SC
DP
de-
the city
Along with this partially bare views to roots, some authors have also argued that
roots are totally bare, having no grammatically relevant information whatsoever, just
encyclopedic content (see Acedo-Matellán and Mateu 2009, this volume, Borer 2005,
2008, and De Belder and van Craenenbroeck 2011). In such an approach, even the
semantic type of roots would be induced by an external functional category, much
like categorial information.
There are interesting similarities between the framework of roots that the
aforementioned authors have developed and the theory of phases of Chomsky
(2000 et seq.), two of which are particularly salient: (i) category-inducing heads are
phase heads, and (ii) the nature of non-phase heads is determined by the nature
of phase heads. The first similarity concerns the hypothesis (put forward by
Chomsky 2000 and adopted by Marantz 2001; see also Lowenstamm this volume)
that phases have natural semantic and phonological correlates. In the case of
Marantz (2001), this claim is made with respect to the complement of categoryinducing heads (v, n, a), which impose a boundary for morpho-phonological and
semantic interpretation:
also grateful to Artemis Alexiadou for editorial help. Usual disclaimers apply. This research has been
partially supported by grants from from the Ministerio de Ciencia e Innovación (FFI2011-29440-C03-01)
and the Generalitat de Catalunya (2014SGR1013).
194
Ángel J. Gallego
xP
(3)
morpho-phonology
x
√ROOT
semantics
The advantage of this kind of approach is that the (PF and LF) interpretation of the
complements of phase heads is done in isolation—i.e., ignoring the phase head itself.
This predicts certain locality patterns of allosemy and allomorphy (see Marantz 2001
et seq.).
The second point of contact between these two frameworks, which I would like to
focus on in this chapter, alludes to the possibility that all the properties of roots (nonphase heads) derive from phase heads, including category (n, v, etc.), φ-features
(gender, number, person), and argument structure. That category and φ-features are
derived from phase heads is fairly standard in the current literature (see Chomsky
2007, 2008, Richards 2007), but the possibility that argument structure is somehow
parasitic on phase heads has not been, to the best of my knowledge, considered.
In what follows I explore this perspective and argue that it can be made to follow
straighforwardly from the assumptions in Chomsky’s (2000 et seq.) Phase Theory
(see Gallego 2010, 2012). In particular, I will argue (with Marantz 2001, 2003, 2005,
but for independent reasons) that argument structure is projected after the relevant
category-inducing morphemes are merged with roots. Notice that saying that it is v,
n, or a that provide roots with the relevant argument taking properties is reminiscent
of Chomsky’s (2007, 2008) claim that phase heads provide non-phase heads with all
its properties through a process of feature inheritance. A key distinction in Chomsky’s framework—which I will capitalize on—is that feature-inheritance is only
forced in the case of unvalued φ-features. Since by hypothesis only light verbs
contain this type of features (which act as what Chomsky refers to as Probe), only
the roots they dominate will take arguments: “argument”, therefore, is here a
misnomer for a DP whose φ-features can value those of the Probe (what Chomsky
dubs Goal). As the reader can see, this crucially leaves the light category n with no
argument taking properties, as n contains interpretable φ-features, which, by
hypothesis, do not require valuation.1
1
The same analysis carries over to adjectives, if these are analyzed as the combination of a prepositionlike element and a noun (or root), as Mateu (2002) originally argued for. If this is correct, then adjectives—
much like nouns—do not take arguments: It is the P that is part of their l-syntax that does. This hypothesis
Roots and phases
195
I will pursue this formal approach to argument-taking, making a distinction
between “syntactic” and “conceptual” arguments, linking the former to Chomsky’s
Probe–Goal system and the latter to the conceptual content of roots. To be more
explicit, I will take the DP the city to be an argument (a syntactic argument) of
destroy in (4a), but not in (4b) or (4c), as n and a heads in destruct{-ion/-ive} require
no DP Goal for valuation purposes.
(4)
a. John destroyed *(the city).
b. John’s destruction (of the city)
c. John is destructive (to the city)
I hasten to add that I am not arguing that the city cannot be called an ‘argument’ of
the root component shared by destroy, destruction, and destructive. All I am saying is
that it cannot be an argument in the syntax. In other words, I have nothing against
the (widely accepted) view that the DP the city is related to the conceptual content of
the root √(DE)STROY in all cases, but it will only be required in the syntax when it
is placed below a light category endowed with unvalued φ-features. To the extent
that this parallelism between the totally bare view of roots and Chomsky’s Probe–
Goal system can be maintained, it will not only provide us with a better understanding of what a syntactic—not conceptual—argument is, but will also support
Chomsky’s (2007, 2008) Agree, which is not required in the case of functional heads
lacking unvalued φ-features (namely, n).2,3
has independently been pursued by Kayne (2011), who in turn builds on Amritavalli and Jayaseelan (2003)
and treats adjectives as nouns incorporated into a covert Case morpheme. In the context of the present
discussion, one could take P to be the relevant Case morpheme (a K(ase) head, in Bittner and Hale’s 1996
terms).
2
An anonymous reviewer asks me to elaborate on the distinction between conceptual and syntactic
argument. I am afraid that I cannot offer a precise formulation of how such a distinction should be
understood, largely for space limitations. I would nonetheless sketch what I think such a formulation should
emphasize. The notion “argument”, together with that of “predicate”, was inherited from the logic tradition
and reinterpreted in linguistic terms, but since there is no consensus on what a predicate is, there can be no
consensus on what an argument is either. Typically, arguments are intuitively characterized as expressions
that complete the meaning of a predicate, so that their absence gives rise to ungrammaticality. This view is
problematic, since it provides no unbiased strategy to determine what arguments (and adjuncts) are—
everything boils down to the ontology one adopts. Worse yet, this view allows all the burden of a theoretical
move to fall on the controversial notion of “(un)grammaticality”, which covers a wide range of effects, not all
of which have a truly syntactic source (see Chomsky 1955[1975]: ch. 5, 1965: ch. 4, and much subsequent
literature). A more promising way to approach the argument/adjunct distinction is to rely on formal criteria
(not intuitions), such as the need for DPs to receive Case through a feature valuation procedure. It goes
without saying, this may turn out to be insufficient (just one piece of the puzzle), but I believe it puts us in a
better situation to approach the ill-understood argument/adjunct distinction, which, being borrowed from a
non-linguistic tradition, should perhaps be entirely dispensed with.
3
An anonymous reviewer points out that the same argument can be provided from the opposite angle.
Since a DP will always have to satisfy the Case Filter, and this will only happen in the context of a φ-Probe,
then it follows that neither n nor a will be capable of taking arguments—since if they did, their Case would
remain unvalued (unless a(n often empty) preposition was inserted).
196
Ángel J. Gallego
Discussion is divided as follows: Section 9.2 introduces some current approaches
to roots and their relationship with arguments. In section 9.3 I sketch Chomsky’s
(2007, 2008) analysis of cyclic transfer, which is motivated by the presence of
uninterpretable φ-features on v. Section 9.4 puts forward the hypothesis that v and
n differ with respect to argument-taking processes, as only the former requires a
Goal for valuation to take place. Section 9.5 summarizes the main conclusions.
9.2 Roots and arguments
Distributed Morphology and related frameworks that study the interaction among
the lexicon, morphology, and syntax (l-syntax, nanosyntax, and others; see Hale and
Keyser’s 2002, Borer 2005, 2009, Ramchand 2008, or Starke 2009) endorse the
hypothesis that there are two basic units of syntactic combination: (i) functional
elements and (ii) roots (referred to as ‘f-morphemes’ and ‘l-morphemes’ in Harley
and Noyer 1999, 2000). Whereas the former are related to the Case-agreement and
discourse-related systems (roughly, the TP and CP layers, putting aside the precise
status of DPs, DegPs, and PPs) and form closed-clases, the latter are involved in
word-formation processes and form the open-class vocabulary. For operative reasons, I will assume the following definition of roots, taken from Harley (2009a).
Roots carry the non-grammatical, encyclopedic semantic content of a given message. It is
perhaps easiest to think of them as the lexicalization of a pure concept, though their interpretations can vary dependending on the syntactic contexts in which they find themselves, as
in, for example, idioms. It is thus more precise to understand them as instructions to access
certain kinds of semantic information, which may vary dependening on the morphosyntactic
context of the Root in question. (Harley 2009a:131)
It is tempting to relate this take on roots to Embick and Marantz’s (2008) claim that
these units must be categorized before they are transferred to the interpretive components. More precisely, one could regard the little heads as “categorizers” that turn
concepts (non-linguistic units; see (5)) into lexical items (linguistic units; see (6)):4
(5)
√CAT, √LIKE, √SAD, √SING, etc.
(concepts)
xP
(6)
x
(lexical item)
√ROOT
Notice that this point of view is saying that, without the involvement of x heads,
roots cannot participate directly in dependencies of the Merge sort. This is assumed
4
See Chomsky (2001:14). In Boeckx (2009) concepts that are lexicalized are endowed with Chomsky’s
(2007, 2008) “edge feature”. What I am assuming here is that Marantz’s x heads play the role of turning
concepts into units that syntax can manipulate.
Roots and phases
197
in Marantz (2001 et seq.) and Embick and Marantz (2008), where roots are regarded
as modifiers of the semantic type of the relevant little head, creating a head adjunction structure before any argument taking process occurs. This is crucially different
from the view held by Alexiadou (2001, 2009), Alexiadou and Grimshaw (2008),
Harley (2004, 2009a, 2009b), Marantz (1997), and others, where roots can take
internal arguments (represented as α below) before combining with categoryinducing heads. The distinction can be seen in (7):
(7)
vP
a.
v
vP
b.
v
√ROOTP
√ROOT α
Marantz (1997)
v
α
√ROOT
Marantz (2001 et seq.)
What is the reason to adopt (7b) instead of (7a)? In Marantz (2001 et seq.), (7b) is
suggested as part of a proposal where all arguments are severed from the verb
(actually, root), and introduced by applicative like heads (analogous to the “&”
operator of (neo-)Davidsonian approaches to argument structure; see Pietroski
2005). In this scenario, “all composition [can be said to be] syntactic (not “Lexical”)”
(Marantz 2005:1). This shift is in turn tied to the assumption that roots have different
denotations/semantic types:5
(8)
√ROOT denotations
a. States = √CLEAN
b. Manners = √HURRY
c. Entities? = √HAMMER
d. Relations? = √AWAIT
(Marantz 2001:20)
In Marantz (2001 et seq.), only state-denoting roots such as √CLEAN, √OPEN or
√DEAD can merge with a DP in order to form a small clause (see (9a)). Other roots
can directly merge with v, if they denote a manner or entity (see (9b) and (9c)).
What interests us here is that no root can take internal arguments directly (notice,
incidentally, that the DP that merges with √OPEN in (9a) is not its internal
argument, but a subject of predication).
5
See Harley (2005), Levinson (2007b), and references therein for discussion. The idea goes back to
Marantz (1997:216), where the issue whether it is roots or light categories that involve semantic categories/
types is mentioned, but left unsolved.
198
Ángel J. Gallego
vP
(9) a.
vP
b.
v
v
SC
vP
c.
v
SC
DP √OPEN v √PAINT
the door
DP
AP
the wall
red
SC
v √HAMMER DP
AP
the metal flat
A proposal that achieves similar results is Mateu’s (2002) (developed in AcedoMatellán and Mateu this volume). According to Mateu (2002), structure-building
units divide into two classes: “relational” (which can further be eventive or noneventive) and “non-relational” (nouns or roots), only the former being argumenttaking elements. Assuming this key distinction, Mateu (2002) derives lexical categories
as involving the three basic patterns in (10):
(10) a.
X
X
b.
Z
Y
X
c. Z
Y
Y
Z
For readers familiar with Hale and Keyser’s (2002) framework, the relationship
between lexical categories and the templates in (10) should not be odd. Following
these authors, Mateu (2002) argues that the structures in (10) typically correspond
with the categories verb, prepositions, and nouns in languages like English. This
author further claims that (10b) is the structure deployed not only by prepositions,
but adjectives and adverbs as well, which are analyzed as the result of merging a nonrelational element with a relational element. The structure that interests us here is of
course (10c), which Mateu (2002) assimilates to non-relational elements: nouns and
roots (see Acquaviva this volume for discussion).
To recap so far, the most relevant trait of these recent analyses of roots is their
assumption that all arguments are severed from the verb, not just the external ones
(see Chomsky 1995, Kratzer 1996, and related literature).
Let us now consider nominal domains (DPs or nPs). Here authors assume that
arguments are introduced in a similar fashion: They are merged after roots and n
combine. Yet again, things are different in Alexiadou (2001, 2009), Alexiadou and
Grimshaw (2008), and Harley (2005, 2009b), where roots do take internal arguments, which are Case marked in the phonological component by means of
of-insertion, as (11) shows:6
6
This preposition, like structural Case and agreement morphemes, is treated as a “dissociated
morpheme” in Distributed Morphology (see Bobaljik 2008), to be inserted at the morpho-phonological
component.
Roots and phases
199
nP
(11)
n
-ent
√STUDP
√STUD
DP
(of)
DP
Chemistry
In the following pages I outline an approach that aligns with both Marantz’s (2001
et seq.) and Mateu’s (2002) in that internal arguments are not taken by roots. I will
however phrase my analysis in phase-based terms, taking the phase heads to be
crucial in deciding argument-taking properties. I will endorse the idea that the
lexicon is not formed by units that can take arguments and units that cannot
(‘relational’ and ‘non-relational’, to use Mateu’s 2002 terminology). All the elements
present in the lexicon are relational (mergeable) by assumption—because syntax is.
I will thus claim that the selection of internal arguments is forced due to feature
valuation. Before doing so, I need to step back a little bit in order to introduce the
technical assumptions that I will be adopting from Chomsky’s (2007, 2008) framework.
9.3 Phase heads and feature valuation
As noted at the outcome, Marantz’s and Chomsky’s frameworks assume that there
are dedicated syntactic domains that determine the pieces that will be interpreted at
the interfaces, morpho-phonologically and semantically: phases.
xP
(12)
morpho-phonology
x
√ROOT
semantics
For Marantz (2001 et seq.), the domains of interpretation correspond to the complement of all category-inducing heads, whereas for Chomsky only categories bearing
unvalued φ-features can force the system into the Transfer/Spell-Out mode (see
200
Ángel J. Gallego
Chomsky 2007, 2008). I would like to review how the latter system works, as it will
be the basis for the proposal I will put forward in section 9.4.
Chomsky (2000 et seq.) proposes that derivations are chunked down in small
units, the phases, which must be transferred to the interfaces in order to reduce
computational load. Although, as just said, phases are meant to have semantic
and morpho-phonological correspondences (to use Chomsky’s words, phases are
“propositional/independent units”), Chomsky (2001, 2004) defines phases as
domains of uninterpretable φ-feature valuation. Since these features make no
semantic contribution (they enter the syntax in an unvalued fashion), the system
must delete them as soon as they show up. In particular, Chomsky (2004:107, 2005:17,
2007:18, 2008:143) assumes that dedicated lexical items, C and v, contain unvalued φfeatures, which make them act as a seeker (what Chomsky calls Probe) looking for a
matching element (a Goal). Let us consider this in the case of a transitive vP:
(13)
[vP John vφ [VP destroy [DP the city ]]]
Given the structure in (13), the φ-features in v probe and match the internal
argument the city, which is used as a feature value provider. An important question
that arises at this point is when Transfer (Spell-Out) takes place. As Chomsky (2007,
2008) argues, it cannot take place after valuation (for φ-features on v would be
treated as interpretables by the semantic component), and it cannot take place before
valuation either (for the φ-features would remain unvalued). Chomsky thus concludes that valuation (and deletion) of φ-features take place when Transfer does. Let
us state this as in (14):
(14)
Valuation (and deletion) of φ-features is part of Transfer/Spell-Out
An important aspect of Chomsky’s derivation by phase model, also present in
Marantz (2001 et seq.), is that Transfer only targets the complement domain of
phases, leaving the edge intact (but see Gallego 2012 and Ott 2011b for discussion).
However, if the transfer zone includes the domain where the φ-features are, then
note that transfer should affect the whole vP, which is headed by v, the head that
contains the φ-features.
In order to solve this problem, Richards (2007) argues that the mechanism of
feature inheritance put forward in Chomsky (2008), whereby the φ-features on v are
passed down to V (or root), is used precisely to avoid a crash at the semantic
interface.7 If φ-features did not go down to the VP/√P domain, they would remain
in v, where they could indeed be valued, but they would then be regarded as
interpretable in the next phase. The entire process is as indicated below (the same
7
Chomsky’s (2008) original motivation to postulate the feature-inheritance mechanism is to derive the
A/A-bar distinction.
Roots and phases
201
scenario emerges in the CP phase, where φ-features of C are inherited by T before
matching the external argument):
(15)
a. [vP John vφ [VP destroy [DP the city ]]]
b. [vP John v [VP destroyφ [DP the city ]]] v-V inheritance
c. [vP John v [VP destroyφ [DP the city ]]] deletion - transfer
Note that this predicts that there can be no v without a VP-internal Goal, or
otherwise the derivation would crash at the semantic interface.8 Now the question
is what other heads apart from v (and C) require the Probe–Goal dependency that
we have just discussed. I will address this issue in the next section.
9.4 Types of roots and light categories: Why nouns
do not take arguments
So far we have assumed the existence of two types of phase heads: category-inducing
heads (Marantz’s x heads) and structural-Case-assigning heads (v and C). In some
cases, these heads coincide (v), but not in others (n and a do not (obviously) assign
structural Case, and C does not have categorizing effects). In this chapter I will not
investigate whether a unified account of phase heads is possible, I will merely assume
the Chomsky–Marantz approach in order to determine what domains qualify as
domains for interpretation. Nonetheless, I would like to argue that there are interesting consequences if we adopt Chomsky’s (2000 et seq.) feature valuation based
formulation, where phases are determined by uninterpretable features alone. In
particular, I would like to claim that this can derive the observation that nouns,
contrary to verbs, do not take arguments.
The idea that nouns do not take arguments is highly controversial, and there are
many accounts that attribute similar argument-taking properties to verbs and nouns
8
This will hold even with unacussative and passive vPs, as long as v contains some uninterpretable
φ-feature. Of course, in those cases, the Probe will be φ-partial (or ‘defective’, in Chomsky’s terms, having
just number or gender), but it will nonetheless require a DP Goal. As a consequence of the incomplete
status of the Probe, the Goal will not receive Case, which will force it to move in order to be visible by a
higher Probe. As a reviewer observes, this will depend on how locality constraints (e.g., Chomsky’s Phase
Impenetrability Condition) are defined. As an example, the reviewer mentions expletive and unaccusative
constructions like (i) and (ii), where the DP internal argument (many players and los embajadores)
remains in situ.
(i)
There are many players in the field.
(ii) Han llegado
los embajadores.
have arrived-3.PL the ambassadors
‘The ambassadors have arrived.’
(Spanish)
There are different strategies for the DPs in (i) and (ii) to receive Case. They could of course be assigned
partitive Case by V in situ. However, it is also possible that they have moved to [Spec, vP], being accessible
to C–T to receive nominative Case. A third option, suggested in Chomsky (2008), is that Transfer be
weakened, so that the domain it targets is not removed from the syntax, allowing Agree to override the
Phase Impenetrability Condition (see Gallego 2012 for discussion).
202
Ángel J. Gallego
(see Alexiadou 2001, 2009, Alexiadou and Grimshaw 2008, Grimshaw 1990, Harley
2004, 2009a, 2009b, Marantz 1997, Picallo 1991, to appear, and references therein).
Interestingly enough, the claim that nouns take bona fide arguments is only made in
the case of deverbal nouns that involve an eventive reading (16a), not in noneventive/result nominals (16b), nor referential ones (16c).
(16)
a. The examination of the patients lasted two hours.
b. The examination is on the table
c. The exam is on the table
Most approaches defending the idea that eventive nouns take arguments assume the
existence of a verb-related shell within the nP in order to derive their special (verblike) interpretation. However, there are two relevant observations that must be
highlighted: first, even event nominals can appear without their alleged arguments
(e.g., The examination lasted for hours), and second, it cannot be the case that v is
present in the nP structure of event nominals, or otherwise structural Case would be
assigned (e.g., *The examination the patients).9 This will not hold for gerund
nominals like (17), which do involve a vP layer, as accusative Case indicates (see
Pesetsky and Torrego 2004):
(17)
Mary’s reading the books surprised us.
What I would like to argue is that the claim that nouns do not take arguments in
the syntax can be seen as a consequence of Chomsky’s (2000 et seq.) approach to
cyclic derivations: Since n does not bear unvalued φ-features, the system will not
force any Probe-Goal process, and therefore n will merely have a categorizing effect
on the root it merges with. On the contrary, v must involve a Probe-Goal dependency, or otherwise its φ-features will remain unvalued and will cause a crash. From
this, the asymmetry in (18) follows without any further ad hoc assumption.10
9
Some clarifications are in order. One could of course claim that the Case-assigning properties of v
are suppressed by higher nominal projections (say, n) in (i):
(i) [nP n -tion [vP v Ø [√P √DESTR(UC) (of) [ the city]]]]
This would be problematic, as it will require for a head within a different phase (n) to alter what has already
happened in a previous derivational cycle (vP). This is perhaps clearer in (ii) through (v):
(ii) [vP vφ Ø [√P √DESTR(UC) [ the city]]]
(iii) [vP v Ø [√P √DESTR(UC)φ [ the city]ACC ]] inheritance + Case assignment
(iv) [vP v Ø [√P √DESTR(UC)φ [ the city]ACC ]] transfer
(v) [nP n -tion [vP v Ø [√P √DESTR(UC)φ of [ the city]GEN]]] of insertion + Acc supression
To be sure, further adjustments can be proposed to make such a proposal work (one could argue that v involves
just category, not φ-features), but none of them seems immediately obvious. In the proposal I am making here,
eventive readings are not parasitic on a v projection, which is supported by the fact that even non-deverbal
nouns can deploy such a reading. For a different approach, see Harley (2009b, sec. 4), who proposes that v
is present on nominalizations, and accusative Case follows from the presence of Kratzer’s (1996) voice.
10
The same claim must be made for relational nouns (e.g., brother, sister, friend). I assume that these
nouns are relational on conceptual grounds, not in the syntax. Therefore, the fact that nouns like brother
appear to refer to an argument (e.g., brother of John) is a consequence of the conceptual (i.e., syntactically
irrelevant) content of the root √BROTHER.
Roots and phases
(18)
203
a. John destroyed *(the city).
b. I saw the destruction (of the city).
If on track, this proposal might also explain the contrast in (19), which Hale and
Keyser (2002:91) account for by arguing that light verbs have a semantically defective
status that forces them to take a complement:
(19) a. *He made (vs. He made trouble/fishtraps/mistakes).
b. *She did (vs. She did a jig/pirouettes/the MCATs).
Put differently, Hale and Keyser (2002) take it that the examples in (19) are ruled out
because of semantic reasons.
[W]e propose that [the] inability [of these verbs] to take nonovert objects follows from the fact
that they are “light verbs”—in other words, verbs without any semantic component that could
enter into a classificatory binding relation capable of licensing an empty complement.
(Hale and Keyser 2002:93)
I would like to argue against this explanation, suggesting that the source of the
ungrammaticality in (19) is the same as that in (18a): the φ-features of make and do
cannot find a matching Goal, which triggers a crash in the semantic component.
A tougher case is provided by what Bosque (2001) calls ‘light nouns’. According to
Bosque (2001), the Spanish data in (20) are unacceptable if the adjective is dropped,
a fact he takes to indicate that nouns like circunstancia, momento, or situación are
‘light’ (meaning, again, semantically defective/weak),11 and therefore need the
semantic support of the adjective to yield a legitimate NP.
(20)
a. La nación pasa
por
circunstancias *(críticas).
the nation go-3.SG through circumstances critical
‘The nation is going through critical circumstances.’
(Spanish)
b. Estamos atravesando
un momento *(difícil).
be-1.PL go-through-GER a moment
tough
‘We are going through a difficult moment.’
(Spanish)
c. Se hallaban al
borde de una situación *(peligrosa).
(Spanish)
CL find-3.PL to-the edge of a
situation dangerous
‘They were on the brink of a dangerous situation.’
(Bosque 2001:32–33)
As Bosque (2001) points out, the examples are grammatical if the supporting
adjective becomes a noun.
11
Bosque’s (2001) light nouns are not to be regarded as a variant of Marantz’s (2001) n, as they are not
derivational morphemes, but full-fledged nouns.
204
(21)
Ángel J. Gallego
a. La nación pasa
por
una crisis.
the nation go-3.SG through a
crisis
‘The nation is going through a crisis.’
(Spanish)
b. Estamos atravesando
una dificultad.
be-1.PL go-through-GER a
difficulty
‘We are going through a difficulty.’
(Spanish)
c. Se hallaban al
borde de un peligro.
CL find-3.PL to-the edge of a danger
‘They were on the brink of a danger.’
(Spanish)
Although intuitively appealing, I think Bosque’s (2001) analysis is not entirely
satisfactory. To begin with, his account could not be extended to cases like (22),
where, to my ear, (b) is ruled out.
(22)
a. He
pasado por
una situación similar.
have-1.SG gone through a
situation similar
‘I went through a similar situation.’
(Spanish)
b. #He
pasado por
una similitud.
have-1.SG gone through a
similarity
‘I went through a similarity.’
(Spanish)
The oddity of Bosque’s (2001) data is akin to that in the following examples, where
the nouns día and manera appear to require some modifier:
(23)
a. Hoy hace
un día *(soleado).
today do-3.SG a day sunny
‘Today it’s a sunny day.’
(Spanish)
b. Habla
de una manera *(extraña).
talk-3.SG of a way
odd
‘She talks in an odd way.’
(Spanish)
I believe, though, that just like in the cases offered by Bosque (2001) in (20), it is
hasty to talk about ungrammaticality, as these sentences become acceptable if the
relevant context is found. For instance, the examples in (23) become perfect if they
are uttered with a suspended intonation (see 24). I take this to argue in favor of
pragmatic accommodation.
(24)
a. Hoy hace
un día . . .
today do-3.SG a day
‘What a day . . . ’
(Spanish)
b. Habla
de una manera . . .
talk-3.SG of a
way
‘She talks in such a way . . .’
(Spanish)
Roots and phases
205
We can leave the discussion here. In this section I have argued that the lack of
argument taking properties by nouns readily follows from the formal fact that n does
not bear unvalued φ-features (which, if pushed to its limits, should be taken to signal
that DPs/nPs are not phases in Chomsky’s 2000 et seq. sense; see Bosque and Gallego
2012 for more arguments in favor of this idea). This, and nothing else, predicts that
there need not be any matching Goal in the complement domain of n. Since no Goal
is required for valuation purposes, no Goal can be: if it appears, it must be regarded
as a syntactic adjunct. The same idea holds for a, but not v, which predicts the data
in (25):12
(25)
a. John destroyed *(the city).
b. The destruction (of the city)
c. It is destructive (to the city).
Notice that this is not to say that the city is not related to the conceptual content of
the root √(DE)STROY in all these cases (this is what we expect, if the root is the
same). However, the syntax will only demand the city in the context of a head
endowed with features that require valuation.
I would like to conclude by pointing out the fact that the proposal just made
predicts that all functional heads endowed with unvalued φ-fetaures will have to take
arguments in the syntax: that is, Goals. Apart from v (and C, which is of no interest
in this context), is there any other phasal head with unvalued φ-features? A growing
line of literature stemming from van Riemsdijk (1978) has argued that P is a phase
head (see Abels 2003 and Raposo 2002). The antecedent evidence for this idea goes
back to van Riemsdijk’s (1978) claim that P has an escape hatch position and to
Emonds’s (1985) unification of C and P.13 Although it is customary to take prepositions to assign inherent Case (and therefore lack uninterpretable features), some
authors have argued that they act as probes, just like v and C. This claim is made by
Baker (2008) and Kayne (2005):
à (along with similar adpositions in other languages) is part of the Case-agreement system, in
Chomsky’s (2004) sense. The fact that in French it does not actually show overt agreement
with any DP is simply parallel to the fact that v usually doesn’t (even if it is involved in past
participial agreement), nor does T in many languages (including English, where T shows overt
agreement in a very limited class of cases). Nevertheless, like v and T, P (adposition) does
show overt agreement in some (or many languages).
(Kayne 2005:95)
12
Mateu (2002) and more recently Kayne (2011) have also argued that nouns do not take arguments.
Kayne’s (2011) proposal builds on the idea (put forward by Guimarães 2000) that lexical items can selfMerge. Crucially, if self-attachment of x (to itself) takes place, x will not be able to occupy its complement
position: an occurrence of x itself will fill that slot.
13
Further similarities between P and v/C are discussed in Boeckx (2009), Kayne (1984), Pesetsky and
Torrego (2004), and Svenonius (2003, 2007, 2008).
206
Ángel J. Gallego
If on track, then this may explain why prepositions must take an internal argument (see Svenonius 2008), and why these must be DPs (not APs). The latter would
follow, I claim, from the fact that the alleged φ-features of prepositions must have a
Goal with interpretable φ-features.
(26)
a. To {Paris/*intelligent}
b. With {John/*clever}
c. For {your friend/*expensive}
The data in (26) align with the hypothesis that argument-taking properties can be
derived from the formal need of certain heads (C, v, and P) to value their φ-features.14 Needless to say, many details remain to be filled in, as this perspective opens
a controversial (but solid) door that ultimately aims at deriving interpretive effects
from purely formal/syntactic dynamics.15
9.5 Conclusions
Although Chomsky’s Phase Theory and the framework of roots are parallel explorations of language from a minimalist perspective, only few similarities between them
have been studied (and exploited) in the literature. In this short chapter, I have
suggested that there are good reasons to look for more common traits between these
minimalist frameworks. One such common trait concerns the relationship between
phase (category inducing) heads and non-phase (root) heads. There is a general
consensus that roots are categorized by means of phase (little x) heads, but the
literature is divided when it comes to deciding what other dependencies these units
establish.
An important property of roots, adopted by most approaches, is that these
elements can take internal arguments—external arguments being introduced by
Chomsky’s (1995) v (or Kratzer’s 1996 voice). From this perspective, the difference
between (27a) and (27b) is to be found at the morpho-phonological component
14
For relevant discussion with respect to the phasal status of prepositions, see Gallego (2012). A
reviewer asks whether particles should be analyzed together with prepositions, given that they do not
require complements (e.g., Sue was looking around). With Svenonius (2007), I assume that particles can be
analyzed as prepositions that have incorporated a covert complement (what, following Talmy, we could
call ‘Ground’).
15
A reviewer mentions the case of adjectives, pointing out that some (e.g., likely) appear to require
arguments (e.g., John is likely *(to pass)). As above, I assume that nouns and adjectives cannot take
arguments in the syntax. Thus, I take it that the following contrast, provided to me by Jaume Mateu (p.c.),
is due to encyclopedic (grammatically irrelevant) reasons (that’s why I use the # diacritic): The room is full
vs. The sea is full #(of fish, ships, etc.). The example provided by the reviewer should be qualified too, since
likely does not always require a CP complement (in the relevant context, That is not likely is perfectly fine).
If there turn out to be cases where the CP cannot be dropped, I would assume it is not due to likely, but to
the reanalysis between v and likely (much like reanalyses between verbs and nouns in Spanish, giving rise
to a sort of unergative structure where the Case marker of the the CP can be dispensed with): No tenía idea
(de) que no vendrías ‘I had no idea you would not come’.
Roots and phases
207
(where the Case of the evidence is determined): selection-wise, there is no remarkable
asymmetry:
(27)
a. The investigation of the evidence
b. The experts investigated the evidence
Contrary to this general trend, I have suggested that internal arguments are obligatory only if a v head is merged with the root. This follows from the fact that it is v,
not n, that contains a bundle of uninterpretable φ-fetaures, which require valuation
(Chomsky 2000, 2001).16
The claim that roots do not take internal arguments aligns with Borer (2005,
2009), Marantz (2001 et seq.), and Acedo-Matellán and Mateu (this volume), but the
rationale put forward here is different. While, e.g., Marantz severs internal arguments from roots due to his assumption that roots are modifiers of the denotation of
x heads, my argument is based on Chomsky’s (2000 et seq.) analysis of feature
valuation: Given that it is x heads that may encode unvalued φ-features, they will be
the ones determining whether roots take internal arguments or not—roots themselves will thus never take arguments, as they are feature-less units. This predicts that
x heads that fail to be endowed with unvalued φ-features will not take internal
arguments in the syntax. The prediction is borne out, to the extent that the evidence
can be absent in (27a), but not in (27b).17
This approach to roots fits in a framework, like Chomsky’s, where non-phase
heads inherit properties from phase heads. There are good empirical and conceptual
reasons to think that feature inheritance takes place in the case of φ-features and
category labels (see Chomsky 2007, Marantz 2001). To this list, I have added
argument taking. It remains to be investigated whether the same holds for the
semantic types that have been ascribed to roots (see Folli and Harley 2005, Harley
1995, 2005, 2009b, Hale and Keyser 2002, Levinson 2007, Marantz 1997 et seq.,
Pylkännen 2008, and Ramchand 2008), and which, by the very same logic, should
also derive from x heads. This and further issues are left for future research.
16
A reviewer asks me to say something about the fact that, unlike n, v bears uninterpretable features.
The reviewer points to the possibility that it can be an arbitrary grammatical fact or a deeper third-factor
oriented requirement. I am very much sympathetic to Chomsky’s idea that the existence of uninterpretable features themselves (be these related to v, n, or any other category) has a third-factor motivation.
However, it is not so clear to me why they should be related to some categories and not others. Taking into
account the Strong Minimalist Thesis, one could plausibly argue that the association of interpretable
φ-fetaures with nouns and not verbs is determined by the interfaces (ultimately, the ConceptualIntentional systems), and not by Universal Grammar. See Chomsky (2001:5) for some qualifications.
17
An issue raised in a previous version of this chapter is whether this take is compatible with (7a) or
(7b). As an anonymous reviewer points out, if feature inheritance is assumed, then we must adopt (7a), for
otherwise the φ-feature would fail to find a matching Goal in its domain. If features are not removed from
phase heads (as argued for in Chomsky 2012), then one could adopt (7b), but notice that this would require
an independent projection mechanism so that the φ-fetaures within the v head percolate and have the DP
Goal within their range of action.
10
The ontology of roots and verbs
LISA LEVINSON
10.1 Introduction
Verbs can be classified in ways that seem to reflect both semantic and morphosyntactic similarity at the same time, as shown in great detail in Levin (1993).1 However,
we are still far from understanding how exactly to connect the semantic properties
with the morphosyntactic properties. In this chapter, I argue that some such
correlations can be derived from the semantic types of the roots which form the
lexical core of verbs. This idea in itself is not new, as for example Rappaport Hovav
and Levin (1998) argue that the meaning of what they call ‘constants’ determines
some aspects of a verb’s syntactic realization. However, what is novel to the present
approach (a development of the view put forward in Levinson 2007b) is putting this
idea together with Distributed Morphology approaches that give roots a life in the
syntax, rather than simply in the lexicon. This makes additional predictions regarding the compositional semantic interpretation of roots, which I argue in this chapter
are borne out. These findings also provide evidence that such roots are not semantically vacuous in isolation as has been proposed by Borer (2005a), Acquaviva (this
volume), and Harley (2009c).
The key aims of this chapter are to provide an explicit compositional account for
verbal lexical decomposition and to show that this formalization can provide
important insight into this domain. It is also hoped that this chapter contributes
to our understanding of the ontology of lexical roots, and also shows that what may
appear to be verb polysemy sometimes also involves structural ambiguity, potentially
combined with root polysemy. Due to this last fact, great care must be taken in
controlling for such ‘lexical’ ambiguities when classifying and analyzing verbs.
1
I would like to thank Chris Barker, Jon Brennan, Marcel den Dikken, Richard Kayne, Tom Leu, Alec
Marantz, Øystein Nilsen, Liina Pylkkänen, Oana Savescu-Ciucivara, Anna Szabolcsi, Eytan Zweig, and
two anonymous reviewers for very helpful comments on the topic of this chapter.
The ontology of roots and verbs
209
The model of grammar that I am assuming is closest to that presented in Marantz
(1997) and Arad (2005) in connection with the Distributed Morphology framework
(Halle and Marantz 1993). One key assumption from Marantz (1997) is the ‘single
engine hypothesis’, that there is one computational system which generates both
words and larger constituents. In this framework, words are not built in the lexicon,
but rather in the same fashion as phrasal constituents, in the syntax. Words are not
atomic, but are built from roots, which constitute the atomic syntactic terminals
providing the “lexical” content. These roots do not bear categories like “verb” or
“noun” (see also Pesetsky 1995, Barner and Bale 2002, Barner and Bale 2005, Borer
2005a, Borer 2005b). Rather, they seem to “join” these syntactic categories when they
combine with what are considered to be category-specific heads in the syntax.
Throughout the chapter, I use the term “verb” either in the informal, traditional
sense of the word, or as a descriptive term for complex constituents which contain a
v head.
I add to this framework the assumption that the mapping between such syntactic
constituents and interpretation is strongly compositional. The strong interpretation
of compositionality I will take is one that admits no semantic rules that do not
correspond to steps in the syntactic derivation, nor semantic elements that do
not correspond to elements in the syntax. In formalizing the semantic proposals, I
will assume an extensional typed λ-calculus in which variables of type e range over
individuals, type s over eventualities, and type t over truth values. The type s is
subdivided into the sorts se for events and ss for states. I further assume that the only
available modes of semantic composition are functional application (as defined in
Heim and Kratzer 1998), predicate modification (functional intersection) (as defined
in Heim and Kratzer 1998), and event identification (Kratzer 1996).
Given these assumptions, roots must be specified for semantic type (in the sense
of formal type theory) in order to compose with other syntactic constituents. The
type of the root has apparently syntactic ramifications, as it determines the arguments the root combines with and the combinatorial possibilities in semantic
composition. Further, what appears to be the same root conceptually may vary
across languages with respect to verb class membership, which would not follow if
roots represented only universal conceptual information. Thus contrasts in crosslinguistic realization of verbs and argument structure motivate the storage of some
‘arbitrary’ linguistic specifications in association with roots, including semantic type.
This proposal puts the onus of certain linguistic contrasts on the semantic properties
relevant to the root itself, which is in opposition to the positions put forth by Borer
(2005a), Borer (2008), Acquaviva (2009), and Harley (2009b), that roots do not have
grammatical properties of this kind. Such opposition to the semantic specification
of roots is picked up by Acedo-Matellán and Mateu (this volume), while other
contributions in this volume (Roßdeutscher this volume, Alexiadou this volume,
Anagnostopoulou and Samioti this volume, Doron this volume, Rappaport Hovav
210
Lisa Levinson
this volume) share the current view that such root properties are empirically
necessary.
In addition to accounting for the regularities of root environments by association
with semantic types, if the present proposal is correct, another locus for the semantic
idiosyncracies of roots must be identified as well. I will argue that some such
‘idiosyncracy’ is due to cases of root polysemy, given that roots may vary in a
given language with respect to their semantic type. Some more radical cases of
idiosyncracy that fall beyond the scope of this chapter may call for a resuscitation of
the proposal put forward in Marantz (1996), later abandoned in Marantz (2007a),
that words are a potential domain for idiomatic interpretation on a par with larger
phrasal constituents.
The verb classes and the relevant contrasts that I will argue can be explained by
this approach are summarized in Table 10.1:
Table 10.1. Verb Class Patterns
Verb Class
Pseudo-Resultative
Double Objects
Obligatory Theme
Root Creation Verbs
Explicit Creation Verbs
Change of State Verbs
ü
*
*
*
ü
ü
ü
*
ü
How can these contrasts be explained? In brief, the proposal is as follows. Root
creation verbs have the properties they do because they are built from roots which
denote predicates of individuals, or entities (like “common nouns”). This fact
constrains the possible contexts that can build this root into a verb, or predicate of
eventualities. Explicit creation verbs are structurally distinct from root creation
verbs, driven by the fact that they are derived from a root which denotes a predicate
of eventualities, not individuals. Therefore, verbs that occur in both “frames” are
actually ambiguous between two structures, leading to different morphosyntactic
realizations. The different structure of explicit creation verbs does not license
pseudo-resultatives, but does allow for applicatives and optional objects. Change of
state verbs involve yet another structure, where the root is a predicate of individuals
mapping to a predicate of states.
