Вы находитесь на странице: 1из 479

Syntax and its Limits

general editors: David Adger and Hagit Borer, Queen Mary, University of London
advisory editors: Stephen Anderson, Yale University; Daniel Bring, 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 Troms; Moira Yip, University College London.
recent titles
Lexical Semantics, Syntax, and Event Structure
edited by Malka Rappaport Hovav, Edit Doron, and Ivy Sichel
About the Speaker
Towards a Syntax of Indexicality
by Alessandra Giorgi
The Sound Patterns of Syntax
edited by Nomi Erteschik-Shir and Lisa Rochman
The Complementizer Phase
edited by E. Phoevos Panagiotidis
Interfaces in Linguistics
New Research Perspectives
edited by Raffaella Folli and Christiane Ulbrich
Negative Indefinites
by Doris Penka
Events, Phrases, and Questions
by Robert Truswell
Dissolving Binding Theory
by Johan Rooryck and Guido Vanden Wyngaerd
The Logic of Pronominal Resumption
by Ash Asudeh
Modals and Conditionals
by Angelika Kratzer
The Theta System
Argument Structure at the Interface
edited by Martin Everaert, Marijana Marelj, and Tal Siloni
Cross-Linguistic Perspective
edited by Jason Merchant and Andrew Simpson
Telicity, Change, and State
A Cross-Categorial View of Event Structure
edited by Violeta Demonte and Louise McNally
Ways of Structure Building
edited by Myriam Uribe-Etxebarria and Vidal Valmala
The Morphology and Phonology of Exponence
edited by Jochen Trommer
Count and Mass Across Languages
edited by Diane Massam
edited by Alda Mari, Claire Beyssade, and Fabio Del Prete
Strategies of Quantification
edited by Kook-Hee Gil, Steve Harlow, and George Tsoulas
Nonverbal Predication
Copular Sentences at the Syntax-Semantics Interface
by Isabelle Roy
Diagnosing Syntax
edited by Lisa Lai-Shen Cheng and Norbert Corver
Pseudogapping and Ellipsis
by Kirsten Gengel
Syntax and its Limits
edited by Raffaella Folli, Christina Sevdali, and Robert Truswell
For a complete list of titles published and in preparation for the series, see pp. .

Syntax and its Limits

Edited by

Great Clarendon Street, Oxford, ox dp,

United Kingdom
Oxford University Press is a department of the University of Oxford.
It furthers the Universitys 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 Raffaella Folli, Christina Sevdali, and
Robert Truswell
The chapters their several authors
The moral rights of the authors have been asserted
First Edition published in
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
Madison Avenue, New York, NY , United States of America
British Library Cataloguing in Publication Data
Data available
Library of Congress Control Number:
ISBN (hbk.)
Printed in Great Britain by
the CPI Printgroup, UK

General Preface
List of Figures and Tables
Raffaella Folli, Christina Sevdali, and Robert Truswell
. Part I: Architectures
. Part II: Syntax and Information Structure
. Part III: Syntax and the Lexicon
. Part IV: Lexical Items at the Interfaces

Part I. Architectures
Harmonic Derivationalism
Winfried Lechner
. Introduction
. Opacity
. The Condition on Extraction from Copies
. Analysis
. Discussion
. Conclusion
Reconstruction, Control, and Movement
Robert Truswell
. Introduction
. Reconstruction Across Control Dependencies
. Lechners Double Dissociation
. Trapping and Countertrapping
. Conclusion
Linearizing Empty Edges
Terje Lohndal and Bridget Samuels
. Introduction
. A Phonological Account of the ECP
. Towards a New Account of Empty-Edge Effects
. Conclusion



Evidence for the Use of Verb Telicity in Sentence Comprehension
Erin OBryan, Raffaella Folli, Heidi Harley, and Thomas G. Bever
. Introduction
. Prior Experiments Bearing on the Hypothesis
. The Word Maze Experiment
. General Discussion and Conclusions

Part II. Syntax and Information Structure

Focus Intervention in Declaratives

Aviad Eilam
. Introduction
. Intervention Effects in Questions
. Intervention Effects in Declaratives
. Conclusion and Implications

Root Phenomena as Interface Phenomena: Evidence from

Ccile De Cat
. Introduction
. Root Phenomena and their Host
. Non-sententials/Fragments
. First Set of Data: French Dislocated Topics in Fragments
. Second Set of Data: Japanese Politeness Markers in Fragments
. Conclusions
Contrast and its Relation to wa in Japanese and nun in Korean
Reiko Vermeulen
. Introduction
. Topic, Focus, and Contrast
. Mapping [contrast] and [topic]
. Ordering Restrictions between Contrastive Topic
and Contrastive Focus
. Contrastive wa- and nun-Phrases as Foci
. Conclusion

Part III. Syntax and the Lexicon

Adjuncts Within Words and Complex Heads
Glyne Piggott and Lisa deMena Travis
. Introduction
. The Spell-Out of Adjuncts
. Other Accounts

. The Structure of Words
. Conclusion
Still Puzzled by Adjectival Passives?
Berit Gehrke
. Introduction
. Event-Related Modification
. The Proposal
. Conclusion
The Role of Syntax in Stress Assignment in Serbo-Croatian
Boban Arsenijevic and Marko Simonovic
. Introduction
. Two Types of De-adjectival Nominalization by Suffixation
in Serbo-Croatian
. Syntactic Analysis: Arsenijevic ()
. Phonology
. The Emerging Picture: Paradigms as the Place of
SyntaxPhonology Encounters
. Zooming Out: Across Suffixes and Languages
. Conclusion
Allosemy, Idioms, and their Domains: Evidence from Adjectival
Elena Anagnostopoulou and Yota Samioti
. Goals
. Background
. Greek -menos and -tos Adjectival Participles: Outer and Inner
. -tos Attaching to Verbalizers That Are Not Eventivizers:
Inner or Outer Affixation?
. vC, vE, Voice, and Idiomaticity
. vE, vC, and Allosemy
. English Participles from the Perspective of Greek:
vC, vE, and Voice in English
The No Agent Idioms Hypothesis
Heidi Harley and Megan Schildmier Stone
. Introduction: Interface-Related Questions Raised by Idioms
. Compositionality
. Structural Constraints on Idioms
. Is the No Agent Idioms Hypothesis True?
. Conclusions




Part IV. Lexical Items at the Interfaces

On the Syntax and Semantics of the Japanese Comparative
Hazel Pearson
. Introduction
. The Japanese Comparative
. The Proposal
. Consequences
. Conclusion

Bare Number
Theodora Alexopoulou, Raffaella Folli, and George Tsoulas
. Introduction
. The Empirical Picture
. Number
. Some Speculative Remarks on the NMP
. Conclusion

Obligatory Resumption in Greek Free and Restrictive Relatives

Evangelia Daskalaki and Marios Mavrogiorgos
. Introduction
. The Evidence
. The Analysis
. Conclusions and Implications

Ethical Datives: A Puzzle for Syntax, Semantics, Pragmatics,

and their Interfaces
Dimitris Michelioudakis and Eleni Kapogianni
. Introduction
. Ethical Datives are a Class of their Own: Grammar and Meaning
. The Semantics and Pragmatics of the ED
. Conclusions
The Syntacticization of Discourse
Liliane Haegeman and Virginia Hill
. Introduction
. The Empirical Data
. The Syntactic Properties of the Particles
. The Interpretation of Discourse Particles
. The Syntax of the Romanian Speech Act Particle hai
. The Syntax of West Flemish Particles
. General Conclusions

A Syntactic Answer to a Pragmatic Puzzle: The Case of
Asymmetric and
Bronwyn Moore Bjorkman
. Introduction
. Asymmetric Coordination: The Facts
. Links between Structure and Interpretation in Clausal
. Discussion
. Conclusion


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 subdisciplines of linguistics. The notion of interface has become central
in grammatical theory (for instance, in Chomskys 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
syntaxmorphology, syntaxsemantics, syntaxphonology, syntaxpragmatics,
morphologyphonology, phonologyphonetics, phoneticsspeech processing,
semanticspragmatics, and intonationdiscourse 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 interlanguage 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 current volume explores the relationship between syntax and other components of the grammar. It does this by bringing together a range of studies that
investigate whether certain phenomena are best dealt with by enriching the syntactic
component, enriching other components with which the syntax interfaces, or developing new approaches to the interfaces between components. This method of attack
on architectural problems in theoretical syntax ultimately stems from Chomskys
Remarks on Nominalization (), which concerns the relationship between the
syntactic and lexical components. In this volume other domains are also considered,
including information structure, parsing, pragmatic structure, and conceptual structure. A number of points of convergence emerge, as well as new questions as to the
standard views of how syntax interacts with other types of linguistic information.
David Adger
Hagit Borer

In , a conference On Linguistic Interfaces (OnLI) was held at the University
of Ulster, bringing together scholars working on a wide range of interface-related
issues. The success of the first OnLI conference and its uniqueness in addressing the
central, interdisciplinary question of the nature of the overall architecture of language
encouraged our sponsors, particularly the University of Ulster Strategic Fund and the
Institute for Research in Social Science, to support the second OnLI conference, which
would not have been possible without their financial and practical assistance.
Most of the chapters in the present volume started their life as papers presented at
OnLI II, held in December at the University of Ulster. We would like to thank
the abstract committee, our colleagues Alison Henry, Catrin Rhys, and Christiane
Ulbrich, our student helpers Megan Devlin, Frances Kane, and Aveen Hassan, all the
speakers, and our university sponsors and the LAGB for their financial support.
The conference was very successful, and the idea for this volume was born during
the conference dinner. In fact, we strongly feel that Elena Anagnostopoulou deserves
a special mention, as it was in conversation with her that the question we wanted to
address with this volume was uncovered: what are the limits of syntax?
Going from an idea to a volume is a long and winding road. We would like to
thank two anonymous Oxford University Press reviewers and Kleanthes Grohmann
for very useful comments on the proposed volume. Also, all the authors for their
contributions, as well as our reviewers, who gave up their time, often with very tight
schedules, to give us comments on the submissions. Finally, we want to express our
gratitude to OUP and in particular Julia Steer for invaluable, prompt, and expert
R. F., C. S., and R. T.

List of Figures and Tables

. Processing difficulty measured as eye fixation times per word in each
sentence zone
. The word maze task
. Word maze results
. Effects of telicity and transitivity in the word maze


Mean self-paced reading times in post hoc reanalyses of two prior studies
Mean speaker change detection errors for the four verb types
Word maze ANOVA results
The morphological paradigm of weak pronouns/clitics in Greek
Narrow syntactic properties of IO resumptives/doubling clitics in Greek
List of verb-based particles with pragmatic role
Inflectional properties

Theodora Alexopoulou is a Senior Research Associate at the Department of Theo-

retical and Applied Linguistics, University of Cambridge, leading the Education First
Unit, which is devoted to research in second language learning of English and innovation in language teaching through a systematic cross-fertilization between linguistic
research and teaching techniques. After a BA in Greek philology at the University
of Athens, she went to Edinburgh to obtain an MSc. in natural language and speech
processing and then a PhD in Linguistics. Before coming to Cambridge she worked
at the Universities of Edinburgh and York and held an Intra-European Marie Curie
Fellowship at the University of Lille III. Her research interests are in second language
acquisition and theoretical and experimental syntax, focusing on English and Greek.
She has published articles in Language, Cognition, Journal of Linguistics, Natural Language & Linguistic Theory, and Lingua, as well as volumes and conference proceedings.
She lives in Cambridge with her husband and two children.
Elena Anagnostopoulou earned her PhD in linguistics from the University of
Salzburg. She is currently Professor of Linguistics and Director of Graduate Studies
at the Department of Philology, University of Crete. Her research interests lie in
theoretical and comparative syntax, typology, and Greek syntax, with special focus on
the interface between syntax and the lexicon, Case, agreement, clitics, and anaphora.
She is currently working on argument alternations, participles, and compositionality,
long-distance anaphora, and Case theory. She is an external collaborator in the SFB
project Collaborative Research Center (University of Stuttgart) on argument
structure and morphosyntax, and was invited to teach theoretical syntax at the Thermi
International Summer School in Linguistics (TISSL) , the Girona International
Summer School in Linguistics (GISSL) , the ENS (EALing ) Fall School ,
and as a Visiting Professor at the Department of Linguistics and Philosophy, MIT
Boban Arsenijevic received his PhD in linguistics from Leiden University in . His
dissertation attempts to reconcile the quantitative and the decompositional theory of
inner aspect, deriving the effects of the former from the primitives of the latter. He has
since worked at the University of Amsterdam, Universitat Pompeu Fabra in Barcelona,
and CSIC Madrid, and is currently engaged as an Assistant Professor at the University
of Ni. He works in the domains of syntax, semantics, and cognition from a theoretical
perspective, and on Slavic languages, with a focus on Serbo-Croatian. His main fields
of interest are verbal aspect and tense, clausal embedding, Case, syntactic recursion,
and the ontology of syntactic and semantic categories.



Thomas G. Bever is Regents Professor of Linguistics, Psychology, Neuroscience, Cog-

nitive Science, and Education at the University of Arizona. He started the first psycholinguistics PhD programme at Columbia University, is a co-founder of the journal
Cognition, founder of the Center for the Sciences of Language at Rochester University, and recent head of the University of Arizona Linguistics Department. His five
decades of research have focused on the behavioural, maturational, and neurological
foundations of linguistic universals. This research has involved studies of language
processing in adults and children, cross linguistic investigations, studies of cerebral
asymmetries in humans and animals, and studies of aesthetics relating to vision and
Bronwyn Moore Bjorkman is a postdoctoral fellow at the University of Toronto, and

finished her PhD at the Massachusetts Institute of Technology in . Her research

has focused on questions at the interface between morphology and syntax, and questions involving the representation and realization of verbal inflection in general and
auxiliary verb constructions in particular. She has also worked on the effects of morphological structure on the phonology of stress assignment and reduplication, and on
issues at the interface between syntax and semantics, relating to the semantic effects
of syntactic feature markedness, and the interpretive consequences of clause size in
Evangelia Daskalaki received her degree in Greek philology (major in linguistics)
from the University of Athens in . After her bachelor studies she moved to the
UK, where she obtained an MPhil. in general linguistics and comparative philology
from the University of Oxford, and a PhD in theoretical linguistics from the University
of Cambridge. Both her MPhil. and PhD dissertations concern topics related to the
morphosyntax of A -movement dependencies in Greek. Since she has been a
postdoctoral fellow at the Department of Linguistics at the University of Alberta,
and has expanded her data set to include non-Indo-European languages. Her current research project concerns the multiple functions of interrogative pronouns in
Lushootseed (Salishan).
Ccile De Cat is Senior Lecturer in Linguistics at the University of Leeds. She holds
an MA in linguistics from the University of Durham, and a PhD in linguistics from
the University of York. Her research focuses on the syntactic reflexes of information
structure and their acquisition, the division of labour between syntax and the discourse component, and the interplay between language and cognition in monolingual
and bilingual children.
Lisa deMena Travis received her BA in literature from Yale University in . She
taught at the Universit dAntananarivo in Madagascar from to , followed
by two years of graduate work at McGill University in Montreal, Quebec. She completed her PhD in Linguistics at MIT in , writing her thesis on parameters



of word order variation. Since she has taught in the Department of Linguistics at McGill. Her research focuses mainly on phrase structure, head movement,
language typology, Austronesian languages (in particular Malagasy), and the PF
Aviad Eilam completed his PhD at the University of Pennsylvania in . His dis-

sertation, Explorations in the Informational Component, argues that information

structure is an autonomous level of representation in the grammar, based on the
analysis of two phenomena: focus intervention and weak crossover effects. He is now
pursuing a career in writing and communications in Washington, DC, and maintains
an interest in linguistic research.
Raffaella Folli is a Senior Lecturer in Linguistics at the University of Ulster. Her
research interests are theoretical and comparative syntax and language processing,
with special focus on the syntaxlexicon and the syntaxsemantics interface. She has
published in journals such as Linguistic Inquiry, Journal of Linguistics, Lingua, and
Trends in Cognitive Science, as well as in several edited volumes.
Berit Gehrke is a postdoctoral researcher at the Universitat Pompeu Fabra in

Barcelona. In she completed her PhD at the University of Utrecht on Ps in

Motion: On the Semantics and Syntax of P Elements and Motion Events. She is
interested in lexical and compositional semantics and the syntaxsemantics interface,
and has worked on various topics related to event predicates, time and space, the
kind/token distinction, and modification.
Liliane Haegeman was Professor of English Linguistics at the University of Geneva

(Switzerland) from to . Between and she was Professor of

English Linguistics at the University of Lille III (France). Since she has held a
research position at Ghent University, where she is project leader for the Odysseus
project Comparative Syntax: Layers of Structure and the Cartography Project. She
has worked extensively on the syntax of English and Flemish, including her native
dialect of West Flemish. Her research is characterized by a constant attention to the
empirical data, which she examines against the background of the generative formal
Heidi Harley is Professor of Linguistics at the University of Arizona. She is interested
in the interfaces between lexical semantics and syntax, and between morphology and
syntax, and has worked on these questions in English, Italian, Irish, Hiaki, Icelandic,
and Japanese, largely within the Distributed Morphology framework. Her work has
appeared in Linguistic Inquiry, Language, and Journal of Linguistics, among others.
Virginia Hill is a Professor of Linguistics at the University of New BrunswickSaint
John. She works in comparative and diachronic syntax, with focus on Romanian,
Romance, and Balkan languages. She authored two books, edited several volumes



of papers, and published articles in major international journals. She was recently
awarded a Leverhulme Visiting Professorship at the University of Kent.
Eleni Kapogianni completed her PhD on the semantics and pragmatics of verbal irony

at the University of Cambridge (), where she currently teaches introductory and
advanced semantics and pragmatics. Her research focuses on implicature interpretation in discourse, using evidence from an experimental approach to speakers intuitions. Her further interests include philosophy of language, intercultural pragmatics,
and discourse analysis.
Winfried Lechner is Associate Professor of German Linguistics and Theoretical Lin-

guistics at the University of Athens, Greece. His main academic interests are located
in the areas of syntax and semantics, focusing on the interaction between these two
components. In recent and current work he has been investigating the logical syntax
of scope, scope freezing, and scope rigidity; consequences of an LF-transparent theory of reflexivization; the cross linguistic typology of same/different; reconstruction
effects; the analysis of opacity phenomena at the syntaxsemantics interface; and the
architecture of the grammar.
Terje Lohndal is Associate Professor of English Linguistics at the Norwegian University of Science and Technology in Trondheim. He is the editor of In Search of Universal
Grammar, published by John Benjamins in , and he has published widely, among
others in Linguistic Inquiry and Journal of Linguistics. He received his PhD from the
University of Maryland in .
Marios Mavrogiorgos is an independent researcher specializing in Minimalist Syn-

tax. He received a BA degree in Greek philology (major in linguistics) from the

University of Athens in . He also obtained an MPhil. and a PhD in theoretical linguistics from the University of Cambridge. His PhD thesis was published
in a revised form in by John Benjamins. After the completion of his PhD,
he spent some time doing research in Greece and at MIT, and in he
took up a temporary teaching fellow position at the University of Ulster. He has
published a number of peer-reviewed articles, and his work mainly focuses on different aspects of clitics in Greek and other languages. Currently he is working on
Tobler-Mussafia languages, on resumption in Greek, and on verbless sentences in
Dimitris Michelioudakis is a postdoctoral researcher at the University of Cambridge,

where he teaches introductory and advanced syntax and introductory semantics/

pragmatics. He recently completed a PhD in Cambridge on the syntax and interpretation of dative arguments from a macro-/micro-comparative and diachronic
perspective. His research interests include issues of argument structure theory,
especially the typology of non-core arguments, relation-changing constructions,



agreement restrictions, theta-related Case, and the morphosyntax and interpretation

of clitics.
Erin OBryan (Richtsmeier) is an instructor and visiting scholar in the Department

of Speech, Language, and Hearing Sciences at Purdue University. She received her
PhD in linguistics in and a Masters degree in speech language pathology in
, both at the University of Arizona. Dr OBryans doctoral and post-doctoral
research in the area of psycholinguistics centres on the use of predicate event structure
information in sentence comprehension. In the field of speech language pathology, her
research focuses on aphasia treatment, specifically syntactic and lexical treatments,
using statistical meta-analyses and evidence-based practice principles. Her current
research interests are in verb-oriented aphasia treatment for agrammatism and the
neurological correlates of the syntaxsemantics interface.
Hazel Pearson is a researcher at the Centre for General Linguistics (ZAS) in Berlin.
After studying linguistics as an undergraduate at the University of Cambridge, she
obtained her PhD from Harvard University in , with a specialization in formal
semantics. Her doctoral thesis, The Sense of Self: Topics in the Semantics of De
Se Expressions, was written under the supervision of Gennaro Chierchia. In this
work, a variety of phenomena bearing on the proper analysis of attitudes de se are
explored, particularly logophoric pronouns, control, and predicates of personal taste.
Her research spans experimental as well as formal semantics, and incorporates data
from under-studied languages such as Fijian (Oceanic) and Ewe (Niger-Congo).
In addition to her projects on comparatives and attitude reports, she has investigated
topics related to plurality, implicature, and presupposition.
Glyne Piggott received his PhD in linguistics from the University of Toronto in .

He taught for one year () at the University of Western Ontario and moved
to McGill University in , where he is currently Emeritus Professor of Linguistics.
His research focuses on the phonology and morphology of Ojibwe (a Native American
language), dependency relations in phonological representations, the phonology of
nasals and nasalization, and the interface between phonology and morphosyntactic
Panagiota (Yota) Samioti is a PhD student in linguistics at the University of Crete. She

received her MSc. degree in linguistics from the University of Crete and her MPhil.
degree in theoretical and applied linguistics from the University of Cambridge. She is
mainly interested in syntax, linguistic interfaces, and second language acquisition.
Bridget Samuels is Senior Editor at the Center of Craniofacial Molecular Biology,

University of Southern California. She is the author of the Oxford University

Press monograph Phonological Architecture: A Biolinguistic Perspective. Previously,
she held positions at the California Institute of Technology and the University of



Maryland, College Park. She received her PhD in linguistics from Harvard University
in .
Christina Sevdali is a lecturer in linguistics at the University of Ulster. She received

her BA from the University of Crete and her MPhil. and PhD from the University
of Cambridge. Her main area of expertise is Ancient Greek syntax, but her research
interests also include multilingual acquisition and the syntaxmorphology interface,
particularly Case. Her paper Ancient Greek infinitives and Phases will appear in
Syntax, and her collaboration with Artemis Alexiadou and Elena Anagnostopoulou
on Patterns of DativeNominative Alternations will appear in the Proceedings of
Marko Simonovic (MA, born ) is currently in the final phase of his PhD project
Lexicon Immigration Service: Prolegomena to a Theory of Loanword Integration,
supervised by Ren Kager and Wim Zonneveld at Utrecht University, the Netherlands.
This dissertation is an attempt at reuniting the generative approaches to language
contact with the insights from sociolinguistics, creating a model which views borrowing as the creation of a new lexical item in the recipient language but also recognizes the transformational knowledge accumulated in the bilingual community. His
research is mostly concerned with paradigm uniformity, surface-to-surface relations
in phonology, lexicon stratification, mappings between related varieties (the so-called
interlanguage mappings), and Serbo-Croatian prosody. Apart from formal linguistics, his areas of scientific interest include biolinguistics, sociolinguistics (especially
the standardization and emancipation discourses), science and technology studies
(especially the place of language in the current models), and new materialisms (more
specifically Karen Barads agential realism).
Megan Schildmier Stone is a PhD candidate in the Department of Linguistics at
the University of Arizona. She is interested in morphology, syntax, and semantics,
particulary as they pertain to idioms, and has done work on English and Cherokee.
She is currently writing her dissertation, to be completed in May , which uses
both traditional and experimental methods to investigate what idioms can tell us
about the limits of human language.
Robert Truswell is Assistant Professor of Syntax at the University of Ottawa. Prior
to that, he was a PhD student at University College London, and a postdoctoral
researcher at Tufts University and the University of Edinburgh. He has published
on various aspects of the syntaxsemantics interface, including the monograph
Events, Phrases, and Questions (Oxford University Press), and has other research
interests in diachronic syntax and the evolution of language.
George Tsoulas is a Senior Lecturer in Linguistics at the University of York. After an

undergraduate degree in linguistics and literature at the University of Strasbourg he

went on to study for a PhD at the University of Paris VIII. His research to date has



focused on the syntaxsemantics and syntaxpragmatics interfaces, and more specifically on issues of quantification, tense and modality, number and the count/mass distinction, topic/focus articulation, particles, and the nature of pronominal reference.
Reiko Vermeulen is a postdoctoral researcher at Ghent University, Belgium. She
received her PhD in from University College London with a dissertation on
external possession in Japanese and Korean. She spent a further three years at University College London as a postdoctoral researcher in an AHRC-funded project on
the interface between syntax and information structure. Her main research interests
include the interaction between syntax, semantics, and information structure, particularly in English, Japanese, and Korean. Her published work includes articles in
Lingua, Linguistic Inquiry, and The Linguistic Review, as well as chapters in books published by Oxford University Press and Mouton de Gruyter. She is also a co-editor of
The Syntax of Topic, Focus and Contrast: An Interface-Based Approach (, Mouton
de Gruyter, with Ad Neeleman).


st/nd/rd person agreement


Answer (Ch. )


Aboutness-shift topic






Adverbial Phrase


Animate intransitive verb


Alternative question


Analysis of variance




Adjective Phrase












Berkeley Construction Grammar




Ban On Null Edge


Bare subnominal deletion



Contrastive topic


Condition on Extraction from Copies


Condition on Extraction Domain


Contrastive focus


Common ground






Classifier Phrase


Clitic Left Dislocation








Contrast Phrase


Coordinate Phrase




Complementizer Phrase


Contrastive topic



Discourse Linking








Degree Phrase


Grammatical dependency






Distributed Morphology


Direct object


Domain of contrast


Determiner Phrase


Deductive system


External argument


Explicit comparison


Empty Category Principle


Ethical dative


Edge Property


Extended Projection Principle


Evaluative Phrase


Focal stress (Ch. )






Finiteness Phrase


Future Marker







Focus Phrase


Force Phrase


Free Relative


Givenness topic


Government and Binding








Head-driven Phrase Structure Grammar


Intonational Phrase


Indefinite Argument Drop


Implicit comparison


Individual-level predicate








Indicative mood






Indirect object


Inflectional Phrase


Intonational Phrase Edge Generalization


Information Structure


Lexical Conservatism


Lexical de-adjectival nominalization


Logical Form




Lexical insertion




Left periphery
















Nominal mapping parameter




Noun Phrase


Negative polarity item


Not significant




Number Phrase








Overt Structure





Particle Phrase




Person Case Constraint








Phonological Form








Prepositional Phrase






Predicative Phrase





Quantifier Phrase


Quantifier Raising


Question under discussion










Revised Extended Standard Theory


Result Phrase


Recnik srpskohrvatskoga knjievnog jezika


Restrictive Relative


Reduced Relative Effect






Situation variable


Speech act


Speech act Phrase


Structurally complex De-adjectival Nominalization


Semantic Reconstruction


Singular agreement


Subject gap in finite/fronted


Strong implicit comparison


Speaker Deixis Phrase










Syntactic Reconstruction






Telic Phrase




Topic Phrase


Tense Phrase





Uniformity Condition


Vowel (Ch. )






Voice Phrase


Verb Phrase


Weak Cross Over


West Flemish


Weak implicit comparison


With respect to


What you see is what you get


Phrases of unspecified category


Yes/No question



Since its inception, generative grammar has pursued a divide and conquer strategy
with respect to the study of linguistic phenomena, inheriting from its predecessors in
structuralist linguistics and traditional grammar the useful notion that different linguistic phenomena are amenable to modes of explanation which suggest the existence
of clusters of linguistic phenomena, some related to sound, some to grammar, some to
meaning, some to aspects of language use, and so on. To capture this, Chomsky ()
formalized the notion of a linguistic level and insisted on the relative autonomy of
these levels. The differentiation of linguistic levels was far from new in itself, of course,
but it takes on a new significance given Chomskys mentalist perspective, delivering
a view of the language faculty, and by extension cognition, as something structured,
with differentiated subcomponents, including minimally a lexicon, and phonological,
syntactic, and semantic components, each with their own structural characteristics, in
stark contrast to the then-dominant behaviourist view of the mind as a unitary black
Naturally a theory containing multiple distinct levels of representation raises immediate questions. How many levels are there, and what exactly do the different levels do?
Is it an accident that the levels enumerated in the previous paragraph correspond so
closely to the classical levels of linguistic analysis in structuralist and earlier linguistics? What can a level of linguistic representation look like?
Equally, questions arise immediately about interactions between representations,
now commonly discussed under the term linguistic interfaces: what does the output
of each level look like, and how does it feed into the next level(s)? How do the multiple
levels interact to produce the range of empirical phenomena that non-theoreticians
happily label with the umbrella term language?

Connectionism, in many ways the descendent of behaviourism, is similarly distinguished by its rejection of operations over the structured symbolic representations pervasive in classical cognitive sciencesee
Fodor and Pylyshyn () for extensive critical discussion.

Folli, Sevdali, and Truswell

The types of representation employed in linguistic analysis at different levels vary

quite substantially. Such variability is consistent with Fodors () description of the
modularity of mind, where modules are characterized partially by their informational
encapsulation: cognition is modular in part because information legible to one system
may be illegible to others. Such a state of affairs places the question of interactions
among linguistic representations (sometimes referred to as internal modularity), and
between linguistic and other cognitive representations (external modularity), at the
forefront of current theoretical concerns arising from the gross logical architecture of
the language faculty.
Our concern in this volume is primarily with questions of internal modularity,
and in particular the status of the module which is taken to underlie the linguistic
patterns studied under the rubric syntax. Syntax forms a core part of the debate
over the modular architecture of the language faculty, for several reasons. On most
assumptions about the architecture of grammar, syntax is in the centre, between
modules more closely related to sound and modules more closely related to meaning.
Moreover, because of this centrality, syntax is often also considered as more abstract
than semantics (doubtless grounded to an extent in non-linguistic conceptual structure) or phonology (plausibly grounded in phonetics). Finally, following Chomskys
foundational work, it is often assumed that formal properties of the syntactic component are responsible for the creativity of language, often described with reference
to Humboldts infinite use of finite media. This has entailed a continuous focus on
the description of the formal properties of syntactic structures, detailing not just the
creative potential of the syntactic component but also, following Ross (), the
restrictions on that creative potential.
Because of this focus on the formal properties of syntactic structure, there are
now two different objects which can be studied under the heading of syntax. One
is essentially descriptive: we study the attested and unattested orders of words in
phrases and sentences. The other starts from the formal characterization of syntactic
structures initiated by the likes of Chomsky and Ross, and asks which observable
phenomena fall under that formal characterization.
Navely, one might expect these two objects of study to coincide. In fact, though,
dissociations between the two were already observed in the s. For example, Miller
and Chomsky () suggested that an observable, descriptive fact about the order of
words in phrases (the impossibility of multiply centre-embedded relative clauses like
()) follows not from principles of sentence grammar but from the limits of the human
parsers ability to pair strings with structures.
() The rat the cat the dog chased killed ate the malt.
Conversely, and more controversially, but also closer to the concerns of this volume,
Generative Semanticists argued that locality constraints on unbounded dependencies
were also active in lexicosemantic composition of complex meanings from semantic


primitives, and in the calculation of scope and other relations (see Lakoff for
a summary). For example, McCawley (b) proposed that x killed y was derived
transformationally from an underlying structure like (a), but that there could be
no verb flimp with a meaning equivalent to (c), because combining kiss, girl,
and allergic transformationally into a single word meaning would violate Rosss
Complex Noun Phrase Constraint.
() a. cause x [become [not [alive y]]]
b. Bert flimped coconuts
c. Bert kissed a girl who is allergic to coconuts.
On the assumption that the locality constraints are basically syntactic, this underpins
an argument that the empirical domain of syntactic structures extends beyond the
order of words in sentences, to constrain the set of possible word meanings. If the
claims of Miller and Chomsky and of Generative Semantics both turned out to be
correct, we would have a double dissociation: there are phenomena (like centreembedding) relating to the distribution of words and phrases which are outside the
empirical domain of syntactic theory, and there are phenomena (like restrictions
on word meaning) within the empirical domain of syntactic theory, which are not
primarily concerned with the distribution of words and phrases.
Examples along these lines have multiplied in the intervening half-century. For
instance, it has become common, following Chomsky (), May (), Kayne
(), and Huang (), to assume that syntactic transformations derive a level
of Logical Form, representing basically semantic relations such as scope and binding.
Recently, however, more nuanced questions have arisen concerning the interaction
of syntax with pragmatics, and with notions of new and old information, aboutness,
and contrast, typically grouped under the rubric of information structure, following
Halliday (). It is not clear that information structure is a structure in the same
sense that phrase structure, prosodic structure, and so on are structures. Instead,
information structure currently appears to be used more as a cover term for a
(fixed, probably quite small, and probably universal) range of options for marking
distinctions within the semantic content of an utterance on the basis of the immediate linguistic and non-linguistic context (this may, of course, be a placeholder for a
more integrated theory of information structure developed along lines stemming ultimately from Kamp and Heim ). Several papers in this collection pursue the
hypothesis that information-structural notions are extraneous to syntax, constraining
the distribution of words and phrases from afar; others propose that informationstructural notions are more or less directly reflected in syntactic structure, as initially
proposed in detail by Rizzi (), allowing at least in principle for the possibility
that information-structural effects on well-formedness might be reducible to classical
syntactic explanations.

Folli, Sevdali, and Truswell

Likewise, the use of syntactic tools to constrain the distribution of sub-word-level

units, after a post-Generative Semantics hiatus, has flourished both in morphology
(following Baker and Pollock ) and in lexical semantics (following Hale and
Keyser and Marantz ). Meanwhile, a whole raft of researchers have taken
up the challenge of formulating non-syntactic descriptions of constraints on possible
assemblages of words into sentences. These constraints can be based on processing
considerations, following Miller and Chomskys lead (see also Bever , Frazier and
Fodor , Crain and Steedman , and countless others), but they can also be
based on semantic (Szabolcsi and Zwarts ), discourse-pragmatic (Erteschik-Shir
, Morgan , Kehler ), or other concerns.
Clearly, this reflects an analytical tension. If we think that some descriptive generalization G is due to syntactic factors, but it is demonstrated that factors apparently
extraneous to syntax influence the applicability of G, then there are three choices:
either we complicate our model of syntax to accommodate these new factors; or we
shrink the empirical domain of our theory of syntax by excising principle G, and leave
the challenge of accounting for the pertinent empirical facts to some other domain,
possibly with concomitant complications of the theory of that domain; or finally, we
leave the theories of the syntactic and non-syntactic components much as they would
be if considered in isolation, but build a more complex theory of the interface between
those two components. Generalizations like G are therefore particularly interesting
because they raise acute questions concerning the empirical domain of syntax and
its relation to other modules. Such generalizations may entail a degree of irreducible
complexity, but close analysis of such interface phenomena sometimes allows us to
address the overarching question: where exactly are the limits of syntax?
This is the central question addressed, in various ways and with respect to various
empirical phenomena, by the work presented in this volume. The answers are rarely
straightforward, but rather typically take the form of a trade-off: all else being equal,
do we end up with a simpler theory if we analyse the phenomenon in question in
syntactic terms, in non-syntactic terms, or at the interface? This must of course be
decided on a case-by-case basis, and so there is no pretence here (or anywhere) to
have provided a definitive characterization of the limits of syntax. Rather, the chapters
in this collection aim to provide precise and explicit hypotheses about the balance
between syntactic and non-syntactic modes of explanation in several areas where
this question is currently debated particularly intensely, including various facets of
the relationship of syntax to the lexicon, phonology, morphology, semantics, and
information structure.
Hand in hand with this question of the limits of syntax comes the question of what
is beyond syntax. What is the nature of the other levels of representation with which
syntax interfaces, and how do they interface with syntactic structures? Again, these
are big, open-ended questions, and no general theory currently even aims to address
them all in detail. However, they are crucial to the current enterprise: postulating a


non-syntactic theory of some phenomenon entails having a theory of the workings

of some non-syntactic module, and how it relates to syntax. Several chapters in this
collection reason in exactly these terms, by attempting to make precise the limit of
lexical influence on syntactic structure, by investigating the nature of the process of
mapping syntactic structure onto semantic or phonological representations, or by
clarifying the relationships between syntax, semantics, and information-structural
notions such as topic, focus, and contrast.
In the previous discussion, we have mentioned the centrality of syntax to questions
relating to language architecture. In the rest of this section, we concentrate on one
particular interface, namely the relationship between the lexicon and syntax, in order
to illustrate how close analysis of concrete empirical phenomena has been brought to
bear on broad architectural questions. As it transpires, this is an area in which ones
assessment of the empirical facts is particularly directly related to ones assumptions
about the modular architecture of the grammar.
Broadly speaking, two main views regarding the interactions of linguistic modules
currently predominate. One approach proposes that, although syntactic and other
representations do not interact directly, corresponendences between these representations are enforced by a superordinate bermodule, which interacts with syntactic,
semantic, and other representations in parallel. For example, the line of research
developing out of Montagues work (e.g. Montague ) emphasizes parallel computation of multiple representations, as exemplified in Montagues rule-by-rule correspondences between strings of terminal symbols and semantic representations. Today,
this approach is pursued in the different varieties of Construction Grammar (including Head-driven Phrase Structure Grammar as presented in Pollard and Sag ; the
Berkeley Construction Grammar of Kay ; Sign-Based Construction Grammar,
the successor of HPSG and BCG, as described in Boas and Sag ; and the Parallel
Architecture of Jackendoff ). The various forms of Categorial Grammar (e.g.
Steedman ) develop Montagues architecture in quite distinct ways, but share
the provision for the parallel computation of multiple representations, on the basis of
stored associations among syntactic, semantic, and phonological representations. For
such theories, then, the bermodule to which all others are answerable is the lexicon,
the repository of such stored information.
On the other hand, in architectures growing out of the Revised Extended Standard
Theory (e.g. Chomsky and Lasnik ), building on the logical necessity of lessthan-total encapsulation of the information manipulated by different modules, syntax
plays a central role, but it is not an bermodule. Rather, one or more phrase-structural
representations are generated, and phonological and semantic structures are derived
from those phrase-structural representations as permitted by mapping rules and constraints on the representations in question. In other words, each module uses its own
primitives and operations, but certain elements legible to one module must also be
legible to other modules: syntax generates representations which are legible in some

Folli, Sevdali, and Truswell

respects to the modules with which it interfaces, but it does not impose any constraints
on what those other modules subsequently do with the representations that it generates. In other words, the nature of the interfaces for such models is determined
by the information passed along from the phrase structure which is legible to the
phonological or semantic components.
Within this latter model, questions arise about the nature of the lexicon and the
extent of its influence over syntactic structure-building. The lexicon here does not
automatically have the bermodule status afforded it by Construction Grammar and
cognate theories, but there are still numerous reasons to suppose that the lexicon
is not simply a set of items serving as an input to a computational system which
manipulates those items. The first concerns semiproductivity among word forms, as
investigated by Jackendoff (), following Chomsky (). Many alternations look
more or less rule-governed, but admit several exceptions and idiosyncrasies: there are
cases in which a rule fails to apply, or the output of the rule displays irregularities
of form or meaning in some cases. In such cases, the regularity makes us want to
treat these alternations within the computational system (typically, syntax), but the
idiosyncrasies make us want to treat the alternations as stored.
A further compelling reason to suspect that the lexicon is less than maximally simple concerns the treatment of phrasal idioms. The phrasal syntax of idioms is generally
identical to the syntax of non-idiomatic constructions, which appears to suggest a
rule-governed compositional element to the structure of idioms. However, the interpretation of those idioms cannot be determined on the basis of the interpretations of
their component parts: even with apparently compositional idioms such as let the cat
out of the bag, although the cat is in some sense interpreted as the secret, this interpretation is only available in this particular phrasal context. For construction grammarians, such facts provide the impetus for essentially treating all syntactic composition
like the composition of idioms: the lexically guided construction of phrasal units is all
there is. Among Minimalists, in contrast, idioms are not treated as a reason to eschew
lexicon-independent principles of syntactic composition, but rather as an indication
that the meaning of a lexical item can be sensitive to syntactic context within a particular locality domain. For example, it is generally assumed, following Marantz ()
and Kratzer (), that there are no idiomatic external argumentverb combinations
Alongside these major architectural distinctions, numerous hybrids and alternatives exist. For example, the Parallel Architecture of Jackendoff (), as well as making reference to the lexicon as a Construction Grammar-like bermodule, also contains small interface modules that regulate correspondences
between the structures to which they connect. Unlike certain post-Montagovian theories, then, in the Parallel Architecture the derivations of phonological, syntactic, and semantic representations are not only related
indirectly, through the lexicon, but also through direct interfaces. Meanwhile, Representation Theory
(Williams ) postulates numerous levels of representation (at least eight), with no bermodule and no
module with the privileged status of syntax in recent Chomskyan theories, but instead a general constraint
on correspondences between levels, making reference to homomorphisms from one representation into
another and varieties of disruption of those homomorphisms. We set these aside here, although the debate
ultimately involves them just as much as the options outlined in the main text.


that exclude internal arguments (although see Nunberg, Sag, and Wasow for
possible exceptions). On the assumption that the verb is syntactically closer to the
internal arguments than to the external arguments, this could be taken to suggest a
syntactic locality constraint on idiom formation (see below for further discussion).
As with semiproductivity, as described above, there is a clear tension here: idioms
are largely syntactically regular, largely semantically irregular, and there are some
grounds for assuming that there is a syntactic limit on that semantic irregularity. Once
again, this empirical area highlights the complex interactions between lexical storage
of idiosyncratic material and rule-governed computation of complex structures.
How to handle this tension remains an open question. Two major approaches can
be defined. One, the lexicalist approach initiated by Jackendoff () and Wasow
(), and adopted in frameworks such as Lexical-Functional Grammar, relies on
devices such as redundancy rules performing a limited amount of pre-syntactic computation, to capture relations among lexical items without predicting the complete
productivity that we would expect from mechanical application of a regular syntactic
rule. The other, antilexicalist, approach, based on work by Hale and Keyser ()
and subsequently adopted by a range of Minimalist researchers, insists that all rulegoverned behaviour affecting both sound and meaning, even below the word level,
is basically syntactic, and that idiosyncrasies are confined to a special domain at the
bottom of a tree (the domain of L-syntax in Hale and Keysers terms).
The architectural commitments of lexicalist and antilexicalist theories are quite
divergent. For lexicalists (see e.g. Wasow , Reinhart , or Koontz-Garboden
, in addition to the above references), the lexicon is much more than a stored
list of correspondences between sounds, meanings, and syntactic features. Rather, it
has a fairly complex internal structure of its own, deriving many alternations between
word forms, with corresponding simplification of the phrase structure. In contrast,
Hale and Keysers alternative gives rise to the hope that syntactic notions will prove
useful for the analysis of lexical alternations as well as phrasal phenomena (see also
e.g. Travis and Kratzer ). One promising line of inquiry focusing on the
derivation of verb argument structure and stemming from Hale and Keysers work
aims to show that the syntactic position of arguments determines key aspects of their
interpretation, and so ultimately verbs at least in part mean what the syntax allows
them to mean (Rosen : ; see also Ramchand ).
Perhaps the poster child for the L-syntactic approach is the incorporation of locality
effects into theories of word-level structure, as anticipated in the Generative Semantics
literature mentioned above, and subsequent research like Bakers () description
of patterns of noun incorporation. Recent research has focused in particular on the
hypothesis that a particular head delimits the domain of L-syntax. This hypothesis

Following Ramchand (), the domain of L-syntax is sometimes referred to as the first phase, the
idea being that syntactic limits on lexicosemantic phenomena are a special case of the general notion of
phases as units of soundmeaning correspondence (Chomsky et seq.).

Folli, Sevdali, and Truswell

has in turn inspired many attempts to identify the head in question. For instance, that
head could be the same head which introduces an agent, the head which introduces
eventive semantics, the head which converts a category-neutral root into a verb,
some combination of these (if a single head has multiple functions), or none of the
above. All of these possibilities continue to be actively investigated (including by many
chapters in this volume), but none is currently universally accepted. The eventual
choice from among these options will directly inform our understanding of the syntax
of idiosyncrasy, and the idiosyncrasy of syntax.
These issues are particularly foregrounded by research in Distributed Morphology (DM: Halle and Marantz , Marantz ; see also Harley and Noyer ,
Embick and Noyer , and Harley, to appear, for recent overviews). DM adopts
the antilexicalist position, often referred to as syntax all the way down, that there
is no word-formation component distinct from syntax: morphemes are syntactic
terminals, and multimorphemic words are represented by particular configurations
of those terminals. More complex morphological phenomena are handled in a postsyntactic morphological component which allows for further manipulation of terminals prior to insertion of phonological forms, while interpretive phenomena are
handled by the Encyclopedia, which determines the interpretation of a terminal in a
given syntactic context. In a DM approach, then, the lexicon is not an bermodule: the
syntactic, semantic, and morphological computations retain some autonomy. However, in contrast to classical architectures from the Standard Theory through to GB,
the lexicons contribution is also not entirely pre-syntactic: it influences operations
in multiple places throughout the derivation. For DM, in other words, syntactic and
morphological structures are interrelated, because they are partially the same thing:
the generative system responsible for phrase structure is also partially responsible
for word formation. This makes it particularly natural to extend analyses based on
cyclicity and other syntactic locality conditions into the domain of classically lexical
phenomena; as a corollary, operations which clearly violate such conditions, such
as lowering of T onto V in English, must be treated post-syntactically, in a separate
module with its own rules (Embick and Noyer ).
Many chapters in this volume (particularly those in Part III) investigate this
range of architectural alternatives, arriving at pleasingly consistent conclusions on
the basis of diverse data. Many more chapters are interested in the other questions
raised above. The same goal runs through all the chapters, though: to find ways to
bring concrete evidence to bear on important but abstract architectural questions
and in the process to help us come a step closer to understanding the place (and
the limits) of syntax within the architecture of language. The rest of this introduction contains brief summaries of the empirical coverage of each chapter; the
point of the foregoing is to give a taste of why those empirical questions are worth


. Part I: Architectures
We begin with a series of chapters addressing architectural issues concerning the
nature of syntactic representations and the way in which they relate to representations
at other levels. Lechner (Chapter ) is concerned with the relationship between syntax
and interpretation. Building on Gajewski () in particular, Lechner distinguishes
two structural representations of meaning. One (the input to the deductive system, or
DS) represents abstract entailment relations, in the absence of nonlogical lexical items.
Optional operations, such as scope shifting, are all represented at DS. The second level
(Logical Form, or LF) provides the input to model-theoretic interpretation, and is
the locus of last-resort operations such as type-driven QR, where a quantifier moves
locally to avoid a type mismatch. DS and LF are serially ordered: DS precedes lexical
insertion and overt syntax, while derivation of LF follows overt syntax. The empirical
basis for assuming this architecture comes from a generalization, the Condition on
Extraction from Copies (CEC), which states that if a lower copy of a complex phrase
XP is interpreted, then elements contained within XP must be interpreted locally to
XP, and cannot, for example, participate in long-distance binding or scope inversion.
Lechner derives the CEC from the properties of DS and LF: in brief, the CEC holds
because independent conditions related to reconstruction mandate that movement
out of a silent copy must be (local) movement at LF, rather than (possibly non-local)
movement at DS.
Truswell (Chapter ) argues for a chain-based theory of reconstruction effects,
based on an examination of cases where reconstruction effects are found in the
absence of movement relations, and the dissociations between reconstruction for
scope and for binding first documented by Lechner (). In Truswells account,
each type of reconstruction effect is associated with a different type of chain, and the
partial autonomy of the two types of chain derives the dissociations. This approach to
reconstruction further suggests that movement (which, uniquely, displays both scope
and binding reconstruction) can be treated as a composite: movement is nothing
more than the cooccurrence of a scope chain and a binding chain. Although Lechners
and Truswells investigations are based on similar basic observations, their analyses
develop in quite different directions. Reconciling their approaches will be an interesting future challenge.
Lohndal and Samuels (Chapter ) are concerned with the relation between syntactic and phonological structures, and in particular with the analysis of certain
effects (such as the ungrammaticality of extraposition of a zero-relative across an
adjunct) classically reduced to the ECP. The authors group the cases in question
in an interesting way: they assume, building on Chomsky (), that adjuncts are
largely absent from the primary syntactic representation, and are only represented
by a null placeholder. These ungrammatical constructions are then distinguished by
structures conaining strings of adjacent null elements (including null placeholders

Folli, Sevdali, and Truswell

for adjuncts). They can then be prohibited as a special case of a theory of linearization
which cannot linearize two featurally identical sisters. Crucially, that theory is only
viable if a specific ordering of operations in the mapping from syntax to phonology
is assumed. In common with many recent theories of this interface, then, Lohndal
and Samuels advocate a sequential application of operations to transform phrasestructural representations into phonological representations, in which linearization
crucially precedes vocabulary insertion, integration of adjuncts, and marking of
copies for deletion.
Finally in Part I, OBryan, Folli, Harley, and Bever (Chapter ) address the question
of what kind of linguistic information is available during the first stages of sentence
processing. Evidence from three new experiments and a reanalysis of previously published results shows that garden path effects induced by a local ambiguity between
main clauses and reduced relatives (Bever ) are conditioned independently by
both transitivity and telicity. The former finding is quite well known and relatively
unsurprising; the latter is new and surprising, but amenable to explanation if telicity is represented syntactically, as suggested by Ramchand () and others, and
hence is available to the parser in the first stages of processing. This chapter therefore
contributes to the debate about the interaction of syntax and grammatically encoded
semantics in online syntactic processing.

. Part II: Syntax and Information Structure

Part II consists of three chapters exploring the relation between information structure
and syntax. Information structure is of particular interest because it spans at least
three intricately interacting types of phenomena (prosodic, syntactic, and discoursesemantic). A central question is whether information structure is a distinct level, or
an aspect of syntactically determined compositional semantics. The chapters in this
section (in contrast to Chapter , by Haegeman and Hill) all argue for the existence
of an independent, syntax-external information structure. Such a view is particularly
interesting as a counterpoint to the serialist, syntactocentric architecture typically
assumed in early Minimalist work on interfaces, and maintained here with respect
to distinct empirical questions by Lechner and by Lohndal and Samuels.
Eilam (Chapter ) proposes an analysis of focus intervention effects based on
constraints on information structure. His main argument rests on the demonstration of patterns of unacceptability in declarative sentences mirroring putative focus
intervention effects in A -dependencies, which would be surprising if the degradation
was related to the A -dependency. Instead, Eilam proposes a simple theory, building
on earlier work by Satoshi Tomioka, in which the degradation is due to presence
of multiple information-structural foci in an environment in which only one such
focus is permitted. This explains why certain information-structural manipulations


can ameliorate the examples in question, and also why the effects are not restricted
to interrogatives, instead showing up across a wider class of sentences with typically
fixed focus structure.
De Cat (Chapter ) argues for an information-structural characterization of root
phenomena, based on data from fragments containing dislocated topics in French and
from politeness markers in Japanese fragments, both restricted to root environments.
De Cat argues that these cases are not straightforwardly captured by approaches to
root phenomena based on clausal syntax, not least because of the existence of root
fragments with no clausal equivalent, and proposes instead an account based on the
hypothesis that root phenomena are degraded in structures that are thetic, and so not
articulated on the level of information structure.
Vermeulen (Chapter ) provides evidence that the mapping between syntax and
information structure is not fully determined by properties of syntactic structure or
information structure, but must instead be described with reference to the mapping
between the two levels. She assumes three primitive information-structural features,
[focus], [topic], and [contrast], which may be combined to represent contrastive
foci or contrastive topics. In Korean, any contrastive phrase, topic or focus, can be
scrambled to medial or initial position. The same is true for Japanese contrastive foci.
Japanese contrastive topics, however, behave like non-contrastive topics in that they
can only be scrambled to initial position. In other words, Korean contrastive topics
behave like other contrastive phrases, while Japanese contrastive topics behave like
other topics. The surprising architectural implication is that there exists a parameter
that distinguishes languages based not on factors related to syntactic structure or
information structure alone, but on the patterns of sensitivity of syntactic structure
to information structure, a conclusion which appears to fit most naturally with a
conception of these two levels as existing in parallel.

. Part III: Syntax and the Lexicon

Part III explores the relationship between syntax and the lexicon. The main questions,
addressed from many different angles, concern the limit of lexical influence on syntactic structure, and the amount of syntactic structure contained within a word.
Piggott and Travis (Chapter ) refine Piggotts previous work on underlying VV
sequences in Ojibwe. They claim that the choice between vowel deletion, consonant
epenthesis, and the status quo is determined by the interaction of the timing of Spellout. Piggott and Travis propose that Spell-out is triggered when internally complex
modifying heads are transferred from one derivational workspace to another, following Uriagerekas () treatment of phrasal Condition on Extraction Domain
(CED) effects. Because of this, vowels can remain in hiatus across a boundary between
modifier and modifiee: the two units are spelled out separately, and so the phonology

Folli, Sevdali, and Truswell

of the two units is determined independently. Consequently, phonotactic constraints

apparent elsewhere in Ojibwe do not apply.
Gehrke (Chapter ) and Arsenijevic and Simonovic (Chapter ) explore different
aspects of the contrast between lexical and postlexical word formation. Gehrke is
concerned with the distinction between adjectival and verbal passives in German,
standardly taken to reflect formation in the lexicon and in the syntax, respectively.
Gehrke shows that the diagnostics do not fully support this neat distinction. In particular, by-phrases, a hallmark of verbal werden-passives, are also possible in some cases
with adjectival sein-passives. Gehrke proposes instead a primarily semantic analysis: sein-passives, like werden-passives, are formed in the syntax, and the distinction
between lexical and postlexical passives is consequently rejected. Rather, sein-passives
are distinguished in that they describe event kinds (as independently diagnosed by the
distribution of so-proforms) and the modifiers possible with sein-passives are those
which are independently admissible in event kind descriptions.
Arsenijevic and Simonovic distinguish two types of de-adjectival nominal in
Serbo-Croatian: once again, one is more lexical and one is more syntactic. One
type (structurally complex de-adjectival nominals, or SDNs) is productive, syntactically complex, and stem-stressed. The other type (lexical de-adjectival nominals,
or LDNs) is syntactically less complex and insensitive to the surface prosody of the
base. Strikingly, the phonological differences between SDNs and LDNs do not always
correspond to the major semantic division Arsenijevic and Simonovic draw: between
nominalizations that can only describe tropes and nominalizations that can describe
tropes, events, or properties. Typically, SDNs must describe tropes, while LDNs are
less constrained. However, this is to some extent a blocking effect: in some cases, only
SDNs exist, and in those cases, the SDN is not constrained to describe a trope.
The papers by Anagnostopoulou and Samioti (Chapter ) and by Harley and
Stone (Chapter ) both approach the relationship between the syntax and the lexicon
from the same DM, syntax all the way down, perspective. Given common Minimalist
assumptions on the low portion of the clause assuming several functional heads above
V, possibly including a verbalizer vC, an element vE adding eventive semantics to the
root, and another head (Voice) related broadly to agentivity, it is natural to hypothesize
that one of these heads corresponds to the upper bound on possible idiomaticity.
Anagnostopoulou and Samioti and Harley and Stone investigate these hypotheses,
based on data from Greek participles and on idioms in English and other languages,
respectively. The two chapters therefore complement each other nicely, as one is concerned with word-level idioms and the other with phrasal idioms. Greek participles
are particularly useful for investigating these questions, because the morphology of
Greek sometimes gives clear indications of the relative height of attachment of verbalizing vC heads and participle-forming morphemes. The two chapters arrive at similar
conclusions. Anagnostopoulou and Samiotis Greek data show that neither vC nor vE
can be the upper bound on idiomaticity, but that Voice plausibly is that upper bound.


Harley and Stone, revisiting a seminal idea from Marantz (), reach the same
conclusion concerning agents, defending the No Agent Idioms Hypothesis against
a class of putative crosslinguistic counterexamples. Both chapters therefore suggest
that the introduction of functional structure is ultimately responsible for delimiting
the domain of idiomatic interpretation.

. Part IV: Lexical Items at the Interfaces

Our collection ends with a series of case studies describing the distribution of individual lexical items. Often, debates about the lexical items in question can be framed
in a broader perspective: which modules are responsible for constraints on a lexemes
First in this series, Pearson (Chapter ) focuses on the semantics of Japanese
comparatives formed with the postposition yori. Elaborating on Kennedys ()
distinction between explicit and implicit comparison, (which could be thought of
in terms of whether the standard of comparison is an argument or an adjunct at
LF), Pearson introduces a subsidiary distinction between weak and strong implicit
comparison, where strong implicit comparison makes no use of comparative morphology (as in English Compared to Mary, John is tall), but weak implicit comparison
does use comparative morphology (Compared to Mary, John is taller). Pearson shows
that, despite the absence of any overt comparative morphology, and despite the fact
that Japanese does have strong implicit comparatives (formed with kurabetara or
kuraberuto), yori-comparatives fall into the weak implicit class, thereby motivating
an argument for a null comparative morpheme in Japanese.
Alexopoulou, Folli, and Tsoulas (Chapter ) consider a number of differences in
the realization of noun phrases in Greek and Italian. Despite the similarity in the use
of Clitic Left Dislocation to mark topics, the two languages differ in the dislocation
of indefinite topics: in Italian, indefinite topics can undergo CLLD, while their Greek
counterparts can only be topicalized. This correlates with a number of other differences. Greek lacks a partitive construction and productively allows bare NPs, while
bare nominals are restricted in Italian, with partitives being used for the expression
of indefinite meanings. Secondly, Greek allows argument drop exactly for the range
of nominals that can be bare, while internal argument drop is unavailable in Italian.
Thirdly, Greek shows bare subnominal deletion, but Italian does not. Alexopoulou
et al. relate these differences to the structure of bare nouns in the two grammars and
to the semantic contribution of the Num head. They argue that Greek bare nouns
in argument positions are Number Phrases (NumPs) lacking a D head, an option
not available in Italian because of differences in the semantics of number in the two
languages. The Num head in Greek takes on the argumentizing role that D has in Italian, licensing bare singular and plural arguments. Therefore, they argue, the syntactic

Folli, Sevdali, and Truswell

behaviour of nominals in the two languages is partly determined by the semantics

of number in the two languages. Like Pearsons Chapter , this chapter highlights a
case in an unrelated language where a null morpheme has a crucial semantic effect,
affording weak implicit comparison in Japanese and type shifting in Greek nominals.
Daskalaki and Mavrogiorgos (Chapter ) argue against treatments of Greek indirect object resumption based on morphological manipulation of copies at PF, demonstrating that according to a number of morphological and semantic diagnostics, the
resumptive pronoun in question behaves distinctly from regular copies and behaves
instead like a clitic in clitic doubling constructions. Such an account argues for an
attempt to keep the distinction between movement and resumption in the syntax,
and contributes to the project of distinguishing between morphological form determined by post-syntactic manipulation of syntactic structure, and lexically determined
morphological form.
The chapters by Michelioudakis and Kapogianni (Chapter ) and Haegeman
and Hill (Chapter ) both deal with the syntax of non-truth-conditional elements.
Michelioudakis and Kapogianni develop an account of ethical datives, a problematic class of pronominals which have certain syntactic properties in common with
regular argumental clitics, despite not forming part of verbal argument structure.
Their contribution thereby hints at a mismatch between syntactic and lexico-semantic
forms of argument hood. The compromise that Michelioudakis and Kapogianni
endorse involves assigning a specific array of regular pronominal features to ethical
datives (which are responsible for most of the argument-like behaviour of the clitics),
while forcing Late Merge of the clitics, thereby keeping them separate from selected
Haegeman and Hills chapter examines the distribution of discourse particles in
Romanian and West Flemisha topic which raises issues related to the syntax
discourse interface discussed in Part II. In clear contrast to those earlier chapters,
though, Haegeman and Hill use facts pertaining to inflection on particles, selection
by particles, and word order restrictions among particles and vocatives to argue
for a syntactic treatment of the distribution of discourse particlesa phenomenon
which we might expect to be explicable in purely discourse-based terms. In particular,
their approach involves expanding the cartographic methodology of Rizzi () and
Cinque () beyond CP, to encompass an articulated left-peripheral series of Speech
Act Phrases, in a manner broadly reminiscent of Ross ().
Finally, Bjorkman (Chapter ) addresses so-called asymmetric uses of and, where
the coordination carries some extra temporal or causal implication above and beyond
logical conjunction. Contradicting the usual assumption that asymmetric uses arise
from pragmatic enrichment of a basic symmetrical coordination, Bjorkman shows on
the basis of distinct interpretations of CP- and TP-coordination that the distinction
between symmetric and asymmetric coordination is syntactically conditioned, which
is unexpected on the pragmatic enrichment account. This is used as one piece of


evidence supporting the construction of a primarily asymmetric semantics for coordination, where symmetric readings emerge in cases where the asymmetry has no
interpretive consequences. As with the preceding two chapters, then, closer inspection
reveals a syntactic basis to distributional facts where our intuitions may tell us to
appeal to discourse pragmatics.
What do we learn from this collection as a whole? One pleasing observation is
that relatively few substantial incompatibilities are in evidence, and they are outnumbered by the points of substantial consensus, despite the diversity of the data
under consideration. As ever in linguistic theory, the hypothesis space offered by
the near-simultaneous advent of the Minimalist Program, Distributed Morphology,
and Hale and Keysers approach to L-syntax is vast, and the evidence required to
discriminate between alternatives is subtle, and contingent on several other partially
theory-internal choices. It is always encouraging, then, to see independent lines of
argumentation, based on disparate empirical data, converge on the same conclusion
or point to a need for further investigation of a more complex set of phenomena, when
conclusions are not at hand. Of the two most striking examples of such convergence
here, one comes from Anagnostopoulou and Samiotis and Harley and Stones findings
concerning the role of agentivity in delimiting L-syntax. The other striking convergence concerns the general agreement on the utility of using a representation of information structure to constrain syntactic operations. There are many other points of
contact between the chapters, but these two stand out because of the uniformity of the
conclusions drawn, and the diversity of the evidence used to draw those conclusions.
In a sense, these are small victories, but because of the relatively limited options
available within a post-REST architecture, they have significant consequences. Anagnostopoulou and Samiotis and Harley and Stones findings significantly narrow the
options for a syntactic characterization of the scope of lexical operations, while the
postulation of a semi-autonomous representation of information structure functioning to constrain syntactic operations poses clear challenges to a literal interpretation of
the Y-model, according to which information flows from syntax to other models, but
not in the other direction. Once again, options are available to preserve the Y-model,
but such evidence serves to narrow down those options. One important message is
that the investigation of how the different components of a modular grammar interact
is crucial to our understanding of the workings of the language faculty in general, if
we are to account for fine-grained aspects of grammatical phenomena. Overall, the
volume represents an example of classic generative methodology in practice: gradual
progress resulting from strong hypotheses confronted with careful empirical crosslinguistic investigation.

Part I

Harmonic Derivationalism

. Introduction
On a widely held view, cognitive systems consist of discrete, informationally encapsulated subcomponents or modules (Fodor ) operating on mental representations.
As the modules can contain partially incompatible types of information, their expressions need to be modulated by interfaces in order to render representations of one
system interpretable by its neighbouring modules. For the language faculty, the two
central syntax-external modules are model-theoretic semantics and the sensorimotor
system, which are accessed by their respective interfaces, Logical Form and Phonetic
Form. Recently, it has been argued that the modules interfacing with the core syntactic system also include a designated system computing logical inferences, called the
Deductive system (Fox , Gajewski ).
The present chapter outlines a new architecture of the grammar that redefines the
way in which the logical syntax of meanings is derived. Concretely, it is suggested
that the model of the grammar includes two covert components instead of a single
one for deriving the abstract representations which are submitted to model-theoretic
interpretation: an interface to DS, which precedes the overt syntax system, and Logical
Form (LF) which directly interfaces with the semantic component. In what follows
I will, unless ambiguity arises, refer to the DS interface simply as DS.
The assumption of two covert components is justified by empirical as well as theoretical considerations. On the empirical side, the DSLF model is supported by the
fact that it offers an explanation for a class of structural restrictions on covert remnant
movement identified in Lechner (, b), where they are referred to as the
Condition on Extraction from Copies or CEC. Theoretically, the new model offers
I am grateful to the organizers of OnLI II, Raffaella Folli, Christina Sevdali, and Robert Truswell,
for their kind invitation, as well as to the audiences of OnLI and GLOW , Vienna, for valuable
feedback. Robert Truswell provided detailed written comments on an earlier version which helped me to
avoid an inconsistency and led to a radical revision of section .. I am also indebted to Anke Assmann for
email discussion on the cycle and predicate fronting. The title is inspired by the term Harmonic Serialism
(McCarthy ). Finally, responsibility for any errors remains with the author.


at least two attractive features. First, it locates covert scope-shifting operations and
movements that solely serve the purpose of local type adjustment in two different
parts of the grammar, resulting in a clean divide between movement that ensures
interface readability and long-distance procedures which potentially create new interpretations. Second, it will be seen that DS is integrated into the stem of the derivation
by an operation which is already widely used within the syntactic component in
order to model syntactic opacity phenomena (such as Binding Theory obviationsee
below): late insertion of lexical nodes, or Late Merge. Widening the role of Late Merge
to apply not only within narrow syntax but also across modules leads to an increase
of the overall symmetry of the system, while at the same time making optimal use of
the analytical tools already employed in the analysis of opacity.
The chapter is structured as follows. Subsequent to some general remarks on the
relation between interfaces and opacity in section ., section . briefly reviews the
empirical basis for the CEC generalization. Section . attempts an explanation for
the CEC. The analysis to be submitted crucially relies on the new DSLF model of the
grammar together with standard syntactic mechanisms such as Late Merge and the
cycle. Some ramifications of the model will finally be considered in section ..

. Opacity
As opacity occupies a prominent position throughout the discussion of this chapter,
the current section makes explicit some background concepts as well as the probably
not so familiar notion of opacity at the interfaces.
In principle, linguistic interfaces can modulate the signal which is passed from
one system to the next in two ways: qualitatively by signal transduction, or
quantitatively by altering the amount and internal composition of the information
to be transmitted. The current discussion is concerned with changes across the
quantitative dimension only. These changes determine (i) how much information is
transferred at a given point (amount); (ii) at which point the information is released
(timing); and (iii) where this information is precisely located in the representations
built from the signal (location).
() Quality
(changes in type
of information to
be transferred)

Amount (How much information is transferred?)
Timing (When does the transfer take place?)
Location (Where is the information located in the

Examples of signal transduction include microphones, which translate acoustic waves into electrical
signals, or scales, which transduce from mass/weight to numbers.

Harmonic Derivationalism

Quantitative factors do not only have an impact on information transfer across modules, but are also at work within an encapsulated component. It has been argued, for
instance, that information in the syntactic system is packaged in discrete units (factor
(i)), and that sequentially ordering these representations leads to the effect of syntactic
derivations (factor (ii)). In order to show that the computation indeed proceeds in
discrete, derivational steps, it is moreover necessary to find constellations in which
the units are ordered in a particular sequence, subject to conditions which cannot be
inferred from the output. Such constellations, known as opaque contexts, have been
taken to be strongly indicative of an incremental, derivational computation.
The study of opacity in linguistics has generally been restricted to phenomena
located within a specific linguistic component such as syntax, phonology, or morphology, where it has been instrumental in guiding the decision among competing models
of the grammar. According to the standard typology of Kiparsky (), opacity
comes in two varieties, which in modern parlance are usually referred to as over- and
underapplication (for important qualifications see Bakovic ). Overapplication
describes scenarios in which a principle or rule applies even though the context of
that condition is no longer visible in the surface form. By contrast, in environments
that are shaped by underapplication, a principle that should have applied, given the
overt context, failed to do so. Two typical examples of over- and underapplication
in syntax come from the interaction of Binding Theory and movement. Principle A
reconstruction as in () is a manifestation of overapplication because the context of
variable binding, that is c-command, is arguably not met any more subsequent to
() Which book about himself did no one like?
Overapplication is standardly modelled by assuming that the signal is richer than
it appears, and includes devices such as silent lower occurrences of higher nodes
(copies: Chomsky ).
With the assumption of copies, () ceases to be opaque. However, the analysis now
turns Principle C obviation by A-movement, as in (), into an instance of underapplication. Principle C is expected to affect the name in the lower copy located in the
position marked t , barring coreference, but fails to do so. (Copies will be presented
as traces for ease of readability.)
() [This picture of John ] seems to him t to be beautiful.
A popular way to resolve this conflict consists in delaying lexical insertion of the
offending parts (picture of John) to a point of the derivation where the name (John) is
On the derivation vs. representation debate see, among many others, Brody (), Haider (),
Lechner (to appear), Williams ().
For details of the semantics of variable binding under reconstruction see Sauerland (). For alternatives, that do not require reconstruction, see Sternefeld and Konietzko ().


no longer c-commanded by the pronoun he (Wholesale Late Merge; Takahashi ,

extending Lebeaux ).
More specifically, Takahashi () advances the hypothesis that the upper limit on
insertion of nominal restrictor arguments is delimited by abstract Case in such a way
that NP-restrictors need to be added inside their Case domains. Thus, objects have to
be fully assembled inside vP, accounting for obligatory binding reconstruction with
wh-movement in ():
() [Which picture of John ] did she say that he liked t best?
By contrast, Late Merge of the restrictor inside finite subjects can be delayed to
Spec,TP. For the raising configuration (), this entails that subjects which have undergone A-movement only need to be fully represented in their higher Case position.
Thus, Takahashi derives classic instances of syntactic opacity from the interaction of
two analytical strategies: copies and late, delayed lexical insertion by Late Merge.
Opacity runs as a red thread through this chapter in various different guises. To
begin with, parts of the discussion will be concerned with well-known analytical tools
for modelling opacity in syntax, which include late insertion, copies, and semantic
reconstruction by higher type traces. In what follows, it will be argued that manifestations of opacity similar to those attested within core syntax can also be observed across
components, specifically between syntax, DS, and LF, revealing structural similarities
between inter- and cross-componental relations which have, to my knowledge, not
been made explicit in the literature so far. In short, such constellations of opacity at
the interfaces arise whenever the hallmarks of under- or overapplication are detected
across a sequence of operations which spans more than a single component. Recognizing these two distinct types of opacity is relevant inasmuch as both can be modelled
by the same tools once the DSLF model is adopted, furnishing strong support for this
revised architecture of the grammar. I will come back to further details, following a
discussion of the empirical basis for the DSLF architecture in the next section.

. The Condition on Extraction from Copies

The present section reviews two empirical arguments for the generalization in ().
() Condition on Extraction from Copies (CEC)
Covert subextraction out of silent copies is as local as possible.
The arguments, to be discussed in turn, come from DP-reconstruction and scope
freezing with predicate fronting.

Further evidence and discussion of the CEC can be found in Lechner (, b, to appear).

Harmonic Derivationalism

.. Reconstruction
While movement can be undone for the computation of both relative quantifier scope
and binding relations, it has been observed that scope and binding reconstruction
do not necessarily coincide, indicating that reconstruction involves an additional
mechanism apart from copies (see also Truswell, Chapter below). For instance, short
scrambling in German illustrates that scope reconstruction does not entail binding
reconstruction. Even though (a) admits an inverse reading for the scrambled direct
object QP , it is not possible to construe the anaphor inside the scrambled object as
being bound by the indirect object to its right (Lechner , ). (b) demonstrates that in the absence of a higher binder, reconstruction results in ill-formedness,
corroborating the hypothesis that short scrambling cannot be undone for binding
(Frey ).
() a. weil wir [einige Freunde von einander/ ] allen t vorstellen wollten
to all introduce wanted
since we [some friends of each other
since we wanted to introduce some friends of each other to everybody
( > / > )
b. weil ich [einige Freunde von einander/ ] allen t vorstellen wollte
since I [some friends of each other to all introduce wanted
since I wanted to introduce some friends of each other to everybody
The bifurcation between binding and scope in (b) can be given a natural analysis
on the assumption that binding relations are structurally encoded at LF and that
scope diminishment is delayed to the semantic component, such that the scrambled
object reaches its scope position too late for the principles of Binding Theory and
variable binding to apply. Implemented in terms of Semantic Reconstruction (SemR;
Cresti , Rullmann , von Stechow ), as in (), the inverse reading of the
scrambled QP is derived by having QP bind a generalized quantifier-type trace into
which QP is converted in the inverse reading:
() We introduced [[DP some friends of each other ] [to everybody [T,et,t ]]]
Contexts which lend themselves to an analysis in terms of SemR are opaque, and they
are so in two different ways. Note to begin with that a derivation is always opaque
relative to a given representation. If this representation is taken to be the surface form,
SemR represents a case of rule overapplication, because the principles responsible for
determining scope locate the scope of the fronted quantifier in a configuration which
has been destroyed by movement. But SemR can also be seen as a manifestation of
underapplication. If the structural description includes copies in the source position
of movement, a principle (Principle A) that would have been expected to apply fails
See also Truswell (Ch. below), who discusses interesting challenges for the mixed approach towards
reconstruction coming from unexpected reconstruction effects in control contexts.


to be visible in the output. The relations can also be stated as follows: SemR describes
overapplication with respect to quantifier scope, and underapplication with respect
to Principle A.
Moreover, opacity in both cases results from the interaction between rules that
belong to two different grammatical components: movement applies in overt syntax,
while reconstruction is delayed to semantics. Since opacity is diagnostic of a serial rule
ordering, this cross-componental opacity presents a strong argument for the view that
syntax precedes semantics, in support of a serial, syntacto centric architecture of the
grammar (Chomsky ). By contrast, it is not evident how the phenomena characterized by the hybrid reconstruction approach can be made to be compatible with
monostratal models such as categorial grammars in which syntactic and semantic
representations are built up simultaneously.
While () presents a strong argument for a theory of reconstruction which admits
both SynR by copies and SemR, the hybrid approach is challenged by the observation
that there are contexts in which SynR and SemR systematically co-vary, in support of
a pure SynR approach (Fox , Lebeaux , Romero ). Such Scope Trapping
environments appear to suggest that the hybrid approach overgenerates, because it
admits unattested combinations of wide binding scope (of situation variablessee
below) and narrow quantifier scope. The trapping generalization relevant for present
concerns contains the two clauses in ():
() a. If a moved DP is construed de dicto, it reconstructs for Binding Theory.
b. If a dislocated DP reconstructs for Binding Theory, it is construed de dicto.
The paradigm in () confirms the validity of (a) (Lechner , to appear, Sharvit
). Both examples, modelled after Russells () yacht sentences, are ambiguous
between a consistent de dicto interpretation for the raising subject and a contradictory
de re reading. Interestingly, in (a) the fronted name John can be construed coreferentially with the pronominal experiencer him only on the contradictory de re reading
(cf. control (b)). This follows on the assumption that the de dicto construal is produced by binding the situation variable (s-variable) inside the subject Johns height
to the raising predicate, which makes the referentially opaque reading contingent
upon subject reconstruction into a position below seem. Relevant parts of the LF
representation are made explicit in (), while () provides details for the de re
() a. [John s height] seemed to him to exceed his actual height.
( consistent de dicto/contradictory de re)
(i) de dicto: It seemed to John that John is taller than he actually is.
(ii) de re: John obtained the impression: I am taller than I am.
b. [His height] seemed to him to exceed his actual height.
(consistent de dicto/contradictory de re)

Harmonic Derivationalism

() Consistent de dicto reading of (a) violates Principle C

a. [John s height]de dicto seemed to him to exceed his actual heightde re
b. s [seemed [s to him [John s height-in-s ] to exceed his height-in-s ]]
() Contradictory de re reading of (a) does not induce Principle C effect
a. [John s height]de re seemed to him to exceed his actual heightde re
b. s [[John s height-in-s ] [seemed [s to him to exceed his height-in-s ]]]
Russell ambiguities also supply a test for (b). Specifically, (a) can only be understood as a consistent de dicto proposition, indicating that anaphor reconstruction
results in narrow s-binding ((b); (b) serves as control). Thus, if a moved DP
reconstructs for the anaphor licensing, the s-variable must be locally bound to the
next intensional predicate (Romero ):
() a. [Each other s height] seemed to the boys to exceed their actual height.
(consistent de dicto/ contradictory de re)
(i) de dicto: It seemed to each boy that the others are taller than they
actually are.
(ii) de re: Each boy had the impression: the other boys are taller than they
b. [Their height] seemed to the boys to exceed their actual height.
(consistent de dicto/contradictory de re)
() Consistent de dicto reading of (a) abides by Principle A
a. [Each other s height]de dicto seemed to the boys to exceed their actual
heightde re .
b. s [seemed [s to the boys [each other s height-in-s ] to exceed their
height-in-s ]]
() Contradictory de re reading of (a) violates Principle A
a. [Each other s height]de re seemed to the boys to exceed their actual
heightde re .
b. s [[each other s height-in-s ] [seemed [s to the boys to exceed their
height-in-s ]]]
The findings above entail that the hybrid theory can be contained from overgeneration
only if two requirements are guaranteed. First, reconstruction by SemR always has
to result in de re interpretations. Otherwise, it would be possible to derive combinations which reconstruct for scope and referential opacity (generating de dicto
readings) but not for Binding Theory. But such constellations are incompatible with
(a). Conversely, syntactic reconstruction by Copy Theory must invariantly produce


a de dicto reading. Otherwise, movement could reconstruct for binding and scope
while preserving referential transparency (de re interpretation), in contradiction
to (b).
The first task, i.e. ensuring that SemR is compatible with de re interpretations only,
falls out from a general principle restricting higher-type traces to extensional types.
For further details see Lechner (to appear). It is the second condition where we finally
encounter a first manifestation of the CEC, repeated from above:
() Condition on Extraction from Copies (CEC)
Covert subextraction out of silent copies is as local as possible.
Essentially, what (b) and the paradigm in () demonstrate is that a successful theory
of reconstruction must be able to block long-distance binding of s-variables out of
reconstructed subjects, in order to exclude LF-representations such as (), which
underlie the unattested de re reading.

