ABSTRACT
Title of dissertation: VARIATION AND LEARNABILITY
IN CONSTRAINTS ON A-BAR MOVEMENT
Nick Huang
Doctor of Philosophy, 2019
Dissertation directed by: Professor Colin Phillips and Professor Howard Lasnik
Department of Linguistics
A classic problem in linguistics is explaining how learners come to know so
much about their native languages, despite receiving limited and noisy input. This
learning problem becomes especially acute when the linguistic properties in question
are obscure and show subtle variation across languages. Cross-linguistic variation
means that learners must identify the appropriate points of variation for their lan-
guage, even though the direct evidence that they need is often hard to detect or
even non-existent.
This dissertation presents two case studies on constraints in A-bar movement.
Because constraints are by nature abstract and dicult to observe directly, a classic
solution to the learning problem posed by constraints claims that knowledge of
these abstract or negative linguistic properties is innate. However, a number of
these constraints show cross-linguistic variation, raising questions about how they
are represented and how linguistic experience might (or might not) shape linguistic
knowledge.
The rst case study, discussed in chapters 2 and 3, involves cross-linguistic vari-
ation in the constraint that governs A-bar movement from relative clauses: some,
but not all, languages allow A-bar movement from relative clauses under exceptional
circumstances. I argue that these porous relative clauses that permit A-bar move-
ment can be distinguished by a property that I call tense dependence, and discuss
how this tense property might be formally related to A-bar movement. I show that
this particular kind of variation presents a learning problem: in languages like En-
glish and Mandarin Chinese, learners have little direct positive evidence that such
A-bar movement is possible. Using tense dependence, I propose that learners might
circumvent this absence of direct evidence by relying on A-bar movement from a
supercially unrelated structure: non-nite purposive clauses.
The second case study, discussed in chapter 4, involves bridge verbs: within a
given language, some verbs allow A-bar movement and others do not; in addition,
the set of verbs that allow A-bar movement varies across languages. I present an ac-
ceptability judgment experiment that is aimed at clarifying existing generalizations
about bridge verbs in English. With more secure generalizations in hand, I discuss
the extent to which bridge eects have a pragmatic origin, bringing in data from
an informal survey of English and Dutch native speakers that looks at the eect
of context on long-distance A-bar movement. Echoing existing work, the survey
shows what appears to be a case of cross-linguistic variation between English and
some Dutch varieties for cognitive factive verbs. To account for this instance of
cross-linguistic variation, I suggest that English learners might have limited access
to direct evidence, and discuss what learning mechanisms a learner might need to
draw the language-appropriate conclusions based on sparse evidence.
Chapter 5 discusses the consequences these case studies have for our formal
accounts of these constraints. I evaluate existing proposals and argue that the range
of variation observed requires more exibility than what many existing proposals
can oer.
Chapter 6 concludes.
VARIATION AND LEARNABILITY
IN CONSTRAINTS ON A-BAR MOVEMENT
by
ZHIPENG (NICK) HUANG
Dissertation submitted to the Faculty of the Graduate School of the
University of Maryland, College Park in partial fulllment
of the requirements for the degree of
Doctor of Philosophy
2019
Advisory Committee:
Professor Colin Phillips, Co-chair/Co-advisor
Professor Howard Lasnik, Co-chair/Co-advisor
Professor Jerey Lidz
Professor Valentine Hacquard
Professor Robert De Keyser (Dean's Representative)
?c Copyright by
Nick Huang
2019

Dedication
To my parents and grandparents.
ii
Acknowledgments
Although I am credited as the author of this thesis, it has been shaped by the
eorts of many others. I would like to take this opportunity to thank the following:
The dissertation committee. I am indebted to my co-advisors, Colin Phillips
and Howard Lasnik, who have been incredibly generous with their time and energy
and have worked with me to develop my ideas over the past ve years. I couldn't
have asked for better academic role models and mentors. I would also like to thank
Valentine Hacquard and Jerey Lidz for their input on this thesis and other projects;
collaborating with them has been a great learning experience for me. And my
gratitude to Robert De Keyser for coming on board as the Dean's Representative.
My cohort and fellow students. For conversations about linguistics, as well as
friendship and cake-making experiences, I thank Gesoel Mendes, Laurel Perkins,
and Kasia Hitczenko. For taking the time to discuss semantics, morphosyntax,
psycholinguistics, and language acquisition, I thank Annemarie van Dooren, Chia-
Hsuan Liao, Suyoung Bae, Paulina Lyskawa, Anouk Dieuleveut, Aaron Doliana,
Yu'an Yang, Maxime Papillon, and Phoebe Gaston, among many others.
The broader Maryland linguistics and language science community. My work
and training have also been informed, directly and indirectly, by the courses I took,
the courses I TA-ed for, workshops, and the many conversations I have had with
other members of this research community. For extensive discussion on psycholin-
guistics and second language acquisition, I am grateful to Nur Basak Karatas and
Eric Pelzl. The Language Science Writing Group, ably led by Allie Johnson, made
iii
sure I got my writing out in a timely manner. I also owe thanks to many members
of the faculty, such as Peggy Antonisse, Tonia Bleam, Naomi Feldman, Norbert
Hornstein, Bill Idsardi, Ellen Lau, Paul Pietroski, Maria Polinsky, Omer Preminger,
Philip Resnik, Juan Uriagereka, Alexander Williams, Kira Gor, and Jan Edwards.
Data and research assistance. I would like to thank Naomi Feldman and Philip
Resnik for helping me get access to the Penn Treebank and Chinese Treebank. For
research assistance on work for this thesis, especially parts of the corpus analyses
reported in Chapters 2, 3, and 4, I thank Scott Kaplowitz and Phoebe Cao.
My undergraduate teachers. I ended up pursuing graduate work in linguistics
in part because I had many inspiring professors at Yale. In particular, I am deeply
grateful to Raaella Zanuttini and Bob Frank for their encouragement and mentor-
ship.
Grant sources and administrative support. I have been lucky to receive sup-
port from the department as well as support from the National Science Foundation
under Grant No. 1449815. I am also grateful for the help received from Kim Kwok,
the department's tireless administrator, and the Maryland Language Science Center.
My running friends. Mens sana in corpore sano, as they say. Special thanks
to all at College Park parkrun, especially co-founder Andrea Zukowski, the Prince
George's Running Club, and running kakis JJ Su and Alvin Pee.
Family. My family has been innitely understanding and supportive of my
academic pursuits, cheering me on from the other side of the world. I am also very
grateful to the Hsieh, Tomenga, and Lessels families for welcoming me.
And to Jessica, for everything.
iv
Table of Contents
Dedication ii
Acknowledgements iii
List of Figures x
1 Introduction 1
1.1 Constraints in linguistic theory . . . . . . . . . . . . . . . . . . . . . 1
1.2 Exceptions and cross-linguistic variation . . . . . . . . . . . . . . . . 7
1.3 Case studies: constraints on A-bar movement from two types of em-
bedded clauses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3.1 Goals and approach . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3.2 Why these clauses? . . . . . . . . . . . . . . . . . . . . . . . . 14
1.4 Contributions of this thesis . . . . . . . . . . . . . . . . . . . . . . . . 18
1.5 Roadmap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2 Variation in A-bar movement from relative clauses 23
2.1 To what extent do relative clauses block A-bar movement? . . . . . . 23
2.1.1 Porous relative clauses cross-linguistically . . . . . . . . . . . . 24
2.1.2 How porous are porous relative clauses? . . . . . . . . . . . . 28
2.1.3 Characteristics that distinguish porous relative clauses . . . . 33
2.1.3.1 Are porous relative clauses always subject relative
clauses? . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.1.3.2 What environments do porous relative clauses ap-
pear in? . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.1.3.3 The semantics of the complex NP? . . . . . . . . . . 38
2.1.3.4 Tense and temporal interpretation in porous relative
clauses . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.1.3.5 Formal approaches to tense phenomena . . . . . . . . 45
2.1.4 Possible analyses of porous relative clauses . . . . . . . . . . . 49
2.1.4.1 A movement-based analysis . . . . . . . . . . . . . . 50
2.1.4.2 A defective tense, phase-based analysis . . . . . . . . 52
v
2.1.4.3 Further complications . . . . . . . . . . . . . . . . . 55
2.1.5 Interim summary . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.2 A learning problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.2.1 Learning an English-like grammar with direct positive evidence 64
2.2.2 Corpus study 1: How frequent is A-bar movement from rela-
tive clauses in English and Mandarin Chinese? . . . . . . . . . 65
2.2.2.1 Data and methods . . . . . . . . . . . . . . . . . . . 66
2.2.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . 67
2.2.2.3 Discussion: Learning from direct positive evidence
unlikely for English and Mandarin Chinese . . . . . . 68
2.2.3 Retreating from an English-like grammar with direct nega-
tive evidence . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.2.4 Retreating from an English-like grammar with indirect pos-
itive evidence . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2.3 Learning an English-like grammar using indirect positive evidence . . 73
2.3.1 A-bar movement from purposive clauses . . . . . . . . . . . . 75
2.3.1.1 A-bar movement from purposive clauses in Hebrew,
and Mandarin Chinese, and Brazilian Portuguese . . 76
2.3.1.2 Restrictions on A-bar movement from purposive clauses 78
2.3.1.3 Languages that block A-bar movement from relative
clauses block A-bar movement from purposive clauses 80
2.3.2 Alternative hypotheses about indirect evidence . . . . . . . . . 84
2.3.2.1 Long-distance A-bar movement from (nite) comple-
ment clauses . . . . . . . . . . . . . . . . . . . . . . 84
2.3.2.2 How relative clauses are marked . . . . . . . . . . . . 86
2.3.2.3 Raising vs. matching relative clauses . . . . . . . . . 86
2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3 Learning with purposive clauses 93
3.1 How purposive clauses, porous relative clauses might form a natural
class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.1.1 Two types of purposive clauses . . . . . . . . . . . . . . . . . . 95
3.1.2 Tense dependence . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.1.3 So (that), because, and only to clauses . . . . . . . . . . . . . 99
3.1.4 Other issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.1.4.1 Temporal adjunct clauses . . . . . . . . . . . . . . . 102
3.1.4.2 Mandarin Chinese wei(le) adjuncts . . . . . . . . . . 106
3.1.5 Extending the analysis for porous relative clauses to purposive
clauses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.1.5.1 Adapting a movement-based analysis . . . . . . . . . 111
3.1.5.2 A defective tense analysis . . . . . . . . . . . . . . . 115
3.2 Alternative characterizations of purposive clauses and porous relative
clauses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
3.2.1 Backgroundedness . . . . . . . . . . . . . . . . . . . . . . . . 117
3.2.2 Veridicality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
vi
3.2.3 Mood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
3.3 Corpus study 2: A-bar movement from purposive clauses in the input
in English . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
3.3.1 Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
3.3.2 How much input does a learner get? . . . . . . . . . . . . . . . 126
3.3.3 CHILDES Treebank study: Methods . . . . . . . . . . . . . . 128
3.3.4 CHILDES Treebank: Results . . . . . . . . . . . . . . . . . . 129
3.3.5 Penn Treebank study: Methods . . . . . . . . . . . . . . . . . 130
3.3.6 Penn Treebank: Results . . . . . . . . . . . . . . . . . . . . . 131
3.3.7 General discussion . . . . . . . . . . . . . . . . . . . . . . . . 133
3.4 Corpus study 3: A-bar movement from purposive clauses in the input
in Mandarin Chinese . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
3.4.1 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
3.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
3.4.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
3.5 Corpus study 4: How much positive evidence is necessary? Evidence
from four other constructions in English . . . . . . . . . . . . . . . . 138
3.5.1 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
3.5.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
3.5.3 Discussion: Further inferences a learner has to make . . . . . . 142
3.5.3.1 Swiping . . . . . . . . . . . . . . . . . . . . . . . . . 144
3.5.3.2 Verbless no matter construction . . . . . . . . . . . . 146
3.6 What kind of learner do we need? . . . . . . . . . . . . . . . . . . . . 147
3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
4 Variation in A-bar movement from complement clauses 154
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
4.2 An overview of the literature on bridge verbs . . . . . . . . . . . . . . 157
4.2.1 Restrictions on what moves . . . . . . . . . . . . . . . . . . . 157
4.2.2 Construction-specic eects . . . . . . . . . . . . . . . . . . . 160
4.2.3 Cross-linguistic variation . . . . . . . . . . . . . . . . . . . . . 161
4.2.4 Concerns about empirical generalizations . . . . . . . . . . . . 165
4.3 Hypotheses about bridge eects . . . . . . . . . . . . . . . . . . . . . 167
4.3.1 Factivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
4.3.2 An indexical shift hierarchy . . . . . . . . . . . . . . . . . . 170
4.3.3 A comment on manner of speaking verbs . . . . . . . . . . . . 174
4.3.4 Bridge eects are sensitive to ne-grained semantic distinctions177
4.3.5 Bridge eects reect backgroundedness eects . . . . . . . . . 179
4.3.6 Exemplar eects . . . . . . . . . . . . . . . . . . . . . . . . . 181
4.3.7 A frequency-based reductionist account . . . . . . . . . . . . . 182
4.4 Acceptability judgments for 100 attitude verbs . . . . . . . . . . . . . 184
4.4.1 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
4.4.1.1 Verb selection . . . . . . . . . . . . . . . . . . . . . . 185
4.4.1.2 Obtaining better frequency estimates . . . . . . . . . 187
4.4.1.3 Identifying the other independent variables . . . . . 189
vii
4.4.1.4 Collecting acceptability ratings . . . . . . . . . . . . 195
4.4.2 Data analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
4.4.3 Analysis 1: Absolute acceptability of long-distance A-bar move-
ment for the fuller set of verbs . . . . . . . . . . . . . . . . . . 198
4.4.4 Analysis 2: Relative acceptability of long-distance A-bar move-
ment for the full dataset . . . . . . . . . . . . . . . . . . . . . 203
4.4.5 General discussion . . . . . . . . . . . . . . . . . . . . . . . . 205
4.5 Just which dimensions matter for long-distance A-bar movement in
English? A post-hoc analysis . . . . . . . . . . . . . . . . . . . . . . . 206
4.5.1 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
4.5.2 Verb classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
4.5.3 A statistical analysis . . . . . . . . . . . . . . . . . . . . . . . 210
4.5.4 Interim summary . . . . . . . . . . . . . . . . . . . . . . . . . 213
4.6 The source of bridge eects . . . . . . . . . . . . . . . . . . . . . . . 215
4.6.1 Three hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . 215
4.6.2 Evaluating these hypotheses from a conceptual perspective . . 218
4.6.3 Evaluating these hypotheses from an empirical perspective . . 219
4.7 A survey of English and Dutch speakers . . . . . . . . . . . . . . . . 223
4.7.1 Survey materials . . . . . . . . . . . . . . . . . . . . . . . . . 223
4.7.2 Collecting and evaluating native speaker judgments . . . . . . 227
4.7.3 Summary of ratings . . . . . . . . . . . . . . . . . . . . . . . . 228
4.7.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
4.8 Directly learning about the acceptability of long-distance A-bar move-
ment with cognitive factive verbs / Corpus study 5 . . . . . . . . . . 234
4.8.1 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
4.8.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
4.8.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
4.8.4 Generalizing despite sparse input . . . . . . . . . . . . . . . . 241
4.8.4.1 The Suciency Principle . . . . . . . . . . . . . . . . 241
4.8.4.2 An alternative Bayesian-inspired approach . . . . . . 244
4.8.4.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . 246
4.8.4.4 Results and discussion . . . . . . . . . . . . . . . . . 247
4.8.5 Prospects of applying these learning accounts to other in-
stances of bridge eects . . . . . . . . . . . . . . . . . . . . . 249
4.8.6 Additional comments on the results of the corpus analysis . . 251
4.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
5 Cross-linguistic variation and consequences for formal theories of A-bar move-
ment constraints 254
5.1 A critique of unied analyses of A-bar movement from relative
clauses and complement clauses . . . . . . . . . . . . . . . . . . . . . 255
5.1.1 An event semantic approach: Truswell 2007, 2011 . . . . . . . 255
5.1.2 A discourse-based approach: Goldberg 2006 and Ambridge
and Goldberg 2008 . . . . . . . . . . . . . . . . . . . . . . . . 259
viii
5.1.3 A distributional learning approach: Pearl and Sprouse 2013a,
2013b . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
5.2 A-bar movement from relative clauses . . . . . . . . . . . . . . . . . . 266
5.2.1 Implications for derivational approaches to adjunct islandhood266
5.2.2 Remarks on syntactic parameters and restricted hypothesis
spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
5.2.3 A critique of two recent approaches on A-bar movement from
relative clauses . . . . . . . . . . . . . . . . . . . . . . . . . . 270
5.2.3.1 A syntactic approach: Sichel 2018 . . . . . . . . . . . 270
5.2.3.2 A processing approach: Kush et al., 2013 . . . . . . . 273
5.3 Comments on the acceptability of A-bar movement from porous rel-
ative clauses and purposive clauses . . . . . . . . . . . . . . . . . . . 277
5.3.1 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
5.3.1.1 Participants . . . . . . . . . . . . . . . . . . . . . . . 278
5.3.1.2 Materials . . . . . . . . . . . . . . . . . . . . . . . . 278
5.3.1.3 Instructions . . . . . . . . . . . . . . . . . . . . . . . 279
5.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
5.3.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
5.4 Bridge eects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
5.5 Evaluation of existing proposals about bridge eects . . . . . . . . . . 289
5.5.1 Syntactic approaches . . . . . . . . . . . . . . . . . . . . . . . 290
5.5.2 Non-syntactic approaches . . . . . . . . . . . . . . . . . . . . 293
5.5.3 Evaluation of these approaches . . . . . . . . . . . . . . . . . 293
5.5.4 Modelling variation within and across languages . . . . . . . . 298
6 Future directions and conclusion 302
6.1 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
6.2 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
References 309
ix
List of Figures
4.1 Frequencies of verbs with complement clauses in the Penn Treebank
and the British National Corpus . . . . . . . . . . . . . . . . . . . . . 189
4.2 Normalized acceptability of matrix A-bar movement vs. long-distance
A-bar movement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
5.1 A-bar movement from adjuncts . . . . . . . . . . . . . . . . . . . . . 281
5.2 Superadditivity in A-bar movement from porous relative clauses . . . 281
5.3 Distribution of z-scores for A-bar movement from porous relative clauses282
x
Chapter 1: Introduction
1.1 Constraints in linguistic theory
A major achievement of modern linguistics is demonstrating the depth of
knowledge individuals have about their native languages in domains such as phonol-
ogy, morphology, syntax, semantics, and pragmatics. This knowledge, often modeled
in the form of rules and constraints on linguistic representations, enables them to
comprehend and produce novel utterances, and to avoid producing certain kinds of
sentences or phrases.