Thus, the contrasts are argued to be due to the contrasts shown in Table 10.2
between the root types involved in the derivation of these verb classes:
Table 10.2. Root Types and Verb Classes
Verb Class
Root Type
Root Creation Verbs
Explicit Creation Verbs
Change of State Verbs
< e,t>
< se,t>
< e, < ss,t>>
The ontology of roots and verbs
211
This is not meant to be an exhaustive list of root types or verb classes. However,
these classes will be used to illustrate the general approach.
10.2 Root creation verbs
Root creation verbs (Levinson 2010), exemplified in (1), coincide largely with those
called ‘goal’ verbs by Clark and Clark (1979:774).
(1)
Root Creation Verbs:
a. The stylist braided her hair. → At least one braid was created.
b. The decorator piled the cushions. → At least one pile was created.
c. The baker sliced the bread. → At least one slice was created.
d. The barista ground the coffee beans. → Fine coffee grounds were created.
These verbs entail the creation of an individual, without the expression of that
individual as a DP argument. The meaning of (1a) is parallel to that in (2):
(2)
The stylist made/reconfigured her hair into a braid.
In examples like (2), the object of the preposition into names the created individual.
However, in root creation verbs, this individual is named by the root of the verb
itself. That is, in (1a), what is created is a braid, in (1b) what is created is a pile, and so
on. This is why the class is called “root creation”, because the root names the
creation. Another crucial ingredient of this “verb frame” is the presence of the
material which is reconfigured, such as her hair in (1a).
In this section, I will first summarize the proposal for the analysis of these verbs
from Levinson (2010), and then show how this proposal predicts the properties of
root creation verbs summarized in Table 10.1—namely that these verbs occur with
pseudo-resultative predicates, don’t occur in the double object construction, and
require theme objects.
10.2.1 Proposal
It is argued in Levinson (2010) that the created individual contributed by root creation
verbs is present in the syntax and is denoted by the root of the verb. The basic idea is
that, in order to build a verb from such a root, which has a denotation like a common
noun, root creation verbs essentially amount to a conflation of elements similar to
those bolded in (2) into one word, where the created individual is contributed by a
root rather than a DP. For example, the root √braid is argued to contribute a property
denotation of λxe.braid(x), which is type < e,t>. The root is related to the “reconfigured” argument, her hair, by two functional heads, called IN and TO. IN takes the
root as an argument and the result denotes the state of being a braid:
(3)
〚IN〛 = λf< e,t>.λye.λsss.∃xe.f(x) & being-in(s,x) & theme(s,y)
212
Lisa Levinson
TO is a purely syntactic head licensed by the causative v which has the potential to
assign case:
(4)
〚TO〛 = semantically/type-theoretically vacuous
Since the root itself does not contribute any eventuality variable, with such verbs
the sole event variable is introduced by a causative v head with “reconfiguration”
semantics that entail a kind of creation that involves reconfiguration:
(5)
〚vreconfigure〛 = λf< ss,t>.λese.∃sss.f(s) & reconfiguration(e) & CAUSE(s,e)
I assume that there is not simply one v head, but rather that there is an inventory
of heads which can serve as verb categorizers. This is akin to the fact that there are
different T and C heads which vary with respect to their syntactic features, semantics, and selectional restrictions. Folli and Harley (2005) describe this diversity of
verbal categorizers as the availability of different ‘flavors’ of v. Here I use the
descriptive predicate ‘reconfiguration’ to distinguish this head from the v used
with explicit creation verbs, but the meaning of this v can actually be quite light,
as can be seen by its interchangeability in the phrasal context with light verbs such as
make (as in (2)).
The verb produced by the heads detailed above is a complex that can be built
syntactically by incorporation, head movement, or conflation (which I assume has
no semantic import).
(6)
vreconfigure
DP
her hair
TO
IN
√
√braid
The denotation for this whole phrase given in (6) will amount to a predicate of
events as follows:
(7)
Formally: λese.∃sss.∃xe.braid(x) & being-in(s)(x) & theme(s,her hair) & reconfiguration(e) & CAUSE(s)(e)
(8)
Informally: A set of reconfiguration events which cause a state in which ‘her
hair’ is in a braid.
The most crucial aspect of this analysis for the purpose of the current chapter
is that the root denotes a predicate of individuals. This makes predictions for
The ontology of roots and verbs
213
modification possibilities and argument structure which will be detailed in the
following sections.
10.2.2 Root modification
One prediction made by the analysis of root creation verbs given in the previous
section is that the root, syntactically ‘active’ and of type < e,t>, might be available for
modification by categories which modify elements of type < e,t> in other contexts,
such as adjectives. Levinson (2007a, 2010) presents evidence showing that this
prediction is in fact borne out, as evidenced by pseudo-resultative modification.
Pseudo-resultatives are adjectival predicates that are superficially similar to resultative secondary predicates:
(9)
Pseudo-resultatives:
a. The stylist braided her hair tight. → At least one tight braid was created.
b. The decorator piled the cushions high. → At least one high pile was created.
c. The baker sliced the bread thin. → At least one thin slice was created.
d. The barista ground the coffee beans fine. → Fine coffee grounds were created.
However, pseudo-resultatives do not modify the DP object as resultatives do. The
resultative-like semantics found with pseudo-resultatives is contributed not by the
addition of or modification of a resultant state, but by modification of an individual
which is created as a result of the event. In languages like Finnish, there is a
morphological distinction between the two types of predicate, which will be useful
later on. As can be seen in (10) and (11), resultatives have translative case marking,
while pseudo-resultatives have illative case marking.
(10)
Mari
hakkasi
metalli-n litteäksi.
Mari.NOM hammered metal-ACC flat-TRANS
‘Mari hammered the metal flat.’
(11)
Mari leti-tt-i
hiuksensa
tiukka-an.
Mari braid-CAUS-PAST hair-ACC.POSS tight-ILL
‘Mari braided her hair tight.’
(resultative)
(pseudo-resultative)
The proposal regarding these predicates is that the pseudo-resultative adjective, a
property of type < e,t>, can combine with the root via predicate (intersective)
modification, a standard treatment of adjectival modification:
(12)
λxe.braid(x) & tight(x)
λxe.braid(x)
λye.tight(y)
√braid
tight
214
Lisa Levinson
This gives the entailments shown in (9). It is because the root is of type < e,t> that
the pseudo-resultative predicate is able to modify it. Therefore, adjectival predicates
will only modify the root of a verb in this way when the root is of type < e,t>. Such an
adjective will only receive a pseudo-resultative interpretation when the role of the
root in the meaning of the verb is that of a created individual. It will be seen in later
sections that verbs that are derived from other types of roots do not support pseudoresultative modification.
10.2.3 Double objects
As discussed in Pylkkänen (2008), English only has double object constructions
which encode an intended transfer of possession associated with the event introduced by the verb. This is in contrast with other languages which allow for purely
benefactive readings like the Luganda example in (13):
(13)
Katonga ya-kwaant-i-dde
Mukasa ensawo.
Katonga 3SG.PAST-hold-APPL-PAST Mukasa bag
‘Katonga held the bag for Mukasa.’
(Pylkkänen 2008: ex.23b)
However, English does permit benefactive-like readings which also encode an
intended result of possession. Such readings are typically found with creation verbs,
as noted in Levin (1993) and shown in (14):
(14)
a. The chef cooked the customer a delicious soup.
b. The university built the department a new lab.
In these examples, the object of the verb is an effected argument and comes to be in
the possession of the benefactive argument as a consequence of the creation event.
Root creation verbs do not license either type of the double object construction,
even if a transfer of possession of the created object as a result of the event is
plausible:
(15)
a. #The worker braided her boss the rope. (where ‘the rope’ is the material,
not the created object)
b. * The baker sliced the customer the loaf of bread.
Although these ‘verbs’ are elsewhere compatible with benefactive applicative arguments in the double object construction, it is only on the explicit creation reading,
which will be discussed in section 10.3. In these cases, the theme is not the material
from which the created object is created, but is the created object itself, and thus
these are not truly root creation verbs:
(16)
a. The jeweler braided her customer a necklace.
b. The pastry chef sliced everyone a piece of cake.
The ontology of roots and verbs
215
Verbs which are unambiguously root creation verbs more clearly illustrate the
incompatibility of root creation verbs with double objects and for this reason will
be used for the rest of this discussion:
(17)
* The librarian piled the student the books.
The incompatibility between the root creation reading and the double object
construction fact can also be observed in Finnish:
(18)
* Hn leti-tti
minu-lle minu-n tukka-ni.
s/he braid-CAUS.PST 1SG-ALL 1SG-GEN hair-POSS1SG
‘She braided me my hair.’
As in English, this contrasts with an explicit creation reading of the same verb:
(19)
Hn leti-tti
minu-lle pullapitko-n.
s/he braid-CAUS.PST 1SG-ALL braided.bread-ACC
‘She braided me a “braided bread”.’ (explicit creation reading)
The analysis proposed above provides a straightforward semantic explanation for
the lack of double objects with root creation verbs. If we assume an analysis of
double objects as involving applicative heads as in Pylkkänen (2008), these structures
involve an applicative head which introduces a relation between two arguments,
such that the first comes to be in the possession of the second. For example, in (20),
the head APPL would take a book and Mary as arguments and relate them such that
the book is the intended possession of Mary.
(20)
Bill sent Mary a book.
After relating these arguments, the applicative head then takes an eventive verbal
head as an argument, relating the arguments to that event. Formally, the APPL
head takes two individual arguments before combining with a constituent of type
< e,< s,t>>:
(21)
Low-ApplTo (recipient applicative):
λx.λy.λf< e,< s,t>>.λe.f(e,x) & theme(e,x) & to-the-possession(x,y) (Pylkkänen
2008:ex. 137b)
Pylkkänen (2008:ex. 37) provides the following analysis of the sentence Mary
bought John the book (setting aside for current purposes merge of the voice head
and external argument):
216
Lisa Levinson
(22)
λe.buying(e) & theme(e, the book) &
to-the-possession(the book, John)
λf<e,<s,t>>.λe.f(e, the book) &
theme(e, the book) &
to-the-possession(the book, John)
λx.λe.buying(e)
& theme(e,x)
buy
John
λx.λy.λf<e,<s,t>>.f(e,x) &
the book
theme(e,x) & to-the-possession(x,y)
APPL
Recall that root creation verbs must be built up from a root which denotes a
predicate of individuals, which is then embedded in the necessary structure which
builds it into a verb. The only position in such a structure which could accommodate
the semantic composition of the applicative phrase is as in (23b):
(23)
a. ∗The librarian piled the student the books.
b. λsss.∃xe.pile(x) & being-in(s,x) & theme(s, the books) &
theme(s, the books) & to-the-possession(the books, the student)
<e,<s,t>>
<<e,<s,t>>,<s,t>>
e
the
student
<e,<<e,<s,t>>,<s,t>>>
APPL
e
the books
TO
IN
√
√pile
The ontology of roots and verbs
217
The denotations of the ApplP and PP nodes would be as follows:
(24)
〚ApplP〛 = λf< e,< s,t>>.λes.f(e, the books) & theme(e, the books) & to-thepossession(the books, the student)
(25)
〚PP〛 = λye.λsss.∃xe.pile(x) & being-in(s,x) & theme(s,y)
There are two types of problems that would arise upon composition of ApplP and
PP. On the one hand, there is no straightforward analysis for the syntactic merge of
these two constituents, given that they are two maximal projections and neither is an
adjunct. If the complex PP syntactically selects for a DP in the specifier which
usually creates a position for the reconfigured material argument, this selection
would also fail to be satisfied. These syntactic problems have their source in the
semantic type of the root, since it is the status of the root as a predicate of individuals
which requires it to be embedded in this more complex structure to produce a verb.
Even if there might be a solution to this syntactic problem such as the availability
of an additional null linking element, the resulting interpretation is not consistent
with either the pure possession or benefactive possession readings that are found
with English double object constructions. The denotations for the v head and the
composed vP would be:
(26)
〚vreconfigure〛 = λf< ss,t>.λese.∃sss.f(s) & reconfiguration(e) & CAUSE(s,e)
(27) 〚vP〛 = λese.∃sse.∃xe.pile(x) & being-in(s,x) & theme(s, the books) & theme(s,
the books) & to-the-possession(the books, the student) & reconfiguration(e)
& CAUSE(s,e)
Given the fact that the created object is denoted by the root, not the direct object,
there is no way for the created entity to be related to the intended recipient or
benefactive argument. That is, the relation established by the applicative head is
between the student and the books, not the student and the pile. Since the verb is not
inherently a verb of transfer, a benefactive interpretation would be required, but
there is no way to establish the appropriate relationship between the created object
and the benefactee. Thus semantically the relation established doesn’t fit the general
condition for English benefactives that the direct object be created with the intention
of resulting possession for the benefit of the indirect object argument.
Put in other words, if composition is syntactically licensed at all, a benefactive
reading of (23a) would require that the books, not the pile, were created for the
benefit and directed to the possession of the student as a result of the piling event.
(15a) would mean that that rope was created to benefit the boss, and (15b) that the
loaf was created to benefit the customer. Yet these readings are not compatible with
the entailments of the verbs which require that the root denotes the created object.
Since it is impossible to establish such an applicative relation with root creation
verbs, the prediction is that low applicative heads and pseudo-resultatives should be
218
Lisa Levinson
in complementary distribution; the latter require a root of type < e,t>, while the
former are not compatible with syntactic and semantic environments necessary to
derive a root creation verb from that root type.
10.2.4 Obligatory theme
Root creation verbs require direct object themes. This is difficult to illustrate directly
because of the same caveat that popped up in the previous section—that in English,
many roots that produce root creation verbs also can be realized in explicit creation
verbs due to root polysemy. However, again the facts are clear when using an
unambiguous root:
(28)
* Mary piled.
With roots that are ambiguous, we must control for the interpretation. It is difficult
to control for interpretation without the presence of an object. However, since it has
been established that pseudo-resultatives occur with root creation verbs, but not
explicit creation verbs, we can use them to help in this task. One way to control for
the right class of verb without the presence of the object is by using the pseudoresultative predicate, which can only occur with the root creation verb:
(29)
The stylist braided *(hair) tight all day. (root creation)
This example shows that true root creation verbs require an object. This is not an
effect of the pseudo-resultative itself, since, unlike resultatives, the pseudo-resultative
would not modify or depend on the object in any case.
This behavior also follows from the analysis given. The head IN is obligatory in
the decomposition of such verbs, and this head introduces an empty argument slot.
The “PP” constituent is of type < e,< s,t>>, looking for an individual argument. This
argument slot must be saturated first before this lower material can combine with
the v head, and this is the role of the direct object. If the direct object is missing, the
PP cannot compose with the v head, which is of type < < s,t>,< s,t>>, as indicated by
the asterisk on the root node in (30):
∗
(30)
λf<ss ,t>.λes.∃sss.f(s)
& reconfiguration(e) & CAUSE(s,e)
vreconfigure
λye.λsss.∃xe.braid(x)
& being-in(s,x) & theme(s,y)
TO
IN
√braid
The ontology of roots and verbs
219
10.2.5 Root creation verb summary
In this section it has been shown how the internal structure of root creation verbs
accommodates pseudo-resultative modification, but not applicatives or optional
objects, due to the semantic type of the root and the elements it must combine
with to “become a verb”.
10.3 Explicit creation verbs
While braid as a root creation verb relates the object DP to a created individual
denoted by the root, braid can also appear as an explicit creation verb, where it
expresses no such relation and is parallel instead to verbs like bake and build:
(31)
a. The jeweler braided a necklace (out of strands of silver).
b. The pastry chef baked a cake.
c. The contractor built a house.
In Levin (1993), these are called “build verbs”. In these examples, the object DP
expresses the created object and the root contributes a manner specification for the
creation of that object. This interpretation of braid can be paraphrased as in (32):
(32)
The jeweler made/created a necklace (out of strands of silver) by braiding.
As summarized in Table 10.1, explicit creation verbs do not occur with pseudoresultatives, do occur in the double object construction, and do not require a theme
(more specifically, they systematically alternate with activity verbs which lack a
theme). In this section, I will first present an analysis for explicit creation verbs,
and will then show how this analysis explains these properties.
10.3.1 Proposal
Marantz (2005) argues that creation verbs like bake have the structure in (33), where
the root combines with a v head and then the resulting constituent takes the DP
object as an argument:
(33)
vP
v
v
object
√
Here the theme DP explicitly refers to the created object. This is in contrast with root
creation verbs, where no DP argument refers to the created object. A minimal pair
can be constructed as in (34):
(34)
a. The stylist braided her hair. (root creation)
b. The jeweler braided a necklace. (explicit creation)
220
Lisa Levinson
One might assume that, structurally, these verbs are equivalent. After all, they
seem to be derived from the same root. However, this does not seem to be correct, as
there are syntactic contrasts which correlate with these different readings, as already
seen in the previous section. (34a) patterns with verbs that are unambiguously root
creation verbs, and (34b) with those that are clear cases of explicit creation verbs.
Semantically, in this environment, √braid cannot be of type < e,t>—there is no set
of braids denoted, just as there is no set of “bakes” denoted by the verb bake. I
propose that the composition of such verbs based on the structure above is as
follows:
(35)
λese.making(e) & braiding(e) & theme(e, a necklace)
λxe.λese.making(e) & braiding(e) & theme(e,x)
a necklacee
a necklace
λxe.λese.making(e) & theme(e,x)
λes.braiding(e)
vcreate
√braid
Here the root is not of type < e,t>, but rather of type < se,t>, a predicate of events,
and combines with the v via Event Identification (Kratzer 1996). Given the fact that
braid can be either a root creation verb or an explicit creation verb, the root √braid
must be polysemous in English between an < e,t> and an < se,t> interpretation. In
this case, this root polysemy corresponds with a structural difference as well. The
vcreate proposed should have a meaning similar to that of make as used in (32) above.
Here it is the v that introduces the argument position for the object.
The realization of the root in this type of structure is related to its possible
denotations. The proposal is that such roots denote predicates of events (of type
< se,t>), and this is able to combine with the relevant v heads directly, rather than in a
relational structure. Vcreate must be of type < e,< se,t>>, since it takes a DP complement, and thus requires an individual argument. This vcreate and the event-predicate
root √braid can combine via event identification (Kratzer 1996). Although this mode
of composition is different from that by which predicates of events combine with
manner adverbs, the result is equivalent, in that both the root and vcreate come to take
the same event as an argument. The difference is that here the resulting expression
has an open individual argument slot. That this argument slot is introduced by the v
head, not the root itself, is another desirable property of this analysis. This complex
head can then semantically compose with the DP of type e.
The ontology of roots and verbs
221
The correlation of the root type < se,t> with the structure in (35) is what ultimately
leads to the generalizations cited above.
10.3.2 Root modification
The analysis above predicts that pseudo-resultatives should be unavailable with
explicit creation verbs, since the root is not of type < e,t>. In English, this is somewhat difficult to test precisely, since given a sentence like (36), one cannot reliably
determine whether the sentence-final predicate is modifying the root or the object:
(36)
Mary braided the necklace tight.
Is it the necklace itself that is “tight”, or the braid that constitutes the necklace? Since
the necklace created by braiding is itself a braid, these two things cannot be teased
apart easily. However, languages which distinguish between pseudo-resultatives and
resultatives morphologically show that this predicate is likely to be a resultative in
English, not a pseudo-resultative. In Finnish, translative (resultative) case occurs on
the secondary predicate with the verb for braiding when the created object is realized
as the DP object argument of the verb:
(37)
Mari leti-tt-i
leti-n
tiuka-ksi.
Mari braid-CAUS-PAST braid-ACC tight-TRANS
‘Mari braided the braid tight.’
Norwegian similarly shows resultative morphology, in the shape of adjectival agreement, on such resultative predicates with explicit creation verbs, whereas agreement
is not possible with root creation verbs which combine with pseudo-resultatives.
Thus agreement morphology on the predicate disambiguates between the root and
explicit creation verb readings:
(38)
a. Marit skjærte kaka
tynn-;.
Marit cut
cake-DEF.F thin-M/F.SG
‘Marit cut the cake thin.’ (Resultative—one whole, thin cake is created)
b. Marit skjærte kaka
tyn-t.
Marit cut
cake-DEF.F thin-NEUT.SG
‘Marit cut the cake into thin slices.’ (Pseudo-resultative)
Default neuter agreement forces a pseudo-resultative interpretation, and thus also a
root creation verb reading in (38b). Feminine agreement, on the other hand, gives a
resultative interpretation, and thus an explicit creation verb, as in (38a).
10.3.3 Double objects
As seen in section 10.2, explicit creation verbs do allow for benefactive applicatives:
(39)
She braided Mary a necklace. (explicit creation—the necklace is the creation)
222
Lisa Levinson
As shown in the section on root creation verbs, the same is found in Finnish:
(40)
Hn leti-tti
minu-lle pullapitko-n.
s/he braid-CAUS.PST 1SG-ALL braided.bread-ACC
‘She braided me a “braided bread”.’(explicit creation reading)
The fact that these verbs can occur with applicative arguments is also predicted,
given that there is a direct object slot of type e which combines with a complex v
head of type < e,< s,t>>. This provides a site for inserting an APPL head. The
combination of the applicative head with this structure would be as in (42):
(41)
The artisan braided her customer a necklace.
(42)
<s,t>
<e,<s,t>>
vcreate
√braid
<<e,<s,t>>,<s,t>>
e
her customer
<e,<<e,<s,t>>,<s,t>>>
APPL
e
a necklace
The denotations of the ApplP, root-modified v and vP nodes would be as follows:
(43) 〚ApplP〛 = λf< e,< s,t>>.λes.f(e, a necklace) & theme(e, a necklace) & to-thepossession(a necklace, her customer)
(44)
〚vcreate + √braid〛 = λxe.λes.making(e) & braiding(e) & theme(e,x)
(45)
〚vP〛 = λese.making(e) & braiding(e) & theme(e, a necklace) & theme(e, a
necklace) & to-the-possession(a necklace, her customer)
This sentence would then entail that a necklace was directed to Mary’s possession by
an event of creating it, which is a natural low applicative interpretation. The analysis
is the same as Pylkkänen’s (2008) analysis of buy above in (22) except for the fact
that the root meaning is separated from the functional v head.
10.3.4 Optional theme
Given that roots of explicit creation verbs do not themselves introduce individual
argument slots, it is correctly predicted that such roots should be possible in
intransitive contexts. This prediction appears to be borne out, as noted in Levin
The ontology of roots and verbs
223
(1993). This can be seen from the examples in (46) which contain verbs based on the
same roots as explicit creation verbs:
(46)
a. The chef baked all day.
b. The workers were building for days.
This is in contrast with root creation verbs, which are obligatorily transitive, as
discussed in section 10.2. This contrast is another which can be explained by
contrasts in root type.
Above, vcreate is analyzed as introducing the object argument position for explicit
creation verbs. However, the root could just as easily combine with a different v
which does not have such an open argument slot, and still retain the same activity
interpretation:
(47)
λese.making(e) & braiding(e)
λes.making(e)
λese.braiding(e)
v
√braid
Thus, it is correctly predicted that roots of type < se,t> should easily occur in
intransitive contexts. Technically, in such a context, these are not explicit creation
verbs anymore, if the verb type is taken to be determined by the structure in which
the root is embedded. However, these are activity verbs built from the same roots as
explicit creation verbs, and the proposal here predicts a systematic alternation
between these two verb classes.
10.3.5 Explicit creation verb summary
In this section, it was shown that deriving explicit creation verbs from a root of type
< se,t> can explain certain facts about their morphosyntactic behavior. It was also
seen that some verbs, such as braid, are nevertheless ambiguous between a root
creation and an explicit creation verb reading. This means that roots like √braid can
be either of type < e,t> or < se,t>. The proposal is not that there is homophony
between two roots, but that one root is able to have more than one associated
denotation, leading to ‘allosemy’ of the root. This would be parallel to the ability of
one root to have multiple allomorphs. As with allomorphy, it is language-specific
what denotations are associated with which roots. Thus in English, √braid can
denote either a predicate of type < e,t>, or of < se,t>, while in another language,
only one of these denotations might be available. The available denotations would
determine the possible environments for insertion of the root.
224
Lisa Levinson
10.4 Change of state verbs
Change of state verbs are those that denote a change of state of the theme argument.
Dowty (1979) describes them as being derived from stative adjectives, though classes
of verbs like Levin’s (1993) ‘break verbs’ seem to pattern similarly despite not
appearing to be derived from adjectives. The change of state uses of the verb open
are illustrated in (48):
(48)
a. The door opened. (inchoative/anti-causative)
b. The bellhop opened the door. (causative)
(48a) is an example of an inchoative, and (48b), where an additional causative
relation is introduced, a causative.
In the introduction, it was indicated that change of state verbs do not occur with
pseudo-resultatives, but require themes and participate in the double object construction. The following section lays out the proposal for these verbs and is followed
by an explanation of how this proposal explains these generalizations.
10.4.1 Proposal
Dowty (1979), Parsons (1990) and Hale and Keyser (1993) all propose that change of
state verbs are deadjectival, while Harley (1995) proposes the structure in (49), where
an inchoative head selects a √P:
(49)
vP
v
√P
INCH
√open door
Pylkkänen (2008) and Alexiadou et al. (2006) also treat the “adjectival” root as
category-neutral. In these analyses of change of state verbs, a stative root combines
first with a theme argument before combining with a v head.
A stative root can occur in a structure like (49) because it has an argument
structure such that it combines first with a type e argument, and then with a state
argument. That is, it is of type < e,< ss,t>>. This is in contrast with the roots of root or
explicit creation verbs, which can only take arguments indirectly via a relational
structure or an argument-taking v head. Therefore, the type of v that roots of type
< e,< ss,t>> combine with will be different than that which combines with these other
roots. The fact that the unaccusative variants are possible can be explained by the
fact that change of state v heads require state arguments in their complements. The
composition of the vP of an intransitive change-of-state verb would be as in (50b):
The ontology of roots and verbs
(50)
225
a. The door opened.
b.
λese.∃sss.open(s) & theme(s, the door)
& change-of-state(e)
& CAUSE(s,e)
λf<ss ,t>.λese.∃sss.f(s)
& change-of-state(e)
& CAUSE(s,e)
vinch
λsss.open(s)
& theme(s, the door)
λze.λsss.open(s)
& theme(s,z)
e
a beer
√open
10.4.2 Root modification
Like explicit creation verbs, change of state verbs do not appear to license pseudoresultative modification. The modifiers that occur with them are resultatives, where
the secondary predicate modifies the direct object:
(51)
The river froze solid.
(52)
The wind froze the river solid.
Semantically, the sentence-final predicate is generally considered to modify the
theme, in these examples ‘the river’. One could argue perhaps that it is the ‘freeze’
that is solid, which would be a pseudo-resultative interpretation. However, in
Finnish these are clearly marked as resultative (with translative case), not pseudoresultative (illative case):
(53)
Joki jäätyi kiinte-ksi.
river froze solid-TRANS
‘The river froze solid.’
The unavailability of pseudo-resultatives with change of state verbs follows from
the fact that again there is no subconstituent of the verb which is a predicate of
individuals for the pseudo-resultative to modify. Resultatives are possible as modifiers of the theme instead.
However, despite the unavailability of pseudo-resultatives, change of state verbs
seem to license another kind of root modifier. Pylkkänen (2008) argues that certain
cases of apparent ‘verb’ modification found with lexical causatives in English are
actually root modification, such as with the bolded modifiers in (54) (examples
originally from Tenny 2000):
226
(54)
Lisa Levinson
a.
b.
c.
d.
e.
John closed the door partway.
John partly closed the door.
Roger half filled the glass.
Roger filled the glass halfway.
Nicolas mostly filled the glass.
These modifiers have lower scope than verbal modifiers, such that partway in (54a)
modifies the “closed” state, not the causing event. This can be seen by the fact that
partway receives the same interpretation in the intransitive variant in (55):
(55)
The door closed partway.
This is in contrast with verbal modifiers which seem to lack the lower scope reading
in causatives, as illustrated in (56) from Pylkkänen (2008:ex. 45):
(56)
a. Bill awoke grumpily.
b. John awoke Bill grumpily. (false if John wasn’t grumpy)
Pylkkänen argues that the low scope illustrated in (54) is root modification, and
therefore change of state verbs must be root-derived. I propose that this modification
is possible because it is compatible with the root type of < e, < ss,t>>.
10.4.3 Double objects
As mentioned above, the roots found in change of state verbs often participate in the
causative–inchoative alternation, whereby they also appear in transitive frames.
Thus, one can ask whether such roots are compatible with double objects in such
a transitive context. As with root creation verbs, these verbs do not inherently
encode any kind of transfer or motion of the object, and thus they are not expected
to be automatically compatible with pure possession double objects. The transitive
variants also do not allow for double objects with a pure benefactive type reading, as
shown in (57):
(57)
a. * The mother cooled her daughter the room.
b. * The tenant opened the landlord the door.
c. * The assistant cleared the executive her schedule.
This is to be expected, as English does not generally allow for pure benefactives.
However, intended possession benefactives could be available. When stative roots
embedded in transitive change-of-state contexts are compatible with a similar
creation interpretation, they also license benefactive double objects:
(58)
The host opened the guest a beer.
(59)
The father cracked his children some walnuts.
(60)
The bartender cleared the patron a spot at the bar.
(61)
Her neighbor grew her a Christmas tree.
The ontology of roots and verbs
227
These benefactives can be derived with the low applicative head as in (62),
representing the vP for (58):
(62)
λess.∃e′se.open(e′) & theme(e′, a beer)
& theme(e′, a beer) & to-the-possession(a beer, the guest)
& change-of-state(e) & CAUSE(e′, e)
λf<ss,t>.λess.∃e′se.f(e′)
& change-of-state(e)
& CAUSE(e′, e)
λes.open(e) & theme(e, a beer)
& theme(e, a beer) &
to-the-possession(a beer, the guest)
vinch
λze.λsss.open(s)
& theme(s,z)
√open
λf<e,<s,t>>.λes.f(e, a beer)
& theme(e, a beer)
& to-the-possession(a beer, the guest)
e
the guest
λye.λf<e,<s,t>>.λes.f(e, a beer)
& theme(e, a beer)
& to-the-possession(a beer,y)
λxe.λye.λf<e,<s,t>>.
λes.f(e,x) & theme(e,x)
& to-the-possession(x,y)
e
a beer
APPL
The applicative head can take the direct and indirect object directly as arguments
and then compose with the root. Unlike root creation verbs, transitive change of
state verbs can have direct objects that are construed as created objects. In this case,
what is construed as being created for the benefit of the indirect object is an open
beer. This is not encoded directly in the semantics, given that change of state verbs
do not always have such entailments, but in the context of an applicative head, the
object is compatible with being construed in this way. For root creation verbs, the
direct object cannot be interpreted as a created object because the root itself denotes
the entity or entities created by the event, and the object is interpreted as the source
from which the object is created.
228
Lisa Levinson
On this analysis, the applicative head takes as its third argument a predicate of
eventualities which is a predicate of states rather than a predicate of events. This
should be possible if the applicative head selects for a head of type < e< s,t>>, with the
eventuality being unspecified for eventive or stative sort. At first this might seem to
conflict with the position in Pylkkänen (2008:18) that, in English, “since low applicatives imply a transfer a possession, they are nonsensical with verbs that are
completely static.” However, in Pylkkänen’s analysis, this requirement is not
encoded directly into the applicative head, and there does not seem to be a reason
that the requirement for eventivity should be stated as a requirement holding
specifically for the argument of APPL. If the constraint can rather be formulated
as a requirement on the vP, this can be satisfied with the introduction of the change
of state v head which introduces dynamic inchoative semantics. This formulation of
the constraint would still allow us to rule out sentences such as (63) which have
stative vPs:
(63) * I held him the bag. (Pylkkänen 2008:ex. 20)
10.4.4 Obligatory theme
Change of state verbs can be intransitive, but since they are unaccusative, their sole
argument in such cases will be the theme, and themes are thus obligatory. This can
be explained by the fact that the root itself selects for an individual argument, as it is
of type < e,< ss,t>>. However, unlike root creation verbs, which have < e,t> roots, it
is possible for the root to combine directly with this individual argument because
it returns a predicate of type < ss,t>, not type t, and thus can combine with the
appropriate kind of v.
10.4.5 Change of state verb summary
In this section, it was shown that change of state verbs do not license pseudoresultative modification, but do require themes and allow applicative arguments. It
was argued that these facts can be attributed to the fact that the root of such verbs is
of the type < e,< ss,t>>.
10.5 Conclusions and future directions
In the introduction to this chapter, it was suggested that, in order to capture
contrasts between verbs which are all derived from category-neutral roots, it is
necessary to make recourse to their semantic type. This has been shown to be a
desirable move, since both apparently syntactic and semantic generalizations have
been shown to depend on the root type. The semantic generalizations are directly
derived from the argument structure and composition of the root with other
elements, while syntactic generalizations arise because of the semantic restrictions
on the composition which lead to different structural contexts for different types of
The ontology of roots and verbs
229
roots. More specifically, I have argued that certain morphosyntactic verb classes
differ due to the type of their lexical roots, as summarized in Table 10.3:
Table 10.3. Summary of Root Types and Verb Behavior
Verb Type
Pseudo-resultatives
Double Objects
Root Creation
Explicit Creation
Change of State
ü
*
*
*
ü
ü
Obligatory Theme
Root Type
Yes
No
Yes
< e,t>
< se,t>
< e,< ss,t>>
The analysis shows that attempting a formal compositional analysis of such decomposition can provide important insight not only into syntax, but also the ontology of
lexical roots and the functional elements they combine with. This work raises the
question of what other kinds of syntactic and argument structure generalizations can
be explained or discovered by this kind of approach, and hopefully future work can
help to answer this question. To the extent that this approach proves to be an
empirically desirable approach to explaining such data, it casts doubt on approaches
which argue that roots are only loosely associated with non-linguistic conceptual
meaning or semantically vacuous in the absence of additional structure.
It was also shown that it is crucial in determining the properties of verbs and verb
classes to carefully control for structural ambiguities “below” the word level—for
example, it was seen that there are two structurally distinct verbs braid, and there is
polysemy with respect to the root √braid. This predicts that what we have previously
considered polysemy is tied up with structure in a way that should also affect
linguistic processing. From this perspective, polysemy breaks down into three
different categories: ‘normal’ polysemy, structural ambiguity, and polysemy plus
structural ambiguity.
11
Derivational affixes as roots: Phasal
Spell-out meets English Stress Shift
JEAN LOWENSTAMM
11.1 Introduction
This chapter1 is devoted to the elucidation of a puzzle: under current assumptions,
Distributed Morphology (henceforth DM) stalls when confronted with a great
classic of English grammar, possibly the most central fact around which the theory
of SPE was built, Stress Shift.2 English Stress Shift, documented in (1), is the
phenomenon whereby stress can be seen to move progressively rightward as affixes
are added to a base.
(1)
átom, atómic, atomícity
I claim that two assumptions, both unnecessary, indeed foreign to DM, are responsible for the apparent inability of DM to handle Stress Shift. When those assumptions are discarded and DM is left to draw on the resources of its own conceptual
toolbox, not only can it handle Stress Shift; it can actually do a better job of it than
previous theories. The two assumptions to be done away with appear in (2).3
(2)
i. “derivational” affixes are categorial exponents
ii. domains of Phasal Spell-out are the same thing as the cycles of SPE
1
For help and encouragement, I am grateful to Artemis Alexiadou, Hagit Borer, Radwa Fathi, Noam
Faust, Patty Garet, Brenda Laca, Victor Manfredi, Andrew Nevins, Khadija Qandisha, Tova Rapoport,
Nick Tasker, to the reviewer for this volume who remained anonymous, and to audiences at Universität
Stuttgart, the University of Tromsø, Tel Aviv University, MIT, Universität Wien, and Université d’Orléans. The influence of the seminal and inspiring work of Lisa Selkirk (Selkirk, 1982), Paul Kiparsky
(Kiparsky, 1982a), and Alain Kihm (Kihm, 2005) will be felt throughout this chapter.
2
For background relevant to the interface issues dealt with here, cf. Borer (this volume), De Belder,
Faust and Lampitelli (this volume), Embick and Marantz (2008), Embick (2010), Gallego (this volume),
Piggott and Newell (2008), and references therein.
3
(2ii), in fact, follows from (2i) in an obvious sense, but that relationship is not further explored in the
context of this chapter.
Derivational affixes as roots
231
Because of space limitations, the assumptions in (2) will be confronted neither
directly nor extensively: (2i) will be treated briefly in the remainder of this introductory section, and (2ii) in the conclusion. Rather, the argumentation will take a
different form: I will offer an account of Stress Shift which, while well within the
spirit of DM, is entirely incompatible with (2i) and (2ii). To the extent that this
account is convincing, it ipso facto carries a refutation of both assumptions in (2).
Exponence is the relationship between a phonological string and the set of
features it signals. While there is a vast and sophisticated literature on types of
exponence (zero exponence, cumulative exponence, parasitic exponence, etc.), many
of the fundamental initial assessments that define problems in this area rest on gut
feeling plausibility, most notably as regards exactly what a particular string actually
signals. For instance, based on the Spanish inflectional sample in (3), most investigators would agree without much debate that < o> and < a> signal gender, while < s>
signals plural.
(3)
a. perro ‘dog’
b. perra ‘female dog’
c. perros ‘dogs’
d. perras ‘female dogs’
There seems to be a comparably confident consensus around the exponence of
categories. Thus, < ian> in reptilian is reputed to signal adjectivalness, < ory> nounness in promontory, etc. Yet, in sharp contrast with the gender and number exponents in (3), such ‘categorial exponent’ strings often carry much more information
than would be necessary for the strict expression of the morphosyntactic features
they supposedly signal. This is clear for instance with < ful> or < less>. Moreover, the
categorial connection of some of those strings is often ambiguous. Thus, while < ian>
signals adjectivalness in reptilian as we just saw, it also “signals” nounness in
librarian; similarly, < ory> can signal adjectivalness in rotatory, the fact that it was
seen to signal nounness in promontory notwithstanding, etc.
The reasoning behind the idea that categories receive expression seems to have
been something like this: a) atom is not an adjective, b) atomic is an adjective, c)
therefore <ic> signals adjectivalness (and, in this case, < ø> signals nounness).
Perhaps, for a word based theory, this is as good a starting point as any. But things
are rather different with √&c (“root and category”) theories. 4 In such theories, roots
are selected by a category defining head, as shown in (4) with the example of the
selection of root √FAT by adjectival head a.
4
For short, I call √&c (root and category) theories those theories, such as Borer’s (Borer 2005a, 2005b)
as well as DM, that a) subscribe to the view that roots undergo categorization as the consequence of their
selection by a category defining head, a, n, or v, and b) explore the consequences of this idea in the context
of a theory of word formation countenancing no active lexicon. See Williams (2007) for a forceful rejection
of the idea that there is no such thing as an active lexicon endowed with properties such as described in
Chomsky (1970), and Borer (1998) for an overview of issues connected with the nature of morphological
operations.
232
Jean Lowenstamm
(4)
aP
a
√FAT
In the scheme in (4), the presence of a is the necessary and sufficient condition for
string < fat> to be the exponent of an adjective. This has the potential to completely
reverse the perspective on say, a sample such as (5).
(5)
foppish, sexy, courageous, Obama-esque, atomic, brutal, golden, fat
In most accounts, fat would be viewed as the odd man out because it is unsuffixed
(or ø suffixed). By contrast, from the point of view of a √&c theory, fat must be
viewed as the archetypal adjective as its makeup involves nothing but what such
theories explicitly define as the necessary and sufficient ingredients of adjectivalness.
In the rest of this chapter, I will pursue the hypothesis that categories have no
exponents. Thus, the view I will put forth differs as follows from the classic take
represented in (6a) where < ic> is the spell-out of a: while I endorse the view that
atomic owes its adjectivalness to a, I reject the view that < ic> signals a. Rather, I will
claim that < ic> is itself a root, viz. √IC.5 This alternative is represented in (6b).
(6)
a.
aP
a
<ic>
b.
√ATOM
aP
a
√P
√IC
√ATOM
Saying that nothing commits √&c theories to the view that affixes signal categories
hardly entails that affixes make no contribution of their own. For instance, there is a
clear difference in meaning between adjectives such as siltic and siltous whereby the
presence of silt in a geological layer will be seen as more fundamentally characteristic
of that layer if the layer is said to be siltic than siltous.6 But how much of that
contribution is bound to the adjectivalness of siltic or siltous? Not much, evidently.