[seemed [s to the boys [each other s height-in-s ] to exceed their

height-in-s ]]

On the assumption that s-variable binding comes about by silent movement of a

situation pronoun (Percus and Sauerland ), () is straightforwardly excluded
by the CEC. In (), variable s has moved across a closer potential binder (s ) in
violation of the minimality condition of the CEC. Thus, the CEC derives an important
restriction on syntactic reconstruction.
.. Predicate fronting
Further empirical motivation for the CEC from a different direction comes from
scope freezing effects with predicate fronting. While (a) is scopally ambiguous, VPtopicalization in (b) makes unavailable the inverse scope interpretation (Barss ,
Huang , Sauerland and Elbourne ):
() a. No one will teach every student
b. and [vP teach every student], no one will t

( > / > )
( > / > )

Assume that inverse scope with distributive operators is the result of covert scopeshifting operations (Fox ) and that object QR is triggered by the need to
avoid type incompatibility (Heim and Kratzer ). Then, the object wide-scope
reading for (b) is contingent upon wide object QR out of the lower vP copy
across the subject, documented in (b). But this movement operation is not the
most local one, because short QR across the subject trace, as depicted in (c),
This requirement needs to be secured both by the hybrid and the pure SynR approach, and is therefore
independent of the decision between these two theories.

Harmonic Derivationalism

also yields an interpretable output representation. Thus, wide QR is prohibited by

the CEC.
() a. and no one will [vP t teach every student ]
(LF, after vP-reconstruction)
b. and [every student [no-one will [vP t teach t ]]]
(blocked by CEC)
c. and [no-one will [every student [vP t teach t ]]]
(admitted by CEC)
Just as with DP-reconstruction, the CEC can be seen as a tool to model opacity,
or more specifically underapplication. The representation (a), which is an output
configuration from the perspective of the interpretive branch of the model, fails to
license an operationscope-shifting QRthat should have been able to apply given
the information supplied by the local context. It is only by looking at the derivational
history that it becomes possible to define the pertinent restriction. Thus, the LF (a)
is opaque. Note finally that the CEC applies only to silent copies. As a consequence,
scope shifting in regular transitive environments (a) is not subject to the stricter
locality conditions imposed by the CEC.
To summarize, it was argued that covert movement operations are subject to a
new condition limiting movement out of silent copies, the CEC. The CEC in essence
describes contexts that fit the profile of underapplication opacity: an operation that
would in principle be expected to affect a given representation at LF fails to do so, and
it fails to do so due to properties visible only at earlier stages of the derivation (overt
syntax). Next, it will be seen that the most plausible explanation for the CEC resides
with a new architecture of the grammar, and not with a reduction to other, known
principles within syntax.

. Analysis
Various analytical options can be explored in looking for an explanation of the CEC,
among others the candidates in ():
() Possible explanations of the CEC
a. Silence of the containing node blocks non-local movement
b. Properties of the movement procedure
c. Feature structure of silent copies (locality as a result of an impoverished
feature inventory on silent nodes)
d. Properties of the linearization procedure (locality as a result of order preservation)
e. Architecture of the grammar: the covert component is divided into a part
which admits non-local movement and a part which only licenses local


Hypothesis (a) is not likely to succeed given that elliptical VPs are silent, yet
nonetheless admit non-local movement (An American flag is hanging in front of every
embassy, and a British one is, too). Options (bd) are either evidently too strong
(b) or vacuous, in the absence of further details. A plausible venue is given by
hypothesis (e), though, according to which the stricter locality conditions expressed
by the CEC reflect the point in the derivation at which these movements occur.
With the information provided in section ., it also becomes possible to be more
specific about how the local processes should be separated from the non-local ones.
In essence, the CEC defines a last-resort strategy for movement, barring all optional
displacement which is not motivated by rendering representations interpretable at the
semantic interface. Moreover, CEC-conforming operations are local since Economy
dictates that they target the closest possible landing site which guarantees interpretability. On this view, the CEC is not a construction-specific principle of the
grammar, but rather exposes a fundamental characteristic of the interface with modeltheoretic semantics: LF admits only local operations which serve to repair type mismatches and possibly other processes that are needed to secure LF-transparency (in
the sense of von Stechow ).
The present proposal entails two consequences. First, it presents a more natural
conception of the interface with model-theoretic interpretation than the orthodox
version of LF, which includes both local adjustment operations and non-local scopeshifting movement. Second, it follows that since local CEC-conforming movement
applies at the LF interface and LF is the last component in the stem of the derivation,
optional, non-local scope-shifting operations must precede LF. Thus, covert movement that applies early in the derivation is subject to more liberal locality conditions
than covert operations that apply late. The grammar therefore has to make available a second silent component apart from LF for hosting non-local, scope-shifting
A potential candidate for such a second silent component emerges in form of the
deductive system (Fox , Gajewski , Fox and Hackl ), or, to be precise,
the grammar-internal interface to DS. The following subsection briefly characterizes
some core features of DS and suggests a new way to integrate the DS interface into the
architecture of the grammar.
.. The DSLF model
Fox () introduces DS as a formal calculus that derives logical inferences which
are, among others, used by the economy metric regulating possible scope relations
(Scope Economy). A conception of DS particularly conducive to the one to be used
below can be found in Gajewski ().
According to Gajewski, the analysis of various semantic phenomena, including the
definiteness restriction and well-formedness conditions on exceptives, needs to make

Harmonic Derivationalism

reference to properties of the logical skeleton of the expressions involved. Roughly,

the logical skeleton is obtained by replacing all non-logical, non-permutationinvariant constants by variables of identical type, as illustrated for the existential
construction in ().
() a. There is no man/a man/ every man
b. Logical skeleton: [theree [be [no/a/every Pe,t ]]]
Gajewski explores the idea, originating with Barwise and Cooper (), that illformedness ensues if the logical skeleton of an expression yields non-contingent
truth-conditions under all possible assignments, resulting in l(ogically) analytical
statements. For instance, () can be shown to be l-analytical just when a strong
determiner is inserted in the logical skeleton (b).
The concept of the logical skeleton dovetails with a group of proposals which
have recently been advanced for various domains according to which certain formal
principles of the grammar operate on impoverished representations. Such incomplete linguistic objects can either be categorially underspecified (Richards ,
Lechner ) or they may lack part of their descriptive content. Also, different
theories have operationalized lexical impoverishment with the help of different analytical strategies, including late insertion in Distributed Morphology and countercyclic (Wholesale) Late Merge in the analysis of Principle C (Takahashi ; see ()
There is now an intriguing synergy between DS and Late Merge, which can be
exposed by assigning to DS the role of a second, grammar-internal covert component
with three properties: (i) DS precedes overt syntax; (ii) scope-shifting operations apply
at DS (if licensed by Scope Economy); (iii) the impoverished representations generated at DS are supplied with lexical content by Late Merge in course of the syntactic
derivation. In short, Late Merge fills the logical skeleton with lexical content. For
reasons of expository convenience, I will from here on reserve the acronym DS for
the grammar-internal component and refer to the language-external deductive system
in the sense of Fox and Gajewski by its full name (by analogy with the distinction
between the linguistic level LF and the non-linguistic, philosophical notion logical
At the moment, the exact nature of the formal objects DS operates on is not entirely
clear. In principle, DS representations could be taken to correspond to logical skeletons, as alluded to above. On this view, scope-shifting operations at DS change the
order of quantified expressions whose restrictors are property variables throughout
DS, to be lexicalized by Late Merge only once the derivation reaches overt syntax.

See Barwise and Cooper () for existentials; see von Fintel () for exceptives.

Gajewski () attributes an idea along these lines to Danny Fox.


While this implementation offers the advantage of optimizing the interface with the
(grammar-external) deductive system, as the latter operates on logical skeletons, I will
adopt the more liberal position that lexical insertion by Late Merge can take place at
any point of the derivation, throughout DS and overt syntax. This has the advantage
of rendering the integration of DS representations into the overt syntactic system
more efficient and simpler because movement copies can be generated by standardly
sanctioned mechanisms. Hence, in what follows it will be assumed that DS representations canbut do not have tobe completely lexically impoverished. Other
options are conceivable, but as far as I can see, the particular analyses to be advanced
below do not inherently depend on any particular choice as to how representations are
Subsequent to DS and overt syntax, the derivation reaches the model-theoretic
interface LF. Whereas DS operations are by nature non-local, as they test for different
scope options, all LF movement, prototypically exemplified by type-driven QR, is
strictly local and exclusively serves the purpose of rendering representations readable
at the interpretive interface. Thus, the resulting DSLF model, schematized in (),
partitions scope-shifting QR and operations required for type repair into two separate

The DSLF model

Lexical Insertion by (Wholesale) Late Merge
DS Overt Syntax


optional scope-shifting rules

LF model-theoretic interpretation

obligatory local adjustment
deductive system
rules (type-driven QR, . . .)
With the assumptions about the DSLF model above, the sample derivation for a
scopally ambiguous example such as () proceeds as follows. Assume for reasons
of concreteness that in the case of (), the DS representations resemble the logical
skeleton. The DS for the surface scope reading, given in (), then includes two
determiners and variables in the position of the restrictor arguments and the verb.
In addition to the logical skeleton for the overt scope order, the model also makes
available a second DS representation for the inverse reading (b), in which the object
has been exported across the subject:
() Two policemen spied on every boy.
a. [every Qe,t [two Pe,t [Re,et [every Qe,t ]]]
b. [every Qe,t [two Pe,t [Re,et [every Qe,t ]]]]

Harmonic Derivationalism

Scope Economy now tests whether the wide-scope option derives a new interpretation
that is logically independent from (a) (Fox ). To that end, both (a) and (b)
are transferred to the grammar-external deductive system. Since the external system
confirms the logical independence of the two readings, both DS representations are
further processed by overt syntax. Note that the scope of an expression is not fixed
or interpreted at DS. DS merely explores which among the logically possible scope
relations are legitimated by the grammar-external deductive system, passing licit
configurations on to overt syntax and, eventually, to LF.
In overt syntax, roots and functional morphemes, including formal syntactic features, are merged. For the surface scope reading, there is only a single option for lexical
insertion, detailed in (), which correctly derives scope as well as word order:
() Lexical insertion in object narrow-scope representation

[two Pe,t [Re,et [every Qe,t ]]]

two policemen spied on every boy

As for the inverse reading, it is at least at first sight tempting to correlate long object
movement at DS with Late Merge of the restrictor into the scope position of the object,
as in ():
() Lexical insertion in object wide-scope representation
a. [every Qe,t [two Pe,t [Re,et [every Qe,t ]]]]
b. every boy [two policemen spied on every Qe,t ]
However, this simple conception is contradicted by two facts. First, QR does not
obviate Principle C effects ( Two policemen spied on every boy who trusted the guys ).
Hence the restrictor must be inserted low, inside the VP (for motivation see (c),
below). Second, Late Merge into the scope position fails to account for the correct
linearization, because QR does not have an effect on word order. Thus, the system has
to insert the descriptive content in a cyclic fashion, starting with the lower occurrence
of the object, proceeding from there to the higher one, as in (). The higher copy is
then marked as PF-inert (typographically symbolized by strikeout), just as in more
traditional single output models that locate all movement operations, irrespective of
whether they are overt or covert, in the stem of the derivation (Bobaljik , Groat
and ONeil , Pesetsky ).

Transfer to the grammar-external deductive system minimally involves translation of the logical vocabulary of a natural language into the logical constants of a formal system. Alternatively, one could assume
that DS includes logical constants only (e.g. symbols such as ), and that the morphophonological matrix
(every) is provided by Lexical Insertion.


() Lexical insertion in object wide-scope representation

a. [every Qe,t [two Pe,t [Re,et [every Qe,t ]]]]
b. every boy two policemen spied on every boy
More complex structures such as subject-to-subject raising and other long-distance
displacement phenomena require additional assumptions about the mechanisms generating DS representations, some of which will be made explicit below. A more complete study of the typology of movements in the new architecture has to await another
occasion, though. The next section precedes to the analysis of CEC effects in the
DSLF model.

.. The DSLF model and the CEC

Turning first to scope freezing with VP-fronting, recall that the CEC admits local QR
of QP below the subject QP , schematized in (a), but bars wide, scope-inverting
QR across the subject QP , as in (b):
() a. [VPPF [QP [TP QP [QP [VP t ]LF ]]]
(local QR out of reconstructed VP)
b. [VPPF [QP [TP QP [QP [VP t ]LF ]]]
(wide QR out of reconstructed VP)
As will be seen below, scope freezing falls out from the claim constitutive of the DSLF
model that LF operations are strictly local, together with the assumptions in () and
(). The propositions in () are known from the literature. I will briefly comment
on each in turn.
() a. The Extension Condition (Chomsky ): non-LF movement extends the
b. Lexical Insertion can take place throughout the derivation, at DS or in overt
syntax. LI can in principle target any position of a chain.
c. Takahashi (): LI into a node is admissible only if is c-commanded by
its Case-assigning head (see also Takahashi and Hulsey on Wholesale
Late Merge).
d. Takahashi (, ): Non-lexicalized chain positions are filled by (possibly silent) determiners. These positions are converted into individual variables in semantics.
() Uniformity Condition (UC)
If a node is fully lexicalized, and is the highest node of a movement chain,
c-commands its lower occurrences throughout the derivation.

Harmonic Derivationalism

(b) captures the default hypothesis that LI is free to apply at any point of the overt
derivation, subject to additional conditions. These conditions are spelled out in (c)
and (d). Following Takahashi, I assume that restrictor arguments of determiners
can be added only within their Case domain ((c); for evidence see discussion of ()
above). Moreover, if a determiner remains unlexicalized throughout the derivation,
it is translated as an individual variable (d). This requirement, which arguably
follows from the hypothesis that natural language quantification is restricted, ensures
that bare determiner insertion precludes not only syntactic but also semantic reconstruction. Finally, the only new addition to the catalogue of principles regulating
the interplay between LI and the derivation is the Uniformity Condition (UC), ().
The UC in () is meant to express the idea that DPs which are fully assembled and
move need to c-command their lower copies throughout the derivation. In essence,
the UC is a descendant of Fiengos () Proper Binding Condition or some version of the cycle, relativized to DPs that are complete in the sense that both the
determiner and the restrictor NP have been merged. As this in turn presupposes
that both D and NP have been assigned Case by a higher c-commanding head, ()
is intuitively closely related to Takahashis Case condition (c). Both conditions
impose c-command requirements on elements that are in need of Case. Hopefully,
future research will reveal directions towards a unification of these two principles (see
Lechner a).
Returning to the analysis of scope freezing with VP-fronting, consider how the
system excludes the unattested object > subject scope order first. For ease of readability, I will subscript DPs which include a lexical restrictor with LI and nodes that
consist of D only and are therefore translated as traces (i.e. individual variables) with
t. Offending constellations are underlined. Furthermore, nodes are labelled to guarantee better orientation in the tree even though, strictly speaking, DS is categorially
In (a), the object has shifted over the subject at DS, landing in a position above TP.
Subsequent VP-fronting in (b) alters the hierarchical relations between the object
and its copies, as the TP-adjoined occurrence of every student no longer c-commands
the copy inside the higher occurrence of VP. As a result, constellation (b) violates
the UC:

On current assumptions, determiners are translated as bare individual variables (d). Hence, the UC
is a c-command condition on copies, and not on variables. This is interesting because it is usually believed
that the Proper Binding Condition and other generalizations about remnant movement can be reduced
to conditions on licit variable binding configurations (a variable needs to be bound at LF/throughout the
derivation). If the UC is on the right track, it falsifies conditions on variables-style analyses.

( > / > )

() and [vP teach every student], no one will


every studentli


no oneli



every studentli
covert, wide DS movement


not teach
every studentli



no oneli



every studentli
overt vP-fronting

Harmonic Derivationalism

Observe also that a minimal variant of (), not depicted here, in which the object
starts as a determiner (every) and the restrictor (student) is late merged above vP
abides by the UC but is excluded by the Case requirement (c), because objects have
to be assigned Case within vP.
() illustrated that complete copies created by movement need to preserve
c-command relations. Evidently, what applies to the object should also hold for the
subject. That is, one is led to expect that if the DP to be carried along by predicate fronting is a subject copy, the base position (Spec,vP) can be filled by the bare
determiner, inducing Late Merge in Spec,TP, but not by the fully assembled DP. The
derivation in () demonstrates that this prediction leads to important, empirically
verifiable consequences.
Suppose that at DS, the object does not cross over the subject but undergoes short
movement only, adjoining to what will later, in syntax, become vP (a). Given this
alternative parse, VP-fronting may target the highest vP segment in (b), removing the complication which afflicted (b). In (b), the object no longer enters
into ambiguous c-command relations with its copies. However, as can also be seen
from (b), the UC now penalizes the chain created by subject raising, because the
c-command relations for the subject are disrupted by VP-fronting.


() a.

no oneli


every studentli


no oneli

every studentli

local DS movement
Wholesale Late Merge of restrictors is licensed by Case. This entails that if LI is to apply at DS, a
level which by definition lacks syntactic features such as Case, the Case requirement must be satisfied
retroactively, in overt syntax. A simple way to implement retroactive licensing is to instruct overt syntax
that all complete DP copies reside within the c-command domain of their Case assigner.


no oneli teach no oneli
every studentli


no oneli teach
every studentli

overt movement of highest vP segment

It follows that the only way for the derivation to abide by the UC is to merge only the
determiner in Spec,vP, and to add the restrictor of the subject late, in Spec,TP.
What is of particular importance for present purposes is the fact that applying the
UC to subjects derives the generalization that in contexts of predicate fronting, the
unattested inverse scope order cannot be generated by subject reconstruction below
the object. Specifically, it must be ensured that at LF, which closely resembles (a),
the scope of the subject is not read off its lowest copy. Otherwise, the subject could
be assigned narrow scope with respect to the object, which resides in the lowest typecompatible node for quantifiers. The UC weeds out such configurations, though, by
requiring that the Spec,vP initially starts as a determiner only. From this it follows that
the subject cannot reconstruct below the object in syntax. Thus, the analysis derives
the fact that unattested scope inversion with predicate fronting cannot be produced
by subject reconstruction into Spec,vP.
The prohibition on inserting complete DPs into remnant-moved predicates has
another desirable consequence. Since bare D s always translate into individual variables (d), semantic reconstruction below the object is also blocked. Thus, the UC
accounts not only for the unavailability of wide object shift across the subject at DS but
also for the observation that subjects of fronted vPs do not reconstruct into their base
position, either in syntax or by SemR. A single principle accordingly derives both of
the two requirements which must be met by any successful analysis of scope freezing
with predicate fronting.
Finally, the trees in () document the derivation of the legitimate subject > object
scope order. As seen in (a), the object is fully assembled within VP, in accordance
with the Case condition (c), while the restrictor of the subject DP is late merged in
An interesting alternative proposal to restrict subject reconstruction in the same environments based
on the cycle is explored in Heck and Assmann ().

Harmonic Derivationalism

Spec,TP in order to avoid the conflict with the UC observed in (). In overt syntax,
the vP is fronted (b). Then, at LF, the object undergoes local type driven movement,
adjoining to vP. Subject reconstruction into Spec,vP is furthermore prohibited, as
Spec,vP is occupied by the determiner only. As a result, the analysis correctly predicts
that the derivation unambiguously generates the subject > object order.
() a.

no oneLI


every studentLI



no oneLI
not teach
every studentli


every studentli

overt VP-fronting
Turning next to DP-reconstruction, recall that if a DP raises and reconstructs for the
evaluation of binding relations, the CEC admits the local de dicto interpretation, in
which the s-variable moves locally, as in (a), but blocks the constellation (b),
where the s-variable skips the closest possible landing site right below the intensional
operator (seem). Nodes subscripted by PF signal spell-out positions, while LF subscripts denote reconstructed positions.


() a. [s [DPPF [seem [s [DP s ]LF ]]]

b. [s [DPPF [seem [s [DP s ]LF ]]]

(SynR of DP, de dicto)

(SynR of DP, de re)

Again, the CEC effect can be reduced to the DSLF model and the cycle.
To begin with, s-variables must be assumed to be part of the logical skeleton.
Otherwise, the existence of narrow-scope de re readings, exemplified by () (from
Buerle ), would remain mysterious.
() Georg believes that a woman from Stuttgart loves every member of the VfB
(believe > a woman > every member, de re)
In (), the restrictor member of the VfB team can be understood de re, while the containing universal QP is at the same time construed within the scope of the indefinite
a woman from Stuttgart (Fodors third reading: Fodor ; see Heim and von Fintel
, Sternefeld for discussion). Thus, the s-variable of the embedded in situ
object canunlike the s-variable of overtly moved noun phrasesbe construed longdistance. This indicates that situation pronouns are already present at DS, the only
level which allows non-local silent movement operations in the present system.
Furthermore, I assume that movement at DS (s-movement, wide QR) does not
proceed in a successive cyclic manner, but directly lands in the scope position of
the affected expression. This is natural given that DS does not yet include formal
syntactic features, which are arguably responsible for determining the location of
phases and other syntactic locality domains. Also, regular QR at LF has, to the best of
my knowledge, never been conceived of as a successive cyclic operation.
With these qualifications in place, it becomes apparent why the LF-representation
(b) is ill-formed. Moving the s-variable long distance at DS, as shown in (a),
creates a configuration that, if combined with subsequent subject raising in overt
syntax (b), leads to a violation of the Extension Condition. Concretely, the domain
within which the first, local step of subject raising to the lower Spec,TP applies in
(b) is properly contained within the domain established by s-movement.
() a. [s [seem [TP [DP-s ] [DP-s ]]]


(long s-movement at DS)

[seem [TP [DP-s ] [t

(subject raising in overt syntax)

What () demonstrates is that the cycle blocks long s-movement prior to raising.
Assume now alternatively that the s-variable remains in situ throughout DS (a) and
the subject raises to its surface position in overt syntax (b), followed by reconstruc Local movement of the object of () at LF does not conflict with the cycle because the movement is
obligatory, i.e. type-driven.

Harmonic Derivationalism

tion at LF. Finally, the s-variable is locally moved in order to ensure interpretability
(c). Then, one obtains the legitimate narrow-scope reconstructed de dicto interpretation for overtly moved DPs.
() a. [DP-s [seem [s [DP-s ]]
(no s-movement at DS)
b. [DP-s [seem [s [DP-s ]]]
(subject raising in overt syntax)
c. [DP-s [seem [s [DP-s ]]]
(reconstruction and local s-movement at LF)
To summarize, the DSLF model offers a natural analysis of phenomena which fall
under the descriptive generalization of the CEC. Silent operations can either apply
early in the derivation, at DS, in which case they may change scope orders and apply
long-distance, or they are delayed to LF, where they are subject to stricter locality
conditions, since as a true interface, LF only admits local displacement that serves
the purpose of repairing type mismatches.

. Discussion
Although at this point the contours of a theory become visible, many details are still
missing and important questions need to be asked. I will briefly add a few clarifying
First, even though parts of the procedure-fixing interpretation take place prior to
syntax, it is important to observe that the DSLF model is not a variant of generative
semantics. This is so because all model-theoretic interpretation takes place in a single
component. DS merely contributes another set of restrictions on the possible shape
the logical syntax of natural language expressions can take. For instance, while scope
relations are represented at DS, where they are subject to conditions such as Scope
Economy, it is LF that prepares them for compositional interpretation. As a result,
overt movement canunlike in theories advocated by generative semanticistsmake
a difference to interpretive properties, e.g. by extending binding domains.
Second, one might wonder if DS and LF ever generate contradictory statements
about scope relations. Two questions are relevant here. Can a wide-scope DSfor
instance, a DS encoding inverse scopeend up as a narrow-scope LF? And can
There is a third logically possible derivation to be considered, in which the subject raises, either at DS
or overtly, followed by local movement of the s-variable and reconstruction of the DP.

(i) a.
[DP-s [seem [DP-s ]]]
b. [s [DP-s [seem [DP-s ]]]]
c. [s
[seem [DP-s ]]]

(subject raising)
(local s-movement)

This Duke of York derivation (Lechner, to appear) is presumably blocked by the fact that s-movement
in (ib) would have to target both subject copies in order to ensure that the s-variable in the lower copy is
bound by s , too. This kind of Across-The-Board movement would be a hapax legomenon, though.
I am grateful to Rob Truswell for discussion of this and other issues addressed in this section.


narrow-scope DS representations be transformed into wide-scope LFs? Both constellations will be exemplified below. Assume that an object quantifier QP covertly
crosses the subject at DS, as in (a). Then, one might suspect, the lower, VP-internal
occurrence of QP has to be translated as an individual variable, and must not be
reconstructed, as in (b). Otherwise, DS derives the inverted reading, while LF
interprets the surface scope order, apparently leading to inconsistent predictions of
the system.
() Illicit configuration I: wide-scope DS paired with narrow-scope LF
a. DS: [QP [QP [VP QP ]]]
b. LF: [QP [QP [VP QP ]]

(wide-scope shift of object at DS)

(object reconstruction)

The problem turns out to be only apparent, though. DS provides a particular initial
hypothesis about a possible scope order, withoutand this is crucialbeing committed to the correctness of this claim. The system is constructional just like the standard
model in that it admits scope relations which have been established at an earlier point
to be changed by operations which apply later on in the derivation. Moreover, in
(a), inversion at DS is motivated by the need to assess the legitimacy of the inverted
order in addition to the surface order. Thus, the surface reading which is eventually
generated by LF is independently available. Possible mismatches between DS and LF
which arise from undoing DS movements at LF are therefore inconsequential.
The converse relation is illustrated by (). Here, the object undergoes local QR,
but subject reconstruction below the vP-adjoined object results in object wide scope
at LF. Again, the derivation seems to lead to inconsistency, in this case because LF
produces a non-surface scope option which is not made available by DS. (Given that
LF movement is always maximally local, such a combination can be arrived at by
reconstruction, but not by wide object QR at LF.)
() Illicit configuration II: narrow-scope DS paired with wide-scope LF
[vP QP [VP QP ]]
(surface scope)
a. DS: [QP
b. OS: [QP [QP [vP QP [VP QP ]]]
(overt subject movement to Spec,TP)
(local object QR)
c. LF: [QP [QP [vP QP [VP QP ]]]]
(subject reconstruction)
d. LF: [QP [QP [vP QP [VP QP ]]]
Now, recall from section .. that DP-reconstruction is contingent upon the presence
of a restrictor, and that the choice whether such a restrictor is merged is made as early
as at DS. Hence, DS already has access to the information whether the subject of ()
reconstructs or not. This is relevant, as the choice between early and late merge of the
restrictor is also reflected in differences in the DS representations to be transferred
to the deductive system. For instance, if a DS with a full subject copy in Spec,vP is
assembled, DS does actually not encode the surface scope option, but the inverted

Harmonic Derivationalism

interpretation. Thus, the conflict posed by scope mismatches between DS and LF

representations simply turns out to be illusoryand disappears.
Finally, the proposal makes no predictions as to how the boundedness of scope is
to be accounted for. One option might be to group DS operations together in cyclic
domains, replicating phase-bound analyses of scope boundedness. Alternatively, the
domain of DS could be the full numeration, filtering out illicit wide-scope options
during overt syntax. At the moment, I remain agnostic about this, as well as numerous
other issues.
The remainder of this section briefly comments on the role of opacity and the
architecture. In section ., it was noted that CEC can be understood as a description
of phenomena which display the signature of underapplication opacity. The absence of
a wide-scope option for object quantifiers and s-variables inside reconstructed nodes
could only be expressed by reference to previous stages of the derivationlocality of
movement is conditioned by overt displacement of the containing categories. In the
DSLF model, opacity can be given a different characterization: it comes as a result of
the model in conjunction with a general principlethe cyclewhich guides the order
in which operations apply. Moreover, CEC effects are not analysed within a single
component, but are resolved by distributing movement across two components, DS
and LF.
Taken together with Copy Theory and Late Merge, the intra-componential and
cross-componential strategies for opacity resolution above yield the Square of Opacity
in ().




a. Within a component

Copy Theory

Late Merge (of adjuncts)

b. Across components

SemR (w.r.t. scope)

SemR (w.r.t. binding)

Late Merge (of restrictors)

As () reveals, opacity is not an exception, but manifests itself in all cells of the matrix
generated by a model that packages information within and across components. This
finding supports the hypothesis that natural language is derivational in the strongest
possible way, as expressed by Harmonic Derivationalism in ():
() Harmonic Derivationalism
Natural language employs procedural signal manipulation and information
packaging within informationally encapsulated systems as well as across components.

Late Merge of adjuncts (Lebeaux ) applies within syntax. Late Merge of restrictors across components supplies DS representations with lexical content. These restrictors can be inserted high if the
determiner has moved within its Case domain, either at DS or in syntax.


The term Harmonic Derivationalism combines the ideas of derivational theories of

syntax with the belief that language is best understood as a set of ordered encapsulated
systems that contain discrete information which is passed on from one system to the
next by interfaces. Harmony derives from the fact that intra-componential and crosscomponential interactions are sequentially ordered.

. Conclusion
I have proposed a new architecture of the grammar, in which overt syntax is preceded by a second, covert component, DS, that branches off to the grammar-external
deductive system (Fox ). DS hosts all non-local movement operations (among
them scope-shifting QR and long s-movement) which potentially have an impact
on truth-conditional interpretation. In the current model, it becomes possible to
assign the traditional repository of covert movement (LF) the role of a true interface, whose function is restricted to rendering otherwise syntactically well-formed
representations afflicted by local type incompatibilities interpretable by the rules of
The logical skeletons which DS operates on consist of logical, permutationinvariant symbols and can bebut do not have to belexically impoverished. Underspecified nodes are provided with lexical content by Wholesale Late Merge, either at
DS or in syntax. Admitting Late Merge both within and across components increases
the symmetry of the system by making optimal use of the Late Merge resource independently used by the grammar.
Empirically, the DSLF model provides a set of effective and comprehensive strategies for the analysis of syntactic opacity effects. () repeats the main components of
the analyses from above. Recall that only (e) is new, and can probably be related to
independent properties of Case or the derivational system:
() a. Lexical Insertion can take place throughout the derivation, at DS or in overt
syntax. LI can in principle target any position of a chain.
b. Takahashi (): LI into a node is admissible only if is c-commanded
by its Case-assigning head.
c. Takahashi (, ): Non-lexicalized chain positions are filled by (possibly silent) determiners. These positions are converted into individual traces.
d. All syntactic movement that does not take place at the LF interface observes
the Extension Condition. As a repair strategy, LF movement is exempt from
the cycle.
e. Uniformity Condition. If a node is fully lexicalized, and is the highest
node of a movement chain, c-commands its lower occurrences throughout
the derivation.

Harmonic Derivationalism

Even though various aspects of the system look promising, this chapter has presented only an outline of a theory, with many details still to be added in the future.
For instance, the particular assumptions about how DS interacts with the grammarexternal deductive system and overt syntax will most likely turn out to be in need of
substantial revision. In the present chapter I have assumed that DS representations are
underspecified, so as to minimize changes in the interface from DS to the grammarexternal deductive systems. Other options are conceivable, though, as already alluded
to in section .. For instance, DS might operate on fully lexicalized terms throughout,
which are then stripped off during transfer to the deductive systems. Crucially, these
theoretical choices do not affect the main claim that scope-shifting silent movement
precedes overt syntax.
Also, various open questions remain, some of which have already emerged in
course of the discussion. What, for example, is the underlying explanation for the
UC, i.e. how does Case constrain remnant movement? In some sense, the UC has
the flavour of a condition which should ultimately fall out from the architecture of the
system, similar to the strict cycle condition. Another issue in need of clarification is
that the proper division of labour between DS and overt syntax must be made explicit.
Are optional movements such as scrambling, which are known to have an effect
on scope, computed in syntax, or pre-empted by DS? (See Lechner a for some
speculations.) At the moment, I have to defer answers to these and other questions to
future investigations into the nature of the DSLF architecture.

Reconstruction, Control,
and Movement

. Introduction
We currently suffer from an embarrassment of riches when confronting the proper
analysis of reconstruction effects, the term given to phenomena in which a constituent is interpreted in a lower position than that in which it is pronounced. The
copy theory of movement (Chomsky , building on earlier work by Barss
among others) provides a minimal syntactic approach to reconstruction effects: there
is no operation of Reconstruction, but instead, simply variation in which copy of a
constituent is interpreted at LF. Complementing this is an approach to reconstruction effects based on higher-type traces, situated within Heim and Kratzers ()
approach to movement dependencies (e.g. Cresti ). On this approach, variation in
the type of trace derives reconstruction effects without any Reconstruction operation,
exploiting instead the power of the compositional semantic mechanism. Several other
possibilities also exist: reconstruction effects could arise as a result of a lowering
operation (May ), of PF-movement with no effect on LF (Sauerland and Elbourne
), of relativizing interpretation to chains rather than single positions (Barss ),
or of a distinct reconstruction relation (Sternefeld ). All of these fit within modern syntactic theory (some more naturally than others), and most can be distinguished
empirically (see e.g. Fox for arguments against an entirely semantic approach to
scope reconstruction).
Most Minimalists have settled on some combination of the first four options, with
questions remaining about the division of labour between them. Chomsky (: )
Thanks to Caroline Heycock, Ray Jackendoff, Ad Neeleman, and two reviewers (one of whom selfidentified as Winnie Lechner), as well as audiences in Edinburgh, Oxford, and Ottawa, for comments on
these ideas in various stages of development. Portions of this material were first developed, but never published, in Neeleman and Truswell (). Subsequent research was conducted as a postdoctoral associate
at Tufts University (), and as a British Academy Postdoctoral Fellow at the University of Edinburgh

Reconstruction, Control, and Movement

included examples of literal lowering alongside cases of interpretation of lower copies

(see also Boeckx b), while Lechner (, Chapter above) argues that a grammar must include both a mechanism of Syntactic Reconstruction (e.g. lower copy
interpretation) and a mechanism of Semantic Reconstruction (like Crestis highertype traces), an argument we will return to at length below.
This chapter, in contrast, argues for a chain-based theory of reconstruction effects.
The argument hinges on cases where reconstruction effects are found in the absence
of movement relations. If we grant for now that such cases can be found, then the
argument is straightforward. Analyses of reconstruction effects in terms of copy theory or higher-type traces tie reconstruction closely to movement, and so struggle
to account for reconstruction without movement. As for the alternatives, a distinct
reconstruction operation is clearly a last resort on grounds of parsimony, while lowering operations are severely constrained by the Proper Binding Condition (Fiengo
). No one, to my knowledge, has suggested literal lowering as a general model of
reconstruction effects, for this reason.
I will argue that cases of reconstruction without movement favour a chain-based
analysis, based on a conception of chain which is not limited to movement relations.
Nothing forces chains to reflect movement relations alone (indeed, a system including chains determined exclusively by movement, alongside movement itself, is seen
as unparsimonious, following Chomsky and Brody ), and more inclusive
types of chain can be defined, as was common throughout the GB era. We will make
extensive use of that flexibility here, by defining two distinct types of chain, both more
expansive than the movement relation, and by tying different reconstruction effects
to those more inclusive types of chain.
The two types of chain are necessary to capture the dissociations among reconstruction effects documented by Lechner (, Chapter above): scope reconstruction
can be found without binding reconstruction, and vice versa. In the account we
sketch below, each type of reconstruction effect is associated with a different type of
chain, and the partial autonomy of the two types of chain derives the dissociations.
Such dissociations are less clearly explicable on a theory which makes movement a
precondition for all reconstruction effects.
An apparent empirical challenge to this approach comes from the trapping effects
discussed by May () and Fox (). Trapping effects arise where the position in which a constituent enters into binding relations determines the position
in which it takes scope. Such an interaction suggests that scope reconstruction and
binding reconstruction march in lockstep: one is found where the other is. This
is in contrast to the dissociations that Lechner observes, which could be seen as
countertrapping effects: in Lechners examples, just because a constituent enters
into scope relations from a given position, there is no guarantee that it could also
enter into binding relations from that position, or vice versa. The final task in this
chapter, then, is to demonstrate that a chain-based approach to reconstruction can


account for trapping effects, and can capture the balance between trapping and
In what follows, section . demonstrates the existence of reconstruction across one
non-movement dependency, obligatory control. Section . compares reconstruction across obligatory control dependencies to the countertrapping effects observed
by Lechner () in A-scrambling and extraction from weak islands. Section .
addresses the balance between trapping and countertrapping effects. Section .

. Reconstruction Across Control Dependencies

Sentences such as (), featuring obligatory complement control dependencies, can be
scopally ambiguous.
() Someone tried to read every book in the library.

( > , > )

Historically, the ambiguity of such sentences has been a matter of contention (they
were prominently claimed to be unambiguous in May ), but a consensus is
emerging that the scope ambiguity is real and cannot be fully explained by reference
to restructuring (Kennedy , Johnson , Hornstein , Wurmbrand ).
Of course, once we decide that the ambiguity is real, we have to decide what to
do about it. We could, like Kennedy (), postulate an operation of long-distance
Quantifier Raising, allowing every book in the library to take matrix scope ().
() [Every book in the libraryi [someonej [tj tried to read ti ]]]
Alternatively, the scope ambiguity may be due to a reconstruction effect: someone
takes scope in the embedded clause. We could provisionally represent this as in (),
employing ideas from Hornstein (, , , ). However, () represents
control as movement and treats reconstruction effects as cases where different copies
are interpreted from those that are pronounced. We will reject both of these claims
() [TP Someone [vP someone tried [TP someone to [vP every book in the library
[vP someone read every book in the library]]]]]
(pronounced copies in italics, interpreted copies in bold)
This chapter will argue that the scope ambiguity is a reconstruction effect. This, of
course, leaves the question of the status of Kennedys () arguments, based on
Antecedent-Contained Deletion, for a QR treatment, but that will have to wait for
another time.
The arguments that obligatory control exhibits scope reconstruction were initially presented in Neeleman and Truswell ().

Reconstruction, Control, and Movement

A major difference between QR and reconstruction is that the QR analysis predicts

that the embedded quantifier every book in the library can take scope over everything
in the matrix clause, while the reconstruction analysis is selective: it predicts that only
someone, among the constituents of the matrix clause, can take scope under every book
in the library. The evidence clearly favours the reconstruction analysis.
Firstly, an embedded quantifier cannot take scope over a matrix adverbial. (a) is
ambiguous, while (b) is not.
() a. John frequently checks every calculation.
( > frequently, frequently > )
b. John frequently tries to check every calculation.
(frequently > , >frequently)
(a) has a reading where all calculations are frequently checked, but not necessarily
en bloc. (b) lacks this reading. This is evidence against the QR analysis: if every
calculation can take matrix scope, why can it not take scope over frequently?
As suggested by a reviewer, a natural way to attempt to capture these data on a
QR approach would be to stipulate that QR of every calculation in (b) can target a
mid-size clausal position (Johnson and Tomioka ) between frequently and the
VP-internal trace position of John, but cannot target a position above frequently, as in
(b) (with irrelevant copies omitted). Meanwhile, in (a), QR would be able to target
a position above frequently, as in (a).
() a. [John (every calculation) frequently (every calculation) [John checks every
b. [John ( every calculation) frequently (every calculation) [John tries [pro to
check every calculation]]]
However, this analysis runs into conceptual and empirical difficulties. Conceptually,
there is the question of why QR would cover a maximum of . clauses, or alternatively, why frequently acts as a barrier to QR in (b), but apparently not in (a).
Empirically, problems arise when trying to order frequently with respect to mid-size
clausal elements across which QR is attested. Johnson and Tomioka () show that
two thirds of the questions on the exam can QR across negation in (a), while (b)
shows that negation can precede (and so presumably c-command) frequently.
() a. Some student or other hasnt answered two thirds of the questions on the
(some > > )
b. John (frequently) doesnt (frequently) check every calculation.
If QR can cross negation (a), and moreover Johnson and Tomiokas mid-size clausal
position is above negation, which is in turn above one position where frequently
occurs, the impossibility of long QR across frequently in (b) is surprising. The most


obvious conclusion, of course, is that long-distance QR out of control complements

does not exist.
Complementary evidence against a QR approach to scope inversion in obligatory
control constructions comes from transitive control constructions, with two matrix
arguments. In such cases, only the controller can take scope under an embedded
quantifier. Given an object control predicate like persuade, an embedded quantifier
can take scope over an object but not a subject.
() a. Mary persuaded someone [to read every book on the reading list]
( > , % > )
b. Someone persuaded Mary [to read every book on the reading list]
( > , > )
In contrast, given a subject control predicate like promise, the object cannot scope
under the embedded quantifier.
() a. Someone promised Mary [to read every book on the reading list]
( > , % > )
b. Mary promised someone [to read every book on the reading list]
( > , > )
Finally, scope inversion in complement control constructions displays trapping
effects: if a quantifier must bind a variable in the matrix clause, then it cannot take
scope under an embedded quantifier, even if this would be otherwise possible.
() Someone promised himself [to read every book on the reading list]
( > , > )
All of these data are compatible with an analysis of scope inversion in complement
control constructions as reconstruction effects: the controller is the only constituent
of the matrix clause which interacts scopally with constituents of the embedded clause.
The data call into question a QR-based analysis, though, as such an analysis does
not predict such selective patterns of scope inversion. We conclude that complement
control constructions display scope reconstruction effects: the controller can enter
into scope relations as if it were in the embedded clause.
Some individuals do not permit even this. To my knowledge, though, no one is more permissive than
this, once matters such as information structure are controlled for.
Winfried Lechner (p.c.) suggested a further potential argument that control complements are scope
islands. If every window must scope under forgot in (i), it would provide further evidence that QR out of
control complements is unavailable.

(i) John forgot to close every window.

My intuition broadly concurs with this prediction. However, here, as elsewhere, emphasis can lead to freer
scope possibilities.
(ii) John forgot to close Every Single Window!

(forgot > , > forgot)

Pending better understanding of this interaction, it is prudent not to read too much into (i).

Reconstruction, Control, and Movement

A natural hypothesis would be that the controller can take scope in the embedded clause because of the dependency between the controller and pro, the empty
subject of the embedded clause. For example, we could adopt the Movement Theory of Control (Hornstein , Boeckx, Hornstein, and Nunes ), according
to which obligatory control is a species of A-movement, distinguished by the fact
that it targets a -position in the matrix clause, rather than a Case position (see ()
above). The possibility of scope inversion in sentences like () would then result from
scope reconstruction across an A-movement dependency, plus whatever mechanism
is responsible for local scope inversion (for Hornstein, a combination of vP-internal
subjects and object shift, as in ()).
However, there are grounds for scepticism concerning the Movement Theory of
Control, as discussed in Culicover and Jackendoff (), Landau (, ),
and elsewhere. Without going too far into the details of this ongoing debate, the
major arguments centre around similarities and differences between the grammatical
dependencies underpinning A-movement and many obligatory control relations. A
large number of differences between the properties of the two types of dependency
have now been documented, militating against a reduction of obligatory control to
Firstly, subject control predicates such as promise notoriously violate minimality
constraints on controller choice. Under normal circumstances, pro is controlled by
the closest c-commanding argument: the matrix object if there is one, and the matrix
subject if not.
() a. Johni wanted [proi to blow up the bridge]
b. Johni persuaded Billyj [proj/ i to blow up the bridge]
However, promise allows the subject to control pro, despite the presence of an intervening object ().
() Johni promised Billyj [proi/ j to blow up the bridge]
Boeckx and Hornstein () attempt to capitalize on this by pointing to the peripheral nature of subject control verbs in the grammar of English: they are acquired
late, and informal surveys suggest that they are often not acquired at all. For
Boeckx and Hornstein, then, the exceptional non-locality of subject control dependencies is reflected in their late acquisition and omissibility from the grammar of
English. However, an irreducible difference remains between locality of obligatory
control and of regular A-movement: the locality constraints on A-movement are
absolute, while those constraining obligatory control are a matter of markedness,
and show some variability in controller choice (a phenomenon known as control
This is most clearly seen, again, with verbs like promise. In (), the subject controls
pro, while in (), following Hust and Brame (), the object controls pro.
() Johni promised Billyj [pro#i/j to be allowed to blow up the bridge]


There are also cases, like (), where the antecedent of an A-trace is not the closest
c-commanding NP.
() a. Billyi strikes me as ti weird.
b. Billyi seems to me ti to be losing the plot.
However, these cases have very different properties from the variable control exhibited
by promise. The experiencer argument of strike or seem is never a landing site for Amovement. This might indicate that the experiencer does not enter into the calculation
of locality in English A-movement dependencies. If so, an analysis can be maintained on which the antecedent of an A-trace is the closest c-commanding potential
antecedent. However, such an analysis will not work for promise: () demonstrates
that the object of promise is a potential antecedent of pro, but that antecedent is
nevertheless skipped in the subject control construction in ().
These considerations weigh against a movement-based analysis, building on Larson (), in which the object of promise is inaccessible as a landing site for some
reason. Even if such an analysis were tenable for promise in the general case, it
would fail to explain the discrepancy between raising (which always targets a unique
designated landing site) and control (which exhibits variability in cases like ()
and ()).
In contrast, thematically based theories of controller choice (Jackendoff ,
Farkas , Sag and Pollard , Jackendoff and Culicover ) predict that subject
control exists, and predict which predicates exhibit this exceptional pattern. Such
theories propose that the control properties of a predicate directly reflect its thematic
properties: very roughly, the exceptional control behaviour of promise is related to
the fact that if X promises Y to V, X is committed to bringing about an instance of
V-ing, while if X persuades Y to V, it is Y who comes to have such a commitment.
The controller is the person who is committed to the V-ing, and exceptional cases
of object-control promise can be accommodated as instances of coercion. To the
extent that there is no uniform syntactic representation of such commitments, the
theory of control is also independent of particular syntactic structures, and the fact
that such dependencies do not share the properties of A-movement dependencies is
Thematic theories of control are also corroborated if control properties, stated
in thematic terms, remain invariant when a predicate appears in different syntactic configurations. Culicover and Jackendoff demonstrate such invariance in

These statements are all complicated by the consideration of dative intervention effects in A-movement.
However, even here, there is an irreducible difference between control and movement: Beaven () shows
that dative intervention constrains A-movement to different extents in different languages, but there are no
dative intervention effects in obligatory control in any language Beaven examined.

Reconstruction, Control, and Movement

the nominalizations of control predicates. Nominalization preserves the control

properties of these predicates, but the syntactic realization of the controller is more
variable in a nominal construction, as in ().
The promise to Susan from John to take care of himself/ herself.
John gave Susan some sort of promise to take care of himself/ herself.
Susan got from John some sort of promise to take care of himself/ herself.
A: John made Susan a promise.
B: What was it?
A: I think it was to take care of himself/ herself.
(all Culicover and Jackendoff : )
e. Johns promise to Susan to take care of himself/ herself.
f. The promise from John to Susan to take care of himself/ herself.
g. The promise that Susan received on the part of John to take care of
himself/ herself.

() a.

It is possible in principle that such variability belies an underlying syntactic uniformity

on which the control relations dependindeed, Boeckx et al. () sketch such an
account of data like these. To my knowledge, however, there is no current general
theory relating such variability in the realization of thematic relations to a uniform
underlying syntax. Examples like (), together with the behaviour of promise, therefore stand as evidence in favour of theories of obligatory control as a thematic, and
not narrowly syntactic, dependency: when the syntax varies and the thematic relations
stay the same, control relations track the thematic relations rather than the configurational syntax.
This picture is complicated somewhat by the fact that obligatory control into the
complement of a control verbexamples such as ()show all the hallmarks of
grammatical dependencies. Koster () claimed that in every grammatical dependency, a unique, obligatory antecedent locally c-commands a dependent element.
Following Hornstein (: ), we can see that complement control shows all these
properties. (a) shows that the controller is obligatory, (b) that it must be local
to pro, (c) that it must c-command pro, and (d) that there must be a unique

Moreover, examples like those in () suggest that c-command is not a necessary condition on the
relationship between controller and controllee, casting further doubt on the analysis of obligatory control as
a species of grammatical dependency. This doubt is perhaps strengthened by examples such as The promise
to leave, where the same person is the agent of promise and leave. However, the latter data, at least, are
amenable to analysis in terms of a null subject of promise.
Landau () discusses partial control at length (The director decided pro to meet at , with pro
interpreted as referring to the director and others). However, partial control does not so much invalidate
the claim that the controller is unique as refine our notion of what it means to be a unique controller.


() a.

It was expected pro to shave himself.

John thinks that it was expected

pro to shave himself.

Johns campaign expects pro to shave himself.


told Maryj proi+j to wash themselves/each other.

It seems, then, that the configuration of pro and controller is regulated by a grammatical dependency in complement control constructions, but that this dependency
is independent of the mechanisms underlying controller choice, which are also active
in cases (particularly control in nominals) where there is no evidence of a grammatical dependency. Such a bipartite system is defended by Landau (), who proposes a series of Agree relations between heads in complement control constructions,
partially independent of the lexicosemantic factors determining controller choice.
Although we do not need to follow the specifics of his implementation, we must adopt
this separation of the grammatical dependency from factors relating to controller
This raises the question of whether scope reconstruction is contingent on this
grammatical dependency or on the thematic control relation. The evidence points
to the former. We have seen, for example, that obligatory control in nominals is not
accompanied by a grammatical dependency. And unlike obligatory control in clauses,
there is no long-distance scope inversion in such cases.
() a. Someone tried to read every book in the library.
b. Someones attempt to read every book in the library.

( > , > )
( > , > )

Likewise, there is no scope inversion in cases of non-obligatory control, where there is

widespread agreement (e.g. Williams ) that there is no grammatical dependency.
() To read every book in the library would be nice for someone.

( > , > )

The presence of scope reconstruction in obligatory control constructions is tied to the

grammatical dependency, then, rather than the control relation.
More remarkably, such scope reconstruction effects are not accompanied by binding reconstruction effects, as far as we can tell. Straightforward testing of the binding principles is marred by various confounds (for Condition A, it is impossible
to tell whether a reflexive in the embedded clause is bound by the reconstructed
If this is accurate, it undermines a reviewers suggestion that scope reconstruction across obligatory
control dependencies could be handled by a higher-type pro, analogous to the higher-type traces of Cresti
(). There are independent grounds for scepticism about this suggestion: a major difference between
Crestis cases and obligatory control constructions is that two distinct -roles are implicated in obligatory
control, which significantly complicates the statement of a compositional semantics involving a higher-type
pro. Regardless of this, though, if there is a pro in nominal, as well as verbal, obligatory control, then the
availability of a higher-type pro would make scope reconstruction equally available in either case, contrary
to fact.

Reconstruction, Control, and Movement

controller or by pro; configurations which would test reconstruction for Condition

B are independently ruled out as surface violations of Condition C; and the failure of
reconstruction for Condition C across a control dependency is unsurprising, given
the demonstration in Fox that A-movement similarly fails to reconstruct for
Condition C).
However, () shows an apparent failure of reconstruction for Principle A with a
subordinate psych-predicate, relying on the familiar property of psych-predicates that
the object can bind into the subject. (a) shows binding of the reflexive in a raising
construction, while (b) shows that binding is not possible in the same configuration
in a control construction.
() a. Stories about himself appear to have upset every senator.
b. Stories about himself have managed to upset every senator.
Meanwhile, (a) demonstrates the absence of reconstruction for variable binding in
control constructions, in contrast with the raising example (b). We will tentatively
take this as evidence that control constructions do not reconstruct for binding.
() a. Heri friends promised every girli to meet her after school.
b. Hisi mother seems to every boyi to be wonderful.
Likewise, () demonstrates the absence of NPI-licensing under reconstruction, further evidence in the same direction if Progovac () is right to relate NPI-licensing
to binding.
() a. [A doctor with any knowledge of acupuncture] wanted not to be
b. [A doctor with any knowledge of acupuncture] seemed not to be
Patterns like these are further evidence against reducing obligatory control to
A-movement: the pattern of reconstruction effects is different in the two constructions. However, it also poses an analytical challenge in its own right: if we see some
reconstruction effects here, why do we see so few? That is the topic of the next

. Lechners Double Dissociation

The occurrence in complement control constructions of scope reconstruction unaccompanied by binding reconstruction recalls seminal data from Lechner (,
a, Chapter above). Lechner showed that German Mittelfeld scrambling displays
scope reconstruction effects, but not binding reconstruction; while the opposite holds
of extraction from weak islands.


() demonstrates scope reconstruction without binding reconstruction. Scope in

German typically reflects surface c-command, so it is natural to relate the ambiguity of
() to the fact that the theme has scrambled over the goal: if the theme is interpreted
in its surface position, it takes wide scope; if it is interpreted in the trace position (we
will modify this below), it takes narrow scope.
() weil sie [ein Bild von seinem i Auftritt]j jedem Kandidateni tj zeigte
appearance every candidate
since she [a picture of his
since she showed every candidate a picture of his appearance
( > , > , Lechner : )
However, despite the fact that the theme can take scope in the trace position, the
pronoun seinem contained within the theme cannot be bound in that position, and
must be interpreted as a free pronoun. () is therefore a second case of scope reconstruction without binding reconstruction.
In contrast, following Longobardi (), Lechner shows that extraction from whislands displays a full range of binding reconstruction effects. (a) demonstrates
reconstruction for Condition A (a reflexive pronoun can be bound from the trace
position), as well as obligatory reconstruction for Condition B (a regular pronoun
must be free in the trace position). (b) demonstrates that reconstruction for Condition C is obligatory (she must be disjoint in reference from Mary), and (c) shows
that reconstruction for variable binding is possible.
() a. It is to herselfi / heri that I dont know whether Maryi wrote.
(Lechner : )
b. It is to Maryi that I dont know whether shei wrote.
c. It is to heri mother that I dont know whether every girli wrote.
However, there is no scope reconstruction into weak islands. How many-questions
can give rise to scope ambiguity: on the standard analysis, the two readings of ()
arise because how many contains two operators. The scope of one is fixed, as in
standard analyses of operator movement (e.g. Chomsky ), while the other is free
to reconstruct.
() How many books does Chris want to buy?
a. For what n, there are n-many books b , . . . , bn , and for every i, i n,
Chris wants to buy bi .
b. For what n, Chris wants that [there are n-many books b , . . . , bn , and for
every i, i n, he buys bi ].
In principle, () should be ambiguous in the same way. However, only the reading
corresponding to (a) is available, with specific books in mind. As this is the reading
where many books takes scope above want, and the reading where many books takes

Reconstruction, Control, and Movement

scope below want is unavailable, we conclude that scope reconstruction into whislands is not possible.
() How many books do you wonder whether Chris wants to buy? (many > want,
want > many)
a. There are three books that I wonder whether Chris wants to buy: Dubliners,
Ulysses, and Finnegans Wake.
b. I wonder whether Chris might be interested in buying three books, but
I dont particularly care whether Chris might be interested in buying two
books or four books.
Combining these two findings gives a pair of examples in which the requirement that
a many-phrase takes high scope (because there is no scope reconstruction into a weak
island) bleeds the possibility of a pronoun being interpreted within the weak island,
a variant on classical trapping effects.
() a. Every boy must see five pictures from the teachers childhood before he can
go home. The teacher must oversee this process, but hasnt paid attention,
so he is now unsure who has seen which pictures.
Question: How many pictures from hisi childhood does the teacheri need
to know whether every boy has seen?
b. Every boy must see five pictures from his own childhood before he can go
home. The teacher must oversee this process, but hasnt paid attention, so he
is now unsure who has seen which pictures.
Question: How many pictures from hisi childhood does the teacher need
to know whether every boyi has seen?
Extraction from weak islands therefore shows the opposite pattern to Mittelfeld scrambling or complement control: binding reconstruction is possible, but scope reconstruction is not. Scope reconstruction and binding reconstruction are therefore doubly dissociable.
The next question is how to account for this. Lechners approach is to make use
of two different technologies for capturing reconstruction effects, and associate each
type of effect with a different technology. Binding reconstruction would then be tied
to copy-theoretic Syntactic Reconstruction, while scope reconstruction would be tied
to the use of higher-type traces, as in Cresti () (Semantic Reconstruction). The
implication is that a theory of the distribution of the two types of reconstruction
technology gives a theory of the distribution of the two types of reconstruction effect.
Lechners approach is adequate in principle to capture the data he describes.
However, both of his technologies crucially implicate a movement relation. The
demonstration in section . that reconstruction effects exist without movement
dependencies therefore leads us to consider a different approach. In fact, I will suggest


that neither of the constructions Lechner describes involves actual movement, but that
both constructions, like complement control, bear similarities to movement relations.
We begin to investigate this possibility by comparing complement control and
Mittelfeld scrambling. Both have the same reconstruction profile: they allow scope
reconstruction, but not binding reconstruction. Moreover, in both cases there are
reasons to suspect that the grammatical dependency in question is not an instance
of movement. For complement control, this has already been discussed above; for
scrambling to the Mittelfeld, I refer the reader to Bayer and Kornfilt (), Neeleman
(), Neeleman and Weerman (), and Fanselow (, ). All of these
researchers propose that the relation between a predicate and a scrambled argument
is direct, rather than mediated by a trace or copy. For Neeleman, and Neeleman
and Weerman, for example, -roles are encoded as dependencies of a verb on its
arguments, and those dependencies form a record of which argument bears which
role. Instead of the movement approach, then, where the intended interpretation is
recovered by virtue of a dependency between a scrambled argument and a trace in
that arguments -position, Neelemans system recovers largely equivalent information
by virtue of a dependency between a scrambled argument and a member of a grid.
Nevertheless, both control and scrambling relations share many properties with
movement: all three dependencies require a unique, obligatory antecedent, which
locally c-commands the dependent. Following Koster (), I take these to be the
core properties of grammatical dependencies. For the remainder of this chapter, I will
Winfried Lechner (p.c.) suggests the following argument that scrambling configurations must be
derived by movement. Pair-list readings of wh-questions only arise if a quantifier c-commands the whtrace (May , Aoun and Li , Chierchia , Beck b).

(i) a. Which student did John give every book to t? (pair-list)

b. Which book did John give t to every student? (single pair)
The same facts occur in German, despite the fact that German, unlike English, allows scrambling (Lechner
cites Pafel in this connection).
(ii) a. Welches Buch hat Peter jedem
Studenten empfohlen?
which book has Peter every.dat student.dat recommended
Which book did Peter recommend to every student? (pair-list)
b. Welchem Studenten hat Peter jedes
which student has Peter every.acc book.acc recommended
Which book did Peter recommend to every student? (single pair)
This is surprising if German scrambling allows internal arguments to freely permute; less so if the indirect
object must always c-command the direct object or a trace thereof. However, an alternative interpretation of
these facts is that A-scrambling bleeds A -movement: A -movement cannot apply to an A-scrambled constituent and disrupt the marked, scrambled order of internal arguments. Motivation for such a constraint
can be found in the discussion of scrambling and information structure in Neeleman and van de Koot
(), orthogonal to the question of whether A-scrambling leaves a trace. This evidence for scrambling as
a movement relation therefore seems to me inconclusive.

Reconstruction, Control, and Movement

assume that neither complement control nor scrambling to the Mittelfeld represent
instances of movement, but that (like movement) they are instances of grammatical
dependencies. This is partly for the independent reasons given in the references
above, and partly because complement control and Mittelfeld scrambling do not
show the same reconstruction effects as movement. Movement allows reconstruction
for both scope and binding, while complement control and short scrambling only
show scope reconstruction effects. If we were to analyse complement control and
Mittelfeld scrambling as instances of movement, we would have to find some auxiliary
explanation for the absence of binding reconstruction effects, and I, personally, dont
see where to start.
We may then ask how movement is different from other grammatical dependencies. Given Kosters characterization of grammatical dependencies, as used repeatedly
above, the major extra property of movement is that it relates positions which share
a thematic role (to be understood, if necessary, in an extended sense that encompasses adjunct participant roles). In contrast, the antecedent and dependent in a
complement control relation bear different thematic roles, while the dependent in
a base-generated scrambling relation is a predicate (a -role assigner), rather than an
We can represent these distinctions, in the spirit (if not the letter) of proposals in
Neeleman and van de Koot (), by thinking about the grammatical dependencies
introduced by the elements in these constructions. A -role assigner is dependent
on the -role assignee, whether that argument is a sister of the predicate (as in nonscrambling constructions) or asymmetrically c-commands it (as in scrambling constructions). A trace is dependent on an antecedent with which it shares a -role, while
pro is dependent on a controller which bears a distinct -role. We represent these
three types of dependency as species of the general class of grammatical dependency,
abbreviated as Dep in ().
The claim is more controversial for A-movement. Following May (), some scope reconstruction
effects clearly obtain, often analysed as instances of Quantifier Lowering (Chomsky , , Boeckx
b). Meanwhile, some potential Condition C violations fail to occur under reconstruction, which is
sometimes taken to indicate an absence of binding reconstruction effects (Fox ), apparently contradicted by prima facie cases of reconstruction for Condition A (Takahashi and Hulsey ). The empirical
picture is still somewhat unclear and inconsistent, but I will assume that A-movement can reconstruct for
scope and binding, just like A -movement, with apparent failures of binding reconstruction to be explained
by independent factors, along the lines of the above references. If this turns out to be incorrect, alternative
theories of A-movement can be envisaged which are compatible with the approach being developed here.
In (), simplifying aspects of Neeleman and van de Koot () which are irrelevant to us, a
constituent marked with Dep introduces a dependency, while a constituent marked with Dep# satisfies
that dependency. A constituent subscripted with Dep# satisfies that dependency. Subscripted i and j are
diacritics to identify which constituents share or do not share a -role.
This bottom-up approach to dependencies, inherited from Neeleman and van de Koot as well as HPSG
and other frameworks, is the reverse of the standard probegoal Agree configuration. I leave it as an open
question how much of the present approach could be translated into Agree-based terms.


() a. Movement: [XPi Dep# [ti Dep ]]

b. Complement control: [NPi Dep# [proj Dep ]]
c. Mittelfeld scrambling: [NPi Dep# [VDep ]]
Movement, complement control, and scrambling to the Mittelfeld all involve a grammatical dependency, then. The distinguishing feature of movement dependencies
is that the antecedent and dependent share a -role. In complement control constructions, the antecedent and dependent have different -roles; in base-generated
scrambling constructions, the dependent assigns a -role to the antecedent, but the
dependent does not itself bear that -role. This then suggests a composite definition
of movement relations.
() Movement =df
a. A grammatical dependency (obligatory, unique antecedent locally ccommanding a dependent), where;
b. The antecedent and dependent share a -role.
The similarities in the scope reconstruction behaviour of movement, complement
control, and Mittelfeld scrambling could then be related to the fact that all three are
grammatical dependencies; the differences between them with respect to binding
reconstruction can be related to the fact that movement is the only one of these three
constructions where the antecedent and dependent share a -role. We arrive at the
following hypothesis:
() a. Scope reconstruction effects
A constituent may take scope from the position of a constituent to which it
is related by a grammatical dependency.
b. Binding reconstruction effects
A constituent may (Condition A) or must (Condition B/C) enter into binding relations from the position of a constituent with which it shares a
This hypothesis would, of course, be strengthened if we could demonstrate a double
dissociation: binding reconstruction effects are found without scope reconstruction
precisely where multiple nodes share a single -role, but are not related by a grammatical dependency. This suggests an analysis of the second part of Lechners double
dissociation, namely extraction from weak islands.
Adapting ideas from Cinque (), we could hypothesize that successive-cyclic
A -movement cannot cross a weak island boundary, but that a null element within a
The disjunctive may or must reflects the demonstration in Lebeaux () that Conditions B and
C, unlike Condition A, are everywhere conditions, which must be obeyed in every relevant position. It is
straightforward to adjust this if it should turn out that Lebeauxs analysis is inaccurate, and that Fox ()
was correct to locate all binding-theoretic effects at LF. I make no different predictions from those of anyone
else in this respect.

Reconstruction, Control, and Movement

weak island can be A -bound by an operator outside the island. According to such a
hypothesis, all A -dependencies crossing weak island boundaries are base-generated.
As is well known, binding relations do not behave like grammatical dependencies: for
example, the antecedent need not c-command the dependent.
() a. [[Every boy]i s mother] loves himi .
b. I danced [with [every boy]i ] on hisi birthday.
If extraction from a weak island is a base-generated binding relation, it is not a grammatical relation. On the hypothesis outlined in (), we would then predict an absence
of scope reconstruction. However, in a base-generated A -binding configuration, the
-criterion insists that Wh and pro share a -role, as Wh is never in a -position. Again
according to (), then, we expect that binding reconstruction will be displayed.
There is much scope for future research along these lines. Postulation of pro at
the foot of an A -chain has the effect of making extraction from weak islands look
like an instance of null resumption, as mentioned by both reviewers. A recent strand
of research has concentrated on interactions between island-sensitivity and reconstruction in a variety of resumptive constructions across languages (Aoun and Benmamoun , Aoun, Choueiri, and Hornstein , Aoun and Li , Guilliot
and Malkawi , , , Rouveret ). Typically (see in particular Rouveret
, and Guilliot and Malkawi ), differences among resumptive constructions
in these respects are cashed out in terms of a typology of pronominal elements, itself
derived from independent research into the syntax and semantics of pronominals,
such as the distinction between e-type and bound readings (Evans , Elbourne
). A logical next step for this research would therefore be to situate these claims
about reconstruction into weak islands, with respect to this typology of resumptive
There are many competing analyses of extraction from weak islands, and I have
no new evidence directly supporting the A -bound pro analysis over any competitors.
Equally, I cannot presently explain why particular reconstruction effects are tied to
This is actually Cinques analysis of strong islands, which (on his definition) can only be escaped by
NP. Weak islands, such as the wh-islands described above, can be escaped by NP or PP. Adapting Cinques
ideas in this way therefore has consequences for locality theory, which I have not presently investigated.
Similar questions arise concerning the syntax and semantics of copy-raising, in which two A-positions
are related and the lower position is occupied by a pronoun. Fujii () demonstrates limited binding
reconstruction, without scope reconstruction, in copy-raising sentences. However, Asudeh and Toivonen
() demonstrate that even the full range of binding reconstruction effects is not attested. (Asudeh and
Toivonen is a pre-publication version of Asudeh and Toivonen . The relevant discussion was
omitted from the published version.) Clearly, more research is needed to reconcile the two sets of facts.
A further interesting piece of theoretical housekeeping for the future concerns the integration of this
line of research with the claims of Takahashi and Hulsey (). Their combination of wholesale late merger
and trace conversion has the effect of rendering many copies indistinguishable from pronouns. Given that
the resumptive literature assumes that pronouns at the foot of chains are independently necessary, it is
natural to ask how much of Takahashi and Hulseys analysis can be reformulated using pronouns instead
of copies.


particular relations in the way I have suggested. However, I believe that the dissociation of scope reconstruction from binding reconstruction in this way has a good deal
of promise. Lechners double dissociation shows that reconstruction effects are not
a monolithic class, although there is a clear tendency in GB and Minimalism to treat
them as such. A major property of copy-theoretic approaches to reconstruction effects
is that, all else being equal, they treat all reconstruction effects as being of a piece,
and this dissociation should cast doubt on that project. On the other hand, Lechners
own analysis of the double dissociation is still closely tied to movement dependencies,
and the foregoing considerations should call this into question.
On the present approach, reconstruction effects are not limited to movement relations, and this appears to be accurate. However, movement does retain a privileged
position in the taxonomy of reconstruction effects: movement is unique in that it is
a grammatical dependency which relates positions which share a -role, as a matter of definition according to (). According to (), then, only movement should
display both scope and binding reconstruction effects: there is more reconstruction
in movement dependencies than anywhere else. However, Lechners double dissociation is also predicted, as a consequence of the independence of the two types of

. Trapping and Countertrapping

The previous section sketched a theory which partially dissociates reconstruction
effects from movement relations, in a way which is sufficiently flexible to account
simultaneously for the attested dissociations between scope reconstruction and binding reconstruction, and for the presence of reconstruction effects in the absence of
movement. We should now ask how to represent the relations which underpin reconstruction effects.
One option which can be immediately discarded is a single level of representation,
like the LF of May () and much subsequent work, in which scope and binding
relations are unambiguously represented by the position of a constituent. The reason
is that a single position cannot represent two dissociable pieces of information: the
fact that XP can take scope in a given position is no guarantee that it can also enter
into binding relations from that position, and vice versa. If, for example, a controller
in a complement control construction can take scope from the position of pro, then
we might represent that by having the controller in the position of pro at LF, as we
provisionally did in (). However, given standard assumptions about the interpretation of LF, this would also predict that the controller enters into binding relations from
the position of pro, contrary to fact.

Of course, all else is not always equal. Explanatory devices like partial reconstruction account for
certain discrepancies between scope and binding reconstruction, but do not touch the patterns discussed

Reconstruction, Control, and Movement

Disregarding this traditional conception of LF, there are two major possibilities:
either different types of relation are represented at different levels, or they are represented at a single level, in such a way as to allow the representation of scope and
binding more autonomy than in traditional LFs, for example by defining multiple
types of chain over surface representations. Neither of these options suffers from the
problem just described for traditional LFs, because in each case, the position in which
a phrase takes scope does not automatically guarantee that a phrase can enter into
binding relations from that position, or vice versa.
To choose between these two options, we must decide how independent scope
reconstruction and binding reconstruction are from each other. To assign different phenomena to different levels of representation is an analytical move
reflecting a hypothesis that the phenomena in question are encapsulated from
each other. Representing them at the same level predicts that they will interact
In fact, a demonstration that scope reconstruction and binding reconstruction
interact substantially was given as early as May (), concerning trapping effects
such as ().
() a. Two people seem to be dancing with every senator.
( > , > )
b. Two people seem to themselves to be dancing with every senator.
( > , > )
() shows that scope reconstruction interacts with binding: the fact that Two people
enters into binding relations in the matrix clause in (b) bleeds scope reconstruction
into the embedded clause.
The existence of trapping effects suggests a representation of scope and binding at
the same level: if we were to adopt the alternative, according to which scope relations
are determined at level L and binding relations are determined at a different level L ,
we would face the extra complexity of stating that the position of a constituent in L
constrains the position of the counterpart of that constituent in L , and vice versa.
The whole point of levels of representation is to avoid having to state such intricate
global similarities as these.
We must represent scope and binding at the same level, then, but that level should
not be LF in the traditional sense. As an alternative, I suggest a chain-based theory
of reconstruction, as in Barss (), or Aoun and Li (). Such approaches have
two main advantages, compared to approaches which treat reconstruction as the
inverse of movement, whether as lowering or as interpretation of lower copies. First,
chain-based approaches need not be so closely tied to movement, depending on
the definitions of chain that we adopt. Secondly, the possibility exists of defining
multiple types of chain, and relating different types of chain to different types of


reconstruction effect. This gives us the power to account for Lechners dissociations
between reconstruction effects.
We define the following, maintaining the agnosticism about the precise formulation
of the binding theory from footnote :
() a. X and Y are members of the same scope chain iff X and Y are related by a
grammatical dependency.
b. X and Y are members of the same binding chain iff X and Y share a -role.
() a. X can take scope from the position of Y only if X and Y are members of the
same scope chain.
b. (i) An anaphor or bound variable X can be bound from the position of Y
only if X and Y are members of the same binding chain.
(ii) A pronoun or R-expression X must be locally free or free, respectively,
in the position of Y if X and Y are members of the same binding chain.
This means that the syntactic representation fed to the semantic component encodes
information about potential scope and binding relations, but this representation does
not unambiguously represent information about these relations, in contrast to standard assumptions about copy-theoretic LFs after deletion of uninterpreted copies (see
May for the original proposal that LF did not fully disambiguate scope and binding relations). However, no great problem arises from this: the information encoded in
a representation like the one sketched here determines predicateargument relations,
as well as information on which elements may, or must not, be referentially dependent on which other elements, and which elements may be in the scope of which
Clearly, this is only an outline of a theory of reconstruction. Particular priorities for
the future are a sharpening of the definition of scope chain in the light of the patterns
of fixed scope discovered by Aoun and Li () in double object constructions; an
investigation of Lebeaux effects, or failures of Condition C reconstruction, within
these terms; and an expansion to further apparent cases of reconstruction without
movement, such as the pseudocleft connectivity discussed by Higgins () among
many others.
However, immediately, several analytical opportunities open up which are unavailable on other approaches. The system is designed to accommodate dissociations such
as those described by Lechner, which are so puzzling on more standard approaches.
Moreover, a promising treatment of movement as a composite relation emerges:
movement dependencies are simply simultaneously scope chains and binding chains,
or in other terms, grammatical dependencies which hold between elements which
share a -role.