But the discovery of rules and constraints also raises questions about how
learners come to acquire them. These rules and constraints are highly abstract
and specic. They are also often not explicitly taught to learners. To illustrate
the problem, consider the following fact, to be discussed in greater detail in this
dissertation: English appears to have a constraint that has the eect of blocking
A-bar movement (in this case, movement of a wh-phrase) from relative clauses. By
positing this constraint, we can straightforwardly explain why native speakers of
English nd a sentence like (1) unacceptable.
(1) *Who did you read a book [RC that was about]? (cf. Who did you read a book
about? )
1
But there seems to be no easy way for a child to learn explicitly that such
a sentence is unacceptable in English, and more importantly, no way for the child
to learn that there is a constraint in English that rules out this sentence and other
structurally similar ones. From a learner's perspective, there might not be easily-
observed direct (positive) cues that speakers cannot produce such sentences, or that
speakers have trouble parsing them. Further, other speakers (parents, older siblings,
other caregivers, and teachers) do not tell children to avoid these kinds of sentences;
direct negative evidence is thus also absent. Supposing that these older speakers
had the desire to do so, the abstractness of the constraint  involving notions
such as A-bar/wh-movement ' and relative clauses  makes it dicult for them
to describe the constraint or the consequences of violating it. And even if these
older speakers had the linguistic training to provide a description, it is dicult to
imagine that learners always have the ability to understand what is being conveyed
or observe the prohibition faithfully.
Because of their negative nature and widespread distribution cross-linguistically,
it is unsurprising that these constraints have played a central role in linguistic the-
orizing. A classic solution to the learning problem they pose is to posit innate
domain-specic knowledge: learners are born knowing these constraints (or more
abstract principles that derive these constraints), so the absence of direct evidence
is not a concern. In addition to eliminating the learning problem, this solution is
also attractive as it explains the cross-linguistic distribution of these constraints.
For example, the earliest discussion of examples like (1), as far as I can tell, is
in Chomsky, 1964, where the low acceptability was accounted for using what is now
2
called the A-over-A principle (2). Chomsky suggests that this principle might . . .
be proposed as a hypothetical linguistic universal (p. 931). Expanding on this issue
in his ground-breaking dissertation, Ross (1967, chapter 4) suggests these examples
are better handled with the Complex NP Island Constraint, which essentially rules
out A-bar movement from a clause embedded within a noun phrase. Like Chomsky,
Ross suggests that his constraint is universal.1
(2) A-over-A principle: If the phrase X of category A is embedded within a
larger phrase ZXW which is also of category A, then no rule applying to the
category A applies to X (but only to ZXW). (Chomsky, 1964, p. 931)
This kind of innate knowledge proposal constitutes a strong claim about the
human language faculty. This point does not necessarily only refer to the proposal
that learners are born with the knowledge that certain dependencies are subject to
specic constraints, such as the Complex NP Constraint. Depending on how the
constraint in question is stated, it can also entail that learners are born with knowl-
edge of certain linguistic categories or properties. For example, Ross' Complex NP
Island Constraint is stated in terms of clauses and NPs, and therefore presupposes
knowledge of these syntactic categories. Further, as mentioned, they predict that
languages should behave alike with respect to certain phenomena, since the speakers
of these languages are all born with the same knowledge.
Subsequent research following up on Ross's work has proposed a variety of
abstract principles, such as subjacency (in terms of bounding nodes or barriers)
(Chomsky, 1973, 1977, 1986; Lasnik and Saito, 1992; Rizzi, 1982), the Empty Cat-
1Strictly speaking, this characterization of Ross's proposal is inaccurate: these constraints
do not actually prevent the movement of a constituent per se; they block the application of a
chopping rule that blocks the moved element from being pronounced in its original position.
3
egory Principle (Chomsky, 1981), relativized minimality (Rizzi, 1990), the Phase
Impenetrability Condition (Chomsky, 2000, 2001; Grano and Lasnik, 2018; Huang,
2018a), the Condition on Extraction Domain (Huang, 1982), restrictions on syntac-
tic derivations (Hunter, 2015; Stepanov, 2007; Uriagereka, 1999), discourse-based
concepts like dominance (Erteschik-Shir, 1973) and backgroundedness (Goldberg,
2006), as well as event structure principles (Truswell, 2007, 2011). However, like
Ross's and Chomsky's proposals, these proposals come with the (implicit) assump-
tion that the principles are part of one's innate knowledge of language; the principles
are too abstract to be taught or deduced from one's linguistic experience.
Another logically coherent response is to deny the above learning problem.
There are two ways mutually exclusive ways of doing so. The rst is to claim
that these constraints are actually not grammatical in nature. Instead, the low
acceptability that has led linguists to propose constraints is merely an artifact of
sentence processing (e.g. Deane, 1991; Hofmeister, 2007; Hofmeister and Sag, 2010;
Kluender and Kutas, 1993). For example, many constraints on A-bar movement
are eectively bans on A-bar movement from syntactically complex structures, such
as various kinds of embedded clauses. One could argue that the low acceptability
associated with A-bar movement from these structures simply reects the processing
challenges that arise from simultaneously processing a wh-dependency and a struc-
turally complex sentence. Or it might reect a failure to imagine an appropriate
context in which a native speaker might produce such a sentence. Of course, for such
an argument to be credible, one will need to provide an independently-motivated
theory of sentence processing. While this is no easy task, a successful account would
4
be theoretically welcome, as it would eliminate the need to posit grammatical con-
straints, and therefore eliminate the need to explain how learners come to possess
knowledge of them.
Although this processing-based approach is conceptually starkly dierent from
the classic innate knowledge solution, it is similar in that it also predicts cross-
linguistic uniformity. This prediction stems from the fact that these constraints are
explained as processing artifacts, reecting factors such as memory limitations and
the way we process sentences using our knowledge of grammar, context, etc. It is
hard to see how such eects might vary across languages, barring eects that are
ultimately attributable to dierences in grammar (e.g. wh-questions are obviously
processed dierently in wh-in-situ languages and overt wh-movement languages).
Consequently, we expect to observe the same processing artifacts across languages.
In contrast, the second option takes seriously the notion that these constraints
are grammatical in nature. Instead, it challenges the assumption that learners have
no access to evidence about them. In this view, learners do get access to evidence;
it is just that the evidence they use happens to be supercially unrelated to the
constraints (or whatever linguistic properties that derive the constraints) that are
to be acquired (see, among many others, Chac?n, 2015; Chomsky, 1981; Pearl and
Sprouse, 2013a,b; Pearl and Mis, 2016; Perfors et al., 2010; Reali and Christiansen,
2005; Snyder, 2001; Viau and Lidz, 2011). It is important to note that such learning
accounts do not necessarily mean that learners are blank slates or that the learning
task is straightforward. For one, for these accounts to work, they must assume
that learners have prior knowledge of hypotheses about the kind of grammars their
5
language might have. In fact, these hypotheses must be stated in fairly abstract
terms, since, by denition, the evidence that supports one hypothesis (or one set of
hypotheses) over another is supercially unrelated.2
As an example, consider one particular account: using indirect negative evi-
dence to learn that a certain construction is ungrammatical (Chomsky, 1981; a vari-
ant of this approach that has received much attention is statistical preemption, see
Boyd and Goldberg, 2011; Pinker, 1989; Wexler and Culicover, 1980, etc.). In such
an account, learners begin with at least two distinct competing hypotheses: (i) the
construction is grammatical, vs. (ii) the construction is not. All else being equal,
these hypotheses make divergent predictions about how frequently the construction
should appear: the rst hypothesis predicts non-zero observations, while the second
predicts no observations. Supposing further that they can track how frequently a
given structure occurs in the input, learners then compare these predictions with
actual observations. If the construction fails to appear, learners use its absence to
conclude in favor of the second hypothesis.
It is also important to note that the availability of a strategy like the use of in-
direct negative evidence does not eliminate the need for other prior domain-specic
knowledge in the language acquisition process. In the context of indirect negative
evidence, for every utterance, there are many other grammatically perfect utter-
ances that could have been produced in the same context but were not. Learners
presumably somehow know which hypotheses a particular utterance bears on and
2It is also logically possible that learners might formulate these constraints from scratch, but it
is dicult to see how this scenario might explain convergence across speakers. In such a scenario,
it is quite likely that learners might end up formulating wildly dierent constraints.
6
which ones it does not (a point made by Pinker (1989) and Yang (2017)).
1.2 Exceptions and cross-linguistic variation
It should be clear from the above discussion that these constraints have impor-
tant implications for our theories of the language faculty. It is thus not surprising
that they have been the object of intense study. For illustration, consider the claim
that these constraints are innately specied. To the extent that the goal of linguistics
is to uncover insights about our knowledge of language, adopting this claim about
innate knowledge of A-bar movement constraints obliges syntacticians to provide a
detailed description of these constraints, as they might be instantiated in the mind.
Alternatively, if one wishes to pursue a reductionist claim, one would need to start
with similarly detailed descriptions of these constraints and show that the eects
associated with them can plausibly be derived as processing artifacts.
However, as a survey of the literature since Ross's dissertation shows, devel-
oping a precise and comprehensive statement of these constraints has proven to be
a challenging task. Two related but distinct problems emerge, and I illustrate them
using examples from Ross's work and subsequent research.
The rst problem is within-language exceptions: for almost every constraint
proposed in the literature, exceptions have been pointed out for that same language,
as though falsifying the constraint. Strictly speaking, this is only a technical issue,
in that we can resolve it by rening our generalizations and formal descriptions,
although perhaps at some theoretical cost. Examples of this kind of problem were
7
observed as early as Ross, 1967. The starting point of Ross's dissertation is the
observation that Chomsky's A-over-A principle (Chomsky, 1964) is too strong for a
language like English. Ross then proposes to replace it with a set of more specic
island constraints. We may see this move as complicating a general theory of gram-
mar. Another example occurs later in Ross's dissertation, with his own Complex NP
Island Constraint. Ross observes that the constraint predicts that a sentence like
(3) should be unacceptable, contrary to what speakers report. He does not provide a
solution to this problem, but tentatively suggests that the acceptable sentences are
actually syntactically dierent, in a deep sense (derive from quite dierent sources,
p. 140), from unacceptable sentences.
(3) The money which1 I have [NP hopes [S that the company will squander t1]
amounts to $400,000]]. (Ross, 1967, p. 140, ex. 4.45a, reporting judgments
by him and other speakers)
(cf. *The money which1 I read [NP reports [S that the company will squander
t1] amounts to $400,000]].)
In more recent work, Truswell (2007, 2011) looks at exceptions (1) to the claim
that adjuncts are islands (for specic proposals, see Cattell, 1976; Chomsky, 1986;
Huang, 1982; Hunter, 2015; Stepanov, 2007; Uriagereka, 1999).
(4) [What song]1 did John walk in [adjunct whistling t1]?
Truswell argues that the prior generalization about adjunct islands is a com-
posite of at least two constraints: a constraint that rules out A-bar movement from
adjuncts that adjoin to a relatively high position in a clause, and another constraint
that makes reference to event structures (his Single Event Grouping Condition).
Again, a simple but perhaps too powerful constraint is replaced with a more nuanced
8
analysis that refers to notions that are relatively novel in the literature, in this case,
event structure. To the extent that the event structure condition is semantic and
does not have a syntactic correlate, this proposal also implies that extra-syntactic
factors can interfere with the grammaticality of A-bar movement, which one might
take as a violation of the autonomy of syntax hypothesis, thus pointing to a more
complicated theory of syntax (I return to Truswell's proposal at several points in
this dissertation.)
The second problem is cross-linguistic variation: a constraint might hold in
one language but not another, often in ways that are hard to observe or charac-
terize. This is a much more vexing problem, as cross-linguistic variation seems
inconsistent with the innate knowledge claim: if constraints were so hard to learn
that knowledge of them (or the knowledge of more abstract principles that give rise
to constraints) were best modeled as innate, why would they vary across languages?
This kind of variation is also apparently inconsistent with an approach that treats
these constraints as processing artifacts: it is much harder to see how cross-linguistic
variation can be derived from a processing angle. Instead, it ts naturally with the
self-evident fact that grammars vary.
Returning to the issue of A-bar movement from relative clauses as an example:
recall that early descriptions by Chomsky and Ross claimed that a universal con-
straint blocks the derivation of these sentences. However, as early as Erteschik-Shir,
1973, this claim was found to be problematic, based on examples of A-bar movement
from relative clauses in Danish (39).
9
(5) Is kender jeg mange der kan lide.
ice cream know I many who like
`I know many who like ice cream.' (Danish, Erteschik-Shir, 1973, p. 48, ex.
48)
Erteschik-Shir also observes that equivalent examples in English are also rel-
atively acceptable, an observation that is also made in subsequent research (Chung
and Ladusaw, 2003; den Dikken, 2007; McCawley, 1981, 1998).
(6) This is a paper1 that we really need to nd someone [who understands t1].
(from Chung and McCloskey, 1983)
Although similar exceptions are found in a variety of languages, these excep-
tions are by no means universal; Cinque (2010) reports that a number of languages
lack them.
Logically, there are two mutually exclusive responses, both implicating non-
trivial learning problems. The rst response is to claim that the language faculty
does not impose the constraint in question on all grammars, therefore the constraint
is not expected to apply to all languages. Instead, the choice of the constraint
(assuming that learners are selecting the most appropriate constraint out of a pre-
dened set) can be inuenced by one's linguistic experience. Alternatively, one could
imagine a scenario where learners formulate a language-appropriate constraint from
scratch, using their linguistic experience. While this scenario might be less plausi-
ble, the broader point about drawing abstract generalizations using one's linguistic
experience, still stands.
The second response is to claim that the cross-linguistic dierence is illusory:
the structures involved are actually dierent. An example of this reasoning is found
10
in work by Kush and colleagues (Kush and Lindahl, 2011; Kush et al., 2013) for
Swedish, which, like Danish, also allows A-bar movement from relative clauses. In
these accounts, thanks to an idiosyncratic property of Swedish, the Swedish equiv-
alent of mange der kan lide in (39) can receive either a relative clause and a small
clause analysis. A-bar movement only takes place when such a string receives a small
clause analysis (but see Christensen and Nyvad, 2014; M?ller, 2015a for arguments
against this proposal). The upshot of this analysis is that Swedish (and possibly
other Scandinavian languages) have the same constraint against A-bar movement
from relative clauses as other languages. The challenge here, for Swedish learners,
is that they must somehow determine using their linguistic experience that certain
strings are structurally ambiguous between these two analyses, even though this
distinction might not be salient or occur frequently.
1.3 Case studies: constraints on A-bar movement from two types of
embedded clauses
1.3.1 Goals and approach
An investigation of both kinds of variation can thus improve our understanding
of the syntax of A-bar movement, while simultaneously addressing related issues re-
garding cross-linguistic variation and language acquisition, specically, how learners
come to determine the appropriate points of variation in their language for dicult-
to-observe phenomena, such as A-bar movement constraints. In this dissertation, I
adopt the following approach:
11
First, I study both types of variation, taking as case studies constraints on
A-bar movement from two types of embedded nite clauses: relative clauses and
complement clauses of verbs. (I explain my choice of case studies in the next section.)
The goal here is to identify the dimensions along which we observe variation. Taking
relative clauses as an example, the goal is to understand the circumstances in which
A-bar movement from relative clauses becomes more acceptable, whether within
languages or across languages. I do so by drawing together existing descriptions
and analyses in the literature, supplementing them with my own data collection.
Next, I consider the kind of linguistic experience learners might reasonably
have. The goal here is to determine what kind of evidence learners have about the
constraints in their language; of particular interest is the extent to which learners
get direct evidence for the relevant constraints in their language. For example,
in a language that allows A-bar movement from relative clauses under exceptional
circumstances (e.g. Danish), do learners actually get to observe A-bar movement
from relative clauses in naturalistic environments? If learners do, the learning task
becomes a little clearer for them: they get direct evidence that their grammar is
one that generates A-bar movement from relative clauses, although the evidence
obviously does not necessarily tell them exactly which grammar that is.
Within generative syntax, it is usually assumed that such direct evidence is
non-existent. In the context of these constraints, this assumption is not an unrea-
sonable one, since the constraints are of a negative nature and most utterances that
learners observe are syntactically simple in the sense that they do not involve embed-
ded clauses (see, e.g. Chomsky, 1971, for similar comments about subject-auxiliary
12
inversion in English).
However, in this dissertation, I look at both child-directed and adult-directed
corpora, following in particular Pearl and Sprouse, 2013a,b, in order to get a more
realistic picture of the sparsity (or richness) of linguistic experience at dierent stages
of the language acquisition process (for other research that take a similar empirical,
corpus-based perspective on linguistic experience, see, among many others, Dudley
et al., 2017; Huang et al., 2018; Hyams, 1986; Legate and Yang, 2007, 2002; Pullum
and Scholz, 2002; Valian, 1990, 1991; White et al., 2018; Yang, 2002, 2016).