Consider the ingredients involved: a, < ic> and < ous>, and √SILT. If the intuition
that < ic> and < ous> behave as operators of restriction on √SILT is correct, then the
scope relations of the relevant ingredients are as in (7a), not as in (7b).
5
For implementations of this idea, see Arbaoui (2010), Lampitelli (2011), Faust (2012), Fathi (2013),
De Belder, Faust and Lampitelli (this volume), Molu (in preparation). For an extension to the treatment of
inflection, see Lowenstamm (2011).
6
This resource was put to systematic use by Guyton de Morveau et al. (1787).
Derivational affixes as roots
(7)
a.
233
b.
a
<ic>
√SILT
a
<ic>
√SILT
Indeed, it is not even crucial for the restriction operation performed by < ic> and
described in (7a) to hold, that it be implemented as an adjective.7
Note, more generally, that if derivational affixes signal categories as claimed by
most, it is a mystery why their performance as exponents is marred by such rampant
ambivalence as exemplified in the sample in (8). But on the view that they are not
categorial exponents, the mystery vanishes: why should they be unambiguous with
respect to category?
(8)
Exponent
able
al
an
ant
ary
ate
ible
ic
ive
ory
ous
esque
y
ish
en
ful
Noun
constable
mammal
librarian
defendant
functionary
consulate
crucible
tunic
incentive
promontory
focus
arabesque
parsimony
rubbish
warden
handful
Adjective
endurable
normal
reptilian
defiant
legendary
intricate
credible
magic
auditive
rotatory
mucous
grotesque
airy
foppish
golden
colorful
The rest of this chapter is devoted to showing that the dissociation of category and
affix advocated in (6b) paves the way for a successful account of Stress Shift in
7
A sense of the proliferation of nouns ending in < ic> alongside homophonous adjectives (the basics or
honorifics type) can be gathered from a cautionary note by John Geissman, Vice-President of the
Geological Society of America (Geissman, date unknown):
An increasing number of GSA members lament the general deterioration in the quality and clarity of
writing by earth scientists [ . . . ] Insofar as it is one of the duties or prerogatives of editors to educate
potential or eventual authors, when necessary or appropriate, we offer this commentary as some of our
suggestions to authors [ . . . ]
• We may say volcanics, clastics, metamorphics, [ . . . ], and granitics to each other in the field, but it is
quite improper grammatically to add an s to an adjective to make a plural noun. It may be tedious or
repetitious to read, but it is correct and unambiguous to write volcanic rocks, clastic rocks, [ . . . ], and
granitic rocks [ . . . ]
234
Jean Lowenstamm
English. It comprises three sections, a conclusion, and a short appendix. In the first
section, I establish that current versions of DM cannot handle Stress Shift. In section
11.2, I develop a DM compatible alternative. In a third section, the alternative is put
to the test of an old riddle of English grammar: why does affix +al attach to X-ment
type nouns if X is not a verb (segment/segmental), but not if X is a verb (employment/
*employmental)? Why is it not just the opposite? It is shown that this apparently
puzzling state of affairs in fact follows as a prediction of my proposal. In the
conclusion, I return to the difference between phase and cycle. The appendix briefly
deals with a generalization put forth in Fabb (1988).
11.2 Stalling
11.2.1 Phase Impenetrability, Head Movement, and Phasal Spell-out
In this preliminary subsection, I am concerned with two technical aspects of Phase
Theory and how they crucially interact with Spell-out: Phase Impenetrability and
Head Movement. With most investigators, I assume the following.
At a given phase, the complement of the phase head is spelled out. Thus, Z in (9a)
will be spelled out at Phase 1, but not X and Y. Moreover, by Phase Impenetrability,
the spell-out of the complement of a phase head cannot be influenced by material
located in a superordinate phase. Thus, L can play no role in the spell-out of Z. On
the other hand, phase edge material (the phase head itself, its specifier and possible
adjuncts) can be accessed from the next higher phase. Accordingly, the spell-out of
Y, X, and W can take into account material contained in L. As Marvin (2003)
correctly notes, Head Movement can potentially interfere in undesirable fashion
with Phase Impenetrability as just defined. For instance, in (9b), Z has undergone
Head Movement, left-adjoined to Y, and consequently been removed from the scope
of Y and brought into the scope of L. Two consequences follow: a) Z can no longer
be spelled out at Phase 1, b) its spell-out can be influenced by L.
(9)
LP Phase 2
a.
L
YP
L
YP Phase 1
W
LP Phase 2
b.
X
YP Phase 1
W
Y⬘
Y
YP
X
Z
Y⬘
tz
Y
Z
Y
Derivational affixes as roots
235
In order to make sure that repeated Head Movement will not totally bleed the
combined effects of Phase Impenetrability and Spell-out, actually rendering the
former vacuous with respect to the latter, Marvin (2003) proposes that Z continue
to be viewed as belonging to the domain of Y, even if it has undergone Head
Movement and left-adjoined to Y, as in (9b). This move is clearly unfortunate as it
now empties the intended combined effects of Head Movement and Phasal Spell-out
of any empirical content: Head Movement and Phasal Spell-out stand in an asymetrical relationship of potential bleeding whereby the former can potentially bleed the
latter. Either it bleeds it, in which case Bleeding is expected to obtain and Spell-out is
thwarted, or it doesn’t (Counterbleeding) and Spell-out can take place, but you can’t
have both Bleeding and the effects of Counterbleeding.8
While the next subsection establishes that Stress Shift stalls, it is shown in the
following subsection that the operation of Head Movement is an entirely orthogonal
issue.
11.2.2 Stress Shift stalls
Consider átom, atómic, atomícity where stress moves forward as affixes, first +ic,
then +ity, are added to atom. For the sake of completeness, two possible analyses of
atomicity will be considered, and it will be shown that, under either analysis, the
phasal scenario blocks the derivation of the correct output, viz. main stress on the
antepenult. The two analyses differ with respect to the < atomic> substring: under
one (10a), atomic is a denominal adjective; under the alternative (10b), atomic, this
time construed as a deradical adjective, directly results from the merger of √ATOM
with little a. Z in (10b) is a phase head merely brought in to provide context.
(10)
nP Phase 3
a.
aP Phase 2
n
ity
nP Phase 2
Z
nP Phase 1
a
ZP Phase 3
b.
ic
aP Phase 1
n
ity
n
ø
√ATOM
a
√ATOM
ic
Suppose, following Marvin (2003), that phase heads trigger the spell-out of their
complement. In that case, both in (10a) and (10b), the root will spell out. [ÁD@m]
will result, with initial stress frozen there, and no possibility of moving it forward at a
8
See Kenstowicz and Kisseberth (1977) for discussion of Bleeding.
236
Jean Lowenstamm
further phase, hence *[ÁD@mĭkĭDĭ].9 Suppose alternatively, following Embick
(2010), that phase heads trigger the spell-out of their cyclic/phasal complement
only. (10a) and (10b) now produce distinct outputs, both ungrammatical. Since the
complement of Phase 1 in (10a) contains no cyclic/phasal material, spell-out only
takes place at phase 2, and stress is frozen on [ÁD@m] again. Again, *[ÁD@mĭkĭDĭ]
will ultimately result. Under the alternative view in (10b), spell-out is delayed until
Phase 2 for the same reason as in (10a). When it takes place, [@thómĭk] results, with
stress frozen on the penultimate syllable. This time, *[@thómĭkĭDĭ] is the outcome.
Both sets of outputs are summed up in (11).
(11)
Marvin (2003)
Embick (2010)
a. [aP a [nP n √ATOM]]
*[ÁD@mĭkĭDĭ]
*[ÁD@mĭkĭDĭ]
b. [aP a √ATOM]
*[ÁD@mĭkĭDĭ]
*[@thÓmĭkĭDĭ]
As the derivation of atomicity along the lines of what precedes increasingly looks like
a slip knot that binds closer the more it is drawn, I will attempt to loosen the noose by
shifting the perspective, thereby raising the stakes to some extent. That is, I will try to
understand why Stress Shift blocks by means of a comparative discussion of atomícity
with another staller, viz. atómicness. Head Movement will come out exculpated.
11.2.3 Head Movement is out of the loop
Here, I consider the derivations of atomícity and atómicness, and I show that,
paradoxically, each derivation requires a generalization to be true, which the other
requires to be false, viz. (12).
(12)
i. Head Movement should be allowed to bleed spell-out.
ii. Head Movement should not be allowed to bleed spell-out.
Let us see how Head Movement might be invoked in order to rescue the derivation
of atomícity from an input such as (13).10
aP Phase 3
(13)
aP Phase 2
a
{less}
aP Phase 1
n
{ity}
a
{ic}
√ATOM
9
For the sake of clarity, a discussion of the patterns of vowel reduction and velar softening (both of
which would only make the point even more dramatic), has been left out.
10
For easier identification of the various positions in the structure, the sites of eventual insertion of
vocabulary items have been filled in with the items themselves, the curly brackets denoting the anticipatory nature of this mention.
Derivational affixes as roots
237
Suppose Head Movement left-adjoined the root to a, as shown in (14). The root,
now having joined the edge of Phase 1, can only be spelled out at Phase 2.
aP Phase 3
(14)
nP Phase 2
a
{less}
aP Phase 1
n
a
√
t
a
√ATOM
But, in order to see how a desirable scenario would proceed, let us just allow Head
Movement to take place again, and left-adjoin the contents of aP to n, the head of
Phase 2, as shown in (15): Head Movement has now removed aP from the complement of Phase 2 and into the latter’s edge, thus delaying spell-out until Phase 3. At
that point, i.e., at Phase 3, the ingredients of atomicity can be spelled out without any
of its pieces having already been frozen by spell-out at earlier phases.
aP Phase 3
(15)
nP Phase 2
a
{less}
aP Phase 1
n
a
n
√ATOM
a
atom
ic
t
t
ity
Clearly, in this case, the correct result is attained when Head Movement is allowed to
proceed unimpeded, and left-adjoined structure is duly treated as edge material for
purposes of spell-out.
But tampering with Phasal Spell-out, in effect letting Head Movement bleed it, as
was just done, will not help even a little bit in view of the need to derive not just
atomícity, but atómicness as well. In order to derive atómicness, left-adjunction of aP
238
Jean Lowenstamm
to n (16b), the very same scenario with double Head Movement which successfully
delayed spell-out of the ingredients of atomícity until Phase 3, now appears to be
precisely what has to be avoided, lest *atomícness result.
aP Phase 3
(16) a.
nP Phase 2
a
{less}
aP Phase 3
b.
aP Phase 1
n
{ness}
a
nP Phase 2
a
{less}
aP Phase 1
n {ness}
a
√
n
t
t
√ATOM
a
√ATOM
a
∗[atomícness]
[atómicness]
That is, the < atomic> substring of < atomicness> MUST be spelled out at Phase 2
(16a)—no later—to ensure that < atomicness> firmly bears stress in the same place as
atómic.
As we can see, the set of decisions that bring about the success of one derivation
stand in the way of the other, and vice versa. Of course, the paradox arises not
because of Head Movement, but because the respective inputs to atómicness and
atomícity are not distinct: in both cases, the input is (13), repeated in (17).
aP Phase 3
(17)
a
{less}
aP Phase 2
aP Phase 1
n
{ness/ity}
a
{ic}
√ATOM
The next subsection is a brief elaboration of the need for more flexibility than is
afforded by inputs exclusively defined in categorial terms, and then subjected to too
crude a version of Late Insertion.
t
Derivational affixes as roots
239
11.2.4 Backing up a bit
Much of the work on English Stress Movement is informed by two assumptions,
most influentially propounded in Chomsky and Halle (1968) and subsequent elaborations.
First, the view that the structure of a complex word such as atomicity is as in (18),
or some version of (18), whereby the noun atomícity “contains” the adjective atomic,
and, perhaps, the adjective atomic “contains” the noun atom.11
(18) a. [N ity [Adj ic [N atom]]]
So with atómicness, represented in (19).
(19) [N ness [Adj ic [N atom]]]
The second assumption has to do with the proper treatment of the differential
behavior of the two classes of affixes that so strikingly pervade the accentual pattern
of the language, cf. Newman (1946). In pre-Phasal Spell-out theories, the differential
impact on stress of the various affixes is encoded in a variety of ways: by means of
different boundaries, by assigning affixes to different lexical strata, etc. But, to the
best of my knowledge, all authors assume that the manner of attachment or location
of affixes are properties of the affixes themselves: again, some attach close; some
don’t; some are cyclic; others are not, etc. For instance, consider adjectives such as
governmental, objectionable, leaderless, and representationary. While their makeup is
the same as regards the categories involved and their hierarchical arrangement, each
adjective represents a different configuration of cyclic and non-cyclic domains,
where the cyclic or non-cyclic character of a particular domain is directly linked
to the specific affix heading that domain.12 Here, Distributed Morphology MUST
make a different assumption. Indeed, in a framework endorsing Derivation by Phase
and Late Insertion, such richness of information as is packed in (20a) cannot be
available. For, by the time spell-out takes place, all four adjectives have exactly the
same structure, viz. (20b).
(20) a. [[[govern V] ment N] al Adj]
b. [a [n [v √]]]
[[[object V] ion N] able Adj]
[[[lead V] er N] less Adj]
[[[represent V] ation N] ary Adj]
At the risk of belaboring the obvious: in pre-Phasal Spell-out theories, domains of
phonological interpretation (cycles) are projected from properties of affixes. In DM,
11
√&c theories do not necessarily endorse the second part of the conjunct here, but I leave it as such
for the sake of the argument.
12
The domain of cyclic affixes has been noted by large, boldface square brackets in (20a).
240
Jean Lowenstamm
in sharp contrast, domains of phonological interpretation (phases) are defined in
strictly categorial fashion, and irrespective of what particular Vocabulary Item may
eventually ornate a given category.
Since discriminations of the kind illustrated in (20a) are undeniably crucial
(however they may be handled), the challenge to DM includes—to return to the
crucial data of the previous section—two facets, not just one: a) to be able to crank
out atomícity, at all; but also b) to find an alternative way of building into the
grammar a distinction such that both atomícity and atómicness be derived. In the
next section, I offer just such an alternative. As announced, it rests on a radical
elaboration of the divorce between category and ‘affix’.
11.3 An alternative
The alternative is (21).
(21)
Affixes are roots.
Because atom, atomic, and atomicity have figured prominently in the previous
section, I begin the exposition of my proposal with the same data.
11.3.1 Affixes as roots: A first pass
For the sake of comparison, I represent my proposal for atomic in (22a,b), along with
the more classical take of mainstream √&c work in (22c).
(22) a.
b.
√P
√IC
√ATOM
aP
a
c.
√P
√IC
aP
a
√ATOM
√ATOM
ic
In (22a), two roots √IC and √ATOM have merged, leading to the formation of a
complex root, √P. Further mergers must take place. For instance, the complex root
can merge with a category-defining head, say a, as in (22b), leading up to the
formation of an adjective, atomic. Alternatively, the complex root can merge with
another root, say √ITY, and the even more complex root in (23a) is formed. If that
root, in turn, merges with a category-defining head, n in (23b), a noun is formed,
atomicity.
Derivational affixes as roots
(23) a.
√P
ZP Phase 2
b.
√IC
nP Phase 1
Z
√P
√ITY
241
n
√ATOM
√P
√ITY
√P
√IC
√ATOM
It can readily be seen that no ingredient of the complex root will undergo spell-out
until merger with n. When spell-out takes place at Phase 1, the rules of English
phonology kick in, and apply cyclically on each root. All roots are cyclic domains,
though as we will soon see, it is a theorem of the system proposed in this chapter that
cyclic phonology, in any derivation, will be exclusively observable at the first phase,
and nowhere else.
Before I offer a more systematic and detailed presentation of root types and what
makes roots stick, in the next subsection, the reader may note what my proposal,
(24b), shares and does not share with the classic view (24a).
(24)
a. classic view:
b. proposal:
[N ity [Adj ic [N atom]]]
[n[√ ity [√ ic [√ atom]]]]
The embedding in (24b) parallels that in (24a) minus the intermediate categorial
labels present in (24a). That is, I claim that the radical material of atomicity contains
the radical material of atomic, and that the radical material of atomic contains the
radical material of atom; not that the noun atomicity contains the adjective atomic,
or that the adjective atomic contains the noun atom. Just enough, in other words, to
capture lexical relatedness, no more. This appears to fit with an important observation of √&c inspired work, namely that configurations involving a local relationship
with the root are often assigned non-compositional meaning. The data in (25) shows
that non-compositionality is indeed rampant where +ic, +al, and +ity are involved.
This follows from my proposal, though not from (24a).
(25)
composition
globe
mode
form
atomic
compositional
global
final
modal
formal
atomicity
globality
finality
modality
formality
242
Jean Lowenstamm
virtue
feud
class
tone
fundament
consequence
congression
emotion
margin
fate
semen
orient
substance
function
origin
person
province
sentiment
neuter
convention
universe
verb
artist
meter
virtual
feudal
classic
tonal
fundamental
inconsequential
congressional
equal
emotional
marginal
fatal
seminal
oriental
bestial
substantial
functional
moral
liberal
mental
oral
original
personal
principal
provincial
sentimental
superficial
technical
neutral
conventional
universal
verbal
vital
artistic
metric
tonality
equality
marginality
fatality
bestiality
functionality
morality
liberality
mentality
orality
originality
personality
principality
superficiality
technicality
vitality
In the next subsection, I turn to nuts and bolts.
11.3.2 Free roots, bound roots, and what makes them stick
Affixes are usually called “bound” morphemes. If affixes are roots, as I claim, they
must be “bound” roots. I propose to capture the difference between bound and free
roots as in (26).
Derivational affixes as roots
(26)
243
i. Some roots can project to the phrasal level on their own, e.g., √BOTTLE,
√RUG.
ii. Other roots, e.g., √AL, √MENT, √NESS, etc., cannot project to the phrasal
level without the help of a complement.
The boundedness of a root will be captured as follows: a bound root bears an
uninterpretable feature which it seeks to check by merging with a complement.
Not until the uninterpretable feature has been checked, can the bound root project at
the phrasal level, and merge with a category-defining head. Two roots appear in (27),
one is free, (27a); (27b) the other, is burdened with an uninterpretable feature, [u √].
(27)
a. √RUG
b. √IC
[u √]
√RUG as such is fit for phrasal status, hence for merger with a category-defining
head, as shown in (28a). √IC alone cannot undergo merger with a category-defining
head as long as it has not rid itself of its uninterpretable feature (28b).
(28)
a.
b.
n
∗
a
√P
√IC
[u √]
…√RUG…
On the other hand, when the uninterpretable feature has been checked owing to
the presence of an appropriate complement, phrasal status is attained, and
merger with a category-defining head can take place, (29a). However, merger with
a category-defining head is not the only option, at that point. Indeed,
[√P √IC √ATOM] can alternatively merge with another “bound root” also in need
of checking its uninterpretable feature, for instance [√P √ITY], as shown in (29b). In
turn, [√P √ITY [√P √IC √ATOM]] will merge with a category-defining head, say n.
(29) a.
aP
a
b.
√P
√IC
[u √]
nP
n
√ATOM
√P
√ITY
[u √]
√P
√IC
[u √]
√ATOM
244
Jean Lowenstamm
The reader will have noticed that the affixes discussed in this subsection are all
typical stress shifters, the Class 1 affixes of Siegel (1974), or the Level 1 affixes of
Kiparsky (1982a). In my proposal, they are [u √] affixes. What is the difference? The
difference lies in the source of the label. The usefulness of recognizing Class 1 affixes
is the possibility it affords of capturing their impact on the stress pattern of the
language. But, at the same time, much of the evidence on which membership in that
class is decided comes from the accentual system of English itself. As a result the
distinction between Class 1 and Class 2 incorporates a measure of circularity. By
contrast, the proposal put forth here, while it also aims at capturing significant
generalizations about stress, rests on considerations that have nothing to do with
stress, namely the selectional behavior of affixes: an affix (strictly speaking a root)
carries a [u √] feature because it selects roots. That +al, +ic, and +ity select roots can
be determined by inspection of a sample such as (30).
(30)
frugal, drastic, calamity
That the characterization of the selectional targets of +al, +ic, and +ity was carried
out in total independence of stress facts can be verified by means of a comparison
with French: inspection of the sample in (31) indicates that French +al, +ique, and
+ité also select roots. Of course, the stress system of the language, exceptionlessly
final, could not possibly have provided any clue as it is indifferent to affixation type,
or even affixation at all.
(31)
frugal, drastique, calamité
A more detailed comparison with Lexical Phonology will be offered in 11.2.4 and
11.2.5. In the next subsection, I turn to the place of Class/Level 2 affixes in a system
such as advocated here where affixes are construed as roots.
11.3.3 Another type of bound root (Class/Level 2 affixes)
Any theory must specify the relationship between the following characteristics of
Class/Level 2 affixes with regard to the accentual system:
(32)
i. They attach outside Class/Level 1 affixes.
ii. They have no impact on the stress pattern of their complement.
I propose that Class 2, or Level 2 affixes, select xP’s. Accordingly, their uninterpretable feature is [u xP].
In addition, I accept (33).
(33)
Categories head roots, not vice versa.
(33) merely reflects the canonical ordering of projections whereby it is the fate of
roots to become categorized (34a), whereas it is not the fate of categories to be turned
into roots (34b).
Derivational affixes as roots
(34)
a.
xP
x
b.
√
245
√P
xP
√
The inclusion of (33) hardly adds to the cost of my account, as (33) or something to
the same effect has to be part of any √&c theory. On the other hand, as the reader
will note, it installs a tension at the heart of my proposal to the extent that (33)
appears to rule against the claim that an entire class of roots—Class 2 or Level 2
affixes—are specified, as was just suggested, as selecting little xP’s.
Take the example of moneyless, for instance. (35a) is ruled out by the fact that the
putative merger of √LESS and √MONEY does not lead up to the elimination of the
uninterpretable feature of the head root, thus correctly capturing the fact that -less is
non-cohesive. But note that when √LESS successfully checks its uninterpretable
feature by merging with the little noun money (35b), a configuration is created
which runs directly counter to (33), with a root now heading a category.
(35)
a.
b.
√LESS
[u xP]
√P
√LESS
[u xP]
√MONEY
nP
√P
n
√MONEY
I submit that the resolution of the tension inherent in (35b) proceeds as in (36): upon
merger with a category-defining head (36a), here a, √LESS left-adjoins to it (36b).
(36) a.
a
aP
b.
aP
a
√P
√LESS
[u xP]
nP
√LESS
√P
a t√
nP
√P
n
√P
n
√MONEY
√MONEY
This move is desirable on two counts: a) √LESS, having been removed from the
scope of spell-out at aP, will be stressed separately from money, b) moneyless is
246
Jean Lowenstamm
entirely compositional, as expected in view of the non-local relationship between the
two roots involved.13
As a preliminary summary, it can be noted that roots and categories are arranged
as in (37), the structure corresponding to, e.g., atomicitylessness.
xP Phase 3
(37)
x
√P
xP Phase 2
√
[u xP]
x
√P
√
[u xP]
xP Phase 1
x
√P
√
[u √]
√
[u √]
√P
√
…√…
The organization in (37) and its consequences for cyclicity will be returned to and
discussed specifically in the conclusion. For the time being, I only wish to draw
attention to the formal similarity between (37) and the results of Lexical Phonology.
In Lexical Phonology, Class 2 affixes are ordered outside of Level 1 affixes (as
a consequence of the ordering of the respective levels to which they pertain). In
(37), roots equipped with a [u xP] uninterpretable feature correspond to Level 2
affixes, roots equipped with a [u √] correspond to Level 1 affixes. They form blocks
ordered as in LP: first [u √] roots, then [u xP] roots. It might thus seem at this point
that my proposal has merely succeeded in reproducing the classic Level Ordering
segregation of Lexical Phonology. That is correct, but in part only. In the next
13
This captures the basic insights of Kaye (1995) with respect to the distinction he draws between
analytic and non-analytic domains.
Derivational affixes as roots
247
section, I directly show how my system does well where LP did well, yet does well
too, where LP did less well.
11.3.4 The bane of Lexical Phonology
It has repeatedly been pointed out that The Level Ordering Hypothesis inherent in
Lexical Phonology was too strong. As noted by Kaisse (2005), it follows from the
architecture of Lexical Phonology that *happy#ness+al or *sing#er+ous are impossible English words. But, by the same token, the Level-ordering Hypothesis incorrectly rules out governmental or neutralization.
In order to gain perspective on the issue (and to be fair to LP), it must be said that
while LP incorrectly rules out govern#ment+al, it correctly rules out *belittle#ment
+al. Note that the problem of how to rule one in and the other out, would be readily
solved if [governmental] could be analyzed as /govern+ment+al/, and *[belittlemental] as /belittle#ment+al/. This is, of course, difficult to contemplate in an SPE type
theory or in LP because in such theories, affixes are cyclic or non-cyclic (or, Level 1
or Level 2) by virtue of a lexical stipulation to that effect. It is clear that those theories
would have lost much of their content, had they stated that an affix could be cyclic,
non-cyclic, or both, or Level 1, Level 2, or both. And yet, the sort of flexibility just
sketched out seems to be called for in a number of other cases, as shown by Mark
Aronoff.
Aronoff (1976) makes a number of extremely interesting observations regarding
the ambiguous behavior of able/ible.14 He notes that a number of -able adjectives can
be stressed in more than one way, to wit (38).
(38)
a. cómparable
réparable
réfutable
préferable
dísputable
b. compárable
repá(i)rable
refútable
preférable
dispútable
The adjectives in (38a) are stressed according to a classic generalization (Halle 1973)
whereby the affix is ignored on account of its short vowel, and then the Primary
Stress Rule affects a heavy penult (refrángible), or the antepenult in case the penult is
light (córrigible). By contrast, the adjectives in (38b) directly contravene this generalization. Rather, they are stressed exactly like the verbs from which they are presumably derived. Aronoff observes that the optional patterns evidenced in (38)
would follow if the affix were preceded by a + boundary in (38a), e.g., compar
+able, but by a # boundary in (38b), e.g., compar#able.
14
The discussion of -able/ible conducted here is entirely drawn from Aronoff (1976).
248
Jean Lowenstamm
Moreover, as Aronoff notes, cómparable and compárable do not mean exactly the
same thing: compare#able (38b) is entirely compositional, whereas compare+able
(38a) can also mean ‘equivalent’. Hence the contrast in (39).
(39)
a. This is the cómparable model in our line
b. *This is the compárable model in our line
Further, Aronoff notes that the patterns of allomorphy in (40) are untypical: -able
normally patterns like -ion, -ive, -ory, and -or with respect to the selection of the
allomorphs of their base, with the difference that -able optionally selects marked
allomorphs (40b) as well as bona fide verbs (40c). Indeed, -able can also select fullblown verbs, as shown in (40c). Again, this dual behavior can be readily accounted
for if the affix is separated from its complement by a + boundary in (40b), but by a #
boundary in (40c).
(40)
a. circumscribe
extend
defend
perceive
divide
deride
b. circumscriptible
extensible
defensible
perceptible
divisible
derisible
c. circumscribable
extendable
defendable
perceivable
dividable
deridable
As it turns out, as Aronoff shows, this correlates in striking fashion with another
generalization due to Ross (1974, 1979): ible/able adjectives “with no lexical base”
(= deradical) frequently allow prepositional complementation, (41).
(41)
a. I am amenable to a change in plans
b. He is eligible for reappointment
This is in contradistinction with the behavior of productively derived deverbal
formations which are much stricter in that respect (cf. Aronoff 1976). In the light
of this last observation, it is significant that, when put to the test of whether they
tolerate prepositional complements, the sample in (40) clearly shows how the
adjectives in (40b) pattern like deradicals such as amenable or eligible, in sharp
contrast with those in (40c):
(42) a. divisible into three parts
b. *dividable into three parts
Furthermore, the adjectives in (40b), unlike those in (40c), have non-compositional
meaning, as can be seen in (43).
(43)
a. There is a flaw in the grain, but it’s imperceptible
b. *There is a flaw in the grain, but it’s unperceivable
To sum up, Aronoff ’s observations on the ambiguous behavior of -able strongly
highlight the need to recognize two modes of attachment for that affix: +able
Derivational affixes as roots
249
and #able. In pre-√&c theories, this has the status of a paradox. In the next
subsection, I show how my proposal makes room for such an ambiguous mode of
attachment.
11.3.5 The third kind of root
One of the threads running through the proposals put forth so far has been that the
distinction between Level 1 and Level 2 affixes is not a raw fact to be captured by
stipulation. Rather, the distinction is entirely derivative from an independent phenomenology, namely the selectional targets of bound roots: root selectors, e.g., √IC,
√ITY, √AL, etc., correspond to Level 1 affixes, while xP selectors, √NESS, √LESS,
√FUL, etc., correspond to Level 2 affixes.
A prediction ensues: if another type of selectional behavior than has been recognized up to this point (selecting roots vs. selecting xP’s) can be identified, that third
type will motivate a principled tripartite nomenclature of bound roots. Hopefully, it
will shed some light on the interaction between affixes, presumably affording
insights not available under the excessively rigid, classic, two-pronged Level 1/
Level 2 (or cyclic/non-cyclic) distinction.
Such a third type indeed exists, the universal selector. It selects both roots and
xP’s. Accordingly, its uninterpretable feature is [u X] where X stands for an
underspecified complement (i.e., of either kind, √ or xP). -ment, -able, -ize are
examples of the universal selector. -able was discussed above, and it is clear how
Aronoff’s observations directly translate into a √&c framework. -ment, the topic of
the next section, is another example of a universal selector. That -ment selects both
roots and vP’s can be seen from the sample in (44). In (44), I have deliberately
restricted the range of examples to cases where the complement of ment can only
be a root, or a verb (the argumental apparatus contributed by the prefixes being the
guarantor of full-blown verbhood).
(44)
√
liga-ment
monu-ment
medica-ment
frag-ment
instru-ment
seg-ment
supple-ment
vP
an=nul-ment
be=little-ment
en=throne-ment
dis=courage-ment
ap=praise-ment
de=fraud-ment
en=force-ment
In the next section, I show how the intricate and apparently paradoxical selectional restrictions controlling the distribution of -ment, are actually predicted by my
proposal.
250
Jean Lowenstamm
11.4 A case study: Ornamental vs. *belittlemental, and the morass
that lies in between
Aronoff (1976), Fabb (1988), and others, note that +al readily attaches to nouns
ending in +ment, to wit instrument/instrumental, segment/segmental, etc. However,
+al resists attaching to an X+ment noun if X is a verb, thus, *contain-ment+al,
*attach-ment+al, *pay-ment+al, etc. This generalization is a very serious challenge
to the locality proviso of ‘syntactic morphology’ inasmuch as successful selection of a
complement by +al would require +al to have access, across +ment, to properties of
what the latter has attached to, viz. a verb or something else. The non-local character
of the alleged dependency is illustrated in (45). Again, in order to attach to +ment,
+al has to check what lies below +ment: if it is a non-verb, say seg in (45a), a wellformed object results, segmental; on the other hand, if it is a verb, say contain in
(45b), an ungrammatical adjective, *containmental, is derived.
(45)
a.
b.
al
ment
ment
~V
seg
al
ment
ment
V
contain
In pre-√&c frameworks, the relationship can be construed as non-local or bottom-to-top. Thus, Aronoff (1976)—quite plausibly—sees the restriction under discussion as requiring inspection of the internal constituent structure of the
complement of < al>, while Fabb (1998) or Plag (2003) explicitly contemplate the
option of having the complement select its own head. Neither option is available in a
Minimalist framework.
Within the confines of the proposal advocated here, the solution is straightforward, as will be shown by means of a comparative discussion of the successful
derivation of segment+al, vs. the unsuccessful derivation of *belittlement+al.
When √MENT heads √SEG, a √P is formed. Subsequent merger with n turns the
whole into an nP (46a). But another option, demonstrated in (46b), is available. The
complex object [√P SEGMENT], by virtue of being a root, can itself be selected by
√AL. In such a case, an even more complex root is formed, [√P SEGMENTAL].
Upon merger with a, an adjective is formed, segmental.
Derivational affixes as roots
(46)
a.
nP
b.
aP
a
√P
n
√MENT √SEG
[u X]
251
√P
√P
√AL
[u √]
√MENT √SEG
[u X]
Where and why does the derivation of *belittlemental crash? First, √MENT merges
with vP belittle (47a). While this is in accordance with its underspecified uninterpretable feature which allows it to select both a root or an xP, it gives rise to the sort
of tension discussed in connection with (33) whereby a √ heads an xP in violation of
the canonical ordering of projections assumed earlier. Further merger with nP (47b)
offers the required escape hatch, allowing √MENT to move up to the head of nP,
thereby ceasing to head an xP.
(47)
a.
√P
√MENT
[u X]
b.
vP
XP
X
nP
n
belittle
√P
√MENT
spellout
n
t
vP
belittle
Two consequences follow. First, having moved up to the head of nP (47b),
√MENT has now been removed from the scope of spell-out of the nP phase.
It will be spelled out separately. Second, √MENT now being firmly tucked into the
head of nP can no longer be selected as such: further selection can only target the nP
which now hosts √MENT. This will be the exclusive privilege of an xP selector, for
instance √LESS (48b), hence belittlementless. By contrast, on account of the uninterpretable feature they carry, √AL or other √ selectors such as √IC or √OUS, will
never even be considered for merger with nP belittlement (48a).
252
(48)
Jean Lowenstamm
a.
b.
√AL
[u √]
n
√MENT
nP
√P
n t
√LESS
[u xP]
nP
belittlement
vP
belittle
What recommends this analysis of the ungrammaticality of *belittlemental is the
way in which it captures the fact that the verbhood of belittle impacts the behavior of
a hierarchically non-adjacent object such as √AL: the fact that belittle is a vP, while it
does not preclude its selection by √MENT, ultimately forces √MENT into a position
such that it can no longer be selected by √AL, or any other Level 1 or cyclic suffix. At
no point in such a scenario do < al> and < belittle> need to be aware of each other. As
a result, selection can be kept both strictly local and strictly unidirectional (a head
selects its complement, not vice versa).
Next, the question arises of the grammaticality of governmental, and whether it is
really unexpected. The problem is usually posed as in Aronoff (1976), viz. -al attaches
to Xment, but not if X is a verb. This is illustrated in (49a,b), along with the puzzling
exceptions in (49c).
(49)
a. Xment
ornament
regiment
segment
b. Xment
employment
discernment
agreement
basement
shipment
c. Xment
government
development
judgement
X is not a verb
orn
reg(i)
seg
X is a verb
employ
discern
agree
base
ship
X is a verb
govern
develop
judge
Xment-al is viable
ornamental
regimental
segmental
Xment-al is not viable
∗employmental
∗discernmental
∗agreemental
∗basemental
∗shipmental
Why is Xment-al viable?
governmental
developmental
judgemental
Derivational affixes as roots
253
What does the system advocated here have to say about the evidence in (48)? Let us
examine it block by block.
First, (49a). ornamental, regimental, and segmental are unproblematic: < orn>,
< reg(i)>, and < seg> are not xP’s, or a corresponding noun, verb, or adjective would
exist. Therefore, < orn>, < reg(i)>, and < seg> are roots. Accordingly, the analysis of
ornamental, regimental, and segmental is straightforward, i.e., a complex root eventually merged with a category assigning head. The analysis of segmental appears
formulaically in (50), though the reader can go back to (46b) for the corresponding
demonstration.
(50)
[a [√P √AL
[u √] [√P
√MENT
[u X]
[√SEG]]]]
Apparently more problematic, is the case of the “ungrammatical” adjectives in
(49b), such as *employmental, *discernmental, etc. Here, my proposal can assess
them in clear fashion, although—as we will soon see—the assessment does not lead
up to the outright elimination of employmental (a good thing too, as we will see).
The assessment is: in order for employmental to be ungrammatical, < employ> must
have been a vP. If < employ> in employmental is a vP, the derivation will crash
exactly for the same reasons as that of *belittlemental (48a): the merger of √AL with
an nP makes it impossible for √AL to check its uninterpretable feature.
(51) [a [ √P √AL [nP √MENT n [√P t√MENT [vP v √EMPLOY]]]]
[u √]
[u X]
Of course, the reason employmental is not ruled out altogether is that it’s not the
case that < employ> could EXCLUSIVELY have been a vP. Nothing indeed rules out
the possibility of < employ> being √EMPLOY, as Acquaviva (2009) points out. In
this case, successive mergers could give rise to the legitimate object in (52), and
employmental should be entirely well-formed, i.e., with the same architecture as
segmental. We return to this momentarily.
(52)
aP
a
√P
√AL
[u √]
√P
√MENT √EMPLOY
[u X]
The “three exceptions” of the third block, judgmental, developmental, and governmental, are unambiguously assessed as well-formed adjectives, with an analysis
254
Jean Lowenstamm
which is exactly that of segmental (46b) or, for that matter, employmental (52). Hence
the conundrum in (53).
(53)
i. There is a source for items reputed to be ungrammatical such as employmental; how can they be ungrammatical?!
ii. The same source yields three perfectly grammatical ‘exceptions’; why aren’t
they vastly more numerous?
I submit that the answer to both questions can only come from a thorough
reconsideration of the data. Indeed, the data in (49) is typically the corpus of a
word-based theory. For a word-based theory, it is important whether employmental
or discernmental exist or not. But, the question I am asking is not at all rooted in a
word-based theory. For clarity, that question is reformulated in (54).
(54)
Does √AL merge with a complex root headed by √MENT?
The existence of adjectives such as employmental or discernmental would directly
answer the question, but their absence does not. For, in a theory such as DM, there
are two possible sources for the non-existence of a “word”. One is its non-generation
by the system. The other, is its idiosyncratic absence from the Encyclopedia. This
can be illustrated by means of the two examples in (55).
(55) a. blueberry
b. cranberry
Three roots are involved in the sample in (55): √BLUE, √CRAN and √BERRY. All
three are perfectly good roots, but only [a a √BLUE] and [n n √BERRY] are recorded
in the Encyclopedia. If we extend this view to complex roots of the type proposed
here, we can make sense of accidental gaps such as in (56b).
(56)
a. atom
b. motor
atomic
*motric
atomicity
motricity
*motric is not attested as an adjective, but its absence need not be interpreted as
meaning that √IC fails to select √MOTOR. On the contrary, the existence of motricity
shows just that: √IC does select √MOTOR. But, [a/n a/n [√P [√ √IC [√ √MOTOR]]]
itself, much as [a/n √CRAN], is simply not recorded in the Encyclopedia.
It is in just this sense that the absence of employmental, discernmental, and the
like falls short of being the negative answer to (54). What is required is positive
evidence, of the same kind as was discussed in the case of motricity. In fact,
attestations of sequences of type < √X+√MENT+√AL>, where there exists a verb
[v √X], can be found fairly easily outside of dictionaries. Adverbs are a fruitful
ground, (57).
Derivational affixes as roots
255
(57) [ADV _ ly]
Colloquialisms such as (58) are a rich source of information, in this respect.
(58) a. . . . ly speaking (informally/loosely/professionally speaking)
b. . . . ly challenged (vertically/romantically/rhythmically challenged)
Consider the data in (59–66) gathered from the Internet.15
(59)
Stupid jerk who continually forgets to include a specified attachment within
an email. Don’t be alarmed if Bob had to send you that spread sheet a few
times, he’s a little attachmentally challenged.
(60)
On behalf of accomplishmentally-challenged Americans everywhere . . .
(61)
But it isn’t just celebrity couples who are accoutrementally mismatched.
(62)
GLITTER CAMP will be a place for the adornmentally-challenged to get
gussied up with glitter.
. . . I mean imagine, if there ever came a time to . . . change any contents, like
amendmentally speaking, it would like totally . . . suck to have to . . .
(63)
(64)
Although the terms and reasons for the imposition of the “DefermentalProbationary Contract” are unclear, Dyson’s argument appears to challenge
the legality of his present confinement.
(65)
Here are some links to get your ass back in the saddle employmentally
speaking.
(66)
What should we do if Obama is elected and does not support Israel . . .
deploymentally, financially, or otherwise?
11.5 Concluding remarks
In this concluding section, I return to the consequences of my proposal for the place
of cyclic domains in the system and their relationship to mechanisms of Phasal
Spell-out.
Roots are arranged into two blocks as proposed at the end of subsection 11.2.3.,
and repeated in (67) for convenience.