Reconstruction, Control, and Movement

Finally, the chains we propose, unlike standard chains, arguably respect Chomskys
() Inclusiveness principle, which states that syntactic properties of complex structures project from the lexicon. Chains normally violate this principle because they
involve annotation of syntactic structures with lexicon-external diacritics. However,
the present conception of scope chains and binding chains arguably escapes this
problem, because the information about chain membership, on this view, must be
recoverable from the syntactic structure in any case: we must know in any case which
syntactic dependencies hold in that structure, and which constituents share which
-role. Given that scope chains are just syntactic dependencies and binding chains
are just shared -roles, information about scope and binding chains is already implicitly represented in the structure. The explicit labelling of these relations as chains
is redundant, and carried out purely for expository clarity. There is no violation of
Inclusiveness in the fundamentals of the analysis being developed.
We can now return to the significance of trapping effects. The core of trapping
effects is that, if a constituent enters into binding relations in a given position, it also
takes scope in that position, and vice versa. That core is independent of the relation
of movement to reconstruction effects, and can be stated in the following way.
() The Trapping Generalization
Each constituent is interpreted in only one position.
However, we can now also state a further generalization concerning the distribution of
Lechners countertrapping effects, the dissociations discussed in section ., although
() is strictly superfluous, being implicit in the foregoing.
() The Countertrapping Generalization
Individual types of reconstruction can sometimes target more positions than
multiple cooccurring types of reconstruction.
The balance between these two generalizations is determined by the interaction of
scope chains and binding chains. To illustrate, let us begin by considering (). In
(), the story is essentially as in May (): the scope chain and binding chain of two
people each consist of the surface position and the trace position, and the constituent
can in principle be interpreted in either position. In the lower position, it can be
interpreted within the scope of every senator, so scope ambiguity is predicted, unless
Principle A forces two people to bind a reflexive from its surface position, as in (b).
Because of the Trapping Generalization, if the binding theory forces interpretation
of a constituent in a given position, that constituent must also take scope in that
Essentially the same is true in (): the scope chain of two people contains the
surface position and pro. If two people may take scope from its surface position or
the position of pro, scope ambiguity arises, as in (a). However, if it must bind a


reflexive from its surface position, as in (b), then it must also take scope from its
surface position, and the scope ambiguity disappears.
() a. Two people want to dance with every senator.
( > , > )
b. Two people have persuaded themselves to dance with every senator.
( > , > )
Where raising and complement control differ is in the nature of their binding chains:
binding chains are determined by thematic relations. The trace of a raised constituent shares a -role, and so a binding chain, with its antecedent, while pro and its
antecedent have distinct -roles, and are therefore not members of the same binding
This distinction is reflected in the possibility of being bound from the foot of
the chain. The examples in (), repeated from (), demonstrate this. (a) shows
that a raised reflexive can be bound in its base position, while (b) shows that a
reflexive contained within a controller cannot be bound in the position of pro. This
is as expected if raising dependencies are binding chains but control dependencies
are not.
() a. Stories about himself appear to have upset every senator.
b. Stories about himself have managed to upset every senator.
Likewise, because a binding chain can cross a weak island but a scope chain cannot,
a failure of scope reconstruction into a weak island can bleed binding possibilities
within the weak island (see () above).
We see, then, that the Trapping Generalization, as formulated in () without reference to movement, is all that is necessary to capture trapping effects, but something
like the distinction between scope chains and binding chains is required to capture the
cases described by the Countertrapping Generalization: the Trapping Generalization
forces scope chains and binding chains to interact to determine a single locus of
interpretation for each element, but scope reconstruction is possible to positions that
binding reconstruction cannot touch, and vice versa, because of the non-identity of
scope chains and binding chains.
It follows as a theorem from this approach to trapping and countertrapping that
dissociations such as Lechners will only be found in non-movement constructions,
because scope chains and binding chains always coincide in movement dependencies.
This approach therefore rests on the correctness of theories of complement control,
scrambling, and extraction from weak islands as base-generated. Only time will tell
whether such theories are tenable, but the prize, if they can be made to work, is a
theory of the distribution of trapping and countertrapping effects, an account of cases
of reconstruction without movement, and a new reductionist approach to movement

Reconstruction, Control, and Movement

. Conclusion
The major aim of this chapter has been to sketch a theory of reconstruction effects
which takes dissociations between those effects seriously, while still allowing for intimate interactions between different types of reconstruction, such as trapping effects.
Dissociating reconstruction effects involves relating each to a different type of relation
(syntactic in one case, purely thematic in the other), while capturing the interactions
involves representing those two types of relation at the same level. Finally, the consideration of two different types of relation makes it natural to consider movement
relations as the cooccurrence of the syntactic and thematic relation.
This approach suggests several immediate avenues for exploration. Of course, there
is much work to do in sharpening this analysis beyond the outline given here. Hand
in hand with this comes the requirement that the general validity of the approach be
demonstrated, by showing that it at least does not damage our current understanding
of other patterns of scope and binding relations, and other types of reconstruction
effect. Perhaps most intriguing, though, are the implications for the general theory
of grammar. One major reason for the adoption of the copy theory of movement in
Chomsky () and much subsequent work is that it allows for a natural treatment
of reconstruction effects, building on work on layered traces in Barss () and
elsewhere. However, there has been a degree of unanswered criticism of the copy
theory (see Neeleman and van de Koot and references therein). Moreover,
Neeleman and van de Koot have also demonstrated the invalidity of claims that copy
theory is indispensable to the Minimalist analysis of movement dependencies, given
the existence of alternative Minimalist approaches to movement, such as lexicalized
traces. The present chapter reinforces Neeleman and van de Koots criticisms, by
demonstrating that an approach such as the copy theory, which ties reconstruction
effects to movement, fails to capture all instances of reconstruction effects. Both
putative advantages of the copy theory are therefore called into question, while the
criticisms of that theory remain unanswered. This chapter therefore lends support to
the programme of developing non-copy-theoretic forms of Minimalism (Brody ,
Neeleman and van de Koot ).

Linearizing Empty Edges

. Introduction
Syntax creates soundmeaning pairs. Within Chomskyan generative grammar, this
is typically described in terms of transferring (portions of) syntactic structures to
the phonological/phonetic and semantic/conceptual interfaces through PF and LF,
respectively. However, to paraphrase Chomsky and Halle (), whether the output
of the syntactic component and the input of the phonological component are the
same thing is an empirical issue. (The same, of course, holds for the semantic side.)
Indeed, since work by Selkirk and others during the early s, much work on the
syntaxphonology interface has focused on how to derive phonological domains from
syntactic structures under the assumption that these two representations are related
but non-isomorphic. This issue of domains is, however, but one piece of the much
larger puzzle of how to characterize the transformations at PF which turn hierarchical,
phonology-free morphosyntactic structures into linear phonological representations.
Within the Distributed Morphology framework, Harley and Noyer (), Embick
and Noyer (), and more recently Embick (), Samuels (), Arregi and
Nevins (), and Idsardi and Raimy () have argued that PF comprises a number
of crucially ordered operations which serve to introduce, concatenate, and group
phonological content while reducing the dimensionality of the syntactic representation. The goal of the present chapter is to further our understanding of the syntax
phonology interface by considering the interplay of linearization, copy deletion, and
vocabulary insertion within a model which assumes multiple transfer, the copy theory of movement, and direct correspondence between phonological and syntactic
The specific case study that we will focus on relates to what we will call empty
edges: phonologically contentless edges of certain domains. Whether to characterize
The authors would like to thank audiences at Syracuse University, the University of Oslo, OnLI II at

the University of Ulster, and GLOW . We are also grateful to the editors of this volume, as well as Artemis
Alexiadou and two anonymous reviewers for their helpful comments.

Linearizing Empty Edges

these domains in syntactic or phonological termsif there is indeed any distinction to

be made herewill be one of our primary concerns. Traditionally, crashes associated
with empty edges have been analysed by appealing to the Empty Category Principle
(ECP); see in particular Stowell (). Pesetsky and Torrego () and Bokovic
and Lasnik () have looked at these facts from a syntactic point of view, while An
(a, b) has argued for a prosodic account in which the notion of an Intonational Phrase plays a pivotal role. Our contribution to this debate will be to argue that
we should return to a syntacticor at least, pre-phonologicalaccount of emptyedge phenomena, and that such an account hinges on exactly how we understand the
PF interface. We will make a specific proposal concerning this architecture, show how
this new account can capture the data discussed in the literature, and extend it to cover
additional data which have not previously been addressed in treatments of the ECP.
The chapter is structured as follows. Section . gives an overview of the data and
provides a brief outline of a competing phonological proposal, namely the Intonational Phrase Edge Generalization (IPEG) by An (a, b). We offer a critique
of the IPEG and then, in section ., give a new account where the observed emptyedge phenomena follow from conditions on linearization. Section . concludes.

. A Phonological Account of the ECP

In this section we will present the data that we are going to discuss and consider Ans
(a, b) account. Because of space considerations, we will refrain from discussing previous syntactic accounts of these data; see An (a) for a comprehensive
discussion of why these accounts (in particular, Stowell , Pesetsky and Torrego
, Bokovic and Lasnik ) suffer from serious shortcomings. After the brief
data presentation, we will critically discuss Ans theory and argue that a different style
of interface account is warranted. The details of the proposal which we motivate here
will be presented and defended at length in the next section.
.. The data and a phonological account
Alternations like those in () show that, in certain circumstances, it is possible to produce sentences with restrictive relative clauses which lack any overt material higher
than TP in the embedded clause.
() a.

I saw the train [CP [IP Mary was waiting for]]

I saw the train [CP that [IP Mary was waiting for]]
I saw the child [CP [IP Mary was waiting for]]
I saw the child [CP who [IP Mary was waiting for]]

However, in very similar cases this generalization does not hold. An (a: )
considers the contrast between the above sentences and the minimally different
paradigm in ().

Lohndal and Samuels

() a. ? I saw the train yesterday [CP [IP Mary was waiting for]]
b. I saw the train yesterday [CP that [IP Mary was waiting for]]
c. ? I saw the child yesterday [CP [IP Mary was waiting for]]
d. I saw the child yesterday [CP who [IP Mary was waiting for]]
While it has long been known that an overt C (b) can ameliorate a sentence like
(a), An points out that filling Spec,CP can also license a postposed relative clause,
as in (d).
The paradigm in () is not the only one where an empty edge seems to be associated
with degradation. The following examples illustrate that we find such cases in many
() a. I believe very strongly [CP that [IP John liked linguistics]]
b. I believe very strongly [CP [IP John liked linguistics]]
(An a: )
() a. [CP That [IP the teacher was lying]], Ben already knew.
b. [CP [IP the teacher was lying]], Ben already knew.
(Stowell )
() a. I distrust [NP the claim [CP that [IP Bill had left the party]]]
b. I distrust [NP the claim [CP [IP Bill had left the party]]]
(Stowell )
() a. Mary believed that Peter finished school and Bill, [CP that [IP Peter got a job]]
b. Mary believed that Peter finished school and Bill, [CP [IP Peter got a job]]
(Bokovic and Lasnik )
The question is, what renders these sentences with phonologically empty CP-edges
unacceptable? For some of the cases above, the adjunct yesterday or very strongly seems
like the obvious culprit; for example, () and () differ only in the presence or absence
of such an adjunct. But that does not account for the data we just presented in ()().
Nor is CP the only category involved in this phenomenon: unacceptability correlated
with empty edges also extends to other phrases, including vP (), DP (), and AP ().
() a. Eat the cake John did and eat the cookie Mary did.
b. [vP Eat the cake] John did and [spec v the cookie] Mary did.
() a. John likes this book of linguistics and Mary, that book of physics.
b. John likes this book of linguistics and Mary, [spec D book of physics]
() a. Eager to win the Pulitzer Prize, John is, and eager to win the Nobel Prize,
Mary is.
b. [AP Eager to win the Pulitzer Prize], John is, and [spec A to win the Nobel
Prize, Mary is]

Linearizing Empty Edges

Based on these facts, An (a: ) argues for the following generalization:

() Intonational Phrase Edge Generalization (IPEG)
The edge of an I-phrase cannot be empty (where the notion of edge encompasses the specifier and the head of the relevant syntactic constituent).
Let us provide some background. For Selkirk (), an I-phrase is a sequence of pitch
accents that are flanked by a boundary tone or pause. Nespor and Vogel () observe
that grammatically related pauses coincide with the end of an I-phrase. Specifically,
Nespor and Vogel () argue that certain constituent types are obligatorily (and
cross linguistically) parsed as I-phrases. These include root clauses, parentheticals,
tag questions, vocatives, non-restrictive relative clauses, and some moved elements
(Selkirk , , , Nespor and Vogel , Bokovic , An a).
An (a) argues that all the environments above where we have seen empty
edges being impermissible are contexts in which the phrase with the empty edge is
obligatorily parsed as a separate I-phrase. An attempts to rationalize this by adopting
the following assumptions (see An a: and references cited there):
() a. I-phrasing must occur at the juncture between two prosodic words.
b. I-phrases are isomorphic with syntactic constituents that are obligatorily
parsed as I-phrases.
c. Prosodic words can be informally defined as phonologically independent
words that bear stress.
Consider the case where the edge of an obligatory I-phrase domainfor the sake of
concreteness, the edge of a CPis phonologically null. The third assumption prevents phonologically null elements from building prosodic words, and by the first
assumption, only prosodic words can host an I-phrase boundary. The first prosodic
word, and hence the first context eligible to host an I-phrase boundary, will be located
somewhere within IP. However, the second assumption says that the edge of the CP
the obligatory I-phrase domainshould coincide with the I-phrase boundary. Thus,
a CP with only null elements in its edge creates a mismatch between the syntactic
domain and the prosodic domain, producing a deviant sentence.
Thus far, we have only discussed what happens when the left edge of an I-phrase
is empty. In other work, An (b: ) proposes a symmetrical alternative to the
IPEG that addresses what is permissible at both the left and right edges by adding
the additional condition that only null elements, not unpronounced copies, trigger
() The Ban On Null Edge (BONE)
If a clause is parsed as a separate I-phrase, its left and right boundary positions
cannot be occupied by phonologically empty elements, where the phonological
emptiness stems from non-insertion of phonological content.

Lohndal and Samuels

The dichotomy between traces/copies and other null elements is necessary given that
sentences like (), with right edges that have been evacuated via movement, are
() [CP Who did Mary [vP see [VP see who]]]
This raises the question of which interface condition should be preferred, given that
the two appear to be extensionally equivalent due to the distribution of null elements
and traces/copies: a general but asymmetrical statement like the IPEG, or a symmetrical one which is more restricted like the BONE? We will provide an alternative to
both: a symmetry-preserving solution under which the distinction between copies
and null elements follows from the order of operations at the interface.
.. Problems with prosodic accounts
Despite the empirical mileage of statements like the IPEG/BONE, the justification
for such generalizations is unclear. They do not appear to follow from any independent facts about the architecture of grammar. One could leverage the fact that CPs
roughly correlate with I-phrases and pursue a theory in which IPEG violations are
disallowed because the emptiness of the CP edge conflicts with instructions to create
(and presumably mark the edge of) a new I-phrase, perhaps resulting in prosodic
malformation. Along these lines, An (a: ) suggests that there is a similarity
between the [Phase Impenetrability Condition] and the IPEG: both conditions give a
special status to the edge of CP. One argument against this view is that (), (), and
() show other phrases, namely vP, DP, and AP, exhibiting the same phenomenon; CP
is not special. Note especially that AP, a non-phase category in virtually all theories
(see Baker for an exception), participates. Moreover, we note in passing that we
(admittedly subjectively) find nothing intonationally aberrant about sentences like
(b) or (b); rather, they have the flavour of syntactic violations, and indeed, more
than two decades of literature have treated them as such.
An (a) also gives another justification for the importance of IPEG, namely that
the specifier and head of a phrase are counted as a unit at PF, as demonstrated by the
fact that filling either Spec,CP or C has an ameliorating effect in (). The following
quote illustrates the reasoning behind this idea:
[The current approach] suggests a partial answer to the conceptual question why specifier and
head are special in the syntax. Note that for a long time, it has been taken for granted that
the two positions enjoy a special relationship to each other, frequently instantiated as Spec
head agreement. However, there may not be an a priori reason for this. For example, X-bar
geometrically, the relation between head and complement may in fact be regarded as closer
than that between specifier and head. However, the behavior of specifier and head with respect
to I-phrasing suggests that they are counted as a unit in PF. Of course, this is not an explanation.
However, from this point of view, it seems more natural that the specifier and the head enjoy

Linearizing Empty Edges

a close relationship to each other, because they are treated as one element at the interface.
(An a: )

This argument leads to an undesirable syntactic consequence: in order to achieve

the interchangeability of specifier and head for IPEG-satisfying purposes, An is
forced to assume that specifiers are always projected, which is not the case under
Bare Phrase Structure (Speas , Chomsky , Alexiadou and Anagnostopoulou
). Within Bare Phrase Structure, only those positions that are needed are projected. For most approaches, this entails that the head is always merged and that the
specifier is optional, though Starke () argues that it is possible to have a specifier
without a head.
Related to this, adopting the IPEG forces analyses that are incompatible with other
aspects of mainstream Minimalist theory, even for the most basic sentences. To take
one example, in English, simple matrix clauses like () need not have any overt
material in Spec,CP or C.
() Mary reads books.
The I-phrase for this sentence will be the IP, hence there cannot be a covert C in the
syntax. To circumvent the IPEG at PF, An (a: ) is therefore forced to posit
covert insertion of C at LF. There are two problems with this solution. One is that late
insertion of C violates the Inclusiveness Condition (Chomsky ). This is because,
in most current work, C is needed in sentences like () for clause-typing reasons, in
which case the nature of the C head matters for interpretation (see Lohndal and Pietroski for a recent proposal). Covert insertion of a C would therefore be adding
meaning, which is banned by the Inclusiveness Condition. The second problem is that
if M. Richards () and Chomsky () are correct to argue that phi-features on
T need to be inherited from C, then matrix clauses need to have C present in narrow
Lastly, there is also an empirical concern. In order to allow for (b), An (a)
needs to assume that the formation of a separate I-Phrase for clausal complements of
verbs (and restrictive relative clauses) is optional, thus licensing an empty CP-edge
in (b) just in case the embedded CP does not form its own I-Phrase. But it is not
obvious that there is any difference in phrasing possibilities between (a) and (b).
() a. I believe [CP that [IP John liked linguistics]]
b. I believe [CP [IP John liked linguistics]]
For these reasons, we believe that the IPEG (and the BONE, which shares its shortcomings) raises more questions about empty-edge phenomena than it answers and
faces serious challenges both theoretically and empirically. In the next section, we

There is another concern regarding the BONE in particular. It predicts that object pro-drop should
only be possible in SOV languages. This is not the case, and our theory does not make this prediction.

Lohndal and Samuels

provide a new account which locates the source of unacceptability in failed linearization rather than the ECP or problematic intonation.

. Towards a New Account of Empty-Edge Effects

.. The road to phonology
We argue that the unacceptable cases shown in section . stem from a problem
with linearization: two identical objects, such as the null elements in these cases,
are unable to linearize due to an irreflexivity constraint on linearization (cf. Nunes
, Richards ). This constraint says that it is impossible to establish an order
between two identical objects. Under such an account, identical objects (we will
expand on this notion shortly) will be problematic anywhere, not just at an I-Phrase
The linearization algorithm that we propose takes c-command as the basic determiner of linear order at Spell-out. This is similar to Epstein, Groat, Kawashima, and
Kitaharas (: ) Precedence Resolution Principle and the algorithm suggested
by Richards (). Our specific formulation is provided in ().
() Linearization algorithm: When encountering the merged pair {, }, and
c-commanding each other, upon Spell-out, create ordered pairs ,  and
, . Ignore the pair that does not fit the parameter of the relevant head.
We follow Epstein et al. () and Richards (), who point out that mutual
c-command overdetermines linearization. As precedence statements (ordered pairs)
are created, adjacency information is simultaneously present with hierarchical information, as is required for morphological operations such as Local Dislocation
(Embick and Noyer ). We argue that copies are marked for deletion subsequent to
adding the adjacency information, but prior to rejecting half of the linearization statements (see also Richards ). Vocabulary Insertion (the operation which exchanges
syntactic/semantic features for phonological ones) also occurs before copies are
marked for deletion. Nodes that are silent will have nothing but whatever property
enables them to be merged, for instance Edge Features/Properties (Chomsky ,
Boeckx ) or labels (Hornstein ). In this chapter, we remain fairly agnostic on
what exactly this property is, but we will notate it as EP in what follows.
To preview our analysis: at linearization, the relevant formal features have been
eliminated already by Vocabulary Insertion, and phonological features are now
present. Copy deletion will be argued to follow linearization, so only truly null items
However, note that cases such as Will will come tomorrow show that the constraint cannot be formulated solely in terms of phonological features. The fact that these two instances of will are separate lexical
items, or have different labels, seems to play a role here. This could suggest that it is a label, rather than EP,
which persists to mark an elements position in the workspace.

Linearizing Empty Edges

will create problems for linearization, not copies. Similarly, reduplication occurs at a
later stage (following Raimy et seq.), so identical elements created through that
process are permitted.
The key component to our analysis is that the linearization algorithm only succeeds
in reducing the adjacency statements to a single ordered pair when the elements to
be linearized are distinct. If the elements are non-distinct, the derivation crashes at
that point. A special case of this is when two adjacent elements only have EPs: these
elements are non-distinct, so they will cause linearization to crash. We need to define,
then, exactly what features are present at the time of linearization, because this is what
determines the permissible configurations.
The conclusion and order of operations for which we argue accords with more
general observations about the effects of unpronounced copies and null elements
in morphophonology; see Samuels (: .). The correct generalization seems to
be that morphophonology knows about null elements, which trigger phonological
cycles and produce cyclic blocking effects, but that unpronounced copies are phonologically inert, setting aside phrasal stress, which we will discuss shortly. These wellknown morphophonological effects argue (contra Embick ) that unpronounced
copies are eliminated earlier than null elements, the latter persisting until the handoff
to phonology proper.
We schematize the relevant steps along the path from syntax to phonology as

, , ,

vocabulary insertion

copies marked for non-pronunciation

marked copies deleted


This may be one way of explaining Raimys () arguments that reduplication does not create identity
or dependency between the base and reduplicant, unlike copying of syntactic elements for internal Merge.

Lohndal and Samuels

This sketch is not meant to be an exhaustive list of the operations which occur at PF.
For instance, phrasal stress must be assigned before it has been determined which
copy will be pronounced: as has been noted since Bresnan (), stress assigned to
an unpronounced lower copy (to put it in modern terms) shows up on a higher copy,
as in the object question below.
() What book did you read what book?
Following Kahnemuyipour () and Sato (), stress is initially assigned to the
lower copy of what book on the first derivational cycle and carries over to the higher
copy which is ultimately pronounced. Insofar as c-command is relevant to selecting
which copy to pronounce (Nunes ), all of this must occur before the tree is
linearized. It also seems to imply that copy deletion cannot be characterized as simply
failing to perform Vocabulary Insertion on the node which will not be pronounced;
the node must be filled and its full phonological content copied, with the original
copy only subsequently being deleted. For this reason, we depart from Idsardi and
Raimy (), who (following Embick and Noyer ) view the insertion of phonological features as concurrent with the establishment of linear precedence relations
and elimination of hierarchical relations along an already immobilized structure.
We will see in the next section why it is crucial that marking copies for deletion and
actually deleting them are sequentially separate.

.. Our account: empty-edge effects follow from linearization

With this roadmap in place, we are now ready to reanalyse the data discussed in
section .. A preliminary question concerns what to do with the contrast between
(a) and (b), which seems to hinge on adjunction:
() a. I saw the train [Mary was waiting for]
b. I saw the train yesterday [Mary was waiting for]
Chomsky (: ) argues that an adjunct is attached to [the syntactic object
to which it adjoins] on a separate plane from the primary plane of the main clausal
spine. We suggest more concretely that adjuncts need to be connected to the main
spine through EPs. That is, as Chomsky also notes, adjuncts need a placeholder in the
main structure to keep track of their scopal properties. One way to implement this
is to say that adjuncts are spelled out separately from the main spine in Uriagerekas
() sense, but that the EP shows where in the main structure the adjunct belongs.
Following this reasoning, (b) has the simplified structure in ().

I saw the train [

EP ] [CP

EP Mary was waiting for]

Linearizing Empty Edges

() a.

the train




Mary was waiting for
linearization OK








The structure in (), shown in the tree in (a), is generated as follows. The IP is generated in the usual fashion; we make no special assumptions regarding its structure.
Then a null C is merged, which is indicated by the first EP , i.e. the sister of IP. The
label of the empty head EP projects, and we get the EP that dominates both EP and
IP. Next, another EP is merged to provide a placeholder for the adjunct yesterday.
Then the verb see is merged and the DP the train is merged as its specifier.
We hypothesize that the problem with the sequence of EP nodes in this structure
emerges once it has to be linearized. We assume that a graph is linearized from the
bottom, and that each node, including labels, counts for the purposes of linearization.
The small trees in (b) and (c) show the relevant parts of the linearization. First
the EP and IP sisters are to be linearized, which does not create a problem since they
are distinct. The problem emerges when the adjunct head EP and its sister, EP , are
to be linearized. Because these two nodes are identical, the linearization mechanism
crashes at this point.
There is one potential worry that we need to address, concerning the possibility of
adjacent adverbs. An example is provided in ().
() I saw the train immediately yesterday.
Our claim is that one EP acts as a placeholder to show where the whole series of
adjuncts goes. The adjuncts are, as Uriagereka () and Chomsky () have
argued, on a different plane from the main structure. In cases where there are multiple
adjuncts that are adjacent, these have to be ordered with respect to one another on
this separate plane. There, the null hypothesis should be that the same linearization

Lohndal and Samuels

principles familiar from the regular clausal spine apply, and the ordering of the
adjuncts as determined on their plane is later integrated with the main structure. This
is one way of cashing out Chomskys claim that adjuncts are pair-merged rather than
Notice that the crash above was caused by the adjunct being adjacent to a null C;
the specifier of CP played no role, and indeed there is no reason to assume that the
specifier was projected. Omitting Spec,CPconsistent with Bare Phrase Structure
in the case of all restrictive relative clauses and embedded CPs, there is no need for
special pleading in () or (b), which are problematic for An (a) for reasons
discussed in the previous section.
() a. I believe [CP EP [IP John liked linguistics]]
b. [CP EP [IP Mary was waiting for the child]]
If we follow Starkes () analysis, again consistent with our Bare Phrase Structure
approach, we see why adjuncts to a matrix clause like the ones below (from Selkirk
) pose no problem; in Starkes system, the adjunct in CP and the subject in TP
are adjacent, with no null element intervening.
() a. [CP [AP True to herself], [TP she planned to remain there]]
b. [CP [PP On the fourth of July], [TP well have a parade and fireworks]]
Even though the adjunct in CP is connected through an EP, there will only be one
such bare EP, and hence no problem for linearization. Thus there is no empty C in
these cases, as Starkes analysis allows for there to be a phrasal CP without a C head.
This is shown in the following structures:
() a. [CP EP [TP she planned to remain there]]
b. [CP EP [TP well have a parade and fireworks]]
Such cases are therefore correctly predicted to be acceptable.
Now we must also consider the case of a null edge created by movement, as in (),
repeated below.
() [CP Who did Mary [vP see [VP see who]]]
The acceptability of such sentences follows naturally from the order of operations for
which we argued on independent grounds: marking one copy for non-pronunciation
renders it distinct from its neighbour, so it can be linearized without incident.
Let us also look at cases of crashes involving other categories. First, vP:


Eat the cake] John did and [v [ V the cookie]] Mary did.

Here we follow Selkirk () in treating these as CP adjuncts. Nothing would go wrong if they are
instead TP-level adjuncts, as long as the adjuncts are not adjacent to a null C.

Linearizing Empty Edges

Since both the v and V nodes are instantiated by null elements with only EP, the
linearization algorithm crashes. Note that even if these heads have some semantic
contentfor example, if the v head is really a silent causeit would not make a difference. On the PF side, silent elements only have EPs since their semantic/syntactic
features have already been converted to phonological ones via Vocabulary Insertion
prior to the point of linearization. So a silent cause behaves exactly as any EP on the
PF side, as desired. Similarly, if one wants to argue for a silent that in (), this would
be fully compatible with the analysis we have proposed.

I saw the train yesterday

[that Mary was waiting for]

We can also account for the unacceptability of nominal complementation without an

overt C, as in (b), if the embedded CP is within a covert PP (cf. Kayne, in press).
() a. I distrust the claim [PP P [CP that Bill left the party]]
b. I distrust the claim [PP P [CP C Bill left the party]]
This is motivated in part by languages like Norwegian, in which this structure appears
overtly (a). In fact, in Norwegian, the preposition is obligatorily pronounced, as
shown in the examples below.
() a. Jeg stoler ikke p pstanden [PP om [CP (at) Bill forlot festen]]
*(that Bill left the party
I trust not on the claim
I distrust the claim that Bill left the party.
b. Jeg stoler ikke p pstanden [PP [CP at Bill forlot festen]]
I trust not on the claim
that Bill left the party
I distrust the claim that Bill left the party.
Note that the complementizer at that cannot be dropped.
Our account also gives insight into a difference between finite and non-finite
clauses, if we adopt the movement theory of control (Hornstein , Boeckx, Hornstein, and Nunes ):

A consequence of this account is that there cannot be multiple null heads in the skeleton of a clause
unless these null heads are separated by specifiers.
The analysis does not extend to the following unacceptable sentences, involving a sentential subject
(ib) and a topicalized CP (iib).

(i) a. [CP C [CP That [IP John is a genius]] was believed by many people]
b. [CP C [CP C [IP John is a genius]] was believed by many people]
(ii) a. [CP C [CP That [IP John is a genius]], Mary believed]
b. [CP C [CP C [IP John is a genius]], Mary believed]
Something else must be going on here, which may be related to a semantic need for a complementizer to
encode factivity.

Lohndal and Samuels

() a. I believe very strongly [CP Mary liked linguistics]

b. John wanted at that time [IP John to leave]
The null C in the embedded finite clause (a) causes a crash upon attempted linearization with the adjunct in the matrix clause, whereas due to the timing of copy
deletion, the adjunct in the control case (b) is adjacent to the copy of John at the
time of linearization (though it has already been marked for deletion) and the process
will therefore proceed smoothly.
The sentence in () may appear to create a problem for our approach:
() John had had dinner already.
However, at the time of linearization, the sentence looks roughly like:
() [John had [John [had dinner already]]]
The two phonologically identical copies of had are not adjacent during linearization
because a copy of the subject, even though it is not ultimately pronounced, intervenes
at the relevant stage of the derivation.
Another question concerns ellipsis. If ellipsis involves deletion, why dont all cases
of ellipsis generate a linearization crash? The following example is a case of VP-ellipsis.
() John can play the saxophone, and Mary can play the saxophone too.
There are several solutions one can pursue here, depending on exactly how ellipsis is
analysed. For reasons of space, we will not delve deeply into this subject here. Suffice it
to say that the copies in the ellipsis site will be marked for non-pronunciation, which
still renders the copies distinct. Hence, no linearization crash emerges.
Lastly, one may wonder about pro-drop languages. The C and Spec,TP are both
empty in such cases, so should that not cause a linearization crash? As for expletive
pro, we follow Alexiadou and Anagnostopoulou () and others who argue that
pro-drop means that the T head carries the force of the EPP in these languages. Consequently, no Spec,TP is created, and there is no problem for those languages. Referential
pro-drop is different, and there a pro has to be present. This can be maintained in our
account as long as the verb raises to T, because otherwise both T and Spec,TP would
appear to be empty. There is independent evidence that the verb raises to T at least in
languages like Italian (Dobashi , Frascarelli , Samuels ).

. Conclusion
We argue that the unacceptable sentences in sections . and .., previously
attributed to ECP or IPEG violations, crash due to a failure during linearization.
Specifically, two syntactic objects cannot be linearized if they bear only EP and are
thus non-distinct at the point when the linearization algorithm () applies. Our

Linearizing Empty Edges

account eliminates the troublesome assumptions that plagued Ans account: reliance
on the notion of violations of Bare Phrase Structure, late insertion of C, suspicious
intonational phrasing, and the mysterious relationship between IPEG violations and
unacceptability of a syntactic flavour.
As a final note, we would like to place the account for which we have argued here
in a larger context. Boeckx () argues that the empty edge phenomenon is one
manifestation of a more general filter XX which applies to different types and tokens
of linguistic objects at different levels of representation. This more general principle would then supersede not only the IPEG but also the melui constraint (Kayne
, Bonet ), the Obligatory Contour Principle in phonology (Leben ),
the whatwhat filter familiar from multiple wh-movement languages (Bokovic and
Nunes ), and more. That is to say, non-distinctness is dispreferred across various
linguistic domains, though the levels at which it is disallowed may vary from language to language. The present work also sits alongside arguments by Samuels (:
Chapter ) that phenomena such as the IPEG and second-position clitic placement
need not be described in terms of prosodic phrasing mediating between syntactic and
phonological representations; purely syntactic representations will suffice to describe
the pattern of acceptable and unacceptable configurations.

Evidence for the Use of Verb Telicity
in Sentence Comprehension

. Introduction
One fundamental question for structural models of sentence comprehension which
assume that syntax plays a crucial role in the first stage of parsing is what kind of
syntactic information is available during that stage. The experimental studies reported
here take as a starting point two classifications of verbs which have played a crucial
role in the study of verb alternation and the syntaxlexicon interface: a classification
in terms of argument structure and a classification in terms of aspectual structure. The
first type of classification, which (following Perlmutter ) proposes a tripartite distinction between transitive, unaccusative, and unergative verbs, has been previously
shown in the processing literature (MacDonald , Stevenson and Merlo ) to
be crucial (see sections .. and .. below). In this chapter we concentrate on the
second classification of verbal predicates, in terms of aspectual information (Vendler
), and present evidence that event structure information, in particular telicity, is
accessed by the processor during online comprehension.
.. Event structure and telicity
Aristotle first noticed that verbs can be divided into states and events and that only
a subset of eventive verbs include in their meaning the idea of an endpoint or telos.
Since Vendlers () reanalysis of Aristotles distinction, the linguistic literature on
verbs has made extensive use of a classification of verbal predicates based on aspectual
This project was supported by the University of Arizona Cognitive Science Program, by NSF Grant
No. to OBryan and Bever, by a British Academy Postdoctoral Fellowship, and by the University
of Ulster. We are grateful to Selene Gardner, Roeland Hancock, Benjamin Jones, Andria Ligas, and Hollis
Weidenbacher for assistance with the preparation and running of the experiments and to Jason Barker, Ken
Forster, Ken McRae, Wayne Murray, Christopher Nicholas, Peter Richtsmeier, an anonymous reviewer, and
David Townsend for suggestions and advice. Finally, we wish to thank Ken McRae for providing the raw
data from his study for the post hoc reanalysis.

Verb Telicity in Sentence Comprehension

information. The event denoted by a verb may entail the existence of a process, may
entail an endpoint, may be homogeneous or have distinct stages of completion, and
may be punctual or non-punctual. Accordingly in Vendlers schema, verbs divide
into four classes, a class of verbs indicating states and three types of events:
() a. Activities (e.g. applaud): events that go on for a time, with homogenous subparts and without a necessary endpoint. They are atelic.
b. Accomplishments (e.g. write the book): events that have a process portion
which proceeds via distinct stages to a logically necessary endpoint. They
often involve a transition, followed by a resultant state. They are telic.
c. Achievements (e.g. stop or trip): change-of-state events that culminate at the
moment they occur and lack a preceding process. They are telic.
Vendler established these classifications with a number of syntactic tests. One famous
test which is often used to distinguish between telic and atelic events is the modification by adverbials such as for an hour/in an hour/at oclock.
() a. Mary pushed the cart (in/for an hour)(at oclock).
Activity (atelic)
b. Mary wrote the book (in/for an hour)(at oclock).
Accomplishment (telic)
c. Mary crossed the finish line (in/for an hour)(at oclock).
Achievement (telic)
These tests are based on the fact that adverbials like for an hour modify atelic events,
whereas in an hour modifies telic events. If the activity can extend for some time
period such as for an hour without any endpoint being specified, then this implies
homogeneity and atelicity. In contrast, if the event denoted by the verb can be completed in some delineated time period such as in an hour, then the verbs meaning
implies completion, and thus heterogeneity, an endpoint, and telicity. The third test,
at oclock, distinguishes between telic events which unfold over an extended period,
called accomplishments, and telic events which occur at a single moment, called
Much work in theoretical linguistics has taken this as a starting point and has
convincingly argued that event structure has implications for syntactic structure. For
example, it has been observed that telic events always have objects in their syntactic

The symbol indicates semantic infelicity.

When a verb is composed with an adverbial that is incongruent with its telicity specification, additional
interpretive effects appear. For example, if a telic verb is composed with an atelic modifier such as for
an hour, it may be interpreted as iterated or incomplete (consider e.g. Mary crossed the finish line for an
hour, or Mary wrote the book for an hour). When an atelic verb is modified by in an hour, the event is
interpreted as peculiarly habitual or having some unspecified conventional extent in the context (consider
Mary pushed the cart in an hour). These additional interpretive effects are referred to as aspectual coercion
(see e.g. Piango, Zurif, and Jackendoff ).

OBryan, Folli, Harley, and Bever

structure, albeit sometimes only underlyingly (Tenny ). The question of how to

represent the connection between telic events and syntactic objects has been the focus
of much work on the lexiconsyntax interface. More generally, in the literature the
observation that events might have complex internal structure (possibly involving
causation components, process components, and endpoints) has caused many to
propose that since event structure determines the mapping to syntax, complex event
structure must surface in the syntax as complex phrase structure (e.g. Pustejovsky
, Borer ). In other words, events might contain subevents and these idealized
subparts have consequences for the syntax. For example, an endpoint subpart in
the aspectual semantics of a verb will be represented in the syntax by a particular
functional projection. Verbs denoting different types of events, then, will be associated
with different syntactic structures (Travis , van Hout , Borer , Ritter and
Rosen , Ramchand , among many others).
The syntactic structure related to the subevents associated with a particular verb is
the event structure of that verb. In this perspective, each argument of a verb occupies
a given position associated with a subpart of the event in the syntactic structure.
The interpretation of the argument itself is determined by the aspectual properties
associated with that projection.
Considering now specifically telic events, the correlation between event endpoints
and syntactic objecthood follows from the fact that the endpoint functional projection
is lower in the syntactic structure than the base position of the external argument. The
event in John broke the vase is telic because it implies that the theme the vase reaches
a final change of state. In the event structure analysis, the vase undergoes movement
to occupy the specifier position of the telic functional projection, labelled TelicP
in ().
() . . .




the vase




Verb Telicity in Sentence Comprehension

Because the telic functional projection is situated lower than the position of the
external argument John, only internal arguments (like the vase) are available to fill its
specifier. Hence only internal arguments may affect the telicity of a verbal predicate.
.. Telicity is not reducible to transitivity
From the discussion above, it might appear that transitivity and telicity are indistinguishable, in that both involve the presence of a syntactic object. This is not the case,
however. Although telicity requires the presence of a syntactic object, the converse is
not true. There are transitive verbs with syntactic objects that are not telic, for example
() and ().
() John pushed the cart (for an hour/in an hour).

(Atelic, transitive)

() John drove the bus (for an hour/in an hour).

(Atelic, transitive)

Similarly, there are purely unaccusative verbs which are telic and project a single
internal argument occupying the specifier of the telic functional projection first and
then raising to became a derived subject. These purely unaccusative verbs have no
transitive variant:
() a. John arrived.
b. John arrived Mary.

(Telic, intransitive)

() a. John died.
b. John died Mary.

(Telic, intransitive)

Finally, when a transitive verb is not telic, the object does not occupy the specifier of
the telic functional projection. Likewise, the agentive subject of an unergative intransitive verb like whistle initially occupies the higher external-argument projection, and
can never appear in the specifier of the lower telic functional projection.
These syntactic consequences of telicity led us to investigate the question of
whether the processing system makes use of verb event structure information
during online sentence comprehension. The independence of telicity and transitivity allows us to ask if each of these two factors independently affects processing. We investigated these questions using the main clause vs. reduced relative
.. Investigating the comprehension of structurally ambiguous sentences
The comprehension of reduced relative clauses has been used to examine which factors are immediately relevant to the comprehension process (Bever et seq.). The
reduced relative clause sentence (a) has the same beginning as the simple sentence
in (b). (c) is the immediately unambiguous version of (a).

OBryan, Folli, Harley, and Bever

() a. The actress tripped by the writer was eccentric.

b. The actress tripped the writer.
c. The actress who was tripped by the writer was eccentric.
If, during the comprehension of (a), the sentence-processing system interprets the
initial noun phrase the actress as the subject and agent of the verb tripped, as in (b),
this simple main-clause interpretation must be reversed later in the sentence. The
preposition by, in conjunction with the past tense morpheme on the verb tripped,
is the first possible cue that the sentence could have a reduced relative structure.
When the processor reaches the actual main verb was, the actress must be treated
as the object of tripped. In structural models of parsing, the degree of comprehension
difficulty in reduced relative clause sentences is a result of the processors degree of
commitment to the main-clause interpretation. The unreduced relative clause version
of the same sentence, (c), serves as a baseline measure of sentence difficulty, because
it is unambiguous. Thus, the reduced relative effect can be measured as the processing
time for (a) minus the processing time for (c) in the ambiguous and disambiguating
regions of the sentence (. . . tripped by the writer was eccentric). The degree to which
a factor affects commitment to the main-clause interpretation can be determined by
how much it affects the size of the reduced relative effect.
.. Argument structure and the reduced relative effect: MacDonalds study
Previous research (MacDonald ) has shown that the reduced relative effect is
affected by the argument structure of the verb, as predicted by Pritchett (). In a
reduced relative clause sentence, the initial verb must receive a transitive reading when
the initial noun phrase is eventually interpreted as the direct object. Verbs may differ
in the degree to which they promote a reduced relative reading in English because
transitive verbs can be either obligatorily transitive, such as capture, or optionally
transitive, such as fight. In a self-paced reading experiment, MacDonald found that
less processing difficulty occurred in reduced relatives with obligatorily transitive
verbs as in (a) than with optionally transitive verbs as in (b). This effect was found
on the main verb (was in ()).
() a. The ruthless dictator captured in the coup was hated throughout the country.
b. The ruthless dictator fought in the coup was hated throughout the country.
This result suggests that the information that a verb requires an object may lead the
comprehension system to be less committed to a main-clause reading of the initial
portion of the sentence. One possible explanation is that if the processor accesses the
information that a verb requires an object, it leaves open the possibility that the initial
noun phrase might be the required object. For example, in (a) the verb capture is
stored with the information that it requires a direct object. This information may
influence the processor to leave open the possibility that the ruthless dictator is the

Verb Telicity in Sentence Comprehension

direct object in a reduced relative construction. Thus, in this case the processor would
not immediately commit to the analysis of the ruthless dictator as the subject of a
main clause. The reduction in the reduced relative effect associated with obligatorily
transitive verbs has been found in several subsequent studies (Townsend and Bever
: ).

.. Stevenson and Merlo ()

Stevenson and Merlo () found that different argument structures result in significantly different acceptability judgements of reduced relative clauses within the class
of optionally transitive verbs. Optionally transitive verbs can be sorted into unergative
or unaccusative groups based on their intransitive form. Reduced relatives formed on
unergative optionally transitive verbs such as race are less acceptable than reduced
relatives formed on unaccusative optionally transitive verbs such as melt (one of the
original pairs of examples were the horse raced past the barn fell vs. the witch melted
in the movie deserved her fate). Stevenson and Merlo argue that this difference is the
result of the different syntactic structures associated with the two classes of verb, the
former having an underlying subject and the latter an underlying object argument.
In other words, they find argument structure effects in reduced relatives that are
independent of obligatory transitivity. Given the connection between telicity and the
underlying object position, it is possible that these effects may reflect the structural
properties of telic predicates.

.. Telicity and the reduced relative effect: the Event Structure Processing hypothesis
The previous sections briefly reviewed the syntaxsemantics literature concerning the
connection between a verbs argument structure and its event structure. We reviewed
the hypothesis that telic events are associated with a syntactic structure with a particular functional projection representing the endpoint of the event. The effect of this
superstructure is that telic events must represent an underlying direct object syntactically, filling the specifier of the telic functional projection. This raises the question
of whether verb event structure influences online structural ambiguity resolution (see
Sanz , and Townsend and Bever : ).
Because many transitive verbs are also telic, the possibility that event structure
influences online resolution represents a potential confound in MacDonalds findings.
Most of MacDonalds () optionally transitive verbs were atelic (fifteen atelic verbs
vs. three telic verbs), and most of her obligatorily transitive verbs were telic (fifteen
telic verbs vs. three atelic verbs). Thus, MacDonalds finding could be due to telicity
rather than transitivity. Since telic verbs must have internal arguments, we predict less
processing difficulty in reduced relatives with telic verbs such as (c) and (d) than
in those with atelic verbs such as (a) and (b).

OBryan, Folli, Harley, and Bever

() a.

Atelic, optionally transitive: The actress sketched by the writer left in a hurry.
Atelic, obligatorily transitive: The actress escorted by the writer left in a hurry.
Telic, optionally transitive: The actress tripped by the writer left in a hurry.
Telic, obligatorily transitive: The actress noticed by the writer left in a hurry.

Both telic and atelic verbs can be obligatorily transitive or optionally transitive,
as shown in (). This allows us to tease apart the relative effects of telicity and
The hypothesis of the current research is that event structure is not simply a byproduct of argument structure, but rather is an independent semantic feature of the
verb, with independent consequences for the processing of syntactic structure. We will
refer to this hypothesis as the Event Structure Processing hypothesis. The endpoint
part of a predicates meaning entails having a theme argument filling the specifier
position of the telic functional projection. This information is represented in the verbs
lexical entry (see Folli and Ramchand for proposals along these lines), and
we hypothesize that it is used immediately by the processor in online comprehension.
Our goal in this chapter is to show that telicity has an effect on comprehension that is not reducible to transitivity. We will present results from four different
comprehension experiments, including prior and current studies, designed specifically to investigate this question. First, we review prior experiments bearing on the

. Prior Experiments Bearing on the Hypothesis

.. Self-paced reading results
As a preliminary test of the Event Structure Processing hypothesis, OBryan ()
conducted a post hoc analysis of two prior published self-paced reading experiments on the reduced relative effect (McRae, Spivey-Knowlton, and Tanenhaus ,
Tabossi, Spivey-Knowlton, McRae, and Tanenhaus ). The post hoc analysis
involved categorizing the verbs from the prior studies and reanalysing reading times
in order to examine the effects of telicity and obligatory transitivity. The experimental
task in both prior studies was self-paced reading in moving-window format, with two
words presented at a time. The embedded verb always occurred in the same region
with the preposition by as shown in (a).
() a. The trainee / taught by / the specialists / was better / skilled than / the others.
b. The trainee / who was / taught by / the specialists / was better / skilled than /
the others.
The unambiguous control sentences, such as (b), were the same as the experimental
sentences except for the insertion of who was. The two experiments used each verb
only once. Thus, the number of items in each experiment was equivalent to the

Verb Telicity in Sentence Comprehension

Table . Mean self-paced reading times in post hoc reanalyses of two prior





The N
by the N
mainV. . .






verb-ed by the N

main verb







The N
by the N
mainV. . .


The N
by the N
mainV. . .










The N
by the N
mainV. . .



number of unique verbs. Combining across both studies, there were a total of items
and native English-speaking participants.
For the post hoc analysis, the verbs were categorized as obligatorily or optionally
transitive based on the results of a questionnaire administered to eight monolingual
English speakers. The verbs were categorized as telic or atelic by three judges, and the
categorizations were then confirmed using an in an hour versus for an hour forcedchoice task administered to native English speakers.
Table . shows the mean self-paced reading times in both studies combined for
each of the verb types in each of the sentence regions. In the mean reading times,
each item from the two studies is represented equally.
A repeated measures ANOVA on all items from both studies was conducted for
each sentence region. The independent variables were telicity (telic vs. atelic) and
transitivity (obligatory vs. optional). The dependent variable was the reduced relative
effect (RRE), measured as mean reading time for the ambiguous sentence minus
the corresponding unambiguous sentence in each region. The results show that on
the verb-ed by region, the RRE was significantly smaller when the embedded verb
was telic rather than atelic, F = ., p < .. There was a trend towards a smaller
RRE when the embedded verb was optionally rather than obligatorily transitive,

OBryan, Folli, Harley, and Bever

F = ., p = .. The interaction between telicity and transitivity was not significant on this region, F = ., p > .. There were no significant effects of telicity,
transitivity, or the interaction between the two on the noun phrase or main verb
regions. On the main verb region, there was a trend towards a smaller RRE when
the embedded verb was telic rather than atelic, F = ., p = .. The numerical difference between the transitivity conditions on the main verb region was in
the same direction as the transitivity differences reported by MacDonald ().
All statistical analyses were conducted by item. By-participant analyses were not
possible because telicity and transitivity were not planned factors in the original
In summary, the post hoc reanalysis showed that verb telicity had a statistically
significant effect on the amount of processing difficulty that arose in the earliest
possible disambiguation region, the verb-ed by region. The reduced relative effect
was significantly smaller when the embedded verb was telic than when it was
The results provided preliminary evidence for the relevance of telicity to online
sentence comprehension. This finding is consistent with our hypothesis that the verb
contributes information about its event structure, which is used without delay in
resolving structural ambiguity. These initial results provided the motivation for additional balanced, controlled experiments.

.. Auditory sentence comprehension results

OBryan, Folli, Harley, and Bever (in preparation) found additional evidence supporting the Event Structure Processing hypothesis in an auditory sentence comprehension
experiment. The experiment used sets of reduced relative clause sentences such as
those shown in ().
() a.

Atelic, optionally transitive: The actress sketched by the writer left in a hurry.
Atelic, obligatorily transitive: The actress escorted by the writer left in a hurry.
Telic, optionally transitive: The actress tripped by the writer left in a hurry.
Telic, obligatorily transitive: The actress noticed by the writer left in a hurry.

The corresponding control sentences were structurally unambiguous due to insertion

of that was, as shown in ().
() The actress that was sketched by the writer left in a hurry.
Each set of experimental sentences differed only in the telicity and transitivity of
the embedded verb. The verb type included in each condition was kept as homogeneous as possible. Verbs in the telic optionally transitive condition were necessarily
unaccusative and could undergo the causative alternation. Verbs in the obligatorily
transitive telic condition, e.g. notice, could not undergo the causative alternation; this

Verb Telicity in Sentence Comprehension

fact is critical for separating telicity from the causative alternation. Verbs in the atelic
optionally transitive condition, e.g. sketched, were object drop verbs. Stative predicates
were avoided in all conditions.
The auditory sentence comprehension experiment tested the Event Structure Processing hypothesis using the speaker change monitoring task. Forty-four native
English-speaking participants heard the spoken sentences and indicated whether they
heard a word or part of a word spoken by a different speaker than the rest of the sentence. The rationale for the task is that listeners will have more difficulty detecting the
speaker change when sentence factors impose a greater processing load (Townsend
and Bever ). In the experimental sentences, the speaker change always occurred
on the first syllable of the noun in the by-phrase, e.g. writer in (). The dependent
variable was the reduced relative effect, measured as the percentage of speaker change
detection errors for the ambiguous sentence minus for the corresponding unambiguous sentence.
The results, shown in Table ., followed the same pattern found in the self-paced
reading data: significantly less processing difficulty occurred when the embedded verb
was telic than when it was atelic, F = ., p < .; F = ., p < ..
Effect size was measured using Cohens f as . for the by-participants analysis and
. for the by-items analysis, indicating a medium effect size (Cohen ). There was
not a significant main effect of transitivity or a significant interaction between telicity
Table . Mean speaker change detection errors for the four verb types





The N
by the N
mainV. . .




The N
by the N
mainV. . .



The N
by the N
mainV. . .








The N
by the N
mainV. . .



Detection Errors

OBryan, Folli, Harley, and Bever

and transitivity, all ps > .. Since the speaker change syllable occurred on the noun
in the by-phrase, this result is consistent with the self-paced reading results.
The auditory sentence comprehension results provided further support for the
Event Structure Processing hypothesis. However, a limitation of the study was that
it only allowed examination of effects on the noun in the by-phrase. Thus, an experimental task that would be sensitive to processing effects throughout the sentence was

.. Eye-tracking during reading: results

OBryan et al. (in preparation) further tested the Event Structure Processing hypothesis in an experiment that measured readers eye fixations during sentence reading. The
materials were sets of sentences like those used in the speaker change monitoring
experiment, shown in (). Forty monolingual English-speaking participants read
the sentences silently while their eye fixations were measured. They answered sixteen
yes/no questions during the experiment to ensure comprehension of the sentences.
The hatchmarks in () show the boundaries of the zones that were defined for
analysis of the eye fixations.

a. The actress
tripped by the writer left in a hurry.
b. The actress that was tripped by the writer left in a hurry.

Note that the critical regions for examining processing effects are zones , which are
exactly the same in the ambiguous experimental sentence (a) as in the unambiguous
control sentence (b).
The dependent variable, as in the self-paced reading and speaker change monitoring studies, was the reduced relative effect, used as an indicator of processing difficulty
resulting from the syntactic ambiguity. Processing difficulty was measured as the eye
fixation time in each zone, measured as go past time per word. Go past time is
defined as the sum of all fixations in a region until the eye exits to the right only
(Barker and Bolger ). Trials with no fixations in a region were discarded from
the analysis.
The results are shown in Fig. .. The x-axis in Fig. . shows each sentence zone,
and the y-axis shows the amount of processing difficulty, measured as the reduced
relative effect. The results indicate that the differences between the experimental
conditions were greatest on the main verb, such as left in (). In this zone,
significantly less processing difficulty was found when the embedded verb was
telic than when it was atelic, F = ., p < .; F = ., p = .; Cohens
f = .. Significantly less processing difficulty was found in the obligatorily transitive
conditions compared to the optionally transitive conditions, F = ., p < .,
F = ., p = .; Cohens f = .. There was not a significant interaction between

Verb Telicity in Sentence Comprehension


RRE in ms.






by the writer
Atelic Oblig-Trans
Atelic Option-Trans


in a hurry

Telic Oblig-Trans
Telic Option-Trans

Fig. . Processing difficulty measured as eye fixation times per word in each sentence zone

telicity and transitivity, F = ., p > ., F < . A planned contrast analysis

revealed a significantly smaller reduced relative effect in the obligatorily transitive telic
condition compared to the obligatorily transitive atelic condition, F = ., p < ..
Neither transitivity, telicity, nor the interaction between the two significantly
affected the size of the reduced relative effect in region (the by-phrase) or region
(the adverbial following the main verb), all ps > ..
The eye-tracking experiment revealed the same pattern of results found in the
speaker change monitoring experiment and the self-paced reading reanalysis: less
processing difficulty occurred in the disambiguating region of reduced relative clause
sentences when the embedded verb was telic compared to atelic. The results also replicated the pattern of results that MacDonald () found for the effect of obligatory
transitivity. The planned contrast analysis yielded a result that is important for supporting the Event Structure Processing hypothesis: telicity significantly lessens the
reduced relative effect, even for obligatory transitive verbs, which are not unaccusative
and do not undergo the causative alternation. Thus, all of the results have supported
the hypothesis that event structure, specifically verb telicity, is used during online
sentence comprehension.
Here we return to the question of whether telicity was a confound in MacDonalds
() experiment. The self-paced reading and eye-tracking experiments replicated
her finding that obligatory transitivity affects online comprehension. However, we
also found that telicity had a significant main effect. These effects showed up in the
same zone (main verb) in the eye fixation data, but there was no interaction between
telicity and transitivity. These results increased our motivation to be more precise
about where the telicity and transitivity effects arise in the disambiguating region.

OBryan, Folli, Harley, and Bever

Such precision would require a measure that allows examination of the effects as each
word is integrated into the sentence. The word maze (Freedman and Forster ,
Nicol, Forster, and Veres ) is the appropriate task for this purpose.

. The Word Maze Experiment

In the word maze task, the participant sees a word on the screen followed by a
sequence of word pairs. For each pair of words, the participant is asked to choose the
word that can follow the previous word(s) in a sentence as shown in Fig. .. The word
maze task forces incremental parsing. Thus, the task is well suited for our purpose: it
provides a measure of the difficulty of integrating each new word into the sentence as
it is processed.
The purpose of the word maze experiment is to investigate how telicity and transitivity each affect incremental processing of the structurally ambiguous sentence. Note
that the task is not the same as normal reading or listening comprehension. Rather,
the task provides insight into the kinds of information that are available from each
word in the linguistic input and how this information can affect the interpretation of
the sentence as it is built.
.. Methods
The word maze experiment used the experimental sentences from OBryan et al.s (in
preparation) speaker change monitoring and eye-tracking studies, with a few changes
in the particular verbs used. There were sets of experimental sentences. The verbs
were controlled for lexical frequency using the KuceraFrancis database (Kucera and

The Word Maze Paradigm

Computer Screen











clock silently


Fig. . The word maze task (Freedman and Forster )

Verb Telicity in Sentence Comprehension

Francis ). An ANOVA on verb frequency counts showed no significant difference

in the lexical frequencies between the verbs in the four different verb-type categories.
The participants in the study were monolingual English speakers.
Construction of the maze task stimuli required the use of non-continuation words,
which served as the incorrect choices at each point in the sentences. For example, were,
them, and clock are the non-continuation words in Fig. .. The non-continuation
words were incorrect choices because they resulted in an ungrammatical sentence
fragment, not just an implausible sentence fragment.
The design of the experiment was the same as in the speaker change monitoring and
eye-tracking experiments. The participants were not asked to answer comprehension
questions or to paraphrase the sentences, because successfully navigating through
the word maze requires the participant to comprehend the sentence. We collected
reaction times and errors for each choice throughout the sentence. The dependent
variable was the reduced relative effect in milliseconds.
The predictions of the word maze experiment were the same as in the prior experiments, except that we expected that the time course of the effects would be different
due to the nature of the task. As before, we predicted that less processing difficulty
would occur in reduced relative sentences when the initial verb was telic compared to
atelic. We also predicted that less processing difficulty would occur when the initial
verb was obligatorily transitive compared to optionally transitive.

.. Results
The results of the word maze are shown in Fig. .. The y-axis indicates the size of
the reduced relative effect in milliseconds. The x-axis depicts the words in the target
Fig. . shows the same data presented differently. The x-axis still reflects the words
in the sentence, but the y-axis shows how much the results went in the predicted
direction. Dark bars show the difference between the atelic and telic conditions. Light
bars show the difference between the optionally transitive and obligatorily transitive
The analysis of variance results are presented in Table ..
The effects of telicity and transitivity on the reduced relative effect throughout the
critical regions (by through the next word region, which is one word following the
main verb) were negatively correlated at ..

.. Discussion of word maze results

The results show that telicity had a significant effect on the amount of processing difficulty that occurred on the early disambiguation region (the word by). As in the three
prior experiments, the reduced relative effect (RRE) was smaller when the embedded

OBryan, Folli, Harley, and Bever


RRE in ms.




Atelic Oblig-Trans
Atelic Option-Trans



Telic Oblig-Trans
Telic Option-Trans

Fig. . Word maze results

Difference in RRE (ms.)




atelic - telic





option.trans - oblig.trans

Fig. . Effects of telicity and transitivity in the word maze

verb was telic compared to when it was atelic. Thus the word maze results further
support the Event Structure Processing hypothesis.
The word maze results also show that transitivity had a significant effect on the
RRE. On the noun in the by-phrase and on the main verb (but not on the earlier by
region), the transitivity effect was significant in the predicted direction. That is, the
RRE was significantly smaller when the embedded verb was obligatorily transitive
compared to when it was optionally transitive. As noted above, MacDonald ()
found the same pattern of results in a self-paced reading experiment. One difference
between MacDonalds study and the current studies is that her reduced relative clause
sentences used prepositional phrases such as in the coup in () instead of by-phrases.

Verb Telicity in Sentence Comprehension

Table . Word maze ANOVA results

Sentence regions




main verb











() The ruthless dictator fought in the coup was hated throughout the country.
In the word maze on the early disambiguation (by) region, the transitivity effect was
significant in the opposite direction from the effect on the main verb. MacDonalds
() Experiment A also found the same pattern: the obligatorily transitive condition resulted in a larger RRE than the optionally transitive condition on the prepositional phrase preceding the main verb. MacDonald reports but does not explain
this result, because it is not the focus of her paper. The difficulty that subjects have
in the obligatorily transitive condition in this early region is understandable, because
an obligatorily transitive verb requires a direct object which must immediately follow
the verb in any active sentence. For example, consider the obligatory transitive verb
capture. In the sentence The ruthless dictator captured in the coup . . . , when the preposition in is encountered, it becomes clear that the sentence cannot be a grammatical
simple main clause. In other words, in the reduced relative, the preposition following
the verb signals the processor that the required direct object is absent from the surface
direct object position. Thus, anything other than a noun phrase in this position, such
as by, a preposition, or an adverbial, should result in processing difficulty.
Less processing difficulty occurs in the optionally transitive condition prior to the
main verb, because a simple main-clause interpretation can be maintained. Consider
the optionally transitive condition sentence in () until the main verb disambiguation point (indicated by the pipe |): The ruthless dictator fought in the coup | was
hated throughout the country. This sentence beginning could have a simple main clause
structure with the ruthless dictator as the subject, since fought requires no direct object.
When the main verb was occurs following coup, processing difficulty will ensue, especially because there was not a cue that a reduced relative structure should be posited
earlier in the sentence. Thus, the optionally transitive condition has a small RRE prior
to the main verb and a very large RRE on the main verb. This pattern of results was
found in the word maze experiment as well as in MacDonalds Experiment A.
The word maze results did show an interaction between telicity and transitivity on
by. In order to understand this, it is important to note that the two factors, each with

OBryan, Folli, Harley, and Bever

two levels, make four verb conditions. The interaction indicates that one of these four
verb conditions has the smallest RRE of all: the optionally transitive telic condition.
However, even with the interaction taken into account, telicity and transitivity have
significant main effects on the RREs in the word maze.
The condition with the smallest RREs on by, the optionally transitive telic condition, is the condition in which the verbs are unaccusative in their intransitive usage.
This finding is consistent with Stevenson and Merlos () claim that reduced relatives with unaccusative embedded verbs are easier to understand than those with
unergative embedded verbs. Unaccusatives that can occur in reduced relative clause
sentences also undergo the causative alternation as shown in ().
() a. The writer tripped the actress.
b. The actress tripped.
This finding raises the question of whether the relevant factor in online comprehension is telicity, unaccusativity, or the causative alternation. If unaccusativity or the
causative alternation were the relevant factor, the difference between the telic and
atelic condition should be found only for optionally transitive verbs. The speaker
change monitoring and eye-tracking results revealed a significant difference between
the telic and atelic conditions for obligatorily transitive verbs, which are neither unaccusativen or unergative and do not undergo the causative alternation. Thus, our results
support the claim that telicity is the relevant factor rather than unaccusativity or the
causative alternation.
In order to understand the word maze results, it is important to notice a
general pattern that occurs in all four verb conditions. When a lot of processing
difficulty occurs on the early region, less processing difficulty occurs on the main verb.
Conversely, when very little processing difficulty occurs on the early by region, a lot of
processing difficulty occurs on the main verb. Processing difficulty on the early region
suggests that the processor may have disambiguated the sentence as a reduced relative.
If this is the case, there will be less processing difficulty on the main verb, presumably
because the processor has already determined that the structure is a reduced relative.
Little or no processing difficulty on the early by region indicates that disambiguation
has not occurred and a simple main-clause analysis is still being considered. When
this is the case, there will be a great deal of processing difficulty at the disambiguation
point, on the noun in the by-phrase or on the main verb region. Because of this
pattern in the word maze data, the interaction between telicity and transitivity on
the main verb is probably due to the interaction on by. The optionally transitive telic
condition has the smallest RRE on by and the largest RRE on the main verb.
The pattern just described explains why a negative correlation in the early disambiguation region would lead to the continuation of the negative correlation later in the
sentence on the main verb. However, it does not explain why a negative correlation
exists in the early disambiguation region. It would have been possible for telicity and
transitivity to pattern together, yielding a positive correlation, but they do not.

Verb Telicity in Sentence Comprehension

The negative correlation between the telicity and transitivity effects throughout the
word maze sentence regions depends on the subtraction of the telicity and transitivity
conditions. The telicity effect is computed as the atelic condition minus the telic
condition, and the transitivity effect is computed as the optionally transitive condition
minus the obligatorily transitive condition. The subtractions are done in this way
because of a similarity between the telic verb condition and the obligatorily transitive
verb condition: both require an object. The difference between them is that telicity
involves requiring an underlying direct object (internal argument and theme), while
obligatory transitivity involves requiring a surface direct object. Thus, the negative
correlation in the early disambiguation region can be understood as follows: in the
reduced relative clause sentence, integration of the preposition by is facilitated by the
verbs requiring an underlying direct object (i.e. being telic), and it is inhibited by
the verbs requiring a surface direct object (i.e. being obligatorily transitive).

. General Discussion and Conclusions

Our experimental findings provide support for the Event Structure Processing
hypothesis: that the processor accesses information about verb telicity and uses it
during online comprehension. The hypothesis was supported by evidence from four
different experimental tasks: self-paced reading, speaker change monitoring, eyetracking during reading, and the word maze. All the experiments found the same
pattern of results: less processing difficulty occurred in the disambiguating region
of reduced relative clause sentences when the embedded verb was inherently telic
compared to when it was atelic. The results also indicate that this effect of telicity is
distinctly different from the effects of obligatory transitivity.

.. The time course of the effects

The effect of telicity showed up in different regions of the sentences in the different
experimental tasks. In the speaker change monitoring experiment, the effect is linked
to the noun in the by-phrase because that is where the speaker change was located in
the stimuli. Consequently, this task does not allow inferences to be made concerning
the time course of the effect of telicity. All we can be sure of is that telicity has a
facilitating effect on reduced relative comprehension at that point in the auditorily
presented sentence.
In the eye-tracking experiment, our results showed an effect of telicity appearing
on the main verb only. No significant effects of telicity or transitivity were found
on the by-phrase. The effect on the main verb during natural reading in the eyetracking experiment is consistent with the notion that the main verb is the most
salient disambiguator. In a reduced relative, the processing load increases sharply at
the disambiguation point, and consequently reading times are longest at this point.

OBryan, Folli, Harley, and Bever

The facilitating effect of telicity thus shows up as reduced eye fixation times at the
disambiguating main verb.
In the self-paced reading reanalysis, the facilitating effect of telicity appeared much
earlier in the sentence, on the verb-ed by region, suggesting that a cue to the complex
structure is available to the processor at that point in the sentence, presumably the
word by itself. There are a few reasons why this cue might be more likely to have
an immediate effect on the processor in self-paced reading than during eye-tracking.
In self-paced reading, the sentence is artificially chunked, and the reader presses a
button to progress through the sentence. Once the button has been pressed for the
next chunk, the previous chunk disappears and is no longer available to the reader.
The inability to look back encourages the reader to make maximum use of available
cues in each chunk as it occurs. In contrast, during eye-tracking, the sentence remains
available on the screen, as during normal reading, and backtracking is possible. Consequently less salient cues to complexity may be safely left for later consideration.
Finally, the need for button pressing during self-paced reading gives the processor
additional time to work between the appearance of chunks of text, while the button
press is occurring, which amplifies the effect of early cues to complexity.
.. Telicity and transitivity are independent
In section .. we described MacDonalds () findings, which showed a significant reduction of the reduced relative effect for obligatorily transitive main verbs. This
result was further supported by Townsend and Bever (: ), who attribute
transitive verbs facilitating effect on reduced relative comprehension to their objectseeking character. If telic verbs and transitive verbs are both object-seeking, then one
might expect that they should show the same facilitating effect on the resolution of
reduced relative clauses.
Some approaches to event structure in the syntactic literature would lead to this
expectation as well. Borer (), for example, proposes that argument structure is
derived from event structure information. Under this view, no information regarding
the number of arguments is represented in the lexical entry of the verb. If argument
structure is derived from event structure, or vice versa, transitivity and telicity should
be indistinguishable in their effects on processing, since telicity does not necessarily
imply transitivity.
Our results, however, show that this is not the case. The word maze results show that
telicity and transitivity had distinctly different effects on processing throughout the
disambiguating region. Both the speaker change monitoring and eye-tracking experiments found a significant difference between telic and atelic obligatorily transitive
verbs. Each of these results show the effect of telicity to be independent of transitivity.
In the resolution of a reduced relative clause sentence, the processor must posit the
presence of a theme argument or underlying object, which is subsequently displaced
by movement. We propose that verb telicity gives the processor a head start in this

Verb Telicity in Sentence Comprehension

process of resolution, as it forces the processor to introduce an underlying object

position so that the specifier of TelicP may be filled, regardless of the verbs transitivity.
Verb telicity can be thought of as a lexical semantic feature on the verb which requires
the presence of a particular syntactic structure (e.g. Folli , Folli and Ramchand
). The event structure properties of a telic verb, in other words, are a cue to the
processor to project a syntactic frame containing an underlying object and a TelicP.

.. Incorporating telicity into current processing models

We now turn to the kind of comprehension model supported by our findings: a model
in which telicity information facilitates the resolution of the syntactic ambiguity in
reduced relative clauses. This requires consideration of models which can incorporate information like event structure into ongoing processing. In the case of reduced
relatives, three kinds of factor may govern the strength of the reduced relative effect:
() factors that affect the commitment to the main-clause reading; () factors that
affect the behavioural or computational difficulty of noticing the ambiguity; () factors
that affect the difficulty of recovering from the misanalysis (for extensive discussion
of the possibilities for integrating telicity into several different kinds of processing
models, see OBryan ). Processing models differ with regard to these factors.
Consider first models which explain the effects as a function of the likelihood of
the main-clause analysis. Some of these models already incorporate transitivity as a
factor in online processing of reduced relatives. In particular, constraint-based models
like that of McRae et al. () have explicitly incorporated the facilitating effect
of transitivity. It is straightforward to include telicity as an additional constraint in
such a model. Telicity and transitivity would then be independent factors that would
facilitate the resolution of ambiguity. Such a model predicts that when both factors
are present, the greatest facilitating effect should be found, and when both are absent,
the greatest reduced relative effect should be found. This is precisely what we see on
the main verb in the eye-tracking study, where the reduced relative effect is largest
in the atelic optionally transitive condition, and smallest in the telic obligatorily
transitive condition (see Fig. . above).
In such models, the reduced relative effect in a reduced relative clause is attributed
to the difference in activation strength of two competing structural analyses of the
ambiguous string: a strong main-clause analysis versus a weak reduced relative analysis (Seidenberg and MacDonald ). As new words are integrated into the string,
the activations of the two analyses change. Telicity information is available as soon
as the embedded verb is encountered, as it is represented in the verbs lexical entry.
We might therefore expect this information to have a facilitating effect at the earliest
possible disambiguation point, namely at the word by. Although our results do show
an early effect of telicity in the self-paced reading task, no such early effect occurs in
the eye-tracking study. Furthermore, such models always raise two related questions:

OBryan, Folli, Harley, and Bever

(a) what is the basis for the acquisition of such constraints? In this case, how does
the child learn to incorporate the statistical effect of telicity in the embedded verb?
(b) what is the computational mechanism that causally links telicity to a reduction
in the reduced relative effect in particular? Constraint-based models can capture the
facts but require further explanation of how the constraints are learned and why the
constraints have the effects that they do.
A relevant hypothesis concerning reduced relative clause processing that we have
not yet discussed is that of McKoon and Ratcliff (). They claim that the meaning
of the reduced relative construction is incompatible with some verbs. They assert that
the reduced relative structure indicates participation in an event caused by some
force or entity external to itself . They further claim that events that are internally
caused by the head noun phrase are incompatible with the meaning of the reduced
relative construction. They consider internally caused events to be compatible with
the corresponding unreduced relative clauses because they have a distinct, bi-eventive
meaning. They claim that The horse raced past the barn fell is ungrammatical because
whenever someone races a horse, the horse must internally cause the racing event.
Their account also predicts that The horse raced by the new jockey fell is ungrammatical.
This false prediction leads us to doubt that their account could be accurate. In any
event, McKoon and Ratcliff s account does not explain our findings. Consider the
reduced relative clause sentences from the speaker change monitoring, eye-tracking,
and word maze experiments, repeated in () for convenience.
() a.

Atelic, optionally transitive: The actress sketched by the writer left in a hurry.
Atelic, obligatorily transitive: The actress escorted by the writer left in a hurry.
Telic, optionally transitive: The actress tripped by the writer left in a hurry.
Telic, obligatorily transitive: The actress noticed by the writer left in a hurry.

Only one of our conditions could be considered to include an event internally caused
by the head noun phrase: the optionally transitive telic condition, (c). McKoon and
Ratcliff s account would then predict that (c) is ungrammatical and thus should
show the greatest reduced relative effect. This prediction was not borne out. In fact, the
eye-tracking and word maze experiments showed that (c) had the smallest reduced
relative effect. If none of our verb conditions are considered to involve internal causation, McKoon and Ratcliff s account does not explain why significantly less processing
difficulty occurred in the telic conditions.
Let us now consider how a structurally based model could approach the facilitating
effect of telicity in terms of the ease of recovering from the main-clause analysis. In
such models, the reduced relative effect is attributed to the processing cost associated with revising the structural analysis. In this case, the independent facilitating
effect of telicity could only be accounted for if its structural representation is distinct
from the structural representation of simple transitivity. One example of a processing
model which proposes an important role for lexically projected syntactic structure is

Verb Telicity in Sentence Comprehension

Stevenson and Merlos () Competitive Attachment model. In such a model, distinct and competing structural analyses are ordered in terms of their computational
complexity. The effect of telicity could be accounted for by assuming the lexical projection of distinct object positions in telic and atelic obligatorily transitive verbs, and
consequent differences in the ease of reassigning the correct structure to the initially
incorrect analysis.
If the facilitating effect of telicity is indeed structural, it must be the case that
a smaller amount of reanalysis is necessary to switch to a reduced relative clause
structure from an erroneous main-clause analysis of a telic verb than from a mainclause analysis of an atelic verb. In other words, there must be some respect in which
the structural analysis of a clause containing a telic main verb shares unique structural
similarities with the analysis of a reduced relative clause.
In recent syntactic theory, the Spec,TelicP position has structural and semantic
properties which are closely related to the properties of resultative participles: the
preterite participles used in passives and perfects in English. The subject of the
reduced relative is ascribed the property of having undergone the action described
by the verb; this is the characteristic meaning of such participles (e.g. Embick ,
Kratzer ). The passive participle in a reduced relative is thus a subtype of resultative participle. A tree representing this general idea is presented in ().



The actress




Similarly, with telic verbs, the underlying object in Spec,TelicP is an inner subject
(Hale and Keyser ), entering into a predication relation with the verbal head lower
in the tree. Telicity involves a transition into an end state, and the verb provides
Harley () argues that in fact the predication relation (a small clause) is the key factor in producing
a telic or resultative interpretation in these sentences, and that dedicated functional projections such as
TelicP and ResultP are not required or motivated. On this approach, the structural characteristic held in
common by telic predicates and the resultative participles would be the inner subject structural relation
associated with the predication represented in each.

OBryan, Folli, Harley, and Bever

the specification of what that end state is. Hence, telic verbs structurally represent
a resultant state in very much the same way that a resultative participle does.


the writer


the actress




Notice that the TelicP portion of the tree in () is structurally identical with the
ResultP portion of the tree in (), and indeed its semantic interpretation is very
With atelic transitive verbs, on the other hand, the object does not move into
Spec,TelicP and this special result-state predication relation is not established. This
is illustrated in ().


the writer




the actress

As a result, reanalysis to a resultative participle requires a greater structural adjustment when the predicate is atelic, involving the addition of a ResultP to the structure.
If we now introduce this structural distinction into Stevenson and Merlos model,
we can consider the different reanalysis steps that would be involved in switching
from a main-clause analysis to a reduced relative analysis at a given disambiguation
point. Consider first the main-clause analysis of a telic verb at the point The actress
tripped . . . The presence of a telic feature on the verb trip will motivate the processor

Verb Telicity in Sentence Comprehension

to introduce a TelicP into the derivation, with a place holder waiting for the object to
occupy its specifier. The structure at this point is represented in ().


The actress








When the processor needs to reanalyse this structure as a reduced relative tree like
that in (), due to encountering a disambiguation cue that rules out the main-clause
analysis, the only structural adjustment required is the replacement of the placeholder
DP (indicated with ??? in ()) with the erstwhile subject DP the actress, accompanied
by a featural adjustment to v from Active to Passive.
In contrast, when the processor has developed a main-clause analysis of an atelic
transitive verb, as in The actress chaperoned . . . , the structure at that point is as represented in (), with an object placeholder occupying the sister-to-V position.


the actress






At the disambiguation point, the change to the reduced relative structure in ()

will require the projection of an additional functional projection, ResultP, and
corresponding movement from the sister-of-V to Spec,ResultP, along with reanalysis

In an optionally transitive verb, however, the object placeholder will not be posited by the processor,
requiring an even greater amount of restructuring to recover from the main-clause analysis. Consequently,
the facilitating effect of transitivity is still accounted for.