With characterizations of what adult native speakers know and what kind of
linguistic experience they might have, I identify the gaps between both components.
To what extent is direct evidence missing? If direct evidence is not available or
very sparse, what else might help learners identify the relevant constraints for their
languages? To address this question, I consider the role of indirect evidence 
evidence that is supercially distinct from the relevant phenomenon  by looking
at cross-linguistic correlations and corpus data (Chomsky, 1981; Pearl and Sprouse,
2013a,b; Pearl and Mis, 2016; Perfors et al., 2010; Reali and Christiansen, 2005).
I put quotes around indirect, because strictly speaking, whether a piece of evi-
dence counts as direct depends on what hypotheses learners entertain about their
grammar. For example, suppose that learners can conclude that construction A is
grammatical, by observing tokens of a supercially distinct construction, construc-
tion B. In a descriptive sense, we might label construction B as indirect evidence
for construction A, as I did above. However, if both constructions are licensed by
some abstract syntactic property P, then it is probably more accurate to say that
13
the observation of construction B is direct evidence for property P, which in turn
guarantees that construction A is grammatical.
A discussion about linguistic evidence therefore inevitably implicates ques-
tions about hypothesis spaces and inference mechanisms (see, among many others,
Chomsky, 1957, 1965, 1981; Gibson and Wexler, 1994; Pearl and Sprouse, 2013a,b;
Pearl and Mis, 2016; Perfors et al., 2010; Viau and Lidz, 2011; Yang, 2016). By con-
sidering the kind of evidence that learners might be using, I also seek to spell out
more concretely the kind of hypotheses that learners consider about their grammars
and/or how they use the evidence they observe to identify the appropriate grammar;
for example, which aspect(s) of the evidence they use to draw the right conclusion
about their grammars. In this regard, this study is part of a long line of inquiry
connecting cross-linguistic syntactic variation with learning concerns (among many
others, Chomsky, 1981; Dunbar, 2013; Hyams, 1986; Legate and Yang, 2007; Rizzi,
1982; Snyder, 2001; Valian, 1991; Viau and Lidz, 2011; Yang, 2002).
1.3.2 Why these clauses?
I next address my choice of case studies. As a rst approximation, relative
clauses and complement clauses are held to impose very dierent restrictions on
(overt) A-bar movement, and have been central to the literature on A-bar movement
constraints (see work by Huang (1982) and others on restrictions on covert A-bar
movement). Consequently, there has been much work done on these two types of
clauses and the circumstances under which (overt) A-bar movement from them is
14
possible. Since this dissertation is concerned with cross-linguistic variation, the
richness of this literature is valuable, as it means that there are relatively many
prior descriptions and analyses to draw from.
More specically, within the generative syntax literature, relative clauses are
standardly held to be in incompatible with A-bar movement. There are two well-
regarded independent analyses for why relative clauses are islands. First, they are
clauses embedded inside a nominal projection; A-bar movement from them thus
violates the Complex NP Island Constraint and its formal descendants, such as
subjacency or even the Phase Impenetrability Condition (assuming that clauses
and nominal projections are phases) (e.g. Chomsky, 1973, 2000, 2001; Ross, 1967).
Alternatively, relative clauses are adjuncts, and some principle of the grammar rules
out A-bar movement from adjuncts (Cattell, 1976; Chomsky, 1986; Huang, 1982;
Hunter, 2015; Stepanov, 2007; Uriagereka, 1999).
That said, there are languages for which A-bar movement from relative clauses
is acceptable under very specic circumstances, challenging the above generalization.
In addition to Danish, Swedish, English, as mentioned previously, similar observa-
tions have been made for Hebrew (Sichel, 2018) and Mandarin Chinese (Xu and
Langendoen, 1985; Zhang, 2002).
Complement clauses of verbs, on the other hand, are canonically depicted as
compatible with A-bar movement, starting from Ross, 1967. The claim that A-bar
movement is unbounded (although subject to negative constraints) is almost always
illustrated with A-bar movement from these complement clauses. The notion that A-
bar movement from complement clauses (or complements in general) is always well-
15
formed is made theoretically signicant in a number of inuential proposals, such as
the Condition on Extraction Domain (Huang, 1982, see also an earlier suggestion by
Cattell (1976)), the barriers framework (Chomsky, 1986; Lasnik and Saito, 1992),
and proposals that derive certain island eects from derivational principles (e.g.
Uriagereka, 1999).
That said, it has long been recognized within the literature that A-bar move-
ment is the most acceptable when the clauses are complements of a small set of verbs,
often labeled as bridge verbs (among many others, Ambridge and Goldberg, 2008;
Chomsky, 1977; Erteschik-Shir, 1973). A-bar movement from the complement of
non-bridge verbs, such as manner of speaking verbs, is associated with lower accept-
ability.
(7) a. Who1 did she {think/say} that he saw t1? [Bridge verbs]
b. ??Who1 did she {mumble/realize} that he saw t1? [Non-bridge verbs]
(examples from Ambridge and Goldberg, 2008)
Another reason for studying relative clauses and complement clauses of verbs
is because A-bar movement from these clauses occurs relatively infrequently in natu-
ralistic contexts, which points to quantitative and/or qualitative gaps between what
learners get to observe of their native languages and what they end up knowing
about these languages. I will return to this issue of sparsity of direct evidence in
greater detail in Chapters 2 and 4. For now, I will summarize prior research by
Pearl and Sprouse (2013a,b), who did a corpus analysis of A-bar movement from
English child-directed speech, adult-directed speech, and adult directed texts, us-
ing CHILDES corpora and corpora from the Penn Treebank (MacWhinney, 2000;
16
Marcus et al., 1999).3
Pearl and Sprouse found that the proportion of tokens with A-bar movement
from an embedded nite clause relative to the total number of utterances ranged
between 0.10% to 0.32%; all these tokens involve A-bar movement from complement
clauses of verbs, and not of any other kind of complement clauses (see their Appendix
B); they did not report nding tokens involving A-bar movement from other kinds
of embedded nite clauses. It is in this sense that we may consider the direct
evidence to be sparse, since it seems unlikely that learners can reliably observe A-
bar movement from certain kinds of relative clauses. One might also wonder why,
given the evidence, learners do not conclude that A-bar movement from complement
clauses of all verbs is equally acceptable.
(8) Selected statistics on A-bar movement from Pearl and Sprouse, 2013b4
Child- Adult- Adult-
directed directed directed
speech speech texts
Total utterances 101,838 74,576 24,243
Total wh-dependencies 20,923 8,508 4,230
Total wh-dependencies involving 324 104 24
embedded nite clauses
We can convert these proportions to annualized gures, using estimates for
number of utterances reported by Hart and Risley (2003, 1995). More specically,
they report that low socioeconomic status (what they label as welfare) children
hear about 176 utterances an hour; assuming these children observe about 5,200
3As far as I can tell, it appears that they were specically interested in A-bar movement of
wh-phrases.
4Figures on embedded nite clauses are calculated with data reported in Pearl and Sprouse,
2013b, Appendix B, for any dependency that is analyzed as crossing a CP headed by a null
complementizer and that. This criterion picks out nite clauses but not non-nite ones. In Pearl
and Sprouse's analysis, control verbs take IP complements, not CP complements.
17
hours of speech a year, this works out to about 915,000 tokens a year. Using this
estimate to calculate annualized gures, one concludes that in a single year, English
learners might get to observe up to 9002,900 wh-dependencies that cross a clause
boundary, of which an absolute majority, if not all, involve movement from com-
plement clauses of verbs. (Of course, just because they are exposed to 9002,900
tokens a year does not guarantee that they use all of them for drawing the right
inferences: it is possible that they fail to observe some of them, or misparse others.)
1.4 Contributions of this thesis
One of the deliverables of this thesis is a review of the circumstances in which
A-bar movement from relative clauses and complement clauses of verbs become
more acceptable, by drawing together existing work in the literature; the hope is
that this review can serve as a summary of the state of the art on both phenomena.
As mentioned, I also supplement this review with my own data collection, where
feasible: I report acceptability judgment experiments on A-bar movement from both
types of clauses as well as informal surveys of native speakers. Doing so is important,
especially in the case of complement clauses, where there are dozens, if not hundreds,
of verbs that take complement clauses (Anand et al., to appear; White and Rawlins,
2016, 2018). However, most existing datapoints and hypotheses in the literature are
based on a small number of verbs; the most thorough investigation I am aware of is
Liu et al., 2019, which looks at 48 English verbs.
Better empirical generalizations allow me to identify the linguistic dimen-
18
sion(s) that aect the acceptability of A-bar movement from these two types of
embedded clauses. Doing so is important for two reasons.
First, better characterizations of the dimensions that matter in A-bar move-
ment and the range of cross-linguistic variation are clearly necessary for formal ac-
counts of these constraints, which in turn bear on more general, architectural issues
in syntax. For instance, consider the case of A-bar movement from relative clauses.
As mentioned previously, it is rather common to come across the generalization that
A-bar movement from relative clauses is uniformly unacceptable; this generalization
is taken to be evidence for certain claims about the nature of syntactic derivations.
For instance, Uriagereka (1999) posits that linearization requires that adjuncts (pre-
sumably including relative clauses) be Spelled Out before they are merged, which
derives the fact that adjuncts are incompatible with A-bar movement. Likewise,
Stepanov (2001) proposes that adjuncts are obligatorily merged late (after A-bar
movement), which also derives the incompatibility of A-bar movement from relative
clauses. As elegant as these proposals may be, the fact that A-bar movement from
relative clauses can be acceptable poses a challenge for them, thus weakening the
overall argument for these proposals (see Truswell, 2011 for similar remarks).5
Second, better generalizations of cross-linguistic variation can inform our the-
ories of learning. Ideally, one should do more than just point out that certain
constraints vary cross-linguistically and thus speakers need to somehow learn what
the relevant constraints are in their language. To the extent that the goal of linguis-
5Sichel (2018) proposes to resolve this issue by positing that these exceptional relative clauses
are actually complements, not adjuncts. However, this particular treatment does not accommodate
cross-linguistic variation easily
19
tic theory is to understand the language faculty, and to the extent that our language
faculty contains a component that makes learning possible, we should seek to also
provide plausible learning accounts. In principle, a full account should consist of
answers to the following questions:
(9) a. What is the learner's initial knowledge state?
b. How does the learner's knowledge state change?
(i) What evidence does the learner encounter?
(ii) What inferences does the learner make?
(iii) What mechanisms and knowledge does the learner use to draw
these inferences?
c. What evidence is there for this learning account?
(i) Is the hypothesized linguistic evidence present in the input?
(ii) If it is present, can learners detect it?
(iii) If learners can detect it, can they draw the hypothesized inference?
For scope reasons, this dissertation will only address questions (9a), (9b), and
(9c-i). In particular, I take a quantitative approach toward questions about evidence,
in particular the question of indirect evidence. Conventionally, such accounts are
substantiated using covariation evidence: showing that in a sample of languages,
the phenomenon to be learned and the phenomenon that serves as indirect evidence
are either both well-formed or both ill-formed. This dissertation presents a proof of
concept, using covariation data to point out what kind of evidence learners might
be using, and providing corpus data to show that the evidence is actually present
in the input and hence in principle accessible for learners.
To the extent that the proposals here are successful, they set the stage for
experimental inquiry, i.e. answering questions (9c-ii) and (9c-iii). We can then ask
how these proposals are similar to or dierent from proposals about how learners
20
acquire other aspects of language, e.g. phonology or semantics, or even how these
proposals compare with learning problems in other domains of cognition. These
questions are important, since they bear on debates about domain specicity and
knowledge: to what extent are the learning mechanisms unique to language? How
much domain-specic knowledge do we need to assume for these learning accounts
to work?
1.5 Roadmap
To sum up, constraints on A-bar movement have been taken to provide a
powerful argument for innate knowledge of language. Systematic variation in these
constraints, however, poses potentially serious problems: it is not always easy to see
how learners might acquire the points of variation appropriate for their language, nor
is it always easy for linguists to accommodate variation in their theories of syntax.
In Chapter 2, I present the rst case study, on variation in the constraint
that governs A-bar movement from relative clauses: some languages, like Polish,
Bulgarian, and German, impose a ban on A-bar movement from relative clauses,
while others, like Hebrew, English, Mandarin Chinese, allow it under exceptional
circumstances. I show that learners of the second set of languages do not always
have reliable direct positive evidence that their language allows A-bar movement
from relative clauses. With cross-linguistic evidence, I argue that learners instead
make use of indirect evidence: A-bar movement from purposive clauses. Chapter
3 elaborates on the argument, presenting corpus evidence to show the feasibility of
21
this indirect learning account, and considers the kind of biases a learner might need
for this account to work.
In Chapter 4, I discuss the second case study, on bridge verbs: within a lan-
guage, certain verbs allow A-bar movement from their complement clauses while
others do not. I review existing claims about the nature of bridge verbs, and present
an acceptability judgment experiment that is aimed at clarifying existing general-
izations about bridge verbs in English. I also evaluate how context might inuence
the acceptability of long-distance A-bar movement, through an informal survey of
English and Dutch. I also nd some evidence showing that there are cross-linguistic
dierences between these two languages, at least in the domain of cognitive factive
verbs, in a way that corroborates existing descriptions of Dutch (Broekhuis and
Corver, 2016). I discuss how these cross-linguistic dierences might be derived from
direct evidence, looking at the distribution of long-distance A-bar movement and
cognitive factive verbs in English as a case study.
In Chapter 5, I review existing proposals regarding these phenomena and dis-
cuss consequences cross-linguistic variation has for our formal descriptions and the-
ories of A-bar movement constraints. For example, I argue that the variation we
observe suggests A-bar movement from relative clauses and complement clauses does
not have a unied analysis. I also use the case of A-bar movement from relative
clauses to argue against proposals that seek to derive adjunct islandhood from more
basic syntactic principles, following Truswell (2011).
Chapter 6 concludes.
22
Chapter 2: Variation in A-bar movement from relative clauses
2.1 To what extent do relative clauses block A-bar movement?
Relative clauses, which often bear nite and/or indicative morphology, are
typically thought to be strong islands that block all types of A-bar movement from
within (1) (Chomsky, 1977; Ross, 1967). Perhaps less appreciated is the fact that
there are relative clauses that seem to allow overt A-bar movement of what is typi-
cally an object, as illustrated in (2) with relativization.
For ease of reference, I will use porous as a label for these relative clauses
and their environments (2); these include a relative clause that is part of a complex
NP that is either c-commanded by an intensional predicate (2a) or the predicate of
a copular sentence (2b).
(1) Regular relative clauses
a. *This is the problem1 that Mary hired the programmer [who solved t1].
[Relativization]
b. *Which problem1 that Mary hired the programmer [who solved t1]? [Wh-
question]
c. *This problem1, Mary hired the programmer [who solved t1]. [Topical-
ization]
d. *Mary solved more problems1 than Smith hired the programmer [who
could solve t1]. [Comparative deletion]
(2) Porous relative clauses (adapted from examples in Chung and McCloskey,
23
1983)
a. This is the problem1 that Mary wants to hire a programmer [who can
solve t1].
b. This is the problem1 that Mary is the only programmer [who can solve
t1].
c. This is the problem1 that there is only one programmer [who can solve
t1].
d. This is the problem1 that I know a programmer [who can solve t1].
2.1.1 Porous relative clauses cross-linguistically
Porous relative clauses are not unique to English. Some of the earliest discus-
sion on A-bar movement from porous relative clauses in the literature is centered on
examples in Scandinavian languages like Swedish (3), Danish, and Norwegian (e.g.
Allwood, 1982; Engdahl, 1980; Erteschik-Shir, 1973; Taraldsen, 1982; see Engdahl,
1997; Lindahl, 2017 for reviews). Other languages where they have been observed
include Hebrew (4) (Rubovitz-Mann, 2000; Sichel, 2018), Mandarin Chinese (Xu and
Langendoen, 1985; Zhang, 2002), and Spanish (Cinque, 2010). Brazilian Portuguese
appears to be another such language (Gesoel Mendes and Kayron Bevil?qua, p.c.).
These examples resemble their English counterparts in (2), occurring in similar syn-
tactic and semantic contexts.
(3) Swedish1 (Engdahl, 1997)
a. H?r ?r [en fr?ga]1 som jag inte k?nner n?gon [som kan svara p?
here is a question that I not know anybody that can answer
t1].
`Here is a question that I don't know anybody who can answer.' (ibid.
1Abbreviations used in glosses: ACC: accusative; CL: classier; COMP: complementizer; DAT:
dative; DEF: denite marker; EXP: experiential aspect; IND: indicative mood; INF: innitive;
MOD: modier; NOM: nominative case; PRF: perfective aspect; REL.PRN: relative pronoun;
SBJ: subjunctive mood; 1/2/3: 1st/2nd/3rd person; sg/pl: singular/plural
24
ex. 15a)
b. Garaged?rren1 ?r det bara Kalle [som kan ?ppna t1].
garage door.DEF is it only Kalle who can open
`The garage door, it is only Kalle who can open.' (ibid. ex. 36)
c. [De blommorna]1 k?nner jag en man [som s?ljer t1].
those owers know I a man that sells
`These owers, I know a man that sells.' (Allwood, 1976 ex. 48)
(4) Hebrew (Sichel, 2018)
a. [al lexem Saxor]1 yeS rak gvina axat [Se-kedai limro'ax t1].
on bread black be only cheese one that-worth to.spread
`On black bread there is only one cheese that's worth spreading.' (ibid.
ex. 46a)
b. [al ha-haxlata hazot]2 Yair Lapid haya ha-axaron [Se-yada t2].
about the-decision this Yair Lapid was the-last that-knew
`About this decision, Yair Lapid was the last to know.' (ibid. ex. 47a)
c. [me-ha-sifria hazot]2, od lo macati sefer exad1 [Se-kedai
from-the-library this yet not found book one that-worth
lehaS'il t1 t2].
to.borrow
`From this library, I haven't yet found a single book that's worth bor-
rowing.' (ibid. ex. 49a)
(5) Mandarin Chinese
a. [Zhe-ben shu]1 wo renwei [du-guo t1 de] ren bu duo.
this-CL book I think read-EXP MOD person not many
`This book, I think there are not many people who have read it.' (Xu
and Langendoen, 1985 p. 14 ex. 61a)
b. [Zheme wanpi de haizi]1, wo zhao-bu-dao [yuanyi shouyang t1
so naughty MOD child I can't-nd willing adopt
de] ren.