15
In view of a remark by one of the referees to the effect that some of the examples in (59–66) have a
marginal ring to them, I urge the reader to examine them one by one with an open mind. I submit that the
odd quality pointed out by the reader stems largely from the humor (intended or unintended) inherent in
the politically correct examples, not from their grammatical status. Indeed, where humor plays no role,
e.g., (63, 65, 66) it actually takes the training of a linguist and the specific concerns of a morphologist to
notice anything.
256
Jean Lowenstamm
(67)
xP Phase 3
x
√
√P
x
t√
[u xP]
xP Phase 2
√P
x
√
x
t√
[u xP]
xP Phase 1
x
√P
√
[u √]
√P
…√…
The first block at the bottom, call it the radical core, is the complement of Phase 1. It
consists exclusively of roots which will bunch up in the way described earlier.
The second block is the rest. Because of the prohibition against a root heading a
category and the ensuing left-adjunction of √ to the next higher categorial head
exemplified in (67), it consists of alternating radical and categorial layers.
This makes it possible to revisit cyclicity in fundamental fashion. Cyclicity can
now be redefined as in (68).
(68)
i. Roots are the domains of application of phonological rules.
ii. Rules apply on the most deeply embedded root, then reapply on the
domain defined by the next adjacent higher root, and so forth.
It follows from the organization in (67) that cyclic phonology will be limited to Phase
1. Indeed, adjacent roots can only be found at Phase 1, as any root located above
Phase 1 will end up squished between two categorial layers. Consequently, no root
needs to be viewed as cyclic (or non-cyclic) as such. Indeed, there is no sense in
which √ATOM, √IC, √NESS, √MENT, or any of the roots that have been quoted so
far, differ from each other in their intrinsic ability to trigger the application of cyclic
rules. Rather, the cyclicity of an affix can be derived from the position in which its
Derivational affixes as roots
257
uninterpretable feature will cause it to be located, either in the radical core, above in
the layered zone, or in either place.16
(69)
W [√ . . . X . . . [√ . . . Y . . . ]] Z
Stipulating which affixes are cyclic or non-cyclic can be dispensed with altogether.
Their behavior in this respect depends on their structural position, and their
structural position directly reflects their selectional behavior. Of course, this is
especially striking in the case of universal selectors which can appear anywhere.
For a synoptic view, the entire gamut of selectional patterns for English is summed
up in (70) along with corresponding positional examples in (71).17
(70)
(71)
Feature
Relationship to a root
[u√]
[uxP]
[uX]
always local
never local
a) possibly local
b) possibly non-local
a.
Involvement in cyclic
phonology
always
never
yes, in such case
no, in such case
b.
√IC
[u √]
√
a
√MENT
[uX]
√
(71a)
(71b)
(71c)
(71d)
d.
c.
√NESS a P
[uxP]
Example
√
√MENT vP
[uX]
v
Appendix
In a much-quoted article, Fabb (1988) puts forth a number of generalizations about
English affixation. Most of Fabb’s generalizations are incompatible with the proposals contained in this chapter, and they obviously deserve to be addressed more
extensively than is possible in the context of a brief appendix. One example only will
be discussed here in an attempt to sort out the issues involved.
16
Thanks to Victor Manfredi for pointing out to me the relevance of Giegerich’s important work
(Giegerich, 1999).
17
Affixes are reputed to be a closed class. While the class is not totally closed judging from the
numerous borrowings of derivational affixes from Romance and Slavic by languages such as English and
Yiddish respectively, it is true that borrowing and creation remain limited. In the context of the proposal
made in this chapter, a rationalization for the difference between open and closed classes is available: free
roots can be borrowed instantly, but bound roots are significantly more complex objects inasmuch as a
decision is required as to the associated uninterpretable feature they will necessarily carry, be it readily
borrowed from the source language or assigned by the borrowing language.
√
258
Jean Lowenstamm
Fabb claims that < ism> never attaches to an already suffixed base. The existence
of radicalism constitutes no ground for a comparison between his and my account,
for Fabb’s account is couched within a word-based theory and < radic> is certainly
not a word. Nevertheless, Fabb’s contention that < ism> will not attach to a suffixed
base is an impossible generalization in the context of the proposals articulated here.
Indeed, < ism>, strictly speaking √ISM, is an xP selector (or a non-cyclic or Level 2
affix). As such, it can only be oblivious to the internal composition and possible
complexity of the √P that lies below the little x it will merge with. If this is correct,
Fabb’s generalization is either accidentally true, or false. As it turns out, it is false as
evidenced by the existence of nouns such as gangsterism, tricksterism, pornsterism,
and other formations along the same pattern, or structuralism, wharfianism, etc.
12
Building scalar changes
MALKA RAPPAPORT HOVAV
12.1 Scales and lexicalization
It has been known from the earliest studies of lexical aspect, that many verbs
participate in event descriptions with varying telicity.1 The earliest studies (Verkuyl
1972, Dowty 1979, Mittwoch 1982, Hinrichs 1985) focused on verbs which produce
event descriptions whose telicity depends on properties of the direct object, as in:
(1)
a. We ate popcorn for ten minutes/*in ten minutes.
b. We ate a bag of popcorn ?for ten minutes/in ten minutes.
The contrast between (1a) and (1b) shows that the sentence in (a) is atelic, as it is not
compatible with ‘in ten minutes’, but compatible with ‘for ten minutes’, while the
sentence in (b) is telic, being compatible with ‘in ten minutes’ and only marginally
compatible with ‘for ten minutes’. The telicity of the predication is determined by
the nature of the direct object as specifying a (non)-delimited quantity of its referent.
Dowty (1991) termed verbs which show variable telicity in this way “incremental
theme verbs” and characterized them as establishing a homomorphism from an
argument to the event they denote (cf. Krifka’s 1989 “Mapping-to-Events”). Tenny
(1994) describes the direct object as being an argument which “measures out” the
event: progress in the development of the described event can be monitored through
the extent of the referent of the direct object. Subsequent research (e.g., Tenny 1994,
Ramchand 1997, Krifka 1998) has extended the notion of “measuring out” to cases
where something other than the part structure of the direct object functions as the
measure. Hay, Kennedy and Levin (1999) and Kennedy and Levin (2008) (see also
Beavers 2008) proposed a unified approach to the determination of telicity for three
previously recognized types of verbs with variable telicity. For each verb type, a
1
I am indebted to Beth Levin, Edit Doron, Anita Mittwoch, and two anonymous readers for comments
on the draft of this chapter, and to them, Chris Kennedy, and Olga Kagan for useful discussion of the
issues discussed in this chapter. This work is supported by Israel Science Foundation grant 370/07.
260
Malka Rappaport Hovav
multi-valued property of the referent of an argument serves as the measure—though
the particular property and the particular argument depend on the verb type—and it
is a change in the value of that property which measures out the event. In the case of
incremental theme verbs, the spatial extent of the referent of the direct object is
relevant, in that the progress of the event is monitored by a change in its spatial
extent. With change of state (COS) verbs, as in (2), it is a gradable property (often
named by a base adjective) associated with the referent of the direct object, and with
direct motion (DM) verbs, as in (3), it is the location of the theme argument along a
path.
(2)
a. The soup cooled for ten minutes.
b. The soup cooled to 30 degrees in ten minutes/*for ten minutes.
(3)
a. The plane descended for ten minutes (before gaining altitude).
b. The plane descended 100 meters in three seconds/*for three seconds.
Since the values of the relevant properties associated with these verbs can be
represented as a scale, a number of studies (e.g., Tenny 1994, Ramchand 1997, Hay,
Kennedy and Levin 1999, Kennedy and McNally 2005, Beavers 2008, Kennedy and
Levin 2008) have suggested that the intuitive notion of “measuring out” can be
formalized using the notion of a scale. Thus, incremental theme verbs are said to be
associated with volume/extent scales, COS verbs are said to be associated with
property scales, and directed motion verbs are said to be associated with paths,
taken to be a kind of a scale. Unifying the three classes with the notion of a scale
allows a generalization regarding the determination of telicity. A somewhat simplified version of this generalization is: a predication asserting a specified change in the
value of the scalar property is telic (as in 1b, 2b and 3b), while a predication asserting
an unspecified change in the value of the scalar property (as in 1a, 2a, and 3a) is
atelic.
While the studies just mentioned showed the relevance of an associated scale to
the determination of telicity, in Rappaport Hovav (2008) I suggested that once we
distinguish between scales which are lexicalized in the meaning of a verb from
those which are not, the notion of a scale can be shown to be relevant to argument
realization, as well as telicity. Although the three classes of verbs behave in a similar way
with respect to the determination of telicity, they differ in their argument realization
properties. In particular, COS verbs and DM verbs share certain basic constraints on the
realization of their arguments, and strikingly these constraints are not shared by
incremental theme verbs. I suggested that this difference between the classes of verbs
can be attributed to the fact that COS and DM verbs lexicalize all or most of the
components of the scale associated with them, whereas incremental theme verbs do not
lexicalize any components of a scale; the scale is introduced by the direct object. In
Rappaport Hovav and Levin (2010), we suggested that the distinction commonly
drawn in lexical semantics between manner and result verbs can be reduced to a
Building scalar changes
261
distinction between verbs lexicalizing a scalar change and verbs which lexicalize a
non-scalar change.
In this chapter, I will support these assumptions by illustrating how change of
state verbs (as in 2) and directed motion verbs (as in 3) are built semantically in a
parallel way. In particular, I will show that for any verb in both classes it is possible
to isolate a stative scalar attribute which serves as the semantic core of the verb. The
events in the denotation of the verb all involve a change in the value of that scalar
attribute. In the case of COS verbs, the focus of section 12.2, the stative scalar
attribute is often named by an adjectival base, though there are many COS verbs
for which this property is actually derived from the base verb. In the case of directed
motion verbs, the focus of section 12.3, the stative scalar attribute is location on a
path, where this stative attribute is generally lexicalized by a preposition. Since there
is no rule converting prepositions into verbs in English, the relation between DM
verbs and the stative attribute from which they are derived is not morphologically
evident. Nonetheless, DM verbs involve a change in the value of the scalar attributes
encoded by prepositions. The parallel between the two classes of verbs is strengthened by the fact that there is a similar typology of scales in both classes and that the
type of scale lexicalized by a verb in either class determines the telicity potential of
the verb. Despite the similarity between the two classes of verbs thus unified, I show
that in English, the components of the different kinds of scales are lexicalized
somewhat differently in the two classes. In particular, the components of property
scales are fully lexicalized in verbs, while for many DM verbs, not all the components
of the scale are lexicalized. This difference has syntactic effects: many DM verbs, but
not COS verbs, require the presence (or recoverability from context) of a constituent
which provides a crucial component of the path scale.
The more general conclusion which emerges from this investigation is that a
scalar analysis of event descriptions and of the lexical semantics of verbs is appropriate. In section 12.4, I show that beyond the striking parallelisms in the structure of
property and path scales, verbs which lexicalize property and path scales share many
interpretive and argument realization properties.
In section 12.5, I show that the changes lexicalized in incremental theme verbs
cannot be given the same analysis. There is no stative scalar attribute that can be
shown to serve as the semantic base of these verbs, nor is there a typology of scale
types associated with this class of verbs. Concomitantly, this class of verbs shows a
range of semantic and syntactic behavior which distinguishes it from the other two
classes under investigation.
Before proceeding with the analysis, a few words are in order concerning
the distinction between semantic components that are lexicalized in the verb and
those that are not lexicalized but can be inferred from the properties of the event
description composed of the verb and its arguments. I will assume that lexicalized
meaning components are those that are specified and entailed in all uses of the verb,
262
Malka Rappaport Hovav
regardless of context (Rappaport Hovav and Levin 2010); this can be seen as an
elaboration of the lexical entailments of Dowty (1989, 1991). A verb which lexicalizes
a scale involves some change in a lexically specified scalar property which is entailed
in all uses of the verb.2 To see the relevance of the distinction between a lexicalized
and a non-lexicalized scale, consider the following three sentences:
(4)
a. I just wiped the counter clean/dry.
b. I just wiped the counter, but it’s as wet/dirty as it was before.
c. I dried the counter (*but it’s no drier than it was before).
In (4a) the APs clean and dry specify a scalar attribute, but this scalar attribute is not
lexically encoded in the verb. The verb wipe describes an action which is typically
associated with an intended result—a change in a scalar property of the surface
wiped (e.g., dryness or cleanliness); however, the use of the verb does not entail this
result, as shown by (4b). In contrast, the use of the verb dry entails some change in
the value of the property of dryness and thus lexicalizes the scalar attribute of
dryness. Similarly, while a verb like run is typically used to describe an event
which involves the traversal of a path—and this is probably part of our knowledge
of the meaning of this verb—it is possible to run in place (5a). Thus, the traversal of a
path is not entailed in all uses of run and is not lexicalized by the verb. It goes
without saying, then, that the verb run does not specify anything about the nature of
the path traversed, and is compatible with traversal of any kind of path (5b). In
contrast, the verb ascend is always used to denote an event which involves change
along a specified path, a rising path (6), though the particular manner can vary.
(5)
a. John ran in place.
b. John ran up the hill/down the hill/into the cave/out of the cave.
(6)
a. The plane ascended from an altitude of two miles to an altitude of three
miles/*from an altitude of three miles to an altitude of two miles.
b. The backpackers ascended the mountain. (traversal must be up the mountain)
This discussion of course presupposes that there is some invariant element of
meaning which can be used to identify a particular sense of a verb. This is not a
simple assumption, especially given the commonly held belief that many of our
linguistic concepts are family resemblance or prototype in nature. Nonetheless, I will
continue to make this assumption, since here, and in other studies, it helps deepen
our understanding of the nature of lexical organization.3
2
In RH&L (2010) and L&RH (2013) we discuss cases in which there is no constant component of
meaning in the context of manner/result complementarity. These cases are circumscribed and have
specific properties which we explore.
3
Note that I do not distinguish here between the semantics of the verb and the semantics of the root,
since I do not present a theory of how verb meanings are derived from the meanings of roots. However, it
should be clear that I intend lexicalized meaning to be a property of roots, determined in the lexicon, and
Building scalar changes
263
12.2 Scalar changes in COS verbs
In this section, I will show how COS verbs are built on the basis of stative scalar
attributes. There is a fair amount of literature which treats the relation of deadjectival verbs to their adjectival base (e.g., Hay, Kennedy and Levin 1999, Kearns
2007, Winter 2006, Kennedy and Levin 2008), and the analysis in this section is
based on this literature. I do not provide a full analysis of these verbs, but only
mention those aspects which will help bring out the parallels to be drawn between
COS verbs and DM verbs, an exercise which, as far as I know, has never been done in
a systematic way. I will review the facts which show that there is a typology of scales
associated with COS verbs corresponding to the typology of property scales associated with the adjectival predicates. However, it is important to point out that the
same typology emerges from COS verbs that are not derived from adjectives. For
these verbs the stative scalar attribute is a deverbal adjective.
COS verbs are often derived from adjectives such as warm, cool, long, short, deep,
fast, old, large etc., associated with a range of values, which is the source of their
gradability. The use of such an adjective asserts that an entity manifests the gradable
property to the degree represented by a particular value or range of values. The set of
degrees on a particular dimension (e.g., height, temperature, cost, depth) with an
ordering relation between them constitutes the scale associated with such an adjective. Adjectives can differ with respect to the dimension they lexicalize (e.g., cool
lexicalizes the dimension of temperature, whereas expensive lexicalizes the dimension of cost) or they can differ only with respect to the lexicalized ordering relation
of the degrees. For example, cool and warm are both used to refer to a range of values
on the temperature scale, but the degree of coolness increases as the temperature
decreases, whereas the degree of warmness increases along with the temperature.
The value of the property attributed to an entity may be explicitly represented as
in John is three feet tall; it may be asserted to be greater or lesser than the degree to
which another entity manifests the property, as in Martha is taller than James.
When there is no explicit measure phrase or comparison, the use of such an adjective
invokes a contextually determined standard of comparison, which represents a
particular value on the scale, and there is an assertion that the entity manifests the
property either to a greater or lesser degree than that represented by the standard of
comparison, depending on the adjective. For example, saying that someone has long
eyelashes is to say that the value of the attribute on the dimension of length is greater
than the standard for eyelashes. Long eyelashes are a lot shorter than long hair on
not structurally. This approach assumes an ontological classification of roots which is grammatically
relevant, a position assumed as well by Alexiadou, Anagnostopoulou, and Samioti (this volume), Doron
(this volume), Levinson (this volume) and Roßdeutscher (this volume). The assumption that roots have an
ontological classification which is relevant to the determination of their grammatical properties is
explicitly denied in Borer (2003, 2005) and Acedo-Matellán and Mateu (this volume).
264
Malka Rappaport Hovav
one’s head; hair on one’s head has a different standard, and long bridges are of course
a lot longer, according to their standard length. Intuitively speaking, because the
scale associated with a predicate like long has no maximal value, the standard is
necessary for specifying the range of the scale the particular use of the predicate
covers. The relation of the predicate to the standard of comparison is determined by
the polarity of the scale: the degree of length increases with an increase in spatial
extent and something is long if it is above the contextually determined standard
on this dimension; the degree of shortness increases with a decrease in the
same dimension and something is short if it is below the contextually determined
standard.
Gradable properties can appear with degree modifiers and in the comparative and
superlative:
(7)
longer, longest, quite long, very long, 3 inches long, 3 hours long . . .
Studies of property scales have distinguished a number of different subtypes (Cruse
1980, Hay, Kennedy, and Levin 1999; Winter and Rotstein 2004, Kennedy and
McNally 2005, Kearns 2007, Kennedy and Levin 2008). The adjectives we have
discussed so far are associated with open scales, i.e., scales with no maximal degree.
In contrast, adjectives such as flat, empty, clear, dry and straight are associated with
upper closed scales, i.e., scales with a maximal degree. These adjectives - but not
those associated with open scales - can be modified by phrases such as completely,
entirely, halfway, fully, etc.
(8)
completely full, halfway empty, entirely dry, fully closed
Finally, some adjectives are associated with a scale with a lower bound which
represents a non-zero degree of the measured property. Adjectives of this class
include wet, impure, dirty, and dangerous. Attributes with a lower bound property
scale hold of an entity when the smallest value is instantiated. For example, something becomes dirty as soon as there is some dirt on it.
COS verbs are predicates which express a scalar change in a stative scalar attribute
from the property domain. A scalar change involves a change in the value of a scalar
attribute in a particular direction, where the direction of the change is determined by
the polarity of the scalar base. Since with the adjective cool, the degree of coolness
increases as the temperature decreases, a cooling event is one in which there is a
decrease in temperature. With warm, the degree of warmness increases along with
the temperature, and concomitantly, a warming event is one in which there is an
increase in the temperature.
If we wish to preserve the notion that the use of a scalar predicate always involves
some comparison, we can say that the use of a verb denoting a scalar change involves
the comparison of the degree to which a scalar property holds of an entity at the
beginning of the event and the degree to which the same property holds of the entity
Building scalar changes
265
at the end of the event (Kennedy and Levin 2008). Since the standard of comparison
is inherent in the meaning of the verb—the value of the attribute at the beginning of
the event—it doesn’t have to be contextually specified. Therefore, a sentence such as
(9) can be appropriate even if a person’s hair got a bit longer but is still considered
short.
(9)
His hair lengthened.
The use of a verb derived from a property associated with an upper-closed scale,
typically implies a change to the maximal value of the attribute. That is, (10a) is
normally understood to mean that we completely emptied the tub, though this
inference is defeasible (10b).
(10)
a. We emptied the tub.
b. We emptied the tub, but not completely.
Kennedy and Levin (2008) suggest that, as a pragmatic principle, lexically specified
standards of comparison are to be preferred over contextually supplied ones, unless
explicitly denied.
We have looked at scalar change verbs derived from gradable adjectives associated
with multi-valued attributes. These adjectives enter into contrary oppositions (e.g.,
Cruse 1980, Horn 1989). There are predicates naming attributes which enter into
contradictory oppositions. Examples include odd, even, female, dead, cracked, etc.
They are associated with a property that does not hold of entities to varying degrees.
Some of these are related to verbs:
(11)
crack, die, faint, explode . . .
Following Beavers (2008), I will consider these verbs to lexicalize a two-point scale
(see also Rappaport Hovav 2008). A scalar change in this case is a transition from an
entity possessing a specified property to its not possessing this property, or a
transition from not possessing to possessing a specified property. For example, the
verb die lexicalizes the transition from not being dead to being dead. Therefore,
unlike the scalar verbs we have been considering, they give rise to punctual events,
involving the near instantaneous transition between the state of being not associated
with a property and being associated with that property. Since the vast majority of
adjectives are gradable, there are not many de-adjectival verbs which give rise to
COS verbs with two-point scales.
With the generalization of the notion of a scale to include two-point scales, we
derive a notion of scalar change that is not equivalent to gradable change. There are a
number of reasons for placing punctual and gradual changes in the same linguistically relevant class (see Beavers 2008). Below, in section 12.5, I will provide additional
support for this move.
While many verbs of scalar change are derived from simple adjectives either with
or without morphological marking (e.g., dry vs. sweeten), there are verbs of scalar
change that are morphologically simple verbs, i.e., verbs not derived from adjectives.
266
Malka Rappaport Hovav
In these cases, it is a deverbal adjective which expresses the stative scalar attribute
and the non-derived verbs which express the scalar change. These verbs can also
be distinguished according to the type of the scalar properties they lexicalize. More
specifically, they may name a change along a two-point scale (e.g., crack), a change
along a multipoint scale with a maximal degree of change (e.g., thaw, freeze—
completely/halfway frozen) or without such a maximal degree (e.g., stretch,
shrivel—?completely/halfway shriveled). There are, however, no such predicates
that are associated with a scale that is open on the lower end (that is, all such
predicates are associated with a lower-closed scale). This is because the onset of the
event always represents the beginning of the manifestation of the scalar attribute and
thus constitutes the lower bound of the scale (cf. Kennedy and McNally 2005:365).
Scales associated with verbs do not have to be fully specified. For example, there is
a small set of verbs which indicate movement along a scale without specifying any
dimension, so that they can easily be used to describe changes of many sorts. In fact,
for verbs of this sort, the argument is often not the entity possessing a scalar
attribute, but rather the attribute itself which takes the entity it is an attribute of as
an argument. It is thus the argument of the verb which helps flesh out the content of
the scalar change.
(12)
a.
b.
c.
d.
The
The
The
The
cost (of gold) increased. (cf. Gold increased in value.)
size (of his waistline) increased with age.
speed (of the plane) decreased.
popularity (of the singer) decreased.
There are also verbs like grow, develop and evolve, the last two pointed out to me by
Jens Fleischhauer, which are also not specific about the dimension, though, interestingly, verbs like increase and decrease tend to apply to a different range of entities
than develop and evolve. It seems that growing, evolving and developing involve
changes which take place in the natural course of events, while increasing and
decreasing involve dimensions which are more readily quantified. Moreover, develop
and evolve seem to be predicated mainly of the entity undergoing the change and not
of the scalar attribute, where the exact attribute is contextually determined.
Table 12.1. illustrates the classes of verb discussed so far:
Table 12.1.
Scale type
Two-valued
Multipoint
Upper Bound
Multipoint
Lower Bound
Multipoint
Open
Adjectival Property
Base
De-adjectival Verb
Deverbal Property
Base
Verb
dead
flat
wet
long
die
cracked
flatten
thawed
lengthen
stretched
crack
thaw
wet
all deverbal
scalar properties
are lower
bounded
stretch
Building scalar changes
267
12.3 Scalar changes in directed motion verbs
The second class of scalar verbs is directed motion verbs (DM). As in the case of the
COS domain, we begin with a scalar attribute that does not involve change. While
there are many dimensions which serve as the basis for COS verbs, the dimension
which serves as the basis for all DM verbs is that of spatial location. Being located
relative to a reference object (RO) is the relevant scalar attribute in all cases. When
the theme and the RO are displaced in space, the set of contiguous points of location
between them form a path. The path can be considered a multipoint spatial scale
consisting of this set of points (Dowty 1991, Jackendoff 1983, Krifka 1998, Zwarts
2005) and being located at one of these points on the path is the relevant gradable
property. As we will see, while the property domain is inherently structured into
dimensions correlating with various gradable properties, the spatial domain is not so
structured and the scale (path) must be built with the help of the RO. Since all spatial
prepositions involve the same dimension (spatial location), the prepositions can vary
in terms of the ordering of the points on the path or in terms of the nature of the
scale as two- or multipoint, and when multipoint as bounded or not on either end.
For presentational purposes, it will be simpler to begin with two-valued predicates.
As mentioned, languages express the location of an entity (here referred to as the
theme) relative to a reference object (Jackendoff 1983, Levinson 2001, Talmy 1985,
among others). Predicates expressing location are typically prepositions which
specify something about the relation between the theme and the RO.
(13)
Where are my keys? [They]theme are [next to]relation of theme to RO [my wallet]RO
The most basic type of preposition, that which is instantiated in all languages with
prepositions representing spatial relations, represents some kind of static relation of
near contiguity between the theme and RO, or expressing further a relation of
superadjacency or containment, the latter often characterized as topological (Levinson 2001).
(14)
at, next to, on, in, . . .
The attribute of being located somewhere is built compositionally from the preposition and its object. The attributes built from these prepositions are not gradable and
cannot appear with degree expressions. They involve a relation of contiguity between
a point and a region and so there are only two values to the attribute: that of being or
that of not being at the location.4
(15)
*very at, *very in, *further at, *further on, *two meters at, two meters on . . .
4
In sometimes does appear with measure phrases since it appears with ROs that can be analyzed as
having an inner space and it is possible to measure the distance from the periphery of that space to the
position of the theme, as in: the pole is buried two meters in the ground.
268
Malka Rappaport Hovav
We turn, now, to multi-valued predicates of spatial location. These are predicates
which encode a relation between a theme and a RO that are displaced from each
other in space. The most basic type is a predicate which merely encodes distance in
space between the theme and the RO and nothing else.
(16)
a. far from, near, close (to), apart, away from
b. The school is far from/close to home.
As with gradable adjectives the use of one of these predicates asserts that an entity
manifests the gradable property (location along a path) to the degree represented by
a particular value. The value can be explicitly specified by a measure phrase as in:
(17)
3 feet away from the house; 5 feet apart
An indication of the semantic similarity between scalar property predicates and
scalar location predicates can be found in shared restrictions on measure modification with the two kinds of predicates (Winter 2005).
(18) a. 3 feet tall vs. *3 feet short
b. 10 meters away from the house vs. *10 meters close to the house
In parallel to scalar adjectives, the value of the prepositional scalar attribute can be
compared to that of another entity. However, since these predicates are prepositions,
most (except for near and close, which may be adjectives) cannot appear in the bare
comparative form, but must be compared with the help of further.
(19)
a. nearer, closer, further, further away from
b. My house is further away from the school than yours.
As in the case of gradable adjectives, in absence of an explicit comparison, the use of
this kind of gradable spatial attribute invokes a standard of comparison. The point
on the path which serves as the standard of comparison is contextually determined,
just as in the case of a property scale. For example, the use of the predicate near
involves the assertion that the theme is at a point on a path which is closer to the RO
relative to some standard, where the standard is contextually determined. The
standard of nearness for two cities is different from the standard for the sun and a
planet.
(20)
a. Boulder is near Denver.
b. Mercury is near the sun.
The points on a scale must be ordered, and indeed the points of location on the
path are inherently ordered, depending on the predicate. For predicates like far,
apart, and away from the value increases with increased distance from the RO. For
near and close the value increases with decreased distance from the RO.
Building scalar changes
269
As mentioned, scales can be distinguished in terms of their structure. Predicates
like near and far specify paths that are relatively underspecified: any path that can be
drawn with the RO as the lower bound of the path is in the denotation of these
predicates. Since the RO constitutes the lower bound, and the theme must be
allowed to vary in its position along the path, all the paths built with these prepositions are upper-unbounded: there is no maximal degree to distance.
Predicates of distance are the simplest of the path predicates of location; there are
predicates which further restrict the nature of the relation between the theme and
the RO. These are predicates of angular location (Levinson 2001) which provide a
more specific direction to the path. The paths that fall within their denotation are
defined through the specification of a part of the RO and the definition of a space
relative to that part of the RO. These predicates may be used felicitously to place a
theme somewhere along one of the paths so defined. As in the case of predicates of
distance, the RO serves as the lower bound of the path.
(21) a. above, away from, below, behind, in front of, under, . . .
b. The cat is behind the TV.
These predicates are also gradable: they may appear in comparatives and take degree
modifiers. These modifiers specify something about the distance between the theme
and the RO along one of these paths.5
(22)
two meters above/below/behind, far/further above, far/further below . . .
With angular location predicates the ordering is such that points that are further
from the RO have greater values; there are no predicates which order the points on
the path such that points closer to the RO have greater values.
(23)
a. The TV is three feet behind the bookcase.
b. The bookcase is three feet in front of the TV.
Though both sentences in (23) describe the same situation, the different prepositions
dictate a different ordering of the points between the TV and the bookcase, since the
preposition in (a) dictates that the location of the bookcase represents the lower
bound of the path while the preposition in (b) dictates that the location of the TV
represents the lower bound of the path.
In the case of scales in the property domain, the structure of the scale—for
example, whether or not it has a minimal or maximal value—is determined by the
lexicalized property. Tallness, by its very nature has no maximal value, while dryness
does. But the spatial domain is not inherently structured. The RO is needed to build
the path which is the foundation of the scalar property; the RO serves as the lower
bound in all cases. Most stative path predicates are thus upper-unbounded, with
5
Comparatives such as more behind also have another meaning (Johannes 2011): more directly behind.
270
Malka Rappaport Hovav
no maximal value. However, the RO itself may have internal spatial extent, defining
a path that can be bounded on both ends, and can give rise to an upper-bounded
path:
(24)
John is all the way across the bridge.6
Finally, The RO does not have to be represented as a point; along specifies that the
RO is spatially extended, defining a path of its own, and it expresses a relation
between theme and the spatial extent of the RO.
(25)
The flowers grow along the side of the garden.
We now turn to DM verbs which are built from these stative scalar attributes.
Being verbs of scalar change, they express a change in the value of the spatial scalar
attribute of location on a path: change of location on a path—i.e., movement. Since
there is no productive rule which converts Ps to Vs in English, the scalar Ps cannot
be turned directly into verbs, in the way that scalar adjectives can be turned directly
into verbs, an exception being the verb near. There are two ways of lexicalizing
movement along path in English. The first is with manner of motion verbs: here, the
verb encodes the fact of motion and directional PPs encode the path of motion
(Jackendoff 1983, Talmy 1985, 2000, Zwarts 2008, among others). There is a typology
of directed paths which is similar in many respects to the typology of scales (see
Zwarts 2008 for insightful discussion). But since manner of motion verbs themselves
do not encode the path of motion, they are not of direct interest to us here. It is with
directed motion verbs that the path of motion is directly encoded in the verb.
We saw above that the use of a verb denoting a scalar change in the property
domain involves the comparison of the degree to which a scalar property holds of an
entity at the beginning of the event and the degree to which the same property holds
of the entity at the end of the event. In parallel, the use of a DM verb involves the
comparison of the location of the theme along a path at the beginning of the event
with its location at the end of the event. Therefore, the location of the theme at the
beginning of the event constitutes the lower bound of the path for all DM verbs. The
path along which the theme traverses must be built with the help of the RO: I call
this the reference path. All DM verbs specify how the reference path is built with the
help of the RO; they differ, as we will see, in terms of whether the RO is lexicalized in
the verb, whether it is specified by a constituent in the sentence, or recoverable from
context. In addition, each DM verb specifies two things about the movement of the
theme. First, it specifies whether the movement is towards or away from the RO.
This is the spatial correlate of the ordering relation between the points on a property
6
The felicity of this sentence depends on there being the perspective of someone at the other end of the
bridge.
Building scalar changes
271
scale. Second, the verb specifies what portion of the reference path the theme
traverses in the course of the event under description. I call this the path of motion.
We begin our illustration of the semantic components of DM verbs with those
which lexicalize a non-gradable spatial property: that of being located at a point
contiguous with a RO (corresponding to the non-gradable predicates listed in (14)
above). Punctual verbs like arrive, reach, leave, and depart specify a two-point path
built with the help of the RO. The RO is not lexicalized in the verb and must be
specified in the sentence (or else recoverable from context.)
(26)
a. Maxine arrived (at the meeting.)
b. Maxine reached *(the meeting).
c. Maxine left (the meeting)/departed (from the meeting).
(26a) above can be felicitously used without the PP only if the content of the PP is
recoverable from context. Arrive and reach specify that the location of the theme at
the beginning of the event is different from that of the RO and the same as that of the
RO at the end of the event (coming to have the non-gradable property of being at
the location of the RO). Verbs such as leave and depart specify that the location of
the theme at the beginning of the event is identical to that of the RO, and different
from it at the end of the event (coming to have the non-gradable property of being
not at location of the RO). Therefore, these verbs are the spatial correlate of verbs
such as die and crack, based on two-valued scales. Like them, they are both telic and
punctual. Previously, we saw that in addition to predicates like at and next to, there
are predicates such as on, under, and in, expressing superadjacency, subadjacency,
and containment. In English, verbs such as enter and exit express coming to have or
not to have the property of being in the area of the RO.
We now move to verbs that lexicalize a path built from a RO displaced in space
from the theme. Since the RO is not lexicalized in the verb, it must be specified by a
constituent in the sentence, and the reference path will be bounded at the lower end
by the location of the theme at the beginning of the event and at the upper end by
the location of the RO. For verbs like come and go, the RO is lexicalized to be the
“deictic center” (Fillmore 1982, Levinson 2006), the difference between them being
the ordering of the points of location. For go, the scalar change increases with
increased distance from the deictic center and with come, it increases with decreased
distance. The deictic center is either expressed by a constituent in the sentence, or
else recoverable from context. Having an upper-bounded reference path, these verbs
behave somewhat similarly to upper bounded verbs of scalar change in the property
domain: unless specified otherwise, the use of such a verb implies traversal of the
entire path to the RO (27a), but the verb is also compatible with motion in the
direction of the RO, without actually reaching it (27b).
272
(27)
Malka Rappaport Hovav
a. John came. (inference: reached the deictic center).
b. After a silence of several minutes, he came towards her in an agitated
manner . . . (http://www.enotes.com/topics/pride-and-prejudice/etext/
chapter-xxxiv; inference: deictic center not reached)
We find similar behavior with take and bring, the transitive counterparts of come
and go.
For verbs like approach, advance, near, and return,7 the RO is also either specified
by a constituent in the sentence or else recoverable from context, and both specify
that the movement of the theme is toward the RO. However, there is an interesting
difference between them. Though the reference path with approach and advance is
bounded by the RO, the verb specifies that the theme does not traverse the entire
path to the RO; i.e., the path of motion must not include the upper bound of the
reference path. As in the case of COS verbs based on upper-bounded scalar attributes, the telicity of the predication is established not by the boundedness of the
reference path but rather by whether there is a distance specified for the path of
motion.
(28)
a. The army advanced slowly toward the border for two months.
b. The army advanced three miles in six weeks.
The verbs we have discussed so far qualify as verbs of scalar change because the
events that fall in their denotation involve change in a scalar property: location on a
path. However, as we have seen, not all the components of the path are fully
specified in the verbs: the verbs specify that a constituent in the sentence, or a
constituent recoverable from context, will serve as the RO, which is necessary for
constructing the reference path. In contrast, there is a set of verbs which have all the
components of the scale lexicalized in the verb: verbs which express motion along a
vertical axis:
(29)
rise, fall, descend, ascend, drop, lower . . .
For these verbs the direction of motion can be fully determined by the verb, which
means that the RO is lexicalized in the verb, and they need not appear with a
preposition specifying anything about the direction of motion. Perhaps it is best to
consider the RO for these verbs to be the source of gravity, or the earth. Verbs such
as rise, soar, and ascend specify movement away from the RO, whereas verbs such as
fall, descend, drop and lower specify movement toward the RO. The path for verbs
like rise and soar is unbounded at the upper end; they lexicalize a scale similar to that
lexicalized in a preposition such as above, bounded at the lower end by the RO and
7
The verb return also specifies that the theme’s location at the end of the event was also the location of
the theme at some time before the event. It is worth pointing out that the scalar analysis is not an
exhaustive analysis of the meaning of the verbs.
Building scalar changes
273
unbounded at the upper end. In the case of verbs like fall and descend, the path of
motion is usually bounded because a drop or a fall typically ends by contact with
some surface below the theme. However, sentences like those in (30) show that the
path of motion is not lexically bounded.
(30)
The plane descended for two minutes before gaining altitude.
Since there is really only one dimension for path scales (on their non-metaphorical uses), namely, location, there are many fewer verbs lexicalizing path scales, than
those lexicalizing property scales. These verbs are differentiated mainly in terms of
the kind of nature of the scale (path) they encode.
Summarizing, we may say that DM verbs are instances of verbs which encode
scalar change. There are striking similarities in the structure of the scales encoded in
COS verbs and those encoded in DM verbs, which justifies considering them to be
two instantiations of the same kind of change. Table 12.2 illustrates the kinds of
stative and non-stative scalar predicates in the spatial domain sorted by scale-type to
bring out the parallel with the predicates in the property domain. Note that all scales
are lower bounded.
Table 12.2.
Multipoint
Upper Bound
Mulitpoint
Unbounded
at, next to, on, in
none; only if the
path is internal
to the RO
away, above,
below
arrive, reach,
depart, leave
return, approach,
cross,
rise, fall, distance oneself
Scale Type
Two-point
Prepositional
Base
DM Verb
12.4 Nonscalar changes
Tenny (1994) stressed that an important characteristic of a scalar change is its
simplicity: it is a directed change in the values of a single simple attribute. Most
verbs which are typically characterized as manner verbs involve non-scalar changes.
A non-scalar change is any change that cannot be characterized in terms of an
ordered set of degrees along a dimension representing a single attribute. Given the
nature of scalar changes, we can isolate two properties which contribute to making a
change non-scalar: the lack of an ordering relation and the complexity of the change.
I illustrate how each of these properties can contribute to a verb not being a verb of
scalar change.
Consider first verbs like rotate and roll. Both of these verbs describe a motion on
an axis, and roll in addition usually describes an event of movement along a path
274
Malka Rappaport Hovav
propelled by such a motion. The crucial point for making both of these verbs nonscalar is that the change (motion), though simple, is not directed. Movement on the
axis in one direction or the other constitutes rolling or rotating. In the case of roll,
while the verb is often used with a PP expressing traversal of a path, the direction of
the path is not specified by the verb.
(31)
The ball rolled to the door/away from the door.
To illustrate this point, if one presents video clips representing scenes described in
sentences like: The ball rolled along the path, or The image rotated, it is not possible
to tell if the video is running in reverse or not. This is in contrast to directed motion
verbs which determine direction of the motion of the theme along the path built
with the use of a reference object, as in:
(32)
a. The ball entered the room.
b. The ball exited the room.
Other verbs that express a non-scalar change in an entity because the change lacks
an ordering relation include spin, roll, slide, move, revolve, and whirl.
Most non-scalar change verbs involve a complex combination of changes at once,
and this complexity means that there is no single, privileged scale of change which
can ‘measure out the event’. These verbs include:
(33)
jog, waltz, crawl, gallop, leap, laugh, grimace, shiver, wink, flutter . . .
For example, jog involves a specific pattern of movement of the legs; while there may
be an internal sequence of changes specified by the verb, collectively they do not
represent a change along a simple dimension. Walk differs from jog in the specification of the pattern of movement. However, not all non-scalar change verbs are so
specific about the precise changes: e.g., exercise. Here, the verb requires an unspecified and unordered set of movements.
Human activities usually involve many co-occurring changes; these activities,
then, do not qualify as scalar changes. Thus, changes that are typically predicated
of animates are non-scalar in nature, while those predicated of inanimates are very
often scalar. However, non-scalar changes may be predicated of inanimates: e.g.,
flutter, flap, wave, and rumble.
12.5 Differences between scalar and non-scalar verbs
In the last two sections, I illustrated how scalar changes are built, and showed how
COS verbs and DM verbs qualify as verbs of scalar change. The relevance of such a
classification is justified if it can be shown that verbs of scalar change show properties which distinguish them from verbs of non-scalar change. In this section,
Building scalar changes
275
I discuss some such properties: interpretive properties and those having to do with
argument realization possibilities.
The first property has to do with an inherent affinity of scalar change verbs to
telicity. We have seen that when a predication includes a specified degree of change
along a scale it will be telic. Both verbs which I have classified as scalar change verbs
and those I have classified as non-scalar change verbs can participate in such telic
predications.