OBryan, Folli, Harley, and Bever

of the position of the actress and the features of the v . This is clearly a more taxing
In other words, current syntactic theory provides structures for telic and atelic
transitive predicates and resultative participles which provide the basis for a structural reanalysis that predicts the facilitating role of telicity in online processing of
ambiguous obligatorily transitive sentences. Without such a structural distinction,
however, such syntactically driven models will erroneously predict that the two types
of transitive verb should behave identically at reanalysis.
Our data do not unambiguously choose between these explanations. Furthermore,
the wide range of studies of reduced relative constructions suggest that both structural
and probabilistic processes may be involved. Superficial features of the reduced relative verb sequence that correlate with a simple declarative sentence have been found
to increase the strength of the reduced relative effect (e.g. McRae et al. ). The
structural differences between the erroneous main-clause analysis and the correct
analysis may also affect the overall difficulty (e.g. Stevenson , Frazier and Clifton
). This invites consideration of models that explicitly integrate the operation of
immediately available pattern-completion constraints with the assignment of syntactic derivations.
One such model is the analysis-by-synthesis framework (Townsend and Bever
). In such a model, comprehension involves the formation of meaning representations based on both probabilistic information and structural derivation. The model
predicts that the initial comprehension of sentences is based on pattern recognition,
reflecting statistically and informationally dominant patterns, without the benefit of
a complete or correct derivational syntax. In this model, semantic event structure
information carried by individual verbs contributes directly to the initial meaning
representation, as suggested by our results. The model then involves an almost parallel assignment of derivational structure, followed by a comparison of the probabilistic and structural analyses. The contribution of such a model is that it provides
an architecture for integrating pattern-based analysis, derivation-based analysis, and
reconciliation of the two kinds of information.
At a more general level, our findings are relevant to the longstanding debate about
the interaction of syntax and grammatically encoded semantics, and its relevance for
sentence comprehension. Telicity, and more generally event structure, is an intrinsic
lexical and categorical semantic property of verbs, but as we have seen, it has strong
syntactic reflexes and lends itself to a structural characterization. We have shown that
online sentence comprehension is sensitive to this factor. Because of its position at
the cusp of the syntaxsemantics interface, event structure will prove to be a fruitful
testing ground for experimental investigation of the role that syntactic computations
and semantic categories play in online comprehension.

Part II
Syntax and Information Structure

Focus Intervention in Declaratives

. Introduction
Although much current work in generative linguistics recognizes the significance of
information structure (IS), questions remain regarding its precise representation and
role in the grammar. This chapter goes some way toward answering these questions,
by showing that IS plays a crucial part in a specific phenomenon, namely, focus
intervention effects in declarative sentences.
The phenomenon of focus intervention is illustrated in the wh-questions in () and
(), where a certain type of operator preceding the wh-phrase yields an unacceptable
sentence. In () the operator is the Korean particle man only, while in () it is
the negative polarity item (NPI) daremo anyone in Japanese. Man and daremo are
therefore categorized as intervention triggers.


nuku-ll po-ass-ni?
Minsu-only who-acc see-past-q
Who did only Minsu see?

() ? Daremo nani-o
anyone what-acc read-neg-past-q
What did no one read?

(Beck : )

(Tomioka : )

In questions involving a disjunction, intervention effects are manifested in the absence

of an alternative question reading, rather than degradedness (Beck and Kim ).
This chapter is based on my dissertation (Eilam ) and updates a short paper from the proceedings
of NELS (Eilam, to appear). I first and foremost thank Tony Kroch for calling my attention to the novel
data reported here. I am grateful to Dave Embick and Satoshi Tomioka for extended discussions, as well
as Catherine Lai, Laia Mayol, Satoshi Nambu, Florian Schwarz, Yanyan Sui, the editors of this volume,
and audiences at Bar Ilan University, the Workshop on Prosody, Syntax, and Information (WPSI) at the
University of Delaware and OnLI II at the University of Ulster.
Here I discuss only elements which are crosslinguistically stable intervention triggers, i.e. the operators
corresponding to English only, even, and also, and NPIs (see Kim , Beck ). I label these only-type
operators; the phrase consisting of the operator and its associate (Minsu-man only Minsu in ()) is the


Thus, in (), the intervention trigger only preceding a disjunctive phrase in English
leaves the question with just a yes/no interpretation, allowing it to be answered as in
(b), but not (a).
() Does only John like Mary or Susan?
a. Mary.
b. Yes.

[ AltQ]

Until recently, the general consensus was that these crosslinguistically robust patterns indicate syntactic ill-formedness, although the precise description of this illformedness remained a matter of debate (see e.g. Pesetsky , Kim , Grohmann
). Beck (), however, notes that all of the syntactic analyses fail to accurately
define the set of intervention triggers, or make use of a definition which is ultimately
semantic. Accordingly, she pursues a semantic approach to intervention. The basic
idea is that focus-sensitive operators interfere with the semantic relation between the
question operator high in the clause and the wh-phrase or disjunction, because they
take the place of the question operator in evaluating the alternatives introduced by
the latter. Although this idea overcomes the above-mentioned weakness of syntactic
analyses, Tomioka () points out various observations which both Becks analysis
and the syntactic alternatives are unable to explain. These include a great deal of interspeaker variability in judgements, and the amelioration of intervention effects when
the intervener is an embedded subject or not a subject.
Eilam () adds further evidence against syntactic and semantic analyses, showing that they derive incorrect predictions. First, intervention effects are not read off
the hierarchical structure, as demonstrated by the fact that they may arise even when
an intervener follows a wh-phrase or disjunction, i.e. when the allegedly necessary
c-command relations do not hold. Second, intervention effects can be ameliorated or
eliminated in certain contexts, which render the potential intervener a non-focus in
IS terms, but do not change its semantic status. If this were a syntactic or semantic
phenomenon, it should be immune to contextual changes. Moreover, this finding
indicates that what underlies intervention effects is not the semantics of only-type
operators, as Beck () claims, but rather the IS notion of focus.
Building on the above-mentioned observations, Eilam () develops an IS analysis of focus intervention first proposed in Tomioka (). Wh-questions and alternative questions are associated with a rigid informational articulation, in which the
wh-phrase and disjunctive phrase, respectively, constitute the IS focus. Since just
one such focus is allowed per sentence, any phrase including an only-type operator
can be accommodated only if it is a non-focus in the informational articulation of
the sentence. This accommodation may be achieved via the syntax, prosody, and/or
context; various strategies and relevant examples of accommodation are provided

Focus Intervention in Declaratives

The goal of this chapter is to corroborate the IS approach to intervention, by applying it to a set of data which has yet to be noted in the literature. That is, intervention
effects are not limited to questions, but also arise when an only-type operator and IS
focus cooccur in a declarative sentence. This is illustrated in the answer to a question
in (b, c) and the corrective context in (), where italics mark associates of only and
small caps indicate the pitch accent on the IS focus.
() a. What did only John drink?
b. Only John drank only BEER.
c. ??Only John drank beer.
() a. Its not true that only John drank wine, only John drank only BEER.
b. ??Its not true that only John drank wine, only John drank beer.
As in the case of questions, the interveneronly John in () and ()can be accommodated through strategies which alter the informational articulation of the sentence.
In English these include structures which place the IS focus before the erstwhile intervener, such as passivization () and specificational copular constructions (). When
the IS focus is the subject, the accommodation comes for free: the configuration in
() and () is identical to that of () and (). The properties of this configuration will
be discussed in section ..
() a. What did only John drink?
b. Only BEER was drunk by only John.
() Its not true that only John drank wine, beer was the only thing that only John
() a. Who drank only beer?
b. Only JOHN drank only beer.
() Its not true that Mary drank only beer, only JOHN drank only beer.
The structure of the chapter is as follows: section . reviews the existing literature
on intervention in questions, focusing on Becks () semantic theory and the
IS theory developed in Eilam (). I show that the patterns found are consistent
with the latter, rather than the former. In section . I analyse the novel data from
declaratives, which is crosslinguistically robust and bears not only on speaker judgements but also on truth-conditions. I demonstrate that the IS theory developed for
questions naturally extends to declaratives, while syntactic and semantic approaches
to intervention are inadequate. Section . concludes by assessing the implications
of the findings for theories of focus association and for the conception of IS in the


. Intervention Effects in Questions

As noted in section ., the discussion of intervention effects in the literature has centred on questions until now. Beginning with Hoji (), researchers have grappled
with the proper analysis of this phenomenon, proposing syntactic, semantic, and IS
explanations. A comprehensive review of this literature can be found in Eilam (,
); here I only highlight the primary types of approaches taken, in order to assess
where we stand before moving on to the novel data from declaratives.
Recall that an only-type operator preceding a wh-phrase in a question yields
degradedness ((a), (a)), and one preceding a disjunctive phrase eliminates the
alternative reading of the question (). The examples below also illustrate possible
repair strategies, used to convey the meaning of the unacceptable intervention configurations: scrambling of the wh-phrase over the intervener in (b) and (b), and
clefting of the alternative question in ().
() a. Minsu-man nuku-ll po-ass-ni?
Minsu-only who-acc see-past-q
b. Nuku-ll Minsu-man po-ass-ni?
who-acc Minsu-only see-past-q
Who did only Minsu see?
() a. ? Daremo nani-o
anyone what-acc read-neg-past-q
b. Nani-o daremo yom-ana-katta-no?
what-acc anyone read-neg-past-q
What did no one read?
() Does only John like Mary or Susan?
a. Mary.
b. Yes.
() Is it Mary or Susan that only John likes?

[ AltQ]

Focus intervention effects of this type were initially treated as syntactic in nature.
Beck (a), for example, argues that in situ wh-phrases are prohibited from covertly
moving across interveners, while for Kim (), the obstructed relation between C
and the wh-phrase does not involve movement but rather Agree. All of the syntactic
analyses, however, suffer from a range of problems. First, they generally capture no
more than a subset of the data; Pesetsky (), for example, is specifically geared
for wh-questions. Second, syntactic analyses often assume that in situ wh-phrases
undergo some type of movement; this assumption, however, is not adopted by much
of the current work in syntactic theory (Cole and Hermon , Reinhart ), and it

Focus Intervention in Declaratives

is not clear how to extend a movement-based account to questions with a disjunction

(Beck and Kim ). Third, it remains a mystery why the patterns observed differ
from those triggered by other well-known syntactic constraints, such as islands: in the
case at hand, only covert movement is proscribed. Fourth, as noted in the previous
section, syntactic analyses either avoid defining the class of intervention triggers or
provide a semantic definition.
Given these problems, Beck () proposes a semantic theory of intervention in
wh-questions. The key element in Becks theory is the semantic notion of focus, i.e. the
presence of alternatives in the semantics, which is a property common to interveners
and wh-phrases. In intervention configurations, the focus operator that comes with
the intervention trigger applies to a complement containing a wh-phrase. If the latter
were a regular focused phrase, this would reset the focus value of the c-commanding
node to its ordinary semantic value, and the semantic computation would proceed
along the tree. Wh-phrases, however, have no ordinary semantic value. Thus, the
c-commanding node has an undefined value, which is inherited by the larger structures in the course of the semantic derivation. Ultimately, the entire question ends up
undefined, since the licensing question operator in C requires a focus semantic value
as its argument. A structure with an undefined semantic value is uninterpretable, and
hence, according to Beck, ungrammatical.
Beck and Kim () show that the same semantic account is applicable to alternative questions, where the element introducing alternatives is not a wh-phrase, but
rather a disjunctive phrase. Again, the focus operator interferes with the evaluation
of alternatives by the question operator, and the structure as a whole cannot receive
a defined question semantics. The yes/no reading is not affected because it does not
involve alternatives to begin with.
Becks semantic approach to intervention effects seems preferable over syntactic
alternatives to the extent that the property underlying the effects is indeed the semantic notion of focus. However, there exists another notion of focusinformation structural focuswhich, despite often coinciding with the semantic category in a given
sentence, is not identical to semantic focus. It is thus crucial to consider whether this
might be the type of focus relevant to intervention effects.
A brief excursus on the informational articulation of sentences is in order. IS focus
is one of the two major primitives making up the informational articulation of the
sentence, along with the ground, following Vallduv (). The focus constitutes
the informative part of the sentence, adding to or modifying the hearers knowledge
store. It is realized across many languages, if not universally, by obligatory intonational
prominence. The ground is the complement to the focus; in informational terms, it is
the portion of the sentence already established in the hearers knowledge store. The
ground is divided into a topic and a tail; the former points to the specific address
I replace Vallduvs term link with the more widely known notion of topic, because the category under
discussion here, unlike a link, is not restricted to shifting topics or to sentence-initial position.


in the hearers knowledge store where the information contributed by the focus is
to be entered. The tail signals how the information is to be entered under a given
address, and necessarily lacks intonational prominence. This informational articulation is illustrated with the English example from Vallduv () in (), where the
boss constitutes the topic, hates is the focus, and broccoli is the tail. In this case, the
hearer is instructed to go to the entry the boss and substitute the new information
hates for V in the existing record The boss V broccoli.
() The boss hates broccoli.

(Vallduv : )

Returning to the issue of different notions of focus, we find that a semantic focus
involving an operator like only or even tends to also be the IS focus of the sentence it
occurs in, but does not have to be. This claim is discussed at length in Vallduv ()
and Dryer (), and is illustrated in example (b) above, where the first only-phrase
is the IS focus and the second is not.
The question, then, is what happens when an only-phrase, i.e. a potential intervener,
is not the informative portion of the sentence. If intervention effects persist, we can
conclude that semantic focus is the underlying cause, while absence or amelioration
of the effects indicates that the IS notion of focus is the relevant one. The findings
presented in Tomioka () and Eilam () are unequivocal: intervention effects
are ameliorated or eliminated in contexts where a potential intervener is not the IS
focus. These contexts include wh-questions in which the intervener is an embedded
subject or a non-subject, which were mentioned above, as well as the English example
in () and (). Here the potential intervener is provided in a context preceding an
alternative question, as opposed to the absence of such a context in the unacceptable
example in ().
() Context: The graduate students in linguistics took two preliminary exams, in
syntax and phonology, last week. The results were surprising: there was one
exam that all the students, including John, passed, but no one except John
passed the other.
() Did only John pass syntax or phonology?

(Eilam : )

Now compare () with (), repeated from ():


Does only John like Mary or Susan?

Crucially, () does not differ from () in its syntax or semantics, but rather only
in its information structure and concomitant prosody. The semantic focus only John
is part of the backgrounded material in (), as indicated by the lack of prosodic
prominence, enabling it to be accommodated in the alternative question and thus
avoiding an intervention effect.

indicates that the alternative question reading is unavailable.

Focus Intervention in Declaratives

Having established that IS focus, rather than semantic focus, plays a role in intervention effects, the obvious question is why this is the case. In other words, what wellformedness condition(s) do intervention structures violate? The answer originally
given in Tomioka () and elaborated on in Eilam () is that an IS focus cannot
be integrated in the informational articulation of wh-questions and alternative questions, which already has one such focus, namely, the wh-phrase or disjunctive phrase.
Given that a sentence contains one and only one focus (Lambrecht ), any element
outside the focus must be a topic or a tail. However, a phrase involving an only-type
operator can be neither of the two. It cannot be a topic because it is non-referential
(cf. Reinhart ), and it is incompatible with tailhood as long as it bears prosodic
prominence (cf. Vallduv ). The resulting mismatch between the components of
the sentence and the informational articulation derives judgements of degradedness,
i.e. intervention effects.
In order to express the intended meaning of the sentences in question, it is necessary to transform the potential intervener into a non-IS focus, and this is precisely
what the various repair strategies noted above do. For example, scrambling as in (b)
and (b) places the intervener in the phonologically reduced domain which follows
foci, including wh-phrases, allowing the intervener to function as (part of) the tail.
The effect of phonological reduction, or deaccenting, can also be illustrated in English
alternative questions. When the disjunction precedes the potential intervener only
Mary in (), the sentence is acceptable because the latter does not have a pitch accent,
and can thus be in the tail. Attempting to mark Mary with a pitch accent, as would
be expected of the associate of only, eliminates the alternative question reading ().
This refutes the claim of Beck and Kim () whereby () bears out the role of
c-command in intervention effects; on the contrary, together with () it confirms
that the effects are not sensitive to hierarchical relations. This is obviously unexpected
under a syntactic or semantic approach.
() Did John or Susan invite only Mary?

Did John or Susan invite only

(Beck and Kim : )


All in all, the case for an IS approach to intervention effects, based on wh- and alternative questions, is rather strong. What would make the case even more compelling
is data from environments other than questions. That is, if the ill-formedness of intervention configurations reflects basic constraints on the informational articulation of
sentences, there is no reason to assume that intervention effects should be limited
to questions. Indeed, we have already observed that they appear in declarative sentences; the next section takes up the relevant examples and shows that they decisively
adjudicate in favour of an IS analysis.


See Eilam (, ) for a range of additional evidence for the IS approach and against its competi-


. Intervention Effects in Declaratives

Examples of intervention effects in declaratives given in section . are repeated below,
where the IS focus is a question/answer focus () or corrective focus ().
() a. What did only John drink?
b. Only John drank only BEER.
c. ??Only John drank beer.
() a. Its not true that only John drank wine, only John drank only BEER.
b. ??Its not true that only John drank wine, only John drank BEER.
The presence of the effect in these sentences is exactly as predicted under the IS
analysis of intervention effects put forward in the previous section. That is, the effect
is correlated with the position of the potential intervener: when the latter precedes
the IS focus, as in () and (), the result is illicit, but if it follows the IS focus,
the sentence is acceptable. The potential intervener can follow the IS focus in one
of two constellations: when the IS focus is the subject (() and ()), and when the
IS focus is the (underlying) object, but the syntactic structure reverses its order vis-vis the potential intervener. In English this can be done inter alia via passivization
() and specificational copular constructions (), while in Catalan, for example,
right-dislocation is possible (). These examples also show that there is nothing
semantically or pragmatically wrong with the sentences in () and (): the meaning
can be conveyed via a different structure.
() a. Who drank only beer?
b. Only JOHN drank only beer.
() Its not true that Mary drank only beer, only JOHN drank only beer.
() a. What did only John drink?
b. Only BEER was drunk by only John.
() Its not true that only John drank wine, beer was the only thing that only John
() a. Qu va beure noms el Joan?
what drank only John
What did only John drink?
b. Noms el Joan va beure noms cervesa.
only John drank only beer
Only John drank only beer.
To the best of my knowledge, this data has not been previously discussed in the literature. The sole
exception is Rooth (), who mentions an example analogous to () in the context of his theory of focus
realization. However, he erroneously attributes the degradedness of () to an unspecified phonological
constraint, and does not note its acceptable counterparts in ()().

Focus Intervention in Declaratives

c. Noms cervesa, va beure noms el Joan.
only beer
drank only John
Only beer, only John drank.

(Laia Mayol, p.c.)

The parallelism between the declarative examples and the questions examined in
section . is clear. First, the configurations are the same in the two types of examples,
as schematized in (). The difference between the unacceptable and acceptable configurations is correlated with a word order change in the case of questions, whereas
in the declarative examples it can either be a function of the IS alone, dictated by the
preceding context and reflected in the prosody, or a structural change.
() a. /??[ only ] [ XP ]
b. [ XP ] [ only ]
The set of intervention triggers in questions and declaratives is also the same; in
addition to the examples with only above, Eilam () shows that intervention effects
in declaratives are caused by NPIs in languages like Japanese, although I omit the
demonstration here. Lastly, as illustrated in Eilam () with respect to questions,
intervention effects in declaratives can be ameliorated in certain structures by virtue
of their pragmatic properties, not necessarily mediated by the syntax or the prosody.
Thus, placing the potential intervener in a pseudocleft derives a well-formed sentence,
as in (), because the content of the free relative, including the only-phrase, is presupposed and hence backgrounded.
() a. What did only John drink?
b. What only John drank was beer.
To complete the description of the declarative data and further illustrate its import, I
introduce an observation made by Atlas () regarding the truth-conditions of the
pair of sentences in (). As noted by Atlas, (b) entails that no one other than John
eats rice (and that rice is the only thing that no one but John eats), while (a) does
() a. Only John eats only rice.
b. Only rice is eaten by only John.

 No one other than John eats rice.

No one other than John eats rice.

Atlas attributes the difference between the sentences to a putative semantic difference
between the active and passive voice in English. Although there may be a variety of
differences between the active and passive, I know of no independent evidence for a
truth-conditional distinction between the two. Moreover, there is no need to ascribe
the pattern in () to properties of the active or the passive; rather, it falls out from the
possible questions under discussion (QUD) each sentence can answer, in accordance
with the configurations described in (). Thus, while (a) can be an answer to the
QUD Who eats only rice? (), and thus does not entail that John alone eats rice, it


cannot answer the QUD What does only John eat?; as illustrated in (b), this would
require the illicit configuration of (a). (b), however, can be the answer to the latter
QUD, as shown in (c), since the configuration is well-formed (cf. (b)).
() a. QUD: Who eats only rice?
b. Only JOHN eats only rice.
 No one other than John eats rice.
() a. QUD: What does only John eat? / What is eaten by only John?
b. Only John eats only RICE.
c. Only RICE is eaten by only John.
No one other than John eats rice.
Decisive evidence that this difference between the sentences is not a function of the
active/passive distinction is provided in (), a specificational construction which has
the same entailment as the passive in (c).
() Rice is the only thing that only John eats.
No one other than John eats rice.
The configurations schematized in () thus affect not only the judgements speakers
give for a sentence but also potentially its truth-conditions. These configurations are
information-structural in nature, as they make crucial reference to the notion of IS
focus, and should therefore be accounted for under the IS approach to intervention.
However, there are certain aspects of the declarative data which distinguish them from
questions and thus require further attention.
The key difference between the question and declarative contexts is the fact that
in the latter case, the problematic only-phrasesi.e. the intervenersare not IS foci.
Rather, since they are material repeated from a lead-in sentence, whether an interrogative or a declarative, they are not potential candidates for focus-hood. Why are
these only-phrases nevertheless incompatible with the informational articulation of
the declarative sentences under discussion? I argue that they cannot fulfil the role of
the two available IS categories in these examples, topic and tail. They cannot be topics
because of their non-referentiality, as mentioned in section .. The only-phrases in
the declarative examples cannot be tails as well because subjects are default topics
(Reinhart , Lambrecht , Erteschik-Shir ), and subjects preceding the
nuclear stress are specifically incompatible with tailhood (Vallduv ). In order
for a subject to be a tail, it typically has to appear in a postnuclear position, which is
by and large not possible in English. In a language like Catalan this option is available
and utilized, as in (), where the object de pa is the topic, and the subject mon germ
the tail. The awkward English translation, using right-dislocation, reflects the lack of
an appropriate equivalent in this language.

Focus Intervention in Declaratives

() De pa
no eni MENJA, mon germ.
of bread no obj eat.sg my brother
Bread he doesnt eat, my brother.

(Vallduv : )

Independent evidence for the putative relation between subjecthood, topic status, and
the occurrence of intervention effects comes from two types of sentence: those in
which the potential intervener is the subject but not the topic, and those in which
it is not the subject, and therefore has no inclination to be the topic. The first type is
provided in (), a there-existential sentence, in which the subject, which also happens
to be a potential intervener, is not the topic; rather, this role is filled by the implicit
spatiotemporal parameters of the sentence (see Erteschik-Shir ).
() a. Where are there only skyscrapers?
b. There are only skyscrapers only in TOKYO.
The second type of sentence is illustrated in (), where the potential intervener is the
indirect object rather than the subject.
() a. What did Mary give only John?
b. Mary gave only John (only) A BOOK .
As expected, both examples are considerably better than sentences in which the
intervener is the subject and interpreted as the topic by default. The first example
shows that elements which are not default topics do not create intervention effects,
and the second that non-subjects do not trigger these effects; in other words, insofar
as a potential intervener can be a tail, it will be innocuous in terms of intervention.
Note also that the difference between the previous sentences and () and () constitutes further evidence for the IS analysis. There is no obvious syntactic or semantic
distinction which would predict the attested patterns of acceptability.
The proposed analysis for the unacceptable declarative examples allows us to also
explain what underlies their acceptable counterparts in various languages. The case
of subject IS foci, as in (), repeated from (), is clear: the potential intervener is an
object and unquestionably deaccented, meaning that nothing bars it from being the
() a. Who drank only beer?
b. Only JOHN drank only beer.
The English passive in (), repeated from (), places the potential intervener in the
domain of deaccenting following the IS focus and demotes it in structural terms to a
by-phrase, enabling it to be the tail. In the Catalan example in (), repeated from (),
right-dislocation is specifically used to mark elements as tails (see Vallduv ).

I thank Satoshi Tomioka for suggesting this type of sentence.


() a. What did only John drink?

b. Only BEER was drunk by only John.
() a. Qu va beure noms el Joan?
what drank only John
What did only John drink?
b. Noms cervesa, va beure noms el Joan.
only beer
drank only John
Only beer, only John drank.
There remains one example which must be considered in light of the IS analysis of
intervention in declaratives. This analysis, if correct, seems at first glance to run into
difficulties given () and (), repeated from () and (). In this case, an onlyphrase is permitted in an alternative question because it has been introduced in a
preceding context and thus does not compete for the status of IS focus.
() Context: The graduate students in linguistics took two preliminary exams, in
syntax and phonology, last week. The results were surprising: there was one
exam that all the students, including John, passed, but no one except John
passed the other.
() Did only John pass syntax or phonology?
If prenuclear subjects generally resist being tailsand this is a key factor underlying
the intervention effects in declarativeswhy does it not yield ill-formedness in ()?
The reason, I argue, is that the context in () and the question itself unambiguously
establish the exams, rather than John, as the topic of the question. The question is
a request for information regarding the exam which only one student, John, passed,
and not a request for information about John. Accordingly, the potential intervener
only John can be the tail, despite its prenuclear subject status, and a clash between its
semantic properties and topic-hood is avoided. The same is not true of the declarative
examples, where no alternative to the subject as topic is set up.
The information-structural approach to intervention has been successfully applied
in this section to the novel data from declarative sentences. On a par with questions, intervention effects in these sentences result from incompatibility between
the IS categories available in a sentence and the elements making up the sentence.
The fact that the IS analysis proposed for wh- and alternative questions also captures the declarative examples strongly supports its validity as a general approach to
Let us briefly consider alternative syntactic and semantic approaches to intervention in light of the declarative data. First, this data refutes approaches which refer
specifically to wh-phrases, such as Pesetsky (), because it shows that intervention is not a property of questions per se. More generally, it is not clear how any

Focus Intervention in Declaratives

movement-based syntactic theory of intervention would account for the declarative

examples. Given the well-motivated claim that licensing of IS focus does not involve
covert movement (Rooth , Newmeyer ), the acceptable and unacceptable
declarative examples are syntactically identical and therefore should yield identical
speaker judgements. There is also a class of syntactic theories of intervention effects
which do not invoke movement, but rather view them as a mismatch between properties of the intervener and its position in the tree (Simpson and Bhattacharya ,
Grohmann ). While these theories are closer to the IS approach in that they
define interveners in IS terms, they adhere to a hierarchical view of intervention,
and therefore generate incorrect structural predictions. In the case of the declarative
examples, these syntactic theories are forced to assume that subjects in English generally occupy a topic position in the C-domain, which is inaccessible to only-phrases.
I am not aware of evidence that subjects are higher than Spec,IP in standard English
As for the semantic approach to intervention, Beck () notes its potential relevance for contexts other than questions and thus extends it to all phenomena involving
the evaluation of alternatives in the semantics, including alternative questions and
disjunctions in general, NPIs, and focus (see Beck , Beck and Kim , Beck
and Vasishth ). The basic idea applicable to all these cases is that a focus operator
blindly evaluates alternatives in its scope; thus, if this operator comes between a higher
operator and its intended argument, the higher operator is unable to evaluate the
alternatives introduced by the argument. Consequently, interpretations which depend
on the higher operator evaluating the alternatives are unavailable, and when this is the
only interpretation possible (e.g. in wh-questions), the result is ungrammatical. The
generalization capturing all these cases is given in (), where the operator (Op) can be
the operator found in questions, the operator evaluating question/answer congruence,
or the operator associated with a particle like only. The intervening operator is any
element that can give rise to a focus-affected reading, i.e. that comes with the
() General Minimality Effect MIN: The evaluation of alternatives introduced by an
XP cannot skip an intervening operator.
[ Op [ C [ XP ]]]
(Beck : )

This approach seems to fare no better than syntactic analyses, since it does not make
the IS distinctions necessary to tell apart well-formed vs. ill-formed configurations
involving potential interveners. In fact, given a further assumption Beck makes, which
is crucial for her theory, the focus structures described here refute the constraint in
(). In the analysis of intervention in wh-questions, Beck assumes that only-type
operators always attach to verbal projections and clausal nodes, even in cases of
apparent DP adjunction. This derives the same LF representation, in (), for both the


acceptable and unacceptable examples in () and () (Q/A in these representations

is the operator evaluating question/answer congruence).
() [Q/AC [C [onlyD [D [onlyE [E [JohnF drank beerF ]]]]]]]
() Only JOHN drank only beer.

Only John drank only BEER.

As the representations are identical, there is no obvious way to predict the wellformedness of (). Although Beck gives independent motivation for the assumed
positions of the only-type operators in these representations, they are needed in any
case in her theory. If only-type operators could attach locally to DP subjects, intervention effects in wh-questions would not be expected under the theory to begin with.
That is, the potential intervener would evaluate the alternatives introduced by the DP
subject and would not interfere with the relation between the higher question operator
and wh-phrase. Furthermore, because focus association is not selective according
to Beck (see footnote ), the semantic computation cannot be fixed so as to avoid
unintended associations. All in all, there seems to be no way to salvage the theory and
still be able to capture the entire range of attested data.
Two types of potential non-information-structural explanations for intervention
in declaratives have been assessed here. Both fall short, because the acceptable and
unacceptable sentences do not seem to differ in any relevant syntactic or semantic
property. Rather, it is only their informational articulation, created by different preceding contexts and structures, and possibly reflected in the prosody, which teases
them apart. Accordingly, only an IS approach of the type first proposed to account for
intervention in questions is able to cover this data.

. Conclusion and Implications

This chapter has sought to substantiate the information-structural analysis of intervention effects, by demonstrating that it extends to a set of data hitherto unnoticed
in the literature. The same basic phenomenon observed in wh- and alternative questions, where an only-type operator influences their acceptability or interpretation, is
also found in declarative sentences. However, in order to detect this phenomenon
in declaratives, it is necessary to set up a particular lead-in context, which in turn
forces a specific informational articulation on the declarative sentence. Mismatches
between the elements making up the sentence and their IS category then give rise to
degradedness, i.e. intervention effects.
In order to get the theory off the ground, Beck is forced to make the debatable assumption that
is not selective in choosing which foci to evaluate. See Wold () for an alternative model, involving

Focus Intervention in Declaratives

Although the context-sensitivity of intervention effects is by itself an indication

of their information-structural nature, a range of potential syntactic and semantic
explanations were nonetheless considered. These were found to be lacking, at times
overly strongpredicting ungrammaticality for perfectly acceptable sentencesand
in other cases weak, unable to rule out illicit configurations. By appealing to the IS
approach proposed for intervention in questions in Tomioka () and Eilam (),
it is possible to cover the entire range of data.
The IS approach espoused here requires a strict separation between the semantic
and IS notions of focus. A semantic focus, such as that associated with the only-type
operators discussed above, is not always an IS focus. If it were, we would not be able
to ameliorate intervention effects by backgrounding the only-phrase, as illustrated in
English alternative questions, nor would we be able to differentiate the acceptable and
unacceptable declarative examples. This supports pragmatic, or strong, theories of
focus association (e.g. Vallduv , Rooth , Dryer ), since they make the
distinction between semantic and IS notions of focus. It contests the rival semantic,
or weak, class of theories (Rooth , Krifka ), which argue that operators
like only need to associate with an IS focus because it alone provides the required
quantificational structure. In such a theory, the crucial distinction between semantic
and IS foci becomes moot.
The findings of this study also have significant implications for the conception of
IS in the grammar. A particular phenomenonintervention effectswas shown to
reduce to well-formedness conditions which are exclusively information-structural
in nature. The uniqueness of focus and the incompatibility between only-phrases and
topic-hood, for example, are generalizations which can be stated only in IS terms. The
existence of such IS well-formedness conditions is robust evidence for the claim that
IS is an autonomous level of organization of linguistic information (Vallduv ,
Eilam ). This level interacts with other levels of representation in the linguistic
system, and imposes restrictions of its own on the output of the system.

Root Phenomena as Interface
Evidence from Non-sententials

. Introduction
To what extent should syntax encode meaning? The way this question has been
addressed within the generative framework has fluctuated over the past four decades,
but the general tendency has been to complexify the syntactic component to account
for much of the complexity of natural language. In the early days of Generative Semantics, all meaning was assumed to be present at D-structure. The current Minimalist
view is that many combinatorial properties of semantics are derived from syntax, and
that (at least some) systematicities of compositional semantics are associated with
phrase structure (Ramchand ). The Cartographic approach in particular assumes
transparent mapping between form and interpretation, and syntacticizes as much as
possible the interpretive domains (Cinque and Rizzi : ), based on evidence
from word order and the order of overtly realized functional morphemes.
This chapter questions the extent to which syntax is necessary to account for two
phenomena with discourse/pragmatic properties, with the following preoccupation
in mind: how should we use Occams Razor to dispose of undesirable theoretical
entities? More specifically: should we rely on syntax (as opposed to other components)
to explain restrictions on the distribution of phenomena with a discourse import, as
advocated by the Minimalist Program and implemented in the classic Cartographic
I would like to thank Shigeru Miyagawa, Reiko Vermeulen, Mika Kizu, and Hidekazu Tanaka for
comments and enlightening remarks about the Japanese data (which doesnt imply they agree with my
analysis), and Mika Takewa for her generous help in gathering these data. Special thanks to Kristine Bentzen
for comments and discussion. Many thanks also to the audiences at the Department of Linguistics at KU
Leuven (), the GIST workshop (Ghent, ), OnLI II, Belfast, , and the SLE workshop on
New Forays into Root Phenomena.

Root Phenomena as Interface Phenomena

approach? Or should we minimize the role of syntax when these restrictions can
be captured by other components of grammar, as advocated e.g. by Culicover and
Jackendoff ()?
The empirical basis for this discussion is root phenomena in non-sententials. The
root properties of non-sententials have been discussed in the literature (Stainton
, Culicover and Jackendoff ) but to my knowledge the extent to which nonsententials can host root phenomena has not been investigated per se.
The chapter is organized as follows. Section . defines root phenomena in general
and lays out the central research question (i.e. to what extent syntax should encode the
root properties of clauses/utterances). Section . introduces the syntactic structure
relevant to the subsequent discussion: fragments (or non-sententials), i.e. non-clausal
structures with root properties. The main evidence is provided in the following two
sections, demonstrating that French fragments can host dislocated topics (section
.) and that Japanese fragments can host (root) politeness markers (section .).
Conclusions are drawn in section ..

. Root Phenomena and their Host

Root phenomena are those that normally occur in matrix clauses but are also allowed
in a restricted set of embedded (root-like) clauses (Heycock ). Typical examples
in English include auxiliary inversion, argument fronting, locative inversion, and tag
questions, as illustrated in turn below. As shown in the (b) examples, these phenomena
tend not to be possible in embedded clauses.
() a. Man, are we in for it!!
b. He discovered that boy, was I in over my head.

(Green : (a))
(Green : (b))

() a. Her regular column, she began to write again. The other ones, she never
b. When her regular column she began to write again, I thought she would be
(Haegeman a: (b))
() a. In the deepest part of the forest lived a scary Gruffalo.
b. He told me that in the deepest part of the forest lived a scary Gruffalo.
() a. Acupuncture really works, doesnt it?
b. I suppose acupuncture really works, doesnt it?
Since Emonds (), many have attempted to capture root phenomena as a syntactic
property of clauses. That approach associates the ability for a clause-type to stand
alone with its ability to host root phenomena. Root-like clauses are generally assumed
The extent to which embedded clauses allow root phenomena will not be explored here. See De Cat
(in preparation) for an analysis of French dislocation as a root phenomenon and its licensing in embedded

De Cat

to be essentially finite embedded clauses selected by verbs of assertion (such as told in

(b)see e.g. Hooper and Thompson , Emonds ), although some adjunct
clauses have also been found to allow root phenomena (see Heycock for an
overview). Under a strict version of the syntactic approach, clauses are endowed with
some kind of Boolean feature [root], and only those belonging to the type [+root]
can host phenomena such as those illustrated in ()(). It has recently been proposed
that root properties could be captured syntactically by positing a reduced structure
in non-root clauses (as in Haegeman ), and/or by appealing to restrictions on
syntactic movement across elements endowed with properties inherent to rootness
(Haegeman a, b).
Syntactic accounts of rootness acknowledge the key role of the interpretive properties of the host: it is generally assumed, following the seminal work of Hooper and
Thompson (), that root phenomena are confined to non-presupposed, asserted
clauses. I do not propose to question the need to identify the interpretive properties
of the host, but the need to associate such properties with particular syntactic structures exclusively.
I will assume, based on De Cat (), that the essential properties licensing root
phenomena are indeed interpretive in nature, and that licensing results from the
interaction of the interpretive properties of root phenomena themselves with the
properties of their host, yielding variability across and within clause types. Two types
of interpretive properties are involved: (i) those that indicate speaker involvement
and (ii) epistemic or information-structural properties. All root phenomena involve
epistemic or information-structural properties (requiring their host to be a locus of
truth evaluation la Reinhart ); speaker involvement is only required in a subset
of cases. Note, however, that nothing in the argument developed in the present chapter
hinges on that particular analysis. The main point of this chapter is to show that some
root phenomena can occur in hosts that are not clausal (even if sluicing is considered),
and that consequently, rootness is not necessarily a property encoded syntactically in
the C-domain.
Before introducing the two root phenomena of interest in this argument, I turn to a
brief introduction to the relatively understudied structure which we will see can host
(some of) them.

. Non-sententials/Fragments
Spontaneous speech prominently features utterances that, in spite of not manifesting
themselves as full clauses, are fully interpretable by their addressee. As an extreme
example of what can be achieved using such fragments, consider the exchange in

Although see Wiklund, Bentzen, Hrafnbjargarson, and Hrarsdttir () and Bentzen () for
evidence that the key interpretive feature of the host may be the ability to convey the Main Point of the
Utterance rather than being asserted or non-presupposed.

Root Phenomena as Interface Phenomena

(), involving three people. Imagine Person A looking out of the window and uttering (a) (signalling the presence of a strange sheep) and fainting immediately afterwards. Person B then comments (b) to Person C, who concludes (c) with regard
to A.
() a. Oh, un mouton
oh a sheep with dots
Oh, a dotted sheep!
b. Difficile imaginer.
hard to imagine
[Thats] hard to imagine.
c. Do
son malaise.
from where his faintness
Hence his faintness.
Fragments (also known as non-sententials, as in Progovac, Paesani, Casielles, and
Barton , or Bare Argument Ellipsis, as in Culicover and Jackendoff ) are
verbless utterances interpreted as full propositions with assertoric force. The whole
exchange in () takes place without a single full clause being uttered: the utterances
appear to consist of just a DP (a), an AP (b), and a PP (c). Note also that the first
fragment occurs out of the blue, i.e. without any preceding linguistic context.
A question that has been hotly debated since the s is whether such nonsentential utterances should be analysed syntactically as full clauses or not.
Recent proponents of the yes option include Merchant (, , ); recent
proponents of the no option include Culicover and Jackendoff (), Stainton
(), and Progovac ().
Merchant () proposes that fragments are fully propositional in the syntax,
but that most of their structure remains unpronounced. He analyses fragments as
derived by ellipsis, and more specifically as a form of sluicing. Merchant argues that
the fragment occupies the specifier of a left-peripheral phrase (which he suggests may
be FocusP) whose head is endowed with an E feature. This feature has two functions:
(i) it instructs PF not to parse its complement (and hence not to pronounce it); and
(ii) it consists of a partial identity function over propositions, which is intended to
ensure that the complement of an E-endowed head has an appropriate antecedent in
the discourse, essentially ensuring that the content of the unpronounced structure
is identifiable or recoverable. Crucial for the argument developed below is the fact
that, under the ellipsis analysis, the full structure must be inherited verbatim from
the immediately preceding discourse. Merchant (b) uses island sensitivity as key
evidence for his sluicing analysis: he attributes the unacceptability of the answer in
() to the illicit movement the pronounced constituent would have had to undergo to
reach the Spec,FocusP position.

De Cat

() Q: Does Abby speak the same Balkan language that Ben speaks?
A: No, Charlie.
No, she speaks the same Balkan language that Charlie speaks.
However, Merchant also recognizes that a direct interpretation approach is needed for
at least some fragments, for which there exists no full-blown syntactic counterpart (as
in (c)).
Culicover and Jackendoff () and Stainton () argue that the syntactic ellipsis approach cannot account for all fragments, and that the direct interpretation
approach should be postulated as default, because it is more economical. Key evidence
(to which we will return later) comes from cases where the fully pronounced structure
would be ungrammatical, for instance as a result of an island violation. Examples of
such fragments are provided below (all are adapted from Culicover and Jackendoff
). In each example, the context is provided in (a), the fragment in (b), and the full
syntactic structure one would have to postulate for it following Merchants analysis is
given in (c).
() a. Haruko has been drinking sake all weekend.
b. Yes. And shochu.
c. And shochu, Haruko has been drinking sake t all weekend.
() a. Whose sake has she been drinking?
b. Her mothers.
c. Her mothers, she has been drinking t sake.
() a. Haruko drinks sake that comes from a very special part of Japan.
b. Where?
c. Where does Haruko drinks sake that comes from a very special part of
() a. Yasu met a child who speaks Urdu.
b. With a Japanese accent?
c. With a Japanese accent, Yasu met a child who speaks Urdu t.
In addition to these, some examples of fragments have no full-clause equivalent at all:
() a. Would you like a drink?
b. How about tea?
Furthermore, Merchants sluicing analysis predicts that fragments cannot occur out
of the blue (as they require a context sentence to be copied from), contrary to what
we observe in (a) (see also () below).
Stainton () beautifully argues that such fragments have pragmatic Force
in spite of not having syntactic Force, which implies that there is no one-to-one
correspondence between what [is] done (a full-fledged speech act i.e. propositional,
force-bearing and literal) and what [is] used (lexical projections, not of semantic type

Root Phenomena as Interface Phenomena

t, not having [an expression encoding] force, and not embedded in any higher tree)
(Stainton : ).

. First Set of Data: French Dislocated Topics in Fragments

.. The evidence
Spoken French features complex fragments, consisting of a verbless phrase
(henceforth labelled the nucleus) and a peripheral (also verbless) phrase (henceforth
labelled the satellite). The nucleus conveys the main information of the utterance
(and could thus be thought of as focused). It is highlighted in bold in examples ()
and () below.
() a. [ Focus XP ] satellite
b. satellite [ Focus XP ]

(XP = nucleus)

The satellite is set off from the nucleus by dislocation prosody: left-dislocation prosody
when it appears on the left of the nucleus; right-dislocation prosody when it appears
on the right. Roughly put, left-dislocation prosody makes the right boundary of a
phrase salient in terms of pitch (its melody is characterized by a rise in fundamental
frequency (F) culminating on the nucleus of the accented (final) syllable), intensity
(there is a marked stress on the last syllable of the left-peripheral element), and
duration (the stressed syllable is lengthened)see Deshaies, Guilbault, and Paradis
(), Rossi (), Mertens, Goldmann, Wehrli, and Gaudinat (), and De Cat
(a). A phrase uttered with right-dislocation prosody is destressed and tends to
feature a low, flat pitch, or a lower copy of the preceding prosodic contour (Delattre
, Ashby , Rossi , Mertens et al. , De Cat a).
The satellite cannot be seen as dislocated in the traditional sense of the term, as it
is not associated with a position or a resumptive element inside a clause. However,
it displays a number of core characteristics of dislocated elements in spoken French
(see De Cat a): (i) it receives similar prosody; (ii) it can be omitted (provided
its referent is salient enough in the context); and (iii) it appears to fulfil the same
informational function as dislocated elements: either it expresses what the utterance
is about (as in ()), or it restricts the (temporal or spatial) domain within which the
predication holds (as in ()).
A noteworthy exception is (), which shows that satellites can in principle be resumed by a clitic (just
like dislocated elements, as shown by the indexing in ()), but only if there is a syntactic host for the clitic.
Exploring what makes the presentative voil a possible clitic host is, however, beyond the scope of this
The data set used in this study comes from two French corpora of spontaneous interaction between
children and adults: the York corpus (available on childes) and the Cat corpus. They contain speech from
four children (between the ages of ; and ;ages are given in year;months.days), and adults from
Belgium, Canada, and France. Recordings took place fortnightly (York corpus) or monthly (Cat corpus) and
lasted minutes. For details regarding the transcription and coding procedures, see De Cat and Plunkett
(), and De Cat ().

De Cat

les petits copainsi .

() a. Lesi voil,
them presentative the little friends
Here (are) the little friends.
b. Je lesi vois, les petits copainsi .
I them see the little friends
I see the little friends.
() a. Et maintenant, de la tomate.
and now
part the tomato
And now (lets add) some tomato.
b. Et maintenant, on met de la tomate.
and now
one puts part the tomato
And now lets add some tomato.
Under a WYSIWYG (what you see is what you get) approach, complex fragments
can be analysed as adjoined structures, just like their full-clause counterpart (De Cat
a, b). Adjunction is a primitive (Lebeaux ) or default operation (Boeckx
) that does not provide explicit mapping instructions and does not involve feature
checking. It allows for a loose semantic relation between the satellite and the nucleus,
the exact content of which is recovered from the context.
.. Do they host true root phenomena?
One might object that the satellites are not the manifestation of a root phenomenon.
The jury is still out as to the root status of right-dislocated topics (see e.g. de Vries ,
pace De Cat a), and left-dislocated topics in Romance have been shown to be
acceptable in contexts normally hostile to root phenomena (such as central adverbial
clauses, as in ()see e.g. Haegeman a).
() Si ce livre-l
tu le trouves la Fnac, achte-le.
if that book-there you it find at the Fnac buy-it
If you find that book at the Fnac, buy it.
(Haegeman : ())
Haegeman (a) explains the unexpected acceptability of () by the availability of
a topic position below FocusP in Romance but not in languages like English. Central
adverbial clauses are argued not to have a root projection in their C-domain (i.e.
ForceP or SpeakerDeixisP), and this blocks the projection of the higher TopicP, leaving
the lower TopicP (which is exclusive to Romance) unaffected. Crucially, those low
topics are claimed not to have root properties. Hence their occurrence in central
adverbial clauses in Romance is argued to be compatible with the non-root status
of these clauses cross linguistically.

I will leave right-dislocation for subsequent work, and concentrate below on left-peripheral satellites.

Root Phenomena as Interface Phenomena

In fragments, however, under a cartographic approach, one would have to assume

that the satellite occupies the higher TopicP position, given that everything below
FocusP is unpronounced. This would automatically confer to the satellite the status of
root phenomenon under Haegemans analysis.
Could it be that not all topics higher than FocusP are root? This is the claim of
Bianchi and Frascarelli (), who postulate the existence of three types of leftperipheral topic: Aboutness-shift topics (A-topics), Contrastive topics (C-topics) and
Given topics (G-topics). The proposed cartography is as follows (with the asterisk
indicating recursivity):
() [ShiftP A-Topic [ContrP C-Topic [FocusP [FamP G-Topic [FinP [IP
A-topics display the following set of essential characteristics:
() A-topics
a. have an impact on Common Ground management: They are associated
with a shift in the conversation and newly propose or reintroduce a topic in
the discourse (Bianchi and Frascarelli : );
b. precede FocusP (and may in fact be outside the C-domain altogether) and
have a distinctive prosodic contour (in Italian);
c. are merged in the structure as an independent speech act and can only
be conjoined with a root(-like) clause implementing another speech act
(Bianchi and Frascarelli : ); as a result they can be combined with
a clause of any illocutionary type (contrary to C-topics, which are restricted
to declarative clauses);
d. are restricted to clauses that express non-reported speech acts, which are
necessarily unembedded.
The satellites in the examples below clearly fit key A-topic criteria: they occur in interrogatives () and exclamatives (), they express a shift in perspective () and (),
and of course they occur in fragments, the unembeddable speech act par excellence:
() conveys a request, (a) and (b) an assertion, (c) an acknowledgement, and
() an order.
() Celui-l,
comme a?
that one-there like
[Shall we put] that one like this?
Common Ground management is defined as the sequence of conversational moves performed by participants [. . .] that determines the way in which the CG content develops (Bianchi and Frascarelli : ).
It is unclear whether they can be embedded at all (Bianchi and Frascarelli speculate that the
embedded A-topics in their corpus might be performance errors).
The satellites in these examples are not contrasted with a referent in the linguistic context, which rules
out a C-topic analysis.

De Cat

() a. Moi, avec lponge.

me with the sponge
I [will do it] with the sponge.
b. Moi aussi, une cape
me too a cape
Me too, [I want] a cape.
c. Ah toi aussi, un petit peu de jus dorange.
ah you too a little bit of juice of orange
Ah, you also [want] some orange juice.
() Toi, dans ta chambre!
you in your bedroom
[Go to] your bedroom!
It is unclear whether the three-way topic distinction proposed by Bianchi and Frascarelli () applies to French (see De Cat, in preparation). But even if it does, the
evidence above indicates that satellites display the characteristics of topics with root
.. Are these elided full clauses?
How much structure should we postulate for complex fragments? The sluicing
approach (Merchant , b) could in principle accommodate dislocated topics
as long as these are hosted higher than FocusP. The nucleus (in bold in () and ())
would move to Spec,FocusP, and because the topic is merged higher than FocusP, it
would remain unaffected by the silent Spell-out of Fs complement.


that one


comme a
like that


on le colorie?
we it colour

Root Phenomena as Interface Phenomena




de la tomate
some tomato



on met
we put

However, a sluicing analysis is impossible for those complex fragments that do not
have a full sentential counterpart, such as (c), (), or (), and for those whose full
sentential counterpart would be ungrammatical, such as (a) (as French does not
allow the fronting of quoi what).
() La boutonnire de ce bouton-l, la voil.
the button hole of that button-there it presentative
Theres that buttons buttonhole!
() Les bleus, oui. Les rouges, non.
the blue yes the red
The blue ones can. The red ones cant.
() a. Quoi, le bleu?
what the blue
What [is the matter with] the blue one?
b. Quoi (est-ce qu) il a, le bleu?
what (is-it that it has the blue
Similarly, a sluicing analysis is ruled out in cases where the full sentential equivalent
would violate an island constraint, as in the examples below. A sluicing analysis of
the perfectly acceptable (b) would require extracting aussi du shochu also some
shochu from a coordinate structure, as shown in the non-elided (c). However, the
latter is ungrammatical, and therefore cannot be considered the source for (b).
The parentheses indicate that the presence of a dislocated topic does not alter the
grammaticality of the example.
The examples in this case are made up and were checked with eight native speakers for acceptability.

De Cat

() a. Haruko a bu
du sake tout le weekend.
Haruko has drunk some sake all the weekend
Haruko has been drinking sake all weekend.
b. Oui. Et (sa soeur), aussi du shochu.
yes and (her sister also some shochu
(i.e. Harukos sister drank sake and (also) shochu all weekend.)
du sake t tout le
c. Et (sa soeur), aussi du shochu, elle a bu
and (her sister also some shochu she has drunk some sake all the
.. Summary of findings
The satellite in French complex fragments is the non-sentential counterpart of dislocated topics in full clauses. It has similar interpretive, prosodic, and syntactic properties. A sluicing analysis has been shown not to capture complex fragments in cases
where there is no full-clause equivalent and where the full-clause equivalent would
violate a syntactic island constraint. Under a cartographic approach, the satellite
would be equivalent to a root topic, even under the three-way distinction proposed
by Bianchi and Frascarelli (): it is necessarily higher than [FocusP], and akin to
an A-topic. I conclude that these non-clausal structures can host a root phenomenon,
despite not projecting a C-domain.

. Second Set of Data: Japanese Politeness Markers in Fragments

.. The evidence
Miyagawa () shows that the politeness marker -mas- (underlined in ()) occurs
exclusively in root-like clauses.
() Watasi-wa piza-o
pizza-acc eat-mas-pres
I will eat pizza.

(Formal register)

In verbless clauses, another politeness marker is used (-des-), which is functionally

equivalent to -mas- (Miyagawa, p.c.). In De Cat (), I show that -des-, like -mas-,
is restricted to root clauses. The simple, verbless clause in (a) hosts -des-; its embedded equivalent in (b) doesnt (it cannot), and as a result politeness is marked in the
matrix clause only in that sentence.
() a. Tomoko-san-wa atama-ga ii
Tomoko-hon-top head-nom good des
Tomoko is clever. (Lit: Tomokos head is good.)

Root Phenomena as Interface Phenomena

b. [Tomoko-san-wa atama-ga ii
to] omi-masu.
[Tomoko-hon-top head-nom good quot believe-mas
I believe Tomoko is clever.
The root politeness marker -des- can appear in fragments that cannot be accounted
for by the syntactic ellipsis approach, i.e. fragments that do not have a full clausal
structure. (b) and (b) show that -des- can be used in fragments that, under
Merchants () analysis, would have to be extracted out of an island (complex
NPs or coordinate structures). The examples below are adapted from Culicover and
Jackendoff ().
() a. Haruko-wa [nihon-no tokubetu-na tokoro-de tukurareteiru]
Haruko-top Japan-gen special-conn place-at produce.past.asp
sake-o nomi-masu.
sake-acc drink-mas
Haruko drinks sake that comes from a very special part of Japan.
b. Doko desu/ masu ka?
where des/mas q
() a. Syusyoo-wa
tomato-jyuusu to nani-ga suki-desu ka?
prime. minister-top tomato-juice and what-nom fond of-des q
The Prime Minister likes tomato juice and what?
b. Biiru-desu.
The examples in (b) and (b) show that -des- can be used in fragments for which
the non-elided counterpart (following Merchants analysis) would be ungrammatical.
() a. Nanika
nomi masu ka?
something drink mas q
Would you like a drink?
b. Otya-wa doo desu/ masu ka?
tea-top how des/mas q
How about tea?

Incidentally, note that the impossibility of using -mas- in the fragments above provides additional
evidence against a sluicing analysis: if these were syntactically full clauses copied from an antecedent in the
discourse, they would involve verbs and the use of -mas- rather than -des- would be obligatory. Merchants
variable island repair strategy, which is invoked to account for the extractability of question words out of
elided islands, can therefore not be invoked here, on account of the presence of des rather than the expected

De Cat

() a. Pizza-o tor-oo.
pizza-acc order-vol
Lets get a pizza.
b. Margherita desu/ masu ka?
margherita des/mas q
.. Are these in fact reduced clefts?
An alternative analysis of the Japanese fragment data presented above is to derive
them as reduced clefts, along the lines of Saito () or Kizu (). However, as
pointed out in Merchant (), the interpretive properties of clefts differ from those
of fragments. Crosslinguistically, it is widely acknowledged that the object of the
matrix copular clause is in focus and that the embedded clause is presupposed (see
e.g. Kizu for Japanese). The focused element itself is normally presupposed in
clefts, and tends to require an exhaustive interpretation. Fragment answers are not
subject to such restrictions, as illustrated in () (from Merchant : ):
() Q: What did the burglar take?
A: Nothing. / It was nothing at all that the burglar took.
The requirement for an exhaustive reading in clefts but not in fragments is also visible
in the examples below. () illustrates this with interrogative clefts vs. interrogative
fragments (van Craenenbroeck ). The modifier for example is impossible in the
cleft question (as it is incompatible with an exhaustive reading) but perfectly fine in
the fragment.
() You should talk to somebody in the legal department for help with that.
a. Who, for example?
b. Who is it, for example?


A similar effect is observed in (), where the presence of the modifier else signals a
non-exhaustive answer, violating the expectation of exhaustivity arising from the use
of the cleft.
() Harry was there, but
a. I dont know who else.
b. I dont know who else it was.
The contrast below indicates that English and Japanese carry the same semanticofunctional restrictions on clefts, and that fragments are immune to those in both

There are exceptions, in which new information does appear in the embedded clause: see Delin ().

Root Phenomena as Interface Phenomena

() Kanojo-wa anata-yori
sinsetsuna tomodachi-ga ooi.
she-top you-more than nicer
friends-nom many
She has nicer friends than you.
a. Kanojo-ga anata-yori
dono sinsetuna
she-nom you-more than many-nmlzr-top which nicer
Lit.: Which nicer friends is it that she has more than you?
b. Dono tomodachi-desu ka?
which friends-des
Finally, Merchant () notes that clefts are attested (albeit infrequently and
possibly not in all languages) in out-of-the-blue contexts, in which case the object
of the matrix clause receives an existential interpretation. However, the contrasts in
() show that there are restrictions on the availability of an existential reading in
putative clefts, depending on the (overt) presence of the embedded clause. This is not
predicted under a sluicing approach, as the ellipsis occurs post-Spell-out and in the
PF component only.
() a. Oh, a sheep!
b. Oh, its a sheep!
c. Oh, its a sheep (that I see / that is there)!
The fragment in (a) cannot do more than signal the existence/presence of a
sheep. The unacceptable cleft reading of (b) implies some kind of ambiguity resolution in the context (e.g. the speaker may have been wondering what was making a
strange noise, or what had been making holes in the curtains): it signals identification
and not mere existence. The same appears to hold for Japanese:
() a. Ara, hitsuji desu!
oh sheep des
Oh, a sheep!
b. Ara, asoko ni mieru no
wa hitsuji desu!
oh there at is visible nmlzr top sheep des
Lit.: Oh, its a sheep that I see.
c. Ara, asoko ni iru no
wa hitsuji desu!
oh there at exists nmlzr top sheep des
Lit.: Oh, its a sheep that exists.
.. Summary of findings
Japanese fragments have been shown to host politeness markers, a phenomenon
restricted to root hosts. Politeness markers are attested in fragments that elude a

De Cat

sluicing analysis. An analysis in terms of reduced clefts has been shown to be inadequate on interpretive grounds. Here again the conclusion has to be that non-clausal
structures can host a root phenomenon, in spite of not projecting a C-domain.

. Conclusions
The data discussed here provide clear evidence for the existence of syntactically nonclausal structures with root properties. The observation that non-sententials (or fragments) have pragmatic Force is not new (Stainton ). What is new is evidence that,
in spite of lacking syntactic Force, fragments can host phenomena that are restricted
to root hosts when they occur in full clauses. These include root phenomena with
an information-structural import (the dislocated topic in French complex fragments)
and root phenomena expressing speaker anchoring (politeness marking in Japanese
Two general conclusions can be drawn from the above. First, the empirical domain
of syntax should not necessarily be coextensive with the sentence or clause: smaller
structures also need to be accounted fornot just because they are pervasive in
spontaneous speech, but also because they display syntactic phenomena in need of
an explanation. This does not automatically call for a revision of core minimalist
assumptions (as demonstrated by Progovac ). Secondly, the existence of root
phenomena in fragments that cannot be captured by a full-clause analysis shows that
what confers root properties to a host is not necessarily syntactic. If the interpretive
component needs to be relied on to confer root status to a syntactic structure in those
cases, should we not confer the same role in full clauses? I would be tempted to follow
the wisdom of William of Occam, especially if it leads us towards a better account for

Contrast and its Relation to wa
in Japanese and nun in Korean

. Introduction
This chapter is concerned with the syntactic and interpretive properties of contrastive
focus and contrastive topic in Japanese and Korean, and their interaction with the
particles contrastive wa and nun. As illustrated below, a contrastive wa- or nunphrase may appear in situ, in (a) for Japanese and (a) for Korean, or undergo
movement, as in (b,c) and (b,c). The kind of contrast implied is indicated by the
possible continuation in parentheses.
() a. John-ga Sue-ni ano hon-wa ageta.
John-nom Sue-to that book-wa gave
b. John-ga ano hon-wai Sue-ni ti ageta.
c. ano hon-wai John-ga Sue-ni ti ageta.
John gave that book to Sue. (But, Im not sure about this other book.)
Earlier versions of parts of this chapter were presented at GLOW , Japanese/Korean Linguistics
, and the Workshop on Formal Altaic Linguistics in , the Colloquium on Generative Grammar in
, and at the University of Potsdam in . I thank the participants for their helpful comments. I would
also like to thank Daniel Bring, Jieun Kim, Ad Neeleman, Kriszta Szendroi, and Michael Wagner for the
numerous inspiring discussions. Thanks are also due to my informants for their immeasurably patient
help. Parts of this research have been funded by the FWO (No. G), the AHRC (No. ), and
the British Academy (No. SG).
There are non-contrastive wa and nun as well, which appear on non-contrastive topics (Kunos
theme). They have a different set of grammatical properties and are generally assumed to be separate lexical
items from contrastive wa and nun, respectively. This chapter is not concerned with non-contrastive wa and
nun. (See Heycock for an overview on wa and C. Lee and Hetland for nun.)
For reasons not entirely clear to me, a wa- or nun-marked phrase sometimes prefers not to surface
adjacent to the verb. This is perhaps due to the phonological fact that a non-contrastive wa-/nun-phrase
forms a separate prosodic domain, but the verb typically forms a domain with an immediately preceding
phrase (Nagahara , Nakanishi , Jun ). An adverbial can be inserted to circumvent this issue
without affecting the claims being made.


() a. John-i
Sue-hantey i chayk-un
John-nom Sue-to
this book-nun gave
b. John-i i chayk-uni Sue-hantey ti cwuesse.
c. i chayk-uni John-i Sue-hantey ti cwuesse.
John gave this book to Sue. (But Im not sure about that other book.)
Despite the vast amount of literature, there is no consensus on whether contrastive
wa- or nun-phrases should be treated as contrastive topics or contrastive foci. For
Japanese, the general view is that any contrastive wa-phrase is a contrastive topic (e.g.
S. Tomioka , Heycock , Kishimoto ), but some researchers have argued
that contrastive wa-phrases behave more like foci (Kuroda , , Endo ,
Oshima ). For Korean, the predominant view is that a contrastive nun-phrase is
a contrastive topic in clause-initial position, but is a contrastive focus elsewhere (Choe
, Han , Choi , , Gill and Tsoulas ). Some others have argued
that they are always contrastive foci (Choe , M. Lee ), while C. Lee (,
) has argued that they are invariably contrastive topics. A factor that contributes
to the general confusion for both languages is the absence of clear definitions of
contrastive topic and contrastive focus in the relevant discussion.
This chapter examines the syntactic distribution of contrastive topic and contrastive
focus in Japanese and Korean, adopting the definitions of contrastive topic and contrastive focus that are motivated independently in the literature on information structure. I show that a contrastive focus on this definition must be case-marked in both
languages and cannot be wa- or nun-marked. I argue furthermore that the meaning of
contrastive wa and nun is in fact compatible with both focus and topic. The evidence
is based on a set of predictions regarding the distribution of contrastive topic and
contrastive focus with respect to each other, made by a particular theory of mapping
between syntax and information structure, proposed by Neeleman and van de Koot
() and Neeleman, Titov, van de Koot, and Vermeulen (). Neeleman et al.
assume that [topic], [focus], and [contrast] are autonomous notions of information
structure that syntax may target and argue that there are specific mapping rules pertaining to each notion. They argue furthermore that contrastive topic and contrastive
focus should be analysed as composites of [topic] and [contrast], and [focus] and
[contrast], respectively (see also Molnr , Giusti ). They show with data from
Dutch, Japanese, and Russian that, depending on the language, a contrastive topic
can be subject to the mapping rule for [topic] or one for [contrast], and similarly, a
contrastive focus can be targeted by the mapping rule for [focus] or one for [contrast].
The predictions regarding the syntactic distribution derive from the combination
of the mapping rules and an independent well-formedness condition at the level of

In Korean, the nominative case marker is realized as ka after a vowel and as i elsewhere. Similarly, the
accusative marker is realized as lul if following a vowel and as ul elsewhere, and the particle nun as nun
after a vowel and as un elsewhere.

Contrast and its Relation to wa and nun

information structure. Contrastive wa- and nun-phrases that are interpreted as topics
show the predicted distribution for contrastive topics, but those that are interpreted
as foci do not.
These predictions do not follow from an approach that assumes designated functional projections for discourse-related notions, such as the cartographic approach
(Rizzi ). The data discussed here, therefore, support an approach that assumes
mapping rules at the interface between syntax and information structure, and argues
against an approach that assumes that much of discourse is part of syntax.
In section . I clarify the definitions of topic, focus, and contrast that this chapter
adopts. Section . examines the syntactic distribution of contrastive topics and contrastive foci on those definitions. Section . demonstrates that the set of predictions
that follow from Neeleman et al. () are borne out. In section . I examine the
meaning associated with contrastive wa and nun, and argue that it is compatible with
both topic and focus, providing syntactic evidence.

. Topic, Focus, and Contrast

The present chapter adopts the definition of sentence topic in the sense of Reinhart
(), and not that of discourse topic. A sentence topic is a syntactic category that
introduces a referent as what the rest of the sentence is about, while a discourse topic is
what the whole discourse is about and can be more abstract. The referent introduced
by a sentence topic often functions as a discourse topic, and continues to do so in the
subsequent discourse. The grammatical relevance of the distinction between the two
types of topic is widely recognized (Givn , Vallduv , Lambrecht , Choi
, Frascarelli and Hinterhlzl , Vermeulen ).
A sentence topic can be identified as the item X in the reply to the request tell me
about X (Reinhart ). Such a request instructs the hearer to give a response that
encodes X as the sentence topic. Thus, John in Speaker Bs utterance below is a sentence
topic. (A topic is indicated by double underlining.)
() A: Tell me about John.
B: Well, John is a student from Canada.
That John in Bs utterance in () is indeed a sentence topic, rather than John in As
request, is suggested by three facts. First, native speakers do not interpret the imperative uttered by A as being about John. Secondly, Bs utterance is also felicitous if the
request is less specific about what is to be the topic of discourse, such as tell me about
someone in your class. Finally, a well-known property of a sentence topic is that it
must be referential (Reinhart ). The infelicity of Bs reply in () demonstrates this
I have consulted four native speakers regarding the English data reported here. The reported judgements were shared by all.


property. Non-specific quantifiers are not referential and therefore cannot function as
sentence topics (Endriss ). The fact that As request in () is possible, where the
target of the request is not specific, also shows that, by analogy, John in As request in
() is not a sentence topic.
() A: Tell me about some people in your class.
B: Few students are from Canada.
For [focus], I adopt the standard definition that it provides a highlighted piece of new
information with respect to the rest of the sentence. As such, it can be identified as
the item corresponding to the wh-part of a preceding question (Rooth ).
[Contrast] is a notion that can combine with focus or topic (Molnr and Winkler
, and references therein). There are some interpretive differences between contrastive topic and contrastive focus, which, I argue, derive from the view that focus is a
propositional notion, while a topic is an utterance-level notion. I assume that contrast
implies the rejection of at least one alternative in the set of relevant alternatives
generated by the contrastive item. Thus, a contrastive focus implies the rejection of at
least one alternative proposition, while a contrastive topic implies the rejection of at
least one alternative utterance (S. Tomioka ). I will not examine the details of
the semantics of the notion [contrast], due to lack of space, although I will discuss
proposals for contrastive wa and nun in section .. However, I will introduce below
some discourse contexts which require a contrastive focus and a contrastive topic.
A context requiring a contrastive focus is an instance of correction, such as (),
where B corrects the statement by A. In Bs sentence, Rosa is a contrastive focus.
B asserts that the alternative proposition expressed by As utterance, namely that
John saw Claire yesterday, is false. (Focus is indicated by small caps and contrast
is indicated by italics.)
() A: John saw Claire yesterday.
B: No, he saw ROSA yesterday.
{[John saw Rosa], [John saw Claire]}
Another context which requires a contrastive focus is provided by a disjunctive question. In (), Rosa in Bs utterance is a contrastive focus. The explicit mention of the
alternative Claire has the effect that the reply implies that the alternative proposition,
John saw Claire yesterday, is false. The contrastive reading can be further strengthened by the use of the focus-sensitive particle only.
() A: Did John see Claire or Rosa yesterday?
B: He saw (only) ROSA yesterday.
A contrastive topic is required in a context like () (Bring , , S. Tomioka
). As question is about Bill, but Bs reply is about Maxine. Maxine is therefore a

Contrast and its Relation to wa and nun

topic that is interpreted contrastively. B implies that an alternative assertion regarding

Bill cannot be made. Here, what is rejected is an utterance, rather than a proposition, and hence the reason for this rejection could be pragmatic, and not necessarily
because the proposition expressed is false. It could be that the speaker does not know
the truth value of the proposition expressed by the alternative utterance, as made
explicit in (B), or that the speaker may know the relevant fact but is unwilling to share
it. In English, contrastive topic and contrastive focus are distinguished by the different accent they bear (Bring and references therein), and they are distinguished
from their non-contrastive counterparts by their movement possibilities (Neeleman
and Vermeulen and references therein).
() A: Tell me about Bill. Did he read The Selfish Gene?
B: Well, I dont know about Bill, but Maxine read The Selfish Gene.
{Assert[Maxine read The Selfish Gene], Assert[Bill read The Selfish Gene], }
Below, I will use the above definitions and discourse contexts for identifying contrastive topics and foci.

. Mapping [contrast] and [topic]

.. [Contrast] in Korean
The following exchange shows that a contrastive topic in Korean is marked by contrastive nun, and it can remain in situ (a), or scramble to a clause-medial position
(b), or to a clause-initial position (c):
() John-i nwuku-hantey ku CD-lul cwuesse?
To whom did John give this CD ?
() Hmm, ku CD-nun molu-keyss-ko
Well, I dont know about this CD, but
In the literature, the topic in (B) is sometimes referred to as a shifting topic, and there are other
contexts in which a contrastive topic is found, such as those below:

(i) A: What did John and Mary buy?

B: John bought a book and Mary bought flowers.
(ii) A: What were the pop stars wearing?
B: The female pop stars were wearing kaftans.

(Bring : )

The topics in (iB) are sometimes referred to as list-topics. I assume that these topics are all contrastive
in the sense discussed in the main text: they all imply the rejection of an alternative utterance and they all
show the same syntactic behaviour in Japanese and Korean as discussed in the main text. The first clause
of (iB) implies that the speaker is unwilling to make the assertion that Mary bought a book, and similarly,
the second clause implies that the speaker is unwilling to make the assertion that John bought flowers.
In this case, it is most probably because the speaker knows that the propositions expressed by the salient
alternative utterances are false. (iiB) implies that the same assertion cannot be made about the male pop
stars, possibly because the speaker is not sure about the male pop stars.

a. John-i

Sue-hantey i chayk-un


John-nom Sue-to
this book-nun gave
b. John-i i chayk-un Sue-hantey t i cwuesse.
c. i chayk-un John-i Sue-hantey t i cwuesse.
as for this book, John gave it to Sue.
A contrastive focus shows the same distribution, as illustrated below. Notice that the
contrastive focus, (ku) chayk (that) book, must be case-marked and cannot be nunmarked (C. Lee , ).
() John-i Sue-eykey CD-ul cwuess-e.
John gave a CD to Sue.
() a. Ani, John-i
Sue-eykey KU CHAYK-UL/ UN cwuess-e.
no John-nom Sue-to
that book-acc/nun gave-decl
b. Ani, John-i KU CHAYK-UL/ UN i Sue-eykey ti cwuess-e.
c. Ani, KU CHAYK-UL/ UN i John-i Sue-eykey ti cwuess-e.
No, John gave that book to Sue.
() John-un Sue-eykey chayk-kwa CD-cwung enu kes-ul cwuess-ni?
Which of the book and the CD did John give to Sue?
() a. John-i
Sue-eykey CHAYK-UL/ UN cwuess-e.
John-nom Sue-to
book-acc/nun gave-decl
b. John-i CHAYK-UL/ UN i Sue-eykey ti cwuess-e.
c. CHAYK-UL/ UN i John-i Sue-eykey ti cwuess-e.
John gave the book to Sue.
Recall that the predominant view in the literature is that contrastive nun-phrases are
contrastive topics in clause-initial position, but are contrastive foci elsewhere (Choe
, Han , Choi , , Gill and Tsoulas ). The above data show
clearly, however, that this characterization is not correct on the standard definitions
of contrastive topic and contrastive focus: a contrastive topic can appear in non-initial
positions and a contrastive focus cannot be nun-marked.
Following Neeleman et al. (), I argue that [contrast] licenses scrambling of the
type seen in (), (), and (), and it is motivated by its effects at the interface between
syntax and information structure. Specifically, a contrastive item is quantificational
For some speakers, a focus corresponding to the wh-part of a simple wh-question, such as What did
John give to Sue?, can also undergo the same kind of scrambling as in () and (). The same holds for
Japanese (see ()() below). Thus, one may wonder whether there is a grammatically relevant distinction
between contrastive focus and ordinary focus in these languages and, even if there is a distinction, whether
the notion licensing the above kind of scrambling is [focus] rather than [contrast]. There are reasons to
believe that contrastive focus and ordinary focus are indeed distinct and that it is [contrast] that is the
relevant notion. First, all my informants report that a contrast between the book and another item must
be implied if scrambling takes place, as in the (b), (c) examples, but not if the book remains in situ, as

Contrast and its Relation to wa and nun

and takes scope, and scrambling has the effect of marking the sister constituent as
the scope of contrast, which Neeleman et al. call the domain of contrast. The idea is
schematized in (). I take this scope-marking operation to be an instantiation of
a mapping rule relevant for [contrast]. In other words, the information-structural
notion [contrast] that is common to contrastive topics and contrastive foci can be
marked syntactically in Korean.
() Syntax:


Information Structure:

[YP. . . ti . . . ]]
CF/CT Domain of Contrast

(CF = Contrastive Focus; CT = Contrastive Topic)

The domain of contrast (DoC) for a contrastive item contains material used to calculate the size of the set of relevant alternatives. The difference between (b) and (c) is
that the domain of contrast for the contrastive topic in the former does not contain
the subject, while the one for the latter does:
() b. [TP John-i [VP I CHAYK-UN i [VP=DoC Sue-eykey ti cwuess-e]]]
() c. [TP I CHAYK-UN i [TP=DoC John-i Sue-eykey ti cwuess-e]]
I assume that contrast is always based on an expression that is minimally a proposition
(Schwarzschild ). Thus, existential closure must apply in the case of ( b). Considering that the contrastive item is a topic, the set of alternatives will contain utterances,
yielding the set in (). For ( c), the subject is included in the set of alternative utterances, yielding the set in (). For () and (), which involve contrastive foci, the set
of alternatives will contain alternative propositions, calculated in the same manner.
() {Assert[someone has given the book to Mary], Assert[someone has given a CD
to Mary], }
() {Assert[John has given the book to Mary], Assert[John has given a CD to
Mary], }
Considering that a contrastive topic is analysed here as a composite of [contrast] and
[topic], one may wonder whether a contrastive topic in Korean could be subject to
the rule for [topic], rather than [contrast]. However, this is unlikely to be the case, as
non-contrastive topics, which share with contrastive topics the feature [topic], must
occupy a clause-initial position in this language, as widely reported (Choe , Choi
, , Han ).
in the (a) examples (see also Choi , ). Second, long-distance scrambling is known to require a
contrastive interpretation in the two languages. Indeed, my informants allow long-distance scrambling in
cases of correction and disjunctive questions, but not in answering a simple wh-question. It is not clear to
me at present why a contrastive interpretation is more easily accommodated for clause-internal scrambling,
but clearly a distinction between contrastive focus and ordinary focus is necessary to account for such data,
and it is [contrast] that licenses scrambling. See Vermeulen () for further discussion and relevant data.


Like other scope-marking operations, the effects of marking the domain of contrast
can best be observed in sentences with more than one scope-taking item. I examine
sentences that contain both contrastive focus and contrastive topic in section ..