MOD person
`Such a naughty child, I cannot nd someone who is willing to adopt
him/her.' (Zhang, 2002 ex. 5c)
(6) Brazilian Portuguese (sentences provided by Gesoel Mendes, judgments from
Mendes and Kayron Bevil?qua)
a. (?)[Esse ? o projeto de lei]1 que o presidente quer encontrar o
this is the project of law that the president wants to.nd a
senador [que suporte t1].
senator who supports.SBJ
25
`This is the bill that the president wants to nd a senator who supports.'
b. Viol?ncia1 ? uma coisa que tem muitas pessoas [que perdoam t1].
violence is one thing that there.is many people who forgive
`Violence is something that there are many people who condone.'
Since relative clauses can block A-bar movement in these languages, one cannot
explain the acceptability of the above porous relative clause examples by claiming
that these languages lack a Complex NP constraint or an adjunct island condi-
tion. Rather, what is more likely is that these languages exceptionally allow A-bar
movement from porous relative clauses.
(7) *[De blommorna]1 kysste jag en man [som salde t1].
those owers.DEF kissed I a man who sold
`Those owers I kissed a man who sold.' (Christensen and Nyvad, 2014 p.
31, ex. 5, citing Kush and Lindahl, 2011) [Swedish]
(8) *[Me-ha-sifria hazot]2, od lo macati [et ha-sefer1 [Se-kedai
from-the-library this yet not found ACC the-book that-worth
lehaS'il t1 t2]] Se-Dina mac'a.
to.borrow that-Dina found
`From this library, I haven't yet found the book that's worth borrowing that
Dina found.' (Sichel, 2018, ex. 67a) [Hebrew]
(9) *[Zheme wanpi de haizi]1, wo zhaodao-le [shouyang t1 de] ren.
so naughty MOD child I nd-PRF adopt MOD person
`Such a naughty child, I found the person who adopted him/her.' (Zhang,
2002 ex. 5c) [Mandarin]
(10) *Esse ? [o problema]1 que a Maria contratou o programador [que
this is the problem that the Mary hired the programmer that
resolveu t1].
solved
`This is the problem that Mary hired the programmer who solved it.' (sen-
tences provided by Gesoel Mendes, judgments from Mendes and Kayron
Bevil?qua) [Brazilian Portuguese]
In an short but important paper, Cinque (2010) reports cross-linguistic vari-
26
ation in A-bar movement from porous contexts. Looking also at relativization and
topicalization, he observes that equivalent structures in German and Bulgarian are
unacceptable. The Bulgarian data are striking, as they contain resumptive pro-
nouns/clitics (12), whose presence is traditionally held to ameliorate island viola-
tions (Ross, 1967) but fails to do so here.2 As further evidence of cross-linguistic
variation, I give examples from Polish (Marta Ruda, p.c.) and French (Valentine
Hacquard, Anouk Dieuleveut, p.c.). These languages appear to block A-bar move-
ment of objects from relative clauses.3
(11) German (Cinque, 2010, ex. 8)
a. *Diese Schrift gibt es niemand, der gelesen hat.
this writing there be nobody who read has
`This writing there is nobody who has read.'
b. *Dies habe ich nie jemand getroen, der getan hat.
this have I never anyone met who done has
`This, I have never met anyone who has done.'
(12) Bulgarian (Cinque, 2010, ex. 9)
a. *Ivan, na kojto njama nikoj, {kojto/deto} mo?e da
Ivan, to whom there-isn't nobody who/that can to
mu ka?e novinata.
him.clitic.DAT tell news.DEF
2That said, recent experimental work on resumptive pronouns, sparked o by papers like Alex-
opoulou and Keller, 2007, has found that they do not improve (but nor further degrade) the
acceptability of island violations.
3My classication of French together with German runs counter to that by Cinque (2010), who
claims that French patterns like English, based on examples like (i). However, the acceptable
examples all contain sentential negation, and they are the only instances of A-bar movement
of objects from relative clauses that were perceived to be more acceptable by the French native
speakers I consulted with. Other examples, explicitly modeled on acceptable non-negated sentences
in English (Scandinavian, Hebrew, Mandarin Chinese), were judged to be unacceptable. It is clear
that French is signicantly more conservative than English (Scandinavian, etc.) is, although I do
not yet fully understand the role of negation, nor why German and French might diverge in the
acceptability of sentences like (i).
i. Jean, ? qui il n' y a personne qui puisse s' opposer.
Jean, whom there not is nobody who can.SBJ self to.oppose
`Jean, whom there is nobody that could oppose, . . . ' (Cinque, 2010, p. 84, ex. 5a)
27
`Ivan, who there isn't anybody who can tell him the news . . . '
b. *Ivan, na kojto njama nito edin prijatel, {kojto/deto} iska da
Ivan, to whom there-isn't not one friend who/that wants to
mu pomaga.
him.clitic.DAT help
`Ivan, who there isn't anybody who wants to help . . . '
(13) *To jest ten problem, kt?ry Jan chce znale?? programist?, kt?ry
this is the problem which Jan wants nd.INF programmer who
b?dzie potra? rozwi?za?.
will be.able solve.INF
`This is the problem that Jan wants to nd a programmer who can solve.'
[Polish (Marta Ruda, p.c.)]
(14) *C' est le probl?me que Jean veut trouver un programmeur qui
this is the problem that Jean wants nd a programmer who
puisse r?soudre.
can.SBJ to.solve
`This is the problem that Jean wants to nd a programmer who can solve.'
[French (Valentine Hacquard, Anouk Dieuleveut, p.c.)]
2.1.2 How porous are porous relative clauses?
Acceptable instances of A-bar movement from relative clauses reported in the
literature  whether for English or for other languages  typically feature rela-
tivization or topicalization from a porous relative clause. It is less common for an
example to feature other types of A-bar movement, such as the movement of a wh-
phrase or tough movement. As far as I can tell, there is no explicit discussion of
why certain constructions are omitted. The literature seems to implicitly suggest
that the omission is just out of convenience, but it is also possible that this issue
has not been studied in detail. At least in English, the compatibility of dierent
A-bar movement constructions with porous relative clauses seems to vary (15), but
28
I will leave a more precise characterization for future work. Nevertheless, I take the
relativization and topicalization facts to show that there are certain constructions
in some (but not all) languages like English or Hebrew that are inconsistent with
standard generalizations about relative clauses.
(15) a. [(?)Which problem/?*What]1 does Mary want to nd a programmer
[who can solve t1]? [Wh-question]
b. *John can solve more problems1 than Mary wants to nd a programmer
[who can solve t1]. [Comparative deletion]
c. ?This problem1 is hard to nd someone [who can solve t1]. [Tough-
movement]
That said, there are also restrictions on what kind of syntactic object can
relativize or topicalize from porous relative clauses. The examples above show that
NP objects of verbs can do so. There is also evidence from English and Swedish that
NP objects of prepositions may extract (16). It is less clear to me whether non-NP
arguments can extract; they can topicalize (in at least some contexts) in Hebrew
and in Brazilian Portuguese (17).4
(16) a. That's one trick1 that I've known a lot of people who've been taken in
[by t1]. (Chung and McCloskey, 1983, p. 708, ex. 9a)
b. Der1 vet jag m?nga [som har fastnat i t1].
that know I many that have gotten.stuck in
I know many people who have gotten stuck on that. (Swedish, naturally
occurring conversation cited in Lindahl, 2017, p. 90, ex. 84)
4It is harder to say whether PP arguments may move from porous relative clauses in English
(i); A-bar movement of argument PPs is at best stilted, even in syntactically simple contexts (ii).
i. a. ??That's the diplomat [to whom]1 they need to nd an ambassador [who can give the
award t1]
b. ??That is the small town [in which] the mining company wants to nd someone [who is
willing to live t1].
ii. a. ?[To whom]1 will the ambassador give the award t1?
b. ?[In what town]1 is Mary living t1?
29
(17) a. Be-yeruSalayim1 hem hayu hayexidim [Se-hiskimu lagur t1].
in-Jerusalem they were the.only.ones that-agreed to.live
Intended: `They were the only ones who agreed to live in Jerusalem.'
(Sichel, 2018, p. 345 ex. 19a)
b. [Nessa cidade]1, a Maria ? a ?nica pessoa [que est? disposta a
in.this city the Maria is the only person who is willing to
viver t1].
live
`Maria is the only person who is willing to live in this city.' (transla-
tion by Gesoel Mendes, judgments due to Gesoel Mendes and Ana N.
Mendes)
c. ??[Nessa cidade]1, eles querer encontrar algu?m [que esteja disposto
in.this city, they want to.nd someone who is willing
a viver t1].
to live
`They want to nd someone who is willing to live in this city. (ibid.)
The precise extent to which A-bar movement of adjuncts from porous relative
clauses is possible, both within a language and across languages, is not entirely
clear. There is not a lot of discussion of this issue in the literature, and it seems
to me that this perhaps reects the diversity of adjuncts (one could distinguish
between adjuncts that describe manner, time, location, and so on). To summarize
some general patterns reported in the literature: in a number of languages, porous
relative clauses do not seem to allow overt A-bar movement of at least some adjuncts,
including temporal or locative PPs (18) or wh-adjuncts like how and why (19).5
(18) Movement of PP adjunct degraded
a. *[In the next session of congress], the president wants to nd a senator
[who will support the law t1].
b. *[Na semana que vem]1 o presidente quer encontrar um senador
in.the week that comes the president wants nd a senator
[que suporte a lei t1].
who supports.SBJ the law
5Since there is no obvious gap in the relative clause in these examples, the acceptability of these
examples is derived by asking native speakers whether the examples can have the same meaning
as an example where the adjunct is found in its intended position. To the extent that they cannot,
the A-bar movement examples are unacceptable.
30
Intended: `The president wants to nd a senator who will support the
law next week.' (Brazilian Portuguese; Gesoel Mendes, p.c.)
c. *Be-yeruSalayim1 hem hayu hayexidim [Se-hiskimu le'exol t1].
in-Jerusalem they were the.only.ones that-agreed to.eat
Intended: `They were the only ones who agreed to eat in Jerusalem.'
(Hebrew; Sichel, 2018, p. 345 ex. 19b)
(19) Movement of wh-adjuncts degraded
a. *How1/[How quickly]1 does Mary want to nd a programmer [who can
solve this problem t1]?
b. *Porque1 o presidente quer encontrar um senador que apoie
why the president want nd a senator who support.SBJ
a lei t1]?
the law
Intended: `What is the reason such that the president wants to nd a
senator who supports the law for that reason?' (Brazilian Portuguese;
sentence and judgement from Gesoel Mendes, p.c.)
c. *Varf?r1 k?nner du m?nga [som har skrivit b?cker t1]?
why know you many that have written books
Intended: `What is the reason do you know many who have written
books for that reason?' (Swedish; adapted from Lindahl, 2017, p. 199
ex. 53)
That said, Engdahl (1997) gives an example of acceptable A-bar movement of
a manner adjunct from a porous relative clause in Swedish; Lindahl (2017) further
points out that similar structures occur naturally. She also provides preliminary data
(her chapter 6) showing that under the right circumstances, a variety of adjuncts,
such as manner, time, reason, amount adjuncts, can be fronted as topics from porous
relative clauses. This contrasts with the analogous English examples (21), which
seem degraded. Likewise, in Sichel's description of Hebrew, A-bar movement of
adjuncts in general is reported to be not possible. If these contrasts are robust, they
point to even deeper cross-linguistic dierences, but more careful work is necessary
to determine whether that is the case.
31
(20) Acceptable A-bar movement of adjuncts in Swedish
a. S?1 k?nner jag ingen [som kan m?la t1]
like.that know I no.one who can paint
`I don't know anybody who can paint like that.' (Engdahl, 1997, p.
57)
b. . . . [genom Per]1 ?r det m?nga [som f?tt jobb t1].
through Per are there many that got job
`. . . there are many who have gotten jobs through Per.' (naturally
occurring example from Radio Sweden, cited by Lindahl, 2017, p. 175,
ex. 15, my translation)
(21) a. *In that style, I don't know anyone [who can paint t1].
b. ??[Through Per]1, there are many [who have gotten jobs t1].
Given the lack of clear generalizations about A-bar movement of adjuncts
cross-linguistically, I will not have much to say about the argument/adjunct split for
porous relative clauses. For ease of presentation, for the rest of the discussion, I will
restrict my discussion and examples to A-bar movement of arguments (specically,
of objects), abstracting away from the argument/adjunct subtleties.6
I note that when the porous relative clause in question is one where a non-
subject (e.g. object) is relativized, its subject cannot undergo A-bar movement
(22). This might receive an independent explanation, as Sichel (2018) also observes.
Assuming a bottom-up syntactic derivation, when such a porous relative clause is
formed, there are two phrases that eventually undergo A-bar movement: the subject
6I would like to also point out that at least within English, there appear to be more general
restrictions on how easily wh-adjuncts can move. Cattell (1978), for example, observes that wh-
adjuncts can move from embedded nite clauses that are complements of a class of attitude verbs
which he calls volunteer stance, such as think, believe, say and so on. Movement of wh-adjuncts
from the complement clauses of non-volunteer stance verbs like deny or regret is more degraded.
For scope reasons, I will not discuss Cattell's proposal (or subsequent proposals, like Hegarty,
1992, Szabolcsi and Zwarts, 1993, or Abrus?n, 2014, to give a few examples), but in light of these
restrictions on A-bar movement of adjuncts from complement clauses, the fact that porous relative
clauses impose restrictions on A-bar movement of adjuncts is thus in a way the rule rather than
the exception.
32
and the non-subject (22c); the subject is structurally higher of the two, but it is
the non-subject that rst undergoes A-bar movement (to form the relative clause).
Movement of the non-subject before the subject is taken to violate a general syntactic
principle like the Path Containment Constraint, Relativized Minimality or Shortest
Move (Chomsky, 1995; Pesetsky, 1982; Preminger, 2010; Rizzi, 1990).
(22) a. *This is the criminal1 who I know some people2 [t1 cheated t2].
b. *AnaSim ka-ele1, ra'iti [maSehu2 [Se-t1 moxrim t2].
people like-that, saw.I something that sell
`I saw something that people like that sell.' (Hebrew; Rubovitz-Mann
(2000), fn. 2, ex. ib, cited in Sichel, 2018, p. 347, ex. 28b)
c. [RC . . . subject . . . non-subject . . . ]
2.1.3 Characteristics that distinguish porous relative clauses
In this section, I turn my attention to the question of what features distinguish
these relative clauses. I start with an overview of examples found in the literature,
summarizing existing descriptions about porous relative clauses from various lan-
guages. I also point out that the relative clauses have distinctive tense and temporal
interpretation properties, an observation that I believe has not been discussed in de-
tail before in the literature. With these descriptive generalizations in hand, I discuss
how to characterize these relative clauses, discussing several alternatives.
2.1.3.1 Are porous relative clauses always subject relative clauses?
One often comes across claims that porous relative clauses are necessarily sub-
ject relative clauses (Chomsky, 1986; den Dikken, 2007; Engdahl, 1980; Erteschik-
Shir, 1973; Kush et al., 2013; Taraldsen, 1982), supported by judgments for non-
33
subject relative clause examples such as (23).
(23) *This is a paper2 that we really need to nd someone1 [that we can intimidate
t1 with t2]. (judgment by Chung and McCloskey (1983))
However, recent work on Hebrew and Swedish has reported acceptable A-bar
movement from non-subject relative clauses. These examples can be sorted into
two categories based on the niteness of the relative clauses. It is not clear to me
that the non-nite examples genuinely counter-exemplify the point about subject
relative clauses (pace Sichel), since classic island eects involving relative clauses
(and the lack thereof) are based on examples with nite relative clauses, not non-
nite ones. The examples that pose a bigger problem for the claim that porous
relative clauses are subject relative clauses are those in (24b), which involve nite
relative clauses. Overall, I agree with Lindahl (2017) and Sichel (2018) that porous
relative clauses do not always have to be subject relative clauses, although, again,
more work is necessary to determine the circumstances in which A-bar movement
from non-subject relative clauses becomes acceptable.7
(24) a. Non-nite relative clauses
(i) [im ha-balSan ha-ze]2, od lo macati be'aya1 [PRO ledaber
with the-linguist the-this yet not found.I problem to.talk
t2 aleya1].
about.it
`With this linguist, I haven't yet found a problem to talk about.'
(Sichel, 2018, p. 347, ex. 29a)
7That said, we may draw some preliminary generalizations based on the acceptable Hebrew
and Swedish examples: in the non-nite cases (24a) and (24b-i), the subject of the relative clause
appears to be bound by the matrix subject, recalling Grano and Lasnik's observations that locality
restrictions can be lifted when the subject of a locality domain is a bound pronoun. In (24b-ii),
however, the subject is an indenite pronoun man. I have yet to come across an example where
the subject of a porous relative clause is a R-expression referring to an individual. As for (23), its
low acceptability might partly reect the fact that it describes an odd scenario: the speakers are
describing their obligation to look for someone who is for some reason afraid of a paper.