(34)
a. John cooled the soup to room temperature (in ten minutes).
b. John rolled the ball into the corner (in three seconds).
Nonetheless, it is possible to show that a scalar change verb such as cool has a
stronger affinity to telicity than a non-scalar change verb such as roll.8 Scalar verbs
with multipoint scales—whether they are lexically associated with bounded scales or
not—can give rise to telic predications without an explicit specification of the degree
of change (Rappaport Hovav 2008).
(35)
a. The soup cooled in ten minutes.
b. The plane descended in two minutes.
While activities can also enter into telic predicates without explicit specification of
degree of change (e.g., Dowty 1979), a sentence such as (36) below demands very
heavy contextual support in order to be felicitous. Sentences such as those in (35)
above, do not need any contextual support to be used telicly: any context will do.
(36)
Ben swam in ten minutes. (Normally anomalous; can mean that he finished
his contextually specified swimming in ten minutes).
In fact, scalar change verbs have two telic readings: an achievement reading and an
accomplishment reading (Kearns 2007). With an accomplishment predicate, the in
adverbial expresses the amount of time in which the event takes place and reaches its
conclusion. In contrast, with an achievement predicate the in adverbial expresses the
amount of time that elapses before the entire event occurs: some change in the
degree to which the property holds, conceptualized as an instantaneous change. As
Kearns (2007) points out, non-scalar change verbs do not have this last reading.
(37)
a.
b.
c.
d.
The price increased/decreased in a month. (= (11a) in Kearns)
They chatted in an hour. (= (10b) in Kearns)
#The ball rolled in a second.
The plane descended in an hour.
8
To the extent possible, I have chosen to contrast verbs such as cool and roll, since they are alike in
terms of classifications involving unaccusativity (both are unaccusative in their intransitive variants) and
type of causation (both are typically considered externally caused). Therefore, any difference between
them cannot be attributed to unaccusativity or type of causation.
276
Malka Rappaport Hovav
(37a) can mean that a month passes and then there was a price increase/decrease; but
the only marginal reading available to (37b) is that after an hour, they finally began
to chat. Likewise (37d) can mean that after an hour, some descent on the part of the
plane took place (in addition to the reading that the descent took an hour).
Second, non-scalar change verbs can appear with a wide range of scale-denoting
XPs (Rappaport Hovav 2008:22).
(38)
a. Place butter square between two sheets of waxed paper and roll it smooth.
< http://books.google.nl/books?id=esrXej99JiYC&pg=PT351&lpg=PT351>
b. We rolled it flat.
c. We rolled the ball into the corner.
d. Rinse out all the soap and roll it dry between two towels, squeezing out the
water as you go. < http://www.ehow.com/how_2077057_cross-stitchsampler.html>
In contrast, the kinds of XPs that can appear after scalar change verbs are more
restricted. In particular, they are restricted to specifying a point on the same
dimension that is the basis of the scalar change lexicalized in the verb. This is
illustrated for both COS and DM verbs:
We froze the ice cream solid.
The chocolate melted into a messy goo.
The biologist dimmed the room to the level of starlight.
We arrived at the airport.
The leaves fell to the ground.
(39)
a.
b.
c.
d.
e.
(40)
a. *We dimmed the room empty. (cannot be used to describe a situation in
which we dimmed the room and as a result the people left the room,
resulting in an empty room)
b. *I thinned the soup tasteless.
c. *We broke the vases worthless.
d. *The vase fell broken.
This follows from a constraint which has been given a variety of formulations and
names but which essentially states that an event description can have only one
measure or scale (Tenny 1994, Goldberg 1995, Levin and Rappaport Hovav 1995).
Third, as Piñón (2000) has pointed out, verbs which lexicalize a scale appear with
the adverb gradually to indicate gradual progress along a scale.
(41)
a.
b.
c.
d.
e.
The sky gradually darkened.
The crack gradually widened.
The prices gradually increased.
He gradually approached the spot.
He gradually fell toward the ground.
Building scalar changes
277
In contrast, non-scalar change verbs do not on their own appear with gradually, but
require the addition of some constituent introducing a scale:
(42)
a.
b.
c.
d.
*(More and more of) the flags gradually waved in the wind.
Mary gradually ran *(more and more quickly).
The ball gradually rolled *(faster and faster).
The top gradually spun *(faster and faster).
Fourth, COS verbs and DM verbs show similar lexicalization properties. In
particular, the scalar change verbs in both domains typically do not encode a manner
component, demonstrating what Rappaport Hovav and Levin (2010) call manner/
result complementarity. This can be seen by examining the COS verbs in (43a) and
the DM verbs in (43b).
(43) a. break, crack, freeze, melt, clear, shatter, dissolve . . .
b. arrive, enter, come, go, rise, ascend, descend . . .
The COS verbs above specify the change that the theme argument undergoes, but do
not specify anything about how this change comes about as illustrated to (44a and b).
(44)
a. I cleaned the tub by wiping it with a sponge/by scrubbing it with steel
wool/by pouring bleach on it/by saying a magic chant.
b. We melted the chocolate by putting it out in the sun/by putting it into the
microwave/by stirring it over a low flame.
Like COS verbs, DM verbs specify a direction of motion, but they do not specify the
manner in which the motion is affected.
(45)
I came to the theater, running/hopping/walking/strolling/by bus/by bike/by
camel.
As Rappaport Hovav and Levin (2010) show, even verbs—such as climb—which
encode manners that are prototypically associated with a particular direction do not
strictly entail this direction:
(46)
a. “Bring the Governor’s reply straight back,” shouted Master Mace as
Mungo climbed the rope ladder into the ship’s rowing boat. (James
Riordan and Beaula Kay McCalla, Rebel Cargo, Frances Lincoln, 2007,
p. 149; L&RH 2014 (28))
(Note: climbing is from a ship to the ship’s rowing boat, i.e., downward.)
b. Marian climbed the rope ladder into the ship unaided, and was back on
board within 15 minutes of jumping. (L&RH 2014 (29))
The fact that both two-point (punctual) scalar change verbs and multipoint scalar
change verbs display manner/result complementarity is partial justification for
considering them both as belonging to the same category of scalar change verbs.
278
Malka Rappaport Hovav
Finally, there is a striking constraint on the argument realization patterns of COS
and DM verbs: the argument of the scalar change must be realized and must be
realized as a direct object (Rappaport Hovav and Levin 2005; Rappaport Hovav 2008;
Levin to appear). This is manifest in the fact that with COS verbs, the direct object
cannot be omitted in non-contrastive, non-habitual, or non-generic contexts (47a);
the argument undergoing the COS cannot be expressed as an oblique (47b) and these
verbs do not participate in the range of object alternations (48). Likewise, if we look
at the transitive variants of DM verbs,9 bring and take, they follow the same
argument realization constraints (49).
(47) a. *John dimmed/cooled.
b. *He broke at the table (cf. He hit at the table)
(48)
a. *My kids broke me into the poorhouse.
b. *We cooled the people out of the room.
(49)
a. * Casey brought/took to the party.
b. *Casey brought/took her parents’ house bare.
c. *Casey brought/took her arms sore.
This property also holds of both multipoint and two-point scalar change verbs,
further justifying analyzing them as belonging to the same larger class of scalar
change verbs.
12.6 Incremental theme verbs do not lexicalize a volume/extent scale
The third type of scale which figures in accounts of telicity phenomena is the
volume/extent scale (see section 12.1 above). These scales typically appear with
verbs that are characterized as incremental theme verbs (Dowty 1991). I claim,
however, that these verbs do not lexicalize any scalar structure and therefore exhibit
systematic differences from verbs which I have considered scalar verbs. In particular,
I suggest that the scalar properties associated with events denoted by predicates
headed by incremental theme verbs do not come from the verb proper, but from the
direct object. Therefore, to the extent that these verbs show variable telicity, this
comes from the way in which they combine with direct object, which introduces a
volume/extent scale. At the outset, it is worth pointing out that there is no recognized classification of incremental theme verbs according to the subtypes of scales
they take (two-point, multipoint, bounded, unbounded, lower-bound, upper bound,
etc.), nor is it at all obvious how one could come up with such a classification. This
does not stem from the nature of the scale: perhaps a volume extent scale by its very
nature cannot be a two-point scale; however, such a scale could be bounded or
9
For some reason, most DM verbs do not have transitive variants in English.
Building scalar changes
279
unbounded. Nonetheless, it appears that there are no verbs which can be distinguished from one another in terms of the boundedness of the volume/extent scale
lexicalized, in the way that COS verbs and DM verbs can be thus distinguished.
When we examine incremental theme verbs with respect to the properties that
were shown to be shared by COS and DM verbs, we find that the verbs from the
former class systematically differ from the verbs in the latter two classes, as expected,
if these properties diagnose verbs which lexicalize a scale. Many incremental theme
verbs can be used as regular activity verbs, and we can show that without the direct
object the verb does not act like a verb of scalar change and only with the introduction of the object it does.
First, we find that incremental theme verbs do not have the same natural affinity
to telicity as COS and DM verbs.
(50)
a. I mowed *(the lawn) in ten minutes.
b. I drew *(a picture) in ten minutes.
c. John carved *(a statue) in three months.
The verb mow can appear with or without a direct object. (50a) does not have the
achievement reading that the scalar change verbs in (37a,d) above have, nor does it
have the accomplishment reading when it appears without the direct object.10 Thus,
without the direct object, it does not show properties of a scalar change verb.
Second, it was shown above verbs of scalar change are restricted in the range of
scale-denoting XPs that they can appear with. They can only appear with XPs which
further specify something about their lexicalized scale. If the volume/extent scale of an
incremental theme verb were lexicalized in the meaning of the verb, we would expect
that incremental theme verbs could not appear with scale-denoting XPs which do not
refer to the volume/extent scale. This is contrary to fact, as illustrated below.
(51)
a. I wiped the entire table clean in three minutes. (Direct object provides
extent scale; property scale predicated of it.)
b. We painted the barn red. (Direct object provides extent scale; property
scale predicated of it.)
c. I read myself to sleep. (Unexpressed extent scale; property scale predicated
of another argument.)
d. They drank the teapot dry. (Unexpressed volume scale; property scale
predicated of another argument.)
10
There are, of course, examples such as John ate in two minutes/John packed in ten minutes. However,
such examples can only be used with specific kinds of direct objects which are lexically implied. When the
verb eat is used to describe a situation in which John ate a specified or non-specified quantity of popcorn,
the sentence cannot be used telicly without a direct object. Likewise John packed, cannot be used to
describe the packing of a freezer with meat.
280
Malka Rappaport Hovav
Third, incremental theme verbs display different properties with the adverb
gradually when they appear with and without the direct object. Specifically, verbs
which appear with or without a direct object can felicitously appear with gradually
only with the direct object. This of course makes sense if it is the direct object which
introduces the scale.
(52)
a. John gradually mowed *(the lawn).
b. We gradually ate *(all the cake).
Fourth, we saw that verbs of scalar change do not at the same time lexicalize
manners (non-scalar change). Prototypical incremental theme verbs (verbs of creation and ingestion) do lexicalize a manner component.
Finally, incremental theme verbs as a class do not show the same restrictions on
argument realization properties displayed by COS and DM verbs. In particular, the
direct object does not have to be expressed, and does not have to be expressed as a
direct object (though there is variation among verbs in terms of allowing the internal
argument to be expressed in a PP).
(53)
a.
b.
c.
d.
John read/ate/mowed.
John read us all to sleep./John read his eyes sore.
The goats ate the hillside bare.
We ate from the cake./He read from the book.
For many incremental theme verbs, there is no change entailed in the entity denoted
by the direct object; therefore, it is not possible to specify the change in that object
along any scalar dimension.
(54)
sing, dance, act, play, read, study, peruse, memorize . . . , sweep, wipe, rub . . .
Therefore, even if these verbs were lexically specified as being associated with some
sort of scale, these would not be verbs of scalar change. For verbs like read, peruse,
study, and memorize, if there is a change specified for any entity, it is the entity
denoted by the subject; when you want to know if a sign was read, you do not
examine the sign, but rather examine the reader.
It is striking that degree modification with these verbs works differently than
degree modification with verbs of scalar change (see also Gawron 2007). In particular, the degree morphology shows up on the direct object rather than on the verb
itself.
(55)
a.
b.
c.
d.
We
We
We
We
cooled the soup some more: → The soup is cooler.
wrinkled the shirt some more: → The shirt is more wrinkled.
lowered the rope some more: → The rope is lower.
descended some more: → We are further down.
Building scalar changes
(56)
281
a. Margie read the book some more: → *The book will be more read.
→ More of the book will be read.
→ Margie read the book again.
b. We perused the document some more: → *The document is more perused.
→ More of the document is perused.
→ We perused the document again.
Verbs like mow, used by Dowty (1991) as an example of an incremental theme verb,
do involve change in the direct object. I cannot have mowed the lawn without any
change occurring in the lawn. However, mowing involves a complex interaction
between the instrument and the theme, rendering the change complex and nonscalar in nature. In particular, if mowing involves changing the grass in the lawn
from being tall to being short, there is no relation between the parts of the event of
mowing and transition from (any of the grass) being tall to being short. That is, the
grass does not get shorter as the event of mowing progresses.
Strikingly, verbs like mow do not allow degree modification the way verbs of scalar
change do (Gawron 2007, Rappaport Hovav 2008, Levin and Rappaport Hovav
2008).
(57)
a. *My lawn was more mowed than yours. (cannot be used if my lawn in
shorter than yours).
b. More of my lawn was mowed than yours.
These examples show that mow patterns like incremental theme verbs that do not
involve change and does not pattern like scalar verbs. The conclusion I draw from
this is that although the direct object can serve as a scale for measuring the progress
of the event, the verb itself does not lexicalize a scalar change.
12.7 Conclusion
What has emerged from this study is that a scalar analysis of the lexical semantics of
verbs is a useful one. The scalar analysis of verbs brings out certain structural
parallels in the semantics of COS verbs and DM verbs, and the linguistic relevance
of these parallels is supported by grammatical and interpretive properties shared by
both classes of verbs. More important, perhaps, is the demonstration that it is
possible to isolate those components of the scalar semantics of event descriptions
that are lexicalized in the verb and those that are contributed by constituents in the
syntactic context of the verb. Distinguishing those aspects of the event description
which are lexically encoded in the verb from those which are not has lead to a deeper
understanding of the argument realization and interpretive properties of various
classes of verbs and lends further credence to the claim that it is possible to explicate
the grammatically relevant meaning components associated with a verbal root.
13
When roots license and when
they respect semantico-syntactic
structure in verbs
ANTJE ROßDEUTSCHER
13.1 Introduction
In this chapter I will investigate the contribution of roots to the syntactic and
semantic properties of verbs.1 From the point of view of a formal semanticist like
me the leading question of the investigation is “How can the semantics of verbs be
constructed from their roots?”. My way of tackling this question is by constructing
Discourse Representation Structures (DRSes), (see Kamp and Reyle 1993), from
word-syntactic representation in the spirit of Distributed Morphology (DM). Some
of the important word-structural assumptions I owe to Embick (2004) and Marantz
(2005). As far as the semantics of roots is concerned my assumptions aren’t much
different in spirit from those of Marantz (2009), Pylkkänen (2007), and Levinson
(2007b, this vol.): roots, as Marantz put it, come in three ontological flavors: events,
stative properties, entities.2 Subtle differences show up when it comes to deciding
how to implement the consequences of these differences in a syntax–semantics
interface. The leading idea of the current approach is this: roots have a semantics
which is the source of argument structure and which determines whether they can
be selected by certain functional heads such as v(verbalizer), n(nominalizer), a
(adjectivizer) in large numbers. For instance, eventive or “manner” roots like
√run are simple event types (directly) merging with v; the property root √dry
creates an argument for the bearer of the property ‘dry’ in the de-adjectival verb to
1
This work developed from joint work with the members of the projects B4 and of the long-term
research project Incremental Specification in Context, funded by the Deutsche Forschungsgemeinschaft.
I wish to thank an anonymous reviewer, Boris Haselbach, Hans Kamp, Florian Schäfer, Sylvia Springorum, and other members of the SFB.
2
With the latter two authors I also share the methodological commitment to model-theoretic
semantics.
Semantico-syntactic structure in verbs
283
dry. In contrast, sortal roots—or, as we also call them, entity roots—generalizing
over sorts like physical object, regions, and others typically fill argument slots that
are created by other roots. E.g., the sortal root √line (‘line’) in to underline a word
satisfies one of the two argument slots created by the preposition-like root √under.
(The other argument is contributed by the direct object of the verb underline.) In
German, where verb formation of this kind is common, it is common for sortal roots
to fill argument slots of preposition-like heads.
Roots, however, may also contribute to words in ways that do not match their
basic semantic categorization. In these cases I will speak of “unexpected contributions” of roots, whereas “expected contributions” are those that are in accord with
the word formation principles sketched above. According to my experience, the
unexpected contributions are comparatively rare, and I will demonstrate the restrictions in the course of the chapter.3
So the leading concern of the chapter is contrasting expected and unexpected
contributions of roots in German verbs. Admittedly, what counts as expected and
what as unexpected is a delicate matter, and for two reasons: (i) We must rely on
hypotheses on rule-based verb formation. (ii) We must rely on collateral evidence
indicating which rules apply when. As a consequence, in order to reveal the
contribution of the semantics of the roots to the semantics of the verb we need to
make use of analyzes that are hypothetical in at least two ways: they involve
hypotheses concerning patterns of word formation in German generally, and, (ii)
they must rely on hypotheses about which patterns are instantiated by which
particular verb.
13.1.1 Background hypotheses
Some hypotheses on verb formation rules emerged in the context of formulating
constraints on the formation of deverbal German nouns ending in -ung. The inititial
hypothesis was stated and defended in Roßdeutscher and Kamp (2010:182), (see also
Roßdeutscher 2010).
Hypothesis 1: Verbs with a bi-eventive structure allow for corresponding -ung nouns, verbs
with a mono-eventive structure do not.
In formulating this constraint, we rely on the syntactic implementation of bieventivity and mono-eventivity in Marantz (2005), which we understand as an
implementation of the dichotomy between “manner” and “result” in the verbal
lexicon as presented in Levin (1999). Simplifying a bit, our empirical hypothesis
can be re-phrased as follows: German -ung-nominalization mirrors the“manner”–
“result” dichotomy in verbs. Result-verbs like den Tisch säubern (‘to clean the table’)
3
As a consequence my approach is in line with the contributions of Alexiadou (this vol.), Levinson
(this vol.) and Rappaport (this vol.) and supports this view and contra Borer (this vol.), Acquaviva (this
vol.) or Acedo-Matellán and Mateu (this vol.).
284
Antje Roßdeutscher
have -ung-nouns (Säuberung); “manner”-verbs like laufen (‘to run’) or den Tisch
wischen (‘to wipe the table’) do not (*Laufung, *Wischung). Postponing the details,
we can say this much at this point:
If one of two verbs constructed from the same root has an -ung-noun while the other hasn’t,
this fact indicates that the root contributes to a “result”-verb in one of them and to a
“manner”-verb in the other. Following Hypothesis 1, we take this as an indication that the
root alternates in contributing to bi- vs. mono-eventively constructed verbs.
Moreover, if a root belongs to a class which, according to the predicted rules, forms verbs
bi-eventively and yet happens to be part of a verb that has no -ung-noun, then this root
occurrence counts as unexpected. Verbs alternating with respect to bi-eventivity
have proved a suitable starting point for discussing the role a root can play in a verb.
13.1.2 Two introductory examples
In German there is the following alternation between the forms (1a) and (1b) (The
corresponding verb in English does not alternate (see Dowty 1991).)
(1)
a. die
the
Kellnerin
waitress
füllte
filled
ein
a
Glas
glass
b. die
the
Kellnerin
waitress
füllte
filled
Tequila
tequila
mit
with
in
into
Tequila.
tequila
ein
a
Glas
glass
But the alternates are not equivalent in all contexts. Compare (2a) and (2b).
(2)
a. Johnny
Johnny
Depp füllte
Depp filled
die Kinokassen
the cinema’s.tills
(mit Millionen an Dollars).
(with millions of dollars)
b. (#) Johnny Depp füllte Millionen an Dollar in
die Kinokassen.
Johnny Depp filled millions of dollars into the cinema’s.tills
(2b) is grammatical, but sounds a bit strange compared with (2a), which seems
perfectly natural. It suggests a scenario where Johnny Depp moves from one cinema
to the next with a gigantic sack of money, pouring its contents into each cinema’s till.
We all know that Johnny Depp is a famous actor and that his way of filling the
cinema’s tills is by attracting customers, who buy tickets for film screenings in which
they show the film or films in which his participation has won a high reputation, and
(2b) just isn’t the right way to express that Johnny Depp filled the cinema’s tills in
that way. (2a) is different. When it is the tills that are the direct object of füllen (and
what they are filled with is expressed by an optional mit-PP), then any way of
causing them to be filled can be reported in this way.
In terms of a “manner”–“result” dichotomy füllen in (1a) and (2a) is a “result”verb, whereas it is a “manner”-verb in (2b). (1a) and (2a) are built bi-eventively, (1b)
and (2b) are built mono-eventively; e.g., füllen in the (a) examples has an -ung-noun
die Füllung des Glases/der Kassen; füllen in the (b) examples has no ung-noun,
compare *die Füllung des Tequilas/der Millionen. According to the implementation
Semantico-syntactic structure in verbs
285
of bi-eventivity in the present framework the root √full denotes a property in (1a)
and (2a) and an event type in (1b) and (2b); in the latter, word formation follows
direct merge in the sense of Embick (2004). Direct merge leads to unergative
(intransitive) and non-core-transitive verbs. Evidence for this is provided by the
fact that mono-eventive füllen passes the V-und-V test from Kratzer (2004): monoeventively built verbs are widely assumed to enter syntax as intransitive verbs. So if
füllen can be used intransitively we can take this as an indicator that the root √full
contributes to an unergative verb. The second conjunct of (3) is indeed of this form.
Moreover, the unergative verb must be an instance of füllen in (1b) and (2b) and
cannot be an instance of füllen in (1a) and (2a), because verbs following the deadjectival formation pattern have internal arguments without exceptions.
(3)
die Kellnerin stellte zwei Gläser auf den Tisch und fing an,
the waitress placed two glasses on the table und
started
den Tequila einzufüllen, und füllte und füllte.
the tequila
[in][to][fill] and filled and filled.
‘[ . . . ] and started to pour in the tequila and was pouring and pouring.’
We account for this fact that bi-eventively constructed füllen has no instransitive use
by assuming that the property-denoting root √full contributes a referential and a
non-referential argument slot of which the former must be satisfied obligatorily.
“Adjectival roots” in de-adjectival verbs contribute argument slots. We have seen
with füllen in (3) that √full can also contribute “manner” and has no argument slot
for any internal argument. If it had, the V-und-V construction in (3) would be
ungrammatical. This is so also for √full in (1b) and (2b), in particular, √full
contributes no argument slots in this case for any internal arguments of the verb.
This non-core-transitive (see Levin 1999, Marantz 2005) füllen exemplified in the
(b)-cases, is built from an event type denoting root √full by “direct merge” with v—
in the way that unergative and non-core-transitive verbs are generally built. But
there is more to this. On first sight in (1a), (2a) and (1b), (2b) it looks like whatever is
selected as direct object in the bi-eventive alternates becomes a goal argument
realized as an in-PP in the mono-eventive alternates, while the mit-PP of the bieventive alternate becomes the direct object. A closer look, however, shows that this
isn’t all. Whereas the bi-eventive verbs allow for any direct object that can be
described as voll (instantiating √full (‘full’)) and the DP in the mit-phrase may
be whatever the direct object ends up being full of, the direct object of the unexpected mono-eventive alternate imposes strict selection restrictions: As regards what
is poured into the container, only stuff is admitted that can be poured in a more or
less literal sense, i.e., liquids or solid matter consisting of parts that are able to move
relative to each other (such as sand, salt, pebbles, coins or apples rolling like
pebbles). The aggregates (4a,b) as opposed to (5a,c) demonstrate this. (4a,b) are
felicitous, whereas (5a,c) are odd. The asymmetry in (2a) vs. (2b) demonstrates
the same point.
286
(4)
(5)
Antje Roßdeutscher
a. eine Gans mit Äpfeln
a
goose with apples
‘to stuff a goose with apples’
füllen
fill
b. das Zimmer mit Rauch
the room
with smoke
‘to fill the room with smoke’
füllen
fill
a. # Äpfel in eine Gans füllen
apples into a
goose fill
b. Äpfel in einen Sack füllen
apples into a
sack fill
‘to sack apples’
c. # Rauch in einen Raum füllen
smoke into a
room fill
As will become more obvious in the course of the chapter, this change of selection
restrictions typically goes hand in hand with different contributions of the root. There
are two points relating to this I would like to make at this early stage of discussion.
First, the description of an action that is constructed by “direct merge” (see Embick
2004) thereby using a property root is suitable for describing behavior which can also be
described with the help of corresponding bi-eventive descriptions; for instance monoeventive füllen describes actions suitable for making something full. But not all
actions suitable for making something full can be described with the help of
mono-eventive füllen, only actions that are like pouring. Stuffing a goose with apples
is not like that. To let a heap of apples roll into a bag comes close enough to pouring and
that is why (5b) is felicitous. This restriction also explains why (2b) is odd. The hearer is
inclined to understand Millionen an Dollar in Kassen füllen as a case of pouring.
Second, the selection restrictions of direct objects in the mono-eventive descriptions of filling events must be suitable for direct manipulation by the agent. With the
inclination to interpret filling as pouring comes the inclination to interpret Millionen
an Dollar ‘millions of dollars’ as concrete—heaps of bills or coins—rather than
abstract (money as a figure in a bank account).4 What, out of all actions of making
This effect makes itself felt more drastically with the root √leer in alternations analogous to (1),
compare (i).
4
(i)
a. die Heizungsrohre leeren
the heating.pipes
empty
‘to drain the heating pipes’
b. # die Heizungsrohre ausleeren
the heating.pipes
[out]empty
‘to drain the heating-pipes’
The particle verb ausleeren is built mono-eventively. (There is no *Ausleerung der Heizungsrohre.) (i.b) is
infelicitous, because the pipes of a heating system cannot be handled in a direct way, like you do with
pouring water out of a bucket. They must be drained, i.e., emptied in some indirect way.
Semantico-syntactic structure in verbs
287
something full, is special about pouring? Pouring liquid or stuff of fine grained
granularity determines a definite point of termination for the activity under the
agent’s control, namely the point when the container is full of the liquid or stuff.
Apples with respect to geese in (4a) and smoke with respect to rooms in (4c) don’t
qualify. Pouring is a prototypical filling action, stuffing isn’t.
The two points above, for which more evidence will be provided, support the
following theoretical perspective on these alternations: the property root √full in
the mono-eventive alternate is “coerced” into a predicate of events. As a result of this
coercion the semantics of the root must obey the semantico-syntactic requirements
of the structural elements the root interacts with.
For the unexpected occurrences of √full in (2b) and (3) this means
(i) The root √full (‘full’) has no selective power of its own.5
(ii) “Direct merge” of √full and v can only denote prototypical actions of
making something full or fuller.
(iii) All direct objects must be under the agent’s direct control.
Arguably (ii) and (iii) can be viewed as a reflection of certain structural requirements: The “verbalizer”, i.e., the head “little v”, selects for agentive event types with
which it directly merges.
I would like to dwell on these observations on structural determination a bit
longer and discuss another example of two different semantic contributions of the
same root that has received considerable attention in the literature (e.g., see Kiparsky
1997 in this light; a similar example was recently discussed by Koontz-Garboden and
Beavers 2010 and Levin 2009, Rappaport Hovav and Levin 2010, Levin and Rappaport Hovav 2013, 2014). (6a–c) are fine, whereas (6d) is ungrammatical.
(6)
a. der Drachen stieg
(zum
the kite
ascended (to the
‘the kite flew up into the sky.’
Himmel) auf.
sky)
[up]
b. der Mann stieg
(zum Gipfel) auf.
the man ascended (to the summit) [up]
‘the man climbed to the summit.’
c. der Mann stieg
(vom
Gipfel)
the man
acscended (from the summit)
‘the man climbed down from the summit.’
d. *der
the
5
Drachen
kite
stieg
ascended
(vom
(from the
ab.
[down]
Himmel) ab.
sky)
[down]
Of course (i) doesn’t exclude the possibility that containers like sacks in (5b) occur in the sentence
predication. But when they do, they are selected by the preposition in, not by the root √fill
288
Antje Roßdeutscher
Kiparsky (1997) discusses this and similar examples as cases of disjunctive meaning.
Verbs like German steigen or English climb contain in their conceptual representation both a manner component (“clambering motion”) [for climb, A.R.] and a
direction component (“upward”); however, the lexicalization constraint (LC) permits only one of these to be lexicalized as part of the Semantic Form of the verb.
(LC) The lexicalization constraint: A verb can inherently express at most one semantic role
(theme, instrument, direction, manner, path, . . . ) (Kiparsky 1997:490)
The analysis in Levin and Rappaport Hovav (2014) is in the same vein, except that
climb alternates with respect to “manner” vs. “result”; direction, which instantiates
scalar change, is subsumed under “result”. According to these explanations it seems
an accident that (6c) lexicalizes “manner” and (6a) lexicalizes “direction”. But there
is a structural point: As a matter of fact it is only agentive uses of steigen, as in (6c),
where the semantics of √steig (‘climb’) isn’t in conflict with “downwards”. And this
restriction correlates with a syntactic difference between (6a) and (6b). The latter
allows for impersonal passives, whereas the former doesn’t, something that is shown
by (7), which only has an agentive “manner”-interpretation. This supports a syntactic analysis according to which the verb in (6b) has a vP-external subject; the agent
enters syntax in voice (see Kratzer 1996).
(7)
Je
höher gestiegen wird, desto
schöner ist die Aussicht.
The higher climb.imp.pass, the.more beautiful is the view
‘the higher people climb, the more beautiful their view is.’
The verb in (6a) has a verb-internal subject. The syntactic differences show that the
root √steig that forms the kernel predicate of the verb in (6a) and (6d) is not the
same as the root in (6b) and (6c). The root in (6a) is a relational predicate which
relates a situation (or, in the terminology I will be using, an event) to the verb
internal argument that manifests itself as subject. The verb steigen formed with this
version of the root √steig says of its event argument that it consists in the rising of
its theme argument as an effect of some external natural force such as buoyancy, as
we find with kites, balloons, smoke, etc. This is all the verb says, upward movement is
an inseparable part of its meaning. (If we tried to subtract it, nothing would be left.)
So when we try to combine it with ab- (down) we get an irreparable contradiction
and the combination aborts. The root √steig of (6b,c) on the other hand is a true
“manner”-predicate, a one-place predicate of events, that says of an event e that it
involves the effort and movements typical of someone who moves upwards, and thus
in a vertical direction, by his or her own strength. The default meaning in this case is
also that of an upward motion. But the manner features of such events are quite
similar to those of voluntary downward movements, enough for them to be able
to become the sole semantic contribution of √steig when upward movement is
taken away from it. This makes it possible for this instance of √steig to combine
Semantico-syntactic structure in verbs
289
meaningfully with ab- (down): the upward meaning of √steig gets overwritten by
“downwards”, contributed by ab-, but the remainder is retained (and there is enough
to survive on its own).
There are some aspects in which the two types of alternations ((1a) vs. (1b) and (2a)
vs. (2b), respectively) on the one hand as opposed to (6a) vs. (6b) on the other are
analogous and some in which they differ. The alternations are analogous in that in
each the two alternates share aspects of meaning: both alternates describe eventualities that are suitable for making something full or for a change of location. They are
also analogous in that there is change in selection and semantico-syntactic behavior.
But the differences are different. The füllen-alternates show bi-eventivity vs. monoeventivity; the steigen-alternates show verb-internal vs. voice subjects. This means
that our indicator of “expected” vs.“unexpected” contribution of a root, namely bi- vs.
mono-eventivity is applicable for füllen, but not for steigen. This is so, because
according to our tests6 all verbs in (6) are constructed mono-eventively. As a consequence the basic semantico-syntactic categorization of √steig must rely on other
heuristics. (N.B.: The refined version of the “manner”–“result” opposition from
Rappaport Hovav and Levin (2010), Levin and Rappaport Hovav 2013, 2014 is
orthogonal to the notion of bi-eventivity vs. mono-eventivity as used in the present
chapter (see (15)). The present notion is a kin of that of Levin (1999), but incompatible
with Levin’s and Rappaport Hovav’s current treatment of the dichotomy.)
13.1.3 Overview of the chapter
In the next section I will put some formal substance into the ideas presented above.
To that end, I will review the implementation of bi-eventively and mono-eventively
constructed verbs proposed in Roßdeutscher and Kamp (2010). I will present a
construction algorithm for the semantics of the expected and the unexpected uses
of the individual property denoting root in füllen and for the expected and unexpected use of the event property denoting root in steigen, both with and without
verb-external subject. In section 13.3 the same will be done for sortal roots. Section
13.4 will be devoted to examples that point in the opposite direction from those
discussed in section 13.3: event type denoting roots (which I will refer to as “manner
roots”) contribute (i) properties of individuals, (ii) entity denoting roots of the sortal
type “event” or (iii) entity denoting roots of the type “material object” to the verb.
If we ignore the two syntactic alternates of eventive roots, with the three basic
types of roots, property roots, entity roots, and “manner” roots, six pairs of expected
and unexpected contributions of a root are logically possible, see (8).
6
(i) None of the verbs have -ung-nouns: *die Steigung des Mannes, *die Steigung des Drachens. (ii)
Both pass the V-und-V test: der Mann stieg und stieg (O.K.); der Drachen stieg und stieg (O.K.); (iii) both
allow for resultative constructions (cf. Kratzer 2004 on verbs that start out as as unergatives): der
hochgestiegene Mann, (the man who has reached a high point in climbing) der hochgestiegene Ballon
(the balloon that has risen to a high level).
290
(8)
Antje Roßdeutscher
Expected Contribution
Unexpected Contribution
individual property
entity
event type (‘manner’)
event type (‘manner’)
individual property
entity
event type (‘manner’)
event type (‘manner’)
individual property
entity
entity
individual property
}13.2
}13.3
}13.4
}13.4
—
—
seldom
common
common
seldom
From these six possibilities only four will be discussed in detail. In section 13.5 I will
give reasons for why the remaining two possibilities are not discussed. In that section
I will also give an outlook on further research, to be focused on the question of
relative frequency: “How common are unexpected contributions of roots?”. The
fourth column in (8) only lists tendencies.
13.2 Syntactic structure and semantics construction
13.2.1 Property roots are coerced to “manner” roots
In (10a,b), I display the syntactic representation of the alternates in (9a,b) which are
slightly modified versions of (2a,b). In order to keep the representation as simple as
possible I have changed the description of the money into a quantized description.
This makes the mono-eventive description telic, whereas the original (mono-eventive) description (2b) was atelic. Note that both (bi-eventive) descriptions (2a) and
(9a) are telic, irrespective of whether a mit-phrase is present.
(9)
a. die
the
Kinokassen
cinema’s.tills
b. eine
one
Million
million
(mit
(with
Dollar
dollars
einer
one
in
into
(10)
a.
Million
million
die
the
Dollar)
dollars)
Kinokassen
cinema’s.tills
füllen.
fill
füllen.
fill
b.
vP
vP
v
rP
pP
r⬘
DP:akk
die
Kinokassen
PP
DP:dat
mit
einer
Million
Dollar
p
PP
r
√full
DP
eine
Million
Dollar
P⬘
P
in
vP
DP
die
Kinokassen
v
√full
Semantico-syntactic structure in verbs
291
In (10a) the root √full (‘full’) functions as the head of a root phrase, which is
selected by v.7 As alluded to in the introduction the root contributes two argument
slots, one for the container to become full, syntactically realized by the direct object
and one for a theme in prepositional case. (The second is syntactically optional and
when it is not overt it can often be reconstructed from context.) The derivation fails
if the former argument slot isn’t saturated.
(10b) shows the syntactic analysis of mono-eventive füllen. We have direct merge
(see Embick 2004) of the root with the verbalizer v, forming the unergative füllen.
The direct object and the goal phrase are assumed to be part of a single prepositional
phrase.8
Figure 13.1 and Figure 13.2 show the steps of the bottom up semantics construction
on the basis of (10a) and (10b).
The semantic representations I am using in this chapter have the following form.
In general they are triples consisting of (i) a string of lambda abstraction operations
(such as e.g., λx.λy.); (ii) a store containing one or more discourse referents; and (iii)
a DRS. An example is (11).
(11) λz. λy.
s,
s: full(y,z)
The λ-bound symbols “z” and “y” are “argument slot holders”. Their only other
occurrence in the representation is as argument positions of some predicate in some
DRS condition—here the first and second argument position following the predicate
full in the DRS to the right. The symbol “s” to the immediate left of the DRS is a
discourse referent, which fills the “referential” argument slot of the predicate ‘full’,
which is here assumed to be singled out by its position to the left of the predicate and
separated from it by “:”. Representations with a non-empty lambda prefix or a nonempty store (or both) can only occur as intermediate representations. It is a wellformed constraint on final representations of complete sentences (and larger pieces
of discourse) that (i) all lambda-bound slots have been instantiated by discourse
referents (through “lambda conversion” or “functional application”) and (ii) no
discourse referents remain in the store. (A discourse referent is removed from the
store when it is bound (in any of the ways allowed for by the given construction
algorithm). Binding of a discourse referent always involves placing it in some DRS
universe. The most common form of binding is that in which the discourse referent
7
An alternative is to represent the head of rP as a functional adjectival head that is modified by √full.
As the differences between the two solutions for the syntactic representation are not relevant to the
arguments in this chapter, I have opted for the simpler representation. A detailed discussion on principles
of the syntax–semantics interface is left for another occasion.
8
I follow the split-P hypothesis of Svenonius (2003) and van Riemsdijk (1990).
292
Antje Roßdeutscher
vP
c⬘s-tills(Y) 1 M. dollars(Z)
e⬘ cause s
s:full(Y,Z)
e⬘,Y,Z,s,
v
rP
Y,Z,s,
e⬘,
r⬘
die
Kinokassen
Y,
c’s-tills(Y) 1 M. dollars(Z)
s:full(Y,Z)
λy.
c’s-tills(Y)
Z,s,
1 M. dollars(Z)
s:full(y,Z)
r
PP
Z,
√full
1 M. dollars(Z)
λz.λy.
s, s:full(y,z)
DP:dat
mit
einer
Million
Dollar
Figure 13.1 Semantics construction for die Kinokassen (mit einer Million Dollar) füllen
is simply transferred from the store to the universe of the main DRS to its right; this
imposes on the discourse referent the existentially quantified interpretation familiar
from the version of DRT of, for instance Kamp and Reyle (1993).
The semantic representations of lexical items I will be using in this chapter are
simplifications of the general schema in (11) in one of two ways: either they have an
empty store or they have an empty lambda prefix. But when such representations are
combined into the representations of syntactically complex expressions, these latter
ones may have non-empty stores as well as non-empty lambda prefixes.
Crucial for the bi-eventive structure in Figure 13.1 as compared with the monoeventive structure in Figure 13.2 are (i) the denotation and (ii) the position of the root
√full. As already alluded to in section 13.1.2, the individual property is the head of a
r(oot)P which denotes a state s. The verbalizer “little v” selects rP. The referential
argument e0 (contributed by “little v”) is interpreted as the causal antecedent of the
Semantico-syntactic structure in verbs
293
vP
e⬘,Y,Z,s,
res(s,e⬘)
1 M. Dollars(Y) c’s tills(Z)
s:Y ⊆int(Z) full(e’)
pP
vP
λe. Y,Z,s,
e’,
res(s,e)
1 M. Dollar(Y)
c’s-tills(Z)
s: Y ⊆int(Z)
e⬘,
full(e⬘)
√full
λe. full(e)
p
PP
Y,Z,s
eine Million
Dollar
Y,
1 M.D(Y)
λs.λe.
1 M.D.(Y)
c’s-tills(Z)
s: Y ⊆int(Z)
res(s,e)
P’
P
in
λz.λy. s, s:y ⊆i(z)
d.Kino-K
Z,
c’s tills(Z)
Figure 13.2 Semantics construction for eine Million Dollar in die Kinokassen füllen
state s, contributed by rP. The semantics of vP is to be read as follows: there exist e0 ,
s, Y, and Z, such that e0 brings about s, where s is a state of the cinemas’ tills Y being
full of one million dollars Z.