.. [Topic] in Japanese
Contrastive topics in Japanese are also typically marked by contrastive wa, but the
topic must move to clause-initial position:
() Dare-ga Sue-ni ano CD-o ageta no?
Who gave that CD to Sue?
() Hmm, ano CD-wa doo-da-ka siranai kedo
Well, I dont know about that CD, but
a. John-ga Sue-ni ano hon-wa ageta.
John-nom Sue-to that book-wa gave
b. ??John-ga ano hon-wai Sue-ni ti ageta.
c. ano hon-wai John-ga Sue-ni ti ageta.
as for that book, John gave it to Sue.
On the other hand, contrastive focus in Japanese shows the same distribution as in
Korean. Notice that in Japanese too, a contrastive focus must be case-marked. Thus, a
contrastive wa-phrase cannot be a contrastive focus on the standard definition, contra
Endo ():
() John-wa Sue-ni CD-o ageta.
John gave a CD to Sue.
() a. Ie, John-wa Sue-ni ANO HON-O/ WA ageta.
no John-wa Sue-to that book-acc/wa gave
b. Ie, John-wa ANO HON-O/ WAi Sue-ni ti ageta.
c. Ie, ANO HON-O/ WAi John-wa Sue-ni ti ageta.
No, John gave that book to Sue.
() John-wa Sue-ni hon-to CD-to dotti-o ageta no?
Which of the book and the CD did John give to Sue?
() a. John-wa Sue-ni HON-O/ WA ageta.
John-wa Sue-to book-acc/wa gave
b. John-wa HON-O/ WAi Sue-ni ti ageta.
c. HON-O/ WAi John-wa Sue-ni ti ageta.
John gave the book to Sue.
I argue that contrastive foci in Japanese are subject to the same mapping rule for
[contrast] as proposed for Korean above (see Vermeulen for further arguments).

Contrast and its Relation to wa and nun

Contrastive topics, however, are clearly subject to a different rule. I propose that
they are subject to the mapping rule for [topic] in Japanese. It is widely reported
that non-contrastive topics typically occupy clause-initial position in Japanese too
(see Heycock for an overview), and in Vermeulen () I provide arguments
that non-contrastive topics must occupy clause-initial position. Following Neeleman
and van de Koot (), I propose that the placement of a topic in clause-initial
position is motivated by its effects at the interface, as was the case for scrambling of
contrastive topics and foci in Korean. Specifically, it marks the rest of the sentence as
the comment, allowing for a transparent mapping between syntax and information
structure, as shown below. In other words, the information-structural notion [topic]
that is common to contrastive and non-contrastive topics is syntactically marked in
Japanese. Contrastive topics show movement properties, while non-contrastive topics
show properties of being base-generated in its surface position, binding a pro (Hoji
, Saito ).

[YP XPi [YP ti /proi ]]


Information Structure:


Recall that a contrastive wa-phrase can appear in non-initial positions, as in ().

However, the data in () and () show clearly that contrastive topics are restricted to
clause-initial position. Those contrastive wa-phrases in non-initial positions therefore
cannot be topics on the standard definition of contrastive topics.
In sum, as far as the syntaxinformation structure interface is concerned, Korean
contrastive topics are subject to the rule relevant for [contrast], while those in Japanese
are subject to the rule for [topic], but contrastive foci in the two languages are subject
to the rule for [contrast]. The data discussed in this section show patterns that differ
from what is assumed in the literature. The following two sections provide further
arguments that the standard characterizations are inadequate, and offer an alternative

. Ordering Restrictions between Contrastive Topic

and Contrastive Focus
At the level of information structure, an utterance is partitioned into topic and comment, and the latter can be further partitioned into focus and background (Lambrecht
In Vermeulen (), I argue that the particle non-contrastive wa is also not a marker for sentence
topics in the sense defined in section .. It can also mark items that refer to discourse topics, such as He in
(i) (Vallduv and Engdahl ). Thus, the fact that a non-contrastive wa-phrase may sometimes occupy
a non-initial position, as in the example in (c), and in (c), is not problematic. The interested reader is
referred to the article for further discussion. See also fn. .

(i) A: Who did Max see yesterday?

B: He saw Rosa yesterday.


; see also Rizzi ), as schematized below. I assume that a comment is a proposition that is predicated of a topic via an assertion operator (Krifka ), capturing
the idea that focus is a propositional notion.
() [Utterance topic [Comment focus [Background ]]]
I have argued above that scrambling of a contrastive item marks the sister constituent
of the scrambled item as the domain of contrast, (). If focus is a propositional
notion, the domain of contrast for a contrastive focus, which is relevant for the interpretation of a contrastive focus, must also only contain propositional material. Considering that topic is an utterance-level notion, it follows that it cannot be contained
inside a proposition and hence also not inside the domain of contrast for a contrastive
focus. It is predicted then that a contrastive focus cannot move to a position above a
topic. This is illustrated in (a).
On the other hand, when a contrastive topic undergoes scrambling in Korean,
what is in the domain of contrast for this contrastive topic is material taken from
its comment, the propositional content expressed by the utterance. As such, it can
contain a focus. Thus, in Korean, we predict that a contrastive topic can move to
a position above a focus. Similarly, in Japanese, movement of a contrastive topic to
clause-initial position marks its comment (). Thus, the structure in (b) should be
well-formed in both languages.
If there is no scrambling of a contrastive topic or focus, no domain of contrast
is overtly marked. There is no mapping instruction in such cases and the ordering
between the two items is predicted to be free, as shown in (c) and (d).
() a.

[ CF [ CT tCF ]]
[ CT [ CF tCT ]]
[ CT CF ]
[ CF CT ]

These predictions allow us to test the claims made in the previous section, namely
that Korean contrastive topics may appear in a non-initial position, while contrastive
topics in Japanese are limited to clause-initial position. Moreover, if these predictions
are correct, they provide strong support for the current interface-based approach
to the syntax of contrastive focus and contrastive topic. These patterns are difficult
to capture on a cartographic approach which assumes that topics and foci move to
designated syntactic positions (Rizzi ), or that contrastive items are licensed by
moving to the specifier of a functional projection for contrast (Vallduv and Vilkuna
, Molnr and Winkler ), either overtly or covertly. We saw in section . that
a contrastive focus can undergo scrambling in Japanese and Korean. Thus, putting
aside the issue of optionality, overt movement of a contrastive focus to a designated
position must be possible in these languages. It is unclear then why (c), where
the contrastive focus has remained in situ, is well-formed, but (a), where it has

Contrast and its Relation to wa and nun

undergone movement, is not. Below, I show that these predictions are correct for
both Japanese and Korean.
The following exchange shows that the predictions in (a) and (c) are borne
out in Korean. The question in () is about John, but the reply in () is about Bill,
making Bill a contrastive topic. The presence of -man only on the object khong beans
ensures that it is a contrastive focus. In (a), the contrastive topic and contrastive
focus appear in their base-generated positions in that order, bearing out the prediction
in (c). In (b), the contrastive focus has scrambled across the contrastive topic and,
as indicated, this results in infelicity, bearing out the prediction in (a). (Henceforth,
I will provide the context-setting material in English, if similar contexts have been
used previously, for space reasons.)
() What did John eat at the party?
() Well, I dont know about John, but
a. Bill-un KHONG-MAN mekesse.
Bill-nun beans-only ate
b. KHONG-MAN i Bill-un ti mekesse.
as for Bill, he was eating only beans.

(CFi CT ti )

The following exchange shows that the predictions in (b) and (d) are also correct:
() Who ate the pasta at the party?
() Well, I dont know about the pasta, but
a. BILL-MAN khong-un mekesse.
Bill-only beans-nun ate
b. khong-uni Bill-man ti mekesse.

(CTi CF ti )

as for the beans, only Bill ate them.

I now turn to Japanese. Due to the fact that contrastive topics in Japanese must appear
in clause-initial position, it may appear that none of the predictions can be tested.
However, the prediction in (a) can be shown to be correct by examining data
involving an embedded clause. A contrastive topic can appear in an embedded clause,
What is at the core of the predictions in (a) and (b) is whether the structures are mapped to an
ill-formed or well-formed structure with respect to (). Thus, the same reasoning for (a) and (b)
should apply to structures where movement of one contrastive item does not cross the other contrastive
item, giving rise to another pair of predictions in (i).

(i) a. [ CF [ tCF CT ]]
b. [ CT [ tCT CF ]]
Due to lack of space, I cannot discuss these predictions. However, Vermeulen () shows that these
predictions are borne out for both Japanese and Korean. The fact that (ia) is ill-formed shows that the
ill-formedness of the structure in (a) cannot be reduced to a Relativized Minimality violation (Rizzi
) either, by assuming that the feature composition of a contrastive topic may be richer than that of a
contrastive focus, for instance.


as shown in (), uttered in the context in (). The context makes kono CD this CD a
contrastive topic, as it shifts the topic of discourse from the book. The presence of kare
his, which is coreferential with the matrix subject Bill, ensures that the embedded
clause is not a direct quotation.
() John finds a book on Sues desk and he asks Bill to tell him something about
the book. Bill does not know anything about the book, but he knew how Sue
obtained a CD that was also on the desk. So, he decides to tell John about the
CD. In describing this situation, you utter ().
() Billj -wa [CP kono CD-wai Mary-ga karej -no mise-de Sue-ni ti
Bill-wa [CP this CD-wa Mary-nom he-gen shop-at Sue-to
ageta to] omotteiru.
gave that thinking
Billj thinks that as for this CD, Mary gave it to Sue in hisj shop.
Independently, a contrastive focus can undergo long-distance scrambling in cases of
() Bill-wa [John-ga Sue-ni CD-o ageta to] omotteiru.
Bill thinks that John gave Sue a CD.
() Ie, ANO HON-Oi Bill-wa [John-ga Sue-ni ti ageta to] omotteiru.
no that book-acc Bill-wa John-nom Sue-to gave comp thinking
No, Bill thinks that John gave a book to Sue.
The prediction here is that it should be impossible to combine these two operations,
as this will result in the ill-formed structure in (a). The prediction is borne out: the
example in (), uttered as a correction of the statement in () is infelicitous.
() Billk thinks that Mary gave this book to Jenny in hisk shop.
() No, Bill didnt know anything about the book, but
SUE-NI i Billk -wa [CP kono CD-waj Mary-ga karek -no mise-de ti tj
Sue-to Bill-wa
this CD-wa Mary-nom he-gen shop-at
ageta to] omotteiru.
gave that thinking
Lit.: its to Sue that Billk thinks that as for this CD, Mary gave it to her in hisk
Crucially, the sentence is acceptable if the focus remains in situ, which is possible in
the same context, in accordance with the prediction in (b):
Slight unnaturalness arises here due to the repeated mention of Bill-wa in (), but this does not affect
the argument here, as the same informants found () acceptable.

Contrast and its Relation to wa and nun

() Billk -wa [CP kono CD-waj Mary-ga karek -no mise-de SUE-NI tj ageta to]
Billk thinks that as for this CD, Mary gave it to Sue in hisk shop.
The data in this section provide support for the idea that overt displacement of topics
and contrastive items in these languages has direct consequences for the mapping
between syntax and information structure. Moreover, the Korean data provide further
evidence against the standard view that a contrastive nun-phrase can function as a
contrastive topic only in clause-initial position.

. Contrastive wa- and nun-Phrases as Foci

We saw in section . that contrastive topics in Japanese and Korean are typically
marked by contrastive wa and nun, but not contrastive foci. In this section I provide
arguments that not all contrastive wa- and nun-phrases are contrastive topics. Contrastive wa and nun can also attach to a focus. Such a focus is not interpreted as a
contrastive focus in the standard sense discussed in section ., but as a focus whose
meaning is further modified by the meaning encoded by contrastive wa or nun, in
a similar fashion to a focus whose meaning is further modified by focus-sensitive
particles such as only and even.
There has recently been much work on the precise interpretation of contrastive waphrases in Japanese (Kuroda , Hara , Oshima , S. Tomioka ). Hara
(), for instance, argues that a contrastive wa-phrase induces the presupposition
that a scalar alternative stronger than the assertion of the sentence exists, and also has
the implicature that the speaker thinks the stronger alternative could be false. Let us
consider the following example. (Here and below, contrastive wa- and nun-phrases
that I argue are foci are marked by small capitals and single underlining.)
() John-ga nanninka-wa tasuketa.
John-nom some people-wa helped
John helped some people. (Implicature: John didnt help everyone.)
() has the meaning in (a). It presupposes that there is a stronger scalar alternative,
such as (b), and implicates that the speaker thinks this alternative could be false,
resulting in the implicature indicated above.
() a. (x)[[person(x)][John helped (x)]]
b. stronger scalar alternative: (x)[[person(x)][John helped (x)]]
The data considered in the literature regarding contrastive wa involve predominantly cases where the
subject bears contrastive wa. However, the same contrastive interpretation obtains when an object in situ
is marked with contrastive wa in similar contexts, and Haras analysis can be extended straightforwardly to
these cases (Yurie Hara, p.c.).


Hara extends the analysis to non-quantified DPs. In a context where there are two
relevant individuals, Mary and Bill, that John could have helped, the example in ()
has the implicature that John probably did not help Bill.
() John-ga mary-wa tasuketa.
John-nom Mary-wa helped
John helped Mary.
The stronger alternative that () induces is that John helped both Mary and Bill,
which the speaker thinks could be false. However, the speaker just asserted that John
helped Mary. The hearer can therefore infer that the intended implicature is that John
did not help Bill.
Contrastive nun in Korean has a meaning very much like that encoded by contrastive wa in Japanese (Choi , C. Lee , , , M. Lee , Hetland
). Thus, where only Yenghi and Inho are the relevant individuals that Swuni could
have met, the sentence in () implies that Swuni did not meet Yenghi.
() Swuni-ka inho-nun manna-ss-e.
Swuni-nom Inho-nun meet-past-decl
Swuni met Inho.

(Choi : , slightly modified)

Similarly, C. Lee () proposes that contrastive nun evokes a scalar implicature like
that induced by contrastive wa in Japanese on Haras proposal. The sentence in ()
implies that no more than one person came: a stronger scalar alternative is that two
people came and the implicature is that it could be false.
() han saram-un w-ass-ta.
one person-nun come-past-decl
One person came.

(C. Lee : (b), slightly modified)

There are obviously differences amongst the proposals offered in the literature. However, one common feature is that a contrastive wa- or nun-phrase generates a set
of alternatives and has a particular implicature, akin to uncertainty, regarding the
alternatives. I believe that this line of analysis provides a correct characterization of
the interpretation of contrastive wa- and nun-phrases in general. However, although
contrastive topics are typically marked by contrastive wa and nun in Japanese and
Korean respectively, nothing inherent in this interpretation makes a contrastive wa- or
nun-phrase a topic, i.e. what the rest of the sentence is about, introducing the referent
as the topic of discourse. I propose that contrastive wa- and nun-phrases can also
function as foci. Contrastive wa and nun attached to a focus has the function of further
modifying the meaning of the focus, in a similar fashion to focus-sensitive particles
such as dake/man only, and sae/to even (Kuroda , , Oshima ). Contrastive wa- and nun-phrases have prosodic properties that are identical to contrastive

Contrast and its Relation to wa and nun

foci in the respective languages (S. Tomioka , Kim , and references therein).
Thus, if they are not interpreted as contrastive topics, the most natural analysis is that
they are foci.
There are other reasons to believe that contrastive wa- and nun-phrases that are
not topics are indeed foci. First, recall that topics must be referential (section .).
The fact that quantifiers such as nanninka some people in () and han saram one
person in () can be marked by contrastive wa and nun, and receive a non-specific
reading, suggests that they are not topics. Secondly, a contrastive wa- or nun-phrase
can correspond to a wh-expression in the preceding question. Thus, the sentences in
() and () can be uttered as answers to Who did John help? and Who did Swuni
meet? respectively (Choi , Hara ). If we take this diagnostic test for focus
seriously, then one must conclude that the contrastive wa- and nun-phrases are foci.
Finally, the predictions discussed in the previous section provide syntactic evidence
that contrastive wa- and nun-phrases need not be contrastive topics. Specifically,
contrastive wa- and nun-phrases that are foci should not be subject to the restriction
for contrastive topics in (a), repeated below:
() a. [ CF [ CT tCF ]]
The prediction is correct for both languages. The Japanese utterance in () contains a
contrastive wa-phrase in situ in the embedded clause, which by virtue of its non-initial
position and non-specificity cannot be a contrastive topic. In correcting this statement, the contrastive focus Sue-o can undergo long-distance scrambling, as shown
in (). The contrast between (b) and () would be unexpected if all contrastive
wa-phrases were contrastive topics, as standardly assumed in the literature.
() Billj -wa [CP Mary-ga sukunakutomo -nin-ni-wa Jane-o karej -no mise-de
syookai-sita to] omotteiru.
Bill thinks that Mary introduced Jane to at least three people in his shop.
() a. Ie, Billj -wa [CP Mary-ga sukunakutomo -nin-ni-wa
no Bill-wa
Mary-nom at least
SUE-O karej -no mise-de syookai-sita to] omotteiru.
Sue-acc he-gen shop-at introduced that thinking

(foc-WA CF)

b. ?Ie, SUE-Oi Billj -wa [CP Mary-ga sukunakutomo

no Sue-acc Bill-wa
Mary-nom at least
-nin-ni-wa ti karej -no mise-de syookai-sita to] omotteiru.
-cl.-to-wa he-gen shop-at introduced that thinking
(CFi foc-WA ti )
No, it is Sue that Bill thinks that Mary introduced to at least three people in
his shop.


The prediction is also correct for Korean, as shown below in a similar context. The
felicity of () shows furthermore that a contrastive nun-phrase in clause-initial
position is not necessarily a contrastive topic, contra the standard description in the
() Ceketo sey-myeng-uy haksayng(tul)-un Chelswu-lul mannass-e.
At least three students saw Chelswu.
() a. Aniyo. ceketo sey-myeng-uy haksayng(tul)-un
at least -cl-gen
YENGHI-LUL mannass-e.
Yenghi-acc saw-decl
b. Aniyo. YENGHI-LULi ceketo sey-myeng-uy
Yenghi-acc at least -cl-gen
haksayng(tul)-un ti mannass-e.
No. At least three students saw Yenghi.

(foc-NUN CF)

(CFi foc-NUN ti )

I have proposed that contrastive wa and nun are not inherent markers for contrastive
topics. So why are contrastive topics typically marked by these particles in Japanese
and Korean? I argue that this is because there is a significant overlap in the implicature
induced by contrastive wa and nun and the interpretation of a contrastive topic. The
implicature induced by the particles, as discussed above, is that the speaker thinks
that a stronger scalar alternative to the proposition expressed by the sentence could
be false, and the interpretation of a contrastive topic is that the speaker is rejecting an
alternative assertion for some reason. If this is on the right track, one might expect
that contrastive topics can be marked by other means than contrastive wa or nun. This
is true at least for Japanese: they may be marked by nara (Munakata ):
() Who ate the pasta?
() Well, I dont know about the pasta, but
suupu-nara Mary-ga tabete ita yo.
soup-nara Mary-nom eating past part
as for the soup, Mary was eating (it).
Note that in English, too, the so-called B-accent can mark items that are not topics in the sense of
Reinhart (), as shown below (modified from McNally : ):

(i) A: How many people expressed interest in your house?

B: Well, [lots]B of people called, and [three]B looked at it, but [nobody]B made an offer.
In English too, therefore, it seems plausible that the B-accent only indicates a contrast of the type that is
compatible with contrastive topics and the topic status of a contrastive topic is identified by discourse (see
Molnr , Hetland , and Wagner for related ideas).

Contrast and its Relation to wa and nun

. Conclusion
I have examined the syntactic distribution of contrastive topics and contrastive foci
in Japanese and Korean, adopting the definitions independently motivated by discourse considerations for these notions, and their interaction with contrastive wa in
Japanese and contrastive nun in Korean. In doing so, I have argued that the meaning
encoded by contrastive wa and nun is compatible with both topic and focus. I have
provided a number of syntactic arguments against the standard characterizations
offered in the literature that contrastive wa-phrases are invariably contrastive topics,
and that contrastive nun-phrases are contrastive topics in clause-initial position and
contrastive foci elsewhere. The arguments were based on predictions that follow from
a particular theory of mapping at the interface (Neeleman et al. ), which does not
postulate designated functional projections for topic, focus, or contrast in the syntax.
The observed data are difficult to capture on an approach that assumes much of the
discourse is represented in the syntax, such as the cartographic approach (Rizzi ;
see also Haegeman and Hill, Chapter below).

Part III
Syntax and the Lexicon

Adjuncts Within Words
and Complex Heads

. Introduction
If one assumes that the layers of morphology mirror the hierarchy of syntax (Baker
, ) and that there is a correlation between phonological and syntactic cycles
(Embick , Marantz , Michaels , Travis , among others), then the
treatment of vowels in hiatus (i.e. VV sequences) in Ojibwe, an Eastern Algonquian
language, offers a perfect laboratory to examine different syntactic architectures that
are imprinted on the phonology. We argue that the three different possible outcomes
of vowel hiatus mirror three different configurationsall created through head movement. Piggott and Newell () previously proposed that, while VV sequences are
allowed to remain across Spell-out domains, a VV sequence within a Spell-out domain
is resolved through vowel deletion and a post-syntactic operation creates the context
where consonant insertion applies. In this chapter we investigate a particular problem
for the proposals of Piggott and Newell and solve it by allowing complex adjunction
within a word. While these adjuncts can be complex, we argue that they are not
phrasal. We propose that tools of the grammar allow the creation of complex heads in a
separate workspace that are spelled out and then merged into the structure. Crucially,
we maintain the standard position in linguistics, stated explicitly in Baker (: ),
that words cannot contain phrasal material.
.. Theoretical background
We begin with some theoretical preliminaries and some background on the relevant
aspects of Ojibwe phonology. The data is taken from a variety of sources, primarily
Baraga (), Bloomfield (), Piggott (), Piggott and Grafstein (), and
We thank the audiences at OnLI II and McSIRG (McGill Syntactic Interfaces Research Group) and
two anonymous reviewers for helpful comments, and FQRSC SE- for funding.

Piggott and Travis

Valentine (). We adopt the theory of Distributed Morphology (Halle and Marantz
), which postulates that words are generated by principles that are essentially the
same as those that generate sentences, a proposal also adopted by Baker (). These
principles manipulate abstract items, consisting of bundles of features; concrete realization is determined post-syntactically. We also adopt the theory of cyclic derivation
now referred to as derivation-by-phase (Chomsky , ). Phonetic interpretation is determined at the Spell-out of elements of the phase to the PF interface, while
meaning is assigned at the simultaneous transfer to the LF interface. Since phase-byphase interpretation is cyclic, phonological and semantic properties determined on
cycle n are transferred to cycle n + . Cyclic derivation has been a fundamental element of linguistic theory for decades. It has informed developments in phonological
theory such as SPE (Chomsky and Halle ), Lexical Phonology (Kiparsky ,
), and Stratal OT (Kiparsky , Bermdez-Otero, in preparation).
While syntactic operations feed Spell-out, Distributed Morphology assumes that
structure can be adjusted post-syntactically. Two relevant post-syntactic operations
are Lowering and Local Dislocation (Embick and Noyer ). Lowering occurs
before the operation that inserts vocabulary items and adjoins a head to the head of its
complement, while Local Dislocation applies after vocabulary insertion and adjoins a
head to an immediately adjacent head.
.. Wordhood in Ojibwe
In order to set up the problem to be investigated, we need to understand what constitutes a word in Ojibwe, and what tests can be used to determine the word domain.
We present three diagnostics that show that the prefixal material on a verb is part of a
complex word that includes the verb. The first involves stress, the second morpheme
order, and the third conjunction.
First we can see that stress assignment helps us to delimit what constitutes a word
in Ojibwe (Piggott , ). Within the domain that we identify as a word there
is only one main stress. This main stress can be located on the verbal or nominal
root, or on any of the morphological entities that precede it, except for monosyllabic
pronominal prefixes that are always word-initial.
However, stress is just part of the evidence for wordhood. We also appeal to the
fact that the elements that precede roots, but which are within the word domain, do
not display the distributional independence expected of separate words. Pre-radical
elements include tense morphemes and a subset of modifiers.
While word order in Ojibwe is relatively free, the linear order of a tense morpheme
and a verb never changes. The evidence from the occurrence of adverbs in relation to
tense affixes is also instructive. Ojibwe has several types of adverbs, e.g. temporal (ai
already, long ago; wiba soon, early; nitam at first), locational (owadi over there;
mamapi here; i+piming above, on high), manner (egad slowly; gega almost;
geskana suddenly). Comparison of (a) and (b) supports the observation that an

Adjuncts Within Words and Complex Heads

adverb can appear in different positions in a sentence without affecting the meaning.
However, (c) shows that an adverb cannot separate a tense morpheme from a verb.
() a. owadi
over there past-sit
He sat over there.
b. ginamadabi owadi
gi:-namadabi owadi
over there
He sat over there.
c. gi ( owadi)
gi: ( owadi
past ( over there sit
He sat over there.
The occurrence of the negative particle gawi not is further proof that the tense
morpheme and the verb are not treated as separate words. While the negative particle
always precedes the verb, it cannot appear between the tense morpheme and the verb
(e.g. wi( gawi)namadabisi he will not sit).
Restrictions on the mobility of adverbs also provide evidence that another preradical morpheme, the modifier, is not an autonomous word. Adverbs may occur
before or after verbs, but they never occur between a modifier and a verb.
() a. wiba giinidago+in
wi:ba gi:-ini-dagoSin
early past-away-arrive
He arrived there early.
b. giinidago+in
gi:-ini-dagoSin wi:ba
past-away-arrive early
He arrived there early.
wiba dago+in
c. giini
wi:ba dagoSin
past-away early arrive
He arrived there early.
d. gi wiba ini dago+in
gi: wi:ba ini dagoSin
past early away arrive
He arrived there early.
The negative particle cooccurs with an obligatory negative suffix /-si:/ attached to the verb.

Piggott and Travis

Additional evidence that tenseverb and modifierverb/noun combinations are complex words rather than phrases comes from restrictions on conjunction. First we can
see that a tense morpheme and a following verb are a single word through restrictions
on coordination with ellipsis. Valentine (: ) identifies a number of
different patterns of coordination in Ojibwe. When two verbs occur in a coordinate
construction, they must be associated with independent tense markers, even when the
marked tenses are identical. In the following examples, adapted from Valentine (:
), the past morpheme /gi:/ must precede each verb. The sentence is ill-formed
if the second occurrence of the past tense morpheme is missing.
() a. ni-gi-debinig
gaje (gi)-gwabiginid
-past-she hold me and past-she pull me from water
She held me and drew me out of the water.
b. mi da+ gi-giwebitojan, (gi)-windamawag
and then past-.run home past-.tell her about it
Then I ran home and told her about it.
If the Ojibwe tense morpheme were an autonomous word, we might expect that it
could be omitted from the second verb in a coordinate structure. We know that in
English a word like will can have scope over two coordinated verbs (e.g. I will arrive
and depart on the same day).
Another argument from coordination will be discussed later (section ..). We will
see that two or more preverbal/prenominal modifiers cannot be conjoined by using
the conjunction /gaje:/. The ban on conjoined elements within words seems to be a
robust, crosslinguistic condition on word formation.
.. Hiatus in Ojibwe
Having shown that Ojibwe clearly delimits wordhood, we return to the issue of hiatus
resolution within words. When morphology creates VV sequences within a word,
three phonological consequences are attested: (i) vowel deletion as shown in (a), (ii)
consonant epenthesis as shown in (b), or (iii) retention of VV as shown in (c).
The absence of the subject agreement prefix in (a) is a consequence of the fact that the second verb is
in the Conjunct mode, where subject agreement is encoded in suffixes rather than prefixes.
It has been noticed (e.g. Valentine : ) that an emphatic particle may appear immediately after
a tense morpheme or even after a modifier. It is interesting to note that Valentine considers these to be slips
of the tongue. Nevertheless, this observation does not undermine our claim that tense morphemes and
modifiers are not autonomous words, because the occurrence of emphatic particles may be prosodically
controlled, in a manner similar to the insertion of the expletive fuckin in an English word such as fanfuckin-tastic. In others words, emphatic particles are inserted at prosodically well-defined boundaries in
Ojibwe (and in other Algonquian languages). Dahlstrom () cites some Fox examples showing that
preverbs can be separated from verbs by both subject and object DPs. All of the examples come from old
sources (Michelson , Bloomfield ) and have not been replicated in any recent fieldwork that we
are aware of. These examples need to be studied more closely.

Adjuncts Within Words and Complex Heads

() a. Vowel deletion
b. Consonant epenthesis
I will (probably) walk in snowshoes.
c. Vowels in hiatus
I walked in snowshoes.
For Piggott and Newell (, ), vowel deletion indicates that the two relevant
morphemes are spelled out within the same domain. The appropriate context will
emerge when two morphemes are generated in the same phase. For example, the
two roots /gi:we:/ go home and /a:daga:/ swim combine to form a complex root
/gi:wa:daga:/ swim home, in which the hiatus is resolved by vowel deletion. This
process also applies when a lower head is adjoined to a higher head through head
movementpresumably to check a feature on the higher headand both heads are
spelled out together. The example we see in (a) shows an N that moves into a Number
head and is spelled out in the same phase as that head.
Piggott and Newell argue that consonant epenthesis occurs when a higher head
is spelled out in situ but then undergoes Local Dislocation and combines with a
previously spelled-out domain (see also Newell ). In (b) the spelled-out domain
is vP, and the material in Tense is first spelled out and then dislocated to the complex
head of vP. As argued in Embick (), Local Dislocation occurs after Vocabulary
Insertion. We propose that there is always a PF trigger for Local Dislocation; in
Ojibwe it is stress. The location of stress is determined by exhaustively parsing syllables
into feet. This language manifests a common preference for the foot to be minimally
bimoraic. The tense morphemes in (b) and (c) are spelled out in a different domain
than the verb, but the monomoraic prefix /ga/ in (b) is too small to be assigned foot
structure where it is inserted. It is forced to undergo Local Dislocation. Since both
vowels have already been sent to the phonological component at Spell-out, hiatus
cannot be resolved via vowel deletion, therefore consonant epenthesis is triggered.
In contrast, the past tense morpheme /gi:/ in (c) is bimoraic and can be parsed (as a
stress-bearing unit), explaining why there is no Local Dislocation and no consonant
epenthesis. The subject marker /ni/ is also subject to Local Dislocation and, under

Piggott and Travis

the appropriate conditions, would be expected to trigger consonant epenthesis, as

illustrated by the word /nida:gamose:/ I will walk in snowshoes, where the epenthetic
consonant separates the prefix /ni/ st person from the verb /a:gamose:/ walk in

.. The problem
A different situation occurs, however, when, instead of a selecting head such as the
tense morpheme in (b), a modifying element is merged into the structure. We see
examples of this below where /bi/ and /ini/ are preverbal ((a) and (b)) and /gitSi/ is
prenominal (c).
() Vowels in hiatus
a. nibiagamose
-towards speaker-snowshoe-walk
I walk here in snowshoes.
b. nidiniagamose
I walk there in snowshoes.
c. git+iesiban
big raccoon
The important example to note is the one in (a) where the modifier is the
monomoraic morpheme /bi/. Given the preference for a bimoraic stress unit, we
would expect this morpheme to undergo Local Dislocation after Vocabulary Insertion, and thereby trigger consonant epenthesis. However, in this example as well as
the others, we see instances where hiatus is not resolved and a VV sequence is allowed
to appear word-internally.
A further complication is that these modifiers may be complex as the example in
() below shows.

The preverbal modifier position is occupied by elements with a range of meanings. Valentine ()
asserts that the range of meanings associated with such elements include desire, obligation, direction,
quantity, quality, manner, and number. Hence, in addition to directional elements such as /bi/ here, towards
speaker, /ini/ there, away from speaker, and /biba:/ around, there are also attributives (e.g. /wa:bi/ white)
and locatives (e.g. /ipa:/ above, /nitami/ in front), manner (e.g. /gi:mo:dZi/ quietly, /ginibi/ quickly), etc.
While here and there may seem like goals, we are treating them all as modifiers.

Adjuncts Within Words and Complex Heads

() a. gimodianimitagozi
gi:mo:d-i a:nimita:gozi
quiet-fin talk
He whispers.
b. [[gi:mo:d-i ] a:nimita:gozi ]
The preverbal element in () consists of a root /gi:mo:d/ and the final which we assume
instantiates the category-defining suffix /i/. Together these form the modifier.
Taking these two complications togetherthe existence within a word of vowel
hiatus and the existence within a word of complex materialwe are forced to confront a larger question: what are the restrictions on word formation? Since Baker
(), the assumption within the Chomskyan paradigm (and Distributed Morphology: Halle and Marantz ) has been that concatenation of heads creates words.
These heads may be combined through a variety of means such as head movement
to a higher head, head lowering, and Local Dislocation (see Embick and Skinner ), but in all these cases, the word is an X which dominates only other
X and no phrasal material. Baker () encodes this restriction in the following
() Morphological well-formedness condition (Baker : )
Xn , where n is greater than
Forms consisting of a modifier plus a verb or noun such as the ones we have just seen,
however, challenge such a restrictive view of word formation. Below we first present
our view of how such forms can be constructed without giving up a restricted view
of word formation, and then we examine other accounts that allow phrasal material
within words and discuss some concerns raised by these accounts.

. The Spell-Out of Adjuncts

We propose that the difference between (b) and (a) can be explained through a
difference in the syntax of the two structures. In particular, we argue that selecting
heads are externally merged with their complements directly into the main structure
while adjuncts must be spelled out in a different workspace independently of the

Note that the same root is independently attested in the verb /gi:mo:dizi/ to sneak around.
As will be mentioned again below, more recently the theoretical viability of head movement has been
questioned. We follow Baker and others (e.g. Bobaljik , Embick , Halle and Marantz , Harley
and Noyer ) in assuming there is head movement and that complex heads spell out as words.
See Embick (: ) for a discussion of the packaging of heads into words.

Piggott and Travis

structure that they adjoin to. Further, in order to account for complex adjuncts, we
propose that complex heads can be created under certain configurations of Merge.

.. The structure of adjuncts

Architecturally, the difference between (b) and (a) is that in (b) the prefix /ga/ is
an inflectional head that selects a complement. In other words, the prefix is part of the
spine of the syntactic tree. As a selecting head, it is merged onto the existing structure
before it is sent to Spell-out along with its complement. At the time of Spell-out, then,
there is a host (i.e. the verb /a:gamose:/) in the environment of /ga/ and /ga/ is able to
locally dislocate to this host.
In (a) and (b), /bi/ and /ini/ are modifiers that are adjoined to their sisters. As
adjuncts, they are created in a separate workspace before they are adjoined to the
structure. Further, as adjuncts, they undergo Spell-out independently of the structure that they are adjoined to (see a similar analysis for XP adjuncts in Nunes and
Uriagereka ). Because of this, even though /bi/ is monomoraic, when it undergoes
Vocabulary Insertion, it is in a separate workspace and has no host to locally dislocate
to. The conditions for hiatus resolution by consonant insertion are, therefore, not
available. Note that in (b), the prefix /ni/ denoting agreement with the argument of
the verb is subsequently merged after the spelled-out modifier, /ini/, has been adjoined
to the structure. This agreement marker, however, is a selecting head. At the time
this head undergoes Vocabulary Insertion, it has a host to locally dislocate to and
consonant epenthesis is triggered between this prefix and the following adjunct.
The fact that the adjunct /bi/ is retained in situ, although it is monomoraic, seems
to engender a paradox. However, it is well documented in the linguistics literature,
especially in the OT framework, that constraints can be violated. In the present case,
the requirement of minimally bimoraic feet is enforced whenever possible. If the
conditions do not allow for construction of a bimoraic foot, a monomoraic foot is
tolerated and the derivation does not crash.
Important issues arise, however, with any analysis that would go in this direction.
While modifiers do undergo adjunction, it is usually assumed that they themselves
are maximal projections that adjoin to maximal projections and not to heads (though
see Travis ). As maximal projections in an adjoined position, it is not clear how a
modifier of a verb could become part of a word containing a verb and its inflectional
material. We can see the problem more clearly by looking at the structure in ()
below. We would not expect head movement to pick up the adverbial material as it
moved to T.

Note that this modifier cannot undergo Late Adjunction since its phonological realization must be
available to trigger consonant epenthesis.

Adjuncts Within Words and Complex Heads






So the question is how the resulting structure, a word, is able to have an adjunct
embedded within it. We turn to this problem now.
.. Complex adjuncts
In (), we saw that the modifier is adjoined to a maximal projection. Given this
structure, then, we should not be surprised to find out that these modifiers can be
complex. We saw in (a) above that the added modifier can have internal structure.
It is clear, then, that the constituency of this form can be represented as in () where
the modifier is structurally complex.





Constructions such as these have led other researchers to propose analyses where
phrasal material may appear within words, and to propose that words in languages
like Ojibwe are not formed through head movement. Below we will discuss two such
accounts, Mathieu () for Ojibwe and Wojdak () for Nuu-chah-nulth; but
first we propose a solution that maintains a restrictive definition of word. We argue
that it is possible to create complex heads in a separate workspace and then headadjoin this already complex (and spelled-out) head into the existing structure. In the
end, this gives the appearance of having word-internal phrases, but we will argue that
this analysis has both empirical and conceptual advantages.
.. An account
Our account for this construction contains three stepsall independently motivated.
The first step is to accept the traditional view of head movement that creates a headadjoined structure such as that given below through internal merge even though
this involves waiving the Extension Condition on Merge of Chomsky () (see
Matushansky and Harley for alternatives).

See Mathieu (), and Barrie and Mathieu (), for similar data and discussion.
We say the traditional view because many have claimed that there are theory-internal reasons to
want to exclude head movement as a possible mechanism of Universal Grammar (see e.g. Chomsky ,

Piggott and Travis






In the second step, we accept the proposal that head adjunction may occur by external merge as well as internal merge. It has long been accepted that the merging of
phrasal material has two possible formsexternal merge, where material is taken
from another workspace, and internal merge, where material is taken (copied) from
the structure that is being merged to. Ideally the merging of heads and the merging
of phrasal material should have parallel possibilities, and it is not clear how external
merge of a head to create a structure similar to () could be ruled out. Therefore, one
should expect to find head adjunction structures created by external merge as well,
resulting in a structure such as () below.




Here, the material adjoined to the head X is taken from the numeration and is not
a head copied from the existing structure. Naoko Tomioka (, ) posits such
adjunction structures to explain the verbal properties of resultative verb complexes in
Japanese. A relevant example is given in ().
() Kotaro-ga hi-o
K-nom fire-acc blow-extinguish-past
While it is tempting to give a translation such as Kotaro blew on the fire until it was
extinguished with huki blow as the main verb, Tomioka gives two arguments against
such an analysis. First, Japanese words are head-final, so one would expect the order to
be keshi-huki extinguish-blow if huki blow were the main verb taking the result keshi
extinguish as its complement, but this is not the correct order. Second, she shows that
huki blow cannot take hi fire as a direct object with the meaning intended in this
construction. The example below shows that the direct object of huki blow is what is
being blown out of the mouth, not what is being blown on.
Koopman and Szabolcsi ). This is still a matter of debate, however (see e.g. Lechner , Roberts
). Given that we adopt the basic assumptions of Distributed Morphology, we also are proponents of
head movement.
This is especially the case in views of head movement such as Matushansky (), where the first
merge position of the head is, in fact, a specifier position.
See also Travis () for pre-head modifiers; Mateu (), Harley (), and Haugen () for
manner conflation/lexical subordination structures.

Adjuncts Within Words and Complex Heads

() Kotaro-ga hi-o
K-nom fire-acc blow-past
Kotaro blew fire. (= blew on the fire)
Tomiokas solution is to create a structure where keshi extinguish is the main verb
taking the direct object hi fire. Hui blow is a manner adjunct that is externally
merged with the main verb. A more appropriate translation, then, would be Kotaro
extinguished the fire by blowing (his breath) and the appropriate structure is given
in ().





These structures are constructed through Pair Merge (as opposed to Set Merge),
where, in Chomskys words we might intuitively think of as attached to on a
separate plane, with retaining all of its properties on the primary plane (Chomsky
: ).
The third step of this argument would allow such complex head structures to be
created in a separate workspace. In other words, it is accepted that specifiers can be
built in a separate workspace, be sent to Spell-out, and then added to a more complex
structure as a single spelled-out phrasal unit (e.g. Nunes and Uriagereka ). We
argue that complex heads of the sort that were created by external merge in ()
above may also be built in in a separate workspace, be sent to Spell-out, and then
added to a more complex structure as a single spelled-out head unit. These complex
heads will not, however, be able to select complements because they will not be able
to have unsatisfied selectional features at Spell-out. They will, then, only be able to act
as adjuncts. It is exactly this structure that is found in the Ojibwe examples we saw in
() and (). We repeat () in () below and give the appropriate derivation in ().
The complex head of /gi:mo:d/ and /i/ is created in a separate workspace. Then this
complex head is head adjoined to the V via external merge.
() a. gimodianimitagozi
gi:mo:d-i a:nimita:gozi
quiet-fin talk
He whispers.
b. [[gi:mo:d-i ] a:nimita:gozi ]

Piggott and Travis


I. Workspace




II. Workspace









.. Consequences
We have proposed first that complex heads may be created both within a structure
via external merge and further that complex heads may be created in a separate
workspace and then externally merged with another head in the structure. We have
done this for the conceptual reason of maintaining a strict view of wordhood. One
might ask, however, if we are creating a less restrictive system by allowing merged
heads not to project to a phrasal node. In other words, when two heads are merged,
how do we know if this is Set Merge or Pair Merge? While this question requires
more careful consideration, we tentatively propose that these structures are created
only in modifier relationships. More specifically, we propose that phrasal projections
are created whenever merge involves feature checking. If selection is done via feature
satisfaction (e.g. Svenonius ), then the projection is always phrasal. When there
is an adjunct relation, no feature is checked and no phrasal projection is formed. This

has the interesting fall-out that a and little x can create a head only when the has
no arguments.

. Other Accounts
Here we briefly introduce some other accounts that also address the issue of the
complexity of these structures, and explain what we see to be their shortcomings.
Before turning to these, however, we present arguments against an alternative analysis
where what we have been calling words are phrases and what we have been calling
morphemes are in fact words.

Adjuncts Within Words and Complex Heads

.. Phrasal words
As the words that we have seen in Ojibwe are complex and contain complex constituency structures that appear quite syntactic, it is important to rule out the possibility that we are not dealing with words but, in fact, syntactic phrases, and the
morphemes themselves are, in fact, words. Two arguments that can be brought to
bear on this question.
As words, the relevant structures are similar to compounds and are therefore regulated by the same restrictions that apply to compounds. In English, for example,
the elements that make up a compound cannot be conjoined. The combinations
hot-house and green-house are both compounds, but hot-and-green house is not
possible. In Ojibwe, words and phrases can be conjoined.
() nokomis
gaje nimi+omis
gaje: ni-miSo:mis
-grandmother and -grandfather
my grandmother and my grandfather
In contrast, the conjunction of prenominal or preverbal modifiers is impossible. While
(a), (b), and (c) are well-formed, (d) is impossible.
() a. git+iesiban
big raccoon
b. madiesiban
wicked raccoon
c. git+imadiesiban
big, wicked raccoon
d. git+i gaje madi esiban
gitSi gaje: madZi e:siban
big and wicked raccoon
big and wicked raccoon
If modifiers combine with nouns and verbs to form phrases in Ojibwe, we would
expect the language to allow conjunction of modifiers, contrary to the available evidence.
We thank Lisa Selkirk for bringing this possibility to our attention.

Piggott and Travis

Now we turn to two analyses in the literature that account for word-internal complexity by allowing phrasal material to appear within words. We begin with Wojdaks
() account which allows word formation through a process called Local Spell-out.
Then we review Mathieus () account, which allows phrasal movement to create
.. Local Spell-out
Another way to account for the sort of data that we have seen in examples such as ()
is by allowing a language like Ojibwe to have rampant affixation so that all syntactic
structure is turned into words but with no head movement. Wojdak () presents
a solution of this type for Nuu-chah-nulth. She proposes that Spell-out occurs after
every case of Merge (Local Spell-out). If one of the elements being merged is a bound
element, then a word will be formed. As the structure becomes more complex, the
word becomes more complex.
One could imagine Local Spell-out applied to the structure in (), where every case
of Merge would create a word, resulting in a complex word with apparent phrases
within it. The example given below, one that she gives from Nuu-chah-nulth, comes
very close to the examples we have presented (Wojdak : ).
() camaspacuiin
camas -pa -cuq
& -in
sweets -taste -in mouth -perf -pl.imp
Let us put something sweet in our mouths.
We see Local Spell-out occurring four times in the structure below.





There are two reasons why this sort of analysis will not address the issues raised
by the Ojibwe data we are investigating. First, in Wojdaks account, it is crucial that
linearization within the word (and the affixal status) is determined by the affixation
information on at least one of the merged elements. In each of the spelled-out domains
above, there is a suffix being added. This affixal information is encoded on the lexical

An additional problem is that it is not clear to us how the syntactic structure is decided a priori. The
tree given has no labels and it is difficult to determine, without knowing the final morpheme order, what
semantic relationships are responsible for creating this structure.

Adjuncts Within Words and Complex Heads

item and does not percolate up to complex nodes. For example, we examine the
status of the constituent that contains -pa and camas. We might assume that -pa
is the head of this structure and that -pa percolates its features to the projection
that immediately dominates it. One has to assume, however, that the suffixal status
of the lexical item is not percolated to this projection since this projection does not
behave as a suffix with respect to the material that it merges with next, in this case
However, if affixal status cannot project to complex constituents, it is not clear how
two complex structures can be merged as is needed in the structure we are dealing
with repeated in ().





At the point where Advn and vP are merged, as neither one is a lexical category and
therefore cannot bear the necessary affixal information, there would be no subsequent
affixation, contrary to fact.
The second problem arises because within the system of Local Spell-out that Wojdak presents, the three responses to vowel hiatus would be left unexplained. There
should be two types of phonological juncture in her systemthose created by Local
Spell-out and those created at the edge of a saturated constituent (DP and CP). As
Piggott and Newell (, ) argue, Ojibwe clearly has phasal domains that are
not as small as every branching category nor as large as a DP. In other words, there
are phase edges within DPs, something that cannot be accounted for in any obvious
way in the system of Local Spell-out.
.. Phrasal incorporation
In a slightly different type of account, some researchers propose that phrasal material
may be part of a word, allowing for processes such as NP incorporation (e.g. Barrie and
Mathieu , Lochbihler and Mathieu , Mathieu , ). A typical example
used to argue for this analysis is given below (taken from Mathieu ). Here the
incorporated nominal is complex, including diminutive and pejorative affixes.
() ni-gii-ikwe-zhenzh-ish-iw-i
I was a naughty little girl.
While it is clear that the nominal component of the verbal complex has internal constituency, what is interesting, as pointed out by the authors, is that full

Piggott and Travis

extended projection of the nominal phrases cannot be incorporated. For example,

obviative morphology that appears on noun phrases cannot appear in incorporated
constructions. The authors propose that material high in the DP cannot incorporate.
Therefore, while words in Ojibwe can be quite complex, they cannot be as complex as
phrasal syntax.
We argue in the next section that while complex material appears to be able to
incorporate, there are limits to this complexity that have to be explained.

. The Structure of Words

While it is tempting to assume that languages like Ojibwe (and Nuu-chah-nulth) allow
phrasal material to appear within words, there are empirical and conceptual reasons to
resist coming to such a conclusion. It is important to note that not all phrasal material
may appear within a worda fact noted by both Wojdak () and Mathieu ().
For example, in Nuu-chah-nulth, definite determiners, which may appear within a
nominal phrase, may not incorporate into a complex verbal form, as the example
below shows (from Wojdak : ). The comparison between (a) and (b)
shows that an indefinite but not a definite nominal may appear within the verb stem.
The comparison between (c) and (d) shows that a bare indefinite cannot appear
on its own while a definite nominal can.
() Nuu-chah-nulth
a. ucinsiikitsi
I made a dress.
b. ucinisiikitsi
I made the dress.
c. iisiikmitsi
det-make-past-sg.ind dress
I made the dress.
d. usiikitsi
-make-past-sg.ind dress-det
I made the dress.
Wojdak () writes: I hypothesize that DP and CP differ from other projections
in that they are self-contained units in the derivation (cf. the notion of phase in

Adjuncts Within Words and Complex Heads

Chomsky ()). The borders of DPs and CPs therefore constitute edges. Mathieu
() observes a similar restriction in Ojibwe:
Let me point out that, although my claim is that Ojibwe INs [incorporated nouns] are phrasal,
this does not mean that they are necessarily fully phrasal. . . . I want to argue that the very high
functional tier of a nominal phrase in Ojibwe does not get incorporated.

The dilemma, then, is the following. Does one allow for phrasal incorporation and
then seek to explain why not all phrases incorporate, or does one restrict word formation to heads and explain the apparent complexity of some elements within words
in a different manner? We choose the latter route for several reasons. First, there is
an account in the literature that would rule out head-movement in an example such
as (b). Li (), in an effort to explain why head movement does not move lexical
categories through functional categories back to lexical categories, proposes a notion
of improper head movement (much like improper XP movement from an A-position
to an A -position, back to an A-position). Whatever the analysis, the observation is
an important one. Head movement never contains C material between two Vs in,
for instance, productive causative constructions. We should not be surprised, then,
that D material is not found between N and V. When the restrictions on phrasal
incorporation begin to resemble restrictions on head movement, one must suspect
a missed generalization, and we assume that the lack of the higher DP material in
Ojibwe Noun Incorporation structures can be explained through this already existing
restriction on head movement.
Another reason not to accept the idea that only higher material in the phrase
structure is prohibited from appearing within a complex word comes from different
bits of data we have seen in this chapter. In section .. we saw free-standing adverbs
that have variable positions in the clause. In the body of the chapter, we have been
accounting for other adverbs that must appear within the verbal stem. It appears to be
a lexical characteristic whether an adverb belongs to the first or second group, and it
has nothing to do with structure. In other words, high adverbs (/Za:Zi/ already, long
ago) as well as low adverbs (/e:ga:dZ/ slowly) appear independently, and high adverbs
(/mat(w)e:-/ used when the event occurs at a distance from the speaker) as well as low
adverbs (/gi:mo:dZi/ quietly) appear within the verbal complex. Within our account,
adverbs would be lexically specified as to whether they can appear merged with the
category-defining head without projecting. In the Local Spell-out account, it seems
that the adverbs that attach to the verb would have to be merged low, and those that
appear independently would have to be merged high even though there would be no
semantic support for this distinction.
This discussion raises the question of what sort of incorporated structures our
account would rule out. We have claimed that base-generated complex heads cannot
Languages will also specify what is possible. Japanese allows external merge of heads for manner
specification (extinguish by blowing) while English does not.

Piggott and Travis

contain argument structure. Any selection process will entail feature satisfaction,
and any feature satisfaction will entail projection to a phrasal constituent. While

proud-a might be a possible adjunct, proud-a-of-his-son, where the adjective has a

complement, would not be. We also would not expect noun incorporation to include
a full possessor.

. Conclusion
The interface between syntax and phonology highlights particular issues concerning
what words can consist of. To summarize the puzzle, it appears that phrase-like
material can appear within words in Ojibwethere are phonological breaks and the
material can be complex. This suggests that the restriction of what can appear within a
word (and what can incorporate) has to be loosened in some languages. However, we
have argued that there are empirical and conceptual reasons to resist such a conclusion. In both the Local Spell-out account of Wojdak () and the NP incorporation
account Mathieu (), something has to be said about the limits of what elements
can undergo these processes. We argue that these restrictions on word formation
follow from restrictions on head movement. The complexity observed within Ojibwe
words is not a case of phrasal material within a word but rather, complex adjunct
Of course arguments and argument-taking heads can be incorporated into higher heads through
head movement. For example, the head N of an argument can be incorporated into a V, or an argument
taking P can be incorporated into a V leaving behind its complement (see Baker ).
In Mathieu (, ), there is an example with a possessor given below.

(i) obezhgoogzhiimi
He/she has a horse.
It is noteworthy, however, that this construction contains a denominal verb and the overall meaning is
one of possession. This raises questions as to what the accurate structural analysis of this construction
would be, and whether one can have a possessive within an incorporated noun with a full verb. Again,
Dahlstrom () gives one Fox example from Michelson () where a possessed noun seems to have
been incorporated. More work needs to be done to understand the full range of this phenomenon.

Still Puzzled by Adjectival Passives?

. Introduction
The received view for languages like English and German is that in adjectival passives
the external argument of the underlying verb is not syntactically active, based on its
lack of control into purpose clauses and the absence of the disjoint reference effect.
In both respects, verbal passives behave differently, leading to the conclusion that the
external argument is present in the syntax, even in the absence of by-phrases (cf.
Baker, Johnson, and Roberts , Kratzer , among others). This is illustrated
by German, in which a morphological difference is made between adjectival passives,
which combine a past participle with sein be ((a), (a)), and verbal passives, which
appear with werden become ((b), (b)).
() a. Der Reifen war aufgepumpt, um
die Fahrt fortzusetzen.
the tyre was inflated
in order the journey to continue
intended: The tyre was inflated in order to continue the journey.
b. Der Reifen wurde aufgepumpt, um
die Fahrt fortzusetzen.
the tyre became inflated
in order the journey to continue
The tyre was (being) inflated in order to continue the journey.
() a. Das Kind war schlampig gekmmt.
the child was slopp(il)y combed
The child was combed in a sloppy manner.
(i) Disjoint reference: Someone (else) (has) combed the child.
(ii) Reflexive: The child (has) combed him/herself.
For comments on earlier versions of this chapter, I thank Artemis Alexiadou, Boban Arsenijevic,
Edit Doron, Scott Grimm, Louise McNally, Antje Rodeutscher, Florian Schfer, Rob Truswell, and two
anonymous reviewers. This work was supported by grant JCI-- from the Spanish Ministry of
Science and Innovation.
This does not hold universally. Anagnostopoulou (a), for instance, argues that one type of adjectival passive in Greek (the -menos-type) has agentive properties that the German one lacks, including implicit
agents that can control into purpose clauses; this is further refined in Anagnostopoulou and Samioti (Ch.

b. Das Kind wurde schlampig gekmmt.
the child became slopp(il)y combed
The child has been combed in a sloppy manner.
Only disjoint reference: Someone (else) (has) combed the child.

These observations, combined with the common assumption that by-phrases syntactically and semantically express external arguments, lead to the prediction that byphrases in English and German should only be possible with verbal passives. The
puzzling fact, then, is that the literature reports many instances of German adjectival
passives combining with phrases headed by von by (e.g. (): see Kratzer , ,
Rapp , Schlcker , Maienborn a, , Gese , Gehrke , and
also McIntyre , Bruening, to appear, for data from English).
() a. Die Zeichnung ist von einem Kind angefertigt.
the drawing is by a
child produced
The drawing is produced by a child.
b. Das Haus ist von Studenten bewohnt.
the house is by students inhabited
The house is inhabited by students.
The topic of this chapter is how to deal with by-phrases with German adjectival
passives. There are at least three possible ways to analyse them. First, one could argue
that combinations of sein be + participle that allow by-phrases are not adjectival
passives. On one such account the participles could be analysed as genuine adjectives
that combine with argumental PPs, similar to e.g. (a), and the by-phrase could be
treated on a par with by-phrases we find with genuine nouns, such as those in (b).
() a. (un)glcklich ber diese Entwicklung
(un)happy about this development
(un)happy about this development
b. ein Buch von Tanya Reinhart
a book by Tanya Reinhart
a book by Tanya Reinhart
A problem for such an approach is that it ignores the fact that the adjectives involved
are clearly participles morphologically derived from some underlying verb. I will
argue here that, although some by-phrases with adjectival passives can indeed be
analysed as adjectival modifiers, modifying the state associated with the adjectivized
participle directly (e.g. (b)), by-phrases of the type in (a) clearly modify an event
and cannot therefore be analysed as adjectival (state) modifiers or arguments.
A second theoretical option is that the previous literature had it wrong, and adjectival passives do contain external arguments. Bruening (to appear), for instance,
argues (for English) that the lack of the disjoint reference effect is an illusion that

Still Puzzled by Adjectival Passives?

only appears with (potentially) reflexive predicates but not with other ones (see also
McIntyre ), and that, more generally, there are no differences between verbal and
adjectival passives with respect to external arguments: both contain passive Voice
heads that license this argument. This cannot be the whole story, since it does not
straightforwardly account for two facts to be discussed in this chapter: (i) not all byphrases that are possible with verbal passives are possible with adjectival passives, and
(ii) the differences between (apparent) external arguments with adjectival and verbal
passives go beyond purpose clauses and the disjoint reference effect.
The third theoretical option, then, which will be pursued here, is to maintain that
these constructions are adjectival passives, which have implicit external arguments
(in line with Bruening, to appear, McIntyre ) but whose by-phrases are different
from those with verbal passives (contra Bruening, to appear).
The overall topic of this chapter touches on several interface issues. First, there are
well-known issues concerning the lexiconsyntax interface. It is generally assumed
that the participles in adjectival passives are ultimately adjectives, but there is disagreement as to whether such adjectives are formed in the lexicon (as implicitly
assumed for German by Maienborn a and subsequent work) or in the syntax
(e.g. Kratzer ). Here, I will take a syntactic position and treat adjectival passives
as true passives, which contain verbal structure associated with eventivity and which
undergo adjectivization at some point of the derivation. Thus, even though Bruenings
(to appear) account will be argued to be incomplete, his overall conclusion that word
formation in the case of adjectival passives is syntactic in nature also underlies the
current proposal.
The main contribution of this chapter, however, lies at the syntaxsemantics interface. In particular, I will be concerned with the question what it means to have verbal
structure within adjectives and what the consequences are for the nature of the underlying event. My answer to this question will be that as a consequence of the adjectival
categorization of verbal material, the event associated with the verbal predicate does
not get instantiated but instead remains in the kind domain. I will propose that it
follows from this that event participants also remain in the kind domain and do not
get instantiated.
The chapter is structured as follows. Section . discusses the difference between
by-phrases with verbal and adjectival passives, as well as event-related modification
with adjectival passives in general, and establishes that even with adjectival passives
there are two kinds of by-phrase, event- and state-modifying ones. Section . out There are also cases where (English) adjectival passives allow for purpose clauses, as shown by McIntyre
() (see also Alexiadou, Gehrke, and Schfer for similar examples from German). In a footnote,
Bruening dismisses purpose clauses as a diagnostic for external arguments, but does not further address
why they are more degraded with adjectival passives.
As we will see, adjectival participles touch upon another issue concerning word formation, namely
whether or not words can contain phrasal material. For reasons of space, I cannot address this issue in
sufficient detail or from a more global perspective, but see Piggott and Travis (Ch. above) for discussion.


lines the general proposal, according to which adjectival passives involve event kinds
rather than event tokens; event-related modifiers will be argued to modify this kind,
whereas state modifiers modify a state token. Section . concludes.

. Event-Related Modification
I will argue that particular restrictions on event-related modification with adjectival
passives can be accounted for by assuming that such modifiers modify an event kind
rather than an event token. I will first show that (event-related) by-phrases with
adjectival passives are crucially different from those with verbal passives. After that
I will outline the general restrictions on event-related modification with adjectival
passives, and in the final subsection I will distinguish two types of by-phrase with
adjectival passives, event- and state-related ones.
.. Different by-phrases with adjectival and verbal passives
Whereas verbal passives can combine with all kinds of by-phrase (e.g. ()), the availability of such phrases is severely restricted with adjectival passives, as has been
repeatedly noted in the literature (the examples in () are discussed in Rapp
and Maienborn ).
() a. Der Mlleimer wird
(von meiner Nichte) geleert.
the rubbish bin becomes (by my
niece emptied
The rubbish bin is being emptied by my niece.
b. Die Tr wird
(von ihm) geffnet.
the door becomes (by him opened
The door is being opened by him.
() a. Der Mlleimer ist ( von meiner Nichte) geleert.
niece emptied
the rubbish bin is ( by my

The rubbish bin is emptied ( by my niece).

b. Die Tr ist ( von ihm) geffnet.
the door is ( by him opened
The door is opened ( by him).
The restrictions are commonly attributed to the assumption that only those eventrelated modifiers are allowed that relate to event participants that still have an impact
on or are still visible in the consequent state, but a detailed investigation of the
properties of acceptable vs. unacceptable modifiers is rarely found. In the following,
Maienborn (a, ) discusses event-related modification in more detail, but the details of her
account, according to which such modifiers are licensed pragmatically, are not fully spelled out, nor does
she systematically investigate the properties of acceptable vs. unacceptable modifiers or provide a positive
characterization of the acceptable ones.

Still Puzzled by Adjectival Passives?

I will compare the by-phrases with the two passives to work out the differences and
understand the nature of the restrictions.
A first observation is that the complements of the by-phrases that are unacceptable
with adjectival passives (e.g. ()) are definite noun phrases that refer to a particular
entity in the discourse, whereas those in the acceptable ones such as () are indefinite
NPs or bare nouns. Second, only with verbal passives does an indefinite noun phrase
in a by-phrase introduce a discourse referent, which can be picked up by pronominal
anaphora in the subsequent sentence (a), not with adjectival passives (b).
() a. Die Zeichnung wurde von einem Kind angefertigt. Es hatte rote Haare.
the drawing became by a
child produced it had red hairs
The drawing has been produced by a child. He/she had red hair.
b. Die Zeichnung ist/war von einem Kind angefertigt. Es hatte rote Haare.
the drawing is/was by a
child produced it had red hairs
The drawing is/was produced by a child. He/she had red hair.
Third, only the complements of by-phrases with verbal participles (with werden
become) but not with adjectival ones (with sein be) can be modified by common
intersective modification, such as adjectival (token) modification (a) or restrictive
relative clauses (b).
() a. Die Zeichnung {wurde/ ist/ war} von einem blonden Kind angefertigt.
blond child produced
the drawing {became/is/was by a
The drawing {has been/is/was} produced by a blond child.
b. Die Zeichnung {wurde/ ist/ war} von einem Kind angefertigt, das ich
the drawing {became/is/was by a
child produced which I
im Kindergarten getroffen habe.
in the kindergarten met
The drawing {has been/is/was} produced by a child that I met at the kindergarten.
Finally, even in the presence of a by-phrase there is no control into purpose clauses in
adjectival passives, unlike what we find in verbal passives:
() Die Zeichnung {wurde/ ist/ war} von einem Kind angefertigt, um
child produced in order its.dat
the drawing {became/is/was by a
eine Freude zu machen.
parents.dat a happiness to make
The drawing {has been/is/was} produced by a child in order to make his/her
parents happy.
Kind modification is possible (e.g. red ink in (a)), which follows from the general proposal underlying this chapter.


In sum, the complements of by-phrases that appear with adjectival passives in German, unlike those with verbal passives, are non-referential and have a rather generic
flavour. In the following section, we will see that this holds quite generally for eventrelated modifiers of adjectival passives that name a participant of the underlying event.
.. Event-related modifiers modify an event kind
Besides by-phrases, German adjectival passives also combine with other kinds of
modifier that are commonly assumed to rely on the presence of an agent, such as
manner adverbs (e.g. (a)) and instrument phrases (e.g. ()).
() a. Der Brief ist mit roter Tinte / (einem) Bleistift geschrieben.
the letter is with red ink / (a
pencil written
The letter is written with red ink / (a) pencil.
b. Der Ordner war mit einem Passwort gesichert.
the folder was with a
password secured
The folder was password-protected.
Again, not all modifiers are acceptable. For example, replacing the noun phrases in
the modifiers in () by definite noun phrases would lead to ungrammaticality. Some
more unacceptable examples are given in () (after Rapp ).
() Der Mlleimer ist ( langsam / gensslich / mit der Heugabel) geleert.
the rubbish bin is (slowly / pleasurably / with the pitchfork emptied
The rubbish bin is emptied ( slowly / with pleasure / with the pitchfork).
The more restricted availability of such modifiers, again, has been attributed to a
general ban on non-state-related modifiers. For example, Kratzer () notes that
(a) is only acceptable in a context in which the hair also looks sloppily combed,
but not in a context in which the combing of the hair was sloppy but nevertheless
resulted in nice-looking hair. In this context, Anagnostopoulou (a) differentiates
between two kinds of manner modifier: result-oriented ones like those in (a), which
are acceptable with German adjectival passives, and agent-oriented ones like those in
(), which are not. She proposes that the latter modifiers attach somewhere within
the VP (lower), and the former to VoiceP, which she argues to be absent in German
adjectival passives.
Disregarding manner modifiers for the time being, we see that the acceptable withphrases in ()very much like the acceptable by-phrases abovecombine with
indefinite or bare nouns, whereas the unacceptable one in () combines with a definite one. Furthermore, just like the complements of by-phrases with adjectival passives, () shows that the complements of with-phrases with adjectival passives do not
introduce discourse referents and cannot be modified (by token-related modifiers).
See also Anagnostopoulou and Samioti (Ch. below) for an application of this idea to Greek.

Still Puzzled by Adjectival Passives?

() a. Die Karte ist/war mit einem Bleistift geschrieben. Er war blau.

the card is/was with a
pencil written
he was blue
The card is/was written with a pencil. It was blue
b. Die Karte ist mit einem kurzen Bleistift geschrieben.
the card is with a
short pencil written
intended: The card is written with a short pencil.
A second restriction, the generality of which has gone practically unnoticed, is that
the event in adjectival passives lacks spatiotemporal location (see Gehrke for a
more detailed discussion). It has been observed that adjectival passives are incompatible with temporal frame adverbials (() is from von Stechow ; see also Rapp


Computer ist vor drei Tagen repariert.

the computer is before three days repaired
intended: The computer is repaired three days ago.

The ban on temporal modification of the underlying event is more general, though.
For example, a modifier like recently cannot modify the underlying event but only the
state denoted by the construction. In particular, () can only mean that the door was
in the opened state recently (and probably is no longer), but not that the door is in the
opened state, with the opening having taken place recently.
() Die Tr war krzlich geffnet.
the door was recently opened.
The door was opened recently.
What has not been noted is that spatial modifiers that pick up the location of the event
that brought about the consequent state are also generally bad.
() Die Reifen sind in der Garage aufgepumpt.
the tyres are in the garage inflated
intended: The tyres are inflated in the garage.
Based on these facts, I concluded in Gehrke () that the underlying event in
adjectival passives is not instantiated, and proposed that adjectival passives involve

Embick () claims for English adjectival passives that they give rise to an ambiguity with recently,
in allowing the second reading that is impossible in (). However, he shows this only with attributive
participles (e.g. the recently opened door), which actually display the same ambiguity in German. Rapp
() convincingly argues for German that such attributive participles should be analysed as reduced
relatives involving verbal rather than adjectival passives; it is possible that this analysis extends to English.
These examples are acceptable if the spatial frame creates some kind of temporal frame for the state
itself (see Maienborn b for further discussion). Under this reading, however, the PP clearly modifies
the consequent state, not the underlying event.


event kinds rather than tokens. More specifically, adjectival passives instantiate the
consequent state kind of an event kind.
This analysis was further supported by the observation that the restrictions on
event-related modification of adjectival passives match the restrictions on possible
antecedents of adverbial so, as described in Landman and Morzycki (), who treat
so as an event kind anaphora (in analogy to the nominal kind anaphora such; cf.
Carlson a). In particular, spatial and temporal modifiers, which modify an event
token, cannot antecede so, and they are also unacceptable with adjectival passives.
Furthermore, manner modifiers, which Landman and Morzycki propose to treat as
event kind modifiers, are acceptable in both, as they name or establish subkinds.
Finally, instruments, when they are acceptable, also name a subkind and can thus be
treated as manner in a broad sense, as the following example (from the Frankfurter
Rundschau corpus) indicates:
() Ihr weigetnchter Krper ist mit Binden geschnrt.
her whitewashed body is with bandages strapped
Her whitewashed body is strapped with bandages.
The with-phrase here does not name the actual instrument of an actual event that
brought about the state the subject is in, but rather makes the kind of state that the
subject is in more concrete: it is of the kind that would result from a strapping-withbandages-event kind, which in turn is a subkind of strapping event.
We can now make sense of the two-way distinction within manner adverbs proposed by Anagnostopoulou (a): an adverb that relates to the degree to which
an object is in a particular state associated with the consequent state of an event
kind should always be acceptable with adjectival passives. This is arguably the case
in (a). If the modifier relates to the event itself, on the other hand, it names a
particular (sub)kind of event, under the current proposal. We can assume, then, that
only those modifiers are allowed that name established event kinds with identifiable
consequent states. With the unacceptable modifiers in () and (), however, there are
no established event kinds: there is no established event kind of emptying something
slowly or with pleasure, as much as there is no established event kind of emptying
with a pitchfork or by my niece.
Finally, I want to propose that a consequence of the idea that the event remains
in the kind domain and is not instantiated is that implicit event participants, such
as agents or instruments, are not instantiated either and equally remain in the kind
domain. From this, in turn, their more generic character and their inability to introduce discourse referents follow.
Similar restrictions have been observed on kind modification in the nominal domain, which is only
possible if the modified noun refers to a well-established kind (first by Carlson a):

(i) The {Coke / ??green} bottle has a narrow neck.

Still Puzzled by Adjectival Passives?

In sum, state modifiers (such as degree adverbs and temporal modifiers of the
state) are generally available with adjectival passives. Event-related modification, on
the other hand (manner adverbs, instruments), is only available if it modifies an
event kind and names an established subkind. In the next section we will see that
a distinction between state (token) and event kind modification similar to the one we
observed for adverbs is found with by-phrases with adjectival passives. In particular,
there is one type of by-phrase that behaves like the other event-related modifiers in
naming a subkind, whereas another type modifies the state directly.
.. By-phrases as event kind vs. state token modifiers
A closer look at by-phrases with adjectival passives reveals the need to distinguish
two types (see also Gehrke ). Rapp () shows that one type is compatible
with participles that are prefixed by the adjectival negative prefix un-, whereas another
type is not, and that the two types display word order differences. Schlcker ()
independently observes differences in prosody and complementation with two types
of by-phrase. Let us take a closer look at these facts.
The first type of by-phrase (by-phrase I) forms a prosodic unit with the participle,
with neutral stress on the modifier (a). Such by-phrases are incompatible with unprefixation (b) and cannot switch order with the participle (c).
() a. Die Zeichnung ist von einem KIND angefertigt.
the drawing is by a
child made
The drawing is made by a child.

(Schlcker )

( von

Maja) ungewrzt.
b. Die Suppe ist
the soup is ( by Maja unseasoned
The soup is unseasoned ( by Maja).

(Rapp )

{ geschrieben} von

c. Der Brief war

einem Experten {geschrieben}.
the letter was { written
by an
expert {written
The letter was written by an expert.
(Rapp )
In addition, Schlcker () observes that the complements have a more generic
character (e.g. von Feuer by fire, von Bomben by bombs).
Rapps and Schlckers characterizations of these by-phrases extend to other eventrelated modifiers that are allowed with adjectival passives, as illustrated in ().
() a. Der Brief war mit (einem) BLEIstift geschrieben.
the letter was with (a
pencil written
The letter was written with (a) pencil.
One might wonder whether with-phrases can be of two types as well, then. I cannot come up with
examples of state-modifying with-phrases, however, so I assume that such phrases always modify events.

b. Der Brief war mit (einem) Bleistift ungeschrieben.
the letter was with (a
pencil unwritten
intended: The letter was unwritten with (a) pencil.
c. Der Brief war { geschrieben} mit (einem) Bleistift {geschrieben}.
with (a
pencil {written
the letter was { written
The letter was written with (a) pencil.

The second type of by-phrase (by-phrase II) does not form a prosodic unit with the
participle: neutral stress is on the participle, secondary stress on the modifier (a).
Such by-phrases are compatible with un-prefixed participles (b) and can switch
order with the participle (c).
() a. Er ist von der MuSK beINdruckt.
he is by the music impressed
He is impressed by the music.

(after Schlcker )

b. Die Dresdner Brger sind von solchen Problemen unbeeindruckt.

the Dresden citizens are by such problems unimpressed
The citizens of Dresden are not concerned with such problems.
(Rapp )
c. Die Dresdner Brger sind {(un)beeindruckt} von solchen Problemen
the Dresden citizens are {(un)impressed by such problems
The citizens of Dresden are (not) concerned with such problems.
(after Rapp )
Finally, the complements have a more concrete character (e.g. vom Feuer by the
fire, von der Bombe by the bomb, Schlcker ). In all these respects they behave
differently from both by-phrase I (compare ()) and other event-related modifiers,
as in ().
Inspired by Rapps () observation that by-phrase I is action/process-related,
whereas by-phrase II appears with stative predicates, I proposed in Gehrke ()
that the former modifies the event kind and names a subkind; it thus behaves like
other event-related modifiers of adjectival passives in all relevant respects. By-phrase
II, in turn, modifies the (result) state token. Such by-phrases are fully acceptable only
with stative predicates, and sharply contrast with other event-related modifiers in all
The proposal that by-phrase II modifies a state token, rather than an event kind, is
further supported by the fact that it behaves more like by-phrases with verbal passives
in some respects, and differently from by-phrase I (recall ()()). In particular, the
noun in a by-phrase II introduces a discourse referent (a) and can be modified

Still Puzzled by Adjectival Passives?

() a. Er ist von einer Melodie beeindruckt. Er hatte sie gestern im Radio

he is by a
melody impressed he had her yesterday in the radio
He is impressed by a melody. He heard it yesterday on the radio.
b. Er ist von einer Melodie beeindruckt, die
er gestern im Radio
he is by a
melody impressed which he yesterday in the radio
gehrt hat.
heard has
He is impressed by a melody that he heard on the radio yesterday.
To provide a syntactic explanation for the facts regarding prosody, word order, and
un-prefixation, I will first outline the main proposal. This proposal will depart from
Schlcker (), who treats by-phrase I as a V-adjunct (following Maienborn
a) and by-phrase II as a VP-adjunct (in the sense of Kratzer ). One reason
to depart from this account is that the modifiers under discussion are phrases, and
phrases adjoin to phrases (e.g. VPs), not to heads (such as Vs). Hence, both types
of by-phrase should adjoin at least to a VP, or to an even larger structure. Second,
Schlcker overlooked the crucial division between eventive and stative predicates,
i.e. the fact that by-phrase II appears only with stative predicates. Nevertheless, her
(informal) observation about the different nature of the complements fits the proposal
underlying this chapter. With these observations in hand, let us now turn to the

. The Proposal
The previous discussion established three things. First, by-phrases (of type I) with
adjectival passives are crucially different from by-phrases with verbal passives: the
external argument with the former does not introduce a discourse referent, from
which I concluded that it is not instantiated but remains in the kind domain. Second, adjectival passives involve event kinds rather than event tokens; event-related
modifiers apply to this kind. Third, adjectival passives can combine with two different
kinds of by-phrase, one which modifies an event kind and one which modifies a state
token (only with stative predicates).
Piggott and Travis (Ch. above) discuss apparent phrasal adjuncts within words in Ojibwe and instead
analyse them as complex adjunct heads that are formed in a separate workspace and participate in word
formation via head movement. They furthermore argue that restrictions on such modification follow from
restrictions on head movement. If we adopt this proposal, the modifiers under discussion could be seen as
V-adjuncts. A potential problem, though, is that they can contain material that makes them appear more
phrasal than the Ojibwe adjuncts discussed, such as prepositions and adjectival kind modification (recall
red ink in (a)). Hence, I will continue to treat them as phrases modifying phrases (along the lines of
Kratzer ), and to derive the restrictions from the kind semantics.


There is one more thing we have to take into account, which is not discussed in
detail here but rather taken for granted. An adjectival passive commonly ascribes a
stative property to an object, which is the theme of the underlying verb; this stative
property is associated with the result or consequent state of the underlying verb, and
it is commonly assumed that the input has to be a resultative or telic verb, which
could be represented in various ways (see Gehrke for more details). Since the
precise implementation is not relevant at this point, I will simply follow McIntyre
() in representing this by a cause-relation and in labelling the agent or causer
At this point, let us return to the question whether adjectival passives contain a
passive Voice layer, as suggested by Bruening (to appear). Essentially, we need to
account for the differences between the two passives with respect to the (un)restricted
availability of by-phrases and their (in)ability to introduce discourse referents. This
could be captured in two ways. First, we could argue against Bruening (to appear)
and follow the more traditional view that there is no VoiceP in adjectival passives (as
advocated by Kratzer , ). By-phrases with adjectival passives would then be
licensed by the causative semantics of the underlying verb.
Alternatively, we could follow Bruening (to appear) in assuming (something like)
passive Voice heads for both passives, but amend this syntactic account with a semantic account of the differences between the two (as suggested in Alexiadou, Gehrke,
and Schfer ). In the remainder of this chapter, I will spell out such a semantic
account and propose to place the differences in a separate head, solely responsible
for adjectivization. Leaning more towards the second option, then, I will actually not
commit to a separate VoiceP but will represent the Initiator argument as part of the
lexical semantics of the verb. Let us turn to the actual proposal.
.. The composition of adjectival passives
I propose that the derivation of an adjectival passive like the one in () is as in ().
() Die Tr ist geschlossen.
the door is closed
() a. Lexical semantics of schlie- (type e, e, s, ev, t):
yxse[close(e) cause(s)(e) closed(y, s) Initiator(x, e)]
b. Part ge-en/t: ysex[ (e)(s)(x)(y)]
c. A : ysek , xk [ (ek )(s)(xk )(y)]
d. [A geschlossen]:
ysek , xk [close(ek ) cause(s)(ek ) closed(y, s) Initiator(xk , ek )]
e. [AP die Tr geschlossen]:
sek , xk [close(ek ) cause(s)(ek ) closed(the door, s)
Initiator(xk , ek )]

Still Puzzled by Adjectival Passives?