34
(ii) Det2 har jag inte haft n?gonting1 PRO att g?ra t1 med t2.
that have I had not anything to do with
`That, I have not had anything to do with.' (adapted from Eng-
dahl, 1997, p. 57, ex. 16a)
b. Finite relative clauses
(i) . . . det2 har jag inte hittat n?gon ryggs?ck1 [jag inte blir
that have I not found some backpack I not become
t1 av t2].
of
`. . . I have not found any backpack that I don't become that [=
sweaty] from.' (Forum thread cited by Lindahl, 2017, p. 96 ex.
95)
(ii) [just den h?r delen av matten]2 kommer ni inte hitta
precisely the here part of math.DEF will you not nd
n?tt jobb1 [d?r2 man beh?ver t2 t1].
some job where one needs
`You won't nd any job where you need THIS part of math.'
(Converation cited in ibid. ex. 97)
2.1.3.2 What environments do porous relative clauses appear in?
A closer survey of examples of porous relative clauses shows that four types of
constructions make up the bulk of examples. I enumerate and label these construc-
tions below, illustrating them with examples from English.
(25) a. Non de re
(i) The president wants to nd a senator [who supports this bill].
(ii) That is the bill1 that the president want to nd a senator [who
supports t1].
b. Predicational
(i) Stevie is the only person [who wanted to record the song].
(ii) That is the song1 is Stevie the only person [who wanted to record
t1].
c. Existential
(i) There are many people [who condone violence] in this town.
(ii) Violence is something1 that there are many people [who condone
t1] in this town.
d. Evidential existential (Rubovitz-Mann, 2000)
(i) I've known a lot of people [who've been taken in by this trick].
(ii) That's the trick1 that I've known a lot of people [who've been
35
taken in by t1].
On a side note, these examples also involve A-bar movement from within
a noun phrase; work looking at A-bar movement from PP complements of noun
phrases have shown that the deniteness of the noun phrase aects acceptability
(Chomsky, 1973; Davies and Dubinsky, 2003; Huang, 2018a; Simonenko, 2015).
However, (25) shows that A-bar movement in these cases is not dependent on
whether the complex NP is morphosyntactically denite or not; the predicational
example shows that denite complex NPs allow A-bar movement.
The labels should be fairly self-explanatory, except for non de re and eviden-
tial existential. I will start with evidential existential, which is easier to describe:
this construction asserts the existence of individuals described by the complex NP,
although it does so indirectly through a verb of perception (see) or acquaintance
(know) (cf. Sichel, 2018).
As for non de re, this refers to the fact that NPs in such a setting can in
principle receive three readings (Fodor, 1970). I illustrate this ambiguity with (26).
(26) Mary needs to nd [NP a professor who can teach Vata].
a. De re (specic, transparent to speaker), reporting a situation where
Mary said, I need to nd Joe. The speaker knows that Joe is a
professor and Joe can teach Vata.
b. De dicto (non-specic, opaque to speaker), reporting a situation where
Mary said, I need to nd a professor who can teach Vata  I don't
care who it is.
c. Third reading (non-specic, transparent to speaker), reporting a sit-
uation where Mary said, I need to nd a professor who can teach a
non-European adposition stranding language  I don't care who it is or
what language it is. Mary actually doesn't know of any non-European
language with this property, nor has any express desire to nd a Vata
teacher. However, the speaker knows that Vata meets Mary's criteria,
36
and reports Mary's statement as (26).
As mentioned above, Sichel (2018) observes that A-bar movement from a rel-
ative clause is blocked if the complex NP receives a de re reading (i.e. presupposi-
tional, in Sichel's terms), even if the context makes clear the intended reading. In
(27), the NP denotes an actual individual, Joe.
(27) Context: Jack and Jill overhear Mary speaking on the phone: I need to
nd Joe. Jill says to Jack: Sounds like Mary needs to nd a professor who
can teach Vanuatu for the eldwork class. Jack knows that Jill has made
a mistake; Joe specializes in Vata, not Vanuatu. Jack corrects her:
*It's not Vanuatu, but Vata1, that Mary needs to nd a professor who can
teach t1].
A-bar movement becomes acceptable when the speaker (or Mary) does not
have any actual professor in mind, e.g. when the NP gets a de dicto interpretation.
In fact, this is also the case for the third reading, although to the best of my
knowledge, this point has not been explicitly made in the literature. In Schwager,
2009, Kaufman observes that a complex NP that receives the third reading can
refer to an individual that does not exist in the actual world. She does so using an
example like (28), which is judged to be acceptable even though the indenite NP
a building with at least 192 oors is extensionally empty.
(28) Context: Mary is looking at the Burj Dubai, which has 191 oors and is
currently the highest building in the world. Also, no other building has
more oors. Mary doesn't know this. She also doesn't know how many
oors Burj Dubai has. She exclaims, Wow, I want to buy a building that's
even one oor higher! John overhears Mary, and reports Mary's statement
as, Mary wants to buy a building with at least 192 oors. (adapted from
Schwager, 2009, pp. 399400 exx. 11, 12)
When the complex NP has the third reading, it also allows A-bar movement
37
from the relative clause. A scenario that facilitates this reading is given in (29), and
A-bar movement takes place in (29b).
(29) Context: Jack and Jill overhear Mary speaking on the phone: I need to nd
a professor who can teach a non-European adposition-stranding language, if
such a language even exists. Both Jack and Jill know that such languages
exist, although they don't know of anyone who can teach them.
a. Jill mistakenly thinks that Vanuatu is such a language, and says to Jack,
Sounds like Mary needs to nd a professor [who can teach Vanuatu].
b. Jack corrects her: It is not Vanuatu, but Vata1, that Mary needs to
nd a professor [who can teach t1].
2.1.3.3 The semantics of the complex NP?
An interesting idea that has been dicult to pin down concretely is the idea
that porous relative clauses or the NPs they modify have a distinctive reading,
perhaps best captured by the term non-presuppositional (Sichel, 2018); other labels
found in the literature include prominent, presentational, non-backgrounded
(Erteschik-Shir, 1973; Goldberg, 2006; Lindahl, 2017; Rubovitz-Mann, 2000).
For instance, in the non de re and predicational environments, the complex NP
itself does not refer to a particular individual, as if the interpretation of the NP does
not involve the presupposition of a discourse referent. Likewise, in the predicational
environment (25b), the only person who wanted to record the song is a description
of Stevie but not an expression that refers to Stevie. If it were the latter, then the
sentences should be tautological, with a reading like Stevie is Stevie.
These two environments alone suggest that the complex NP in question might
be distinguished by its semantics: in these cases, it is extensionally empty or non-
38
referring. However, attractive as this generalization might be, it runs into problems
with the existential environments (or even the evidential existential construction, to
the extent that these two constructions have similar existential semantics). Although
there are proposals that claim that in existential constructions, the post-copular
NP (the pivot) is semantically not an individual (Keshet, 2008; McNally, 1997;
Schwarz, 2012), the fact remains that in armative frames, these complex NPs are
ultimately associated with some discourse referent.
Setting aside the technical challenges for now, I note that this kind of analysis
has a precedent in the literature on subject contact clauses. Following McCawley
(1998), Doherty (2000) observes that the environments that permit porous rela-
tive clauses also permit what is called subject contact clauses (SCCs) in certain
non-standard dialects of English. SCCs are bare VP-like constituents that appear
immediately on the right of the noun they modify, without a complementizer that or
a relative pronoun like who or which. McNally (2009) summarizes very succinctly the
distribution of SCCs in liberal/permissive dialects like Appalachian English and
Hiberno-English: they are licensed in certain modal contexts . . . ; as complements
to intensional verbs such as seek, where they have a strong or exclusive preference
for de dicto readings . . . ; in a slightly wider variety of copular sentences . . . , and in
the restriction of universally quantied DPs (p. 175). The SCC examples in (30),
taken from Doherty, 2000, bear a strong resemblance to the porous relative clause
examples discussed above.8
8The it-cleft case (30f) is interesting because these sentences are compatible with subject contact
clauses but are incompatible with A-bar movement from the relative clause 8, at least not for
English. However, Engdahl (1997) notes that relativization from cleft structures is acceptable in
Swedish.
39
(30) Subject contact clauses
a. We want someone [SCC knows John]. (Doherty, 2000, p. 91, ex. 112b,
noting that de dicto reading strongly preferred if not obligatory)
b. John is the only one [SCC can do it]. (ibid., p. 91, ex. 110a)
c. There's something [SCC keeps upsetting him]. (ibid., p. 72, ex. 40a)
d. I knew a man [SCC owned twenty restaurants]. (ibid., p. 92, ex. 114)
e. Everybody [SCC lives in the mountain] has an accent. (ibid., p. 92, ex.
113a)
f. It was John [SCC actually did it]. (ibid. p. 100, note 37)
Notably, in these dialects, SCCs cannot appear in what Doherty and McNally
call extensional contexts, where noun associated with the SCC receives an exten-
sional reading (31).
(31) Unacceptable SCC examples
a. *A man [SCC speaks Irish] walked into the bar. (Doherty, 2000, p. 91,
ex. 111a)
b. ??I gave a lift to this man [SCC knows Mary]. (ibid., ex. 111b)
c. {??A / *The} man [SCC speaks Irish] was at the party. (ibid., p. 93,
ex. 117)
In other words, there is a good case to be made that SCCs in liberal dialects
and porous relative clauses have the same distribution. Analyzing SCCs as relative
clauses, Doherty (1993) suggests that the condition licensing SCCs is semantic: the
noun that a SCC modies is non-referential, in the pretheoretical sense that it
fails to denote an individual in the real world (an extensional individual) (as cited
in Doherty, 2000, p. 91). Echoing this intuition, and adopting the analysis of
existential constructions in McNally, 1997, McNally (2009) suggests that SCCs are
i. a. *Which crime1 was it John who committed t1?
b. *That's the crime1 that it was John who committed t1.
I do not have a full explanation of the oddness of these examples, although I suspect that they
can be reasonably blamed on information structure of it-clefts in English. For example, (29b)
might be unacceptable because the sentence tries to focus two elements: the crime and John.
40
licensed in nominals that do not introduce persistent discourse referents into the
discourse model (p. 176, ex. 33).9
2.1.3.4 Tense and temporal interpretation in porous relative clauses
It is also possible to look at porous relative clauses through the lens of tense
and temporal interpretation, to the extent that tense is dissociable from the refer-
entiality/intensionality facts discussed above. In this section, I present novel facts
about English porous relative clauses that distinguish them from regular relative
clauses; some of these facts have been observed in the literature on tense (Abusch,
1988; Ogihara, 1996), but the connection to porous relative clauses has not been
made explicit. The generalization that emerges is that tense in porous relative
clauses is dependent; more specically, it is always evaluated with respect to the
tense in the main clause. This dependence sets porous relative clauses apart from
regular relative clauses, and mean that some of them pattern more like complement
clauses. I will discuss English facts, as the idiosyncrasies of tense in English are
relatively well-studied.
The temporal interpretation facts only become signicant after we consider
the corresponding datapoints for complement clauses and relative clauses. Conse-
9Doherty (2000, p. 92) points out complications for his proposed non-referentiality condition.
First, he observes that in contexts like (30d), it is not clear if the complex NP object of know is
referential, although I believe this objection can be mitigated if one adopts the analysis of evidential
existential constructions of Rubovitz-Mann, 2000 (and Sichel, 2018). Second, he observes that if
SCCs are licensed in all non-referential contexts, it predicts that they are licensed as long as
the NP is used predicatively, which is not the case. Third, as alluded to above, this condition
over-generates for more conservative dialects. In conservative dialects, SCCs are found only in
existentials, have-existentials (I have this friend lives in Dublin, p. 72 ex. 41a), clefts, and
copular sentences with deictic subjects, but not in e.g. intensional contexts.
41
quently, I will rst give an overview of facts related to tense and temporal interpreta-
tions in these contexts, using them as baselines (Abusch, 1988; En?, 1987; Gennari,
2003; Ogihara, 1996; von Stechow and Gr?nn, 2013a, among many others).
I rst consider the interpretations available when a past tense in the matrix
clause appears together with a past tense in a complement clause. These environ-
ments can receive a back-shifted reading: the time denoted by the lower past tense
is in the past of the time interval denoted by the higher past tense, which is in turn
in the past of the utterance time (32a). In certain circumstances, e.g. if the embed-
ded predicate is stative, the reading obtained can also be a simultaneous one, as if
the lower past tense were semantically vacuous (32b); this phenomenon is known as
sequence of tense.
(32) Hillary thought [that her husband was president].
a. OK: Hillary thought at time t, My husband was previously president.
(backshifted)
b. OK: Hillary thought at time t, My husband is now president. (si-
multaneous)
Importantly, the lower past tense cannot denote a time interval that is in
the past of utterance time but in the future of the higher past tense (i.e. a later
than matrix reading) (33). Importantly, in both cases, the embedded tense lacks a
later-than-matrix reading.10
(33) Hillary thought [that her husband was president].
a. Not OK: Hillary thought in 1993, My husband will be president in
1998. (later than matrix)
10The only exception that I am aware of is reported by Klecha (2016): for a set of verbs that
have a future orientation, like hope or pray, an embedded past tense c-commanded by a higher
past tense can get a later-than-matrix reading.
42
The tense literature has observed that tense in relative clauses generally runs
afoul of this generalization. The tense in a relative clause can be independent of
the tense in the main clause, allowing a later than matrix reading. For example,
in (34), the time of marriage is free to precede or follow the time of the husband
becoming president.
(34) Hillary married a man [who became president].
a. OK: Time of marriage precedes time of becoming president. (Hillary
married someone, and that person later became president.)
b. OK: Time of becoming president precedes time of marriage. (For ex-
ample, Hillary married a former president.)
When a past tense c-commands a present tense in a complement clause, a
double access reading becomes available. In (35), the time interval of Michelle's
husband being president must contain both the time of Michelle's speech and the
time of utterance.
(35) Michelle said [that her husband is president].
a. OK: One reports in 2015 that Michelle said in 2014, My husband is
president. (Barack Obama served as US president from 2008 to 2016.)
(double access reading)
b. Not OK: One reports in 2019 that Michelle said in 2014, My husband
is president. (relative present tense)
Having discussed temporal interpretation for complement clauses and regular
relative clauses, I now review temporal interpretation for porous relative clauses,
contrasting them with regular relative clauses. Essentially, in contrast to regular
relative clauses, the time interval denoted must be related to the time interval de-
noted by the main clause's tense, in ways that recall the restrictions that apply to
complement clauses.
43
First, when both tenses are in the past tense, there is no later than matrix
reading available. The time described by the embedded predicate can only be si-
multaneous with the main clause time, if not backshifted (Abusch, 1988; Ogihara,
1996). That said, to be clear, a later than matrix reading is available for (36a) if
the complex NP a senator who voted for this bill receives a de re reading (ibid.).
(36) a. The president was looking for a senator [who voted for this bill].
(When a de dicto reading is intended, this sentence cannot describe a
scenario where the president was looking for any senator who would
vote for the bill in the relative future.)
b. At the start of the language courses, Stevie was the only person [who
studied German].
(Intended but impossible scenario: a group of military ocers were
assigned to learn various languages. Stevie was the only one assigned
to learn German, most ocers were assigned to learn Spanish. The
commanding ocer uttered the above sentence after all the ocers
became uent in their assigned languages. The matrix past tense refers
to a time interval coinciding with the start of the language courses,
while the embedded past tense describes a later time interval that is in
the past of the utterance time.)
c. At the start of the language courses, there were many people [who
studied Spanish].
(Impossible for the scenario above)
d. At the start of the language courses, I knew a lot of people [who studied
Spanish].
(Impossible for the scenario above)
Second, a present tense in the relative clause generally cannot co-exist with a
past tense in the main clause (37). (A main clause present tense is compatible with
either a present tense or a past tense in the relative clause.) The one exception that
I am aware of is for the non de re environment (37a), which must get a double access
reading: the president had a desire to nd a senator who supported the bill at the
time of search through the time of utterance (Abusch, 1988; Ogihara, 1996). In this
respect, the non de re construction resembles complement clauses, possibly because
44
the main predicate in this construction is also a predicate with modal semantics,
e.g. want or look for.
(37) a. The president was looking for a senator [who supports this bill].
(Cannot describe a scenario where the president was looking for a sena-
tor who did not support the bill at the time of search but would support
the bill later at a time that coincides with the time of utterance.)
b. *Last Monday, Stevie was the only person (in the group) [who speaks
German].
(Intended but impossible scenario: Stevie is one of several native Ger-
man speaker in a group. Stevie arrived last Monday, while the other
German speakers arrived only on Tuesday.)
c. *In this school last year, there were many students [who speak Spanish].
(Intended but impossible scenario: last year, there were many native
speakers of Spanish enrolled in this school. This year, after Spanish
language classes were canceled, almost all these students transferred out
of the school. Since these speakers are still students and still Spanish
speakers, the present tense is in the relative clause arguably justied.)
d. *Last year, I knew a lot of students [who speak Spanish].
(Impossible for the scenario above)
As for the other three constructions, the restrictions on main clause past tense
and present tense in the relative clause for the constructions have a parallel in
specicational copular constructions (38), where the phenomenon is known as tense
harmony (Akmajian, 1970; Higgins, 1973; Romero, 2004, 2005; Sharvit, 2003).
(38) Tense harmony in specicational copular constructions
a. [What John was] was/is a fool.
b. [What John is] *was/is a fool.
2.1.3.5 Formal approaches to tense phenomena
In this section, I review two distinct approaches for modeling tense phenomena
in complement clauses and relative clauses. This review sets the stage for discussing
how to model the tense facts for porous relative clauses.
45
I rst discuss a Priorian analysis of tense, in which tenses are relative. In this
analysis, a tense denotes a time interval whose interpretation is determined with
respect to a higher (c-commanding) tense, if there is one. If there is no higher
tense, a tense is interpreted relative to utterance time. My discussion will focus on
Ogihara, 1996, an inuential proposal that covers a wide range of the facts.