The root √full occurring in Figure 13.2 the semantics of which is a one-place
event predicate is the result of an operation of coercion, which is assumed to have
already taken place. The steps involved in this coercion are displayed in (12). The
structure in (12c) cannot be formed in a direct compositional way on the basis of the
semantics assumed for the root √full in (12a). Therefore the semantics of √full is
modified so that it can combine in the way that merge of v and r requires, viz. as a
case of argument insertion (functional application).
294
Antje Roßdeutscher
(12)
a.
root semantic
r
√full
λz.λy. s,
s:full(y,z)
b.
selection failure
c.
reinterpretation
vP
?
v
e⬘,
e⬘,
r
√full
λz.λy. s,
s:full(y,z)
full(e⬘)
r
√full
v
e⬘,
λe.
full(e)
Direct merge (see Embick 2004) of the root represented in (12a) with the verbalizer v
isn’t possible, because v requires specifying conditions for the process e0 . It is this
structural requirement that triggers the shift in the syntactic and semantic contribution of the root.
The property root √full can act as a predicate of e0 if it is reinterpreted as a
“manner” root, i.e., as an event property, as in (12c). As a side effect of the coercion,
the argument slots for y and z are filtered out.
The semantic differences mentioned above are plausible in the light of the
structural differences. In the bi-eventive event description (2a) whatever Johnny
Depp did caused the result state of the tills being full. This is an intrinsic part of the
predication expressed by the bi-eventive structure. The bi-eventive predicate is
silent about what the agent actually did, if indeed he did anything at all. There is
no manner specification of e0 . Quite the opposite is true in the case of (2b) and (9b).
The vP-representation doesn’t entail that the cinema tills are full. The monoeventive description (9b) of the filling event only commits us to the conclusion
that there are one million dollars in the cinemas’ tills. We infer that Johnny Depp’s
money-pouring action had come to an end at that point. As for mono-eventive
filling descriptions, Johnny Depp can stop pouring money into the tills at any point,
the money being under his immediate control. Whether the event is described as
one with a culmination depends on whether or not the direct object is quantized.
If no direct object phrase or PP is adjoined to the unergative structure of the vP
in Figure 13.2 (as is the case in the V-und-V construction sie füllte und füllte in
(14)(=3)), the sentence supports hardly any result state inferences. Let’s look at this
last description from another angle: assume the following situation. The waitress
intends for a certain glass to become full of tequila. She starts pouring tequila into it
and stops when the glass is full. (13), a repetition of (1), truthfully describes this entire
action. The second conjunct of (14) describes stretches of pouring, beginning anywhere between the time she started pouring and ending anywhere before she
stopped.
295
Semantico-syntactic structure in verbs
(13)
die Kellnerin füllte
the waitress
filled
ein
a
Glas
glass
(mit
(with
Tequila).
tequila)
(14)
[die Kellnerin begann den Tequila einzufüllen] und füllte und füllte.
and filled and filled
‘the waitress started pouring tequila poured and poured.’
We can conclude: If there are culmination conditions in mono-eventive structures
they must be contributed by quantized direct object phrases. A direct object phrase
of the form zwei Deziliter Tequila (‘two deciliters of tequila’) instead of Tequila or a
quantized description eine Million Dollar (‘one million dollars’) instead of Millionen
an Dollar (‘millions of dollars’) do.
To give an intermediate summary, (15) displays the dichotomy between
bi-eventive verbal structures (15a) and mono-eventive verbal structures (15b1)
and (15b2).
(15)
bi-eventive
a.
mono-eventive
b1.
b2.
vP
vP
e⬘, s, e⬘ cause s
s:φ
XP
s,
s:φ
v
e⬘,
φ(y)
ϕ(e⬘,y)
e⬘, y,
vP
e⬘, ϕ(e⬘)
v
e⬘,
rP
v
rP
e⬘,
λe. y,
λe. ϕ(e)
comp
y,
φ(y)
ϕ(e,y)
r
λy.λe. ϕ(e,y)
φ(y)
The semantic characterization of XP in (15a) generalizes over rPs, aPs, and PPs. In
the former, the head is a one-place- or two-place relation introduced by property
roots or prepositional roots. Examples involving prepositional roots will be given in
the next subsection. Functional heads like a(djectival) may also play the role of the
heads of such phrases. (15b1) represents unergative verbs as in (6b), (15b2) unaccusative verbs with non-agentive subjects. E.g., the root √steig in (6a), repeated as
(16a) instantiates (15b2).
296
(16)
Antje Roßdeutscher
a. der Drachen stieg
(zum
the kite
ascended (to.the
‘the kite flew up into the sky.’
Himmel)
sky)
b. der Mann stieg
(zum Gipfel)
the man
ascended to.the summit
‘the man climbed to the summit.’
auf.
[up].
auf
[up]
The question that is still to be answered is whether the basic semantic categorization
of the root √steig has the form in (15b1) or has the form (15b2). Assuming that one
direction is more basic than the other, which is indeed the case, then in which
direction does reinterpretation take place? Although the matter cannot be decided
straightforwardly, the reinterpretation direction is probably from agentive and
simple eventive to non-agentive and relational eventive. This is in line with what
has been claimed about English climb in Levin and Rappaport (2013) and work
cited therein. A point in favor is the fact that the range of suitable subjects for
agentive steigen is apparently more restricted than for its agentive counterpart. The
change from less restricted to more restricted selection restrictions seems to be
a general feature of root coercion. Unaccusative steigen exclusively selects subjects
that are physical objects where upward motion is due to buoyancy.9 When this
restriction is satisfied, an interpretation of the contribution of √steig as describing a
downward change of location is excluded, because it is now an intrinsic part of the
semantics of the root that buoyancy is the cause of motion; and such motions cannot
be downward.
The counterpart of (12) for the coercion process of the root occurring in (6a) =
(16a), is displayed in (17).
According to the entry of steigen in Grimm and Grimm (2007) √steig was initially a root suitable to
describing autonomous change of location activities, which, like English climb, involved “resisting the pull
of gravity” (cf. Levin and Rappaport Hovav 2013) but also in the sense of keeping one’s balance. In
Goethe’s work one finds er steigt die gefährlichsten Kanten wie im Schlaf, meaning ‘walking along the most
dangerous edges [next to a chasm] as if asleep’. How the default of upward motion in the agentive use on
the one hand and the strict inference to upward motion with the unaccusative steigen in (6a) might have
come about is impossible to say without an in-depth historical investigation, which I am not in a position
to undertake. But, apparently the use of steigen to speak of flying or swimming animals such as insects,
fishes, and birds that overcome gravity by their own strength has been a factor. Impersonal passives like
that in (7) seem not, or only marginally applicable to flying animals and not to swimming ones. This
suggests that the use of steigen is already the result of the transition from the one-place √steig of (6b) to
the two-place √steig of (6a). This is an interesting fact, given that the motions of animals that steigen can
be used to describe are agentive in the fairly strict sense that (usually) they are voluntary and that the agent
must itself produce the force that propels it. There thus appear to be at least two distinguishable
components to the transition: the transition from a one-place predicate that allows for a voice projection
to a two-place predicate that does not; and (ii) the transition from a predicate that describes its event
argument as an activity performed in a certain “manner” to a predicate that describes its event argument
as a motion event with a specific cause (buoyancy) and a direction (upwards) that is entailed by this
particular kind of causation.
9
Semantico-syntactic structure in verbs
297
(17)
a.
b.
structural
failure
root semantic
c.
reinterpretation
rP
voice
vP
v
e⬘,
r
√steig
resist or
overcome
λe.
gravity(e)
move(e)
autonom.(e)
comp
y is
subject to
buoyancy
r
√steig
λy.λe.
move(e,y)
align(path(e,y),
vert)
¬auton.(e,y)
Besides the syntactic differences between the root in its basic categorization in (a)
and its coerced categorization in (c), direct merge and a voice projection including
an agentive subject—as opposed to a voiceless projection of a verbal predicate which
contains an argument slot for the subject from the start—we have (i) conflicting
conditions between “autonomous motion” in (a) and “non-autonomous motion”
in (c). This difference is decisive for the syntactic restructuring involving vP-internal
subjects. As we have already seen, the coercion operation also involves strengthening
of the conditions conveying the direction of the motion. Whereas in the semantics of
√steig in (17a) there is an implication that the motion is vertical, which correlates
with the manner of motion that animals and humans engage in primarily when they
move up, and to some extent also when they move down, the upwards motion has
become an essential part of the semantics of the reinterpreted √steig in (17b) (cf.
align(path(e,y),vert)).10
13.3 More of the same: Entity-denoting roots are coerced
to manner-denoting roots
There is another alternation in German, for which there is no direct counterpart in
English. An instance is presented in (18), with the sortal root √deck instantiated in
the nouns (i) Decke, (ii) Deck, and (iii) Deckel. Prototypical instances of the nouns
are: (i) either (a) piece of cloth serving as a cover, such as a tablecloth or a blanket or
bedspread or (b) the ceiling of a room or other enclosed space; (ii) a solid plane that
serves as the upper closure of what is below it and at the same time as floor like
10
Levin and Rappaport (2013) assume that the conditions in a. should determine “manner”, whereas
those in (c) should lack this determination. In this chapter “manner” is used as a label for “specifying
agentive actions”.
298
Antje Roßdeutscher
English deck (as in the deck of a ship); (iii) a cover or lid of a container, such as a jar
or a box. √deck too can enter into a verb both as sortal and as manner root. The
most common verb in which it occurs as sortal root is the prefix verb bedecken in
(18a) whereas the “simple” verb decken is primarily used as a verb in which √deck
acts as a manner root, s. (18b)11
(18)
a. den Kopf (mit einem Tuch) bedecken
the head (with a
cloth) pref.cover
‘to cover the head (with a cloth)’
b. ein Tuch über den Kopf decken
a
cloth over the head cover
‘to spread a cloth over the head’
For good measure, I briefly repeat the syntax and the semantics construction from
Roßdeutscher and Kamp (2010) and Roßdeutscher (2010) according to (19a) in
Figure 13.3. The reader is referred to the cited papers for a more detailed discussion
and also for the readings of -ung-nominalizations of this pattern. (I skip the
semantics construction of (19b) because it is much like in Figure 13.2.)
vP
e⬘,s,y, η,
the head(y) cover∗(η)
e⬘ cause s s: bei(y,η)
v
PP
e⬘,
the head(y)
s:bei(y,η) cover∗(η)
s,η,y,
P⬘
den Kopf
y, the head(y)
λy η,s,
P
be(i)
λz.λy.
cover∗(η)
s:bei(y,η)
√deck
η,
s,
cover∗(η)
s:bei(y,z)
Figure 13.3 Semantics construction for den Kopf bedecken
11
Evidence for the different structure is -ung-nominalizability of the verb in (18a), but not of that in
(18b). Compare die Bedeckung des Kopfes to the ungrammatical *die Deckung eines Tuches über den Kopf.
299
Semantico-syntactic structure in verbs
(19)
a.
b.
vP
vP
PP
PP
mit e. Tuch
pP
v
DP
P⬘
DP
d. Kopf
p
PP
PP
P
√deck
e. Tuch
vP
v
√deck
P⬘
P
d. Kopf
über
beA brief comment on this construction: the preposition-like prefix be- denotes a
stative application relation between two individual discourse referents. In (19a) it
selects a sortal root as internal and a DP as external argument. The former contributes both a discourse referent η (of some undetermined mereological type) and
the sortal property cover of η. The semantics of the PP is a state such that the head
is provided with a cover. The rest of the construction follows the bi-eventive
formation rules.
Here too we can observe the change in the selection requirements for direct
objects, as in (20).
(20)
a. den Holzstapel
the stack.of.wood
Tisch mit einem
table with a
‘to cover the stack of
mit einer Plane
bedecken
/ den
with a
tarpaulin pref.cover / the
Tuch bedecken
cloth pref.cover
wood with a tarpaulin/the table with a cloth’
b. eine Plane
über den Holzstapel
decken / ein Tuch
a
tarpaulin over the stack.of.wood cover / a cloth
über den Tisch decken
over the table cover
‘to spread a tarpaulin over the stack of wood/a cloth over the table’
c. den Boden mit Wasser / mit Krümeln bedecken; die
the ground with water / with crumbs pref.cover the
Hand mit Küssen bedecken
hand with kisses pref.cover
‘to cover the ground with water/with crumbs; to cover the hand with kisses’
300
Antje Roßdeutscher
d. *Wasser / *Krümel / über den Boden decken; *Küsse über
water
/ crumbs
over the ground cover; kisses over
die Hand decken
the hand cover
While the bi-eventive verbs (20a) allow arguments in mit-phrases that can be
conceptualized as ending up in some way or another as a cover of the direct object,
the direct objects in (20b) must in addition qualify as two-dimensional material
objects just as the prototypical instances of √deck do. As a consequence, tarpaulins
and cloths, but not water or crumbs, can be direct objects of the verbs with the
coerced root denotation. Again we observe that the coerced root contribution leads
to descriptions of situations that can also be described by verbs containing the noncoerced, sortal version of the root; but it imposes the additional constraint that
the direct objects must be prototypical instances of the denotation of the nominal
root.
The examples in (21) are also built from a nominal root √lad (‘load’) the
denotation of which shows up in German as the entity-reading of the -ung-noun
Ladung (load). But as the nominal root is ambiguous—either denoting (i) stuff to be
transported in or on a vehicle or (ii) ammunition, or (iii) some amount of positive or
negative electricity—all direct objects in (21d) instantiate the root’s denotation and
all sentences are acceptable (even if (21c) is more familiar to most of us).
(21)
a. einen Wagen
a
wagon
b. Heu
hay
auf
onto
mit
with
einen
a
Heu
hay
Wagen
wagon
(be-)laden
(pref).load
laden
load
c. eine Batterie mit 5A(ampere) laden; ein Gewehr mit
a
battery with 5A(ampere) load; a
rifle
with
Munition laden
munitions load
‘to charge a condenser with voltage; to load a rifle with ammunition’
d. 5A in
die Batterie laden; Munition
in
ein Gewehr
5A into the battery load; ammunition into a
rifle
laden
load
A comparison of my presentation of the laden- or load-alternation with solutions in
the literature (see, e.g., Dowty 1991, Basilico 1998, Hale and Keyser 2002) is called for
at this point. However, the details of such a comparison have turned out to be rather
involved. A comparison will therefore has to wait for another occasion.
Semantico-syntactic structure in verbs
301
13.4 Coercion of event denoting roots
13.4.1 Event properties coerce to result state properties of individuals
I will mention these cases here informally because they have already been discussed
to some extent in the context of the theory of -ung-nominalization (cf. Roßdeutscher
and Kamp 2010 and Roßdeutscher 2010). According to (15b) verbs built from
manner roots by direct merge are mono-eventive constructions. Prefix-verbs with
the same root, however, are in certain cases bi-eventively constructed, see (22). The
simplex verbs in (22a,c) have no -ung-noun, (22b) has an -ung-noun. Note however
that the be-verb in (22d) does not.
(22)
a. der
the
Mann
man
arbeitete.
worked
b. der Beamte
bearbeitete
die Akte.
the civil.servant pref.worked the file
c. der Hund bellte.
the dog
barked
d. der Hund bebellte
den Briefträger.
the dog
pref.barked the postman
In cases like (22b) where the simplex verb has no -ung-noun but the prefix-verb
does, judgements often vary as to whether the respective ung-nouns are well-formed.
If the complex verb isn’t telic (which is more often the case than not, see section 13.5)
-ung-nominalization is excluded. This general picture gave rise to the idea that the
event descriptions are coerced into the following productive pattern in which the
prefix be- selects properties, which can be supplied by property roots. The pattern
yields bi-eventively constructed verbs. An instance of this pattern is (23). The twoplace relational prefix selects a property (e.g., √feucht (‘humid’)) which is applied
to the individual discourse referent representing the direct object.
(23)
der Beamte
befeuchtete die Briefmarke.
the civil.servant pref.humid the stamp
‘the civil servant moistened the stamp.’
In brief, (22b) is reconstructed along the lines ‘the file is brought into a state of being
done with’ (just as (23) means ‘the stamp is brought into the state of being humid’).
It should be noted that in this case coercion is triggered by the selection restrictions
imposed by be-12.
Coercion of roots denoting event types to roots denoting resultant state properties
is also accompanied by change of selection properties. This phenomenon doesn’t
12
Coercion from manner to resultant state properties can also be observed for other German prefixes
like ver- and zer-, and of particles such as ein- and ab-.
302
Antje Roßdeutscher
play any role with simple event types where no verb-internal selection takes place.
But it does show up when roots contributing relational event types are coerced into
roots denoting resultant state properties, for instance in (24).
(24)
a. Aus
out.of
der
the
Wunde
wound
quoll
welled
Blut
blood
b. Das Holz quoll
the wood swelled
(24a) describes a change of location of some liquid conceptualized as a simple
natural event; it is described in mono-eventive terms (*Quellung des Blutes is
ungrammatical). (24b) describes an event that is the result of an intervention: the
wood expanded because of the humidity. The structure is bi-eventive, (cf. Quellung
des Holzes). The current account predicts this difference: in (24b) the root √quell
has lost the selective power that it has in (24a).
13.4.2 Event type denotations shift to sortal denotations
13.4.2.1 Shift to a sortal root of type event Roots which denote event properties can
occasionally enter sub-lexical structures much in the same way that entitity denoting
roots productively do. This is shown by alternations like those in (25)13 and (26).
(25)
a. das Konzert endete.
the concert ended
b. der Pianist beendete
das Konzert.
the pianist pref.ended the concert
‘the pianist terminated the concert.’
(26)
a. der Postbote eilte.
the postman hurried
b. der Postbote beeilte
sich.
the mailman pref.hurried refl
‘the postman took a hurry.’
(25a) and (26a) are constructed mono-eventively, (25b) and (26b) bi-eventively.
(Evidence: *die Endung des Konzerts is ungrammatical, whereas die Beendung des
Konzerts is well-formed). The root √end is a relational eventive root like the root
√steig (i.e., it denotes a relation between an event and an individual, the “internal
subject”). The root selects events with natural endings and, more importantly, these
endings are understood to occur without any external intervention: normally concerts, seasons, diseases, punishments come to an end all of their own. But from this
open list only concerts and punishments can be “given” an end. As a consequence
13
The example is taken from Bierwisch (2008).
Semantico-syntactic structure in verbs
303
verbal predicates of the form *den Winter, *die Krankheit beenden (‘to end the
winter, the disease’) are ungrammatical.
This is expected regarding the fact that in (25b) the root does not act as head of a
root phrase. (The root in (26a) is a simple eventive root (‘manner’ root), which is
also coerced to a sortal root in (26b)).
The syntax–semantics interface of (25b) and (26b) is as in (19a) for bedecken (‘to
cover’). The vP-representation for das Konzert beenden (‘to end the concert’) is
displayed in (27). (The structure should be read as follows: “the referential argument
e’ of the verb, which is contributed by (unmodified) v, brings about the state s. s is a
state of the concert e1 having been provided with an end e2).
(27)
the concert(e1)
e⬘ cause s
e⬘, s, e1, e2,
s:bei(e2,e1)
end(e2)
√end undergoes coercion from a relational event type to a sortal root of an
eventuality type. As in cases discussed earlier in this chapter coercion of a root is
triggered by the selection of the other root it merges with. In (25b) the selector is be-,
requiring an entity for its second argument slot. The slot can be filled with the
discourse referent e2 supplied by the sortal root √end (a discourse referent representing an eventuality).
(28) a.
root semantics
b.
selection failure
c.
reinterpretation
P⬘
λy. e2,s, s:bei (y,e2)
end(e2)
?
λy.λe.
end(e,y)
z is entity
be
λz.λy. s,
P
be
s:bei(y,z)
λz.λy. s,
e2,
end(e2)
s:bei(y,z)
In (25), (26) the sortal information, i.e., the ontological sort “event” is preserved by
the coercion. But are there also cases of coercion from eventive root types to other
ontological types? In the next subsection I will discuss examples that can be seen as
instances of such coercion.
304
Antje Roßdeutscher
4.2.2 Shift to sortal roots of non-eventive type There is a small number of manner
roots that can enter syntactic structure like sortal roots denoting material objects,
rather than events. Examples are √bau (‘build’), √schreib (‘write’) and √mal
(‘paint, draw’). Collateral evidence for this type of coercion is provided by entityreadings of the -ung-nominalizations Beschreibung (des Potsdamer Platzes),
Bebauung (des Potsdamer Platzes), Bemalung (des Potsdamer Platzes) of the
respective be-verbs den Potsdamer Platz bebauen, beschreiben, bemalen.
The noun Bebauung is ambiguous. It can describe an event of building, (say,
skyscrapers) on the square, or the square’s state of having those buildings, or even
the skyscrapers themselves. There is a strong empirical correlation between entityreadings and sortal roots in verbal constructions formulated in Roßdeutscher and
Kamp (2010) in the form of as the following hypothesis:
When an -ung-noun refers to an entity, then it has a sortal root and it refers to the entity
contributed by this root. Moreover, this entity must be conceptualisable as resulting from the
event described by the corresponding verb. (Hypothesis 5, Roßdeutscher and Kamp 2010:204)
Our finding is in line with Hypothesis 5. Moreover, there is collateral evidence from
English counterparts of the German roots in question: the English derived nominals
writing, building, and drawing also denote entities as well as events. And note that it
is true of all these roots, both the German and the English ones, that the events they
denote are production activities, activities that are performed in order to bring about
the entities which are denoted by the derived nominals. Nevertheless, √bau,
√schreib, and √mal are outliers. I don’t know of any other manner roots that
allow reinterpretation as sortal roots of the sort “material”.
This leads us to a further dimension of the phenomenon described in the last
three sections, that of frequency.
13.5 How common are unexpected contributions of roots?
13.5.1 Sortal roots functioning as manner roots
13.5.1.1 Reinterpretation of sortal roots as manner roots is not uncommon Of the
“frequency” aspect of what I have presented as cases of root coercion: how
common are verbs that involve such root coercion? To obtain a clearer picture of
this statistical dimension, it is necessary to carefully inspect large amounts of data.
And the problem isn’t just that so much data needs to be inspected. For many roots,
deciding what should be considered their “original” meaning is a delicate matter. First
sight intuition often has to be revised upon closer scrutiny. One example of this kind
is the root √grab (‘dig’) which we find on the one hand in German unergative
graben (to dig) and on the other in the transitive prefix-verbs begraben (to bury) and
untergraben (to undermine). Unergative graben, meaning ‘to dig’ has a monoeventive structure (and therefore no derived -ung-nominalization *Grabung),
Semantico-syntactic structure in verbs
305
whereas the verbs begraben and untergraben do have the derived -ung-nouns
Begrabung and Untergrabung and thus must be bi-eventive. In these verbs √grab
has a semantics similar to its English kin √grave viz. a predicate true only of graves.
As a consequence, a reconstruction of the unergative graben as direct merge of “little
v” with a coerced root as described in section 13.3, seems straightforward. This is not
to say that a speaker of present-day German might be aware of any semantic
connection between the verbal constructions. There is no reason to expect this.
Roots have a history and verbal constructions have a history, too. And speakers may
be expected to know their language without knowing its history. The reader might
doubt whether graben (to dig) should be viewed as a mono-eventively constructed
“manner”-verb in its own right. This, of course, is an option. √grab would be
homonymous, then. Still, reconstructing the connections along the lines of a single
root √grab might reveal insights into how a lexicon may exploit the semantics of its
sub-lexical units. Such insights remain hidden with the homonymy assumption.
There are many alternations along the lines of den Kopf bedecken vs. etwas über
den Kopf decken, recall (20). An extended list of examples from a corpus study is
documented in Roßdeutscher (2010). For each example, I argued there that the
interpretation of the mono-eventively constructed verbs can be viewed as parasitic
on the bi-eventively constructed ones.
13.5.1.2 Coercion of sortal roots to eventive roots is restricted It might be thought
that it is quite common for a sortal root to change its semantics to that of a manner
root. But in fact not all sortal roots undergo sort-to-manner coercion easily. For
instance the root √kleid (‘dress’) enters German verbs as denoting stuff used to
cover the surface of material objects (including persons).14
Other sortal roots that appear to be reluctant to permit coercion into “manner” roots
are √sold (‘wage’) as in die Beamten besolden (‘to give wages to the civil servants’);
√stuhl as in einen Saal bestuhlen (‘to furnish a hall with seats’) and others. (There
are no verbs *solden or *stuhlen nor mono-eventive particle verbs like *absolden or
*abstuhlen). The reason for this resistance to manner-coercion seems to be that for
these sortal roots it is very difficult to abstract away from the specific information
they supply while retaining what is typical of the activities involved. (What could be
the prototypical manner of events described by verbs like bestuhlen or besolden?)
However, many more data have to be explored to get a better grip on the facts
about coercion of particular roots.
14
(i) presents a list of verbal constructions with √kleid, none of which is mono-eventive.
(i) eine Person be-kleiden ‘to dress someone’
ein(e) Ding/Person [ein]kleiden ‘to dress someone, to wrap something entirely’
ein Brillenetui [aus]kleiden ‘to back something with tissue’
eine Person [an]kleiden ‘to dress someone’
eine Person [um]kleiden ‘to change someone’s clothes’
einen Kühlschrank ver-kleiden ‘to cover a (built-in) fridge’
306
Antje Roßdeutscher
13.5.2 Roots denoting properties of individuals acting as eventive
roots is very restricted
Before discussing the percentages of property roots contributing individual properties and those contributing simple event properties, I would like to recall that
de-adjectival verb formation is much more restricted than de-nominal verb-formation. Although de-adjectival verb formation instantiates a productive pattern (in my
conviction), there are tight constraints which to my knowledge are not yet properly
understood.
While the (open) list of de-adjectival verbs is relatively small, the list of those that
permit alternates along the lines of ein Glas (mit Tequila) füllen vs. Tequila in ein
Glass füllen (recall (1)) is a good deal smaller still. According to my findings there is
only a handful of property roots occurring in “manner” verbs. This is my tentative
list, including German and English roots.
(29)
a. German √full (‘full’), as in (1); and German √leer, as in (i) in footnote 4;
English √empty as in to empty water out of a boat.
German √schliess (‘close, shut’) as in the resultative construction die
Tür aufschließen/zuschließen lit. ‘to lock a door open or shut’;15 but not
German √offen (‘open’);
b. German √sauber contributing simple event type (“manner”) in intransitive saubermachen16 (Compare den ganzen Tag saubermachen (‘clean
the whole day’); English (√clean (cf. Levin 2009, Levin and Rappaport
Hovav 2013)
c. English √clear (as in clear the dishes of the table); but not German √klar.
The negative evidence is as significant as the positive evidence: Although schließen
and öffnen are antonymous there is no resultative construction *auföffnen (‘to open
by unlocking’) and German √klar while obviously cognate to English √clear,
doesn’t appear to have mono-eventive alternations either. I am convinced that this
restriction is not accidental. My hypothesis is as follows:
Hypothesis 2. In order for a property root to undergo coercion to an event-type root it must
be (a) relational and (b) involve universal quantification in its non-referential argument.
As we have been assuming, √full is a binary relation full(y,z)—y is full of z.
Moreover, its semantics involves an element of universal quantification: all parts of
15
Resultative constructions indicate that “the verb enters syntax as atelic” (see Kratzer 2000) or is a
“manner” denoting predicate (see Levin 1999). According to the present account secondary predication
indicates mono-eventivity.
16
According to my analysis the root √sauber combines via direct merge with the light verb machen
(‘make’) in v to form saubermachen. There is also colloquial reinemachen from the property root √rein
(‘pure’). Reinemachen is obtained via the same construction, but I know of no other intransitive verbs with
adjectival roots of this pattern.
Semantico-syntactic structure in verbs
307
the denotation of the argument y are occupied by the stuff denoted by z; likewise
empty(y,z) involves “negative” universal quantification: for all parts of y there is no
stuff z (of the relevant kind) occupying that part. The root √schliess (‘closed, shut’)
is silently relational as well: All apertures if any are closed. The roots German
√sauber and English √clean involve the “removal of all dirt, debris, or other
unwanted material” (see Levin and Rappaport Hovav 2013, [boldface A.R.]). English
√clear is also relational (cf. clear of st.) and apparently universal—but as we have
noted German klar, is not.17 In Standard German the adjective klar predicates the
property of transparency or brightness; it describes visibility conditions and is not
relational. The property root √offen (‘open’), though relational, like √schliess,
predicates the existence of a gap z in a region y, and thus does not involve universal
but rather existential quantification. This, I believe, is why German √klar and
√open do not coerce to event properties to combine with v by direct merge.
Why is it the property roots which have a relational meaning involving universal
quantification that allow for the coercion that enables them to merge directly with v
in the way of manner roots? This probably has to do with the fact that monoeventive verbs generally lack culmination conditions. Change of relational properties
that involve universal quantification over the second argument of the relation will
typically be protracted, going on until the quantification domain has been exhausted.
Often such protracted processes take some prototypical form, and these forms can
then be incorporated into the root meaning so that a genuine “manner”-root meaning remains after that step.18
17
According to Grimm and Grimm (2007) there has been a North-German regional variant klären that
is like English clear with the meaning of removing dirt or debris, which licenses a von-phrase for what is
being removed in the same way that English clear licenses an of-phrase.
18
This tentative explanation is somewhat different from the one suggested in Levin and Rappaport
(2014). According to Levin and Rappaport Hovav’s account the decisive feature is change along a “change
scale”. These change scales differ from one verb (or verb root) to the next with de-adjectival “result verbs”
at one end of the spectrum and “manner” verbs, which denote change without any reference to scales, at
the other. So the transition from non-“manner” to “manner” is the loss of “the homomorphic association
between the scale of change lexicalized in the verb and the time course of event” (Levin and Rappaport,
2014:355). To account for the limited occurrence of such shifts Levin and Rappaport Hovav claim “manner
uses do not apparently develop whenever this link is lost. We assume that the frequency and ubiquity of
cleaning as a household task facilitated this process for the verb clean, setting it apart from other
deadjectival verbs” (Levin and Rappaport, 2014:355).
In the light of the suggested explanation we would assume that selected direct objects of saubermachen
(or clean as a manner verb) would be disjoint from those of causative säubern. Indeed such a disjunction
shows up with what the authors diagnose as anticausative uses of clean up (see Levin and Rappaport,
2014:345) as in the motorcycle frame cleaned up. This use goes with material objects, such as metal, guns,
carpets, mirrors, and glass, but not with kitchens, bathrooms, yards, which go with “manner”-readings of
clean. The particle clean-verbs apparently instantiate the same constructions as German aus-particle verbs
such as ausreifen (‘to fully ripen’) which I analyzed in Roßdeutscher (2012). In this chapter I show that deadjectival aus-verbs are achievements and that the particle interacts with the degree to which the property
denoted by the root applies to the theme. Clean something up must be analyzed as providing something
with the highest degree of cleanness. This predicate can probably be more naturally predicated of physical
objects than of places. But apart from these differences I cannot attest any disjunction between “manner”
vs. “result” clean, in the set of selected direct objects, at least not for German saubermachen vs. säubern.
308
Antje Roßdeutscher
13.5.2.1 The missing logically possible pairs Verbs built from property roots, I said,
seem to be comparatively rare, and much less frequent than verbs built from sortal
roots. But sometimes the distinction between property root based verbs and sortal
root based verbs is not easy to draw. Consider the verb befeuchten. It is tempting to
analyze the verb as derived, from the property root √feucht (‘humid’) (cf. (23)). The
fact that feucht is used as an adjective, consisting just of the root plus the
morphology of adjectives in German may be seen as supporting this analysis. But
how can we be sure? Couldn’t √feucht be just as well a sortal root, denoting
humidity (which by the way is the denotation of the German noun Feuchte)? In
that case the semantic analysis of befeuchten would amount to something like
“process which culminates in the state of humidity having been added to the
object” (i.e., “humidity is bei (at) the object”). That comes to the same thing as
when we analyze befeuchten as property root based and as describing processes that
culminate in the object being humid. Similar doubts can be raised in relation to verbs
like überhöhen (to increase by too much) and bestärken (to reinforce, to support).
Should we analyze überhöhen as a process that leads to the object being too high (the
property based analysis) or as a process that leads to the object having been provided
with too much height (the sortal root based analysis)? How can we tell? And the
same goes for bestärken (from √stark (‘strong’)); for all that I know at this point it
could be analyzed as describing processes that render the object strong (or stronger)
but also as describing processes that culminate in strength having been added to the
object. From a purely semantic point of view these are distinctions without a
difference and other evidence has to be brought into play if we are to be able to
choose between these analyzes in a non-arbitrary way.
13.5.3 Coercion of manner denoting roots to resultant states denotation
occurs often, but not normally
So far my understanding of root coercion is based primarily on a study of be-verbs
(cf. Roßdeutscher 2010). We have been looking at a list of two-hundred be-verbs for
which there are corresponding simplex verbs. Of the two-hundred pairs (consisting
each of a simplex verb and a corresponding be-verb, e.g., < bellen, bebellen>, < arbeiten, bearbeiten> about eighty of the be-verbs are on the same (mono-eventive) side
of the mono-eventive/bi-eventive divide as the corresponding simplex verb. In about
fifty cases the be-verb lies on the other side (i.e., it behaves like a bi-eventive verb).
On the account proposed in the present chapter these are all cases in which the
“manner”-root of the simplex verb is coerced into a sortal or property root. It should
be noted, however, that it isn’t always easy to tell whether the be-verb really is
bi-eventive. Informants often diverge in their judgements relating to these verbs.
In particular there is a fair amount of disagreement over their -ung-nominalisalibility. If we are right, then speakers who reject the corresponding -ung-noun as
non-existent or impossible analyze the be-verb as mono-eventive; those who accept
Semantico-syntactic structure in verbs
309
the -ung-noun must assign the be-verb a bi-eventive structure. In the terminology
introduced earlier in this chapter: those who analyze the be-verb as a mono-eventive
verb give an “expected” analysis, those that understand it as a bi-eventive verb assign
it an analysis that is “unexpected”.
13.6 Conclusion
I have presented a number of different patterns of verb-formation. The word
formation pattern of a given verb from a given root is unexpexted, I postulate, if
the formation operations are incompatible with the properties of the root. In such
cases the root is coerced into one with properties that are compatible with the
operations that define the pattern.
The account I have presented is based on the assumption that there are at least
four types of roots that can enter into the construction of verbs: (i.a) simple event
type, (i.b) relational event type, (ii.a) one-place individual property, (ii.b) two-place
relational (existential and quantificational) individual property, and (iii) sortal.
Furthermore the account presupposes that we have ways of telling—at least in
some instances—(a) what the construction pattern of a given verb is, and (b) what
the type of a given root is. The unexpected contributions are those which are
inconsistent with these independently motivated classifications.
As regards the reinterpretations that are needed to make the unexpected constructions possible the following points emerged:
1. Reinterpretation can be reconstructed as triggered by selection requirements of
roots or functional heads.
2. The semantics of the roots resulting from coercion can be traced back to the
original semantic contribution of the roots.
3. Unexpected occurrences are comparatively rare.
At one point I noted that there is an alternative to the root coercion account I have
explored in this chapter: rather than assuming that a root can be coerced so that it
fits an unexpected construction pattern, one could assume that there is another root,
which looks and sounds the same as the first (insofar as roots can look and sound the
same) and whose semantics is also related to the first root, but which nevertheless
bears the properties that make it fit the given “unexpected” pattern. This is a weaker
proposal than the one I have been pursuing here; but it, too, needs to say something
about the semantic relations between the two roots and for that the first of the three
points above seems as relevant as it is to the coercion story presented here.
It is points 2 and 3 that have persuaded me to go for the stronger hypothesis,
according to which one of the two roots can be obtained from the other via general
principles of structural and semantic coercion.
References
Abels, Klaus (2003). Successive Cyclicity, Anti-Locality, and Adposition Stranding. Ph.D.
dissertation. UConn.
Acedo-Matellán, Víctor (2010). Argument Structure and the Syntax–Morphology Interface: A
Case Study in Latin and other Languages. UB dissertation, Barcelona. Available at < http://
hdl.handle.net/10803/21788>.
Acedo-Matellán, Víctor and Mateu, Jaume (2008). “The path from satellite-framed Indoeuropean to verb-framed Romance: A lexical-syntactic account”, paper presented at X Diachronic Generative Syntax Conference, Cornell University, August 7–8.
Acedo-Matellán, Víctor and Mateu, Jaume (this volume). “From syntax to roots: A syntactic
approach to root interpretation”, 14–32.
Ackema, Peter and Neeleman, Ad (2004). Beyond Morphology. Oxford: Oxford University
Press.
Ackema, Peter and Schoorlemmer, Maaike (1995). “Middles and nonmovement”, Linguistic
Inquiry 26: 173–97.
Acquaviva, Paolo (2005). “I significati delle nominalizzazioni in -ATA e i loro correlati
morfologici”, in M. Grossmann and A. Thornton (eds.), La formazione delle parole.
Roma: Bulzoni, 7–29.
Acquaviva, Paolo (2008a). “Roots and lexicality in Distributed Morphology”, unpublished
manuscript. Dublin: University College Dublin. Available at < http://ling.auf.net/lingbuzz/
000654>.
Acquaviva, Paolo (2008b). Lexical Plurals. Oxford: Oxford University Press.
Acquaviva, Paolo (2009). “Roots and lexicality in Distributed Morphology”, York Papers in
Linguistics: 1–21.
Acquaviva, Paolo (this volume). “The roots of nominality, the nominality of roots”, 33–56.
Acquaviva, Paolo and Panagiotidis, Phoevos (2012). “Lexical decomposition meets conceptual
atomism”, Lingue e linguaggio XI/2: 165–80.
Åfarli, Tor A. (2007). “Do verbs have argument structure?”, in E. Reuland, T. Bhattacharya,
and G. Spathas (eds.), Argument Structure. Amsterdam: John Benjamins, 1–16.
Ahn, Byron (2012). The Distribution of Focal Stress and the Syntax of Reflexivity. Manuscript,
UCLA.
Aikhenvald, Alexandra (2003). Classifiers. Oxford: Oxford University Press.
Alexiadou, Artemis (2001). Functional Structure in Nominals: Nominalization and Ergativity.
Amsterdam: John Benjamins.
Alexiadou, Artemis (2009). “On the role of syntactic locality in morphological processes: The
case of Greek nominals”, in A. Giannakidou and M. Rathert (eds.), QP, Nominalizations
and the Role of DP. Oxford: Oxford University Press.
References
311
Alexiadou, Artemis (2010). “On the morpho-syntax of (anti)causative verbs”, in M. Rappaport
Hovav, E. Doron, and I. Sichel (eds.), Syntax, Lexical Semantics, and Event Structure.
Oxford: Oxford University Press, 177–203.
Alexiadou, Artemis (2011). “Plural mass nouns and the morpho-syntax of number”, WCCFL
28. Somerville: Cascadilla Proceedings Project, 33–41. Available at < http://www.lingref.
com>, document #2433.
Alexiadou, Artemis (this volume). “Roots in transitivity alternations: Afto/auto-reflexives”,
57–80.
Alexiadou, Artemis and Anagnostopoulou, Elena (1999). “Tests for unaccusativity in a language without tests for unaccusativity”, Proceedings of the 3rd international conference on
Greek Linguistics. Athens: Ellinika Grammata, 23–31.
Alexiadou, Artemis and Anagnostopoulou, Elena (2004). “Voice morphology in the causativeinchoative alternation: Evidence for a non-unified structural analysis of unaccusatives”, in
A. Alexiadou, E. Anagnostopoulou, and M. Everaert (eds.), The Unaccusativity Puzzle.
Oxford: Oxford University Press, 114–36.
Alexiadou, Artemis and Anagnostopoulou, Elena (2008). “Structuring participles”, Proceedings of WCCFL 26. Somerville: Cascadilla Proceeding Project, 33–41.
Alexiadou, Artemis and Anagnostopoulou, Elena (2011). “Decomposing the Greek verb: The
view from the clear-alternation”, paper presented at Approaches to the Lexicon (Roots II),
The Hebrew University of Jerusalem, June 13–16.
Alexiadou, Artemis and Anagnostopoulou, Elena (2013). “Manner vs. result complementarity
in verbal alternations: A view from the clear-alternation”, Proceedings of NELS 42: 39–52.
Alexiadou, Artemis and Doron, Edit (2012). “The syntactic construction of two non-active
voices: Passive and middle”, Journal of Linguistics 48(1): 1–34.