The lexical semantics of a resultative verb, in this case schlie(en) (to) close, is given in
(a): there is an event e, which brings about a state s, with the holder of the state (the
theme) being y and the Initiator of the event being x. Passive participle formation
in (b) existentially quantifies over the external (Initiator) argument.
Following Lieber (), it is generally assumed that the participle is turned into
an adjective by zero-affixation (e.g. Kratzer , Maienborn a):
() cop [AP [A [geffnet] ]]
The verbs event variable has to be existentially bound (as a kind) at some stage
of the derivation, and this could happen either when the participle is formed or
when the participle is adjectivized. I propose that the event variable is bound as a
kind when the participle is adjectivized, i.e. when the category is changed or specified
(depending on the precise implementation). This allows a uniform analysis of passive
participles, whether they are used in verbal or adjectival passives.
Hence, at the stage of participle formation (b) both verbal and adjectival passive
participles are still identical. What makes adjectival passives different from verbal
passives is the adjectivization of the participle. At this point, it is also determined
that the implicit Agent of the event stays in the kind domain (represented as xk ). The
semantics of the adjectival head and the output of adjectivization are given in (c)
and (d), respectively. The internal argument of the underlying verb, in turn, is the
subject of the AP and thus has to be externalized at some point. McIntyre () and
Bruening (to appear) propose different versions of -abstraction of the verbs object at
the point of adjectivization, and in principle either will do for our purposes. I will gloss
over the details here and assume that the output of adjectivization after externalization
of the verbs internal argument is as in (e).
From this analysis it follows that an adjectival passive refers to the instantiation of
a result or consequent state kind of an event kind. The state is instantiated at some
later stage and temporally located, so that temporal modifiers can access the states
temporal index.
Event-related modifiers (including by-phrase I) have to apply before adjectivization, since they modify an event, not a state, and since the event is closed off and
not accessible any more after adjectivization. From this it follows that adjectivization
I assume that e (and likewise s), without a subscript, ranges over both event (state) kinds and tokens,
and that the extension of any event predicate includes event kinds and event tokens.
Note that McIntyre () and Bruening (to appear) collapse the tasks I distribute over Part (passivization, existential quantification over the external argument) and Adj (stativization, externalization
of the internal argument) into one passive Voice head, which they label Part and Adj, respectively. The
dissociation of passivization and adjectivization is motivated by the idea that at the point of participle formation both verbal and adjectival participles are alike, and more importantly by the need for an attachment
site for event-related modification before adjectivization, which renders the event inaccessible for further
modification (see below).
Some operator binds the state variable and asserts that the state takes place at an instant, which is part
of an interval (see e.g. Truswell for a proposal of such an operator).


can also target phrases (following Kratzer , ), and the relevant phrase in this
case is PartP. Event-related by-phrases, then, adjoin to PartP (the same holds for byphrases with verbal passives, since at this point both passives are the same). Adverbial
modification in general, then, either applies to the event kind (before (c)) or to the
state token (after adjectivization).

.. The syntax of adjectival passive modification

Let us now turn to syntactic assumptions about event-related modifiers with adjectival
passives, and in particular to some of the differences we observed with the two types of
by-phrase with respect to prosody, word order, and un-prefixation. In the following, I
will use somewhat simplified syntactic representations, ignoring PartP-internal structure and externalization of the internal argument, and using a head-final structure.
Recall that I follow Kratzer (, ) in assuming that adjectivization can be
lexical (targeting heads) or phrasal (here: targeting a PartP). The tree in () represents
the lexical case, where an adjectival passive does not involve additional event-related
modification. The input is the participle alone, e.g. geschlossen closed (a complex head
consisting of V and Part ).
() Lexical adjectivization


() represents the phrasal case involving a modifier of the event kind (instruments,
by-phrase I). The input is the PartP and the PP modifier adjoined to it.
() Phrasal adjectivization: event kind modifiers



von einem Kind

by a child


The assumption that event kind modification takes place before adjectivization captures the fact that the PP is prosodically integrated into the participle and cannot
appear after it. Furthermore, as Rapp () points out, un-prefixation is not possible,

Still Puzzled by Adjectival Passives?

since only lexical adjectivization is compatible with the adjectival prefix un- (following
Kratzer ).
Finally, an adjectival passive with a by-phrase modifying a state token (by-phrase
II) is represented in ().
() Modifiers of AP: state modifiers

von der Musik
by the music


(un)beeindruckt -

This analysis accounts for the fact that state-related by-phrases are not prosodically
integrated into the participle and that they can appear before or after the participle,
given that they apply after adjectivization. In addition, the participle itself, which is
the output of lexical adjectivization, is compatible with un- (following Rapp ).
In the following section, I will argue against an alternative account of event-related
modification in terms of state modification.
.. Against the uniform state modification account
Maienborn (a) suggests that the restrictions on event-related modification of
adjectival passives follow from general restrictions on state modification. This suggestion is worked out in Gese (), who argues that such modifiers trigger coercion
of the state into an event and pragmatic enrichment. Underlying such an account is
the assumption that the participles are state-denoting adjectives with the event being
truly absent, and that adjectival passive is actually a misnomer.
However, the uniform account of event-related modification as state modification
raises several problems. For one thing, under this approach it would not be clear how
to account for the difference between the two by-phrases. Even more problematic is
the fact that the restrictions on state modification are not completely identical (cf.
Maienborn b, Katz ). It is true that both states and adjectival passives do
not allow spatial modifiers (unless the spatial frame creates some kind of temporal
frame); compare () and (a) (Maienborn b: ). It is also true that both allow
modifiers temporally locating the state; compare () and (b) (Maienborn b:
() a. Das Kleid ist auf der Wscheleine nass.
the dress is on the clothes line wet
intended: The dress is wet on the clothes line.

b. Carolin war gestern / zweimal / tagelang mde.
Carolin was yesterday twice
for days tired
Carolin was tired yesterday / twice / for days.

However, only adjectival passives allow manner adverbials, by-phrases, instruments;

states do not (cf. (a)), except in very rare cases like (b).
() a. Bardo war friedlich / mit seinem Teddy mde.
Bardo was peacefully with his
teddy tired
intended: Bardo was tired peacefully / with his teddy.
(Maienborn b: )
b. Peter war mit Begeisterung Angler.
Peter was with enthusiasm fisherman
Peter was a passionate fisherman.

(Maienborn b: )

Maienborns take on (b) is that this case involves the coercion of a state into an
activity, based on world knowledge, to resolve the sortal conflict between a state and
a modifier that needs to apply to an event. Like Gese () for adjectival passives,
then, she reverts to event coercion, where the event is non-compositionally inserted.
What is crucial in licensing the modifier in (b), however, is that Angler is actually
a deverbal noun derived from the verb angeln to fish (literally: fisher). So the actual
generalization is that event-related modifiers are possible with states that are derived
from verbs, be it nominal or adjectival states. But then why do we not just take the
morphology more seriously and assume an event with deverbal nouns and adjectives
to begin with?

. Conclusion
In this chapter I have shown that event-related by-phrases with German adjectival
passives are crucially different from by-phrases with verbal passives, since only a
subset of the by-phrases that are acceptable with verbal passives are also acceptable
with adjectival passives, and since furthermore those with adjectival passives do not
introduce a discourse referent. Based on the general restrictions on event-related
modification with adjectival passives, I proposed that German adjectival passives refer
to the instantiation of a consequent state kind of an event kind, and that therefore
event participants of such event kinds do not get instantiated but remain in the kind
domain as well. Finally, I argued that by-phrases with adjectival passives are acceptable
only if they modify either the event kind (naming an established subkind) or the
(consequent) state token (which is only possible with stative verbs).
This is noted in passing by Maienborn, but apparently not seen as relevant.

Still Puzzled by Adjectival Passives?

The key idea underlying this proposal is that (re)categorization of verbal structure
as an adjective has consequences for the semantic nature of the underlying event:
it does not get instantiated. We should expect this idea to have repercussions for
the analysis of other non-verbal categories that have been argued to contain verbal
structure, such as nominalizations. Furthermore, we should expect some parallels
between event participants of deverbal adjectives and deverbal nouns.
A comparison of by-phrases with adjectival passives and with nominalizations
might be revealing for the overall question whether or not adjectival passives contain
Voice. Alexiadou, Anagnostopoulou, and Schfer () argue that, in English, byphrases can be licensed by Voice (e.g. in verbal passives) or by encyclopedic knowledge associated with the (underlying) verb (e.g. in -(a)tion nominalizations). They also
suggest that the German counterpart of -(a)tion is -ung. Unlike adjectival passives,
such nominalizations do not combine with phrases headed by von by, but only with
phrases headed by durch through (introducing both agents and causers). This, in
turn, might indicate that the von-phrases in adjectival passives are licensed by Voice
after all and not just by encyclopedic knowledge. A more detailed investigation would
have to address the general question concerning prepositions across verbal and nominal domains, also from a crosslinguistic perspective; however, this has to be left for
future research.

The Role of Syntax in Stress
Assignment in Serbo-Croatian


. Introduction
The chapter focuses on de-adjectival nominalizations, which show insightful interface
effects in two distinct domains of grammar in Serbo-Croatian (S-C): the morphosyntactic process of derivation of nouns from adjectival stems, and the phonological process of stress/tone assignment. In both domains, the results of the observed processes
come in two general flavours. As for morphosyntax, there is, on the one hand, a class
of nominalizations which are fully compositionally (transparently) derived from a
very broad set of adjectival stems using always the same derivational morpheme
(-ost). Due to these characteristics, we analyse such nominalizations as members of
the (extended) paradigm of the adjectival stems in S-C. On the other hand, there
are nominalizations which pattern together for being strikingly idiosyncratic: they
are derived only from a limited number of stems, use various derivational suffixes,
and come in most cases with a semantic shift with respect to the compositional
interpretation. In the other domain, stress and tone assignment, there is a comparable
dichotomy: there is a class of nominalizations in which the positions of tone/stress are
predictable from their position in the base word, and there is another class in which
another pattern emerges (either the phonologically unmarked pattern or the stress
pattern imposed by the suffix), resulting in a deviation from the prosodic properties
of the base word.
We are grateful to Ren Kager, Tanja Samardic, Marieke Schouwstra, Wim Zonneveld, our two
anonymous reviewers, and the audience of OnLI II for the many useful comments on the previous versions
of this chapter. The research presented here is partly supported by the research projects FFI- by
There are adjectives that do not nominalize at all, such as comparatives and superlatives, and very few
which derive only the other type of nominalizations, and which we briefly mention and discuss in fn. as
well as in section .. But with a marginal approximation, one can say that all adjectives that appear as
predicates also derive this type of noun.

Stress Assignment in Serbo-Croatian

We show that there is a neat match between these two domains. De-adjectival nominalizations which are productive, and part of the paradigm, i.e. which are spelled out
to phonology with a rich and straightforward syntactic structure, have their stress and
tone copied from the base adjective. In idiosyncratic de-adjectival nominalizations, to
which we assign a less rich syntactic structure (a bare concatenation of the two elements), the unmarked prosodic pattern emerges, i.e. these nominalizations undergo
a purely phonological tone/stress assignment, insensitive to any non-phonological
features, most importantly the surface prosody of the base. Finally, even though a
strong correlation as described can be observed, there are cases where particular
stems do not derive one of the two types of nominalizations, i.e. where one of the
types takes on the interpretive effects normally typical of the other type, or where
the phonological environment does not allow for the contrast to be brought out (see
section ..).
The research question can be stated as follows: why does structurally complex,
transparent interpretation go hand in hand with stress/tone assignment sensitive
to the surface prosodic pattern of the base adjective? And why are there gaps in
this correlation? Our account crucially relies on the fact that faithfulness to the
surface prosody of the base word is a tool in the reconstruction of the internal
morphological structure of a complex word. When its components stand both for
their own lexical semantic contribution and for a sequence of syntactic structure,
identification of the elements is vital for the recoverability of the structure. When
the complex word has a non-transparent interpretation and as such constitutes a
lexical entry, its internal structure is not crucial for its interpretation and its syntactic
behaviour, so it also need not be recoverable, and phonology therefore does not take
the surface prosody of the base adjective into account. This analysis is supported
by the experimental findings that modifying the stress of the stem will impede or
slow down access (Steriade , quoting Cutler ). The gaps are ascribed to
particular constraints, mostly phonology-internal but sometimes syntax-internal, on
the derivation of one of the two types of nominalization. As we shall see, there are
various handshakes, or guaranteed alignments, between syntactic structures and the
phonological structures which serve as their exponents, so that the variation explored
here is systematically excluded in many contexts for both semantic and phonological
We make use of Steriades () concept of Lexical Conservatism to establish a
link between the internal syntactic structure of a derived word and the activation
of phonological constraints which require the preservation of the prosodic properties of the base adjective. The constraint family LC militates against the introduction of novel allomorphs by enforcing the copying of the surface properties of
the listed allomorph in every new form. In the case at hand, LC(Stress/Tone)
requires the copying of the surface stress pattern of the base adjective. Crucially,
this constraint is sensitive to whether the form does or does not belong to the

Arsenijevic and Simonovic

paradigm of the stem involved, and militates against discrepancies in paradigm

The rest of this chapter is organized as follows. Section . introduces the two
classes of de-adjectival nominalizations in S-C, with their semantic and morphosyntactic asymmetries. In section . we recapitulate the syntactic analysis from Arsenijevic (). Our phonological model is introduced in section .. Section .
discusses the issue of the status of syntax with respect to the module of phonology, as
indicated by our analysis. In section . we look at the possibility of an extension of
our approach to other languages, focusing on a comparable pattern in Dutch. Section
. concludes this contribution.

. Two Types of De-adjectival Nominalization by Suffixation

in Serbo-Croatian
S-C adjectives nominalize in two general ways: through suffixal derivation and
through conversion. We concentrate here on derivation, and describe two different
types of de-adjectival nominalization derived by suffixation. As foreshadowed in the
previous section, these two different types of nominalization come with two different
sets of semantic, syntactic, and phonological properties. In derivation, the default
suffix used is -ost, but there are also other suffixes with the same syntactic and semantic
effects. The suffix -ost is special in taking part in both types of de-adjectival nominalization discussed here. Only one type, idiosyncratic de-adjectival nominalization, also
makes use of other suffixes. Conversion, which is not directly relevant for the topic of
our chapter, has been discussed in Arsenijevic ().
The most straightforward way of specifying the split between the two groups is in
terms of their stress patternsa dimension which we discuss in more detail in section
. (a general flavour is given in (b) vs. (c)). For the moment, it is worth noting
that the nominalizations in (b) have a stress pattern different from the base adjective,
whereas the ones in (c) fully copy the prosody of the base adjectives (a). Other differences are morphosyntactic and semantic. Arsenijevic (, ) qualifies these
differences in terms of structural complexity. The nominalizations in (b) are similar
to what is commonly referred to as (deverbal) root nominalizationsthey seem to
involve no structural complexity, often have a non-transparent interpretation, and do
not necessarily entail that there is a bearer of the property denoted by the adjectival
stem. In further text, we refer to them as the Lexical (or Lexicalized) De-adjectival
Nominalizations (LDN). Those in (c) are structurally complex, and arguably nominalize a PredP, in certain cases with a full VP embeddedtheir interpretation is
regular and transparent, including an entailment of an instantiation of the property
involved in a bearer (which is preferably overtly expressed). Henceforth, we term them
Structurally complex De-adjectival Nominalizations (SDN).

Stress Assignment in Serbo-Croatian



b. opAAsn-OOst


prazn-InA (?sobe)
empty-ost room.gen
emptiness (belonging to the room)

c. Opaasn-oost


prAAzn-OOst (sobe)
emptiness (of the room)

() a. Opaasan

In the next section, we give an overview of the main syntactic and semantic differences between these two types of nominalization. We argue that the crucial difference is in their productivity, i.e. that one class is part of the paradigm of the
adjectival stem, and the other includes instead idiosyncratic instances of lexical
As mentioned above, another important asymmetry between the two types of
de-adjectival nominalization is in their number and productivity. Almost every
adjective derives a SDN. The number of adjectives deriving LDNs is relatively
small, and subject to certain restrictions. There are actually several types of adjectival stem which only derive SDNs, including those of adjectival (usually resultoriented) passive participles, those of adjectival active participle forms, and the
counterparts of English -able adjectives, which all arguably embed a VP structure.
Forms such as the starred ones in () are not attested in dictionaries and grammars, and they are all judged totally unacceptable by S-C native speakers. In sections
.. we show how a conspiracy of syntax and phonology brings about these
a. UvrEEden-oost
b. posUstAl-oost
c. cItljIv-oost








Capital letters are used to mark prosodic prominence, which is marked by stress and tone in SerboCroatian. The representations of Serbo-Croatian prosody are discussed further in section ..

Arsenijevic and Simonovic

Asymmetries between the two classes of nominalizations are best visible when the
same adjectival stem derives both a LDN and a SDN. Looking at examples of this
type, the following asymmetries can be observed.
When used for properties of particular persons and objects, matching in denotation Moltmanns () type of tropes, SDNs are more suitable, and LDNs are often
() SlUcaajnoost/ slucAAjnOOst ovog uzorka
ne dovodi se u pitanje.
this sample.gen not lead refl in question
The random nature of this sample does not come into question.
More precisely, LDNs are out if they have undergone a lexical semantic shift, becoming
idiosyncraticusually in taking, as a more prominent or even the only one, one type
of reference: to events, situations, concepts, or some other semantic class. For all
meanings other than properties as instantiated in particular referents, some of which
are discussed in what follows, the LDN is a better candidate than its SDN counterpart
(the SDN often cannot receive the relevant interpretation). For instance, LDNs can
be modified by adverbials that select for quantized eventualities, while their SDN
counterparts cannot.
() a. cesta
frequent danger.ldn
frequent danger

nekadanja rudarska solidAArnOost

earlier.adj miners solidarity.ldn
miners solidarity from the past

nekadanja rudarska sOlidaarnoost

b. (njegova) cesta
frequent dangerousness.sdn earlier.adj miners solidarity.sdn
his frequently instantiated dangerousness

c. Jovanov-a/-e povremen-a/-e ljubAAznOst(-i)

J-poss.sg/pl occasional-sg/pl kindness(es)-ldn
an occasional kindness from/by Jovan/occasional kindnesses by/from Jovan
d. Jovanov-a/-e povremen-a/-e ljUbaaznost(-i)
J-poss.sg/pl occasional-sg/pl kindness(es)-sdn
Jovans occasional kindness
LDNs go well with count quantifiers and modifiers, receiving an eventive interpretation (quantification is over events in which the property denoted by the adjective is
instantiated), while with SDNs these constructions are ungrammatical.
() a. nekoliko Jovanovih ljubAAznOsti/ ljUbaaznosti
several Jovans kindnesses.ldn/kindnesses.sdn
several events instantiating Jovans kindness
b. razne opAAsnOsti/ OpAAsnosti
diverse dangers/dangerousnesses
diverse dangers

Stress Assignment in Serbo-Croatian

Another asymmetry relates to a subtle intuition: while LDNs can have a generic
meaning, referring to a concept, an intensional property, SDNs always seem to imply a
bearer of a property, as instantiated in a particular referent. This asymmetry is similar
to the one between unaccusatives on the one hand and transitive verbs that allow for
a zero object on the other, where the former (e.g. sink, float, break) do not imply the
existence of an agent, and the latter (e.g. eat, read, wash) do imply the existence of an
affected participant.
() a. ta mala ljubAAznOOst o
kojoj pricam, koju niko
that little kindness.ldn about which talk.sg which nobody neg.aux
that little kindness that Im talking about, which nobody manifested
b. ta mala ljUbaaznoost o
kojoj pricam, koju niko
that little kindness.sdn about which talk.sg which nobody neg.aux
that little kindness that Im talking about, which nobody manifested
An SDN always denotes a property holding of a referent. An LDN may also have this
interpretation, but it is more naturally interpreted as talking about a generic notion,
intensional or extensional.
() a. KnjIeevnoost ovog
teksta je upitna.
literariness.sdn this.gen text.gen is questionable
The literariness of this text is questionable.
b. KnjIeevnoost ovog
je upitna.
literariness.sdn this.gen author.gen is questionable
The literariness of this author is questionable.
( unless the context makes literariness a natural property of an author)
c. KnjiEEvnOOst ovog
je upitna.
literature.ldn this.gen author.gen is questionable
The literature (by this author) is questionable.
d. KnjiEEvnOOst ovog
teksta je upitna.
literature.ldn this.gen text.gen is questionable
The literature of/by this text is questionable.
( unless a context is made where a certain piece of literature is part of a
certain text)
Finally, while semantic counterparts of LDNs are derived by many different (stemspecific) suffixes, the semantics specific to SDNs only comes with nouns built by the
suffix -ost which is added to an adjectival stem.

Arsenijevic and Simonovic

() a. impotEnc-Ija

knjiEEvn-Ost dobr-OtA dOsad-a

good(ness) boredom


b. impotEntn-oost, knjIeevn-oost, ?dObr-oost, dOsadn-oost, ?Umoor-n-oost

being impotent literariness
goodness boringness tiredness
While LDNs normally have a broader range of interpretations, and SDNs are more
restricted, there are uses in which only the SDN is acceptable.
() Njegova Opaasnoost/ opAAsnOOst ne dovodi se u pitanje
dangerousness.sdn/ldn not lead refl in question
His dangerousness does not come into question.
It is a general fact that when used for properties of particular persons and objects, i.e.
to stand for concrete predications, SDNs are more suitable, and LDNs are sometimes
out. The only class of adjectives that do not build SDNs are adjectives that do not
nominalize at all: relational and certain subsective adjectives, i.e. those adjectives that
cannot appear as predicates.
() a. dOOkAAznii materijal
evidence.Adj material
b. Ovaj materijal je dOOkAAzan
this material is evidence.Adj
c. dOOkAAzn-oost
d. dokAAzn-OOst
The regularity is that all adjectives that can appear as predicates also can derive a SDN,
but not necessarily also a LDN. This property of SDNs is one of the reasons why in
the next section we analyse SDNs as members of the paradigm of an adjective, and
LDNs as new words derived from a certain stem, possibly establishing their own new

That these suffixes are stem-specific means that they are also in mutual exclusion: where one is used to
derive a noun from an adjective, the others derive ill-formed words. Together with their shared semantics,
this implies that they belong to the same class, or even different instantiations of formally the same suffix.
As briefly discussed again in section . below, adjectives that involve an additional argument of
their degree component, such as comparatives, superlatives, or adjectives denoting equal degrees of two
arguments, also cannot derive SDNs. Assuming the reason is the impossibility of embedding the second
argument in the nominalization, this further confirms our analysis.

Stress Assignment in Serbo-Croatian

Finally, it is not necessarily the case that the semantic SDNLDN asymmetry
triggers different phonological output. There are nominalizations like hUmaanoost humanity/humaneness, which have both the SDN and the LDN interpretation. Moreover, even in proper LDNSDN pairs, one finds variation among the
speakers of S-C, such that one of the forms is not actively used, leaving the other
with both interpretations available, as in the case of hUmaanost above. In section
.. we show that the neutralization of the contrast is phonologically conditioned,
but also that the contrast is systematically blocked in some categories where it
could not be meaningful. In sum, the distinction is not always available, but we
shall show that whenever it is systematically unavailable, it is unavailable for a
Before presenting the analysis, we need to point out that a detailed analysis of
a closely related phenomenon in Slovenian can be found in Marvin (: ).
Although the focus of her research is on participant-denoting deverbal nominalizations, Marvin does observe asymmetries in Slovenian that parallel those discussed
in our chapter. Moreover, she sketches a syntactic analysis which shares one crucial
property with the analysis outlined here and discussed in more detail in Arsenijevic (, ): that (in her case Slovenian) LDNs involve a syntactic simplex in
the adjectival component, while SDNs involve phrases (Marvin labels them roots
and APs). However, whereas in Slovenian the SDN/LDN distinction is prosodically
brought out in a small number of nominalizations from simplex native adjectives, in
S-C its scope is much larger and includes most Latinate adjectives (see Simonovic
). On the theoretical side, while we agree in the general intuition that one type
derives from a simple item, and the other from a non-trivial syntactic structure, we
argue here that the latter type involves a rich syntactic structure (an entire PredP),
i.e. that it cannot be accounted for only in terms of categorial heads typical for
the current theory of Distributed Morphology (such as vC, vE, Voice: see Marantz
, Anagnostopoulou and Samioti, Chapter below). Also, the present chapter devotes more attention to the sensitivity of phonology to syntactic structures,
while Marvin is primarily interested in the syntactic, phase-theoretic aspects of the

. Syntactic Analysis: Arsenijevic ()

We introduced the fact that adjectival participles only derive SDNs, never LDNs.
In fact, not even all active participles derive SDNs: only those built from unaccusative
or middle VPs, which have in addition received an adjectival (rather than eventive)
Transitive and unergative verbs make -ost-nominals from the passive participle, and with unaccusative
and middle interpretations build them from the active participle.

Arsenijevic and Simonovic

() a. utrnul-oost
go numb.ActPart-ost
b. udaril-ost

be paralysed.ActPart-ost
fatigue, being asleep, numbness


An explanation is due in respect of what is traditionally referred to as the active

participle in S-C. This form is rather a subject-oriented participle, which assigns a
process or result interpretation of the respective verb to the subject (i.e. the agent)
if the verb is transitive or unergative, and only a result interpretation to the subject
of the verb if it is unaccusative (in this case, to the patient or theme). Only the latter
nominalize, and hence only result interpretations are attested on nominalizations of
the active participle.
The restrictions on participles in deriving SDNs can be formulated in a simpler
way: only those participles which can also be used with a copula, with an internal
argument as the subject, and receive an adjectival interpretation, can derive SDNs
with the ending -ost (see footnote above).
() a. Ruka je utrnul-a
/ obamrl-a.
arm is go numb.ActPart-F.Nom be paralysed.ActPart-F.Nom
An/The arm is numb/paralysed.
b. Jovan je pevao
Aux sing.ActPart song
Jovan sang a/the song.
Pesma je
song Aux sing.ActPart
intended: The song went on.
intended: the property of having sung or the property of having been sung
This, together with the fact that the typical interpretation for an SDN is that of a
property as holding for a particular referent, suggests that a predication actually
underlies each SDN or, in syntactic terms, that SDNs derive from PredPs. As PredP is
a relatively high structural projection, this account also predicts that SDNs will always
have a transparent interpretation, i.e. that they will not be able to establish idiomatic
meanings, while LDNs will have available both options: compositional and idiomatic
interpretations (see Anagnostopoulou and Samioti, Chapter below, and Harley and
Stone, Chapter , for extensive discussion of the syntactic boundaries for idiomatic

Stress Assignment in Serbo-Croatian

Arsenijevic () argues for a syntactic analysis in which the main asymmetry

between SDNs and LDNs is in their syntactic complexity. LDNs are derived from a
bare stem, without any internal syntactic structure (i.e. its structure is eliminated in
the process of lexicalization), while SDNs involve a richer structure, where a PredP is
() A simplified schematic representation of the syntactic structures of SDNs and







SDN: Opaasnost

LDN: opAAsnOst

The semantic asymmetries are derived from the syntactic difference: LDNs nominalize a concept corresponding to the lexical meaning of the stem, while SDNs nominalize a predicative relation between an argument and a verbal or nominal predicate.
Nominalized predications are argued to be restricted to denote property instantiations, i.e. tropes, and nominalized lexical meanings, by default, property types, but
they are also free to syntactically derive any other type of denotation, in particular
tropes and eventualities. While Arsenijevic () merely hints at a phonological
analysis, in terms of the order of Spell-out, in this chapter we aim at providing a deeper
phonological analysis and drawing out the implications that it has for the status of
syntactic structure in a phonological process such as stress and tone assignment.
One anonymous reviewer remarks that our analysis of SDNs incorrectly predicts that they should allow
for accusative complements and expressions like *veoma opasnost very dangerousness, drawing a parallel
with similar effects observed on nominalizations of complex structures (see a summary in Kornfilt and
Whitman ). As for modifiers like veoma very, we explain their ungrammaticality in nominalizations
by the requirement of their scalar semantics for a positive morpheme (Kennedy and McNally ), and
the intervention effect of the nominalizing suffix. As we are discussing de-adjectival rather than deverbal
nominalizations, and adjectives in S-C do not take accusative complements, there is no reason why SDNs
should take accusative complements (in nominalized participles, the participle is already incompatible
with a complement in the accusative case). But it lends further support to our analysis, as asymmetries
in preserving complements of the base adjective do appear in the predicted way:

(i) a. on je bio impotEntan [u tom trenutku], ona je bila dOsaadna [preko svake mere]
he Aux been impotent at that moment she Aux been boring
over every measure
b. njegova [impotEntnoost / impotEncIja [u tom trenutku]] /
impotence.sdn *impotence.ldn at that moment
[preko svake mere]]
njegova [dOsaadnost / dOsaada
boringness.sdn *boredom.ldn over every measure

Arsenijevic and Simonovic

We argue that the relevant notion in this respect is that of paradigm. We take the
paradigm of a stem to include all the forms that can productively be built from each
member of the relevant class of stems, irrespective of the syntactic class of the product
(i.e. belonging to the nominal category in this particular case does not prevent an
item from being a member of the paradigm of an adjectival stem). Under such a
definition, de-adjectival nouns are members of the paradigms of adjectival stems,
as long as they can be derived from each member of one or more morphosyntactic
classes of adjectival stems. Paradigms are characterized by a high degree of transparency: words derived are interpreted strictly compositionally with respect to the
internal structure. Derivations which do not satisfy this condition, and hence stand
outside the paradigm of the stem involved, are rather idiosyncratic in their nature,
and the derived word is prone to a shift in lexical meaning. This is especially likely to
be the case if the derivation proceeds via a relatively simple structural configuration,
as is the case with LDNs.
We have tested our predictions on data extracted from the dictionary (RMS
) and through acceptability judgements of more than ten native speakers
of S-C. It was empirically confirmed that each of the adjectives of the relevant type
can derive a SDN, although in a small number of cases, these came with a degree
of (arguably pragmatic) degradation. Only a restricted number of adjectives were
attested to derive LDN forms. In what follows, we turn to the productivity of the
two types of nominalization. A generalization can be formulated that the more structurally complex the stem is, the more likely that only a SDN will be derived from
it. For instance, as already discussed, participial and other adjectival forms which
arguably involve a VP structure (e.g. S-C counterparts of the English -able adjectives)
can only ever derive a SDN: their LDN counterparts are systematically ungrammatical. Certain phonological restrictions apply as well, and these are discussed in
section ..
What we have seen so far has provided some converging evidence for our initial
proposal that in the default case, the SDN is a member of the paradigm of the adjectival
stem, while the LDN is an idiosyncratic instance of lexical derivation. This is also the
hypothesis on which we build our analysis of the sensitivity of phonology to syntactic

. Phonology
This section brings our phonological account, which makes crucial use of interface constraints, in particular of their capacity to distinguish between members of
paradigms of the base (SDN) and lexicalized derivations (LDN). In order to present
our phonological account, we first need to introduce some preliminaries on S-C
prosody and the representations we use.

Stress Assignment in Serbo-Croatian

.. S-C prosody
The traditional view is that there are four accents in S-C, differentiated by length and
() Traditional diacritics for S-C accents, and the notation used here (in
short rising: vda (vOdA) water.nom short falling: v
odu (vOdu) water.acc
long rising: tga (tUUgA) sorrow
long falling: kla (kOOla) school
The rising accent is a tonal span over the accented syllable (which hosts the stress) and
the syllable following it. All generative analyses of S-C prosody (for a recent overview,
see Werle ) share a few important assumptions. First, the surface rising spans
imply underlying tones on the rightmost syllable of the span, so that vda (vOdA in
the notation used here) is analysed as /vodaH /. Whenever there is an underlying tone
hosted by a non-initial syllable, it will create a rising span over the originally tonebearing syllable and the one preceding it (hence vda/vOdA) and the stress will be
located on the left syllable of the span. This, in combination with the general ban on
final stress, explains why all non-initial accents are rising (at least in native words), so
that a form like vodA is excluded. This is also the reason why minimal pairs between
rising and falling accents can only occur on initial syllables (e.g. vEcEra dinner.nom
vs. vEcera have dinner.aor). Second, the falling accents, which can occur only on the
initial syllable of a word, are the underlying Hs on that syllable, so that vodu (vOdu)
corresponds to /voH du/. Finally, the initial falling accent is assigned post-cyclically to
words which have not acquired a stress/tone in the derivation (Inkelas and Zec ,
Zec ). Consequently, /voH du/ corresponds to the toneless /vodu/ on a yet deeper
The present chapter looks at S-C prosody from the perspective of a model which
focuses on surface prominence and how it is copied throughout paradigms. For this
reason, we are using the following format to represent S-C prosody: capital letters are
used to mark all the syllables which get surface prominence (stress and tone) and double vowels to mark long vowels. Note that single capitalized vowels (the falling accent,
where tone and stress cooccur) will be encountered only on the first syllable (in native
words). Any two adjacent capitalized syllable nuclei stand for a rising span with the
stress falling on the leftmost syllable of the span. In sum, the capital letters can be read
as indicators of tone-bearing syllables, whereas the distribution of stress is predictable
from that of tone (e.g. tatAtA stands for ta"tAtA, whereas tAtata stands for "tAtata).
() Representations of S-C tones used here, preceded by the corresponding underlying structure with marked Hs
(tataH ta ) tAtAta: short rising span over first and second syllable
(taataH ta ) tAAtAta: long rising span over first and second syllable
(tatataH ) tatAtA: short rising span over second and third syllable

Arsenijevic and Simonovic

(tataataH ) tatAAtA: long rising span over second and third syllable
(tatata taH tata ) tAtata: short falling accent on the first syllable
(taatata taaH tata ) tAAtata: long falling accent on the first syllable

.. The locus of variation: -ost prosody and the default pattern

Ost-suffixation generally does not affect the lexical tone/stress of the adjectival stems.
As () shows, both surfacing rising span (a,b) and lexical falling accent (allowed
only in loanwords: see Simonovic ) are preserved in the derived abstract nouns.
From this we can conclude that the suffix -ost either has no underlying tone or has
an underlying tone which gets overridden by that of the base. If the latter is the case,
-ost would be exceptional with respect to all the other nominalizing suffixes, which all
attract stress/tone () and which we further discuss in section ...
() a. gramAtIcn- grammatical + -ost
b. UvElwithered
+ -ost
c. konkrEtn- concrete
+ -ost

gramAtIcnoost grammaticality
konkrEtnoost concreteness

The situation in the adjectival stems which surface with post-cyclic prosody
tone/stress aligned with the left edge of the word, spelled out as an initial falling
accentis much less clear. The few available generative descriptions are unclear as
to what the main pattern is. For instance, Inkelas and Zec () quote the example
in () to show that -ost is one of the suffixes after whose addition a High is added
at the right edge of the word. Note that the initial falling tone of jEdnaak is assigned
() jEdnaak equal + -ost

jednAAkOOst equality

On the other hand, Zec () presents the examples in (), in which the toneless
stems do not have any change of the position of stress/tone.
() a. lUUd
insane + -ost
b. hUmaan humane + -ost

hUmaanoost humanity

It is clear that jednAAkOOst and hUmaanoost are expected to display the same pattern
and that one of them has to be exceptional. As it turns out, the pattern in jednAAkOOst
is the odd one out, but this actually arises due to the specificity of its semantics.
As already hinted in footnote above, adjectives involving an overt degree argument,
We refer the interested reader who is wondering about the rationale behind the assignment of the
initial high tone to toneless words to Zec and Zsiga () for a recent account.
The fact that -ost is productive and applied to virtually every adjectival stem cannot be a basis for
a whole account here. For instance, the comparative suffix -ijii is just as productive, but it does affect
the stress/tone (e.g. the comparative forms of the adjectives in () are gramatIcnIjii, more grammatical,
uvElIjii more withered, and konkrEtnIjii more concrete).
Zec () also has a class of adjectival stems for which a floating lexical High is assumed. It is not
clear to us what the evidence in favour of this representation is. Furthermore, most examples provided by
Zec () are problematic to a sizeable number of our informants.

Stress Assignment in Serbo-Croatian

such as comparatives and superlatives, cannot derive any nominalizations. The adjective jEdnAAk equal is neither comparative nor superlative, but still involves a degree
argument, as it specifies that a certain property holds to the same degree of two or
more referents. In a copular construction, it involves a second argument, which is an
argument of degree, just as in comparatives.
() Nisam ja jednak sa tobom.
neg.am I equal with you
Im not equal to you.
Our model hence predicts for this adjective that it cannot derive a SDN. However,
while comparatives are usually hard to take to the generic level (cf. biggerness,
smarterness in English), the meaning associated to jEdnaak equal can be taken to
the generic level, and hence this stem derives a LDN: jednAAkOOst equality.
In sum, the adjectives with post-cyclic prosody do not, as a group, show any evidence for a H in the lexical representation of -ost. However, as we have seen already
in the examples in section ., there is a sizeable number of nominalizations derived
from adjectives which surface with post-cyclic prosody (e.g. Ispraavan) in which there
is a semantically conditioned variation in prosody between the pattern of copying the
surface pattern of the base adjective (Ispraavnoost) and creating a rising span on the
right edge (isprAAvnOOst).
Consulting the native speakers on our dataset of over nominalizations which
have a base which surfaces with post-cyclic prosody has enabled us to state the following generalizations about the distribution of the rising span (as in isprAAvnOOst):
() It never arises when the syllable preceding the suffix is short. So we only ever
have kIsel+oost sourness and hrApav+oost coarseness, but no comparable
derivates with a rising span.
() Pace jednAAkOost, there are no de-adjectivals which have a single consonant
preceding -ost and rising span (hUmaanost, sImultaanost, Androgiinost, etc.).
() In the de-adjectivals with a long closed penult, the rising span is generally
possible and the variation depends on the semantic opposition between the
SDN and LDN.
The emergent prosodic pattern with a rising span points at a general tendency for
stress and weight to cooccur, formalized in the famous Stress-to-Weight constraint
Note that in words of the type Opaasnoost/opAAsnOOst and ljUbaaznoost/ljubAAznOOst, we are
assuming syllable boundaries cutting though the clusters consisting of fricatives and sonorants (sn, zn,
n, n, etc.) in spite of the fact that they can be syllable onsets in S-C (e.g. sneg snow and znoj sweat).
We claim that these clusters are possible syllable onsets only word-initially. This claim is sustained by their
distribution. First, there are no S-C words in which these clusters are preceded by a consonant (which
would enforce the syllabification of the clusters as onsets): there are words like kasno and grozno as well as
words such as lipsati to die and obzir regard, but no words like *kapsno and *grobzno. Note that we do not
claim anything about clusters of the type fricative + plosive (sp, sk, etc.), which never occur in the context
relevant here.

Arsenijevic and Simonovic

(Kager : ), as well as the fact that coda consonants seem to contribute to weight
in S-C. The fact that a rising accent span is built at the end of the word, attracted
by a heavy syllable, is a well-attested preferred pattern in S-C (see e.g. Werle :
discussion of the Rightward tone shift). The new insight is that this pattern
is not restricted to words which have an underlying tone (as generally assumed), but
actually occurs also in toneless environments under certain conditions and that it is
not only vowel length, but also coda consonants which create the environment for
the shift.
In order to make sense of the emergence of the rising span, it is instructive to
look at the -ost-nominalizations derived from monosyllabic and disyllabic toneless
() cIl(ii)
chirpy + -ost
glAtk(ii) smooth + -ost
pEgav(ii) freckled + -ost
cUUln(ii) fleshly + -ost
bUUrn(ii) stormy + -ost
brIIn(ii) worried + -ost




The generalization is that if the base contains a short syllable, the initial short falling
accent is carried over to the noun. However, if there is a long syllable in the base (so the
base adjective has a falling accent as in cUUlnii), the rising span arises: cUUlnOOst,

cUUlnoost. There is, however, a small class of apparent exceptions to this generalization, which we represent in ().
() jOOnsk(ii) ionic
+ -ost
mIItsk(ii) mythical + -ost



There is one important difference between the adjectives of the type cUUlnii and those
of the type jOOn.skii. The latter class is exceptional in that its paradigm lacks the
indefinite forms, most importantly the nominative singular masculine form, the only
form in the adjectival paradigm which has a null ending. So there is the form cUUlan,
but there is no form joonsk/ joonsak.
This means that the entire paradigm of the adjectives of the type jOOn.ski has the
initial long closed syllable, whereas the paradigm of the words like cUUlnii always
contains a form like cUU.lan in which the long syllable is not closed. This seems to be
the point at which the grammar makes the distinctiononly the consistently closed
syllable with a long vowel is good enough to arrest the H in all the derivates. The
relevance of the syllabification clearly points at the importance of the surface form of
the base in word derivation, which is one of the main ingredients of our analysis. Here
we can see the first interesting handshake between syntax and prosody: due to the

There are no adjectival bases longer than one syllable which have a long syllable with a falling accent.

Stress Assignment in Serbo-Croatian

semantics of the adjectivizing suffix -sk, these adjectives will only ever derive SDNs.
The reason, in short and with some simplification, is that this suffix always attaches to a
noun interpreted as a referential nominal expression, hence always carrying complex
internal syntax, which cannot be represented in the flat structure of LDN. As we shall
see below, the hyperfaithfulness to the surface form of the base in SDNs is one of the
main features of the S-C system, for which our model accounts.
In sum, because most adjectives do have full paradigms, the default pattern in this
class is: the falling accent if it is hosted by a short syllable or a rising span if it is headed
by a long syllable. In this sense, the pattern encountered in polysyllabic derivations
can be seen as an extension of this default pattern.
Both stress-to-weight effects and the influence of syllabification on the prosody
are underexplored for S-C. Given the restricted focus of this chapter, we leave it to
future research to provide a complete picture. For present purposes, it is important
to clearly delineate the locus of variation in de-adjectivals, where phonology favours
different output forms based on semantic/syntactic structure. As we have already seen,
in the case at hand variation is by and large restricted to the many nouns derived from
toneless adjectives (which means that they surface with an initial falling accent) whose
stem ends in a consonant cluster preceded by a long vowel e.g. Ispraavn(ii) correct.
In these cases the options supplied by the phonology are the following:
() the pattern which copies the post-lexical surface prosody of the base adjective:
Ispraavn(ii) + -ost Ispraavnoost
() the emergent unmarked stress pattern with a rising span on the right edge:
Ispraavn(ii) + -ost isprAAvnOost.
This is exactly the locus of prosodic differentiation between SDN and LDN deadjectivals: the pattern copying of the base adjective prosody prevails in SDNs,
whereas the default pattern with a rising span arises in LDNs.
At this point, having defined the locus of variation, we can understand better
another important handshake between syntax/sematics and the phonological form:
the three categories which are not able to derive LDN due to their syntactic and
interpretive properties (as shown in ()) also have phonological structures which
cannot derive a prosodically distinct LDN: participles (both active and passive) and
adjectives in -(lj)iv -able never create the right environment for systematic variation
between the two classes of de-adjectivals. This is because participles never end in a
combination of a long vowel and a consonant cluster (so -ost can never be preceded
by a heavy closed syllable) and (lj)iv has its own lexically specified H which leads to
-ljiv always being part of a rising span (e.g. vId-e-ti to see, vIdljIv visible).

The same default is true of borrowed (Latinate) nouns in -a. Generally, only closed long syllables
arrest the High: nOOrma, sAAlva, fAArma, but mOOdA, lAAmA, fAAmA (see Simonovic, in preparation).
The default short accent is falling: sEpsa, mOka, pAsta.

Arsenijevic and Simonovic

In the following section, we present the formal model in which we show how the
interface (syntactic/semantic) constraints interact with the prosodic ones, leading to
an account of the fact that SDN de-adjectivals copy certain surface properties of the
base adjective, unlike their LDN counterparts.
.. Formalization and analysis
In order to formalize this effect of copying of the surface properties of the base in
the process of paradigm formation/word derivation, we propose an analysis couched
in Optimality Theory (Prince and Smolensky , Kager ). The constraints
essential for our analysis, which militate against the emergence of novel allomorphs,
have been proposed by Steriade () and termed Lexical Conservatism constraints.
In () we give the canonical form of these constraints.
() The form of Lexical Conservatism conditions: LC(P)
Let T() be the allomorph of appearing in a form under evaluation.
Let L() be a listed allomorph of .
Let P be a phonological property.
T() is characterized by P only if some L() is characterized by P.
(Steriade )
At this point we need to explain the notion of paradigm as used in this paper.
Recent generative morphosyntactic theory (e.g. Halle and Marantz , Marantz
, Embick , Travis ) points out that there is no formal distinction
between derivation and inflection. In this light, traditional definitions of the notion of
paradigm as, roughly, a set of forms of one word distinguished by different inflectional
make-ups, no longer can be maintained. We take the paradigm to include all the words
which share the same stem and are derived by a productive mechanism, such that
their semantics is transparent (i.e. such that they have a compositionally interpreted
morphological configuration).
As we have seen, in ost-de-adjectivals, the lexical stress is generally faithfully preserved. So, the faithfulness constraints against deletion or dislocation of the lexical
stress/tone are ranked higher than the prosodic constraints responsible for the emergence of the right-aligned rising span headed by a long closed syllable (such as StressTo-Weight) which we collapse as Prosody. In this case Prosody requires the default
prosody of the class: it is violated whenever () there is a long closed syllable which
does not carry stress, or () there is any other configuration which does not have the
initial falling accent (a good overview of the specific constraints and their interactions
proposed for S-C prosody can be found in Werle ).
Since we have no cases of -ost changing the lexical stress/tone of the base,
Max(stress/tone) will be considered undominated. The relative ranking we arrive
at is Max(stress/tone) >
> Prosody.

Stress Assignment in Serbo-Croatian

Now we turn to the Lexical Conservatism constraints. Crucially, the constraints which require the preservation of the surface properties of the stem stress,
which we term LC(stress/tone), impose stronger requirements on SDNs than
on nouns whose meaning cannot be derived compositionally. So the constraint
LC(stress/tone)paradigm , indexed to SDNs, dominates the general LC(stress/tone)
constraint which is in charge of all derived words. Now, the crucial ranking is
the one which allows the prosodic constraints to influence the stress of the LDNs,
whereas in SDNs the copying of the stem stress is enforced. This ranking is
> Prosody >
> LC(stress/tone).
LC(stress/tone)paradigm >
Since there are no cases which would offer a ranking argument for the constraint
LC(Stress/tone)paradigm with respect to Max(stress/tone), we are leaving these
two constraints unranked with respect to each other. Turning to the tableaux, we need
to remind the reader that the LC constraints here require the copying of the prosody
of the adjectival form Ispraavan, a feature which is not present in the input underlying
representation. The use of a related surface form as part of the input is common in OT
approaches to paradigms, most importantly the outputoutput correspondence (for
an overview of approaches, see e.g. Steriade ). In what we have seen so far, there
were quite a few good reasons to assume that the surface form of the base is not part of
the input of the tableaux for LDN. Still, we are including this representation (and the
violations of LC(stress/tone) which it makes possible to assign) between brackets
in order to see how it affects general evaluation.
() a. /ispraavn + oost/ correctnessSDN MAX(S/T) LC(S/T)PARADIGM PROSODY LC(S/T)

a. isprAAvnOOst
F b. Ispraavnoost

b. /ispraavn + oost/~ correctnessLDN MAX(S/T) LC(S/T)PARADIGM PROSODY LC(S/T)


a. isprAAvnOOst
b. Ispraavnoost


As the tableaux in () show, the Max(stress) constraint is never violated because

there is no lexical stress to be unfaithful to. The constraint LC(stress/tone)paradigm
only plays a role in the evaluation of the candidates for correctnessSDN , whereas in
the case of correctnessLDN , whose meaning has to be retrieved from the lexicon, it is
always vacuously satisfied.
In order to show that our ranking predicts the lack of variation in nouns whose
base has a lexical stress/tone (e.g. konkrEtnoost) and in those whose base has no long
closed final syllable (e.g. hUmaanoost), in () and () we give the tableaux for these

Arsenijevic and Simonovic

() a. /konkrEtn + oost/~ concretenessSDN MAX(S/T) LC(S/T)PARADIGM PROSODY LC(S/T)


a. konkrEtnost
b. kOnkretnoost

b. /konkrEtn +oost/ concretenessLDN MAX(S /T) LC(S /T)PARADIGM PROSODY LC(S/T)


a. konkrEtnost
b. kOnkretnoost


() a. /humaan + oost/ humannessSDN MAX(S /T) LC(S /T)PARADIGM PROSODY LC(S/T)


a. humAAnOOst
F b. hUmaanoost

b. /humaan + oost / humanityLDN MAX(S /T) LC(S /T)PARADIGM PROSODY LC(S/T)


a. humAAnOOst
F b. hUmaanoost


As expected, the general constraint LC(s/t) never chooses the winning candidate
in our tableaux. It may well be the case that this is not accidental, i.e. that surface
properties (such as syllabification and surface prosody) cannot be copied outside
paradigms. If this is the case, it can be explicitly encoded into our basic model. LDNs
(and comparable non-transparent derivations) can be defined as those derivations
which do not have the surface form of the base as part of their input. This automatically
makes the general constraint LC(stress/tone) vacuous and superfluous. Also, the
remaining LC constraint would not have to be indexed to a subset of nominalizations;
it would just be vacuously satisfied in all the cases where it has nothing to compare to
(i.e. in all LDNs). Our dataset is not sufficient to confirm such a generalization, but our
analysis clearly puts this hypothesis on the research agenda and shows how different
outcomes of this research are to be formalized. The important conceptual point here
(as one of our reviewers suggests) is that, with or without this final intervention, we do
not deny that LDNs are related and even synchronically derived from the adjectival
base. However, (at least in S-C) they are crucially unable to copy the surface prosody
of the base. These are all questions which require further research.
Note that our constraint set as it is now also makes clear predictions about the
possible constellations in languages: we exclude the option where the non-paradigm
members copy more properties of the base than the paradigm members. Even in a
language in which LCparadigm is ranked below the general LC, the general constraint
will block allomorphy in both paradigm members and non-members. In this sense,
LCparadigm can never really be ranked below LC, it can either be subsumed under it
or promoted above it. Crucially, since there is no constraint which militates against

Stress Assignment in Serbo-Croatian

allomorphy in non-paradigm members, but tolerates it in paradigm members, such a

constellation cannot arise.
.. Signs of reanalysis? -ost getting High
In this short coda to this section, we briefly return to the exceptional case where
the LDN has the rising span although there are no phonological conditions for its
emergence: jednAAkOOst (related to jEdnaak equal).
Note that according to everything we have said so far, this de-adjectival should
act as hUmaanoostLDN (b) and display post-cyclic prosody. However, this does not
Our tentative account here is that in cases where the SDN is blocked for semantic
reasons, the LDN can actually stop being part of the derivational family of the base. In
other words, jEdnaak and /jednaak/ can stop being part of the input for jednAAkOOst.
Importantly, in the long run, if there are enough cases of this type, the -ostLDN may
be reanalysed as having lexically specified tone. Still, it is very hard to make the claim
that this reanalysis has already taken place, given the fact that there are absolutely no
cases where the H in -ost either overrides the lexical tone or draws the tonal span to a
short stem-final syllable.

. The Emerging Picture: Paradigms as the Place of SyntaxPhonology

The essence of our analysis of the interaction between prosody of S-C de-adjectival
nominalizations and their syntax and semantics is contained in the one phonological
constraint that we label LC(stress/tone)paradigm . This constraint is active only when
the input is a member of a paradigm, and only in these cases is there the effect of
favouring a computationally more expensive solution (it involves a lexical check, or
a richer input, containing also the prosodic specification of the surface base form).
A natural question is why this would be the case. Why should a more expensive
solution win, and why exactly for those inputs that are members of a paradigm? In
this section, we come up with two sets of possible answers.
Answer : The relevant property of members of a paradigm is their interpretive transparency. Paradigms are domains of (guaranteed) interpretive transparency.
Highly context-independent, the interpretation of members of a paradigm is strictly
derived from the semantic contribution of its components in combination with the
semantic contribution of the (complex word-internal) syntactic structure in which
they appear (Fodor and Lepore , Koontz-Garboden ). A word can only
have a non-transparent (shifted) lexical meaning if it has left the paradigm, or never
belonged to it. And vice versa: once a word leaves the paradigm it was a member of

Arsenijevic and Simonovic

(because the paradigm becomes unstable, or for some other reason), its meaning will
tend to become less transparent.
A system that marks whether a complex lexical item is a member of a paradigm
or not yields a higher productive capacity. Members of a paradigm which become
lexicalized, with an idiosyncratic meaning, no longer trigger a blocking effect on the
productive formation of their counterpart paradigm members: the two layers have
independent lives.
Answer : The reason lies in the relation of (a)symmetry. It has been argued that
grammar avoids symmetry (Kayne , Moro ). Asymmetric relations are
more easily linearized, and more easily assigned roles in the interpretive relation
of predication. In the case of S-C de-adjectival nominalizations, structures coming from a paradigm (i.e. SDNs) are characterized by a rich syntactic structure, in
which the stem and the suffix stand in an asymmetric relation. The stem is more
deeply embedded than the suffix, and according to the traditional models of Spellout should be spelled out to phonology first (Uriagereka , Chomsky ; also
Marvin in relation to a similar empirical pattern). In such a configuration,
phonology receives the input in an asymmetric relation, and preserves the prosody
of the element that it receives first (as if the first element were spelled out, as it
were, on its own, i.e. when the base adjective is pronounced), which is the stem
(see the schematic representation in ()). The structure of an idiosyncratic nominalization (LDN) is flat: it involves two items in a symmetric relation. In such a case,
phonology has no handle to resolve the symmetry. It proceeds without prosodic
assignment, and assigns tone and stress only at its own interface with phonetics, as a
last-resort strategy, insensitive to any information which is not contained in the lexical
The two answers above are not mutually exclusive. They may be taken as two
perspectives on the same phenomenon, or even as two forces conspiring to give it
its shape. In any case, the analysis we provide opens many interesting new questions
about the interaction between syntax and semantics on the one hand and phonology
on the other. For instance, while phonology may only stipulate the fact that certain
suffixes come with a lexically specified high tone and some not, our model gives us
a handle to explain why this is the case. Suffixes that most often or always appear in
the shallow type of syntactic structure are most often also the first syntactic head in
the word spelled out to phonology. Hence, if we take it that suffixes by default have no
tone specified in the lexicon, suffixes of the described type will most often still surface
with a toneone that is phonologically assigned. Suffixes that appear in more complex
It is possible that some asymmetry comes from the projecting nature of the nominalizing suffix, and
that this is why the final domain of the noun is targeted by the prosodic assignment. Arsenijevic ()
argues that the stress is indeed assigned to the suffix, and then as a consequence of a general rule in S-C
moves to the penult syllable of the entire word.

Stress Assignment in Serbo-Croatian

structures often enough, where they are not the first head in the word to be spelled
out, will also surface without a high tone often enough. It is expected, then, that the
former type of suffixes will eventually be stored in the lexicon with a high tone, and
those of the latter type will remain toneless. Being lexically encoded with a high tone
also marks for a particular suffix that it has a higher capacity to take the derived word
out of its stems paradigm, and to form a new paradigm from the derived word.

. Zooming Out: Across Suffixes and Languages

In this section we look at three windows opened by the dataset on which we have
based our analysis. In section .. we look at other nominalizing suffixes within
S-C, showing that there is a match between their prosody (which overrides that of
the base) and their semantics and syntax (LDN). Section .. looks at borrowed
abstract nouns in S-C, in an attempt to define an important constraint on borrowing.
Section .. discusses some possibly equivalent constellations in Dutch.
.. The suffixes competing with -ost
Zooming out, we are now looking at other suffixes available in S-C for deriving
abstract nouns from adjectives. In the picture we have outlined here, it would be
surprising to find suffixes which are less stress-affecting than -ost. And indeed, we
have found no such suffixes. All the other productive suffixes available in the lexicon
possess a lexical stress/tone which overrides the original stress of the adjective. In (),
we give some frequent nouns lexicalized with these endings.
() -OtA

dObAr good
skUUp expensive
tEEAk heavy
gOOrAk bitter

lep-OtA beauty
dobr-OtA goodness
skup-OcA expensiveness
fin-OcA sophistication
te-InA weight
gorc-InA bitterness

An interesting consequence of the application of these stress-attracting suffixes is that,

due to the general requirement of S-C to neutralize all pretonic length contrasts, all
long stems get shortened, which incurs extra violations of further LC constraints
e.g. LC(long). It therefore comes as no surprise that these suffixes are restricted to a
limited number of native adjectives and generally have LDN interpretations.
It should also be noted that whenever -ost versions are made of these adjectives
(marginal for most speakers), they always copy the stress of the stem and maintain its
length. In other words, the make-up of the derivational morphology in S-C conspires
to create the optimal circumstances for the distinction between the two classes of deadjectivals.

Arsenijevic and Simonovic

.. Borrowed de-adjectivals in S-C

So far, we have only dealt with native derivations. However, a sizeable number of
abstract nouns are borrowed and belong to the Latinate stratum of the lexicon (see
Simonovic ). Many such nouns have been borrowed alongside the related adjectives. Importantly, there seem to be no productive derivational patterns within the
Latinate stratum. We relate this to the fact that, typically, there is an -ost noun available
for every adjective as well.
() Adjective
apstrAkt-an abstract
elokvEnt-an eloquent
stEriil-an sterile

ost noun

Borrowed noun

stEriilnoostSDN ,


As our examples show, the -ost nouns are preferred in the SDN contexts and usually
restricted to them. In cases like stEriilan, we can see three levels of relatedness to the
stem: ost de-adjectival with the prosody copying that of the base adjective (stEriilnoost), ost de-adjectival with default prosody (sterIIlnOOst), and an unpredictable
derivation (sterilItEEt). In this sense, it is not clear whether there are any real SDNs
derived using foreign suffixes in S-C.
These examples show an interesting restriction on borrowing: derivationally
complex words tend not to be borrowed with their compositional structure. Rather,
they are first borrowed as simplex and then, in the case that some related words
from the source language are imported, the link between them can be reconstructed
(see Simonovic , in preparation). Given that borrowing is initiated in codeswitching, this fact matches the finding that semantic specificity is preferred
in code-switching (Backus ). This means that S-C speakers who borrowed
apstrAkcIja, elokvEncIja, and sterilItEEt did so with a specific and non-compositional
meaning. The moral of the story is that it is important not to rush into concluding
that derivational suffixes are getting borrowed across the board, and that what might
seem to be a class of borrowed derivations can actually turn out to have semantics
which points at its lexicalized status.
The three nouns derived from (or related to) stEriilan, arranged along the
SDN/LDN scale in the predicted way, show that language will make the relevant
distinction on any number of available items. We have already seen that a similar
distinction exists in a micro-class in Slovenian. In the following subsection, we will
see a similar case from Dutch.
.. Beyond S-C: Dutch
So far, we have only seen S-C facts and provided them with a S-C-internal account.
However, it is important to look at the facts from a crosslinguistic perspective. What

Stress Assignment in Serbo-Croatian

are the reasons for the dichotomy in S-C, and are there any comparable phenomena
in other languages? The two features of S-C which have conspired to produce the
dichotomy are the availability of a suffix which does not affect the prosody of the base
(unavailable in say French or Italian) and the ungrammaticality of the other common
means of expressing the trope meaning: Xs being Y:


/ Jovanovo bivanje
being.perf dangerous.nom *Jovans being.imperf

Jovans being dangerous

Before we move on to other languages we need to remind the reader that the
dichotomy is not omnipresent in S-C either. There seems to be a pressure to have
a single nominalization from every adjective. And indeed, in many cases, even S-C
speakers report allowing a single form from a single adjective for all functions even
where there is room for variation. This pressure is even stronger in Dutch, which
allows the conversion of infinitives, so that what is expressed in S-C as Jovanova
Opaasnoost Jovans dangerous-N can be expressed in Dutch as Jovans gevaarlijkheid Jovans dangerous-N, but also as Jovans gevaarlijk zijn Jovans dangerous
Despite the existence of the converted copula, there seem to be tendencies which
point at a distribution similar to that in S-C.
There is a stress-shifting micro-class of nominalizations which matches the S-C
pattern: two adjectives "dakloos homeless and "werkloos unemployed allow a stress
shift in heid nominalization (-heid being comparable to -ostsee below). This stress
shift tends to strongly correlate with LDN contexts.
() a. Daar komt hij weer met zijn "dakloosheid/??dak"loosheid/ "werkloosheid/??werk"loosheid.
Here he goes again with his homelessnessSDN / unemployednessSDN .
(Implication: He is homeless/unemployed.)
b. Daar komt hij weer met zijn ??"dakloosheid/dak"loosheid/ ??"werkloosheid/werk"loosheid.
Here he goes again with his homelessnessLDN / unemployednessLDN .
(Implication: Homelessness/unemployment, as a general issue, is his
favourite subject.)
Furthermore, Dutch also has Latinate nominalizations, which are much more
widespread than those in S-C. Booij (: ) notes that the suffixes -heid (native)
and -iteit (Latinate) are synonymous, but show an interesting asymmetry in distribution: whereas -iteit only combines with non-native bases, -heid can attach to both.

Arsenijevic and Simonovic

His examples are ab"surdheid and absurdi"teit (the base being ab"surd) for non-native
bases and ge"woonheid and gewoni"teit (from ge"woon ordinary) for native bases.
Two remarks are in order here. First, just like in S-C, the productive native suffix is
the one which leaves the stress pattern intact, whereas the borrowed one is stressattracting. Second, ab"surdheid, although dispreferred by most speakers, will surface
in contexts in which S-C has SDNs.
() de duidelijke absurdheid / ?absurditeit van die regel
the clear
absurdness ?absurdity of that rule
Also, for many adjectives for which heid-nominalization is held impossible and absent
from dictionaries, it actually surfaces in SDN contexts, in which the generally accepted
nominalization is blocked.
staat buiten kijf.
() a. De seksueelheid / seksualiteit van hun relatie
the sexual-heid *sexual-iteit of their relationship is beyond dispute
The sexualness/sexuality of their relationship is beyond dispute.
b. Hij moet per se
zijn speciaalheid / specialiteit bewijzen.
he must at any price his special-heid *special-iteit prove
He has to prove his specialness at any price.
In this section we have made an attempt to sketch the broader picture and initiate
some discussion on the general conditions on the prosodysyntax interactions of the
type discussed in this chapter.

. Conclusion
This chapter has examined the role that syntax and semantics play in the domain of
phonology, more precisely in prosody assignment in derived words. We discussed
an empirical domain, that of de-adjectival nominalizations, where a correlation can
be observed between the prosodic pattern of the derived noun and the types of
interpretation it may be assigned. The correlation is very strong in the majority of
cases, but it also faces certain restrictions, yielding considerable paradigm gaps: cases
where only one prosodic pattern is empirically confirmed, and the nominalization
expresses all the available types of meaning. Using a syntactic analysis that explains
the semantic effects, we proposed a model of the phonology of S-C and its interface
with syntax, in which both the phonologysemantics correlations and the gaps receive
a formal analysis and explanation. The central role in accounting for the correlation
effects is played by the tendency of Lexical Conservatism, in interaction with some
relevant syntactic and phonological properties, such as the structural complexity of
SDNs or syllable weight. In the last section we have made an attempt to show the larger

Stress Assignment in Serbo-Croatian

picture in which other factors, such as the availability of different shapes of affixes in
the lexicon and language contact, are situated.
It is our hope that we have opened many exciting windows and closed as few doors
as possible, and that future research will interact with this contribution in ways that
will enable further insights into the prosodysyntax interaction.

Allosemy, Idioms, and their Domains
Evidence from Adjectival Participles

. Goals
The goal of this chapter is to explore what special meaning is and how it arises
in word formation. On the basis of adjectival participles in Greek, we investigate
a distinction between idioms and the concept of contextual allosemy introduced
in recent work by Marantz (, ). Marantz argues that contextual allosemy
is the semantic analogue of contextual allomorphy and amounts to local polysemy
resolution within a specific context. The chapter investigates adjectival participles
from this perspective; as is well known, adjectival participles have been claimed in the
literature to give rise to special and/or non-compositional meanings (semantic drifts,
idioms), in contrast to verbal passive participles (Wasow and much subsequent
literature). Traditionally, this observation has been linked to the lexical vs. syntactic
nature of the two types of participles. In recent literature on English, however, the
lexical nature of adjectival participles has been argued to be an oversimplification
because there is a class of adjectival participles that behave syntactically like verbal
Different versions of this chapter have been presented at the nd GLOW Colloquium, University of
Nantes, April ; the Roots Workshop, University of Stuttgart, June ; the Workshop on Linguistic Interfaces (OnLI II), University of Ulster, December ; and Stuttgart University, June .
We would like to thank the audiences, especially Artemis Alexiadou, Winfried Lechner, Asaf Bachrach,
Orin Percus, Peter Svenonius, Gertjan Postma, Danny Fox, Alec Marantz, David Embick, Andrew Koontz
Garboden, Lisa Travis, Christina Sevdali, Raffi Folli, Heidi Harley, Antje Rodeutscher, Florian Schfer, and
Giorgos Spathas. An extensive discussion of this material which does not take into account the distinction
between allosemy and idiomaticity is presented in Anagnostopoulou and Samioti (forthcoming). The discussion of allosemy and idiomaticity is new and relies heavily on recent work by Alec Marantz (, ).
We are grateful to Alec Marantz for making his paper available to us, and to Roberta dAlessandro, who
gave us the opportunity to present this material at a mini-course held at Leiden University ( September
). We thank Roberta, the students, and participants for their input, and one anonymous reviewer for

Allosemy, Idioms, and their Domains

passive participles, qualifying as syntactic with respect to relevant diagnostics. It is not

our purpose here to settle the matter as far as English is concerned. Instead, we will
focus on Greek participles, hoping that our findings will shed some light on English
and other languages as well.
More specifically, we will investigate the interpretive properties of two morphologically and semantically distinct classes of Greek verb-based adjectival forms (Anagnostopoulou a, Alexiadou and Anagnostopoulou , Anagnostopoulou and
Samioti forthcoming) which correspond to two different groups of adjectival participles that have been identified for English by Kratzer (), Embick (), Bruening
(to appear), and others:
(a) Purely stative participles which lack the implication of a prior event (in Greek,
these are -tos participles); and
(b) resultative adjectival participles which denote a state resulting from a prior
event (in Greek, these are -menos participles).
The semantic and syntactic differences between the two forms have been treated as
resulting from the different height of attachment of participle affixes in Anagnostopoulou (a), Alexiadou and Anagnostopoulou (), and Anagnostopoulou
and Samioti (forthcoming). According to this proposal, the participle affix -tos
attaches directly to the root in stative participles (this is a case of what Marantz
calls inner affixation), while in resultative participles the affix -menos attaches above
a verbalizing head (this is a case of outer affixation). On the basis of a study of the
two types of participle, we will discuss the locality conditions on idiomaticity and
allosemy in participles displaying inner and outer affixation, and whether and how
these locality conditions can be derived from appealing to the notion of a Spell-out
domain at PF and LF (the phase). Drawing on the results of Anagnostopoulou and
Samioti (forthcoming), as these have been interpreted by Marantz (, ), we
will provide evidence in favour of the following points:
(a) Idiom formation and contextual allosemy should be kept distinct, as they show
different interpretive properties and are subject to different locality restrictions.
(b) Both -tos and -menos forms can yield idioms that are not necessarily related to
idioms formed by the corresponding verbs. The boundary for idiomatic interpretations for participles is Voice, just as has been observed for phrasal idioms.
Crucially, stative and resultative participles may give rise to idiomatic readings,
as long as external argument/agentivity features are absent (in accordance with
Marantz ; see Harley and Stone, Chapter below, for discussion of phrasal
idioms subject to the same condition and a reanalysis of potential counterexamples to this generalization).
(c) Greek stative -tos participles very often show special meanings, not in the
sense that they are non-compositional but in the sense that they do not have

Anagnostopoulou and Samioti

a meaning derived from the meaning of the corresponding verb. By contrast,

resultative -menos participles always have a meaning regularly related to the
verb meaning. This difference can be accounted for by appealing to the notion
of contextual allosemy. The boundary for contextual allosemy is the first contentful categorizing head, little a (the adjectivizer) in stative participles and little
v (the verbalizer) in resultative participles.
(d) There exist semantically empty categorizers (in the cases we discuss, verbalizers
which are not eventivizers), which are included in stative -tos participles and
do not count as boundaries for contextual allosemy of the root in the context
of little a.

. Background
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 morphophonologically 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 , Dubinsky and Simango , Horvath and Siloni ),
derivation vs. inflection (e.g. Anderson , Perlmutter , Spencer ), level I
vs. level II (e.g. Siegel , Kiparsky , Mohanan ).
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. For example, stative/adjectival participles
may show special morphology, while eventive/passive participles always show regular
() 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:
() a. the hung jury
b. The jury was being hung.

(Someone hung the jury.)

In English and other languages (though not universally; see Dubinsky and Simango
on Chichewa), the stative vs. eventive participle distinction correlates with a
split between derivation and inflection (Wasow , Marantz , , Horvath
and Siloni ). Stative participles are adjectives and passive participles are verbs, as
shown by very vs. very much modification in () and (). In English, very modifies
exclusively adjectives and very much verbs, as indicated by the contrasts between

Allosemy, Idioms, and their Domains

(b) vs. (c) and (b) vs. (c), respectively. The fact that the stative participle respected
is modified by very in (a) presents evidence that it qualifies as an adjective. On the
other hand, passive respected in (a) is modified by very much, qualifying as a verb.
() a. Your family was very respected.
b. John is very fond of your family.
c. John is very much fond of your family.
() a. Your family was very much respected by the neighbours.
b. John very much respects your family.
c. John very respects your family.



Wasow () 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;
see Horvath and Siloni 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.
Marantz (, ) and Arad (, ) argue that the difference between
regular and irregular affixation can be derived in a syntactic approach to word
formation by appealing to the point in the derivation at which an affix is merged
with its complement. They propose that there are two domains for word formation
delimited by a category-defining head, as shown in (). Attachment of x directly to
the root, as in (a), leads to irregular word formation, while attachment above the
category-defining heads (little v, a, n), as in (b), leads to regular word formation:
() a.




n, v, a

n, v, a

On this view, there is a split between inner and outer morphology:

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 : )
But cf. Strong Lexicalist approaches such as Levin and Rappaport (), 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 () for evidence against this view and Marantz () for discussion of Wasows
analysis, as compared to Levin and Rappaports.

Anagnostopoulou and Samioti

The main properties associated with inner vs. outer affixation are summarized in ().

Inner affixation

Outer affixation

Potential special form and

special meaning
Attaches inside morphology
determining lexical category

Predictable form and predictable meaning

Attaches outside morphology
determining lexical category

The special role of category-defining heads is linked to the hypothesis that they are
phase heads (Chomsky , ) which delimit cyclic domains leading to the
semantic interpretation and phonological Spell-out of the chunk of syntactic structure they merge with, i.e. the root or a more complex constituent formed by inner
affixation. Once a chunk of structure is shipped off to PF and LF for pronunciation/interpretation, its phonological Spell-out and semantic interpretation cannot be
altered. As a result, further affixation cannot access the properties of the Root. The
empirical generalization this hypothesis is based on will be called here the idiosyncrasy generalization in () and is linked to the phase-based hypothesis summarized
in ():
() The idiosyncrasy 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.
() The phase-based hypothesis
Roots are assigned an interpretation in the context of the first categoryassigning head/phase head merged with them, which is then fixed throughout
the derivation.
Arad (, ) presents extensive evidence from Hebrew denominal verbs supporting () and the inner vs. outer architecture.
Marantz () argues that inner vs. outer affixation can correctly derive the differences between adjectival passives, which are taken to involve inner affixation, and
verbal passives, which are taken to involve outer affixation. According to Marantz
(: ):
Merger with root implies properties sensitive to properties of the root:
Negotiated (idiosyncratic) meaning of root in context of morpheme.
Semi-productivity (better with some roots than others).
Meaning of construction must depend on root semantics independent of argu-

ment structure.
Corollary of the above: cannot involve the external argument of the verb.

Allosemy, Idioms, and their Domains

Merger above a category-determining morpheme implies insensitivity to idiosyncratic

properties of the root:

Compositional meaning predicted from meaning of stem.

Complete productivity.
Meaning of structure can involve an operation on argument structure.
Can involve the external argument of a verb.
(Marantz : )

However, the traditional view of the verbal/adjectival passive distinction has proved
to be too simplistic and has been argued to be wrong, most notably by Embick (),
McIntyre (), and Bruening (to appear). Based on Kratzer (, ), Embick
argues that there are two subclasses of adjectival participles in English: pure stative
participles which lack implications of a prior event, and resultative participles which
denote a state that results from a prior event. In recent work, McIntyre () and
Bruening (to appear) go one step further. They argue that many distinctions that
have been claimed to exist between verbal and adjectival participles in English and
are summarized in () (the summary from Bruening, to appear; Bruenings ()) are
empirically wrong:

Adjectival passives

Verbal passives

Category: Adjective
Feed word formation
Cannot be fed by ECM/raising
No external argument present

Category: Verb
Do not feed word formation
Can be fed by ECM/raising
External argument syntactically
Obligatory disjoint reference
Cannot be formed from
Can affect applied arguments
Interpretation is regular
No idioms that are not part of
the input
No missing input

. Permit reflexive interpretation

. Can be formed from unaccusatives
. Cannot affect applied arguments
. Semantic drift
. Idioms that are not part of the input
. Missing input

Bruening (to appear) argues that differences , , and in () are not real, and that
adjectival passives in English do not differ from verbal passives in the presence of
an external argument, reflexive interpretation, and, most importantly, their ability to
involve NPs that are not semantic arguments of the underlying verb (ECM/raising
and also pseudopassives). This means that adjectival passive formation in English
must be a syntactic and not lexical process, contra what is often suggested in the
literature. On the other hand, properties in () indeed distinguish adjectival
from verbal passives in English, and have to be dealt with in a syntactic analysis of
adjectival passives. Crucially, there are differences in interpretation between adjectival

Anagnostopoulou and Samioti

and verbal participles, even though both can be claimed to be formed by outer affixation. We will postpone a more detailed discussion of for later on (section .),
after having investigated the interpretive properties of adjectival participles in Greek
from the perspective of Marantzs (, ) theory of interpretive domains in
word formation.

. Greek -menos and -tos Adjectival Participles: Outer and Inner

Next to adjectives, Greek has two adjectival constructions: the participle in -menos
and what traditional grammars call the verbal adjective in -tos.



opened / open

They have a similar function to adjectives, i.e. they appear in attributive and predicative positions:
() a. to
the.neut.sg.nom open-t-neut.sg.nom window
the open window
b. to
the.neut.sg.nom open-men-neut.sg.nom window
the opened window
() a. To parathiro ine anix-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.
-menos and -tos participles show a number of semantic and syntactic differences
discussed in Markantonatou et al. (), Georgala (), Kordoni (), Anagnostopoulou (a), Alexiadou and Anagnostopoulou (), among others. The
most fundamental one is that -menos participles denote a state resulting from a prior
event, while -tos forms lack event implications. They denote what has been called a
characteristic state by Markantonatou et al. (). Consider the examples in ():
() a. I patates ine tiganis-men-es.
the potatoes are fry-men-f.pl.nom
The potatoes are fried.
Passives are verbal in Greek and therefore there is never an ambiguity between verbal and adjectival
passives in Greek, unlike English.