In this analysis, a past tense that c-commands another past tense always
yields a backshifted reading. To account for the simultaneous reading, some kind
of Sequence of Tense rule is necessary. Ogihara suggests that when a tense c-
commands another tense of the same type, the lower tense can delete; the resulting
tenseless embedded clause is then interpreted as being in the relative present of (i.e.
simultaneous with) the higher tense.
This rule alone cannot explain why relative clauses allow for a later than matrix
reading. Ogihara derives this reading associated with relative clauses by positing
that (i) NPs move at LF from their surface position to a much higher position,
e.g. for quanticational purposes (such as QR); (ii) temporal interpretations are
assigned based on structural congurations at LF, after NP movement. In (39a),
as a result of LF movement, neither past tense morpheme c-commanding the other.
Both PASTa and PASTb are interpreted with respect to utterance time, and so the
time intervals they denote are in the past of utterance time, but not necessarily
with respect to each other. In contrast, complement clauses are not hypothesized to
move. Therefore, in (32), PASTb, which is in the complement clause, is necessarily
interpreted with reference to PASTa.
46
(39) a. LF of (34): [[a man who PASTa become president]1 [Hillary PASTb
marry t1]].
b. LF of (32): Hillary PASTa think [that her husband PASTb be presi-
dent].
An alternative view holds that tenses are absolute, and therefore are not in-
terpreted relative to other tense morphemes. This analysis easily accounts for the
temporal interpretations found for regular relative clause, but overgenerates for em-
bedded tense in complement clauses. All else being equal, it incorrectly predicts,
for example, that Melania said Donald became president (Melania PAST say [Don-
ald PAST become president]) can be used to report in 2019 a later-than-matrix
scenario where Melania said in 2014, Donald will become president in 2017. Con-
sequently, one requires additional stipulations to derive the temporal interpretation
facts for complement clauses. For example, one could imagine stipulating that
tenses in complement clauses are obligatorily bound by some higher tense (for a
similar obligatory binding analysis, although motivated for very dierent reasons,
see Kusumoto, 2005).
This absolute analysis of tense is in particularly compatible with a view in
which tenses are pronoun-like (e.g. Kratzer, 1998; Partee, 1974). In this analysis,
a past tense can refer to a particular time interval that is prior to utterance time,
just as a pronoun, like she, refers to a particular (human) individual that is female.
Consider (40), which Partee uses to argue that the operator approach does not
generate the right interpretation, which can be paraphrased as At a particular past
time interval, I did not turn o the stove.
(40) I didn't turn o the stove! (said shortly after leaving the house)
47
a. Priorian approach in which tenses are operators on propositions
(i) PAST [? [I turn o the stove]] (too weak: There is a past time
interval in which I did not turn o the stove)
(ii) ? [PAST [I turn o the stove]] (too strong: There is no past time
interval in which I turned o the stove)
Both approaches do not naturally accommodate double access readings. Within
a relative tense approach, this reading is usually accounted for by assuming that the
English present tense is exceptional in being a relative tense that also has an index-
ical component. As a result, the time interval it denotes has to be interpreted with
respect to some higher tense, while also containing the utterance time. An absolute
tense analysis models the indexical element much more easily, but requires further
stipulation to explain why the present tense is interpreted relative to a higher past
tense. (See Gennari, 2003 for a pragmatic analysis of this phenomenon, and Alt-
shuler et al., 2015 for discussion of the circumstances that give rise to the double
access reading.)
Conventional analyses of sequence of tense phenomena are also silent on tense
harmony eects. Recent proposals on tense harmony by Sharvit (2003) and Romero
(2004) derive these eects by assuming that the present tense in this construction
is both indexical and relative to the tense in the main clause. In particular, they
stipulate that the tense in the embedded clause is such that the time interval denoted
must contain the utterance time and be contained by the time interval denoted by
the main clause's tense. As a result, a past tense in the main clause cannot combine
with a present tense in the relative clause, as doing so would produce a contradiction:
the time denoted by the present tense would overlap with the utterance time and
48
yet be contained by some past time interval.
2.1.4 Possible analyses of porous relative clauses
To summarize, compared to regular relative clauses, porous relative clauses
are subject to more restrictions on what kinds of tense may appear in them and
what kind of interpretations they receive. More abstractly, we can think of tense in
relative clauses as being independent, and therefore free of restrictions; in contrast,
tense in porous relative clauses is dependent. This generalization raises questions
about how this tense dependency is represented and how it might be related to
A-bar movement.
As I see it, a complete answer to these questions would require a full theory of
tense, including a theory of double access and a theory of tense harmony. Existing
analyses of both phenomena invoke the idea that an embedded present tense stands
in some relation with a structurally higher tense, but implement this idea in ways
that are incompatible with each other. Specically, as mentioned, conventional anal-
yses of double access facts need to ensure that an embedded present tense denotes
a time interval that contains the time interval denoted by the main clause's tense,
while Sharvit's and Romero's accounts of tense harmony require that an embedded
present tense denote a time interval that is contained by the time interval denoted
by the main clause's tense. (Romero (2004) concedes that her analysis does not
account for double access (p. 287).)
For the following discussion, I will abstract away from these technical issues
49
surrounding the semantic representation of tense. Assuming more generally that
there is some tense dependency involved, I will instead focus on the question of
whether it has a syntactic correlate, so that we may connect it more precisely with
A-bar movement. As I see it, there are two distinct approaches of giving a syntactic
analysis of this phenomenon, although other analyses, perhaps more satisfactory
ones, could also be possible. For concreteness, I present both approaches below.
2.1.4.1 A movement-based analysis
The rst approach is a movement-based analysis. In this proposal, complex
NPs modied by regular relative clauses (possessing independent, non-anaphoric
tense) are presuppositional and therefore move (covertly) to a position higher than
the main clause tense projection, as envisioned, for instance, in a classical theory of
QR (pace Sichel, who suggests that this movement might be a kind of dierential
object marking and not for quantication reasons; see her fn. 20, p. 355). As
mentioned above, as a result of this movement, the tense in the complex NP is no
longer c-commanded by the tense in the main clause. In a relative tense framework,
this structural conguration also means that the tense in the complex NP is not
interpreted relative to the tense in the main clause (Abusch, 1988; Ogihara, 1996;
von Stechow and Gr?nn, 2013a; Stowell, 1993), thus deriving tense independence
(which manifests itself in the form of later-than-matrix readings, for example).
In contrast, an NP that does not move remains in situ. A tense in a relative
clause modifying this NP is thus c-commanded by the main clause tense; the em-
50
bedded clause, and therefore the lower tense, is interpreted relative to the higher
tense. This conguration yields tense dependence in the form of simultaneous and
backshifted readings or tense harmony eects. Additional assumptions, such as the
assumption that an embedded present tense can undergo res movement, can derive
the availability of double access readings (but see Gennari, 2003 for a account based
on pragmatics).11
This analysis aligns rather neatly with the movement-based analysis proposed
by Sichel, 2018 (pp. 355.) to account for restrictions on A-bar movement from
relative clauses in Hebrew. In Sichel's proposal, the covert movement operation
causes these NPs to freeze (Uriagereka, 1999; Wexler and Culicover, 1980, among
many others), blocking A-bar movement.12 In contrast, the complex NPs of porous
relative clauses do not freeze, since they do not move.
While the movement-based account can help explain tense dependence and re-
strictions on A-bar movement, it undergenerates. Specically, it predicts that porous
relative clauses modifying subjects should be incompatible with A-bar movement,
if one assumes (standardly) the VP-internal subject hypothesis, in which NPs occu-
pying a subject position move to that position. This prediction is not borne out: in
Mandarin Chinese, A-bar movement from porous relative clauses modifying subjects
is attested (41). (For an argument that subjects move from a VP-internal position
11That said, these proposals are silent about tense harmony in specicational copular construc-
tions; technically, because the relative clause modies the subject, the tense in the relative clause
does not c-command the matrix tense. Nor does the matrix tense c-command the relative clause.
One might thus incorrectly expect the absence of tense harmony eects.
12An important assumption is that covert movement of NPs must precede overt A-bar movement.
While this ordering of operations is not possible under a Y-model of syntax, Sichel notes that there
are other models of syntax that allow it. Jerey Lidz (p.c.) points out that it is also possible that
the operation that moves NPs is overt but string vacuous.
51
in Mandarin Chinese, see Huang (1993).)
(41) [Zhe-ben shu]1 wo renwei [Subject [du-guo t1 de] ren] bu duo.
this-CL book I think read-EXP MOD person not many
`This book, I think there are not many people who have read it.' (Mandarin
Chinese; Xu and Langendoen, 1985 p. 14 ex. 61a) (= (5a))
2.1.4.2 A defective tense, phase-based analysis
An alternative analysis that avoids the undergeneration problem is a phase-
based analysis, on the standard assumption that (nite) clauses (CPs) are phases,
which are locality domains (Chomsky, 2000, 2001). I will begin this discussion with
how a phase-based analysis can explain classic restrictions on A-bar movement from
relative clauses (as described by, e.g. Ross (1967)). First, since relative clauses are
nite clauses, it is reasonable to think that they are also phases. A-bar movement
from relative clauses is eectively ruled out by the Phase Impenetrability Condition
(PIC), which (essentially) requires A-bar movement from phases to proceed through
the left edge of each phase Chomsky (2000, 2001). A standard assumption for many
languages, including English, holds that there is only one position on the left edge for
successive cyclic movement; if this position is occupied, e.g. by a relative pronoun
or some operator, other phrases cannot move out of the phase without violating the
PIC (42). This particular conguration alone can explain why relative clauses are
islands for A-bar movement, since relative clauses are formed by A-bar movement
of either a relative pronoun or phonologically-null operator to the left edge of the
clause.
52
(42) *wh1 . . . [Phase wh2 . . . t2 . . . t1 . . . ]
I now address how the PIC might be voided in porous relative clauses. In a re-
cent proposal, Grano and Lasnik (2018) (also Barros and Frank, 2017, Huang, 2018a)
observe that some dependencies cannot cross the boundary of a nite clause, unless
the clause has a bound pronoun subject. Assuming that clauses are phases and that
these dependencies cannot cross a phase boundary, they suggest that bound pro-
nouns can neutralize the phase. More specically, they argue that bound pronouns
can be syntactically defective and fail to value features on T; unvalued features on
T then neutralize the CP phase. (Barros and Frank, 2017 and Huang, 2018a of-
fer dierent implementations of this eect, but share the core intuition that bound
pronouns (among others) are syntactically distinguished and are responsible for this
neutralization property.
(43) Comparative deletion (Grano and Lasnik, 2018:472 exx. 9-11) (Struck-out
material is silent but interpreted.)
a. More teachers gave the students pencils than gave the students pens.
b. *More teachers claimed that the principal gave the students pencils than
claimed [that the principal gave the students pens].
c. ?More teachers1 claimed that they1 gave the students pencils than claimed
[that they1 gave the students pens].
(44) Too/enough construction (ibid.)
a. This book is too valuable for James to lend _ to Bill.
b. *This book is too valuable for James to claim [that Mark lent _ to Bill].
c. ?This book is too valuable for James1 to claim [that he1 lent _ to Bill].
This phase neutralization proposal provides another way to explain why porous
relative clauses exceptionally allow A-bar movement in languages like English. As
pointed out previously, tense in porous relative clauses is dependent. Suppose that
53
this semantic property is reected syntactically as an unvalued temporal feature on
T. Perhaps because this feature is temporal in nature, it can only be valued by a
higher tense morpheme and not by any syntactic object in the same clause. Borrow-
ing the intuition in Grano and Lasnik's proposal, this unvalued feature neutralizes
the CP phase. Consequently, since the relative clause is no longer a phase, A-bar
movement from porous relative clauses is no longer subject to the PIC. In contrast,
independent tenses have valued temporal features, and therefore do not have the
same neutralization eect. Consequently, the regular relative clauses in which inde-
pendent tense appears are phases, and A-bar movement from these relative clauses
violate the PIC.13
This adaptation also builds on a number of parallels between dependent tense
and bound pronoun subjects. First, by denition, their interpretation depends on
other elements outside of the clause they appear in. Second, structurally speaking,
they occupy positions in the same region in the left periphery of a clause; subjects
are conventionally analyzed as occupying the specier position of T(ense)P. Fur-
ther, there is a tradition of treating tenses as being pronominal, as reviewed earlier
(Kratzer, 1998; Partee, 1974, for example).
13One question that arises is whether the presence of a bound pronoun subject in a nonporous
relative clause would make it porous. A plausible-sounding example is not easy to construct; (45)
is one somewhat plausible example, and A-bar movement from the relative clause is not acceptable.
(45) a. Hillary1 married a man2 [who she1 later caught t2 committing adultery].
b. *?What crime3 did Hillary marry a man [who she later caught committing t3]?
c. *?That's the crime3 that Hillary married a man [who she later caught committing t3].
It's not immediately clear that the A-bar movement examples are bad because of this particular
structural conguration. These sentences seem to be unnecessarily complicated. For example, both
the wh-question and relativization examples seem to be about identifying the crime that Hillary's
husband committed. If these are indeed the discourse goals of uttering these sentences, then it
would be far simpler and more ecient to just say What crime did Hillary catch her husband
committing? or That's the crime that Hillary caught her husband committing.
54
In this defective tense analysis, the porousness of a relative clause depends (in
part) on the properties of T, not on whether the relative clause is part of a NP that
moves. In principle, this analysis is compatible with A-bar movement from porous
relative clauses that modify subjects, as observed in Mandarin Chinese. Of course,
for this analysis to account for porous relative clauses in Mandarin Chinese in the
rst place, one would need to assume the presence of tense projections and anaphoric
tense in Chinese, which does not mark tense or niteness morphologically.14 The
representation of tense in Chinese is a topic that lies beyond the scope of this
dissertation, but for proposals that assume the existence of tense(-like) projections
or tense morphemes in Mandarin Chinese, see for instance Grano, 2017; Huang,
1982; Li, 1990; Lin, 2015; Sybesma, 2007; N. Huang, 2015, 2018b; Chen, 2017).
All else being equal, this approach would overgenerate for a language like En-
glish (46). To avoid over-generation, one would need to assume that unlike Mandarin
Chinese, English imposes (more) restrictions on A-bar movement from subjects.
(46) *That's the book1 that I think [people [who read t1]] are many in number.
2.1.4.3 Further complications
I previously observed that the movement-based analysis undergenerates. Here,
I point out that it might over-generate: all else being equal, it predicts that all NP
adjuncts that remain structurally low at LF are compatible with A-bar movement.
This is not the case: PP adjuncts and bare participial adjuncts (often descriptively
14If one assumes that there are no tense projections in Mandarin Chinese, then one might predict
that Mandarin Chinese relative clauses are always porous. This is not a desirable prediction, since
there are relative clauses that are incompatible with A-bar movement (9) (Zhang, 2002).
55
labeled as reduced relative clauses) of NPs are not compatible with A-bar move-
ment (47), even if the complex NPs they modify appear in environments that should
allow A-bar movement, e.g. a non de re construction (47). (As far as I can tell,
Sichel (2018) does not discuss A-bar movement from these kinds of adjuncts.) If one
wishes to capture the restrictions in using the same freezing mechanism, one must
augment the analysis by assuming that these adjuncts obligatorily freeze, even if
the NPs they modify do not move.
(47) a. PP adjunct of non de re NP
(i) Mary wanted to interview someone [with a special skill].
(ii) *What1/Which skill1 did Mary want to nd someone [with t1]?
(iii) *That's the kind of skill1 that Mary wanted to interview someone
[with t1].
b. Bare present participle adjunct of non de re NP
(i) Mary wanted to interview someone [learning a musical instru-
ment].
(ii) *What1/Which instrument1 did Mary want to interview someone
[learning t1]?
(iii) *That's the kind of instrument1 that Mary wanted to interview
someone [learning t1].
c. Bare past participle adjunct of non de re NP
(i) Mary wanted to interview someone [trained in karate].
(ii) *What1/Which martial arts1 did Mary want to nd a man [trained
in t1]?
(iii) *That's the martial art1 that Mary wanted to nd a man [trained
in t1].
d. Bare adjectival predicates
(i) Mary wanted to interview someone [familiar with a musical in-
strument].
(ii) *What1/Which instrument1 did Mary want to interview someone
[familiar with t1]?
(iii) *That's the kind of instrument1 that Mary wanted to interview
someone [familiar with t1].
To account for these restrictions, the defective tense analysis requires the addi-
tional assumptions that these structures are always phases, (ii) the absence of overt
56
tense morphology mean that they lack defective tense, and that (iii) the presence of
defective tense (as opposed to, for example, the mere absence of independent tense)
is necessary to neutralize a phase.
2.1.5 Interim summary
In this section, I discussed several properties of porous relative clauses, drawing
together existing observations in the literature and novel observations. While there
are still some gaps in our understanding of the syntax of these relative clauses, I
pointed out that there are a number of cross-linguistic generalizations: unlike regular
relative clauses, in certain A-bar movement constructions, they allow the movement
of arguments, although they appear to still impose (some) restrictions on movement
of adjuncts. Syntactically, porous relative clauses tend to be subject relative clauses,
although acceptable A-bar movement from nite non-subject relative clauses has
been observed at least in languages like Swedish. Semantically, certain NPs modied
by porous relative clauses have distinct referential properties; a case study of English
porous relative clauses shows that these relative clauses also have distinctive tense
properties.
2.2 A learning problem
The cross-linguistic survey by Cinque (2010) found that not all languages allow
A-bar movement from porous relative clauses. This discovery raises an interesting
question about language acquisition: how might a learner determine whether he/she
57
is learning a language like English, which allows A-bar movement from porous rel-
ative clauses, or a language like German, which does not?