Alexiadou, Artemis and Grimshaw, Jane (2008). “Verbs, nouns and affixation”, in F. Schäfer
(ed.), SinSpeC (1): Working Papers of the SFB 732, University of Stuttgart, 1–16.
Alexiadou, Artemis and Iordachioaia, Gianina (2012). “Psych nominalizations with causers”,
Proceedings of NELS 42: 53–66.
Alexiadou, Artemis and Schäfer, Florian (2014). “Towards a non-uniform analysis of naturally
reflexive verbs” Proceedings of WCCFL 31: 1–10.
Alexiadou, Artemis, Anagnostopoulou, Elena, and Schäfer, Florian (2006). “The properties of
anticausatives crosslinguistically”, in M. Frascarelli (ed.) Phases of Interpretation. Berlin:
Mouton de Gruyter, 187–212.
Alexiadou, Artemis, Anagnostopoulou, Elena, and Schäfer, Florian (to appear). External
Arguments in Transitivity Alternations: A Layering Approach. Oxford: Oxford University
Press.
Alexiadou, Artemis, Haegeman, Liliane, and Stavrou, Melita (2007). Noun Phrase in the
Generative Perspective. Berlin: Mouton de Gruyter.
Allen, Margaret (1978). Morphological Investigations. Ph.D. dissertation, University of Connecticut.
Amritavalli, Raghavachari (2007). “Syntactic categories and lexical argument structure”, in
E. Reuland, T. Bhattacharya, and G. Spathas (eds.), Argument Structure. Amsterdam: John
Benjamins, 49–61.
Amritavalli, R. and Jayaseelan, K. A. (2003). “The genesis of syntactic categories and parametric variation”, paper presented at the 4th Asian GLOW, Seoul, August.
312
References
Anagnostopoulou, Elena (1999). “On experiencers”, in A. Alexiadou, G. Horrocks, and
M. Stavrou (eds.), Studies in Greek Syntax. Dordrecht: Kluwer, 67–93.
Anagnostopoulou, Elena (2003). “Participles and Voice”, in A. Alexiadou, M. Rathert, and
A. von Stechow (eds.), Perfect Explorations. Berlin: Mouton de Gruyter, 1–36.
Anagnostopoulou, Elena and Everaert, Martin (1999). “Towards a more complete typology of
anaphoric expressions”, Linguistic Inquiry 30(1): 97–118.
Anagnostopoulou, Elena and Somioti, Yota (2013). “Allosemy, idioms, and their domains:
Evidence from adjectival participles”, in R. Folli, C. Sevdali, and R. Truswell (eds.), Syntax
and its Limits. Oxford: Oxford University Press, 218–50.
Anagnostopoulou, Elena and Samioti, Yota (this volume). “Domains within words and their
meanings: A case study”, 81–111.
Αναστασιάδη-Συμεωνίδη, A. (1994). To τεμάχιο –τος στα ρηματικά επίθετα της Νεοελληεοελληνικής. Studies in Greek Linguistics 15: 473–484.
Anderson, Stephen (1982). “Where’s morphology?”, Linguistic Inquiry 13: 571–612.
Anderson, Stephen (1992). A-Morphous Morphology. Cambridge: Cambridge University Press.
Arad, Maya (2003). “Locality constraints on the interpretation of roots: The case of Hebrew
denominal verbs”, Natural Language & Linguistic Theory 21: 737–78.
Arad, Maya (2005). Roots and Patterns: Hebrew Morpho-Syntax. Dordrecht: Springer.
Arbaoui, Nora (2010). Les dix formes de l’arabe classique à l’interface phonologie/syntaxe: pour
une déconstruction du gabarit, Doctoral Dissertation, Université Paris Diderot.
Aronoff, Mark (1976). Word Formation in Generative Grammar. Cambridge: MIT Press.
Aronoff, Mark (1980). “Contextuals”, in T. Hoekstra and M. Moortgat (eds.), Lexical Grammar: Publications in Language Sciences. Dordrecht: Foris, 263–85.
Aronoff, Mark (1994). Morphology by Itself: Stems and Inflectional Classses. Cambridge: MIT
Press.
Arquiola, Elena (2003). “Morphology, argument structure and lexical semantics: The case of
Spanish auto- and co-prefixation to verbal bases”. Linguistics 41(3): 495–513.
Bach, Kent’ (2002). “Giorgione was so-called because of his name”, Philosophical Perspectives
16: 73–103.
Bachrach, Asaf and Wagner, Michael (2007). “Syntactically driven cyclicity vs. output-output
correspondence: The case of adjunction in diminutive morphology”, U. Penn Working
Papers in Linguistics 10: 1.
Baker, Mark (1985). “The Mirror Principle and Morphosyntactic Explanation”, Linguistic
Inquiry 36: 373–416.
Baker, Mark (1988). Incorporation: A Theory of Grammatical Function Changing. Chicago:
University of Chicago Press.
Baker, Mark (2003). Lexical Categories: Verbs, Nouns, and Adjectives. Cambridge: Cambridge
University Press.
Baker, Mark (2008). The Syntax of Control and Agreement, Cambridge: Cambridge University
Press.
Baker, Mark, Johnson, Kyle, and Roberts, Ian (1989). “Passive Arguments Raised”, Linguistic
Inquiry 20: 219–51.
Barker, Chris (2010). “Nominals don’t provide criteria of identity”, in A. Alexiadou and
M. Rathert (eds.), Nominalizations Across Languages and Frameworks. Berlin: Mouton de
Gruyter, 9–24.
References
313
Barker, Chris and Dowty, David (1993). “Non-verbal thematic proto-roles”, Proceedings of
NELS 23: 49–62.
Barner, David, and Bale, Alan (2002). “No nouns, no verbs: Psycholinguistic arguments in
favor of lexical underspecification”, Lingua 112: 771–791.
Barner, David, and Bale, Alan (2005). “No nouns, no verbs? A rejoinder to Panagiotidis”,
Lingua 115: 1169–1179.
Basilico, David (2008). “Particle verbs and benefactive double objects in English: High and low
attachments”, Natural Language and Linguistic Theory 26: 731–73.
Beard, Robert (1990). “The nature and origins of derivational polysemy”, Lingua 81(2): 101–40.
Beard, Robert (1995). Lexeme–Morpheme Base Morphology: A General Theory of Inflection and
Word Formation. New York: Suny Press.
Beavers, John (2008). “Scalar complexity and the structure of events,” in J. Dölling, and
T. Heyde-Zybatow (eds.), Event Structures in Linguistic Form and Interpretation. Berlin:
Mouton de Gruyter, 245–65.
Belletti, Adriana (1982). “Morphological passive and pro drop: The impersonal construction in
Italian”, Journal of Linguistic Research 2(4): 1–33.
Belletti, Adriana and Rizzi, Luigi (1988). “Psych-verbs and theta-theory”, Natural Language
and Linguistic Theory 6: 291–352.
Bertinetto, Pier Marco (1995). “Compositionality and non-compositionality in morphology”,
in W. V. Dressler, and C. Burani (eds.), Crossdisciplinary Approaches to Morphology.
Vienna: Österreichische Akademie der Wissenschaft, 9–36.
Bittner, Dagmar (2002). “Semantisches in der pronominalen Flexion des Deutschen”, Zeitschrift für Sprachwissenschaft 21: 196–233.
Bittner, Maria and Hale, Ken (1996). “The structural determination of case and agreement”,
Linguistic Inquiry 27: 1–68.
Blau, Joshua (1954). “Beynoni pa’ul behora’a ’aqtivit [Passive participles with active interpretation]”, Leshonenu 18: 67–81.
Bloom, Paul (1997). “Intentionality and Word Learning”, Trends in Cognitive Sciences 1: 9–12.
Bobaljik, Jonathan David (2008). “Where’s Phi? Agreement as a post-syntactic operation”, in
D. Harbour et al. (eds.), Phi-theory: Phi-features across Modules and Interfaces. Oxford:
Oxford University Press, 83–129.
Boeckx, Cedric (2009). “Aspects of a theory of phases”, Manuscript, ICREA/CLT-UAB.
Boeckx, Cedric (2010). “Defeating Lexicocentrism”, Manuscript, CLT/UAB. Available at
< http://ling.auf.net/lingBuzz/001130>.
Bolozky, Shmuel (1999). Measuring Productivity in Word Formation: The Case of Israeli
Hebrew (Studies in Semitic Languages and Linguistics 27). Leiden: Brill.
Borer, Hagit (1998). “Morphology and syntax”, in A. Spencer and A. Zwicky (eds.), The
Handbook of Morphology. Oxford: Blackwell, 151–90.
Borer, Hagit (2003). “Exo-skeletal vs. endo-skeletal explanations: Syntactic projections and the
lexicon”, in J. Moore and M. Polinsky (eds.), The Nature of Explanation in Linguistic
Theory. Stanford: CSLI Publications, 31–67.
Borer, Hagit (2005). Structuring Sense I. Oxford: Oxford University Press.
314
References
Borer, Hagit (2005). Structuring Sense II: The Normal Course of Events. Oxford: Oxford
University Press.
Borer, Hagit (2009). “Root bound”, seminar given at CASTL, University of Tromsø. October
2009.
Borer, Hagit (2013). Taking Form, Structuring Sense Vol. III. Oxford: Oxford University Press.
Borer, Hagit (this volume). “The category of roots”, 112–48.
Borer, Hagit and Grodzinsky, Yosef (1986). “Syntactic cliticization and lexical cliticization: The
case of Hebrew dative clitics”, in H. Borer (ed.), The Syntax of Pronominal Clitics: Syntax
and Semantics, 19. New York: Academic Press, 175–217.
Bosque, Ignacio (2001). “On the weight of light predicates”, in J. Herschenson et al. (eds.),
Features and Interfaces in Romance. Amsterdam: John Benjamins, 23–38.
Bosque, Ignacio and Gallego, Ángel (2012). “Against subextraction”, Manuscript, UCM /
CLT-UAB.
Brame, Michael (1983). “Bound anaphora is not a relation between NPs: Evidence for local
word grammar (without trees)”, Linguistic Analysis 11: 139–66.
Brisson, Christine (1994). “The licensing of unexpressed objects in English verbs”, Chicago
Linguistic Society 30: 90–102.
Bruening, Benjamin (2004). “Verbal reciprocals”, in T. Solstad, B. Lyngfelt, and M. F. Krave
(eds.), Demoting the Agent: Proceedings of the Oslo Workshop on Passive and Other VoiceRelated Phenomena. Oslo: University of Oslo, 3–12.
Bruening, Benjamin (2012). “By-phrases in passives and pominals”, Syntax 16: 1–42.
Bruening, Benjamin (2014). “Word Formation is Syntactic: Adjectival Passives in English”,
Natural Language and Linguistic Theory 32: 363–422.
Carey, Susan and Xu, Fei (1999).”Sortals and kinds”, in R. Jackendoff, P. Bloom, and K. Wynn
(eds.), Language, Logic, and Concepts. Cambridge: MIT Press.
Carlson, Gregory (1977). Reference to Kinds in English. Ph.D. dissertation, University of
Massachusetts at Amherst.
Carlson, Gregory (1998). “Thematic roles and the individuation of events”, in S. Rothstein
(ed.), Events and Grammar. Dordrecht: Kluwer, 35–51.
Castella, Marta (2010). The Semantics and Syntax of Auto. MA thesis, Utrecht University.
Chierchia, Gennaro (1984). Topics in the Syntax and Semantics of Infinitives and Gerunds.
Ph.D. dissertation, University of Massachusetts at Amherst.
Chierchia, Gennaro (1998). “Reference to kinds across languages”, Natural Language Semantics 6: 339–405.
Chomsky, Noam (1955). The Logical Structure of Linguistic Theory. Mimeograph, Harvard
University. Published in 1975, New York: Plenum.
Chomsky, Noam (1965). Aspects of the Theory of Syntax. Cambridge: MIT Press.
Chomsky, Noam (1970). “Remarks on nominalizations”, in R. Jacobs and P. Rosenbaum
(eds.), Readings in Transformational Grammar. Waltham: Ginn & Co., 184–221.
Chomsky, Noam (1995). “Categories and transformations”, in N. Chomsky, The Minimalist
Program. Cambridge: MIT Press, 219–394.
References
315
Chomsky, Noam (2000). “Minimalist inquires: The framework”, in R. Martin, D. Michaels,
and J. Uriagereka (eds.), Step by Step: Essays on Minimalist Syntax in Honor of Howard
Lasnik. Cambridge: MIT Press, 89–151.
Chomsky, Noam (2001). “Derivation by phase”, in M. Kenstowicz (ed.), Ken Hale: A Life in
Language. Cambridge, MA: MIT Press, 1–52.
Chomsky, Noam (2004). “Beyond explanatory adequacy”, in A. Belletti (ed.), Structures and
Beyond: The Cartography of Syntactic Structures (vol. 3). Oxford: Oxford University Press,
104–31.
Chomsky, Noam (2005). “Three factors in language design”, Linguistic Inquiry 36: 1–22.
Chomsky, Noam (2007). “Approaching UG from below”, in U. Sauerland and H.-M. Gärtner
(eds.), Interfaces + Recursion = Language? Chomsky’s Minimalism and the View from
Syntax–Semantics. Berlin: Mouton de Gruyter, 1–30.
Chomsky, Noam (2008). “On phases”, in C. Otero et al. (eds.), Foundational Issues in
Linguistic Theory: Essays in Honor of Jean-Roger Vergnaud. Cambridge: MIT Press, 134–66.
Chomsky, Noam (2012). “Problems of Projection”, Manuscript, MIT Press. To appear in Lingua.
Chomsky, Noam and Halle, Morris (1968). The Sound Pattern of English. New York, Harper
and Row.
Clark, Eve and Clark, Herbert (1979). “When nouns surface as verbs”, Language 55: 767–811.
Conti, Sara (2007). Compound Adjectives in English. Ph.D. dissertation, Università di Pisa and
University of Vienna.
Contini-Morava, Ellen and Tobin, Yishai (2000). Between Grammar and the Lexicon. Amsterdam: John Benjamins.
Cusic, David (1981). Verbal Plurality and Aspect. Ph.D. dissertation, University of Stanford.
Cruse, D. Alan (1980). “Antonyms and gradable complementaries”, in D. Kastovsky (ed.),
Perspektiven der lexikalischen Semantik. Bonn: Bouvier Verlag.
De Belder, Marijke (2008). “Size matters: Towards a syntactic decomposition of countability”,
Proceedings of the 27th West Coast Conference on Formal Linguistics. Somerville: Cascadilla
Proceedings Project, 116–122.
De Belder, Marijke (2011a). “A morphosyntactic decomposition of countability in Germanic”,
The Journal of Comparative Germanic Linguistics 14(3): 173–202.
De Belder, Marijke (2011b). Roots and Affixes: Eliminating Lexical Categories from Syntax.
Utrecht: LOT.
De Belder, Marijke and van Craenenbroeck, Jeroen (2011). “How to merge a root”, unpublished
manuscript. CRISSP/HUB/UiL-OTS/Utrecht and CRISSP/HUB/FUSL/KUL. Available at
< http://ling.auf.net/lingBuzz/001226>.
De Belder, Marijke, Faust, Noam, and Lampitelli, Nicola (this volume). “On a low and a high
diminutive: Evidence from Italian and Hebrew”, 149–63.
Déchaine, Rose-Marie, Girard, Raphäel, Mudzingwa, Calisto, and Wiltschko, Martina (2012).
“The internal syntax of Shona Class prefixes”, paper presented at the 15th International
Morphology Meeting, Vienna, February 2012.
Dixon, Robert M. W. (1982). Where Have All the Adjectives Gone? And Other Essays In
Semantics and Philosophy 12. Berlin: Mouton, 39–94.
Doron, Edit (1983). Verbless Predicates in Hebrew. Ph.D. dissertation, The University of Texas
at Austin.
316
References
Doron, Edit (1986). “The pronominal copula as agreement clitic”, in H. Borer (ed.), Syntax and
Semantics 19: The Syntax of Pronominal Clitics. New York: Academic Press, 313–32.
Doron, Edit (1999). “habeynoni hasavil [The Passive Participle]”, Balshanut Ivrit 47: 39–62.
Doron, Edit (2003). “Agency and voice: The semantics of the Semitic templates”, Natural
Language Semantics 11: 1–67.
Doron, Edit (2008). “trumato šel habinyan lemašma’ut hapo’al [The contribution of the
template to verb meaning]”, in G. Hatav (ed.), Modern Linguistics of Hebrew. Jerusalem:
Magnes Press, 57–88.
Doron, Edit (2011). “Agents and causers”, paper presented at Roots III Workshop. Jerusalem,
June 2011.
Doron, Edit (this volume). “The interaction of adjectival passive and Voice”, 164–91.
Doron, Edit and Rappaport Hovav, Malka (2009). “A unified approach to reflexivization in
Semitic and Romance”, Brill’s Annual of Afroasiatic Languages and Linguistics 1: 75–105.
Dowty, David R. (1979). Word Meaning and Montague Grammar: The Semantics of Verbs and
Times in Generative Semantics and in Montague’s PTQ. Dordrecht: Reidel.
Dowty, David (1989). “On the semantic content of the notion of ‘thematic role’ ”, in
G. Chierchia, B. H. Partee, and R. Turner (eds.), Properties, Types, and Meaning II. Kluwer:
Dordrecht, 69–129.
Dowty, David (1991). “Thematic proto-roles and argument selection”, Language 67: 547–619.
Dubinsky, Stanley and Simango, Silvester Ron (1996). “Passive and stative in Chichewa:
Evidence for modular distinctions in grammar”, Language 72: 749–81.
Dupré, John (1999). “Are whales fish?”, in D. Medin and S. Atran (eds.), Folkbiology. Cambridge: MIT Press.
Dugas, André (1992). “Le préfixe auto-”, Langue Française 96: 20–29.
Edmondson, J. A. and Plank, F. 1978. “Great expectations: An intensive self analysis”,
Linguistics and Philosophy 2: 373–413.
Elbourne, Paul (2008). “Ellipsis sites as definite descriptions”, Linguistic Inquiry 39: 191–220.
Embick, David (1998). “Voice systems and the syntax/morphology interface”, in H. Harley
(ed.), Vol. 32 of MIT Working Papers in Linguistics: Papers from the UPenn/MIT Roundtable
on Argument Structure and Aspect. Cambridge: MIT Press, 41–72.
Embick, David (2000). “Features, syntax, and categories in the Latin perfect”, Linguistic
Inquiry 31: 185–230.
Embick, David (2003). “Locality, listedness, and morphological information”, Studia Linguistica 57: 143–169.
Embick, David (2004a). “On the structure of resultative participles in English”, Linguistic
Inquiry 35: 355–92.
Embick, David (2004b). “Unaccusative Syntax and Verbal Alternations”, in A. Alexiadou,
E. Anagnostopoulou, and M. Everaert (eds.), The Unaccusativity Puzzle: Explorations of the
Syntax–Lexicon Interface. Oxford University Press, Oxford, 137–58.
Embick, David (2010). Localism versus Globalism in Morphology and Phonology. Cambridge:
MIT Press.
Embick, David and Halle, Morris (2005). “On the status of stems in Morphological Theory”, in
T. Geerts, and H. Jacobs (eds.), Proceedings of Going Romance 2003. Amsterdam: John
Benjamins, 37–62.
References
317
Embick, David and Marantz, Alec (2008). “Architecture and blocking”, Linguistic Inquiry
39(1): 1–53.
Embick, David and Noyer, Rolf (2007). “Distributed Morphology and the Syntax/Morphology
Interface”, in G. Ramchand and C. Reiss (eds.), The Oxford Handbook of Linguistic
Interfaces. Oxford: Oxford University Press, 289–324.
Emonds, Joseph (1985). A Unified Theory of Syntactic Categories. Dordrecht: Foris.
Fabb, Nigel (1988). “English suffixation is constrained only by selectional restrictions”, Natural
Language and Linguistic Theory 6: 527–39.
Fabricius-Hansen, Cathrine (1975). Transformative, Intransformative und kursive Verben.
Tübingen: Niemeyer.
Farkas, Donka and Henriette de Swart (2003). The Semantics of Incorporation. Stanford: CSLI.
Fathi, Radwa (2013). Vocalic Length in Egyption Arabic: A New View. Ph.D. dissertation,
Université Paris Diderot-Paris 7.
Faust, Noam (2011). Forme et fonction dans la morphologie nominale de l’hébreu moderne:
Études en morpho-syntaxe. Ph.D. dissertation, University Paris Diderot-Paris 7.
Faust, Noam and Lampitelli, Nicola (2012). “How vowels point to syntactic structure: Roots
and skeletons in Hebrew and Italian”, in S. Blaho et al. (eds.) ConSOLE XVII Proceedings:
121–35.
Fodor, Jerry (1998). Concepts: Where Cognitive Science Went Wrong. Oxford: Oxford
University Press.
Folli, Raffella and Harley, Heidi (2005). “Consuming results in Italian and English: Flavors
of v”, in P. Kempchinsky, and R. Slabakova (eds.), Aspectual Inquiries. Dordrecht: Springer,
95–120.
Frege, Gottlob (1884). Die Grundlagen der Arithmetik. Eine logisch-matematische Untersuchung über den Begriff der Zahl. Breslau: Wilhelm Köbner.
Gallego, Ángel J. (2010). Phase Theory. Amsterdam: John Benjamins.
Gallego, Ángel J. (2012). Phases: Developing the Framework. Berlin: Mouton de Gruyter.
Gallego, Ángel J. (this volume). “Roots and phases”, 192–208.
Gast, Volker and Siemund, Peter (2006). “Rethinking the relationship between SELF-intensifiers and reflexives”, Linguistics 44(2): 343–81.
Gawron, Jean M. (2007). “Unifying degree-based and mereological accounts of gradual
change,” talk given at CASTL Workshop on the Syntax and Semantics of Measurability,
Tromsø.
Geach, Peter (1962). Reference and Generality. Ithaca: Cornell University Press.
Geissman, John. “Writing hints from John Geissman”. Available at < http://www.geosociety.
org/grants/04gw_wkshp/04-Geissman_hints.pdf>.
Gelman, Susan (2004). “Psychological essentialism in children”, Trends in Cognitive Sciences
9: 404–09.
Geniušiene_, Emma (1987). The Typology of Reflexives. Berlin: Mouton.
Georgala, Effi (2001). “The translational correspondence between the Modern Greek formations ending in -tos and -menos and their equivalent forms in German”. Manuscript,
Institute for Natural Language Processing of the University of Stuttgart.
318
References
Giannakidou, Anastasia and Merchant, Jason (1999). “Why Giannis can’t scrub his plate clean:
On the absence of resultative secondary predication in Greek”, in A. Mozer (ed.), Greek
Linguistics ’97: Proceedings of the 3rd International Conference on Greek Linguistics. Athens:
Ellinika Grammata, 93–103.
Giegerich, Heinz (1999). Lexical Strata in English. Cambridge: Cambridge University Press.
Goldberg, Adele E. (1995). Constructions: A Construction Grammar Approach to Argument
Structure. Chicago: University of Chicago Press.
Greenberg, Yael (2010). “Event internal pluractionality in Modern Hebrew: A semantic
analysis of one verbal reduplication pattern”, to appear in Brill’s Annual of Afroassiatic
Languages and Linguistics 2.
Griffin, Nicholas (1977). Relative Identity. Oxford: Clarendon Press.
Grimm, Jacob and Wilhelm Grimm, Wilhelm (2007), S. Hirzel (ed), Deutsches Wörterbuch.
Kompetenzzentrum für elektronische Erschließungs- und Publikationsverfahren in den
Geisteswissenschaften an der Universität Trier.
Grimshaw, Jane (1982). “On the lexical representation of Romance reflexive clitics”, in
J. Bresnan (ed.), The Mental Representation of Grammatical Relations. Cambridge: MIT
Press, 87–148.
Grimshaw, Jane (1990). Argument Structure. Cambridge: MIT Press.
Grimshaw, Jane (2005). Words and Structure. Stanford: CSLI Publications.
Guimarães, Maximiliano (2000). “In defense of vacuous projections in bare phrase structure”,
UMD WPL 9: 90–115.
Gupta, Anil (1980). The Logic of Common Nouns. New Haven: Yale University Press.
Guyton de Morveau, Louis-Bernard, Lavoisier, Antoine, Berthollet, Claude-Louis, and de
Fourcroy, Antoine Francois (1787). Méthode de nomenclature chimique. Paris: Cuchet.
Hale, Kenneth and Keyser, Samuel J. (1993). “On argument structure and the lexical expression of syntactic relations”, in K. L. Hale and S. J. Keyser (eds.), The View from Building 20:
Essays in Honor of Sylvain Bromberger. Cambridge: MIT Press, 59–109.
Hale, Kenneth and Keyser, Samuel J. (1998). “The basic elements of argument structure”, in
H. Harley (ed.), MIT Working Papers in Linguistics 32. Cambridge: MIT Press, 73–118.
Hale, Kenneth and Keyser, Samuel J. (2002). Prolegomenon to a Theory of Argument Structure.
Cambridge: MIT Press.
Hale, Kenneth and Keyser, Samuel J. (2005). “Aspect and the syntax of argument structure”,
in N. Erteschik-Shir and T. Rapoport (eds.), The Syntax of Aspect. Oxford: Oxford University Press, 11–41.
Halle, Morris (1973). “Stress rules in English: A new version”, Linguistic Inquiry: 451–64.
Halle, Morris and Marantz, Alec (1993). “Distributed Morphology and the pieces of inflection”, in Kenneth L. Hale and Samuel J. Keyser (eds.), The View from Building 20: Essays in
Linguistics in Honor of Sylvain Bromberger. Cambridge: MIT Press, 111–76.
Harley, Heidi (1995). Subjects, Events, and Licensing. Ph.D. dissertation, MIT.
Harley, Heidi (2004). “Merge, conflation, and head movement: The first sister principle
revisited”, in K. Moulton and M. Wolf (eds.), Proceedings of NELS 34: 239–54.
Harley, Heidi (2005). “How do verbs get their names? Denominal verbs, manner incorporation and the ontology of verb roots in English”, in N. Erteschik-Shir, and T. Rapoport (eds.),
The Syntax of Aspect. Oxford: Oxford University Press, 42–64.
References
319
Harley, Heidi (2007). “The bipartite structure of verbs cross-linguistically, or why Mary can’t
exhibit John her paintings”, paper presented at the 2007 ABRALIN Congress, Belo Horizonte, Brazil. Available at < http://ling.auf.net/lingBuzz/000435>.
Harley, Heidi (2009a). “Compounding in distributed morphology”, in R. Lieber and
P. Stekauer (eds.), The Oxford Handbook of Compounding. Oxford: Oxford University
Press, 129–44.
Harley, Heidi (2009b). “The morphology of nominalizations and the syntax of vP”, in
A. Giannakidou and M. Rathert (eds.), Quantification, Definiteness, and Nominalization.
Oxford: Oxford University Press.
Harley, Heidi (2009c). Roots and locality. Handout from Roots Workshop, University of
Stuttgart, June 2009.
Harley, Heidi (2012). “Lexical decomposition in modern generative grammar”, in W. Hinzen,
M. Werning, and E. Machery (eds.), Handbook of Compositionality. Oxford: Oxford University Press.
Harley, Heidi and Haugen, Jason D. (2007). “Are there really two different classes of instrumental denominal verbs in English?”, Snippets 16: 9–10.
Harley, Heidi, and Noyer, Rolf (1999). “State-of-the-article”, GLOT 4(4): 3–9.
Harley, Heidi and Noyer, Rolf (2000). “Licensing in the non-lexicalist lexicon”, in B. Peeters
(ed.), The Lexicon/Encyclopaedia Interface. Amsterdam: Elsevier, 349–74.
Harrell, Richard (1962). A Short Reference Grammar of Moroccan Arabic. Washington:
Georgetown University Press.
Haugen, Jason D. (2009). “Hyponymous objects and late insertion”, Lingua 119: 242–62.
Hay, Jennifer, Kennedy, Christopher, and Levin, Beth (1999). “Scalar structure underlies
telicity in ‘degree achievements’ ”, SALT 9.
Heim, Irene, and Kratzer, Angelika (1998). Semantics in generative grammar. Malden, MA:
Blackwell.
Hinrichs, Erhart (1985). A Compositional Semantics for Aktionsarten and NP Reference in
English. Ph.D. dissertation, Ohio State University.
Hoekstra, Teun (1988). “Small clause results”, Lingua 74: 101–39.
Hole, Daniel (2002). “Agentive selbst in German”, in G. Katz, S. Reinhard, and P. Reuter
(eds.), Sinn und Bedeutung 6. Proceedings of the Sixth Meeting of the Gesellschaft für
Semantik, Osnabrück.
Horn, Laurence (1989). A Natural History of Negation. Chicago: University of Chicago Press.
Horvath, Julia and Siloni, Tali (2008). “Active lexicon: Adjectival and verbal passives”, in
S. Armon-Lotem, G. Danon, and S. Rothstein (eds.), Current Issues in Hebrew Linguistics.
Amsterdam: John Benjamins, 105–34.
Howell, Jonathan. 2012. Melody and Prosody. Ph.D. dissertation, Cornell University.
Iatridou, Sabine (1988). “Clitics, anaphors, and a problem of co-indexation”, Linguistic Inquiry
19: 698–703.
Jackendoff, Ray (1983). Semantics and Cognition, Cambridge: MIT Press.
Jackendoff, Ray, Bloom, Paul, and Wynn, Karen (1999). Language, Logic and Concepts: Essays
in Memory of John McNamara. Cambridge: MIT Press.
Johannes, Kristen (2011). “The role of scales in directional PPs and motion events”, paper
presented at the Workshop on Scalarity in Verb-Based Constructions, Düsseldorf, Germany.
320
References
Kaisse, Ellen (2005). “Word formation and phonology”, in P. Štokauer, and R. Lieber (eds.),
Handbook of Word-Formation. Dordrecht: Springer, 25–47.
Katz, Jerrold and David Pitt (2000). “Compositional idioms”, Language 76: 409–32.
Kaye, Jonathan D. (1995). “Derivations and interfaces”, in J. Durand and F. Katamba (eds.),
Frontiers of Phonology. London: Longman, 289–332.
Kayne, Richard (1984). Connectedness and Binary Branching. Dordrecht: Foris.
Kayne, Richard (1994). The Antisymmetry of Syntax. Cambridge: MIT Press.
Kayne, Richard (2005). Movement and Silence. Oxford: Oxford University Press.
Kayne, Richard (2009). “Antisymmetry and the Lexicon”, Linguistic Variation Yearbook 8(1):
1–31.
Kayne, Richard (2011). “Antisymmetry and the Lexicon”, in A. M. Di Sciullo and C. Boeckx
(eds.), The Biolinguistic Enterprise. Oxford: Oxford University Press, 329–53.
Kearns, Kate (2007). “Telic senses and deadjectival verbs,” Lingua 117: 26–66.
Kemmer, Suzanne (1993). The Middle Voice. Amsterdam. John Benjamins.
Kennedy, Christopher (2007). “Vagueness and grammar: The semantics of relative and
absolute gradable adjectives”, Linguistics and Philosophy 30(1): 1–45.
Kennedy, Christopher and Levin, Beth (2008). “Measure of change: The adjectival core of
degree achievements,” in L. McNally and C. Kennedy (eds.), Adjectives and Adverbs: Syntax,
Semantics and Discourse. Oxford: Oxford University Press, 156–82.
Kennedy, Christopher and McNally, Louise (2005). “Scale structure and the semantic typology
of gradable predicates”, Language 81(2): 345–81.
Kenstowicz, Michael and Kisseberth, Charles (1977). Topics in Phonological Theory. New York:
Academic Press.
Kihm, Alain (2005). “Noun class, gender, and the lexicon/syntax/morphology interfaces: A
comparative study of Niger-Congo and Romance languages”, in G. Cinque and R. Kayne
(eds.), The Oxford Handbook of Comparative Syntax. Oxford: Oxford University Press,
459–512.
Kiparsky, Paul (1982a). “From Cyclic Phonology to Lexical Phonology”, in H. van der Hulst
and N. Smith (eds.), The Structure of Phonological Representations. Dordrecht: Foris, 131–75.
Kiparsky, Paul (1982b). “Lexical Morphology and Phonology”, in Linguistic Society of Korea
(ed.), Linguistics in the Morning Calm: Selected Essays from SICOL-1981. Seoul: Hanshin.
Kiparsky, Paul (1982c). “Word Formation and the Lexicon”, in Fred Ingeman (ed.), Proceedings of the Mid-America Linguistics Conference. University of Kansas, pp. 3–29.
Kiparsky, Paul (1997). “Remarks on denominal verbs”, in A. Alsina, J. Bresnan, and P. Sells
(eds.), Argument Structure. Stanford: CLSI, 473–99.
Kiparsky, Paul (1998). “Partitive case and aspect”, in M. Butt and W. Geuder (eds.), The
Projection of Arguments. Stanford: CSLI, 265–307.
Klaiman, Miriam H. (1991). Grammatical Voice. Cambridge: Cambridge University Press.
König, Ekkehard (2011). “Reflexive nominal compounds”, Studies in Language 35: 112–27.
König, Ekkehard, and Vezzosi, Letizia (2004). “The role of predicate meaning in the development of reflexivity”, in B. Wiemer, W. Bisang, and N. Himmelmann (eds.), What Makes
Grammaticalization? A Look from its Fringes and its Components. Berlin: Mouton de
Gruyter: 213–44.
Koontz-Garboden, Andrew (2010). “The lexical semantics of derived statives”, Linguistics and
Philosophy 33(4): 285–324.
References
321
Koontz-Garboden, Andrew and Beavers, John (2010). “Manner and result in verbal meaning”,
unpublished manuscript, University of Manchester and The University of Texas at Austin.
Kordoni, Valia (2002). “Participle-adjective formation in Modern Greek”, LFG Meeting, July
2002. Athens, Greece.
Kratzer, Angelika (1994). The Event Argument and the Semantics of Voice, unpublished
manuscript, UMass.
Kratzer, Angelika (1996). “Severing the external argument from its verb”, in J. Rooryck and
L. Zaring (eds.), Phrase Structure and the Lexicon. Dordrecht: Kluwer, 109–37.
Kratzer, Angelika (2000). “Building statives”, in L. J. Conathan (ed.), Proceedings of the
Berkeley Linguistics Society 26: 385–99.
Kratzer, Angelika (2001). “Building statives”, Berkeley Linguistic Society, 26.
Kratzer, Angelika (2005). “Building resultatives”, in C. Maienborn and A. Wöllstein-Leisten
(eds.), Events in Syntax, Semantics, and Discourse. Tübingen: Niemeyer, 177–212.
Krifka, Manfred (1989). “Nominal reference, temporal constitution and quantification in event
semantics,” in R. Bartsch, J. van Benthem, and P. van Emde Boas (eds.), Semantics and
Contextual Expressions. Dordrecht: Foris, 75–115.
Krifka, Manfred (1995). “Common nouns: A contrastive analysis of Chinese and English”, in
G. Carlson, and F. Pelletier (eds.), The Generic Book. Chicago: University of Chicago Press,
398–411.
Krifka, Manfred (1998). “The origins of telicity,” in S. Rothstein (ed.), Events and Grammar.
Dordrecht: Kluwer, 197–235.
Kripke, Saul (1972). “Naming and necessity”, in D. Davidson and G. Harman (eds.), Semantics
of Natural Languages. Dordrecht: Reidel, 253–355.
Labelle, Marie (2008). “The French reflexive and reciprocal se”, Natural Language and
Linguistic Theory 26: 833–76.
Lakoff, George (1987). Women, Fire, and Dangerous Things. Chicago: University of Chicago
Press.
Lampitelli, Nicola (2010). “Nounness, gender, class and syntactic structures in Italian nouns”,
in R. Bok-Bennema, B. Kampers-Manhe, and B. Hollebrandse (eds.), Romance Languages
and Linguistic Theory 2008: Selected Papers from “Going Romance” Groningen 2008.
Amsterdam: John Benjamins, 195–214.
Lampitelli, Nicola (2011). Forme phonologique, exposants morphologiques et structures nominales: Étude comparée de l’italien, du bosnien et du somali. Doctoral Dissertation, Université
Paris Diderot.
Landau, Idan (2010). The Locative Syntax of Experiencers. Cambridge: MIT Press.
Lascaratou, Chryssoula (1991). “How ‘adjectival’ are adjectival passive participles in Modern
Greek and English?”, Glossologia 7–8: 87–97.
Lascaratou, Chryssoula and Philippaki-Warburton, Irene (1984). “Lexical versus transformational passives in Modern Greek”, Glossologia 2–3: 99–109.
Lekakou, Maria (2005). In the Middle Somewhat Elevated: The Semantics of Middles and its
Crosslinguistic Realization. Ph.D dissertation, UCL.
Levin, Beth (1999). “Objecthood: An event structure perspective”, Proceedings of CLS 35,
Volume 1: The Main Session. Chicago Linguistic Society, University of Chicago, 223–47.
Levin, Beth and Rappaport, Malka (1986). “The formation of adjectival passives”, Linguistic
Inquiry 17: 623–61.
322
References
Levin, Beth and Rappaport Hovav, Malka (1995). Unaccusativity: At the Syntax-Lexical
Semantics Interface. Cambridge: MIT Press.
Levin, Beth and Rappaport Hovav, Malka (2008). “A constraint on verb meanings: Manner/
result complementarity”, Cognitive Science Department Colloquium Series, Brown University, Providence, RI.
Levin, Beth and Rappaport Hovav, Malka (2009). “Constraints on the complexity of verb
meaning and VP structure”, Snippets 20.
Levin, Beth and Rappaport Hovav, Malka (2013). “Lexicalized meaning and manner/result
complementarity”, in Boban Arsenijević, Berit Gehrke, and Rafael Marín (eds.), Subatomic
Semantics of Event Predicates. Dordrecht: Springer, 49–70.
Levin, Beth and Rappaport Hovav, Malka (2014). “Manner and result: A view form clean”, in
R. Pensalfini, M. Turpin, and D. Guillemin (eds.), Language Description Informed by
Theory. Amsterdam: John Benjamins, 337–57.
Levinson, Lisa (2007a). Finding arguments for pseudo-resultative predicates. In Proceedings of
the 30th Annual Penn Linguistics Colloquium, ed. Tatjana Scheffler, Joshua Tauberer,
Aviad Eilam, and Laia Mayol, 155–168. Philadelphia: University of Pennsylvania Working
Papers in Linguistics.
Levinson, Lisa (2007b). The Roots of Verbs. Ph.D. dissertation, NYU.
Levinson, Lisa (2010). “Arguments for pseudo-resultative predicates”. Natural Language and
Linguistic Theory 28: 135–182.
Levinson, Lisa (this volume). “The ontology of roots and verbs”, 208–29.
Levinson, Stephen C. (2001). “Space: Linguistic expression”, in N. J. Smelser, and P. Baltes
(eds.), International Encyclopedia of Social and Behavioral Sciences: Vol. 22. Oxford: Pergamon, 14749–52.
Lieber, Rochelle (1980). The Organization of the Lexicon. Ph.D. dissertation, MIT.
Lieber, Rochelle (2004). Morphology and Lexical Semantics. Cambridge: Cambridge University
Press.
Longobardi, Giuseppe (1994). “Reference and proper names”, Linguistic Inquiry 35: 609–65.
Lowenstamm, Jean (2011). “The phonological pattern of phi-features in the perfective paradigm of Moroccan Arabic”, Brill’s Annual of Afroasiatic Languages and Linguistics 3:
140–201.
Lowenstamm, Jean (this volume). “Derivational affixes as roots: Phasal Spellout meets English
Stress Shift”, 230–58.
Manney, Linda (2000). Middle Voice in Modern Greek: Meaning and Function of an Inflectional Category. John Benjamins, Amsterdam.
Marantz, Alec (1984). On the Nature of Grammatical Relations. Cambridge: MIT Press.
Marantz, Alec (1995). “Cat as phrasal idiom”, unpublished manuscript, MIT.
Marantz, Alec (1997). “No escape from syntax: Don’t try morphological analysis in the privacy
of your own lexicon”, in A. Dimitriadis et al. (eds.), University of Pennsylvania Working
Papers in Linguistics, vol. 4.2. Philadelphia: University of Philadelphia, 201–25.
Marantz, Alec (2000). “Roots: The universality of root and pattern morphology”, paper
presented at the Conference on Afro-Asiatic Languages, Université de Paris VII.