Allosemy, Idioms, and their Domains

b. I patates ine tigani-t-es.

the potatoes are fry-t-f.pl.nom
The potatoes are fried.
(a) 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, (b) 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 (), based on Embick
(: ). 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 (), -tos forms expressing characteristic states are licit in such a context,
while -menos participles are not, due to their eventive subcomponent:
() a. I porta
anix-t-i / anig-men-i.
the door.nom built.NAct.sg open

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.
c. Eftiaksa tis patates tigani-t-es / tiganis-men-es.
made I the potatoes 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 (a) and can license by-phrases and instrument PPs (c), while
-tos participles cannot (compare (b) to (a), and (d) to (c)):
() 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
c. 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.
d. O tixos ine xtis-t-os me mistri / apo ton ergati.
the wall is built
with trowel by the worker
Anagnostopoulou (a), following Kratzer (), furthermore points out that
-menos participles can denote target and resultant states (Parsons : ). The
former describe states that are in principle reversible; the latter introduce states that

Anagnostopoulou and Samioti

hold forever after the event that brings them about. Target state participles in () are
compatible with the adverbial akoma still, while resultant state participles in () are
incompatible with it:
() Ta pedhia ine akoma krimena.
the children are still hidden


theorima ine akoma apodedigmeno.

the theorem 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 or agentive adverbials. As shown by (), by-phrases and instrument phrases are
incompatible with akoma still:
() a. Ta lastixa ine ( akoma) fuskomena apo tin Maria.
inflated by the Mary
the tyres are (*still
The tyres are still inflated by Mary.
b. Ta lastixa ine ( akoma) fuskomena me tin tromba.
inflated with the pump
the tyres are (*still
The tyres are still inflated with the pump.
Moreover, note that there are two types of manner adverbial: (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 (), while adverbs denoting the visible result (result-oriented)
are compatible with it ():
() To thisavrofilakio itan ( akoma) prosektika anigmeno.
cautiously opened
the safe
was (*still
The safe was still cautiously opened.
() Ta malia mu ine (akoma) atsala xtenismena.
the hair my is (still
sloppily combed
My hair is still sloppily combed.
On the basis of these considerations, Alexiadou and Anagnostopoulou () proposed in a Distributed Morphology (DM)-based decomposition framework that the
three types of Greek participle have three different structures, depending on the height
of attachment of the participle morpheme. 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:
() [AspP Asp X]

(where X= Root, vP, or VP)

Allosemy, Idioms, and their Domains

-tos participles lack event implications (they are licensed after verbs of creation in
(), they do not permit result-oriented modification) and agentivity (no agentoriented modification, no by-phrases or instruments). Alexiadou and Anagnostopoulou () take this to mean that they involve root attachment:

(root attachment of Asp)




-menos target state participles which include the implication of an event (they are not
licensed after verbs of creation in (), they license result-oriented modifiers) but lack
agentivity (no agent-oriented modification, no by-phrases or instruments) involve v
attachment (v is taken by Alexiadou and Anagnostopoulou to be the eventivizing



(v attachment of Asp)

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):

(Voice attachment of Asp)





Important for present purposes is the proposal that -tos participles instantiate innercycle attachment while -menos participles instantiate outer-cycle attachment:

Root cycle

Outer-cycle attachment



Anagnostopoulou and Samioti

This leads us to expect that -tos forms will show properties of inner affixation and menos forms of outer affixation. Prima facie evidence in favour of the inner vs. outer
division in () 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 ():
() a.






worked out
shampoo bathed
hurt, tampered

. -tos Attaching to Verbalizers That Are Not Eventivizers:

Inner or Outer Affixation?
As discussed by Anagnostopoulou and Samioti (forthcoming), the inner vs. outer
affixation analysis of -tos and -menos participles appears to be challenged by a class
of -tos characteristic state participles that lack event implications but nevertheless
contain an overt verbalizing element. Greek productively employs verbalizing suffixes which have been analysed as root verbalizers by Alexiadou (, ; see
Giannakidou and Merchant , Ralli for discussion of these elements):
() Root-verbalizing elements
Greek: -iz, -on, -en/an, -ev, -az, -a

(Alexiadou , )

Note that out of , verbs listed in the online Lexicon of Anastasiadi-Simeonidi

(http://www.komvos.edu.gr/dictionaries/dictOnLine/DictOnLineRev.htm) that have been checked
so far, verbs form both -menos and -tos participles, , form only -menos participles, and verbs
only -tos participles. These numbers need to be checked more carefully.
As argued for in detail in Samioti (, in progress) and Anagnostopoulou and Samioti (forthcoming), -tos participles display outer-cycle attachment when they denote ability/possibility, in which case they
have a meaning comparable to -able adjectives in English. An example is given in (i):

(i) a. Afti i dikaiologia ine pistef-t-i.

this the excuse
is believe-t-f.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.
We will come back to these later on (section .), when we discuss the question of idiomaticity and the
role of Voice.
Another case where -tos attaches above little v and hence should be analysed as an instance of outer
affixation is with negated perfect of result participles (corresponding to English participles with un- prefixation). These will not be discussed at all here. The reader is referred to Anagnostopoulou and Samioti
(forthcoming) for detailed discussion of -tos participles showing outer affixation.

Allosemy, Idioms, and their Domains

These elements are present on the verbs in ():

() a. aspr-iz-o


b. pag-on-o


c. diaplat-en-o

become sick

d. sten-ev-o


e. diav-az-o

split, share

f. pul-a-o


As expected by the outer analysis of -menos participles in (), verbalizers (and their
allomorphs as in the (c), (f) examples below) are systematically present in -menos
participles, providing morphological evidence for outer-cycle attachment:
() a. aspr-iz-menos


b. pag-o-menos


c. diaplat-i-menos


d. sten-e-menos


e. diav-az-menos

split, shared

f. pul-i-menos


Root verbalizers are often disallowed in -tos participles, a fact which was taken by
Alexiadou and Anagnostopoulou () as evidence for root attachment on -tos
in ():
() a.

















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-istos whitened vs. aspr-os white, sten-ef-tos tightened vs. sten-os tight. Alexiadou and Anagnostopoulou
() analyse this as an instance of blocking. See below for discussion.

Anagnostopoulou and Samioti

e. diav-as-tos


*split, shared

But, quite unexpectedly, many Greek characteristic state -tos participles do include
such verbalizing elements:
() a.

steaming hot
koudoun-is-tos ringing
vathoul-o-tos hollow
magir-ef-tos cooked
angali-as-tos embraced
evodi-as-tos fragrant

koudoun-iz-o ring (a bell)
vathoul-on-o hollow out
magir-ev-o cook
angali-az-o embrace
evodi-az-o be fragrant

Despite the presence of verbalizers, the -tos participles in () do not have event
implications (they denote characteristic states), and they do not license manner modification, agent PPs, or instruments:
() a. To fagito ine kala / prosektika magir-ef-t-o.
the food is well carefully cooked
b. To fagito ine magir-ef-t-o apo tin Maria.
the food is cooked
by the Mary
c. Ta fita ine fit-ef-t-a me diaforetika ergalia.
the plants are planted with different instruments
The participles of the type illustrated in () 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 leads to the question of how to
An anonymous reviewer suggests that the data in () are not sufficient to justify the separation of vC
and vE. According to the reviewer, one could account for these facts by making a parallel with subject demotion accounts of passives and put -tos in the agent position. The reviewer furthermore points out: Splitting a
category in two is costly and shouldnt be done so easily. We are fully aware of the theoretical complications
entailed by the separation of vC and vE we are arguing for here (as well as in Anagnostopoulou and
Samioti, forthcoming), and we do not feel that we have done this so easily. Concerning the data in (),
-tos participles with vC do not permit any kind of modification, not just agent-oriented modification.
For example, they do not permit the result-oriented modifier kala well in (a) which has been argued
for by Anagnostopoulou (a) and Alexiadou and Anagnostopoulou () to be introduced at the v
level and not at the Voice level (see Anagnostopoulou a, Alexiadou and Anagnostopoulou , and
Anagnostopoulou and Samioti, forthcoming, for more details). Moreover, the -tos participles with a vC do
not have event implications, as evidenced by tests like the Embick-based test presented in (), and they
denote characteristic states, just like their counterparts without verbalizers. We believe that Greek readers
would never disagree that the -tos participles with vC lack event implications; it is a very strong intuition.
Non-Greek readers can perhaps get a sense of what the interpretation of such participles is via the English
participles in the first column of () and () discussed in section ..

Allosemy, Idioms, and their Domains

interpret the phase-based hypothesis in (): is the verbalizer vC or the eventivizer vE

the phase head delimiting the domain of inner vs. outer affixation? In section . we
will argue on the basis of allosemy that the latter is correct: the semantically contentful
head vE introducing an event variable is what qualifies as a phase head.
What are the conditions under which vC verbalizers are present in -tos participles,
and when does -tos attach directly to roots? In order to answer this question, we will
adopt Harleys (), Levinsons (, ), and Marantzs () proposal that
roots can be further classified in terms of basic ontological types such as individual,
state, and event. We will follow Marantz (: ) in taking:
roots [] as belonging to semantic types associated with the category heads n, a, and v. If the
usual interpretation of v is to introduce an event variable, of a to introduce a state variable, and
of n to introduce an entity variable, then verbal roots are those that modify events, adjectival
roots those that modify states, and nominal roots those that modify entities.

On the basis of this classification, we can formulate a number of generalizations

concerning the environments where vC verbalizers are present in -tos participles:
Generalization I: -tos does not combine with Rootthing/nominal . It combines with
Rootthing/nominal + verbalizer. When roots are interpreted as modifying an entity,
they can combine directly with nominal inflection forming a noun, as shown in the
first column of (). -tos does not directly combine with such roots. They first have
to become verbal by combining with a verbalizer, and then -tos attaches to the root +
verbalizer complex:
() a.


foamN afr-iz-o
foamV afr-is-tos
steamV axn-is-tos
steaming hot
steamN axn-iz-o
screwN vid-on-o
screwV vid-o-tos
ringN koudoun-iz-o ringV koudoun-is-tos ringing
buttonN koumb-on-o buttonV koumb-o-tos buttoned
coverV skep-as-tos
roofN skep-az-o

Interestingly, there are even some cases in which the denominal verb formed by the
Rootthing + verbalizer is deviant (second column in (), i.e. such expressions cannot
be verbs), while the corresponding -tos participle is perfect:
() kamban-a bellN ??kamban-iz-o bellV kamban-is-tos sounding like a bell
An anonymous reviewer raises the question: if roots are categorized what makes them roots? Our
reply to this concern is that roots are not categorized, but rather they are classified according to the
basic ontological types individual, state, event. This assumption is not only necessary for the narrow
goal of understanding the distribution of vC in -tos participles (see also Anagnostopoulou and Samioti
forthcoming) but, more broadly, in order to explain why e.g. the interpretation of the noun dog (in the
dog) feels more literal and natural than the interpretation of the verb dog (in to dog). See Marantz ()
for a characterization of this type of polysemy as a process of type-shifting allosemy. See Levinson (,
) for extensive argumentation based on this type of polysemy that the ontology of roots must allow for
at least predicates of individuals, predicates of states, and predicates of events.
See Alexiadou () for discussion of such formations.

Anagnostopoulou and Samioti

This shows that there is a formal requirement for -tos to attach to a verbal element,
even when this verbal element is not, at the end, licensed as a description of an
Generalization II: As shown in (), it is often the case that -tos does not combine with
Rootstate + verbalizer because there is an adjective available blocking the -tos form (as
pointed out by Alexiadou and Anagnostopoulou ).
() a.

aspr-iz-o whiten aspr-iz-men-os

mavr-iz-o blacken mavr-iz-men-os
kitrin-iz-o yellow kitrin-iz-men-os
prasin-iz-o green prasin-iz-men-os
kathar-iz-o clean kathar-iz-men-os
megal-on-o grow megal-o-men-os

aspr-os/i/o white is-tos

mavr-os/i/o black is-tos
kitrin-os/i/o yellow is-tos
prasin-os/i/o green is-tos
kathar-os/i/o clean is-tos

In principle, however, the formation Rootstate + verbalizer + -tos is possible. And

indeed, such forms do exist, sometimes with specialized uses. For example, the -tos
form in () is used only for food:

kokin-iz-o redden

kokin-os/i/o red

kokin-is-tos with a red sauce

A few other -tos participles with a verbalizer attaching to a Rootstate (adjective or

adverb/ preposition) are provided in (), with a canonical compositional meaning:
() a.



round/ rounded
misty/ blurred

Crucially, -tos does not directly attach to Rootstate , i.e. to roots that can combine with
adjectival inflection.
Generalization III: -tos directly attaches to roots which can be characterized as
Rootevent , as shown in (). These forms involve roots that are understood as modifying particular types of events. Like nominal and adjectival roots, which directly
combine with nominal and adjectival inflection to yield nouns and adjectives, verbal

An anonymous reviewer asks: What does it mean to be of a verbal category, but not to involve a
description of an event? And why would such a description not be licensed? These questions are exactly the
questions one can raise once a dissociation between vC and vE is introduced. Since vC is semantically empty,
then a form containing vC can, in principle, be verbal without involving a description of an event. Note that
the crosslinguistic distribution of stative expressions shows that it is not always the case that there is oneto-one correspondence between the category of an expression and its ontological type. In many languages,
stative predicates e.g. expressing possession or psychological states are categorized as verbs, while in other
languages they are categorized as adjectives or nouns.

Allosemy, Idioms, and their Domains

roots can directly combine with verbal inflection to yield verbs, as shown in the first
column of ().
() a.

ftiax-n-o make
per-n-o take
klev-o steal
plek-o knit
klin-o close

ftiax-tos made
par-tos taken
klef-tos stolen
do-tos given
plek-tos knitted
klis-tos closed

On the basis of the three generalizations examined so far, we can conclude that -tos
must attach to elements that are verbal. When roots modify an event, as in (), they
qualify as verbal in the relevant sense and -tos combines with them directly. When
roots modify an entity or a state, however, verbalizers are required before -tos attaches
to them. Crucially, these verbalizers do not introduce an event variable, and therefore
-tos participles never have event implications (and the syntactic properties associated
with an event interpretation).

. vC, vE, Voice, and Idiomaticity

We are now in a position to look at the question of the interpretation of Greek
adjectival participles from the perspective of the idiosyncrasy generalization () and
the phase-based hypothesis (). As they stand, () and () raise the question of how
to interpret idiosyncratic meaning and phase head, respectively. In this section, it
will be seen that if idiosyncratic meaning is equated with idiomatic meaning then
neither vC nor vE are boundaries for idiomaticity. The boundary in Greek adjectival
participles is Voice, as in the earlier proposal of Marantz (). See Harley and Stone,
Chapter below, who argue that exactly the same domain restriction applies to
phrasal idioms.
.. Neither vC nor vE are boundaries for idioms
We assume a working definition of idioms along the lines of Svenonius (: ) as
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, i.e. vC) and little v (the head contributing eventiveness,

Two comments on (): (i) Sometimes verbal roots form verbs via the formative -n- plus inflection,
as in (a,c). (ii) As pointed out by Antje Rodeutscher (p.c.), it is interesting that the roots in klin-o close
and anig-o open are classified as verbal and not adjectival in Greek.

Anagnostopoulou and Samioti

licensing arguments, etc., i.e. vE) introduced above, 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 vC in -tos participles yields such an effect.
In other words, are there significant differences between -tos participles with a
verbalizer and -tos participles without a verbalizer with respect to idiomaticity,
in comparison to the corresponding verbs?
(b) Assuming, alternatively, that it is the presence of eventive vE which defines
a cyclic domain for interpretation (phase), it should be tested whether the
presence of an eventive little v fixes meaning in such a way that affixes attaching
above it always lead to predictable interpretations. 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 stative -tos
participles does not correlate in any way with (non-)idiomaticity. Participles 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 () and ()
() Stative -tos participles showing direct attachment of -tos to Rootevent


a. sfing-o
b. ftin-o
c. klin-o

lit. spitted
lit. closed

Idiomatic interpretation
of participle only
careful with money
spitting image

() Stative -tos participles showing attachment of -tos to Root + verbalizer

a. kol-a-o
b. xtip-a-o
bang, hit, whip
c. xon-ev-o


Idiomatic interpretation
of participle only

lit. glued
lit. whipped
no lit. meaning

close friend
inside the wall

Allosemy, Idioms, and their Domains

d. karf-on-o

no lit. meaning

very fast/direct

The answer to the second question is once again negative. -menos participles may have
idiosyncratic readings, just like the -tos participles in () and ():
() striv-o

stri-menos geros
crotchety old man

strif-to tsigaro
lit. twisted (rolled) cigarette

Importantly, the -menos participle in () only has the idiomatic (non-literal) 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 () and () are
understood as referring 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
() and () 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 (). And, in fact,
there is robust evidence that there is no necessary correlation between the meaning of
verbs and the meaning of the corresponding -menos participles. One can distinguish
between the following three cases:

Idiomatic verb

Non-idiomatic participle

lit. clean
idiom. kill

only lit. cleaned

() a. Ton katharisan xtes

to vrady.
him cleaned.sg yesterday the evening
They killed him yesterday evening.
b. ?Aftos o anthropos ine katharismenos.
this the man
is cleaned
This man is cleaned.

Non-idiomatic verb

Idiomatic participle

only lit. pull

lit. pulled
idiom. far-fetched

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.

Anagnostopoulou and Samioti

() Verb
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
Janis pulled the argument.
() Participle
a. ?I porta ine trav-ig-meni.
the door is pulled
The door is pulled.

(Lit. interpretation)

b. To epixirima ine trav-ig-meno.

the argument is pulled
The argument is far-fetched.

(Only idiom. interpretation)

Idiomatic verb: one meaning

Idiomatic participle: another

lit. digest
idiom. like

lit. digested
idiom. understood

() Verb literal, participle literal

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
a. O Janis den xonevi ta mathimatika.
the Janis not digests the maths
Janis dislikes mathematics.
b. Ta mathimatika den ine xone-mena.
the mathematics not are digested
Maths is not understood.

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.

Allosemy, Idioms, and their Domains

To conclude, in this section we have argued that the hypothesis () based on the
generalization () is not a hypothesis about idioms, regardless of how exactly we
interpret hypothesis (), i.e. whether we take vC or vE to be the relevant phase head.
.. Voice is a boundary for idioms
While idiomatic interpretations are not blocked by vC or vE, the presence of Voice
systematically imposes literal/predictable interpretations on -tos and -menos forms.
This can be seen in a number of cases:
(a) As briefly mentioned in footnote , there is a class of -tos participles denoting
ability/possibility, roughly corresponding to English -able adjectives, as shown by the
translations below (Markantonatou et al. ; see Samioti , in progress, and
Anagnostopoulou and Samioti, forthcoming, for extensive discussion):
() a. Afti i dikaiologia ine pistef-t-i.
this the excuse
is believe-t-f.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, theat-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
(, in progress) argues that -tos in ability participles attaches above Voice. Evidence
for this comes from the fact that they license agent (a) and instrument PPs (b),
and agent-oriented adverbs of the type found in English middles ():
() 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.
() 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 (). According to
Lekakou, the type of voice employed in dispositional middles is a variant of the
passive voice, unlike English and other languages. An implicit external argument is

Anagnostopoulou and Samioti

present and, therefore, by-phrases and instruments are licit (see Lekakou for
detailed argumentation). Crucially, -tos participles denoting ability/possibility never
have idiomatic readings. All participles belonging to this class have exactly the same
meanings as the corresponding verbs, and their interpretations are compositional. We
believe that this is not a coincidence but relates to the fact that these forms contain
Voice. The obligatory presence of Voice does not permit idiomatic interpretations of
these participles.
(b) Greek has a productive process of yielding ability/possibility -tos participles from
the verbal root combined with the adjectival prefix aksio- worth- (see Samioti ,
in progress, and Anagnostopoulou and Samioti, forthcoming, for a more detailed
() a.


worth loving
worth admiring
worth studying
worth despising

Semantically, such formations implicate an external argument (a generic/

impersonal argument corresponding to English one), which points to the presence
of a (middle) Voice head similar to the one found with the -able/-tos forms in ().
Interestingly, whenever the prefix aksio- combines with an idiomatic participle, the
non-compositional meaning is lost, as shown in ():
() a. trav-ig-menos
lit. pulled
idiom. far-fetched
b. stri-menos
lit. twisted
idiom. crotchety

only lit. worth pulling
only lit. worth twisting

This constitutes a further piece of evidence that the presence of Voice blocks idiomatic
interpretations of participles.
(c) Agentive adverbs like deliberately and agent-oriented manner adverbs like
carefully systematically prevent idiomatic interpretations. See Alexiadou, Anagnostopoulou, and Schfer () among many others for discussion of these as diagnostics for the presence of an agentive Voice head:
() a. trav-ig-menos
lit. pulled
idiom. far-fetched
b. stri-menos
lit. twisted
idiom. crotchety

prosektika/skopima travigmenos
only lit. carefully/deliberately pulled
prosektika/skopima strimenos
only lit. carefully/deliberately twisted

Allosemy, Idioms, and their Domains

(d) Agentive and instrument PPs, which also modify an agentive Voice head present
in the structure, have the same effect of forcing compositional interpretations in ():
() stri-menos jeros

idiom. crotchety man

stri-menos apo kapion/ me kati
only lit. twisted by someone/
with something

The effects of Voice illustrated in (a)(d) above support Marantzs (, ) generalization that the syntactic head that projects an agent defines a locality domain for
idiomatic meanings of participles, as stated explicitly in prediction (b):

boundary for domain of special meaning

head projecting agent
(Marantzs : ())

() Predictions: (Marantzs : ()):

a. No idioms with fixed agents (root in agent position, context for special
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
In Marantz (), the difference between verbal and adjectival participles with
respect to idiomaticity, i.e. (b), is linked to agentivity. But notice that both eventiveness and agentivity are properties attributed to the same head v in (), and one
could equally well assume that eventiveness is the relevant factor. This is the direction
taken in Marantz (, ), who links the difference between verbal and adjectival
passives to the verbal/eventive vs. adjectival/stative nature of English participles. In a
decomposition system where Voice is separated from a lower little v, this amounts to
claiming that little v is the relevant head. The Greek data discussed in this section,
however, show that Voice and not little v is the crucial head for idioms.

. vE, vC, and Allosemy

In a recent paper, Marantz (, ) revisits () and () and argues that the first
contentful categorizing head (n, a, v) is a phase head which serves as a context for
fixing the meaning of a polysemous root at LF (what he calls allosemy), in a fashion
parallel to the local determination of contextual allomorphy at PF. This leads to a
different interpretation of what idiosyncratic meaning is in (), and to a dissociation
between the domains for allosemy and idiomaticity. Marantz () explicitly states

Anagnostopoulou and Samioti

that we should separate allosemy from idiomaticity, and that the domain for idioms
is larger than the domain for allosemy.
In this section, we will argue that the concept of allosemy correctly characterizes
a set of systematic meaning differences between -tos and -menos participles, which
relate to the absence in the former and the presence in the latter of the contentful
head vE.
.. Background: locality conditions on contextual allomorphy
Marantz (, ) assumes as background Embicks () local theory for contextual allomorphy:
() Contextual allomorphy = the choice of the form of a morphemethe appropriate allomorphwithin a local environment.
According to this theory, local contextual allomorphy is sensitive to phases: once a
designated phase head is merged with a complement, it triggers its Spell-out, and a
higher head can no longer serve as context for the choice of an allomorph located
below that phase head.
A potential problem for this analysis is presented by cases like the English past
tense, where contextual allomorphy does cross the boundary of a phase head. Consider (), where past tense triggers allomorphy of the root (and vice versa, the root
serves as a context for the choice of an irregular past tense morpheme). Crucially,
little v/Voice, which are assumed to be phase heads, intervene between the root and
the Tense head:

TEACH + + past = taught (irregular /t/ or null allomorph of past)
Marantz (, ) points out that there are two crucial locality issues raised by
examples like ():
(a) The first issue concerns what has already been discussed above. If Voice and little
v are phase heads, then why is the root not spelled out in the complement domain of
Voice/v prior to Merger of T? If this were the case, then T would not serve as context
for root allomorphy. In order to resolve this problem, Marantz proposes that each
root be adjoined to the head that types it as a lexical category, prior to any stage of
See esp. Marantz (: ): Apparent counterexamples in the literature to this claim of the local
determination of root meanings conflate idiomatic special meanings with polysemy. [] Clearly complex
words can acquire special meanings and uses; like phrases, complex words can be idiomatic in the sense
of conveying meanings not computable from the meanings of their parts. Marantz () is confusing if
not simply wrong in conflating the question of idiom with the notion of special meaning or meaning
choice associated with polysemy. For the issue of root [] polysemy, the relevant locality domain for
fixing meaning appears to be the phase, while for idioms, the domain is clearly larger []. Idioms,
then, involve a type of meaning thats built on top of polysemy resolution []. Idiom formation seems
constrained to the domain of an external argument.
Implicitly, Voice must also be taken to count as such a head, perhaps in the spirit of the notion of
Extended Projections (Grimshaw , , van Riemsdijk ).

Allosemy, Idioms, and their Domains

the derivation in which either the category head or the root might be phonologically
interpreted. As a result of head movement, a verbal root ends up in the same Spell-out
domain as Voice and little v and is not spelled out in the complement domain of little v
(only the arguments of the root are). Therefore, the Voice/v complex does not interfere
with Tense serving as the context for Vocabulary Insertion (VI) at the root (and vice
versa): all three heads (rootvVoice) are spelled out at the same time, namely in the
complement domain of the next phase head higher up, C.
(b) The second issue concerns adjacency, namely the requirement that the root
needs to be adjacent to the conditioning environment T for VI. Crucially, phonologically null heads do not block adjacency between the root and the conditioning
environment for stem allomorphy, while spelled out little v heads do. Compare (),
which contains an empty head, to (), with the overt v head -ize. The presence of -ize
blocks allomorphic choice at the root conditioned by Tense:

QUANTUM + ize + past = quantized, quintized, quantumized, etc.

.. Contextual allosemy and locality

Marantz () argues that the semantic analogue of contextual allomorphy exists,
subject to the same locality conditions. Just as there is special form choice sensitive
to context at PF (contextual allomorphy), there also exists special meaning choice
sensitive to context at LF. This is what he calls contextual allosemy: the possible
conditioned semantic Spell-out of a morpheme at the LF interface (Marantz :
). A starting example for root allosemy is provided in () (Marantz : ):
() to house
The particular interest of this example is that it shows both contextual allomorphy
and contextual allosemy. The verb to hou[z]e from the root HOUSE shows special
voicing of the final fricative in the environment of the little v head and has a meaning
such that neither a literal house nor even a literal container are present. On the
other hand, it is possible to form a verb from the noun house, with a meaning do
something with houses. This verb has the literal nominal reading of the root HOUSE
as well as the default phonological form with a voiceless fricative /s/. This treatment of
() offers an explanation in terms of allosemy to Kiparskys () generalization that
root-based denominal verbs do not entail the existence of the corresponding nouns
while noun-based ones do (see Arad , for extensive discussion of Hebrew
based on this). Direct attachment of little v to the root permits this root to retain the
range of meanings associated with HOUSE, while a specific choice is made when little
n is introduced. Further affixation above little n cannot alter this choice by going back
to the interpretations available at the root level. Two more examples of allosemy are
() and ():

Anagnostopoulou and Samioti

() a. GLOBE = sphere/sphere-like or the world

b. a globe (context: little n) = the root can mean spherical (a glass globe
Christmas ornament) or this (whole) planet. No choice made.
c. global = little a chooses the world/planet reading of the root.
d. globalize = the planet reading is preserved. No make into a sphere reading.
() a. NOVEL = artistic work or physical volume
b. a novel = both meanings preserved in the context of little n.
c. novelize = little v chooses the artistic work meaning of the root.
In both examples, there is a point in the derivation where a specific meaning is chosen
(when the little a head is introduced in (c) and when the little v head is introduced in
(c)). Once this happens, there cannot be any meaning shift back to an interpretation
available at a previous stage (e.g. sphere in () or physical volume in ()). See
Marantz () for extensive discussion of the absence of this pattern, which he
calls flip-flopping, as an argument in favour of local allosemy and against potential
alternative theories that would not postulate such a strict and local procedure to the
process of polysemy resolution.
Coming to the issue of locality, the null hypothesis is that contextual allosemy is
subject to exactly the same conditions as contextual allomorphy, namely: (a) only
material in the same Spell-out domain can condition allosemy (phase-based locality), and (b) this material must be semantically adjacent to the morpheme whose
semantic value is determined (adjacency). The semantic analogue of a phonologically
overt head like -ize in () would be a semantically overt, contentful head, while the
semantic analogue of a head like the empty v in () would be a semantically empty,
contentless head.
.. -tos vs. -menos participles and allosemy
Looking at -menos and -tos participles from this perspective, we are led to a number
of predictions concerning potential meaning differences between the two:
(i) We expect that -menos will not serve as a context for determining the meaning
of the root because there is always a little v head vE introducing an event
variable below -menos which creates a domain for allosemy.
(ii) As a result of the presence of vE, we expect that -menos participles will inherit
the literal meanings of the corresponding verbs.
(iii) Given that -tos participles do not have event implications, we expect that -tos
will be close enough to the root to serve as context for root allosemy and,
potentially, that the interpretation of the root in -tos forms will be related but
not identical to the interpretation of the same root in a -menos participle.
(iv) Finally, if it turns out that -tos participles have the same meaning properties
regardless of the presence or absence of a vC verbalizer in them, then we will

Allosemy, Idioms, and their Domains

have to conclude that vC is ignored when present in a -tos participle, i.e. it does
not disrupt the semantic adjacency between -tos and the root.
The comparison between -menos and -tos participles reveals that they indeed show
systematic differences in their interpretation, relating to the presence vs. absence of
a vE head. Secondly, that vE counts for allosemy, as stated in predictions (i) and (ii)
above, while vC doesnt (prediction (iv)). In what follows, we first discuss examples
without an overt verbalizer and then proceed to cases with an overt verbalizer.
An example without a verbalizer is ():
() a. spas-tos


b. spas-menos

(a) and (b) both appear to involve direct attachment of -tos and -menos to a
(verbal) root. (b), however, clearly contains an abstract verbalizer/eventivizer vE
between the root and -menos, while vE is absent from (a). The literal meaning
of break is retained in the -menos participle in (b), while (a) has a distinct
interpretation: it means folding and applies to objects consisting of parts that can
break into smaller pieces:

break-ti.f umbrella
folding umbrella


break-meni.f umbrella
broken umbrella

folding table


broken table


The two meanings are closely related but sufficiently distinct to be expressed by different roots in a language like English. Note that both spas-tos and spas-menos can
form closely related (though not identical) idioms, a fact showing that the idiomatic
interpretation of participles is a separate issue:
() spas-ta
break-ta.pl.neut Greek
broken Greek (not very good Greek)
() spas-meni
broke-meni.f.sg voice
broken voice (voice that does not sound clear)
Another pair illustrating a clear meaning difference between the two forms is provided
in (), based on the verbal root FTIAX (make):

Anagnostopoulou and Samioti

() a. ftiax-tos
made up/fake

b. ftiag-menos

And a very similar meaning difference can be seen in () based on the root PLATH
(and its allomorph PLAS- form/create):
() a. plas-tos

b. plas-menos

In both () and () the -menos forms in (b) have the regular meanings of the
corresponding verbs make and create. On the other hand, in (a) and (a)
the root has the interpretation made up, i.e. something made/created and therefore unreal/false/fake. Interestingly, only the -menos participle in (b) can have an
idiomatic reading (it means stoned/drugged or angry), while the other three forms
cannot, once again indicating that idiomaticity is subject to conditions on top of
The data discussed so far verify predictions (i)(iii) above, namely that -menos
participles have meanings based on the corresponding verbs while -tos participles do
not, and that the meanings of the roots in -menos participles are not the same as in
-tos participles due to the presence in the former and the absence in the latter of an
abstract vE head. We now turn to prediction (iv) stated above, exploring the role of
an overt vC head in a -tos participle.
As it turns out, the meaning distinctions between -menos and -tos participles are
as they have been described so far, despite the presence of an overt vC head in a -tos
participle. To illustrate, consider the following examples:
() a. kokin-is-tos
made red


b. kokin-is-menos
made red

In both examples, the root means red. In the context of -tos, however, red acquires a
very specific meaning, red sauce/cooked in a red way. As a result, kokin-is-tos can
only mean something cooked with a red sauce/in a red way. On the other hand,
kokinismenos can apply to anything that has become red as a result of a becoming
red event:
() kokin-is-to kreas / kotopoulo / magoulo
red-v-to meat chicken
meat/chicken with a red sauce
() kokin-is-meno derma / magulo / mati / xroma
red-v-meno skin
cheek eye colour
skin/cheek/eye/colour that has turned red as a result of an event

Allosemy, Idioms, and their Domains

A similar contrast is found with -tos and -menos forms based on the root GEM (full)
in ():
() a. gem-is-tos


b. gem-is-menos

The meaning of the root is similar in (a) and (b), but gemis-tos means stuffed and
applies to stuffed things (tomatoes, chickens, etc., a much more specific meaning)
while gemismenos applies to anything that is filled as a result of a filling event:
() a. gem-is-tes tomates
full-v-tes tomatoes
stuffed tomatoes

b. gem-is-meno bukali
full-v-meno bottle
filled bottle

As with example () without a verbalizer, the two meanings are closely related but
sufficiently distinct to be expressed by different roots in English. The same can be seen
in ():
() a. xtip-a-o

b. xtip-i-tos

c. xtip-i-menos

Note that in all the examples ()() the question of idiomaticity does not arise,
because there is nothing unpredictable or non-literal in the meaning of the -tos and
-menos forms. It is just that the root has a different sense in the context of -tos than
it has in the context of little vE, the head present in verbs and in the corresponding
-menos participles.
A final example illustrating the systematic meaning difference between -menos and
-tos participles with verbalizers is ():
() a. xon-ev-tos
inside the wall


b. xon-e-menos

What is interesting about this example is that the meaning of the root XON is rather
underspecified and becomes fixed in context. In the context of -tos, the final meaning is inside the wall. On the other hand, xon-e-menos means digested, just like the
meaning of the corresponding verb.
.. vC and semantic adjacency
Marantz () proposes that the vC head described above instantiates a semantically
empty head, the zero semantic counterpart of the phonologically empty v head in
cases of allomorphy (see e.g. the English past tense example ()). The -t- participle
See Anagnostopoulou and Samioti (forthcoming) for extensive discussion of such cases that are
reminiscent of what Arad (, ) describes for Hebrew.

Anagnostopoulou and Samioti

morpheme requires a verbal stem syntactically; at semantic interpretation, however,

this v node is ignored and -t- can create a meaning directly from the root. The exact
analysis for Greek -tos participles with a semantically empty verbalizer would look as
follows (Marantz : ):
() Greek root-attaching -tos forms



() Greek -tos verbalized stems




In () the little v head is semantically empty and therefore ignored for allosemy. As
pointed out by Marantz, it is crucial that -tos is not a little a head but rather part of the
inflectional domain. In order to condition contextual allosemy, -tos must fall within
the same Spell-out domain of the root and not define its complement as a Spell-out
domain, i.e. it must behave like the T head in taught and not like a C head. This was
also proposed by Alexiadou and Anagnostopoulou () for independent reasons:
recall that part is analysed as an Aspect stativizing head in ()(). Independent
evidence that -t- is not a little a head but rather an inflectional head comes from the
fact that -t- is present in both adjectival forms and nominals. This would be difficult
to explain if -t- was a little a head. More specifically, -t- can be found in nouns that
have exactly the same structure as -tos adjectives, except that they are nouns (e.g. they
have fixed gender):
() pag-o-to
ice cream
The noun in () has the same internal composition as a -tos participle. It consists
of an entity-/quantity-denoting root PAG- (ice), a verbalizer, and the -tos ending.
This time, however, the result is a noun meaning ice cream, which has fixed gender
(neuter). The same root combined with a verbalizer can give the verb freeze:
() pag-on-o
freeze (used causatively and anticausatively)

Allosemy, Idioms, and their Domains

From pagono in () a -menos participle can be created while a -tos one cannot:
() a. pag-o-menos

b. pag-o-tos

Given the unavailability of (b), it cannot be claimed that the noun in () is created
on the basis of an adjective. Hence, -t- in () cannot be analysed as a little a head.
Some more examples of this type are given below:
() a. kopan-is-ti
a type of cheese
b. kalam-o-ti
a type of roof
c. grap-to
exam paper
In conclusion, the Greek vC head identified by Anagnostopoulou and Samioti (forthcoming) in Greek characteristic state -tos participles is the LF analogue of a zero
head at PF. It does not count for interpretation in two related respects: (a) it does
not introduce an event variable; (b) it is ignored when the meaning of the root is
determined in the context of the part head -tos.

. English Participles from the Perspective of Greek: vC, vE, and Voice
in English
Marantz (, ) argues that similar cases to the ones discussed in the previous
sections are also attested in English (he also discusses Japanese causatives which
present a similar pattern; Volpe ). () contains two types of participles. On the
left, participles are based on nominal roots to which an -ize verbalizer attaches. On
the right, participles are based on nominal roots which become adjectives through the
adjectival head -al; the verbalizer -ize attaches next and, finally, the participle affix -ed:
() a. quantized energy
b. pulverized lime
c. atomized individual

d. globalized universe
e. nationalized island
f. fictionalized account

The adjectives on the left correspond to Greek -tos participles with a contentless
head vC: they may be interpreted as describing a state (in quanta) that was not
necessarily brought about via an event described by the embedded verb (to quantize).

Anagnostopoulou and Samioti

Pulverized lime can mean lime in powder form (no event of crushing implied) and the
root is interpreted as powder. An atomized individual is separated from society not
necessarily as a result of a separation event. On the right, the adjectives describe states
resulting from events, and correspond to Greek -menos participles with a contentful
vE head. Marantz attributes the difference between the two groups of participles to
the adjectival suffix -al, an exponent of little a. Being a phase head, -al intervenes
between little v and the root, preventing the part head from serving as a context for
determination of the meaning of the root.
Bruening (to appear) investigates the interpretive differences between adjectival
and verbal passives and points out that the most striking difference between the two
is the case of missing inputs. There are many examples of adjectival passives in English
that do not exist either as verbal passives or as active transitive verbs. In () we list
Bruenings examples:
() a. bedridden, destined, dilapidated, embattled, hagridden, laden, stricken,
b. bedecked, bedraggled, bejewelled, belated, beloved, bemused, benighted,
benumbed, bereaved/bereft, beribboned, beringed, besotted
According to the Oxford English Dictionary, in some of these cases the active verb used
to exist, mostly with adjectival passives beginning with the prefix be-. There is also
a productive class of English adjectival passives formed from nouns, and especially
body-part nouns:
() a. bearded, mustachioed, one-eyed, long-haired, broad-chested, big-eared
weak-chinned, thick-skinned, club-footed, gap-toothed, pigtailed, sixfingered, long-necked, big-boned, honey-tongued, smooth-cheeked
b. scaled body, maned head, clawed feet, fanged frog, frilled lizard, long-tailed
weasel, flop-eared bunny, duck-billed platypus
c. landed gentry, spiked collar
Bruening hypothesizes that these are derived from an intermediate non-existent verb
form derived from an N and meaning possessing N perhaps via a null affix.
Greek presents evidence that the classes in () and () should not be treated along
exactly the same lines. The participles in () contain a zero vC and the participles
in () a zero vE. Starting from those in (), we have seen that Greek has similar
cases of -tos participles based on nouns which combine with a semantically empty but
phonologically overt vC, leading to forms that are non-existent as verbs. In Greek, the
relevant nouns are not body parts but other types of noun, which do however lead to
forms often conveying the meaning possessing/having the property N. The example
we saw was (), repeated here:
() kamban-a bellN ??kamban-iz-o bellV kamban-is-tos sounding like a bell

Allosemy, Idioms, and their Domains

Other examples are listed in ():


having stamps/stamped

b. riga


having stripes/striped

c. tharos


having courage/courageous

d. kopad-i


many together (like a herd)

e. agkathi



() a. stamba

The -tos participles in () are unambiguously non-eventive in Greek, and it is clear

that the overt verbalizer in them is an exponent of a semantically empty v. On the
basis of the evidence from Greek, then, it could be proposed that the missing input
cases listed by Bruening in () contain a semantically empty v (vC), which happens
to also be phonologically empty.
On the other hand, the cases in () seem to contain an empty vE, which can be
modified etc. (see e.g. http://www.thefreedictionary.com/hagridden: a man hagridden by the futurehaunted by visions of an imminent heaven or hell upon earth
(C. S. Lewis)). These are predicted to correspond to -menos participles in Greek, and
this is correct: e.g. destined is prooris-menos in Greek, embattled is mach-o-menos,
and benumbed is moudias-menos. Note that the Greek counterparts of the English
words in () are based on existent verbs, namely moud-iaz-o (to become numb), prooriz-o (to determine beforehand), mach-ome (to fight).
Turning, finally, to the question of idiomaticity in English, Bruenings (to appear)
investigation leads him to conclude that there are no special conditions on the
idiomatic interpretations of adjectival participles that are qualitatively distinct from
the conditions on the idiomatic interpretations of phrasal idioms. Bruening calls the
former semantic drifts and the latter idioms, and takes the position that Semantic
drift and phrasal idioms are just two names for the same thing: special interpretations
of syntactic structures (Bruening, to appear) . He argues for a frequency-based theory of idioms, according to which there is an inverse relation between frequency and
size of syntactic structure, such that special interpretations become less frequent the
larger a syntactic structure is. Therefore, there are more idioms consisting of just a
verb and its object than idioms consisting of a subject, a verb, and an object. Bruening
proposes that the reason why there are more idioms with adjectival passives than with
The noun stamp can also be used as a basis of the verb stamb-ar-o (stamp-v-sg) in Greek which
means spot. That verb can be used as a basis for both a -tos and a -menos participle, with the usual semantic
and syntactic differences between the two.

Anagnostopoulou and Samioti

actives and verbal passives is because adjectival passives have a stative use that is not
the same as the verbal active or passive. Therefore, speakers will have a reason to
limit a special interpretation just to the adjectival passive, while, when they encounter
a special interpretation of the verbal passive, they will have no reason to assume
that this interpretation is not available to the active, which is truth-conditionally
equivalent to the passive. He argues that this is indeed what we find. Verbal passives
with idiomatic meanings usually correspond to actives with the same meaning, while
adjectival passives have special meanings different from the meanings of actives and
passives. He furthermore points out that there is at least one example of a verb forming
a verbal passive with a meaning not available to the active. The verb be born/be
born again/be reborn, has distinct active and passive uses, as shown in () and
() (Bruenings (), ()). The active only has literal and scientific uses while the
passive has non-literal/metaphoric uses. Note that the progressive in () provides
evidence that this is a case of a true passive:
() a.

Mammals bear live young./ Live young are borne by mammals.

A child was born to/ by Mary and Joseph./ Mary and Joseph bore a child.
She bore him a child./ He was borne a child (by her).
That childs soul was reborn to the same mother/ That mother rebore the
soul of her first child.

() a. At this very moment, the prophesied saviour is being born to unsuspecting

b. At this very moment, the soul of the Dalai Lama is being reborn in a tiny
On the basis of idiomatic reborn, he concludes that passive idioms are not excluded in
principle. They are just very rare. Bruening does not address the agentivity restriction
on idioms, and it is not clear to us what he would say about this restriction in connection to the passive. But notice that the be born examples in () and () clearly lack
agentivity properties. This is evident from the fact that what could be characterized
as an external argument is not introduced by the preposition by but rather by the
preposition to (in (b,d)), and has the interpretation of an affected indirect object
argument (in the active (c) this argument is realized as him). Therefore, the data
he discusses do not actually challenge the generalization/hypothesis that there is an
agentivity restriction blocking idiomatic interpretations of passive participles, just as
it blocks agent-based subjectverb idioms (Harley and Stone, Chapter below). At
this stage, the reason for the agentivity restriction is unclear to us. But notice that, if
the allosemy hypothesis is correct, then this restriction could not simply be attributed
to the status of Voice as a phase head, unless we had a theory that could answer in a
principled manner why the vE phase head defines a domain for allosemy, while the
Voice phase head defines a domain for idiomaticity.

The No Agent Idioms Hypothesis

. Introduction: Interface-Related Questions Raised by Idioms

Every language contains phrasal expressions whose meanings must be learned by rote,
idioms like the English phrases chew the fat chat, gossip or eat crow be proven wrong.
There are two major families of thought concerning the best way to model idiomatic
interpretations in the grammar. According to one approach, the entire phrasal structure is listed as bearing a meaning, considerably expanding the purview of the lexicon.
The second approach treats idiomatic meanings as properly belonging to individual
words within the idiom, which are endowed with very specific context-dependent
interpretations in addition to their regular, context-independent meaning.
Given modern syntactic notions about the cyclic nature of the relationship between
the syntactic computation and semantic interpretation, the context-dependent
approach suggests the existence of constraints on the structural contexts on which
such special meanings can depend. For example, if the syntactic computation is interpreted in chunks, as in phase theory, we might expect that a particular lexical items
special meaning can be dependent on contexts no larger than a phase (Stone ),
since that is the maximum amount of material that can be semantically interpreted at
a single time. In this chapter we explore the status of one such proposed constraint,
the No Agent Idioms Hypothesis of Marantz (, ), and conclude that it has
some substance. We argue that a class of counterexamples to the hypothesis are not
true counterexamples, as their argument structure has significantly different properties from that of agentive structures. In addition, we argue that a context-dependent
approach is more consistent than a stored-structure approach with a central tenet of
linguistic theory, namely the compositionality hypothesis. We take this to indicate that
We would like to thank the audience at OnLI II and the participants in the Fall Composition-

ality Seminar at the University of Arizona for helpful discussion. We would also like to thank Tal Siloni
and Julia Horvath for important commentary, and an anonymous reviewer for many clarifications and
improvements. All shortcomings remain entirely our own responsibility.
As a reviewer points out, the existence of such constraints is not proof that the context-dependent
approach is on the right track.

Harley and Stone

the context-dependent meaning view of idioms has some advantages over the phrasal
lexicon approach.
In section ., we point out the different consequences for compositionality of the
two hypotheses. In section ., we discuss whether there are structural constraints on
idiomatic meanings, and if so, what these might be. We dismiss one overly restrictive
proposal, and in section . we pursue the No Agent Idioms hypothesis, arguing
against an alternative explanation for the paucity of agentverb idiom combinations,
and that one class of counterexamples to the hypothesis are in fact object-experiencer
constructions, lacking true Agent subjects. Other counterexamples have atypical syntactic properties, which we take to indicate the possibility that they contain derived,
rather than base-generated, subjects. Section . concludes.

. Compositionality
Below, we introduce the two conceptions of the nature of idioms described above and
discuss the relationship of each to the compositionality hypothesis. The first, we argue,
treats idioms as true exceptions to compositionality and opens the door to a general
re-evaluation of basic compositionality as a fundamental tenet of the theory. The
context-dependent meaning approach, on the other hand, maintains compositionality
in its traditional form.
The compositionality hypothesis states that the meaning of a complex expression is
constructed from the meanings of its parts and the combinatoric structure imposed
on them by the syntactic computation. Discussions of compositionality frequently
feature a disclaimer about idioms: idiomatic expressions are thought of as cases in
which strict compositionality breaks down. The meaning of chew the fat chat, gossip,
for example, cannot be computed from the context-independent meanings of chew,
the, and fat. This VP has two distinct interpretations: one which is computed from
such context-independent meanings and an idiomatic one which is not.
.. Listing phrases
A typical way of thinking about the problem of idiomatic phrases such as chew the fat
is to allow complex syntactic structures to be listed as lexical items, with idiosyncratic
meanings analogous to those of basic Saussurean signs (see e.g. Williams , Goldberg , Jackendoff ). These structures enter into the syntactic computation
directly as verb phrases, not as word-sized objects. Once in the syntactic frame,
however, they subsequently behave as regular verb phrases: the verb is independently
inflected and these expressions are often able to participate in regular syntactic processes such as passivization and relativization. On such an approach, compositionality
does truly break down within an idiomatic structure: there is more to its meaning than
that contributed by the individual lexical items and their mode of composition.

The No Agent Idioms Hypothesis

Construction-based approaches to grammar take this to the next level: phrase

structures entirely devoid of lexical content can have idiosyncratic meanings attached
to them. Thus, for instance, in construction grammar (e.g. Goldberg : ) the
double object construction contributes an interpretation (causereceive) entirely on
its own, prior to any particular lexicalization; the PRED relation fills in with whatever
lexicalization is chosen:











Ditransitive Construction






On this view, a languages grammar is nothing but a collection of forms or templates

(i.e. constructions), with idiosyncratic meaning attached. Constructions can vary in
their generality. Consider the examples below, the first three from Williams ()
and the others modelled on his notation. In this construction-type view, the structures
that bear special meanings range all the way from the very general, as in (a), to the
very specific, as in (e):


[Wh-phrase] []S
[V S]
[kick the bucket]VP


embedded question
yes/no question
lower trunk wear

On such a view, in which structures themselves can be Saussurean signs, there is no

necessary connection between a given structure and the meaning it can convey. Structural meanings are able to vary without limit in the same way that monomorphemic
lexical items meanings can. Compositionality as traditionally understood is no longer
a desideratum in such theories.
.. Context-dependent meanings
Given that idioms represent a key case which has been taken to support the necessity
of listing structurally complex items in the lexicon, we might want to ask if preserving
strict compositionality, in the traditional sense, is an option. Does the meaning of the
VP in John kicked the bucket involve more than the meanings of kick, the, and bucket?
In fact, a well-established alternative to the construction-based view exists. If
the special meaning is attached to one or more of the lexical items used in the
idiom, and conditioned by a lexical or syntactic context, then, technically, idioms are

Harley and Stone

compositional: their meanings are computed from the meanings of their parts and
the way they are put together. Its just that the meanings of their parts are potentially
dependent on the syntactic and lexical context.
This intuition about the relationship of idiomatic interpretations to individual
lexical entries is reflected in the organization of idioms in English-language dictionaries. The special meanings for idioms are typically listed as submeanings under
the headword for the head of the phrase, not as separate headwords themselves. For
example, the MerriamWebster Dictionary and the Oxford English Dictionary list chew
the fat under the headword chew. On this conception of idiomatic interpretation,
compositionality holds, albeit in a rather indirect way. The verb chew is polysemous,
just like, for example, the noun mouth. The catch is that one of its senses is only
available when it appears in combination with a certain DP object, the fat, which
presumably also is polysemous and receives a special interpretation in combination
with chew. ,
We consider the fact that the context-dependent meaning approach allows us to
preserve a traditional notion of compositionality to be a significant argument in its
favour. It therefore seems important to explore further implications of this view of
idiomatic meaning within modern syntactic theory. In particular, in the next section we turn to the question of the existence of structural constraints on idiomatic

. Structural Constraints on Idioms

As noted above, within Minimalist syntactic theory, the syntactic computation is
interpreted cyclically, in strict chunks, or phases, as the derivation is built up
(Chomsky , , ). If this is the case, it is plausible that phase-based interpretation may place a strict constraint on the conditioning domain for lexical items
special interpretations. It seems intuitively correct to assume that, at the point of interpretation, lexical entries must be accessed to construct the semantic representation,
Note that this idea that individual lexical items contribute content to the interpretation of the idiom is
similar to, but crucially distinct from, the concept of Idiomatically Combining Expressions of Nunberg,
Sag, and Wasow (). Their Idiomatically Combining Expressions are restricted to cases where a clear
metaphorical analogue for each subpart of the idiom exists, as in spill the beans, where spill corresponds to
tell and the beans corresponds to a secret, in which it seems very reasonable to attach these interpretations
as individual subentries in the polysemous semantics for each of these subparts. In our conception here,
however, the same contextually dependent interpretations are employed in idioms where no such neat
correspondence exists between subparts of the idiom and the content it expresses, as in kick the bucket
for die or chew the fat for chatwhat Nunberg et al. term Idiomatic Phrases. The catch in the present
treatment is that in such expressions the special meaning contributed by the lexical items bucket or fat will
be effectively zero. See Stone (in preparation) for further discussion, and Horvath and Siloni (to appear) for
arguments against mutually dependent listing of information concerning a single idiomatic interpretation
in multiple entries (e.g. under both chew and fat).
See Anagnastopoulou and Samioti (Ch. above) for a relevant discussion of polysemy and idiomaticity. Specifically, they discuss how locality restrictions on idiomaticity and polysemy differ.

The No Agent Idioms Hypothesis

and only information within the phase that is being interpreted could condition the
selection of the appropriate truth-conditions needed for this representation; this is
a consequence of the Phase Impenetrability Condition of Chomsky (). We thus
predict that there should be strict structural limitations on the domain relevant for
idiomatic interpretations. In this section we review some proposals concerning the
structural constraints on idioms and identify one that appears to have the properties
predicted by phase theory.
.. The structure-dependence of idiomatic interpretations
Tellingly, idioms first received extensive attention within generative syntactic theory
as diagnostics for underlying constituent relations. For example, idiomatic interpretations played an important diagnostic role in Chomskys () distinction between
raising and control infinitives. The availability of an idiomatic interpretation for the
cat BE out of the bag in (a) and its unavailability in (b) below was taken as evidence
for the hypothesis that the NP cat was base-generated in a local relation with be out
of the bag in (a) but not (b).
() a. [The cat]i seems t i to be out of the bag.
b. [The cat]i wants proi to be out of the bag.
This reflects an underlying hypothesis about idiomatic interpretations: in order for an
idiom to receive its idiomatic reading, it has to have all its parts in an appropriate local
configuration at D-structure. In modern terms, we could understand this as a claim
that all the parts of an idiom must be accessed within a single interpretive domain (e.g.
within a single phase) in order for the idiomatic interpretation to obtain. In (b), the
lexical entry for cat is never part of the same phase as the rest of the idiom, and therefore the interpretive component cannot access the mutually dependent conditioned
Another structural condition on idiomatic interpretations was identified by Koopman and Sportiche (: ). They observe that while idioms frequently comprise
constituents with open positions contained within them, as in () below, the open
position never corresponds to the structural head of the idiomatic constituent. No
idiomatic interpretation can depend on, for example, a pair of cooccurring words one
of which is a modifier of the subject and the other a modifier of the VP. They give the
example in () as a case of an impossible idiom, where special interpretations of pale
and slowly depend on their cooccurrence in these positions:
As noted by a reviewer, alternative possibilities exist. For example, a constraint of the type, Idiom
constituents cannot receive a theta-role from an idiom-external predicate would predict the absence of
idiomaticity in (b) as well. It is perhaps worth noting that the phase-based interpretive constraint described
above, however, is more general; it follows as a consequence of how the phase construct is intended to
function in the theory, and is straightforwardly connected to the other roles phases play in non-idiomatic

Harley and Stone

() a. get DPs goat

b. crane anaphors neck
c. send DP to the showers
() The pale man slowly put flowers next to John.
They write:
Assume the following: only if pale modifies a subject and slowly co-occurs in the same proposition does the following idiomatic interpretation arise: pale means unknown to the speaker and
[slowly means] the action was done in a roundabout way. In other words, when uttering [()],
the speaker means that the man unknown to me put flowers next to John in a roundabout way.
More generally, pale X slowly verbed , stands for X unknown to speaker verbed in a roundabout

In order to rule out the occurrence of such patently impossible idioms, Koopman and
Sportiche propose the following structural constraint on idiomatic interpretations:
() If X is the minimal constituent containing all the idiomatic material, the head of
X is part of the idiom.
This, in a sense, represents a formalization of the intuition that drives lexicographers
to often list idiomatic interpretations as subentries of a single lexical item. See Horvath
and Siloni (, to appear) for a formal treatment of idioms as stored under heads of
phrases. Further, while Koopman and Sportiche do not seek to place limits on the size
of the constituent which can bear idiomatic meaning, their constraint does reflect the
fact that local relations between idiomatic constituents are of paramount importance.
.. The Categorization Hypothesis
Before turning to the generalization about Agents and their (non-)participation in
idiomatization which will be our central focus below, we pause here to discuss another
recent proposal about the locus of idiomatic interpretations which we might call the
First Merge or Categorization Hypothesis (Marantz , Arad , ). This
proposal is couched within a framework of assumptions about the syntactic nature
of lexical structure, which crucially includes the existence of acategorial roots. These
roots acquire their syntactic category by merger with a little x head, a categorizing
element drawn from the set n , a , or v (e.g. Marantz ).
According to the MarantzArad First Merge Hypothesis, the only true domain of
idiosyncratic meaning assignment occurs at the moment of categorization. When n
or a or v , considered to be phase heads under this proposal, are merged with a
root, this constitutes both the first phase and thus the first semantic cycle, and the
interpretation of the root in the environment of this categorizing head is thereafter

fixed. For example, the acategorial root spec occurs merged with v in spec-ify and

The No Agent Idioms Hypothesis

with n in spec-ies, but the interpretations of the two are only tangentially related. On

the Categorization Hypothesis about idiomatic meaning, spec only acquires specific
meanings as the result of such categorization, and such a meaning, once acquired in
a given structure, should carry through for the rest of the derivation.
Obviously, on this view idiomatic interpretations of phrasal constituents must be a
different phenomenon entirely, perhaps simply an extreme form of metaphorical or
figurative language use (Alec Marantz, p.c.). As such, the MarantzArad proposal is of
perhaps limited relevance to the issues being addressed here. However, as a concrete
structural proposal concerning the appropriate domain for idiomatic interpretation,
we feel it appropriate to address the question of whether there is a qualitative difference between the idiosyncratic meanings assigned at the first merge of an acategorial root with its categorizing head, and idiosyncratic meanings that arise later. In
particular, we find it implausible to suggest that meanings acquired upon first merge
must persist in the interpretation throughout the derivation. It seems clear that truly
idiosyncratic semantics can be assigned following the first cycle of categorization, as
also argued by Borer () and Anagnostopoulou and Samioti (, Chapter
Consider, for example, the cases in () below. In each group of words based on a
common root, there is at least one word derived from a derivation of that root which
bears a meaning that does not obviously arise from the meaning of the derived stem
plus the meaning contributed by the affix; this meaning is presented in boldface:
() a. edit
edit-or one who edits
edit-or-ial . of or relating to the editor
. opinion article
b. nature
natur-al of nature
natur-al-ized . made natural
. became a citizen by residing in country

c. sanitsanit-ary clean
sanit-ari-um institution for mentally ill

d. auditaudit-ory to do with audition

audit-ori-um large performance space
e. class
class-ify sort
class-ifi-ed sorted
class-ifi-ed-s small newspaper advertisements

Harley and Stone

natnat-ion community of people possessing a territory and government

nat-ion-al of a nation (not an antonym of private)
nat-ion-al-ize . make national
. government takeover of business
(antonym of privatize)
g. domindomin-ate rule or control
domin-at-rix female top in ritualized sexual domination



institinstit-ute put a system into place

instit-ut-ion . a system that has been put into place
. an organization or building, especially a public
care facility
institut-ion-al . of a system that has been put in place
. grim, like a public care facility
institut-ion-al-ize . make institutional
. commit to a care facility

i. universe
univers-ity institution of higher learning

hospithospit-al institution for medical care
hospit-al-ity welcomingness
It should be clear, for example, that the root of national has merged with at least
two categorizing heads, realized by the suffixes -tion and -al. On the MarantzArad
hypothesis, then, the meanings associated with national should figure compositionally in the meaning of nationalize. However, on at least one of its uses, nationalize
has a considerably distinct range of interpretations which are not compositionally
related to the meaning of national. For example, we can speak of a national company,
intending to refer to a company with operations throughout the country. However,
to nationalize a company is not necessarily to cause it to have operations throughout the country, but rather to implement a government takeover of the company.
In this sense, nationalize is the antonym of privatize. If the meaning of national
figured in the meaning of nationalize, and the same held true for private and privatize, we would expect private and national to be antonyms. Other similar cases
might include the pairs conserveconservation, relaterelation, deducededuction,
protectprotectorprotectorate, economiceconomical, specifyspecifier.
Furthermore, there are clear cases of roots which have no independent meaning following merger with the first categorizer. The meanings of these roots are

The No Agent Idioms Hypothesis

instead wholly dependent on occurrence in a bigger syntactic and lexical conditioning

() a.

kit and caboodle

run the gamut
by dint of
in cahoots with
kith and kin
vim and vigour
high jinks

It would seem logical on the MarantzArad story that if these roots are assigned
semantically independent compositional meaning at the first categorizer, they should
be able to occur independently and contribute that meaning in distinct syntactic
contexts. However, none of these items has a recognizable meaning or occurs freely
in other appropriate syntactic environments; utterances such as What vim he has!
or Those jinks were lower than the ones we got up to yesterday! would qualify as
metalinguistic play.
Of course, on any account involving acategorial roots, the interpretation of the first
combination of a root with a categorizer will have to be idiosyncratic; roots dont
occur in isolation. In that sense, all root meanings will have to be context-dependent.
In support of this view, it is patently the case that some identifiable roots seem entirely
meaningless outside the first merge context (see also Baeskow , Harley ):
() a. -ceive
deceive, receive, conceive, perceive
b. -here
adhere, inhere
c. -port
comport, deport, report, import, support
d. -pose
suppose, depose, compose, repose, propose
These cases, and others like them, are clearly diagnosable as roots by an acquiring
child or linguist, as they exhibit contextual allomorphy and morphological selectional
restrictions regardless of the context they appear in, and furthermore would do so
productively in a nonce-form containing the root:
() a. -ceive -cept + ion
deception, reception, conception, perception
b. -pose -pos + ition (not -ation or -ion)
composition, supposition, proposition, deposition

Harley and Stone

The First Merge Hypothesis is correct in the sense that no root will contribute a
meaning entirely independently of the syntactic context, given that all roots must
be categorized. What we wish to emphasize here, however, is that idiosyncratic
interpretations of complex constituents created after the first categorizer has merged
arise in exactly the same way and are phenomenologically indistinguishable from the
idiosyncratic interpretations assigned at the first categorizing node. The main point
is that interpretations of complex constituents even after the first categorizer can still
be idiosyncratic, and need not contain the meaning specified at the first categorizer
as a proper subpart. The more structure a given phrase contains, the less likely it is
to be idiosyncratically interpreted, simply because of the increasing number of pieces
involved. However, the data do not suggest the existence of a qualitative difference
between an idiomatic phrase and an idiomatic wordthat is, until we arrive at the
domain of the external argument.
.. Marantz (): the special status of Agents
The strongest claim that we know of concerning the availability of idiomatic interpretations is that first articulated in Marantz (). Marantz pointed to an apparent
asymmetry in the frequency with which verbs receive special interpretations which
depend upon the identity of one of their arguments. He observed that, while it is
extremely common for a verb to receive a special interpretation based upon the
identity of its internal argument, it is vanishingly rare for a verb to receive a special
interpretation based upon the identity of its external argument. He gives the examples
in (), (), and () below. In all these cases, the truth-conditions given by the verb
vary significantly depending on the denotation of the internal argument: throwing a
boxing match involves a significantly different set of activities from throwing a party,
and neither of them bears much similarity to throwing a baseball.
() a.

throw a baseball
throw support behind a candidate
throw a boxing match
throw a party
throw a fit

() a.

kill a bug
kill a conversation
kill an evening
kill a bottle
kill an audience


cause the bug to die

cause the conversation to end
while away the time span of the evening
empty the bottle
entertain the audience to an extreme degree

On the MarantzArad approach to idiosyncratic meaning, the question of what an idiom is, and
how to identify one in the wild, does not arise; there just arent any phrasal idioms with linguistically
interesting properties. For us, an important question is how to distinguish idioms and metaphors. One
useful criterion, first mentioned to us by Massimo Piattelli-Palmarini, is (un)interpretability of a given
phrase by second language speakers with similar cultural backgrounds (cf. e.g. Lakoff and Johnson
on widely understood metaphorical mappings from physical to abstract domains, among others).

The No Agent Idioms Hypothesis

() a.

take a book from the shelf

take a bus to New York
take an aspirin
take a nap
take a letter in shorthand

It is worth underlining the point that these are not fixed in the sense that we normally
understand idioms to be fixed (for discussion of this point see e.g. Williams ). The
particular identity of the object can vary, as shown in ():
() a. kill the bottle/the peanuts/the casserole/the wine
b. kill an hour/a few minutes/time
We can think of these as idiomatic interpretations of the verb that arise when its object
is a member of a specific class, not necessarily when its object is a specific word or
phrase. The beauty of these examples is that it is easy to see the subject/object contrast,
given the ubiquity of cases depending on the semantic class of the internal argument
and the absence of such cases with external arguments.
Marantzs claim was that this is completely general, i.e. that special interpretations
involving a fixed verb and constrained Agent, with the object position being unconstrained/variable, do not exist. Marantz () revisits and reiterates this claim. If he
was correct, there should exist no idioms of the following form:
() a. The elephant kicked DP. = An employer fired DP.
b. [Timespan] killed DP. = [Timespan] made DP leave.
A more spectacular kind of evidence that has been adduced in favour of this hypothesis involves examples where otherwise readily available idiomatic meanings become
systematically unavailable when Agents are included in the structure; see Anagnostopoulou and Samioti (, Chapter above) for a convincing example of such a
case from Greek.
Kratzer () provides a formal proposal about the special structural and interpretive properties of Agents which is intended to explain the generalization. Before
proposing theories about why Marantzs generalization might be true, however, we
must first be sure that it is in fact true, since numerous works have presented counterexamples and considered the empirical basis for the claim to be fundamentally
flawed. We turn to this issue in the next section.

See the appendix of Horvath and Siloni (to appear) for some suggested counterexamples involving
unergative verbs with different types of subject. The examples (run, fly) that they discuss in detail are verbs
of motion, however, which have been extensively shown to alternate between unergative and unaccusative
structures (see e.g. Hoekstra and Mulder , among many others; Folli and Harley provide a recent
overview). Fly, for example, undergoes lexical causativization (The pilot flew the plane to New York), not
usually considered possible for unergative verbs ( The boy laughed the baby).

Harley and Stone

. Is the No Agent Idioms Hypothesis True?

We are asking what can be the conditioning context which triggers a special interpretation for a particular lexical item. Candidates for special interpretations range
all the way from interpretations completely dependent upon a single lexical item, as
in a fixed idiom such as kick the bucket, to the kind of truth-conditional variation
(as in ()() above) that was the topic of Marantzs original observation, which
was conditioned by the semantic category of some element in the context, but not
its particular lexical identity. The question is: what kind of syntactic relation can
the conditioning element bear to the conditioned element? Marantzs observation,
restricting conditioned element to the case of verbs, is that it is relatively common
for the conditioning element to be the direct object of the conditioned verb, and
his claim is that it is impossible for the conditioning element to bear an external
argument/Agent relation to the conditioned verb.
It seems clear that there are plenty of verbobject idioms or idiomatic templates,
and comparatively fewer, possibly no, subjectverb idioms or idiomatic templates.
What are some of the alternative explanations that have been advanced for this relative
distribution? We turn to this below.
.. Nunberg, Sag, and Wasow (): interacting prohibitions
Nunberg et al. dispute the claim that the prohibition on subjectverb idioms is categorical in character; rather, they ascribe it to the interaction of two independent
tendencies in figurative uses of language. First, when used figuratively, animate NPs
tend to refer to animate entities (as in figurative uses of dog, chicken, monkey, etc.).
Second, phrasal idioms typically describe an abstract situation in concrete terms.
These two tendencies are in conflict when it comes to the occurrence of animate NPs
in idioms, because animates are necessarily concrete. Therefore, they write, literally
animate NPs are rare in idioms, and since Agents and Goals are characteristically
animate, they too are rare (Nunberg et al. : ).
Nunberg et al.s prediction is thus that animate NPs should be rare in idioms in general, and they do provide significant evidence in favour of this hypothesis. They show
that verbs which overwhelmingly tend to take animate objects in normal discourse
(like hit or kiss) surprisingly nearly always occur with inanimate objects when used
idiomatically. Their examples with hit and kiss are given below:
() Idioms based on hit + NP:
hit the ceiling (get very angry)
hit rock bottom (reach the lowest point/worst situation)
hit the bottle (drink excessively)
hit the bulls-eye (get something exactly right)
hit the deck (take cover)
hit the hay (go to sleep)

The No Agent Idioms Hypothesis

hit the headlines (become public in media)

hit the jackpot (win)
hit the mark (get something right)
hit the nail on the head (guess right, express the precise truth)
hit the spot (something, such as food or drink, being enjoyable)
hit the road (get going)
() Idioms based on kiss + NP:
kiss the canvas (in boxing: fall down)
kiss the dust (fall down due to being shot/hit, be slain)
kiss NPs ass (flatter somebody), (Imperative form: curse)
kiss the cup (drink)
kiss the ground (admire, be grateful)
kiss the rod (accept chastisement submissively)
kiss something goodbye ((will) lose it)
Similarly, it is relatively rare for an idiom to feature a fixed Goal or Possessor argument.
This, like the paucity of fixed Agent arguments, also follows from their generalization,
given that Goals and Possessors tend to be animate.
However, while it seems clear that the tendency for animates not to occur in idioms
is exactly thata (robust) tendencyin the case of internal arguments such as Goals
and Possessors, the prohibition on Agents in idioms seems to be much more of a
categorical contrast. Even without looking in an idiom database, it is quite easy to
come up with idioms involving animate NPs in internal argument position. Some
examples are provided below, a few of which are noted by Nunberg et al.:
() a.

send X to the devil

throw X to the wolves
get Xs goat
give the devil his due
put the cart before the horse
catch a tiger by the tail
bell the cat
let the cat out of the bag
beard the lion in his den
look a gift horse in the mouth

Horvath and Siloni (to appear) argue that Nunberg et al.s observation about animates should be recast
as an observation concerning NPs referring to human entities, which strictly preserve animacy, hence
concreteness, and are therefore less likely to occur in idioms describing abstract situations; thus, they claim,
the paucity of Agent NPs in idioms is attributable to the cognitive bias against abstract interpretations of
human-denoting NPs.

Harley and Stone


pay the piper

separate the men from the boys
keep the wolf from the door
kill the goose that lays the golden egg

None of these involve Agents. Even allowing for Nunberg et al.s observation about
the infrequency of animate NPs in phrasal idioms, the contrast between internal and
external arguments still seems clear. Indeed, if we expect the proportion of fixed
NPs in idioms to roughly mirror the proportion in normal discourse, subject to the
distorting effect of the no-animates tendency, we would expect it to be easier to come
up with AgentV idioms than VGoal idioms, given that the proportion of animate
NPs in normal discourse which are Agents is significantly larger than the proportion
of animate NPs which are Goals. Yet it is easier to find such examples with Goals
than with Agents. According to Nunberg et al.s predictions, there shouldnt be any
substantive difference between the Agent and Goal positions with respect to the ease
of idiomatization of animate NPs.
Therefore, even assuming that Nunberg et al. are correct about why there are fewer
animate NPs than inanimate ones in idioms in general, we feel that there still is a
significant puzzle about the paucity of fixed Agents in idiomatic expressions. Indeed,
we believe that the prohibition is categorical and should be explained by appeal to
fundamental properties of the grammatical system and its interface with encyclopedic/cognitive knowledge. To justify this position, however, we must carefully consider
the data that have been presented as counterexamples to the generalization, that is, the
cases of subjectV idioms, or interpretive idiosyncrasies dependent on the identity of
the subject, that have appeared in the literature.
.. Counterexamples: are there AgentV idioms?
On various occasions when Marantzs generalization is discussed, putative counterexamples are identified. Nunberg et al. present one example, [A little bird tell X Y].
Another English example is provided by Horvath and Siloni (), [Lady Luck smile
on X]. Nunberg et al. contend that these expressions are genuine idioms, not proverbs
or some other type of fixed expression, because they can occur in any tense or mood,
as shown in () below:
() a. Did a little bird tell you that?
b. A little bird must have told her.
Idiomatization may require construing an animate NP as having an inanimate referent, as in (c,e,f),
but that should be just as unlikely as using an animate-referring NP in an abstract context, so either way,
these are predicted to be rare cases, according to Nunberg et al.s hypothesis. Yet rare and non-existent are
still distinct.
In what follows, we omit discussion of subjectV idioms involving derived subjects, e.g. with unaccusative verbs, as it should be clear in the theory presented below that they are predicted to exist. See
Chtareva () and Everaert () for examples of this type.

The No Agent Idioms Hypothesis

Nunberg et al. also describe and give the date of a Linguist List post by Alexis
Manaster-Ramer, saying he gives a number of other idioms of this type (in several
languages) (Nunberg et al. : ); we reproduce Manaster-Ramers post in full
here, thanks to the archival powers of the Internet:
Date: Thu, 28 Jan 93 21:15:18 EST
From: [log in to unmask]
Subject: Subject-Object Asymmetry and Idioms (A Query)
The latest issue of US News and World Report has a nice
example of an apparent subject-verb idiom:
the vultures appear to be circling FBI Director Wm
Another such idiom is:
The spirit move- NP (where move- means you can have
different tenses).
What I am wondering is
(a) whether anybody still believes the subject-object
asymmetry thesis in connection with idioms (Chomsky,
Marantz) or any of the other theories about idioms which
would prevent such idioms from being possible, and
(b) whether anybody has noted other such examples
(N.B. we want idioms consisting of just subject+verb,
not subject+verb+ other stuff).
Also, I have a small number of examples of subjectverb idioms in Polish, Hindi, and German, and in all
three languages it appears that the idiomatic subject
cannot normally come first and the non-idiomatic must
precede it, e.g.,
Ihn reitet der Teufel (lit. Him rides the devil)
He is going bonkers
Now, I think that these word order facts probably
fall out of the normal rules for "topic" fronting in
these languages, but I am wondering if some people would
argue that the idiomatic subject is not really a
subject in such cases.
Horvath and Siloni () give additional examples in Hebrew and Hungarian:
() a. Ha-goral heir lo
the-fate lit up to+him face
He had good luck. (lit: Fate lit up its face to/on/at him.)

Harley and Stone

b. Ha-ruax nasa oto.
the-wind carried him
He disappeared. (lit: The wind carried him.)
c. Bala
oto ha-adama.
swallow him the-earth
He disappeared. (lit: The earth swallowed him.)


() a. Elkapta
pro a gpszj.
caught.pl.def.do him the driving belt
He got roped in. (lit: The driving belt caught him.)
b. Rjr
a rd.
onto him goes the shaft
Hes having a series of misfortunes. (lit: The shaft goes onto him.)
The most extensive set of such examples that we know of are given by Chtareva (,
), in Russian.
() a. Ivana
ne (s)xvat-il.
Ivan.acc almost Kondrashka.nom not grabbed
Ivan was frightened to death. (lit. Ivan was almost grabbed by paralysis.)
b. Ivana
zajela sovest.
Ivan.acc ate up conscience.nom
Ivan had a guilty conscience. (lit. Ivan is eaten up by his conscience.)
c. Ivana
zamuchali somnenija.
Ivan.acc tortured doubts.nom
Ivan experienced serious doubts. (lit. Doubts tormented Ivan.)
d. Ivana
oxvatil strax.
Ivan.acc seized fear.nom
Ivan experienced fear. (lit. Fear seized Ivan.)
She also gives a number of examples involving an external argument containing a variable genitive or dative possessor and a fixed body part, which often involve unergative
The idiomatic quality of these examples varies from case to case, but in all cases the verbs interpretation is recognizably distinct from its literal or typical interpretation, in a way that depends on the semantic
category of its external argument. Fear seized him, for example, involves a special interpretation of seize
which is triggered by the fact that the external argument denotes an emotional state. In this sense, we believe
these examples to be parallel to Marantzs.
About this idiom, Chtareva writes: Kondrashka is a personal male name, but in case of this idiom
it does not refer to a person. In its archaic idiomatic usage the word used to refer to paralysis, but most
native speakers are not aware of this meaning anymore. In the th century literature, the word appears
in an idiom X xvatil Kondrashka meaning X was paralysed. The idiom in (a) is a modern day variant
of this archaic idiom. It is resistant to any word order variations other than OSV with the focus stress on
Kondrashka, since it immediately follows the focus marker chut. It is a frozen idiom in terminology of
Nunberg et al. ().