Before addressing this issue in detail, I wish to address a question about the
extent to which A-bar movement from porous relative clauses is actually acceptable /
well-formed. This question is important in light of recent formal experimental results
reported by Kush et al. (2013). While A-bar movement from porous relative clauses
is usually characterized as being acceptable (examples in Chung and McCloskey,
1983 and McCawley, 1998 are left unmarked), Kush et al. nd that in English,
A-bar movement from porous relative clauses in existential (48a) and evidential
existential environments (48b) is only of marginal acceptability (p. 254), even
though it is still relatively more acceptable than A-bar movement from regular
relative clauses (48c). (I evaluate Kush et al.'s proposal in Chapter 5. In the same
chapter (Section 5.3), I report results from an acceptability judgment experiment
about A-bar movement from relative clauses in a non de re environment and about
A-bar movement from another type of clausal adjunct whose relevance will shortly
become clear  purposive clauses.)
(48) Example stimuli from Kush et al., 2013, experiment 3 (pp. 251253)
a. That was the bill that there were many senators [who supported t1] in
the congress.
b. That was the bill that he knew many senators [who supported t1] in
the congress.
c. That was the bill that he met many senators [who supported t1] in the
congress.
One implication of this nding, then, is that existing reports might have over-
stated the acceptability of A-bar movement from porous relative clauses in English.
58
In fact, Kush et al. (2013) propose that the low absolute acceptability indicates
that A-bar movement from porous relative clauses is ill-formed (ungrammatical).
However, compared to other ill-formed structures, A-bar movement from porous
relative clauses is more acceptable because the parser is able to successfully repair
the ill-formed structure.
This is certainly a coherent argument and Kush et al. (2013) ought to be
recognized for contributing novel judgment data to the literature. That said, I
would like to suggest that Kush et al.'s judgment data do not necessarily mean that
these structures are ill-formed, especially in light of the fact that there is no widely
accepted theory that can relate acceptability judgments to grammaticality.
Further, as noted above, Cinque (2010) reports that languages vary in whether
A-bar movement from porous relative clauses is acceptable to their speakers, al-
though Cinque's sample sizes are much smaller than Kush et al.'s. To the extent
that the dierence reported Cinque are robust, it is straightforward to account for
the dierence by treating it as a grammatical phenomenon, reecting dierent cri-
teria for well-formedness (grammaticality).
The alternative is to posit cross-linguistic dierences in how speakers of dier-
ent languages map ill-formedness to acceptability. Kush et al. (2013) propose that
English speakers are able to repair ill-formed A-bar movement from porous relative
clauses. Taking this argument seriously, one might consider a generalized version of
this proposal: perhaps English (Mandarin Chinese, etc.) speakers are more success-
ful at repairing ill-formed A-bar movement from porous relative clauses than German
speakers are, although it is not immediately clear why such an asymmetry should
59
exist. Similarly, since, by hypothesis, English (and Mandarin Chinese) speakers can
repair some instances of A-bar movement from relative clauses, it is not immediately
clear why they do not easily repair all instances of A-bar movement from relative
clauses. I am not aware of any such proposal on the market; in Chapter 5, I show
that A-bar movement from porous relative clauses in Mandarin Chinese and regular
relative clauses in English poses problems for Kush et al.'s repair account.
One necessary assumption in this line of reasoning is that the judgments and
contrasts reported in the literature are valid. Although it is not feasible for the
scope of this dissertation, a formal survey ? la Kush et al., 2013 for languages
like Hebrew, Mandarin Chinese, German, Bulgarian, Polish, etc. to conrm the
observations of Cinque and others would go some way to determine whether this
particular assumption is valid.
A consideration of the structural properties of relative clauses also provides
some reason for why A-bar movement from porous relative clauses might be less
acceptable in absolute terms, even if they are structurally well-formed. Recall the
assumption made above that relative clauses are formed by A-bar movement of either
a relative pronoun or phonologically null operator to the left edge of a clause. A-bar
movement from a relative clause, whether it is porous or not, involves the syntactic
conguration schematized in (49), where wh1 stands for the syntactic object that
moves from the relative clause, and wh2 the relative pronoun or phonologically-null
operator that is coindexed with the head of the relative clause.
(49) . . . wh1 . . . [RC wh2 . . . t2 . . . t1 ]
60
I suggested that A-bar movement is possible because the complex NP is not
frozen (following in part Sichel, 2018) or because defective tense neutralizes the
phase. Even if the structure is well-formed in this regard, it is not necessarily
easy to parse. To the extent that it is dicult to parse, we predict low accept-
ability. More specically, immediately after processing wh2, the parser has to look
for not one, but two potential gap sites in the relative clause, one corresponding
to wh1 and the other for wh2. Maintaining a distinction between these two A-bar
dependencies might be dicult for the parser, perhaps due to similarity-based in-
terference (Hofmeister and Sag, 2010; Kluender, 2004; Lewis, 1996). In addition,
there are other proposals claiming that the processing of niteness and of A-bar
movement from embedded clauses might interact in ways that reduce acceptability
(e.g. Hofmeister, 2007; Kluender, 2004). I return to these issues in more detail in
Chapter 5 (Section 5.3).
Having argued that there is reason to think that the phenomenon is a gram-
matical one, I now return to the learning question: how learners of English might
learn that their language allows A-bar movement from porous relative clauses, and
how learners of languages like German learn that their language does not.
In a sense, there are really two closely-related questions being asked here. In
Section 2.1.3.4, I argued that a key characteristic that distinguishes porous relative
clauses from non-porous relative clauses is tense dependence. The rst question,
then, is how speakers identify tense dependence in a relative clause, so that they
can draw conclusions about whether the NP modied by the relative clause has
moved (or not) or whether the tense in the relative clause is defective. This is, of
61
course, not a trivial question, as tense dependence involves identifying a particular
relation between tenses in a sentence. Even supposing that learners have innate
knowledge of temporal semantics and thus knowledge of temporal relations, it is
not necessarily easy for them to determine what temporal relations are encoded
in sentences in general and how they are encoded, since the temporal relations
intended by speakers have to be determined from context (see similar remarks by
Wagner (1998) on this problem, which she refers to as the mapping problem).
That said, building o observations in the preceding discussion, it seems to
me that there could also be observable morphosyntactic cues that learners can make
use of. For example, in a language like English, one might expect there to be distri-
butional dierences between matrix and embedded tenses between porous relative
clauses and non-porous relative clauses. More specically, in a number of construc-
tions (e.g. existential constructions), tense dependence means that a present tense
in a porous relative clause cannot co-occur with a past tense in the main clause,
while there is no such restriction on regular relative clauses. These restrictions could
then lead to dierences in how frequently certain tense combinations appear in cer-
tain constructions. Assuming that learners are able to track these frequencies, they
might be able to conclude that in certain constructions, there are tense restrictions
 and hence dependencies  that exist between the relative clause and the main
clause.
Of course, for a language like Mandarin Chinese, which does not morphologi-
cally mark tense, speakers would have to resort to alternative distributional evidence.
Here, it seems plausible that learners could draw upon the distribution of various
62
overt temporal markers, such as the future marker jiang, modals, or aspect markers
(Huang, 1989; Li, 1990, N. Huang, 2015; Huang et al., 2018).
The second question is a question about evidence for A-bar movement: what
evidence do speakers have that their languages allow or disallow A-bar movement
from porous relative clauses? Below, I consider the following learning scenarios.
In the rst scenario, learners of languages like English get very direct evi-
dence about the acceptability of A-bar movement from porous relative clauses: they
actually get to observe that A-bar movement from these relative clauses occurs fre-
quently in their linguistic experience. In contrast, their German counterparts do
not. The exposure to A-bar movement from porous relative clauses, or lack thereof,
thus allows English learners to conclude that their language has a grammar that
allows A-bar movement from porous relative clauses. I will describe prior work that
shows that this very direct form of learning might be possible for a language like
Danish. However, I also report results from corpus studies showing that this scenario
is unlikely to be feasible for a language like English or Mandarin Chinese.
The second and third scenarios ip the logic around, by considering the hy-
pothesis that learners of languages like German receive clear evidence that A-bar
movement from porous relative clauses is actually unacceptable. In the second sce-
nario, these learners are explicitly told that such sentences are unacceptable. In the
third, learners are not explicitly told about the acceptability of these structures.
Instead, they observe some supercially-unrelated cue that leads them to the con-
clusion that the grammar of their language does not allow A-bar movement from
porous relative clauses. I provide arguments against both scenarios.
63
The upshot of this discussion is that learners of English or Mandarin Chinese
cannot count on their linguistic experience providing reliable evidence that A-bar
movement from porous relative clauses is possible. Further, it is unlikely that these
learners begin with a grammar that allows A-bar movement from porous relative
clauses (or perhaps more accurately, A-bar movement from relative clauses with
dependent tense). This leads us to the conclusion that learners of English and
Mandarin must rely on a more indirect source of evidence. I argue that this evidence
takes the form of A-bar movement from non-nite purposive clauses.
2.2.1 Learning an English-like grammar with direct positive evidence
There is reason to think that learners of Scandinavian languages might have
reliable access to A-bar movement from porous relative clauses in their linguistic
experience. Looking at three corpora consisting of a total of 18 hours of Danish adult
speech, Jensen (2002) found 10 tokens of A-bar movement from an relative clause or
relative clause-like constituent (as reported in Lindahl, 2017, p. 34). Assuming that
speakers are exposed to about 5,200 hours of linguistic input a year (using estimates
for child learners provided by Hart and Risley, 2003), this frequency works out to
about 2,900 tokens per year. In other words, at least older speakers of Danish
get to observe A-bar movement from porous relative clauses in naturally occurring
contexts. If similar frequencies are available in child-ambient Danish speech, then
Danish-learning children might be able to acquire A-bar movement from porous
relative clauses at a relatively young age from direct positive evidence. Conrming
64
this point will require further research, as there is no analysis of child-ambient Danish
speech that I am aware of, nor is there any study on the developmental trajectory
of speakers' knowledge of A-bar movement from porous relative clauses, whether
within a language or across languages.
But in any event, there is no a priori reason to expect that what is true for
Danish will automatically hold for other languages, especially for such a syntacti-
cally complex construction. In a study of 101,838 tokens of English child-directed
speech from the CHILDES database (MacWhinney, 2000), and 98,819 tokens of
adult-directed speech and text from the Penn Treebank (Marcus et al., 1999), Pearl
and Sprouse (2013b) reported no instance of A-bar movement from complex NPs
with nite clauses, which they dene as NPs dominating CPs.15 This fact suggests
the hypothesis that A-bar movement from relative clauses appear at dierent rates
across languages, and therefore the direct evidence is not always available for either
child learners (whose experience is approximated by the CHIDLES dataset) or for
older learners (whose experience is approximated by the Penn Treebank data).
2.2.2 Corpus study 1: How frequent is A-bar movement from relative
clauses in English and Mandarin Chinese?
To test the hypothesis that linguistic experience about A-bar movement from
relative clauses vary across languages, I re-ran Pearl and Sprouse's analysis on En-
glish and Mandarin Chinese corpora, making the same assumption that they are
15This fact about English is itself interesting, since English and Danish are genetically and
historically related and typologically quite similar, and therefore the null hypothesis is that the
distribution of constructions should be largely similar in both languages.
65
representative of linguistic experience. These languages were chosen partly because
they are languages where large parsed datasets are available.
2.2.2.1 Data and methods
Following Pearl and Sprouse (2013b), I wrote a Python script using the Natural
Language Toolkit package (Bird et al., 2009) to extract overt A-bar dependencies,
including relative clauses and wh-questions (root and embedded).
For English, I used the entire Penn Treebank dataset, consisting of 251,255
tokens (more than twice the size of Pearl and Sprouse's Penn Treebank data), ex-
cluding the 572 relatively formulaic tokens of the ATIS corpus, which consists of
queries and replies about airline travel. Note that the Penn Treebank data consists
of both written (Brown and Wall Street Journal corpora) and spoken (Switchboard)
registers. For Mandarin Chinese, I used the Chinese Treebank (Xue et al., 2010), a
collection of 51,477 sentences of more formal registers of Mandarin, e.g. newswires,
broadcasts, blogs.16 Since Mandarin Chinese is a wh-in-situ language and the Chi-
nese Treebank does not contain information on covert A-bar movement, the script
eectively extracted instances of relativization and (to a lesser degree) topicalization
from the Chinese Treebank.
Both treebanks are well-suited for testing this hypothesis. First, the tokens
found in these treebanks are syntactically more complex, and hence more likely to
contain relative clauses in the rst place. Second, in both treebanks, constituents
16I would like to thank Scott Kaplowitz for research assistance, and to Naomi Feldman and
Philip Resnik for giving me access to the Penn Treebank and the Chinese Treebank.
66
are labeled and traces are clearly marked. Consequently, the script was able to
straightforwardly identify all the constituents that dominated the trace but not the
head of the A-bar dependency, i.e. what Pearl and Sprouse call container nodes,
or more intuitively, the path of A-bar movement.
2.2.2.2 Results
The algorithm extracted a total of 43,999 A-bar dependencies from a dataset
of 251,255 tokens. Replicating Pearl and Sprouse's ndings for English, I found
no A-bar movement path through a complex NP with a nite subordinate clause,
which in Penn Treebank terms would be an NP that immediately dominates a nite
SBAR.17
For the Chinese Treebank data, the algorithm extracted 24,992 A-bar depen-
dencies. As was the case for the Penn Treebank, I found no dependency in the
Chinese Treebank that crossed into a complex NP with a subordinate clause. Be-
cause the Chinese Treebank data impressionistically seems to have more manual
coding errors, I also undertook a check of a random 20% sample of these depen-
dencies (5,067 sentences). Again, I found no dependency crossing into a complex
NP.
17I note there are actually a handful of instances where A-bar movement takes place from a
non-nite subordinate clause that might be analyzed as subordinate to an NP. However, as I note
in section 3.3, they might be plausibly reanalyzed as VP adjuncts. It is thus unlikely that these
tokens are reliable positive evidence for the acceptability of A-bar movement from a complex NP.
67
2.2.2.3 Discussion: Learning from direct positive evidence unlikely for
English and Mandarin Chinese
I will assume Pearl and Sprouse's CHILDES dataset to be representative of
the variety of syntactic structures encountered by an English-learning child of up to
three years of age, and the Penn Treebank to be representative of the written and
spoken tokens encountered by an older learner or an adult. I will also assume that
the sentences in the Chinese Treebank are representative of what older Mandarin
speakers are familiar with.
Given these assumptions and the above null results, we may conclude that
the input for English and Mandarin speakers is likely to contain very little direct
positive evidence of overt A-bar movement from (porous) relative clauses. This
suggests that speakers of these languages either (i) already begin with a grammar
that allows A-bar movement from porous relative clauses and German (Bulgarian,
etc.) speakers learn a grammar that disallows it, or (ii) speakers of English and
Mandarin rely on some less obvious source of evidence.
2.2.3 Retreating from an English-like grammar with direct negative
evidence
In this section and the next, I consider the logical possibility that learners
of German, Polish, etc. come to learn that all relative clauses are islands in their
languages.
68
We can rule out fairly condently that learners might do so using direct nega-
tive evidence. Such evidence for island constraints is eectively non-existent: native
speaker adults are generally unlikely to tell learners to avoid sentences with ill-
formed A-bar movement in their languages. It is similarly unlikely that learners will
actually produce A-bar movement from relative clauses and then receive some kind
of negative signal, such as explicit correction or implicit disapproval. And even if
negative evidence did exist, it is unclear whether learners make use of it. To take an
example from English, there are old prescriptivist injunctions against preposition
stranding and split innitives, but it is clear from the fact that they have existed
for a long time that these instances of negative evidence are ineective.
2.2.4 Retreating from an English-like grammar with indirect positive
evidence
In the absence of direct negative evidence, another option that a German (or
Bulgarian, Polish, French) learner with an English-like grammar has is to rely on
indirect positive evidence. Some cue that is readily available in their input, but
distinct from A-bar movement from relative clauses, might inform the learner that
their language bans A-bar movement from porous relative clauses. The challenge
is identifying what such a cue might be, given the dierences in the morphosyntax
of these languages: what easily-observed linguistic phenomena in these languages
could lead learners of these languages to conclude that their initial hypotheses about
relative clauses over-generate?
69
One interesting hypothesis, due to an observation by Cinque (2010), is that
learners might rely on how relative clauses are marked. Cross-linguistically, lan-
guages may or may not dier in morphosyntactically distinguishing between rela-
tive clauses and complement clauses. In German, for example, these two types of
clauses are marked distinctly: relative clauses are marked by an article-like parti-
cle, while complement clauses are marked by the complementizer dass. In contrast,
Scandinavian languages can mark both relative clauses and complement clauses with
som/sem and Hebrew marks both relative clauses and complement clauses with Se.
Building o these observations, Cinque suggests that this kind of embedded clause
marking predicts A-bar movement from relative clauses: languages that mark rel-
ative clauses and complement clauses dierently, like German, do not allow A-bar
movement from relative clauses, while languages that mark them uniformly, like
Scandinavian languages, Hebrew, etc. do.
More specically, Cinque speculates that in languages that mark both types
of clauses in a uniform manner, like Scandinavian and Hebrew, A-bar movement
might proceed through Spec,CP (p. 86). Although the original discussion is a little
vague, I believe that Cinque had in mind an analysis where som/sem/Se are heads
of a CP, which leave the specier position of CP free for successive-cyclic movement
of a wh-phrase (50a). In contrast, in languages where relative clauses are marked
distinctly, the relative clause marker indicates that the specier position is occupied,
either by the relative pronoun itself (50b) or by some phonologically-silent operator.
Because the specier position is already occupied, successive cyclic A-bar movement
of another syntactic object is blocked, resulting in ungrammaticality.