Marantz, Alec (2001). Words and Things. Manuscript, MIT and NYU.
Marantz, Alec (2003). Subject and Objects. Manuscript, MIT.
References
323
Marantz, Alec (2005). “Objects out of the Lexicon: Objects as events”, unpublished manuscript, MIT.
Marantz, Alec (2007a). “Phases and words”, in S. H. Choe et al. (eds.), Phases in the Theory of
Grammar. Seoul: Dong In Publisher, 191–222.
Marantz, Alec (2007b). “Restitutive re- and the first phase syntax/semantics of the VP”, paper
presented at the University of Maryland, College Park, April 20.
Marantz, Alec (2009). “Roots, re- and affected agents: Can roots pull the agent under little v?”,
paper presented at the workshop Roots: Word Formation from the Perspective of “Core
Lexical Elements”, Stuttgart, June 10–12.
Marantz, Alec (2014). “Locality domains for contextual allomorphy across the interfaces”, in
O. Motoshansky and A. Marantz (eds.), Distributed Morphology Today: Morphemes for
Morris Halle. Cambridge MA: MIT Press, 95–115.
Marchand, Hans (1960). The Categories and Types of Present-day English Word-formation.
München: C. H. Beck’sche Verlagsbuchhandlung.
Marelj, Marijana and Reuland, Eric (2012). “Deriving reflexives: Deriving the lexicon-syntax
parameter”. Manuscript, OTS, Utrecht.
Markantonatou, Stella (1992). The Syntax of Modern Greek Noun Phrases with a Derived
Nominal Head. Ph.D. dissertation, University of Essex.
Markantonatou, S., Kaliakostas, A., Bouboureka, V., Kordoni, V., and Stavrakaki, V. (1996).
Μία (λεξική) σημασιολογική περιγραφή των ρηματικών επιθέτων σε –τóς. Studies in Greek
Linguistics 17: 187–201.
Marvin, Tatjana (2003). Topics in the Stress and Syntax of Words. Ph.D. dissertation, MIT.
Mateu, Jaume (2002). Argument Structure: Relational Construal at the Syntax–Semantics Interface. UAB dissertation. Bellaterra. Available at < http://www.tdx.cat/handle/10803/4828>.
Mateu, Jaume (2003). “Parametrizing verbal polysemy: The case of bake revisited”, Chicago
Linguistics Society 39: 256–69.
Mateu, Jaume (2008). “On the lexical syntax of directionality/resultativity: The case of
Germanic preverbs”, in A. Asbury et al. (eds.), Syntax and Semantics of Spatial P. Amsterdam: John Benjamins, 221–50.
Mateu, Jaume (2012). “Conflation and incorporation processes in resultative constructions”, in
V. Demonte, and L. McNally (eds.), Telicity, Change, and State: A Cross-Categorial View of
Event Structure, Oxford Studies in Theoretical Linguistics 39. Oxford: Oxford University
Press, 252–78.
Mateu, Jaume and Acedo-Matellán, Víctor (2012). “The manner/result complementarity
revisited: A syntactic approach”, in M. C. Cuervo and Y. Roberge (eds.), The End of
Argument Structure? Series: Syntax & Semantics 38. Bingley: Emerald, 209–28.
McCawley, James D. (1973). “Syntactic and logical arguments for semantic structures”, in
O. Fujimura (ed.), Three Dimensions in Linguistic Theory. Tokyo: TEC Corp, 259–376.
McIntyre, Andrew (2004). “Event paths, conflation, argument structure, and VP shells”,
Linguistics 42: 523–71.
Meirav, Ariel (2003). Wholes, Sums and Unities, Kluwer, Dordrecht.
Meltzer, Aya (2006). “Adjectival passives and adjectival decausatives in Hebrew”, Israel
Association for Theoretical Linguistics 22.
Meltzer-Asscher, Aya (2011). “Adjectival passives in Hebrew: Evidence for parallelism between
the adjectival and verbal systems”, Natural Language and Linguistic Theory 29: 815–55.
324
References
Mendoza, Martha (2010). Spanish auto: Reflexivization and Proxies. Manuscript, University of
Utrecht.
Merchant, Jason (2001). The Syntax of Silence. Oxford: Oxford University Press.
Mittwoch, Anita (1982). “On the difference between eating and eating something: Activities
versus accomplishments”, Linguistic Inquiry 13: 113–22.
Mohanan, Karuvannur (1986). The Theory of Lexical Phonology. Dordrecht: Reidel.
Molu, S. P. (in preparation). Questions d’interface phonologie/syntaxe en bamoun et en
français. Ph.D. dissertation. Université Paris Diderot.
Moro, Andrea (2000). Dynamic Antisymmetry. Movement as a Symmetry Breaking Phenomenon. Cambridge: MIT Press.
Mueller-Reichau, Olav (2006). Sorting the World: On the Relevance of the Kind-level/Objectlevel Distinction to Referential Semantics. Ph.D. dissertation, Universität Leipzig.
Mulder, René H. (1992). The Aspectual Nature of Syntactic Complementation. Holland Institute of Generative Linguistics Dissertation, Dordrecht.
Mutz, Katrin (2004). “Zur Argumentstruktur der deverbalen Ableitungen von auto”, in
M. Hummel and R. Kailuweit (eds.), Semantische Rollen in der Romania: Tübinger Beiträge
zur Linguistik 472. Tübingen: Narr, 355–74.
Νάκας, Θανασης. (1978). Για τη μετοχή και το ρηματικó επίθετο σε –τος óπως εμφανίζονται
στην κοινή Νέα Ελληνική και στις διαλέκτους. Β’ Συμπóσιο Γλωσσολογίας του Βορειοελλαδικού Χώρου (Ήπειρος-Μακεδονία-Θράκη), 241–262.
Newman, Stanley (1946). “On the stress system of English”, Word 2: 171–87.
Newmeyer, Fritz (2009). “Current challenges to the lexicalist hypothesis: An overview and a
critique”, in Lewis, W., S. Karimi, H. Harley, and S. Farrar (eds.), Time and Again:
Theoretical Perspectives on Formal Linguistics in Honor of D. Terence Langendoen. Amsterdam: John Benjamins.
Ouhalla, Jamal (1991). Functional Categories and Parametric Variations. London: Routledge.
Ott, Dennis (2011a). “Diminutive-formation in German: Spelling out the classifier analysis”,
Journal of Comparative German Linguistics 14: 1–46.
Ott, Dennis (2011b). “A note on free relative clauses in the theory of phases”, Linguistic Inquiry
42: 183–92.
Panagiotidis, Phoevos (2009). Categorial Features and Categorisers as Phase Heads. Manuscript, University of Cyprus.
Papangeli, Dimitra (2004). The Morpho-syntax of Argument Realization: Greek Argument
Structure and the Lexicon–Syntax Interface. Ph.D. dissertation, LOT.
Parsons, Terence (1990). Events in the Semantics of English: A Study in Subatomic Semantics.
Cambridge: MIT Press.
Perlmutter, David (1988). “The split Morphology Hypothesis: Evidence from Yiddish”, Theoretical Morphology: 79–101.
Pesetsky, David (1982). Path and Categories. Ph.D. dissertation, MIT.
Pesetsky, David (1995). Zero Syntax: Experiencers and Cascades. Cambridge: MIT Press.
Pesetsky, David and Torrego, Esther (2004). “Tense, case, and the nature of syntactic categories”, in J. Guéron and J. Lecarme (eds.), The Syntax of Time. Cambridge: MIT Press,
495–537.
Pianigiani, Ottorino (1926). Dizionario Etimologico, Firenze: Ariani.
Picallo, Carme (1991). “Nominals and nominalization in Catalan”, Probus 3: 279–316.
References
325
Picallo, Carme (2005). “Some notes on grammatical gender and l-pronouns”, in K. von
Heusinger, G. Kaiser, and E. Stark (eds.), Konstanzer Arbeitspapiere zur Sprachwissenschaft
119, 107–21.
Picallo, Carme (to appear). “Structure of the noun phrase”, in J. I. Hualde et al. (eds.)
Handbook of Hispanic Linguistics. Oxford: Blackwell.
Pietroski. Paul (2005). Events and Semantic Architecture. Oxford: Oxford University Press.
Piggott, Glyne L. and Newell, Heather (2008). Syllabification, Stress and Derivation by Phase in
Ojibwa. Manuscript, McGill University.
Piñón, Christopher (2000). “Happening gradually,” in L. J. Conathan, J. Good, D. Kavitskaya,
A. B. Wulf, and A. C. L. Yu (eds.), Proceedings of the Twenty-Sixth Annual Meeting of the
Berkeley Linguistics Society: 445–56.
Plag, Ingo (2003). Word-formation in English. Cambridge: Cambridge University Press.
Polomé, Edgar (1967). Swahili Language Handbook. Washington: Center for Applied Linguistics.
Putnam, Hilary (1975). Mind, Language and Reality: Philosophical Papers, Volume 2. Cambridge: Cambridge University Press.
Pylkkänen, Liina (2008). Introducing Arguments, Cambridge: MIT Press.
Ralli, Agela (2001). Μορφολογία. Athens: Patakis.
Ralli, Agela (2006). Η Σύνθεση Λέξεων. Athens: Patakis.
Ramchand, Gillian (1997). Aspect and Predication. Oxford: Clarendon Press.
Ramchand, Gillian (2008). Verb Meaning and the Lexicon: A First-Phase Syntax. Cambridge:
Cambridge University Press.
Ramchand, Gillian, and Svenonius, Peter (2002). “The lexical syntax and lexical semantics of
the verb-particle construction”, in L. Mikkelsen, and C. Potts (eds.), Proceedings of WCCFL
21. Somerville: Casacadilla Press, 387–400.
Σετάτος, Μιχαλης. (1984). Παρατηρήσεις στα ρηματικά επίθετα σε –μένος και –τος της κοινής
νεοελληνικής. Studies in Greek Linguistics: Proceedings of the 5th Annual Meeting of the
Department of Linguistics of the Aristotle University of Thessaloniki, Thessaloniki.
Raposo, Eduardo (2002). “Nominal gaps with prepositional modifiers in Portuguese and
Spanish: A case of quick spell-out”, WPL IUOG 9: 127–44.
Rappaport Hovav, Malka (2008). “Lexicalized meaning and the internal structure of events”,
in S. Rothstein (ed.), Theoretical and Crosslinguistic Approaches to the Semantics of Aspect.
Amsterdam: John Benjamins, 13–42.
Rappaport Hovav, Malka (this volume). “Building scalar changes”, 259–81.
Rappaport Hovav, Malka and Levin, Beth (1998). “Building verb meanings”, in M. Butt and
W. Geuder (eds.), The Projection of Arguments. Stanford: CSLI, 97–134.
Rappaport Hovav, Malka and Levin, Beth (2010). “Reflections on manner/result complementarity”, in M. Rappaport Hovav, E. Doron, and I. Sichel (eds.), Syntax, Lexical Semantics,
and Event Structure. Oxford: Oxford University Press, 21–38.
Rapoport, Tova (1993). “Verbs in depictives and resultatives”, in J. Pustejovsky (ed.), Semantics
and the Lexicon. Dordrecht: Kluwer, 163–84.
Reinhart, Tanya and Reuland, Eric (1993). “Reflexivity”, Linguistic Inquiry 24: 657–720.
Rotstein, Carmen and Winter, Yoad (2004). “Total adjectives vs. partial adjectives: Scale
structure and higher order modifiers”, Natural Language Semantics 12: 259–88.
326
References
Richards, Marc (2007). “On feature inheritance: An argument from the phase impenetrability
condition”, Linguistic Inquiry 38: 563–72.
Riemsdijk, H. C. van. (1978). A Case Study In Syntactic Markedness: The Binding Nature of
Prepositional Phrases. Lisse: The Peter de Ridder Press.
Rivero, Maria (1990). “The location of non-active voice in Albanian and Modern Greek”,
Linguistic Inquiry 21: 135–46.
Rivero, Maria (1992). “Adverb incorporation and the syntax of adverbs in Modern Greek”,
Linguistics and Philosophy 15: 289–331.
Rosch, Eleanor (1978). “Principles of categorization”, in E. Rosch and B. B. Lloyd (eds.),
Cognition and Categorization. Hillsdale: Erlbaum, 189–206.
Rosén, Haiim B. (1956). “mefo’al ba’ivrit hayisre’elit. [mefo’al in Israeli Hebrew]”, Leshonenu
20: 139–48.
Ross, John R. (1974). “Wording up”, lectures given at the 1974 Linguistic Institute, University
of Massachusetts.
Ross, John R. (1979). “Five years later”, lecture given at Tel Aviv University.
Roßdeutscher, Antje (2010). “German -ung-formation. An explanation of formation and
interpretation in a root-based account”, in S. Olsen (ed), New Impulses in Word-Formation,
Special issue of Linguistische Berichte. Hamburg, Buske-Verlag, 101–132.
Roßdeutscher, Antje (this volume). “When roots license and when they respect semanticosyntactic structure in verbs”, 282–309.
Roßdeutscher, Antje and Kamp, Hans (2010). “Syntactic and semantic constraints in the
formation and interpretation of -ung-nouns”, in A. Alexiadou and M. Rathert (eds), The
Semantics of Nominalisations across Languages and Frameworks, Interface Explorations 22.
Berlin, Mouton de Gruyter, 169–214.
Rotstein, Carmen and Yoad Winter (2004). “Total adjectives vs. partial adjectives: Scale
structure and higher-order modification”, Natural Language Semantics, 12: 259–288.
Samioti, Yota (2008). “The structure of -tos participles expressing ability/possibility”, Studies
in Greek Linguistics 29.
Samioti, Yota (in progress). Issues in the Lexicon: Syntax Interface and Applications in Modern
Greek as a Second Language, the Case of Participles. Ph.D. dissertation, University of Crete.
Schäfer, Florian (2007). By Itself. Manuscript, Universität Stuttgart.
Schäfer, Florian (2008). The Syntax of (Anti-)Causatives: External Arguments in Change-ofstate Contexts. Amsterdam: John Benjamins.
Schwarzwald, Ora (2008). “The special status of nif ’al in Hebrew”, in S. Armon-Lotem,
G. Danon, and S. Rothstein (eds), Current Issues in Generative Hebrew Linguistics. Amsterdam: John Benjamins, 61–75.
Selkirk, Elisabeth O. (1982). The Syntax of Words, Cambridge: MIT Press.
Siegel, Dorothy (1974). Topics in English Morphology. Ph.D. dissertation, MIT.
Siegel, Dorothy (1979). Topics in English Morphology. New York: Garland.
Siddiqi, Daniel (2009). Syntax within the Word: Economy, Allomorphy and Argument Selection
in Distributed Morphology. Amsterdam: John Benjamins.
Siemund, Peter (2010). “Grammaticalization, lexicalization and intensification: English itself
as a marker of middle situation types”, Linguistics 48(4): 797–836.
Simons, Peter (1987). Parts: A Study in Ontology. Oxford: Clarendon Press.
References
327
Sorace, Antonella (2000). “Gradients in auxiliary selection with intransitive verbs”, Language
76.4: 859–90.
Spathas, Giogos (2010). Focus on Anaphora. LOT Dissertation Series, 264.
Spathas, Giorgos, Alexiadou, Artemis, and Schäfer, Florian (to appear). Middle Voice and
Reflexive Interpretation: Afto-prefixation in Greek. Natural Language and Linguistic Theory.
Spencer, Andrew (1991). Morphological Theory. Oxford: Blackwell.
Starke, Michal (2009). “Nanosyntax: A short primer to a new approach to language”, Nordlyd
36: 1–6.
Stechow, Arnim von (1996). “The different readings of wieder ‘again’: A structural account”,
Journal of Semantics 13: 87–138.
Strawson, Peter (1959). Individuals: An Essay in Descriptive Metaphysics. London: Methuen.
Stroik, Thomas (1992). “Middles and movement”, Linguistic Inquiry 23: 127–37.
Stroik, Thomas (1995). “On middle formation: A reply to Zribi-Hertz”, Linguistic Inquiry 26:
165–71.
Stroik, Thomas (1999). “Middles and reflexivity”, Linguistic Inquiry 30: 119–31.
Stump, Gregory T. (1993). “How peculiar is evaluative morphology?”, Journal of Linguistics 29,
1–36.
Svenonius, Peter (1996). “The verb–particle alternation in the Scandinavian languages”,
unpublished manuscript, University of Tromsø.
Svenonius, Peter (2003). “Limits on P: Filling in holes vs. falling in holes”, Nordlyd 31: 431–445.
Svenonius, Peter (2005). “Extending the extension condition to discontinuous idioms”, Linguistic Variation Yearbook 5: 227–63.
Svenonius, Peter (2007). “Adpositions, particles, and the arguments they introduce”, in
E. Reuland et al. (eds.), Argument Structure. Amsterdam: John Benjamins, 63–103.
Svenonius, Peter (2008). “Projections of P”, in A. Asbury et al. (eds.), Syntax and Semantics of
Spatial P. Amsterdam: John Benjamins, 63–84.
Talmy, Leonard (1985). “Lexicalization patterns: Semantic structure in lexical forms”, in
T. Shopen (ed.), Language Typology and Syntactic Description 3: Grammatical Categories
and the Lexicon. Cambridge: Cambridge University Press, 67–149.
Talmy, Leonard (2000). Toward a Cognitive Semantics. Cambridge: MIT Press.
Tenny, Carol (1992). “The aspectual interface hypothesis”, in I. Sag and A. Szabolcsi (eds.),
Lexical Matters. Stanford: CSLI, 29–54.
Tenny, Carol (1994). Aspectual Roles and the Syntax–Semantics Interface. Dordrecht: Kluwer.
Theophanopoulou-Kontou, Dimitra (1980). Middle Reflexive Verbs in Modern Greek.
Tovena, Lucia M. (2010). “Pluractional verbs that grammaticise number through the part-of
relation”, in R. Bok-Bennema, B. Kampers-Manhe, and B. Hollebrandse (eds.), Romance
Languages and Linguistic Theory 2008: Selected Papers from “Going Romance” Groningen
2008. Amsterdam: John Benjamins, 233–48.
Tovena, Lucia M. (2011). “Issues in the formation of verbs by evaluative suffixation”, in
G. Massariello-Merzagora and S. Dal Maso (eds.), I luoghi della traduzione: Le interfacce.
Atti del XLIII Congresso internazionale di studi della Società Italiana di Linguistica, 913–25.
Travis, Lisa (2005). “Agents and causes in Malagasy and Tagalog”, in N. Erteschik-Shir and
T. Rapoport (eds.), The Syntax of Aspect: Deriving Thematic and Aspectual Interpretation.
Oxford: Oxford University Press, 174–89.
Trépos, Pierre (1968). Grammaire bretonne. Rennes: Imprimerie Simon.
328
References
Tsimpli, Ianthi M. (1989). “On the properties of the passive affix in Modern Greek”, UCL
Working Papers in Linguistics 1, 235–60.
Tsimpli, Ianthi M. (2006). “The acquisition of voice and transitivity alternations in Greek as
native and second language”, in S. Unsworth, T. Parodi, A. Sorace, and M. Young-Scholten
(eds.), Paths of Development in L1 and L2 Acquisition: 15–55.
Tsoulas, George (2006). “Plurality of mass nouns and the grammar of Number”, paper
presented at 29th GLOW Colloquium, Barcelona.
Unsworth, Sharon, Parodi, Teresa, Sorace, Antonella, and Young-Scholten, Martha (2006).
Paths of Development in L1 and L2 Acquisition.
Verkuyl, Henk (1972). On the Compositional Nature of the Aspects. Dordrecht: Reidel.
Verkuyl, Henk (1989). “Aspectual classes and aspectual composition”, Linguistics and Philosophy 12: 39–94.
Volpe, Mark (2004). “Affected object unergatives”, Snippets 8(4): 12–13.
Ware, Robert W. (1979). “Some bits and pieces”, in F.-J. Pelletier (ed.), Mass Terms: Some
Philosophical Problems. Studies in Linguistics and Philosophy 6. Dordrecht: Reidel, 15–29.
Wasow, Thomas (1977). “Transformations and the Lexicon”, in P. W. Culicover, T. Wasow,
and J. Bresnan (eds.), Formal Syntax. New York: Academic Press, 327–60.
Wierzbicka, Anna (1988). The Semantics of Grammar. Philadelphia: John Benjamins.
Wiese, Heike (2003). Numbers, Language, and the Human Mind. Cambridge: Cambridge
University Press.
Wiggins, David (2001). Sameness and Substance Renewed. Cambridge: Cambridge University
Press.
Williams, Edwin (1981). “On the notions lexically related and head of a word”, Linguistic
Inquiry 12: 245–74.
Williams, Edwin (2007). “Dumping Lexicalism”, in G. Ramchand and C. Reiss (eds.), The
Oxford Handbook of Interfaces. Oxford: Oxford University Press, 353–83.
Wiltschko, Martina and Steriopolo, Olga (2007). “Parameters of variation in the syntax of
diminutives”, in M. Radisic (ed.), Proceedings of the 2007 Canadian Linguistics Association
Annual Conference.
Winter, Yoad (2005). “Cross-categorial restrictions on measure phrase modification”, Linguistics and Philosophy, 28: 233–67.
Xu, Fei (2007). “Sortal concepts, object individuation, and language”, Trends in Cognitive
Science 11/9: 400–06.
Zamparelli, Roberto (2000). Layers in the Determiner Phrase. New York: Garland.
Zombolou, Katerina (2004). Verbal Alternations in Greek: A Semantic Approach. Ph.D. thesis,
University of Reading.
Zubizarreta, Maria Luisa and Oh, Eunjeong (2007). On the Syntactic Composition of Manner
and Motion. Cambridge: MIT Press.
Zwaarts, Joost (2005). “Prepositional aspect and the algebra of paths”, Linguistics and Philosophy 28(6): 739–79.
Index
Acquaviva, P. 40, 44, 54, 253
adjectival derivation
denominal adjectives 47–8, 235
Greek 81–2, 86–111
Hebrew passive participles 188–90
adjectival passives 165–7, 183–6
derivation 188–90
stative 170, 172n, 178n
vs. verbal passives 82, 84, 167–9
adjectives
optional patterns 247–9
pseudo-resultatives 213–14
scale 264–5, 268–9
as stative participles 83–4
affixation 47–9
exponence 231–3
inner vs. outer 85, 103
Level Ordering 139–46, 246–7
optional patterns 247–9
reciprocal prefixation 79–80
reflexive prefixation 59–79
roots as 240–7, 250–8
zero C-functors 122–47
see also category-free roots
affixes vs. roots 4, 11–12
agentive voice 108–10
Alexiadou, A. 39, 96
Alexiadou, A. and E. Anagnostopoulou 71,
78, 86–7, 90, 92, 101
Alexiadou, A., E. Anagnostopoulou, and
F. Schäfer 224
Alexiadou, A. and E. Doron 61, 74
Alexiadou, A. and G. Ioardachioaia 67
Anagnostopoulou, E. 86–7, 89, 92
Anagnostopoulou, E. and Y. Samioti 106
Arabic diminutives
Egyptian 160
Tunisian 154
Arad, M. 52, 81, 82, 84–6, 102–3, 105, 110, 209
argument structure 14–18, 194–5, 196–207
Argument Structure Nominals
(AS-nominals) 130–4, 138–9, 144–5
arguments in roots 193
Aronoff, M. 46–8, 247–9, 250, 252
Arquiola, E. 65, 67, 69–70
Baker, M. 33, 205
Baker, M., K. Johnson, and I. Roberts 173n
Basilico, D. 52
Beard, R. 124
Beavers, J. 265
birthing verbs 25
Bittner, D. 44
Boeckx, C. 15n, 37n, 196n
Borer, H. 12, 14, 21, 38–40, 44–5, 46, 50–1,
54–5, 116, 122n, 131
Bosque, I. 203–4
boundedness 25, 28n, 242–3, 269–73, 278–9
Brisson, C. 28n
canonical names 37–8, 39
Carlson, G. 51n, 52
Castella, M. 65, 67
Catalan, telicity 30
Categorial Complement Space (CCS) 114–16,
120–1, 128
categories, exponence 231–3
category-free roots 2–4, 35, 52, 55
English 122–47
Italian 159
causative participles in Hebrew 166, 168–9,
175–82, 186–7, 189–90
change of state (COS) verbs 28–30, 210,
224–8, 263–6
Chomsky, N. 62n, 85, 116–18, 120, 133, 193,
194–5, 199–202, 206–7
Chomsky, N. and M. Halle 239
Clark, E. and H. Clark 126, 211
330
Index
coercion of roots 8, 287, 290–7, 300, 301–9
compositional diminutives 149–63
compositionality, strong interpretation 209
compositionality generalization 85
Compound Stress Rule 142
compounding
noun stems 45–6
reflexives as 76
concepts 50–6
conceptual content of nouns 36–8, 39–43
conceptual properties (C) 17–18
content of roots 54–6
Content-relations 112–13
contextual categorization 113–21
Creole, diminutives 160–1
cyclicity 85, 105, 239, 241, 247, 256–7 see also
Level Ordering
De Belder, M. 11, 39, 40, 44, 150n
derivational morphology 47–9, 85
exponence 231–3
see also adjectival derivation;
nominalization; verbal derivation
directed motion verbs 267–73
discourse referents 34, 40–3
Distributed Morphology (DM) 118–21, 124,
125, 127, 129–30, 136–8, 196, 209
and Stress Shift 230–1, 239
division 43–6
Doron, E. 157, 182–3, 186n
double objects 214–18, 221–2, 226–8
Dowty, D. 224, 259, 262, 281
Dugas, A. 62n, 65, 67
Dutch
diminutives 161–2
noun division 44
reflexives 57–8
Edmondson, J. A. and F. Plank 62n
Elbourne, P. 35
Embick, D. 5, 12, 18n, 60, 64, 72, 76, 79,
88, 136, 166, 169–70, 236, 285
Embick, D. and A. Marantz 196–7
Emonds, J. 205
English
compounding 45
derivational morphology 47
double objects 214–18, 226–8
evaluative morphology 160
manner/result 23
participle formation 83, 110
reflexives 62, 64n, 70
Voice 74
word formation 112
zero categorizers 122–47
entity roots see sortal roots
evaluative affixation 40, 43
evaluative morphology, diminutives 149–63
eventive nouns 44
eventive participles vs. stative participles 83–4
Exo-Skeletal Model (XSM) 125, 137
explicit creation verbs 210, 219–23
exponence 231–3
Extended Projection (ExP) 114, 132, 137–8
Fabb, N. 250, 257–8
Finnish
change of state verbs 225
double objects 215, 222
explicit creation verbs 221
resultatives 213
Folli, R. and H. Harley 212
French
diminutives 160
reciprocal prefixation 79–80
reflexives 65
Gast, V. and P. Siemund 74
German
expected and unexpected roots 283,
290, 304–5
nominalization 283–4
reflexives 62n
verb formation 284–309
Voice 74
Goethe 296n
Greek
adjectival constructions 81–2, 86–111
Index
reciprocal prefixation 79–80
reflexives (afto-prefixation) 58–76, 77–8
Greenberg, Y. 157–8
Grimm and Grimm 296n, 307n
Grimshaw, J. 20, 22n, 64n, 130–1
Hale, K. and S. J. Keyser 15, 16n, 22n, 24–5, 52,
198, 203, 224
Halle, M. and A. Marantz 119n, 134
Harley, H. 5, 11, 15, 24–31, 82, 99, 103, 120n,
132, 196, 224
Harley, H. and J. D. Haugen 27, 141
Haugen, J. D. 27
Head Movement 234–8
Hebrew 82
diminutives 152–3, 154–9
passive participles 167–78
passive participle templates 164–7, 178–91
root meanings 52
Hole, D. 70
homophony 104
idiomatic interpretations 82, 86, 105–10, 210
idiosyncratic meaning 81–2, 84, 104–5
incremental theme verbs 278–81
individuation 34–6
inherently reflexive verbs 57
instrument verbs 24–31
Italian
auxiliary verbs 21
concepts 53
diminutives 149–51, 153, 159, 162–3
eventive nouns 44
reflexives 59–60
Kaisse, E. 247
Katz, J. and D. Pitt 47
Kayne, R. 205
Kearns, K. 275
Kemmer, S. 64
Kennedy, C. and B. Levin 265
Kihm, A. 39–40
Kiparsky, P. 28n, 46, 100, 123, 127, 132,
139–46, 288
König, E. 62n
Koontz-Garboden, A. and J. Beavers 22–3, 30
331
Kratzer, A. 76, 89, 104, 170, 285
Krifka, M. 35n
Labelle, M. 60, 65, 76–7, 79–80
Lampitelli, N. 150n
Latin, concepts 53
Lekakou, M. 92
Level Ordering 12, 139–46, 247, 249
Levin, B. 69, 208, 214, 219, 222–3,
224, 283
Levin, B. and M. Rappaport 84n
Levin, B. and M. Rappaport Hovav 19–20,
288, 296, 297n, 307n
Levinson, L. 211, 213
Lexical Phonology 247–9
Lexical Phonology and Morphology
(LPM) 125, 126–7, 139
Lexicalization Constraint (LC) 19, 288
lexicalized nouns 41–2
Lieber, R. 44, 45
Longobardi, G. 38
Lowenstamm, J. 162–3
Luganda, double objects 214
manner-of-death verbs 22–3
manner/result complementarity 18–24, 72–3
Marantz, A. 8, 15, 17–18, 24, 29, 81, 82, 84–5,
105, 109–11, 118, 120n, 132, 193, 196–7,
199–200, 207, 209, 219, 283
Markantonatou, S. 78
Markantonatou, S., A. Kaliakostas,
V. Bouboureka, V. Kordoni, and
V. Stavrakaki 88, 92
Marvin, T. 234–6
Mateu, J. 198, 205n
Mateu, J. and V. Acedo-Matellán 18–24
Mauritian Creole, diminutives 160–1
meaning
idiosyncratic 81–2, 84, 104–5
non-compositional 10
of roots 5, 52–4
Mendoza, M. 65
morphology, inner vs. outer 85, 103
Multiple Contextualized Meaning
(MCM) 82, 86, 103
Mutz, K. 62n
332
Index
naturally disjoint verbs 58–9
naturally reflexive verbs 57–8
negated participles 94–5
neoconstructionist approach 20–1
neologisms see word formation
Newmeyer, F. 130n
nominal domains 198–9
nominal interpretation 38–43
nominality 33–4, 35, 38
nominalization 38, 41–2, 48–9, 51
deverbal nouns 130–4, 142, 202
German deverbal nouns 283–4
non-compositional diminutives 149–63
non-compositionality 10
non-relational elements 15–17
nonscalar changes 273–4
nonscalar vs. scalar change verbs 274–8
Norwegian, explicit creation verbs 221
nouns
compounding 46
conceptual content 36–8, 39–43
division 43–6
impossible 49–50
individuation 34–6
lack of argument 201–6
obligatory theme 218, 228
optional theme 222–3
Panagiotidis, P. 54
Papangeli, D. 71
Parsons, T. 224
part structure 43–6
participle derivation, Greek 81–2, 86–111
passive participle templates in Hebrew
164–7, 178–91
passive participles
adjectival vs. verbal 167–9
resultative 172–8, 188–9
stative 169–72, 189
passive voice, reflexives 73–5
Perfect of Result 98–9
phase theory 193–5, 200–7, 234–40
phonological properties (F) 17
Piñón, C. 276
Plag, I. 250
plural marking in English 135–6
plurals 44–5, 53
Polish, diminutives 153
polysemy 104
Probe-Goal System 195, 201–2
pseudo-resultatives 213–14
Pylkkänen, L. 214, 215, 222, 224, 225–8
Rappaport Hovav, M. 260
Rappaport Hovav, M. and B. Levin 5, 18–22,
23–4, 28n, 72, 208, 260–1, 277
reciprocal prefixation 79–80
reference objects (ROs) 267–73
reflexive verbs
prefixation 59–79
three classes 57–8
Reinhard, T. and E. Reuland 76
relational elements 15
resultatives 213
Richards, M. 200
Riemsdijk, H. C. van 205
Rivero, M. 75–6
Romance, reflexives (auto-prefixation)
59–61, 64–6, 76–7
root creation verbs 210, 211–19
roots
vs. affixes 4, 11–12
expected and unexpected 283, 290, 304–5
typologies 192–3
Ross, J. R. 248
Roßdeutscher, A. 298, 307n
Roßdeutscher, A. and H. Kamp 283, 289,
298, 304
Russian, concepts 53
Samioti, Y. 92, 93n
scalar vs. nonscalar change verbs 274–8
scalar properties 260–73
Schäfer, F. 75
semantic properties vs. semantic features 17–18
Semitic languages 164 see also Hebrew
single engine hypothesis 209
Sorace, A. 21n
sortal roots 36, 283, 302–5, 308
Index
Spanish
argument structure 201n, 203–4
diminutives 154
exponence 231
reflexives 65–6
Spathas, G., A. Alexiadou, and F. Schäfer 70
Spell-out 200, 234–40
stative participles vs. eventive participles 83–4
stress rules, optional patterns 247–8
Stress Shift 230–1, 235–40
Svenonius, P. 17n, 105
Swahili, concepts 53
synonyms, Hebrew 171
syntactic terminals 1–4
syntax of roots 8–9
Talmy, L. 19
telicity 25–31, 259–61, 275, 278
templates in Hebrew 164–7, 178–91
Tenny, C. 259, 273
tense marking in English 134, 136–7
terminal coincidence relation 15
Theophanopoulou-Kontou, D. 64n, 75
Transfer 200
Tsimpli, I. M. 60, 75, 77
verb classes 210–11
verb formation, German 284–309
verbal derivation
deadjectival verbs 26n, 285
denominal verbs 21, 24–31, 46–7, 132, 140–6
German 306
Greek 96–8
333
Hebrew passive participles 182–7
verbal passives vs. adjectival passives 167–9
verbs
argument structures 14–18
change of state 28–30, 210, 224–8, 263–6
directed motion 267–73
as eventive participles 83–4
explicit creation 210, 219–23
incremental theme 278–81
instrument 24–31
manner vs. result 72–3
manner/result complementarity 18–24
nonscalar changes 273–4
reflexive
prefixation 59–79
three classes 57–8
root creation 210, 211–19
scalar properties 260–73
scalar vs. nonscalar change 274–8
Voice, Greek 73–5
Voice heads 60–1
Volpe, M. 16n
Wasow, T. 84
Wierzbicka, A. 45
Wiltschko, M. and O. Steriopolo 150n, 161
word formation, domains 83–5
Zamparelli, R. 38
zero categorizers/category-free roots 113–21,
122–47, 224
zero derivation 122–30, 134–46
Zombolou, K. 62–3, 66n, 71, 75
OXFORD STUDIES IN THEORETICAL LINGUISTICS
Published
1 The Syntax of Silence
Sluicing, Islands, and the Theory of Ellipsis
by Jason Merchant
2 Questions and Answers in Embedded
Contexts
by Utpal Lahiri
3 Phonetics, Phonology, and Cognition
edited by Jacques Durand and Bernard Laks
4 At the Syntax-Pragmatics Interface
Concept Formation and Verbal Underspecification
in Dynamic Syntax
by Lutz Marten
5 The Unaccusativity Puzzle
Explorations of the Syntax-Lexicon Interface
edited by Artemis Alexiadou,
Elena Anagnostopoulou, and Martin Everaert
6 Beyond Morphology
Interface Conditions on Word Formation
by Peter Ackema and Ad Neeleman
7 The Logic of Conventional Implicatures
by Christopher Potts
8 Paradigms of Phonological Theory
edited by Laura Downing, T. Alan Hall, and
Renate Raffelsiefen
9 The Verbal Complex in Romance
by Paola Monachesi
10 The Syntax of Aspect
Deriving Thematic and Aspectual Interpretation
Edited by Nomi Erteschik-Shir and Tova Rapoport
11 Aspects of the Theory of Clitics
by Stephen Anderson
12 Canonical Forms in Prosodic Morphology
by Laura J. Downing
20 Adjectives and Adverbs
Syntax, Semantics, and Discourse
Edited by Louise McNally and Christopher
Kennedy
21 InterPhases
Phase-Theoretic Investigations of Linguistic
Interfaces
edited by Kleanthes Grohmann
22 Negation in Gapping
by Sophie Repp
23 A Derivational Syntax for Information
Structure
by Luis López
24 Quantification, Definiteness, and
Nominalization
edited by Anastasia Giannakidou and Monika
Rathert
25 The Syntax of Sentential Stress
by Arsalan Kahnemuyipour
26 Tense, Aspect, and Indexicality
by James Higginbotham
27 Lexical Semantics, Syntax, and Event
Structure
edited by Malka Rappaport Hovav, Edit
Doron, and Ivy Sichel
28 About the Speaker
Towards a Syntax of Indexicality
by Alessandra Giorgi
29 The Sound Patterns of Syntax
edited by Nomi Erteschik-Shir and Lisa
Rochman
30 The Complementizer Phase
edited by Phoevos Panagiotidis
13 Aspect and Reference Time
by Olga Borik
31 Interfaces in Linguistics
New Research Perspectives
edited by Raffaella Folli and Christiane
Ulbrich
14 Direct Compositionality
edited by Chris Barker and Pauline Jacobson
32 Negative Indefinites
by Doris Penka
15 A Natural History of Infixation
by Alan C. L. Yu
33 Events, Phrases, and Questions
by Robert Truswell
16 Phi-Theory
Phi-Features Across Interfaces and Modules
edited by Daniel Harbour, David Adger,
and Susana Béjar
34 Dissolving Binding Theory
by Johan Rooryck and Guido Vanden
Wyngaerd
17 French Dislocation
Interpretation, Syntax, Acquisition
by Cécile De Cat
18 Inflectional Identity
edited by Asaf Bachrach and Andrew Nevins
19 Lexical Plurals
by Paolo Acquaviva
35 The Logic of Pronominal Resumption
by Ash Asudeh
36 Modals and Conditionals
by Angelika Kratzer
37 The Theta System
Argument Structure at the Interface
edited by Martin Everaert, Marijana Marelj,
and Tal Siloni
38 Sluicing
Cross-Linguistic Perspectives
edited by Jason Merchant and Andrew
Simpson
39 Telicity, Change, and State
A Cross-Categorial View of Event Structure
edited by Violeta Demonte and Louise
McNally
40 Ways of Structure Building
edited by Myriam Uribe-Etxebarria and
Vidal Valmala
41 The Morphology and Phonology of
Exponence
edited by Jochen Trommer
42 Count and Mass Across Languages
edited by Diane Massam
43 Genericity
edited by Alda Mari, Claire Beyssade, and
Fabio Del Prete
44 Strategies of Quantification
edited by Kook-Hee Gil, Steve Harlow, and
George Tsoulas
45 Nonverbal Predication
Copular Sentences at the Syntax-Semantics
Interface
by Isabelle Roy
46 Diagnosing Syntax
edited by Lisa Lai-Shen Cheng and Norbert
Corver
47 Pseudogapping and Ellipsis
by Kirsten Gengel
48 Syntax and its Limits
edited by Raffaella Folli, Christina Sevdali,
and Robert Truswell
49 Phrase Structure and Argument Structure
A Case Study of the Syntax-Semantics Interface
by Terje Lohndal
50 Edges in Syntax
Scrambling and Cyclic Linearization
by Heejeong Ko
51 The Syntax of Roots and the Roots
of Syntax
edited by Artemis Alexiadou, Hagit Borer,
and Florian Schäfer
Published in association with the series
The Oxford Handbook of Linguistic
Interfaces
edited by Gillian Ramchand and Charles
Reiss
In preparation
External Arguments in Transitivity
Alternations
by Artemis Alexiadou, Elena Anagnostopoulou,
and Florian Schäfer
Modality Across Syntactic Categories
edited by Ana Arregui, Mari Luisa Rivero,
and Andrés Pablo Salanova
Semantic Continuations
Scope, Binding, and Other Semantic
Side Effects
by Chris Barker and Chung-chieh Shan
Phi Syntax
A Theory of Agreement
by Susana Béjar
Stratal Optimality Theory
by Ricardo Bermúdez Otero
Phonology in Phonetics
by Abigail Cohn
Generality and Exception
by Ivan Garcia-Alvarez
The Morphosyntax of Gender
by Ruth Kramer
The Semantics of Evaluativity
by Jessica Rett
Computing Optimality
by Jason Riggle
Pragmatic Aspects of Scalar Modifiers
by Osamu Sawada
Gradience in Split Intransitivity
by Antonella Sorace
Descargar