The No Agent Idioms Hypothesis

() a. U Ivana
ne drognet (chto-to sdelat).
at Ivan.gen hand.nom not shake (to-do something)
Ivan wont hesitate to do something.
b. U Ivana
cheshutsja (chto-to sdelat).
at Ivan.gen hands.nom itch
(to-do something)
Ivan cant wait (to do something).
c. Kak u teba yazyk
povernulsja skazat takoye?
how at you tongue.nom turned
to say that
How could you say that?
d. U Ivana
po spine zabegali.
at Ivan.gen goosebumps.nom along spine ran
Ivan got the goosebumps.
e. U Ivana
at Ivan.gen heart.nom blood.inst pour over
Ivan is suffering. (lit. Ivans heart is bleeding.)
f. Ivanu
medved na uxo nastupil.
Ivan.dat bear.nom on ear stepped
Ivan cant hear music. (lit. A bear stepped on Ivans ear.)
For most of these cases, we will argue below that there is either solid evidence or
suggestive evidence pointing to the idea that the subjects of these idioms are not
true Agents. Alternatively, for a few of the cases, we will argue that the verb is not
involved in the idiom. Before proceeding to this discussion, however, it behoves us to
explain in what sense the subject of a verb like oxvatil seized or zamuchali tortured
may not be an Agent. In their non-idiomatic uses, the subjects of these verbs are
indeed typical, even prototypical, Agents. In a theory in which the lexical content of
the verb fully determines its theta grid, this would be enough to justify identifying
these cases as counterexamples to Marantzs generalization. However, as has long
been recognized, verbs are often not tied to a unique theta grid, and often exhibit
alternating argument structure frames with different interpretations. In modern syntacticocentric approaches to this phenomenon, the consensus solution has been to
divorce verbal content from argument structure frames to a certain degree, either
radically, as in the work of Borer (, ), or partially, as in Ramchand (),
among others. Within such a framework, it is possible to ask, are these idiomatized
verbs occurring in syntactic frames where their external argument is an Agent, or
have they been subject to an argument structure shift, appearing in a frame in which
the external argument is no longer in fact agentive, or external? Our argument below,
following Chtarevas insightful work, is that these idioms, although they use agentive
verbs, do not actually involve any Agent arguments; in the idiomatic structure, they

Harley and Stone

are internal arguments. We will discuss each set of examples from above in inverse
order, beginning with Chtarevas observations about Russian.
Chtareva argues that in all these cases of subjectverb idioms, the external argument is a Cause, rather than an Agent (despite the fact that all the verbs in () can
have an Agent subject in non-idiomatic uses), and that in all the cases in (), the open
object variable bears an Experiencer, rather than a Patient, role. She demonstrates
that the examples in () exhibit the syntactic behaviour associated in Russian with
psych-predicates, rather than the behaviour typically associated with agentive verbs.
We summarize her discussion below.
Chtareva shows that none of the structures in () can undergo passivization, or
be modified by an agent-oriented adverbial like on purpose, although non-idiomatic
interpretations of the same predicates can do so, as shown in examples (c,d)
ne (s)xvatil.
() a. Ivana
spetsialno chut KONDRASHKA
Ivan.acc on purpose almost Kondrashka.nom not grabbed
Ivan was frightened to death on purpose.
b. Ivan
byl sxvach-en
Ivan.nom was grabbed-p.prt Kondrashka.instr
Ivan was frightened to death.
c. Kondrashka
spetsialno chut ne (s)xvatil Ivan.
Kondrashka.nom on purpose almost not grabbed Ivan.acc
Kondrashka almost grabbed Ivan on purpose.
d. Ivan
byl sxvach-en
Ivan.nom was grabbed-p.prt Kondrashka.instr
Ivan was grabbed by Kondrashka.
Chtareva shows that the reciprocal binding possibilities mirror those of objectexperiencer predicates like radujut gladden, as can be seen by comparing (a) and
(a). They do not parallel the binding possibilities of normal transitive agentive
predicates like priglasili invite, illustrated in (b) and (b):
() a. Ivana
i Mariju
drug druga.
[Ivan.acc and Mary.acc]i gladden.pl success.nom.pl [each other]i
Each others success gladden Ivan and Mary.
Nunberg et al. (: , fn. ) note that the literature has not been clear on the point of whether the
relevant thematic role [of the subject] should be determined relative to the literal meaning of the expression,
or to its idiomatic meaning. It should be clear from the preceding discussion that we are using the latter.
As noted by a reviewer, the failure of passivization could be alternatively explained if these idioms
dont fall into Nunberg et al.s class of Idiomatically Combining Expressions; there are, of course, many
non-passivizable idioms. The failure of passivization, then, is not diagnostic of these idioms non-Agentive
status; rather, it is merely consistent with it.
According to Chtareva adverbial chut ne almost generally does not appear in passive constructions.
Since it is not part of the idiom, it is omitted in (b).
The word order differences between agentive vs. non-agentive sentences will be discussed below.

The No Agent Idioms Hypothesis

b. Roditeli
drug druga ne priglasili Ivana
i Mariju.
parents.nom [each other]i not invite [Ivan.acc and Mary.acc]i
Each others parents didnt invite Ivan and Mary.
() a. Strax
za armii drug druga ovladel Novgorodom i Pskovom.
[fear.nom for armies each otheri ] captured [Novgorod and Pskov]i .inst
Fear of each others armies seized Novgorod and Pskov.
b. Armii
drug druga ovladeli Novgorodom
i Pskovom.
[armies.nom each otheri ] captured [Novgorod.inst and Pskov.inst]i
Each others armies captured Novgorod and Pskov.
She also shows that the word order of such idioms in broad-focus contexts mandatorily differs from that of agentive verbs, patterning instead with the backwards word
order typical of object-experiencer verbs in the same context:
() Broad-focus environment: What happened/is happening?
a. Regular transitive verb
poluchil telegrammu.
Ivan.nom received telegram.acc
Ivan received a telegram.


b. Object-experiencer transitive verb:

rasstroili novosti.
Ivan.acc upset.pl news.pl.nom
The news upset Ivan.


c. SubjV idiom:
zajela sovest.
Ivan.acc ate up conscience.nom
Ivan had a guilty conscience. (lit. Ivan is eaten up by his conscience.)
Chtarevas conclusion, with which we concur, is that these idiomatic external arguments are Causers, not Agents. It has long been established that Causer subjects
of object experiencer verbs are base-generated VP-internally (Belletti and Rizzi ,
Pesetsky ). Therefore, such subjects are not base-generated as external arguments,
The idiom involving Kondrashka (() above) shows mandatory AccNomV order. Of this, Chtareva
writes: The [Kondrashka] idiom is a frozen idiom in that it resists scrambling and can only appear in
AccNOMV order with a focus stress on the subject. This focus stress is not identificational in the sense of
. Kiss () since it doesnt involve picking an element out of a set, but rather emotive in the sense of King
(), who describes it as emphatic stress on a constituent in emotive speech. Such focused elements are
marked with sentence stress and occur immediately before the verb, following preverbal topic, which is
exactly the case [in this idiom].
Chtareva argues that the accusative case on the Experiencer object is inherent rather than structural,
and the absence of a true external argument is a Burzios generalization effect, following Belletti and Rizzi

Harley and Stone

but rather acquire their subject hood via movement. Since the important conditioning factor in idiomatic interpretations is the configuration of the base-generated
structure, not the derived structure, idioms involving such subjects do not count as
counterexamples to Marantzs generalization; indeed, they are predicted to exist, since
the surface subject is base-generated as an internal argument. Within a syntacticocentric theory of argument structure, it is not particularly surprising to see normally
agentive verbs used in such object-experiencer frames. It should be clear that when
they are used in such frames, idiomatization is predicted by Marantzs theory to be
A similar set of potential counterexamples in Spanish were brought to our attention by Violeta Vzquez-Rojas (p.c.), and are illustrated in () below. These involve
transitive verbs with fixed nominative subjects and accusative objects. Very tellingly,
however, they all contain the reflexive clitic se, and exhibit unusual word order: the
nominative subject is mandatorily sentence-final, and the accusative object must be
in preverbal position.
() a. A Juan se lo
va a cargar el payaso.
dat Juan se acc.sg go to carry the clown
Juan is going to be in deep trouble. (lit. The clown is going to carry Juan
b. A Juan se lo
llev la trampa.
dat Juan se acc.sg carried the trap
Juan underwent a terrible fate. (lit. The trap took Juan away.)
Omission of se renders the idiomatic reading unavailable, as illustrated in ():
llev la trampa.
() a. ?A Juan lo
dat Juan acc.sg carry the trap
b. A Juan lo
va a cargar el payaso.
dat Juan acc.sg go to carry the clown
The clown is going to carry Juan away. (only literal meaning)
We take the reflexive morphology and the word order preference to indicate again an
internally base-generated position for the subject in the syntax for these examples as
A reviewer notes that Pesetsky () establishes a distinction between two classes of objectexperiencer verbs, ones whose subject is a Cause (worry-type) and ones whose subject is a Theme (appealtype), and wonders which class the Experiencer interpretations of the idioms here fall into. This would
matter if either of the classes had base-generated external arguments in Pesetskys analysis. In fact, it does
not matter. In Pesetskys analysis, even the Causer subject of PsyCaus object-experiencer verbs is basegenerated internally to the VP, as the object of a caus preposition in the complement of V. It then raises
to the external (thematic) Cause position. Its internal base-generation thus explains its backward binding
properties. Pesetskys analysis is thus compatible with the No Agent hypothesis of idiom formation for both
classes of object-experiencer verbs.

The No Agent Idioms Hypothesis

Turning now to the cases discussed by Horvath and Siloni from Hebrew and Hungarian, we have not yet pursued a systematic investigation of their properties along
the lines of the Russian cases discussed above, but an informal discussion with Edit
Doron seems to point in the same direction. None of Horvath and Silonis cases in ()
can be passivized, although their non-idiomatic counterparts can be (again, this may
be due to their not being Idiomatically Combining Expression idioms, rather than
to their non-agentive status: see footnote above). Further, the word order in (c)
is typical for these expressions; Doron comments that (b) is more felicitous in the
inverted order of (c), with the purported idiomatic external argument ha-ruax the
wind appearing at the end of the sentence. This VOS word order is quite atypical
for Hebrew, which most frequently exhibits SVO or VSO order. Although further
investigation is required, these idioms do seem to have a syntax that is atypical for
agentive expressions in the language, which is at least consistent with the possibility
that they involve derived, rather than base-generated, subjects.
We know even less about the Hungarian examples; however, judging by their
translation, they appear to exhibit the semantic qualities of object-experiencer psychpredicates. An investigation of their syntactic properties will have to await future
work. It is also worth recalling Manaster-Ramers remarks concerning word order in
his example from German and parallel cases he doesnt give from Polish and Hindi:
Also, I have a small number of examples of
subject-verb idioms in Polish, Hindi, and German,
and in all three languages it appears that the
idiomatic subject cannot normally come first
and the non-idiomatic must precede it, e.g.,
Ihn reitet der Teufel (lit. Him rides the devil)
He is going bonkers
Florian Schfer (p.c.) provides the following definition of this idiom from a German
source: If the devil rides someone, this person acts blindfold, imprudent or stupid[ly];
the person is obsessed (by the devil) and wanton. Semantically, then, this is similarly
non-agentive in character, like a causative object-experiencer psych-predicate; interestingly, however, Schfer points out that the expression does passivize:
Note that there are many different classes of verbs which have derived subjects; object-experiencer
verbs are only one such class. Our claim here is not that all these idioms have the semantics of objectexperiencer structures. Rather, it is that they do not have base-generated Agentstheir subjects are derived.
For the Hebrew, Hungarian, and German examples we do not have the diagnostics yet to establish this
conclusively. Here we are merely claiming that the atypical word order patterns evident in these expressions
are suggestive that something unusual is going on syntacticallysomething which is consistent with,
though not diagnostic of, a derived subject status for these purported Agents. More broadly, we can ask
why all these purported counterexamples to the No Agent Idioms Hypothesis have such unusual syntactic
structures. If Agent-dependent idiomatic interpretations are in principle freely available in the language
faculty, why arent there more examples that have a syntax that is typical for agentive verbs?

Harley and Stone

() So mancher Motorradfahrer scheint vom Teufel geritten zu sein.

so many motorcycle riders seem by the Devil be ridden to be
Many motorcycle riders seem to be ridden by the devil.
Berit Gehrke (p.c.) notes that this passive is adjectival rather than verbal; if sein
be is replaced with werden become here, the idiomatic reading is lost. See Gehrke
(Chapter above) for discussion of the absence of agentive content with adjectival,
but not verbal, passives in German. It does seem that active versions of this idiom
are possible; Martin Weigand (p.c.) suggests that questioning the identity of the devil
makes the active particularly felicitous: Welcher Teufel hat dich geritten Which devil
has ridden you?
That leaves us with a few English examples to think about, which we collect below
for convenience:
() a.

Lady Luck smiled on him.

The vultures are circling him.
The spirit moved him.
A little bird told me.

(Horvath and Siloni)

(Nunberg et al.)

We suggest that neither (a) nor (b) qualify as idiomatic DPV combinations, for
the following reason: insofar as they are idiomatic rather than metaphorical, they are
clearly DP idioms, like The Big Apple, not SubjV idioms. Both Lady Luck and vultures
can occur with other verbs in other positions, and still receive the same interpretation:
I hate those payday loan vultures; He has never once been visited by Lady Luck, Lady
Luck laughed in his face.
The case in (c) is perhaps more of a challenge, but it seems plausible to consider
it to fall within the range of object-experiencer constructions, at least semantically.
The subject is clearly a Causer, not an Agent, and the object seems equally clearly
to be an Experiencer. In addition, it is not clear that the DPCauser + V combination
is necessarily linked by idiomaticity, rather than simple compositionality, once the
spirit is understood in an idiomatic/metaphoric way. Parallel structures with different
subjects which receive this (object-experiencer) meaning for move are perfectly fine:
A reviewer asserts that although Lady Luck can have independent uses, Lady Luck smiled on X
nonetheless requires listing as an idiomatic expression, as the same meaning cannot be expressed by other
verbs with literal content similar to smile: Lady Luck grinned at him, for example. However, this is expected:
smile (on) has a conferred favour meaning that grin, laugh, etc. do notand it has this meaning quite
independently of cooccurrence with Lady Luck. We can say, with a similar conferred favour meaning:
The fates smiled on him, The king smiled on him, etc. ( The king grinned at him does not have the meaning
conferred favour either). Consequently, the DP Lady Luck is expected to be able to occur (compositionally)
with the verbal expression smile on. No idiomaticity is necessary to explain the meaning of this expression,
as the context-independent meanings associated with each of the elements account for the meaning of the
expression. Similarly, whatever meaning Lady Luck grinned at him has, it has by virtue of the meanings of
the individual components, Lady Luck (a personification of fate, fortune, etc.) and grinned at him (which
can suggest an implication of mockery, perhaps). Again, no idiomatic listing for the whole collocation is
necessary (or, according to the hypothesis here, possible).

The No Agent Idioms Hypothesis

() a. Her plight moved him (to tears/to speak out).

b. The attack moved him to enlist.
Having suggested that there is doubt concerning the status of (ac) as counterexamples, we are left with (d). It seems to meet one of the requirements needed in
a genuine counterexample: the subject is clearly an Agent, and is clearly interpreted
idiomatically. Bruening (), arguing along the same lines for this idiom as we have
done for (a,b) above, claims that (d) also is a DP idiom, citing examples such as
A little bird whispered/emailed it to me, A little bird is broadcasting that, and I heard it
from a little bird. To these observations we will add the point that in this expression,
the head verb tell still means literal tell, unlike the typical idioms discussed above
where the verbs literal meaning is clearly absent (kick, keep, take, throw, hit, kiss).
(It is worth noting that two other English examples mentioned by Bruening clearly
fall into the object-experiencer category: The love bug bit X and Whats eating X?,
further confirmation of the susceptibility of the Experiencer thematic structure to
To summarize, we have argued that there are at least reasonable grounds for suspicion that all counterexamples to the No Agent Idioms Hypothesis admit of one of two
principled analyses. Either they are derived-subject structures, in which the subject is
actually a promoted VP-internal Causer argument, or the idiomatic status is limited
to the subject DP alone, and the verb itself is not fixed in the expression, nor is its
interpretation special to that subject.
It is worth emphasizing again that if Belletti and Rizzi () and Pesetsky ()
are correct concerning the VP-internal status of subjects of object-experiencer predicates, cases of such predicates with idiomatic subjects, rather than being counterexamples, are actually predicted to exist by Marantzs generalization. We expect the
verb to potentially enter into an idiomatic relationship with its internal arguments,
and it would be suspicious, given the independent grounds for believing that such
Causer arguments are indeed internal, if there were no idioms which involved such

. Conclusions
We conclude that Marantzs generalization remains empirically unchallenged, and
that it is reasonable to take it as the foundation from which to draw theoretical
conclusions about the status of Agent arguments in syntax and semantics. In this
section, we consider the implications of this conclusion for theories of the syntax
semantics interface, and predictions for various domains of investigation that seem
to us to be indicated.
Above we have suggested that the explanation for the distribution of idiomatic
domains could be the status of the phase as a Spell-out domain; interpretations can

Harley and Stone

only depend on constituents contained within a single phase (Stone ). Kratzer

() proposes a more purely semantic account. Below, we sketch Kratzers proposal
and draw attention to a theoretical consequence which contrasts with that of the
phase-based approach.
Kratzers (, ) account of Marantzs lists of special Verb+Object interpretations treats verbal idiomaticity as contextually dependent polysemy, as described
above, putting the idiomatic meaning entirely in the truth-conditions specified in the
verbs semantics. As noted above, this move is justified on conceptual grounds if we
want to preserve compositionality and avoid construction grammar-style listing of
phrasal meanings. Idiomatic interpretation must be linked to individual lexical items.
Context-dependent meanings for individual lexical items are justified independently
on empirical grounds, for items like jinks, caboodle, and gamut (see () above), which
only occur in particular lexical and structural contexts. Phrasal meaning remains a
function of the meaning of the lexical items in the sentence and the way they are put
Kratzers semantic proposal restricts the verbal truth-conditions from referring to
the content of the external argument by proposing that the external argument is not
directly semantically selected by the verb, but rather is introduced by a separate predicate, a Voice predicate. In other words, context-dependent meanings in her approach
can only be conditioned by arguments which are directly semantically selected by the
relevant head. As Harley () notes, this has a significant consequence for the syntax
of internal arguments: if dependent idiomatic semantics can be conditioned only
by arguments of the relevant head, this entails that internal arguments must be direct
arguments of their verb roots, contra several proposals according to which internal
arguments, like external ones, are introduced by dedicated functional projections (e.g.
Ramchand , Borer , , Basilico ). This contrasts with the phasebased domains of interpretation view, where selection of the conditioning constituent
need not be relevant at all; rather, any constituent contained within the same phase as
the idiomatically interpreted head is accessible to provide the relevant conditioning,
regardless of whether it is selected for by the head in question. The different empirical predictions made by these two approaches to conditioning of special meanings
is an area ripe for further investigation. For example, it is worth asking whether
Note that Kratzers polysemous analysis is motivated by the original light verb kinds of cases discussed by Marantz, where the particular identity of the object is not fixed, but the idiomatic interpretation
is conditioned by objects which are members of a particular semantic class. This approach allows for a
kind of continuum of idiomaticity, according to which fixed idioms are subject to exactly the same kind
of interpretive process but simply more constrained in the specification of their triggering environment,
which is dependent on a particular DP, rather than on a particular semantic class of DPs. Caboodle items
are the logical extreme, where no context-independent interpretation exists. Note that it is very likely that
caboodle items are very common in the I-language of speakers, where an unfamiliar lexical item is learned
within a particular collocation and is only used by that speaker in that context, although in E-language as
a whole it might have an independent existence. See the Eggcorn Database (http://eggcorns.lascribe.net/)
for several plausible cases.

The No Agent Idioms Hypothesis

non-s-elected adjuncts within the vP may condition special interpretations for the
verb; this would be possible on the phase-based approach, though perhaps not on
Whether the conditioning environment for the idiomatically defined head is
constrained via semantic composition or via syntactic Spell-out, the basic contextdependent meaning approach comports well with several other ideas about the relationship of meaning to structure from various theoretical domains. For example,
the notion that idiomatic interpretation is crucially dependent on the idiosyncratic
content of one or more heads is consistent with Koopman and Sportiches proposal
concerning the headedness of idioms.
Another promising connection can be found in the treatment of lexical meaning in the Distributed Morphology framework. In the DM view, Root elements are
the only locus at which linguistic structure interfaces with encyclopedic knowledge.
If contextually-dependent special truth-conditions are encyclopedic in nature, they
should only be statable for Roots. Idioms, then, should have to contain Roots, and be
unable to consist purely of functional structure. The location of idiomaticity in Root
elements entails that functional material in the clause will still necessarily contribute
its compositional content to the interpretation of the sentence, exactly as argued by
McGinnis ().
Locating idiomatic meaning in truth-conditions associated with Roots also makes
predictions in the psycholinguistic domain. It entails that activation of a roots lexical
entry will at least temporarily activate all its interpretations, idiomatic as well as literal,
until the conditioning environment is detected and the system settles on the most
likely licensed interpretation. Early activation of multiple meanings of this kind has
been detected both behaviourally and in imaging studies in the case of homophones
like batAnimal and batSportingEquip ; see e.g. Gunter, Wagner, and Friederici ().
We consider that investigation of such activation with idiomatic interpretation is a
promising direction for future research.
Another open research question is whether there are other syntactic constraints on
idiomatic structures. We have argued that the No Agents constraint holds for idioms
with open variables which must then compose with other, non-idiomatic material.
However, in proverbial expressions, where the whole sentential structure is fixed, as
in Thats the pot calling the kettle black, it seems to be the case that a genuinely agentive
argument receives an idiomatic interpretation in the context of the whole. Pending a
satisfactory definition of what it means to receive an idiomatic interpretation, it could
be that such fully fixed expressions are relevant to the debate concerning the nature
of the No Agent constraint, bearing, perhaps, on whether it is a syntactic or semantic

Though see Glasbey () and McGinnis () for further debate.

Part IV
Lexical Items at the Interfaces

On the Syntax and Semantics
of the Japanese Comparative

. Introduction
In this chapter, I examine a particular comparative construction in Japanesethe socalled yori-comparative, exemplified in ().
() Watashi-no musuko-wa watashi yori(mo) se-ga
son-top I
yori-mo height-nom tall
My sons taller than me.
(Oda : ())
This is discussed alongside an English comparative construction that as far as I know
has received little attention before nowthe of-phrase comparative, as in Of John and
Mary, John is taller. These classes of comparative are compared with a thirdthe compared to comparative, exemplified by Compared to John and Mary, John is tall (Beck,
Oda, and Sugisaki , Kennedy ). Close comparison of yori, the of-phrase, and
compared to reveals that they are broadly alike in their syntactic properties, but exhibit
fine-grained distinctions in their lexical semantics that have significant consequences
for the interpretations of the sentences in which they occur.
My project is situated in the context of recent investigations of the crosslinguistic
semantics of comparison (Beck et al. , Oda , Kennedy , Beck et al. ,
Hohaus , Pearson ). The goal of these studies is to identify the dimensions
along which languages vary in how they form comparatives. For instance, Kennedy
For Japanese judgements I am grateful to Hiroki Narita; any data not accompanied by a citation were
elicited from him. My understanding of the semantics of the comparative was much improved by a very
enjoyable class on that topic taught at MIT by Sigrid Beck and Irene Heim in spring . For helpful
comments, thanks to an anonymous reviewer, Amy Rose Deal, Michael Yoshitaka Erlewine, Peter Jenks,
Li Julie Jiang, Toshiko Oda, Orin Percus, Maria Polinsky, and to the audiences of the Language Universals
and Linguistic Fieldwork workshop at Harvard University and of OnLI II. I am responsible for all mistakes
and shortcomings.


() considers the possibility that variation may be found in whether a language

employs so-called implicit or explicit comparison. () provides Kennedys definitions.
() a. Explicit comparison
Establish an ordering between objects x and y with respect to gradable property g using a morphosyntactic form whose conventional meaning has the
consequence that the degree to which x is g exceeds the degree to which y
is g.
b. Implicit comparison
Establish an ordering between objects x and y with respect to gradable property g using the positive form by manipulating the context such that the
positive form is true of x and false of y.
This distinction is inspired by the observation that although comparisons are prototypically formed by employing comparative morphology and a than-constituent, as
in (), they need not be, as illustrated by ().
() John is taller than Mary.
() Compared to Mary, John is tall.
() and () exemplify different strategies for communicating that Johns height is
greater than Marys. () is an example of an explicit comparison (EC) sentence,
whereas () instantiates implicit comparison (IC). While English has the resources
to employ either strategy, there is debate over whether there are languages that only
employ IC, with Japanese supplying a test case. I argue that Japanese is indeed an
implicit-comparison-only language. This aligns my position broadly with the analyses of Beck et al. () and Oda (), although I recognize the limitations of
their accounts identified in Kennedy (). I argue that these problems are due
to treating implicit comparison constructions as a homogeneous class. Instead, I
adopt Pearsons () division of IC into strong implicit comparison (SIC) and
weak implicit comparison (WIC). Kennedys arguments, which focus on Japanese
comparatives employing the postposition yori, show only that yori-comparatives do
not instantiate strong implicit comparison; they do not argue against treating yoricomparatives in terms of weak implicit comparison. I argue that this is the correct classification for yori-comparatives, and give a corresponding semantics for this
In order to motivate the distinction between strong and weak IC, it is instructive to examine the of-phrase comparative. Here are some more examples of this
() a. Of John and Mary, John is the tall one.
b. Of John and Mary, only John is tall.

The Japanese Comparative

According to Kennedy, a construction must have two properties in order to belong

to the IC category: (i) the positive form of the gradable predicate must be used (no
comparative morphology should be employed), and (ii) the context should be manipulated so as to render the predicate true of one individual and false of the other. It is
in this sense that the comparison can be regarded as implicit: if P denotes a gradable
property, and A is P and B is not P, then it must be the case that A is more P than B is.
Let us check that the sentences in () satisfy these criteria. The adjective is indeed used
in its positive form. Moreover, while the sentences assert that John is tall, they do not
assert that Mary is not tall. This latter claim is instead established indirectly. Here is
an informal way of thinking about how this comes about. Suppose that the function
of the phrase of John and Mary is to modify the domain of discourse by specifying
that the only individuals that it includes be John and Mary. Now consider (a). By the
uniqueness presupposition of the definite article, if John is the tall one, then Mary is
not tall. Likewise, if only John is tall (b), then Mary is not tall. (a) and (b) therefore
both communicate that John is tall and Mary is not talland hence that John is taller
than Marybut unlike with explicit comparison, this comes about via manipulation
of the context. In this way, the second condition for implicit comparison is satisfied
by these sentences.
Are there constructions in English that meet only one of Kennedys criteria for
implicit comparison? A slight alteration of the sentences in () and (a) reveals that
there are.
() a. Compared to Mary, John is taller.
b. Of John and Mary, John is the taller one.
Crucially, (a) and (b) both involve comparative morphology, violating the condition
that the positive form of the gradable predicate be used. Yet these are not cases of
explicit comparison: (a) tells us that explicit comparison requires a morphosyntactic
form that ensures that the degree to which x is tall (say) exceeds the degree to which
y is tall. To see what lies behind this idea, let us lay out some traditional assumptions
concerning the syntax and semantics of explicit comparison sentences:
() Syntax and semantics of explicit comparison (von Stechow , Heim ; see
Beck for a survey)
(i) Gradable predicates such as tall denote relations between degrees and individuals, type d, e, t.
(ii) In the syntax, a head Deg is present, which projects DegP.
(iii) Material occupying Deg either:

I assume a typed metalanguage where the basic types are e, t, and dthe types of individuals, truth
values, and degrees.

(a) denotes the degree argument of the gradable predicate (e.g. feet in
feet tall), or
(b) combines with another constituent to form a quantifier over degrees
(comparative morphemes e.g. -er, -est, less, too, enough).
(iv) The constituent with which the comparative morpheme combines is typically a than-constituent such as than Mary (is) and denotes the standard of

() illustrates one recent syntactic implementation of these ideas, due to Bhatt and
Pancheva ().
() John is taller than Mary is.







-erj than Mary is tall





The than-constituent than Mary is tall late merges with the comparative morpheme
-er in the position to which it has moved as a consequence of Quantifier Raising,
ensuring that -er and the than-clause form a constituent at LF. Furthermore, Bhatt
and Pancheva argue that the site of the Late Merger is the position in which the thanclause is pronounced; this position obligatorily marks the scope of the comparison.
The than-clause contributes the set of degrees to which Mary is tall. -er is a quantificational element denoting the proper subset relationship, which takes the denotation
of the than-clause as its first argument and the denotation of John is t tall as its second
argument, where t is the trace of -er, interpreted as a variable of type d. Given standard
assumptions about QR, the LF includes an index immediately below the position to
Whether the than-constituent in English is always clausal, or whether it may sometimes be a smaller
phrase is controversial. I adopt the term than-constituent in order to avoid a commitment to a particular
view on this issue.

The Japanese Comparative

which -er moves, which serves as a variable binder. This yields as the second argument
of -er a predicate of degreesthe set of degrees to which John is tall. (a) is a bracketed
representation of (), enriched to show degree abstraction, while the truth-conditions
are in (b).
() a. [DegP -er [than [d Mary is t t tall]]] [d John is t tall]
b. [[(a)]] = iff
{d : Mary is tall to degree d} {d : John is tall to degree d}
The result is a streamlined mapping between LF syntax and semantics: the semantics
of the comparative morpheme specifies that it seeks two arguments, each of which
describe a set of degrees, while the syntactic processes of Quantifier Raising and Late
Merger make these elements available to the interpretive component. Assuming this
characterization of explicit comparison, let us return to my claim that the sentences
in (), repeated in (), are not examples of EC constructions.
() a. Compared to Mary, John is taller.
b. Of John and Mary, John is the taller one.
The requirement that explicit comparison incorporate a morphosyntactic form that
accomplishes a given comparison involves not only comparative morphology, but also
a than-constituent that forms a constituent with the comparative morpheme at LF.
(a) and (b) involve no than-constituent, although one might wonder whether
compared to Mary and of John and Mary perform the same function. This hypothesis
obviously would not work for (b): while than-constituents introduce standards of
comparison, the of-phrase furnishes both the standard of comparison and the other
individual being compared. Moreover, if compared to Mary and of John and Mary
had the syntactic and semantic functions of a than-constituent, then the syntactic
distribution of these phrases would be expected to be confined to positions that let
them mark the scope of the comparison. This is not borne out: the examples below
show that the placement of these constituents is relatively free.
() a. John is taller than Mary.
b. John is taller compared to Mary.
c. John is the taller one of John and Mary.
() a. John is than Mary taller.
b. John is, compared to Mary, taller.
c. John is, of John and Mary, the taller one.
() a. John than Mary is taller.
b. John, compared to Mary, is taller.
c. ?John, of John and Mary, is the taller one.


() a. Than Mary John is taller.

b. Compared to Mary, John is taller.
c. Of John and Mary, John is the taller one.
The relative freedom in the placement of the compared to-phrase and the of-phrase
suggests that they are syntactic adjuncts, and I conclude that they are not good candidates as arguments of -er. Rather, I assume that in (), the first argument of the
comparative morpheme is left implicit, as in ().
() John is taller.
In a context that provides a suitable standard of comparison (e.g. one where Mary
and John are being compared along various dimensions), () is acceptable, despite
the lack of a than-constituent.
I have shown that (a) and (b) are not cases of explicit comparison, but since
they involve comparative morphology they do not conform to Kennedys definition of
implicit comparison either. Moreover, I would like to capture the fact that these sentences are minimally different from prototypical cases of IC: they differ only insofar
as comparative morphology is employed. Following my earlier work (Pearson ),
I therefore propose to distinguish two classes of implicit comparison. The kind that
satisfies both of Kennedys criteria I call strong implicit comparison (SIC). That which
employs comparative morphology, but which does not qualify as explicit comparison,
I call weak implicit comparison (WIC). WIC involves a piece of morphology, the
comparative morpheme -er, whose lexical entry requires it to take a semantic argument that serves as the standard of comparison, but no overt material supplies this
The remainder of this chapter is devoted to arguing that the yori-comparative in
Japanese has the properties just described. I make this argument by establishing two
things: (i) yori-comparatives do not instantiate SIC, which is to say that they involve a
(covert) comparative morpheme; and (ii) yori-comparatives instantiate implicit rather
than explicit comparison, which is to say that the standard of comparison argument
that the comparative morpheme seeks is not provided by a than-constituent. One of
the questions raised by this analysis is why yori-comparatives cannot be interpreted
as strong implicit comparison sentenceswhy is the covert comparative morpheme
obligatory? I answer this question by extending observations concerning the semantics of the English of-phrase comparative to the yori-comparative, so that I am in
a position to provide a close comparison of the syntax and semantics of these two
The chapter is structured as follows. Section . discusses the yori-comparative,
analyses of it in earlier work, and their shortcomings. My proposal is described in
As far as I know, the argument that cases such as () show, that the first argument of -er may sometimes
be left implicit, was first made by Beck et al. ().

The Japanese Comparative

section ., while its consequences are discussed in section .. One theme that
emerges is that although yori, the of-phrase, and compared to are broadly alike in
syntactic properties, the semantics of compared to is significantly different from that
of yori and the of-phrase. These lexical distinctions have consequences for whether the
constructions in which the items participate involve weak or strong implicit comparison, thereby partially determining the inventory of modes of comparision exhibited
by the languages in question. These issues are summarized in section ..

. The Japanese Comparative

.. Yori-comparatives
Kennedy () identifies three interpretive properties of English IC, and shows that
yori-comparatives lack each of them. The first property involves contexts requiring
so-called crisp judgements. In these contexts, the individuals being compared with
respect to some property differ only slightly in the degree to which they possess that
property. Suppose for example that Mary is foot tall and John is foot inch tall.
Kennedy observes that in such a context EC is felicitous, but IC is not:
() Context: Mary is foot tall and John is foot inch tall.
a. John is taller than Mary.
b. Compared to Mary, John is tall.
Kennedy notes that a yori-comparative is felicitous in a context requiring crisp judgements. The one that he considers involves comparing two papers, one of which is
words long, the other :
() Kono peepaa-wa ano peepaa yori nagai.
this paper-top that paper yori long
This paper is longer than that one.

(Kennedy : ())

Kennedys second diagnostic concerns so-called minimum standard gradable adjectives: adjectives P such that all it takes for x to be described as P is for x to possess the
relevant property to some degree. Tall is not a minimum standard gradable adjective,
since possessing some degree of height does not warrant being described as tall. Bent,
on the other hand, is a minimum standard gradable adjective, since an object that has
some degree of bend qualifies as bent. Kennedy identifies a restriction on the use of
Actually, he identifies four properties of IC, but only employs three of these in his investigation of
One might wonder whether the infelicity of (b) is due not to Marys and Johns closeness in height,
but to the fact that any height over foot might be considered tall. Indeed, Kennedy notes that implicit
comparisons carry an implicature that the predicate does not hold of the standard of comparison. The
reader is invited to check that this is not what is at work here by verifying that the sentence is also infelicitous
in a context where Mary is foot inches tall and John is only inch taller than that.


minimum standard gradable adjectives in IC constructions that does not apply to EC.
This restriction reveals itself in a context where the minimum standard adjective is
true of both of the compared objects, as in (). Explicit comparison is felicitous in
this context (a), but implicit comparison is not (b).
() Context: Rod A and rod B are bent, rod B more so than rod A.
a. B is more bent than A.
b. ??Compared to A, B is bent.

(Kennedy : ())

Kennedy notes that in this context, a yori-comparative is felicitous ().

() Ano sao-wa kono sao yori magatte-iru.
that rod-top this rod yori bent-be
That rod is more bent than this rod.
(Adapted from Kennedy : ())
Finally, Kennedy observes that differential measure phrases are available with explicit
but not implicit comparison.
() a. John is one inch taller than Mary/John is taller than Mary by one inch.
b. Compared to Mary, John is one inch tall/Compared to Mary, John is tall
by one inch.
He reports that differential measure phrases are available with yori-comparatives,
() Gozira-wa Rodan yori hyaku-ton omoi.
Godzilla-top Rodan yori -tons heavy
Godzilla is tons heavier than Rodan.
Kennedy concludes that yori-comparatives instantiate explicit rather than implicit
comparison. This conclusion is a little hasty, however. The diagnostics all involve
properties of IC constructions that employ the positive form of the predicate, thereby
diagnosing strong implicit comparison. English weak implicit comparison exhibits
none of the properties Kennedy identified for SIC, as can be verified by adding
comparative morphology in (b), (b), and (b). While Kennedys arguments
show that yori-comparatives do not involve strong implicit comparison, they do not
exclude the possibility that they instantiate weak implicit comparisonthat is, that
yori-comparatives have a covert comparative morpheme, even though there is no
overt material to supply its first argument. I think that a WIC analysis is preferable
In Kennedys original example, magatte-iru is modified by motto, yielding an expression that means be
very bent. While the status of motto is controversial (see fn. below), modification of a minimum standard
predicate by this expression arguably produces a relative predicate, and is therefore best avoided for this
test. The judgement remains unchanged.
In fact, matters are more complicated still. In Pearson () I argued that Kennedys diagnostics
involving crisp judgements and differential measure phrases are not good tests for strong implicit com-

The Japanese Comparative

to an EC analysis. In this, I concur with Oda (), although in the next subsection
I identify shortcomings of her implementation.
.. Oda ()
Oda () accepts Kennedys insight that yori-comparatives cannot be assigned
essentially the same semantics as a sentence of form Compared to A, B is P. However,
she rejects his conclusion that they instantiate explicit comparison. Drawing on Beck
et al. (), she provides evidence that the contribution of a yori-constituent is different from that of a than-constituent. First, she calls attention to the ability of contextual
information to contribute to the recovery of the standard of comparison in yoricomparatives. This property is present for both weak and strong implicit comparison,
but not for explicit comparison, as the dialogue in () establishes.
() A: How does your sons height compare to yours when you were his age?
B: My sons taller than me/than I am.
Compared to me, my sons tall.
Compared to me, my sons taller.
Building on Beck et al. () (who credit personal communication from an audience
member at the University of Delaware), Oda observes that, like English IC, a yoricomparative is a felicitous answer to As question in ().
() Watashi-no musuko-wa watashi yori(mo) se-ga
son-top I
yori-mo height-nom tall
My sons taller than me.
(Oda : ())
Secondly, Beck et al. point out that yori, unlike than, can occur with a non-gradable
predicate ().
() a. Ken-wa yooroppa-yori amerika-ni iku-koto-ni kimeta.
Ken-top Europe-yori America-to go-fact-dat decided
Ken decided to go to America rather than Europe. (Beck et al. : ())
b. Ken decided to go to America than Europe.
Oda concludes that the yori-constituent does not supply an argument for a comparative morpheme, but rather provides information about the context which,
indirectly, enables a comparative meaning to be derived. Moreover, she responds to
Kennedys arguments by proposing that comparative morphology is present in yoricomparatives. In my terms, this amounts to saying that yori-comparatives involve
parison. However, I think that the test with minimum standard gradable adjectives is reliable, and so I set
this issue aside for my present purposes.
As far as I know, this contrast between English explicit and implicit comparison has not been investigated in detail, a task I leave to future research.


WIC: one but not both of Kennedys criteria for implicit comparison is satisfied. This
is the classification of yori-comparatives that I endorse in this chapter.
This view raises the question of where the comparative meaning appears, given
that it does not surface overtly. Here Odas proposal and mine part company. Odas
answer is that there is no dedicated morpheme meaning -er, but rather that Japanese
adjectives are inherently comparative in meaning. A sample lexical entry is in ().
() [[takai]] = x.max(d.tall(d)(x)) > c.
(Based on Oda : (a))
where (i) c is a contextually supplied standard; (ii) for any set X, max(X) is Xs
largest element.
According to this view, takai denotes the set of individuals that are taller than some
contextually supplied standard of height. This predicts that there is no strong implicit
comparison in Japanese: by definition, SIC involves the positive form of the adjective,
yet according to Oda, Japanese adjectives have no positive form. This prediction is
not borne out: there is a class of comparatives, which I call kurabetara/kuraberuto
comparatives, that pass the diagnostics for SIC. () provides examples.
() a. Ziro-ni kurabe-ta-ra
Taro-wa se-ga
Ziro-dat compare-past-cond Taro-top height-nom tall
Compared to Ziro, Taro is tall.
b. Ziro-ni kurabe-ru-to
Taro-wa se-ga
Ziro-dat compare-pres-cond Taro-top height-nom tall
Compared to Ziro, Taro is tall.
(Sawada : ())
Sawada () demonstrates that this construction exhibits the hallmarks of strong
implicit comparison. For reasons of space, I discuss only the crisp judgement diagnostic; the reader is referred to Sawadas work for additional data. Unlike yoricomparatives, kurabetara/kuraberuto comparatives are infelicitous in crisp judgement
() Context: There are two papers. One is pages long and the other is pages
a. Ano peepaa-ni kurabe-tara
kono peepaa-wa nagai.
that paper-dat compare-cond this paper-top long
Compared to that paper, this paper is long.
kono peepaa-wa nagai.
b. Ano peepaa-ni kurabe-ru-to
that paper-dat compare-pres-cond this paper-top long
Compared to that paper, this paper is long.
I conclude that the view that Japanese gradable predicates are inherently comparative
is incorrect. On the other hand, I am convinced by Beck et al.s arguments that
yori-comparatives do not involve EC. Japanese has both strong and weak implicit

The Japanese Comparative

comparison, the former arising with kurabetara/kuraberuto-comparatives, the latter

with yori-comparatives. That Japanese has weak implicit comparison but no overt
comparative morpheme implies that there is a covert comparative morpheme with
roughly the same meaning as more, which finds its standard of comparison argument
contextually rather than being fed it by a than-constituent. , My next task is to
explain the distribution of the comparative morpheme, which I have said is present
in yori-comparatives but not in kurabetara/kuraberuto comparatives.

. The Proposal
My starting point is the of-phrase comparative, which does not let the positive form
of the gradable predicate stand on its own.
() ??Of John and Mary, John is tall.
English has at least four ways of rescuing (). It can employ comparative morphology
(a), embed the predicate within a definite description (b), use the focus-sensitive
item only (c), or apply focal stress to the subject (d).
() a.

Of John and Mary, John is taller.

Of John and Mary, John is the tall one.
Of John and Mary, only John is tall.
Of John and Mary, [John]F is tall.

These strategies have in common that they communicate that Mary is less tall than
John. (), on the other hand, fails to say anything about Mary; it communicates
just what is communicated by John is tall. The of-phrase adds nothing further. This
redundancy is responsible for the degraded status of (). I propose that yori plays
a similar role to the of-phrase. However, Japanese has a rescue strategy that English
lacks: redundancy on a par with () can be avoided in yori-comparatives by means
of covert comparative morphology. To develop the proposal in detail, I firstly give a
semantics for English of-phrase comparatives.

In some work, motto much, very is taken to be the morpheme associated with the comparative of
superiority. See Beck et al. () for arguments that this gloss is incorrect.
An anonymous reviewer suggests that yori might itself be the comparative morpheme. One argument
against this is that yori-constructions do not always express comparisons, as we shall see in the next section.
An anonymous reviewer reports finding () acceptable. I suspect that variations in judgement
concerning this sentence reflect the degree to which speakers are prepared to assign it an interpretation
involving focus or covert exhaustificationin essence, an unpronounced counterpart of the only that
renders the sentence acceptable in (c).
(d) may be less acceptable for some speakers than (ac); however it seems to be an improvement
on ().


Section . introduced the idea that the of-phrase restricts attention to contexts
whose domains of discourse have a certain structure. We can model this in a Kaplanian semantics that distinguishes two levels of meaning. A character is a function
from a context to a content, which is in turn a function from a world to an extension
(entities in the world or truth values). Kaplans notion of context is intended to provide
the elements required in order to furnish referents for indexicals, for example by
specifying who the speaker is in the case of I. I shall enrich this by letting the domain
of discourse U be a contextual parameter too. Additionally, I add to the metalanguage
the basic types k, the type of contexts, and s, the type of worlds. The character of
a sentence is of type k, s, t, that of a one-place predicate type e, k, s, t, and
so on.
We are now in a position to model the semantics of the of-phrase by manipulating
the context parameter. I assume that the of-phrase is a sentence modifier, type
k, s, t, k, s, t. Its semantics is as follows.
() [[of x and y]] = k,s,t ck ws .(c[U = {x, y}])(w) =
Where c[U = {a, b}] is that context c that is just like c except that the domain
of discourse of c consists only of a and b.
Let us see what interpretation this semantics assigns to ??Of John and Mary, John
is tall. I need a lexical entry for tall, along with a semantics for the operator POS,
which is an existential quantifier over degrees, and therefore indirectly saturates the
degree argument of a predicate like tall in the absence of comparative morphology.
The following suffice for my purposes.
() [[tall]] = dd xe ck ws .x is d-tall in w.
() [[POS]] = Dd,t ck ws .d D : d s
Where s is the contextually salient degree standard.
A sentence of form P(x), where P is a gradable predicate used in its positive form,
asserts that there is some degree d to which x is P such that d is at least as great as

See Percus () for extensive discussion of this framework.

Readers may be concerned that by letting the semantics of the of-phrase manipulate the context
parameter, we obtain a system that violates Kaplans () prohibition against monsters; we appear to
be contradicting his claim that operators that manipulate the context parameter are absent from natural
language. On the other hand, recent crosslinguistic work on indexical shift (e.g. Schlenker , ,
Anand and Nevins , Anand ) has undermined the empirical case for banning monsters, which was
originally based only on English. Moreover, the controversy over the status of monsters is really concerned
with whether (i) there are operators that quantify over contexts, much as modal operators quantify over
worlds, and (ii) the reference of an indexical expression can shift in the scope of such operators. The
modification of the context performed by the of-phrase given the semantics in () exhibits neither of
these characteristics.
The semantics is similar to that proposed for compared to in Kennedy () in that both induce a
shift to a context with a constrained domain of discourse.

The Japanese Comparative

the degree standard d made salient in the context. I illustrate this by providing the
interpretation of John is tall.
() a. John is tall.
b. LF: [[DegP POS] [d John is t -tall]]
c. [[(b)]] = ck ws .d {d : John is d-tall in w} : d s
The degraded sentence ??Of John and Mary, John is tall has the following meaning.
() a. ??Of John and Mary, John is tall.
b. [[(a)]] = [[Of John and Mary]](ck ws .d {d : John is d-tall in w} :
d s)
= (ck ws .d {d : John is d-tall in w} : d s)
This says that in a context in which the domain of discourse consists solely of John
and Mary, Johns height is at least as great as the contextually salient degree standard.
But this shifting of the domain does not necessarily have any effect on the degree
standard. Suppose that in the original context, the standard is just the average height
of a human being; indeed, this seems to be the standard with respect to which John
is tall is interpreted out of the blue. Notice that in order for this degree standard to
be available, it need not be the case that the domain of discourse includes the entire
human race; suppose we are discussing whether every syntactician is tall, confining
our attention to syntacticians who attended the conference On Linguistic Interfaces
. The question of interest might nonetheless be whether every such syntactician is
tall relative to the average human, rather than relative to, say, the semanticists and
phonologists at the conference. This suggests that the degree standard in a context and
the domain of discourse in that context are in principle independent of one another.
Consequently, adjusting the discourse domain to include only John and Mary does
not necessarily make any difference to the standard with respect to which John is tall
is evaluated. It is in this sense that the contextual shift induced by the of-phrase is
redundant in the simple case involving the positive form of the gradable predicate.
The next step is to state more explicitly how comparative morphology in (a) saves
the sentence from the degraded status of (). I said that in implicit comparison sentences like (a), there is no overt material that performs the syntactic and semantic
function of a than-constituent; the of-phrase supplies the standard of comparison in
an indirect way, by manipulating the context. Let us suppose then that in such a case,
For simplicity, I omit world and context variables from the LFs. I have in mind an extensional
framework where world arguments of predicates are encoded in the syntax as silent pronominals (Percus
); the same could be assumed for context arguments, as sketched e.g. in Percus ().
This will follow if the domain of discourse, but not the standard of comparison, is a parameter of the
Kaplanian context.
In this I disagree with Kennedy (, ), who explicitly assumes that manipulation of the context
affects which individuals are included in the extension of a gradable predicate.


the standard of comparison is represented syntactically as a covert individual argument, written proi . In section ., I said that the English comparative morpheme
undergoes QR to a position in which the than-clause merges, which contributes a
predicate of degrees as the standard of comparision. I now need a second morpheme,
-er , which takes an individual argument and is interpreted in situ:
() [[-er ]] = xe gd,e,k,s,t ye ck ws .max({d : g(d)(y)(c)(w) = }) >
max({d : g(d)(x)(c)(w) = })
I can now give the meaning of (a).
() a. Of John and Mary, John is taller proi .
b. [[John is taller proi ]] = ck ws .max({d : John is d-tall in w}) > max({d :
gc (i) is d-tall in w})
c. [[Of John and Mary]] = k,s,t ck ws .(c[U = {John, Mary}])(w) =
d. [[[[Of John and Mary] [John is taller proi ]]]] =
ck ws .[[John is taller proi ]](c[U = {John, Mary}])(w) =
ck ws .max({d : John is d-tall in w}) > max({d : gc[U={John,Mary}] (i) is
d-tall in w})
I assume that the referents of pronominal expressions are supplied by an assignment
functiona parameter of the context c, written gc . I also suppose that the range of
values that gc may supply is constrained by the domain of discourse U in c. Since in
the modified context U consists solely of Mary and John, and since it is necessarily
false that John is taller than himself, the only plausible value for proi is Mary. Hence
(a) is true in just those situations where John is taller than Mary.
Now consider the semantics of the yori-constituent. I propose that this is like the
semantics of the of-phrase except that the individuals mentioned in the complement
of of make up the domain of discourse in its entirety. By contrast, the complement of
yori denotes an individual that is in the domain of discourse, but the subject of the
gradable predicate should also be included in that set:
() [[yori]] = xe Pe,k,s,t ye ck ws .P(y)(c[U = {x, y}])(w) =
Let us check how the idea works using as our example (), repeated below.
() Watashi-no musuko-wa watashi yori(mo) se-ga
son-top I
yori-mo height-nom tall
My sons taller than me.
Recall my claim that yori-comparatives involve a covert comparative morpheme. I
firstly explore what would happen if this element were absent. Yori would compose
with its complement, watashi, me to yield a modifier for takai, tall. The resulting

I thank an anonymous reviewer for this suggestion.

The Japanese Comparative

predicate would take as its argument the subject watashino musukowa, my son. The
sentence would be true if and only if, in the context that is just like the actual context
except that the domain of discourse consists of me and my son, my son is tall. The
same thing goes wrong here as with ??Of John and Mary, John is tall. Nothing is
contributed by adjusting the context to one in which the domain of discourse consists
of the speakers son and the speaker. A covert comparative morpheme resolves this
redundancy. The next step is to give a semantics for this morpheme. I follow Kennedy
() in assuming that the Japanese counterpart of -er takes an individual rather than
a degree as its first argumenti.e. it has the semantics of English -er . Since I think
that yori-comparatives involve weak implicit comparison, this argument is supplied
by a covert pronominal proi ; this is analogous to my analysis of English WIC. The
truth-conditions of () are computed below.
() a. My son [[is tall -er proi ] [yori me]].
b. [[is tall -er proi ]] = [[-er ]]([[proi ]])([[tall]]) =
ye ck ws .max({d : tall(d)(y)(c)(w) = }) > max({d : tall(d)(gc (i))(c)(w)
= })
c. [[yori me]] = [[yori]](me) =
Pe,k,s,t ye ck ws .P(y)(c[U = {me, y}])(w) =
d. [[[[is tall -er proi ] [yori me]]]] = [[yori me]]([[is tall -er proi ]]) =
ye ck ws .[xe ck ws .max({d : tall(d)(x)(c )(w ) = }) >
max({d : tall(d)(gc (i))(c )(w ) = })](y)(c[U = {me, y}])(w) = =
ye ck ws .max({d : tall(d)(y)(c[U = {me, y}])(w) = }) >
max({d : tall(d)(gc[U={me,y}] (i))(c[U = {me, y}])(w) = })
e. [[My son [[is tall -er proi ] [yori me]]]] =
[[[[is tall -er proi ] [yori me]]]](my son) =
ck ws .max({d : tall(d)(my son)(c[U = {me, y}])(w) = }) >
max({d : tall(d)(gc[U={me,my son}] (i))(c[U = {me, my son}])(w) = })
(a) is true if and only if, in that modified context that is just like the actual context
except that the domain of discourse is the speaker and the speakers son, the height
of the speakers son exceeds the height of the individual supplied by the assignment
function. Since in such a context the only plausible standard is the speaker, I derive the
intuitively correct meaning, My son is taller than me. The following section discusses
some consequences of my proposal.

. Consequences
.. Non-comparative yori-constructions and of-phrase constructions
Recall that one of the motivations for treating yori-comparatives as examples of
implicit comparison was the ability of yori-phrases to occur in non-comparative
sentences, as in (a), repeated below.


() Ken-wa yooroppa-yori amerika-ni iku-koto-ni kimeta.

Ken-top Europe-yori America-to go-fact-dat decided
Ken decided to go to America rather than Europe.
I noted that this property is shared with of-phrase constructions:
() Of America and Europe, Ken decided to go to America.
I suggest that what makes these constructions available is the possibility of construing
them with focus on America. That this is the interpretation that is assigned to () can
be checked by comparing it with the following variant.
() Of Ken and Sue, Ken decided to go to America.
The occurrences of Ken decided to go to America in () and () cannot be rendered
with the same intonation: In (), focal stress must be assigned to America, while in
(), it must be assigned to Ken:
() a. Of America and Europe, Ken decided to go to [America]F .
b. Of America and Europe, [Ken]F decided to go to America.
() a. Of Ken and Sue, [Ken]F decided to go to America.
b. Of Ken and Sue, Ken decided to go to [America]F .
In general, with an of-phrase construction involving focus marking, the DP that is
focus marked must be mentioned in the of-phrase. I conclude that in these examples,
the role of the of-phrase is to provide the focus alternatives, a strategy that we have
already seen is available with of-phrase constructions involving gradable predicates:
() Of John and Mary, [John]F is tall.
Without going into focus semantics in any detail, lets simply say that a sentence
containing a focus-marked constituent communicates that, for every member of
the set of focus alternatives (excluding itself), that sentence obtained by replacing
with in is false. Furthermore, I assume that the set of focus alternatives may only
include elements of the domain of discourse. Now recall the semantics I am assuming
for the of-phrase:
() [[of x and y]] = k,s,t ck ws .(c[U = {x, y}])(w) =
The meaning of (a) is as follows.
() a. Of America and Europe, Ken decided to go to [America]F .
b. [[(a)]] = ck ws .[[of America and Europe]]([[Ken decided to go to
[America]F ]])(c)(w) =
ck ws .[[Ken decided to go to [America]F ]](c[U = {America, Europe})(w)

The Japanese Comparative

Assuming the view of focus semantics sketched informally above, () communicates

that John decided to go to America, and he did not decide to go to any other location
included among the focus alternatives of America. On the assumption that focus
alternatives are constrained by the domain of discourse, the contextual manipulation
performed by the of-phrase derives that John did not decide to go to Europe, in
accordance with our intuitions.
The counterpart of this sentence involving the yori-construction receives a parallel
analysis. I adopt the simplifying assumption that this sentence involves the twoplace predicate decide to go to; yori-phrases can be predicate modifiers of two-place
predicates as well as one-place predicates like tall. The semantics of yori is therefore
type-flexible; the entry appropriate to the case involving a two-place predicate appears
() [[yoriplace ]] =
xe Re,e,k,s,t ye ze ck ws .P(y)(z)(c[U = {x, y, z}])(w) =
In general, for any n-place predicate Q, yori combines with some individual x to form
a modifier of Q whose contribution is to shift the context to one whose domain of
discourse has n + members: it is that set consisting of x and Qs arguments. I assume
that in (), America is again focus marked:
() a. Ken decided to go to [America]F yori Europe.
b. [[yoriplace ]](Europe)(decide-to-go-to)(America)(Ken) =
cw.(decide-to-go-to)([America]F )(Ken)(c[U =
{Europe, America, Ken}])(w)
() communicates that Ken decided to go to America, and that he did not decide
to go to any other location from among Americas focus alternatives. I said that only
members of the domain of discourse can be included among the focus alternatives.
Given the contextual manipulation performed by the yori-phrase, this domain consists of Europe, America, and Ken. Plausibly, then, the set of focus alternatives is just
{Europe, America}, since Ken is not a location. I thereby derive that Ken did not decide
to go to Europe.
The view that yori-comparatives instantiate implicit comparison makes room for a
unified treatment of those constructions with cases where the yori-construction does
not have a comparative meaning. Adjusting with types where appropriate, a single
lexical entry for yori accommodates both types of case, given the assumption that the
domain of discourse may play various possible roles in recovering the meaning of yorisentences. Where it supplies a standard of comparison, a comparative meaning results,
while a non-comparative meaning is available if the domain of discourse constrains
focus alternatives. By adopting the view that of-phrases likewise constrain the domain
of discourse, I derive that of-phrase constructions also tolerate non-comparative interpretations.


.. Compared to comparatives
My analysis assumes that without comparative morphology, the yori-phrase would
fail to supply a standard of comparison, and would therefore be redundant. This
is responsible for the WIC status of yori-comparatives. Since my task is to explain
why yori-comparatives employ comparative morphology but kurabetara/kuraberuto
comparatives do not, the question arises why the kurabetara/kuraberuto-phrase is
not redundant as the yori-phrase would be in the absence of the covert comparative
morpheme. That is, I take it that in Japanese, -er is introduced as a last-resort strategy
in order to prevent a yori-comparative from having a degraded status on a par with
??Of John and Mary, John is tall. That kurabetara/kuraberuto comparatives pass the
tests for SIC suggests that they do not need comparative morphology, because there
is no question of the kurabetara/kuraberuto-phrase being redundant. In this section,
I sketch an explanation of this.
The first step is to understand why Compared to Mary, John is tall is acceptable
i.e. why the phrase compared to Mary is not redundant, even though tall is used in
its positive rather than comparative form. An option that is not available to us is to
appeal to Kennedys semantics for compared to:
() [[x is A compared to y]] = in context c iff [[x is A]] = in every context c that
is just like c except that the domain of discourse includes x and y.
(Kennedy : ())
Underlying Kennedys semantics is the assumption that adjustment of the domain
of discourse is enough to affect whether x counts as A, where A is the positive form
of some gradable predicate. But this is precisely what I denied in my discussion of
of-phrase constructions involving the positive form.
Instead, I treat compared to as a universal quantifier over worldindividual pairs.
This reflects the standard treatment of when-adjuncts and the antecedents of conditionals as unselective binders; it seems no accident that kurabetara and kuraberuto
bear conditional morphology, while x is A compared to y is equivalent to x is A when
compared to y. The ability of tall to be modified by such a constituent is in fact rather
surprising: individual-level predicates usually resist such modifiers when they take a
proper name subject.
() a. When he is healthy, John is tall.
b. When talking to Mary, John is tall.
() a. If he is healthy, John is tall.
b. If he talks to Mary, John is tall.
Intuitively, what underlies the degraded status of () and () is that since Johns
height is a permanent property, whether he is tall is unaffected by whether he is healthy
or whether he talks to Marysuch occurrences make no differences to his height. In

The Japanese Comparative

fact, whether John is compared to Mary makes no difference to his height either: his
height cannot change as a consequence of such an event. Yet we can say that when
compared to Mary, John is tall. I suggest that the reason for this is that, although
comparing John to Mary cannot change the properties that he has, such comparison
changes ones perceptions of those properties. That is, the constraint on modification
of ILPs with when-adjuncts is circumvented in the case of compared to because the
meaning of tall is coerced to mean roughly seem tall. Since seem tall is not individuallevelI may seem tall in one situation but not in another, depending on who I am
standing next to, saythere is no question of violating this constraint. Evidence in
favour of this view can be found in the equivalence of the sentences in (), which
disappears when compared to is omitted, as in (); compare the continuations in ().
() a. Compared to Mary, John is tall.
b. Compared to Mary, John seems tall.
() a. John is tall.
b. John seems tall.
() a. ?John seems tall compared to Mary, but he isnt tall compared to Mary.
b. John seems tall, but he isnt tall.
I conclude that compared to y is a universal quantifier over worldindividual pairs
w , z such that w involves a comparison with y by z; x is A compared to y asserts that
for every such pair, z perceives x to be A.
() [[compared to y]] = k,s,t ck ws .w, z[Acc(w, w )& w involves a
comparison with y by z] [z perceives (c)(w ) to be true]
Where Acc(w, w ) iff w is like w in all relevant respects.
We are now in a position to understand why Compared to Mary, John is tall is
acceptablethat is, why the phrase compared to Mary is not redundant, even though
tall is used in its positive rather than comparative form. Below I display the composition of the compared to phrase with John is tall.
() a. Compared to Mary, John is tall.
b. [Compared to Mary] [[Deg POS] [d [John is [d -tall]]]]
c. [[[Deg POS] [d [John is d -tall]]]] = cw.d {d : John is d -tall in w} :
d. [[(b)]] = cw.w, z[Acc(w, w )& w involves a comparison with Mary by z]
[z perceives it to be true that d {d : John is d -tall in w } : d s]
This says that in any accessible world involving a comparison with Mary, one perceives
Johns height to be at least as great as the contextually salient standard for tallness.
Strong implicit comparison in Japanese with kurabetara/kuraberuto works in the same


way. The role of the compared to or kurabetara/kuraberuto phrase is to quantify over

situations that have a property that affects ones perception of Johns heightnamely
the property of involving comparison with Maryletting us assert that in any such
situation, one perceives John as tall. The kurabetara/kuraberuto-phrase is therefore
not redundant, and so there is no need to employ a covert comparative morpheme.
Indeed, if covert material is only introduced where necessary, then the comparative
morpheme will never be employed in kurabetara/kuraberuto comparatives.
An advantage of this proposal is that it captures a difference between of-phrases and
compared to in English and between yori-constructions and kurabetara/kuraberuto
constructions in Japanese. There are cases where the main predicate in the compared
to construction need not be gradable:
() Compared to Sue, Mary is Mother Teresa.
The corresponding sentence with an of-phrase rather than compared to is unacceptable.
() Of Mary and Sue, Mary is Mother Teresa.
In Japanese, the kurabetara/kuraberuto construction patterns like (), and the yori
construction like ():
() a. Sue ni/to
kurabetara/kuraberuto, Mary-wa Mazaa Teresa-da.
Sue dat/with compared
Mary-top Mother Teresa-cop
Compared to Sue, Mary is Mother Teresa.
b. Mary-wa Sue-yori Mazaa Teresa-da.
Mary-top Sue-yori Mother Teresa-cop
Lit.: Mary is more Mother Teresa than Sue.
The degraded status of (b) is no surprise given my account: since be Mother Teresa
is not gradable, the rescue strategy of introducing covert comparative morphology is
unavailable. On the other hand, the acceptability of () and its Japanese counterpart
(b) is straightforwardly explained: both sentences assert that in any world involving
comparison with Sue, Mary resembles Mother Teresa.

. Conclusion
Let us summarize what we have learned about the syntax and semantics of the
of-phrase, yori, and compared to. In section ., I presented arguments that the of

An anonymous reviewer observes that conditional morphology improves this example:

(i) Of Mary and Sue, Mary would be Mother Teresa (if we had to pick one).
This seems to have a different meaning from (), however: there is no suggestion that placing Mary next
to Sue makes her resemble Mother Teresa.

The Japanese Comparative

phrase and the compared to phrase are syntactic adjuncts. Space considerations have
prevented me from discussing the syntax of the yori-phrase in detail, but Beck et al.
() propose that this constituent has the syntactic status of a relative clause; I refer
the reader to their paper for arguments in favour of this view. If so, then each of the
implicit comparison constructions this chapter has been concerned with (indirectly)
provides its standard of comparison by means of a syntactic adjunct. It seems then that
the implicit/explicit comparison distinction can be characterized in purely syntactic
terms, with reference to whether the material that provides the standard of comparison is an argument or adjunct. I leave it to future work on a wider range of languages
to confirm or falsify this hypothesis.
I would like to end, however, by observing that when we look at a finer-grained
distinctionthat between weak and strong implicit comparisonthe category of
a given comparative construction does not appear to be deducible from syntactic
properties alone. Consider the yori-comparative and the kurabetara/kuraberuto comparative. I argued that the former employs a covert comparative morpheme, but the
latter does not, thereby predicting that yori-comparatives display the characteristics
of weak implicit comparison, and kurabetara/kuraberuto comparatives exhibit those
of strong implicit comparison. Suppose I am correct that these facts follow from the
different semantics of yori and kurabetara/kuraberutoone producing an adjusted
context, and the other being a modal quantifier. This would imply that in the end, the
question of which mode of comparison a given constituent gives rise to can only be
answered by appeal to properties of the lexical entries in question.

Bare Number

. Introduction
The last fifteen years have seen a resurgence of studies on the structure and interpretation of bare nominals, partly in the wake of Chierchias seminal paper Reference
to Kinds across Languages. Within his framework, Chierchia proposes the existence
of a semantic parameter, the Nominal Mapping Parameter (NMP) that governs the
distribution of bare nouns across languages and explains to a significant degree the
correlations between the bare form and its denotations. The NMP is formulated in
terms of the possible combinations of two binary features [pred], [arg]. If the
parameter is set to [+pred, arg], a positive specification of the former feature means
that lexical nouns in the language have predicative (property, type e, t) denotations,
and as a result cannot be used bare in argument position. To appear as arguments,
nouns in these languages must be accompanied by argumentizing functional structure, typically a D (type: e, t, e or e, t, e, t, t as the case may be). An example
of such a language is French, where bare arguments are altogether disallowed:


ont maltrait Ganelon.

have mistreated Ganelon
Cooks have mistreated Ganelon.

Conversely, languages of the [pred, +arg] type do not require argumentizing functional structure: their bare NPs can serve as arguments referring typically to kinds
(type e). Typical examples of such languages are Chinese, Japanese, and Korean.
We would like to thank the audiences of LAGB , workshop on Bare nouns: Syntactic Projections
and their Interpretation, Universit Paris Diderot, Laboratoire de Linguistique Formelle, UMR CNRS
(), OnLI II (), Workshop on Current Issues on Semantics and Pragmatics at York (), Mediterranean Syntax Meeting (), Syntax Group at Cambridge (), and Workshop Celebrating thirty years
from On the Syntax and Semantics of Resumptive Pronouns by Edit Doron (). We are also grateful
for discussion to Giuliana Giusti, Gennaro Chierchia, Napoleon Katsos, and Giuseppe Longobardi.

Bare Number

A language may also choose a positive specification for both attributes [+pred,
+arg]. This is the case in English, where bare arguments can occur only if they are
plural or mass but otherwise will require a determiner in order to appear in argument
All the different types of language given above have been extensively studied. In
this chapter we will be focusing on languages of the [+pred, arg] type, a category in which the Romance languages have figured most prominently. A significant
observation regarding at least some languages in this category is that despite what
the general setting [+pred, arg] entails, we do observe bare arguments in certain
structural positions. The positions in question are on the one hand those which, using
technical vocabulary corresponding to GB assumptions, are lexically governed (i.e.
complement to a verb or preposition) and, on the other, positions associated with
specific information-structural notions. The literature on e.g. Italian or Spanish has
converged on the idea that these languages have in their inventory of determiners
or D-like elements a phonologically null element which is used in these cases. The
strict syntactic conditions correspond to the requirement for the null element to
be identified. Whether or not a language has a null D seems to escape principled
explanation at least insofar as the NMP is concerned. Italian and Spanish have it,
French does not. Despite the focus on the status and availability of a null D, one
theoretical possibility that has remained under-explored is that languages with (some)
bare arguments which lack a null D may use other functional structure to achieve the
same effect as a null D. Here is a little thought experiment. Imagine a language, call it
Freek, which uses bare arguments fairly liberally, to wit more than Italian, i.e. not only
in the designated lexically governed positions. However, like French, it can be shown
not to have a null D. Imagine also that Freek allows both bare singular and plural
arguments. A question that immediately arises in the context of the classification given
by the NMP is: which class does Freek belong to? Clearly, it is not a [pred, +arg]
language. Nouns in Freek are not all mass and it does not have a developed classifier
system. Freek must be either [+pred, +arg] or [+pred, arg]. Taking a narrow view
of the NMP we must reject the possibility of filing Freek together with English as a
[+pred, +arg] language, since it allows bare singular arguments as well as bare plurals.
Regarding the second possibility, i.e. that Freek is a [+pred, arg] language, it would
appear that taking a very narrow view of the NMP, we would also have to rule this
possibility out since, by assumption, Freek lacks a null D. What to do? It seems that
there are three possibilities:
A. Freek does not, and crucially cannot, actually exist. It is an impossible language.
B. Sub-parametrize the [+pred, +arg] setting so as to allow:
A language that would choose [pred, arg] as its setting would not be able to use nominal arguments
at all. This possibility has generally been excluded from study.

Alexopoulou, Folli, and Tsoulas

(a) the N Pred mapping to yield non-count nouns.
(b) The operator must also be defined for singulars.

C. Recognize as a possible exponent of the [+pred, arg] setting, a case where the
projected argumentizing structure is different than D.
Granting, for the purposes of the thought experiment, that (A) should be set aside,
it seems to us that option (C) is the most attractive. There is indeed nothing in the
system that disallows lexical variation of this type. It is, to all intents and purposes, on
a par with saying that a language may or may not have a null D. In the same way, in
fact, nothing in principle prevents a language from having (alongside D) a category,
, which partakes in the functions of D while also performing other syntactic or
semantic functions (after all, Ds have other meanings apart from type shifting a
property to an argument). Option (B) can be pursued to get the facts right and we
could probably claim descriptive success in the process, but the deeper reasons for the
variation will elude us unless a specific and systematic property of Freek that accounts
for this behaviour is found.
But does Freek actually exist? In the rest of this paper we will show that Greek is in
fact very close, if not identical, to Freek. We will do this by zooming in on a cluster
of properties that find a natural explanation if we assume that Greek, like Freek, is a
[+pred, arg] language, has no null D, and the argumentizing function of D can also
be fulfilled by Num , the head of NumP.
The rest of the chapter is organized as follows. In section . we make the empirical
case that Greek does indeed have the properties mentioned above. We present a
number of striking empirical contrasts between Greek and Italian, both [+pred, arg]
languages, that point to variation in their nominal syntax with regard to the argumentizing head. We also show how a number of different but related properties
cluster around the status of the argumentizing head. Having established that Greek
is both similar to Italian in that it is in the [+pred, arg] category, but different in the
availability of a null D, we then show that Number in Greek is a semantically active
category in all cases. We show this by comparing Greek morphologically numbermarked nouns to those of other languages, and show that the Greek cases never show
number neutrality properties. We then flesh out an analysis of number in Greek that
captures its unexpected semantic properties and provides semantic content to its role
as an argumentizing head. We present our analysis in section .. In the closing
sections we offer some speculative remarks on the nature of the Nominal Mapping
Parameter and suggest that it may be reformulated in terms of the specifications of
the Number head. Section . concludes.
The vast literature on nominal syntax is comprehensively reviewed and discussed by Alexiadou,
Haegeman, and Stavrou (). The nominalizing/argumentizing role of D in Italian has been debated
by Longobardi (), Chierchia (), and Giusti (). We here restrict discussion to bare nouns in
Greek and do not discuss the analysis of the Greek definite article, which has been a matter of debate in
the Greek literature (see e.g. Horrocks and Stavrou , Androutsopoulou , Panagiotidis , and
Kolliakou ).

Bare Number

. The Empirical Picture

Our first task is to show that Greek, like our imaginary Freek, indeed lacks a null D. We
take it as given that Italian does have a null D (Longobardi et seq.). Our strategy
consists in comparing Greek and Italian with respect to a range of phenomena that
can be reasonably taken to indicate the presence of a null D.
In a study on the internal structure of nominals in connection with topic strategies
in Greek and Italian, Alexopoulou and Folli () document a number of differences
between the two languages which are unexpected if Greek and Italian were simply
to be seen as [+pred, arg] languages. The findings of that paper are particularly
relevant to our purposes here, and we summarize them below. We begin by considering left dislocation structures where the topic is an indefinite. Crucially, while the
two languages standardly use Clitic Left Dislocation (CLLD) as the main syntactic
strategy for encoding topics (Philippaki-Warburton , Cinque , Rizzi ,
Anagnostopoulou , Tsimpli , Alexopoulou and Kolliakou ), they differ
when the topic is indefinite:
() a. Una segretaria (la)-trovi facilmente / prima o poi una segretaria
sooner or later a secretary
a secretary (her.cl-find easily
A secretary, you will find her easily/sooner or later you will find her.
b. Gramatea tha ( ti)
secretary will ( her.cl find.pl certainly
A secretary, you will find her certainly.
The examples in () show that in Italian, CLLD can be used with indefinite topics
and that, just as for definite topics, the dislocated element is a DP and it is resumed
by the pronominal clitic la. In contrast, in Greek we see a case of Topicalization
where the topic is bare and the comment cannot contain the clitic ti. The scopal
properties of indefinite topics also differ between the two languages: Greek CLLDed indefinites resist opaque or de dicto readings (Iatridou , Alexopoulou and
Kolliakou ):
() a.

Mia kokini fusta tin psahno

edho ke meres
a red skirt it look for.sg here and days
Ive been looking for a red skirt for a few days

b. =ke de boro
na vro
kamia pu na maresi.
and not can.sg subj find.sg none that subj me please.sg
and I cannot find any that I like.

ke de boro
na thimitho
and not can.sg subj remember.sg where her.cl have.sg put
and cannot remember where I put it. (Alexopoulou and Kolliakou )

Alexopoulou, Folli, and Tsoulas

In Italian, on the other hand, the corresponding example in (a) is ambiguous and
can therefore have both (b) and (c) as a possible continuation:
() a. Una gonna rossa la
da un po
a red skirt her.cl look for.sg for a while
A red skirt Ive been looking for for a while
b. ma non ne
trovata nessuna che mi
but not of them.cl have.sg found none.f that me.cl please.sg.subj
but have not found any one that I like.
c. ma non riesco
a ricordarmi dove lho
but not reach.sg to remember where her.cl have.sg put
but I cannot remember where Ive put it.
Thus, we can draw a preliminary conclusion here and say that while CLLD is employed
in both Greek and Italian to encode discourse topics, Italian encodes topics exclusively
by means of CLLD, irrespective of referentiality/specificity, whereas Greek employs
Topicalization for non-referential topics as in ():
() a. Fetos
i moda ine apesia; idika
i bluzes ine aparadektes.
this year the fashion is awful; especially the blouses are outrageous
I hate this years fashion; the blouses are especially outrageous.
b. Mia kokini bluza psahno
edo ki ena mina ke de boro na
a red blouse look for.sg here and one month and not can subj
puthena kamia pu na maresi.
find.sg anywhere any one that subj me like.sg
A red blouse Ive been looking for for a month now and I cannot find one that
I like.
These examples highlight two crucial differences between the two languages in these
structures: Greek, unlike Italian, uses bare nouns where a nominal has a weak indefinite reading; Greek has a gap where Italian has an obligatory clitic pronoun. We
interpret this type of data as showing that if Greek does not employ the (otherwise
available) clitic which is categorially a D, this is because the moved constituent is not
a DP headed by a null D.
Furthermore, these differences can be taken as a sta