70
(50) a. wh1 . . . [NP . . . [CP t1 Csom . . . t1 [Scandinavian]
b. *wh1 . . . [NP . . . [CP der 2 C t2 . . . t1 [German]
From a language acquisition perspective, this analysis oers an interesting
solution to the problem of how learners of languages like German might nd out
that their languages block A-bar movement from relative clauses. Essentially, in
this scenario, A-bar movement from porous relative clauses is in principle possible in
these languages as it is in English, except that learners learn that all relative clauses
in their languages must be constructed in a way that is fundamentally incompatible
with subsequent A-bar movement.
That said, this analysis poses several problems. First, it is unclear where
English sits in Cinque's typology. English relative clauses can be marked in a way
identical to complement clauses, with the complementizer that, but they can also
be marked distinctly with relative pronouns like which and who. Under Cinque's
analysis (50), one might predict that relative clauses that are marked with that
should be more acceptable than those marked with relative pronouns. This does
not seem to be the case: the literature on A-bar movement from English relative
clauses is replete with examples where A-bar movement has taken place from a
relative clause marked with who or which, although Cinque reports that there are
native speakers who nd them to be relatively worse. Two examples are given in
(51), taken from Chung and McCloskey, 1983.18
18Jerey Lidz (p.c.) observes that one could weaken Cinque's generalization to accommodate
English: languages that can mark relative clauses and complement clauses in the same way are
those that allow A-bar movement from porous relative clauses. As I see it, the challenge for this
weaker generalization lies in providing a concrete analysis of A-bar movement that explains how, in
a language like English, a syntactic object might move from relative clauses marked with relative
pronouns. The heart of Cinque's insight is that relative pronouns occupy the specier position of
CPs, and therefore block subsequent A-bar movement. In a language like English, one needs to
71
(51) a. That's one trick that I've known a lot of people who've been taken in
by.
b. Then you look at what happens in languages that you know and lan-
guages that you have a friend who knows. (naturally occurring exam-
ple, attributed to Charles Ferguson, May 1971).
A related problem for the generalization, also observed by Cinque, comes
from Romanian, where it is reported to be acceptable to move from relative clauses
marked with the relative pronoun care `who, which' (52).19
(52) Ion, caruia nu este nimeni care poate sa-i reziste, . . .
Ion who.DAT not is nobody who can PRT-him resist
`Ion, who there is nobody who can resist . . . ' (Cinque, 2010, p. 87, fn. 11
ex. ib)
French presents the opposite problem: both relative clauses and complement
clauses are arguably marked (relatively) uniformly, with que and its variant qui, but
patterns more like German in having more restrictions on A-bar movement from
relative clauses.
Finally, Cinque's suggestion (or rather, what I take his suggestion to be) seems
to be theoretically inconsistent with a fact of Bulgarian, namely, it allows multiple
wh-phrases to move from a clause. In examples like (53), the presence of a wh-phrase
on the left periphery of an embedded clause does not seem to pose an obstacle for
assume that relative pronouns do not always have this blocking property, unlike relative pronouns
in languages like German.
19Cinque (2010) argues that the acceptability of (52) might be related to the fact that in Ro-
manian, care can behave more like a complementizer than a relative pronoun. As I understand
the argument, the claim is based on the fact that in examples like (i), care, appears without the
marker pe or case marker, which is not typical of relative pronouns (see Grosu, 1994).
i. A venit la noi un elve?ian, care proiectul lui l-a interesat pe
has come.PTCP to us one Swiss who project.DEF his him-has interested DOM
director . . .
director
Lit. `A Swiss came to us, who his project interested the director' (Cinque, 2010, p. 87, fn.
11, ex. ii)
72
A-bar movement of another phrase from the same clause.
(53) a. Koj1 kude2 misli? [?e t1 e oti?ul t2]?
who where think.2SG that has gone
`Who do you think (that) went where?' (Rudin, 1988, p. 450, p. 6a)
b. Vidjah edna kniga, kojato1 se ?udja [koj znae [koj prodava
saw.1SG a book which wonder.1SG who knows who sells
t1]].
`I saw a book which I wonder who knows who sells.' (ibid. p. 457, ex.
19)
Another problem is more theory-internal. Conventionally, relative clauses have
the following analysis: a wh-element, coindexed with the head noun, moves to the
left edge of the clause. This A-bar movement operation has an important semantic
consequence, as it is necessary for lambda abstraction, in turn necessary for ensuring
that the relative clause is interpreted as a predicate that modies the head noun,
and not as a proposition. The analysis presented in (50a), however, does not have
movement of a wh-element that is coindexed with the head noun. To ensure that
the structure is interpreted as a predicate, one would need additional assumptions,
which may or may not be desirable. For instance, one might assume that the
complementizer is an operator that assigns a predicate interpretation to the clause
(Jerey Lidz, p.c.).
2.3 Learning an English-like grammar using indirect positive evidence
In preceding sections, I argued that learners cannot reliably count on their
linguistic experience to provide direct evidence about whether their language allows
A-bar movement from relative clauses. The logical solution is that learners must
73
rely on some other kind of evidence that is more easily observed in their linguistic
experience, even though what this might be is not immediately apparent to linguists.
More specically, I propose that in languages like English and Mandarin Chi-
nese, the relevant evidence for these learners is A-bar movement from (non-nite)
purposive clauses  rationale and purpose clauses (Faraci, 1974; Jones, 1991; Nis-
senbaum, 2005; Whelpton, 1995, among many others). In Chapter 3, using the
notion of tense dependence, I argue that these adjunct clauses form a natural class
with porous relative clauses, so observing A-bar movement from these clauses would
provide reliable evidence that A-bar movement from porous relative clauses are pos-
sible.
To motivate this proposal, I provide an argument from comparative syntax,
similar to what Cinque (2010) does. A comparative approach is valuable here,
because it is hard to identify the evidence that learners use by just looking within
a single language. After all, there are many linguistic properties that could be
plausibly linked to A-bar movement from porous relative clauses. However, since it
is unlikely that all these properties are present cross-linguistically, we can eliminate
the irrelevant ones by looking across languages. For this reason, after presenting the
case for purposive clauses, I also discuss several alternative hypotheses, evaluating
them from a comparative syntax perspective and also from a learnability perspective.
74
2.3.1 A-bar movement from purposive clauses
The argument that the relevant evidence is A-bar movement from purposive
clauses is based on an observation in the literature that tries to derive island ef-
fects from discourse-based factors, such as Erteschik-Shir, 1973 and Goldberg, 2006.
For example, Goldberg (2006) observe that a number of clausal adjuncts in En-
glish, which she describes as being not backgrounded or not presupposed (p.
145), allow A-bar movement. Adjuncts that she refers to using this label include
porous relative clauses (she calls them presentational relative clauses) and purpo-
sive clauses (54). In this view, purposive clauses and porous relative clauses form a
single semantic and syntactic natural class (although characterizing them as non-
presupposed is not precise enough; see Chapters 4 and 6 for more details), despite
their supercial morphosyntactic dierences.
(54) Purposive clauses
a. Mary went to the store [(in order) to buy chocolate]. [Rationale
clause]
b. What1 did Mary go to the store [(in order) to buy t1?]
c. Mary brought a bottle of wine1 [to give t1 to John]. [Purpose clause]
d. ?Who2 did Mary bring a bottle of wine1 [to give t1 to t2]?
If these two types of adjunct clauses do form a natural class, we then expect
languages like Hebrew and Mandarin Chinese, which allow A-bar movement from
porous relative clauses, to have A-bar movement from purposive clauses. Conversely,
languages like German or Polish, which block A-bar movement from relative clauses,
should lack A-bar movement from purposive clauses. In the next section, I provide
evidence that supports this prediction.
75
Of course, the fact that languages vary in terms of whether A-bar movement
from purposive clauses is acceptable shows that the acceptability of this congu-
ration is also learned; I will suggest that it is learned via direct positive evidence.
This fact also raises questions about why this cross-linguistic variation in purposive
clauses exists and how it arose. I do not have a ready answer for these questions,
except to note that this is at heart a question of why there are dierent languages.
2.3.1.1 A-bar movement from purposive clauses in Hebrew, and Man-
darin Chinese, and Brazilian Portuguese
Hebrew has rationale clauses, optionally marked with kedey in order preced-
ing the non-nite predicate. Like porous relative clauses, rationale clauses allow
A-bar movement from within, with appropriate prosodic stress on the non-nite
verb in the rationale clause (55b). A-bar movement is judged to be most accept-
able when kedey is omitted and prosodic stress is placed on the non-nite predicate
(Omer Preminger, p.c.).
(55) Hebrew (Omer Preminger, p.c.)
a. Dan halax la-xanut [li-knot Sokolad].
Dan walked to.the-store to-buy chocolate
`Dan went to the store to buy chocolate.'
b. Ze ha-Sokolad1 Se-Dan halax la-xanut [li-knot t1].
this the-chocolate that-Dan walked to.the-store to-buy
`This is the chocolate that Dan went to the store to buy.'
Mandarin Chinese also has rationale clauses (56) (Lin and Liao, 2019; Paul,
2008; Tsai, 1995). As shown in (57), it is possible to move a wh-phrase from a
rationale clause (Tsai's purpose clause, Lin and Liao's  lai -purposives).
76
(56) Mandarin Chinese
a. Zhangsan qu Bali [fangwen zhe-ming geshou].
Zhangsan go Paris interview this-CL singer
`Zhangsan went to Paris to interview this singer.'
b. Zhangsan mai-le yi-ping hong-jiu [(lai) quyue jingli].
Zhangsan buy-PRF one-bottle red-wine come please manager
`Zhangsan bought a bottle of red wine to please the manager.'
(57) a. [Zhangsan qu Bali [fangwen t1] de] geshou1 . . .
Zhangsan go Paris interview MOD singer
`The singer that Zhangsan went to Paris to interview . . . '
b. [Zhangsan mai-le yi-ping hong-jiu [(lai) queyue t1] de]
Zhangsan buy-PRF one-bottle red-wine come please MOD
jingli1 . . .
manager
`The manager who Zhangsan bought a bottle of red wine to please . . . '
Similar facts hold for Brazilian Portuguese (Gesoel Mendes and Kayron Bevil?qua,
p.c.):
(58) Brazilian Portuguese
a. Eles foram para New York [pra ver o cantor].
They went to New York for to.see the singer
`They went to New York to see the singer.'
b. O Jo?o foi na loja [comprar o chocolate].
the Jo?o went to.the store to.buy the chocolate
`Jo?o went to the store to buy chocolate.'
(59) a. Esse ? o cantor1 que eles foram para New York [pra ver t1].
this is the singer that they went to New York for to.see
`This is the singer that they went to New York to see.'
b. Esse ? o chocolate1 que o Jo?o foi na loja [comprar t1].
this is the chocolate that the Jo?o went to.the store to.buy
`This is the chocolate that Jo?o went to the store to buy.'
77
2.3.1.2 Restrictions on A-bar movement from purposive clauses
In this section, I discuss parallels between A-bar movement from porous rela-
tive clauses and A-bar movement from purposive clauses, in terms of what elements
may undergo movement. A-bar movement from purposive clauses is considerably
less well-documented, whether in English or in other languages. Other than Truswell
(2007, 2011), I am not aware of major, recently-published work that specically looks
at this topic.
For this discussion, I will mostly restrict the discussion to English. As in the
case of porous relative clauses, objects of verbs and prepositions can move relatively
easily from purposive clauses (60), compared to adjuncts, including wh-adjuncts
(61).
(60) a. This is the client1 that Mary traveled to New York [to meet t1].
b. This is the client1 that Mary traveled to New York [to meet with t1].
(61) a. *It is [with this client]1 that Mary traveled to New York [to meet t1].
b. ??In what hotel1 did Mary travel to New York [to meet the client t1]? (a
logically plausible answer: Mary traveled to New York [to meet the
client at the Waldorf Astoria].)
A further parallel lies in the fact that A-bar movement of subjects of pur-
posive clauses is unacceptable, as shown by (62). However, this parallel might be
supercial, as A-bar movement of subjects in this context is likely to be due to the
complementizer-trace constraint (or whatever deeper principle that is responsible
for this constraint): because purposive clauses are non-nite, subjects appear only
in the presence of the complementizer for. A-bar movement of the subject then
78
creates a conguration in which its trace is adjacent to for. Similar congurations
involving for -marked complement clauses are also unacceptable (63). This anal-
ysis thus contrasts with that suggested for porous relative clauses, in which the
unacceptability of A-bar movement of subjects is due to a principle like the Path
Containment Condition, Relativized Minimality, or Shortest Move. (It is also, of
course, possible that A-bar movement of subjects in purposive clauses is subject to
both the complementizer-trace constraint and principles like the Path Containment
Condition.)
(62) Context: a terrorist group kidnapped Mary. To secure her release, the
government paid a ransom.
a. The government had to pay a ransom [in order for Mary to be released].
b. *{Who1 / [Which victim]1} did the government have to pay a ransom
[(in order) for t1 to be released]?
c. *That is the victim who the government had to pay a ransom [(in order)
for t1 to be released].
(63) a. Mary prefers for John to become the next CEO.
b. *{Who1 / [Which executive]1} did Mary prefer [for t1 to become the
next CEO]?
c. *That is the executive who Mary preferred [for t1 to become the next
CEO].
I further note that A-bar movement from a nite clause within a purposive
clause is unacceptable (64).20 It is less clear to me what the cause of this low
acceptability is. It is certainly possible that the low acceptability is because of
some constraint against A-bar movement from an embedded nite clause within a
purposive clause. However, embedding a nite clause in a purposive clause seems
20One might wonder about nite adjunct clauses that have purposive semantics, such as so
(that) clauses. These clauses are incompatible with A-bar movement, as Truswell (2011) observes.
I discuss them in Chapter 3, section 3.1.3.
79
to require a relatively complex context, which adds to the length of the sentence.21
Consequently, the low acceptability could just be a processing eect that reects
the complex context and/or the overall length of the A-bar dependency.
(64) Sally is one of several heirs to a billionaire's fortune. John, Sally's hus-
band, discovers compromising material about the billionaire, and decides to
blackmail the billionaire, so that she would promise to let Sally inherit a
company.
a. John blackmailed the billionaire [to make her promise [that Sally would
inherit the company]].
b. *{Who1/Which heir1} did John blackmail the billionaire [to make her
promise t1 would inherit the company]?
c. *That's the heir1 who John blackmailed the billionaire [to make her
promise t1 would inherit the company].
d. *{What1/Which company1} did John blackmail the billlionaire [to make
her promise (that) Sally would inherit t1]?
e. *That's the company1 that John blackmailed the billionaire [to make
her promise (that) Sally would inherit t1].
2.3.1.3 Languages that block A-bar movement from relative clauses
block A-bar movement from purposive clauses
In this section, I present evidence from Polish, German, Bulgarian, and French
that is consistent with the prediction that languages that do not allow A-bar move-
ment from any relative clauses should not allow A-bar movement from purposive
clauses. In the examples below, the um-headed clauses in German and ?eby-headed
clauses in Polish are purposive clauses.
21More specically, to introduce a nite clause, one requires a communicative verb (say,
promise) or mental state verb (think). However, these verbs cannot be felicitously used as
the main verb of a purposive clause, which expresses some kind of agentivity or intent; in other
words a sentence like Someone did something to say/think that . . . is unfelicitous. Consequently,
these verbs need to be introduced by an agentive verb, which naturally increases the length of the
sentence. A-bar movement of any element from the nite clause produces a considerably longer
dependency.
80
(65) Polish (Marta Ruda, p.c.)
a. Jan poszed? do sklepu, [?eby kupi? czekolad?].
Jan.NOM went to store COMP buy.INF chocolate.ACC
Jan went to the store to buy chocolate.
(Note: the comma between the ?eby-marked clausal adjunct is an or-
thographic convention and does not entail an intonational break)
b. *To jest ta czekolada, kt?r?1 Jan poszed? do
this is this chocolate.NOM which.ACC Jan.NOM went to
sklepu, [?eby kupi? t1].
store COMP buy.INF
This is the chocolate that Jan went to the store to buy.
(66) (Southern) German
a. Der Hans ist zum Aldi gefahren, [um ein paar ?pfel zu
the Hans is to.the Aldi gone COMP a bunch apple to
kaufen].
buy
`Hans went to Aldi to buy a bunch of apples.'
b. *Was1 ist der Hans zum Aldi gefahren, [um t1 zu kaufen]?
What is the Hans to.the Aldi gone COMP to buy
`What did Hans go to Aldi to buy?' (above examples from Aaron
Doliana, Lara Ehrenhofer, p.c.)
c. Er hat das Land verlassen, [um sie zu retten].
He has the land left COMP her to rescue.
`He left the country in order to rescue her.' (Giusti, 1986 p. 119, ex.
8)
d. *Wen1 hat er das Land verlassen, [um t1 zu retten]?
whom has he the land left COMP to rescue
`Who has he left the country in order to rescue?' (ibid. p. 125, ex. 26)
Similar restrictions are found in French (Truswell, 2011, Postal, 1998).
(67) French
a. Marie est all?e au magasin [(pour) acheter des pommes].
Marie is gone to.the store for to.buy the apples
`Marie went to the store to buy apples.'
b. ??Ce sont les pommes1 que Marie est all?e au magasin [(pour)
this are the apples that Marie is gone to.the store for
acheter t1].
to buy
`Marie went to the store to buy apples.' (both examples and judgments
from Anouk Dieuleveut, p.c.)
81
c. Elle y est all?e en avion [pour confronter le directeur].
She there is gone by plane for to.confront the manager
`She went there by plane to confront the manager.' (Truswell, 2011, p.
197, ex. 32c, adapted from examples in Postal, 1998, p. 76)
d. *Le directeur1 qu' elle y est all?e en avion [pour confronter t1]
the manager that she there is gone by plane for to.confront
. . .
`The manager that she went there by plane to confront . . . (ibid., ex.
32d)
Unlike their counterparts in Polish, German, and French, purpose clauses in
Bulgarian are nite in the sense that they bear agreement morphology. Nevertheless,
Stateva (2005) observes that Bulgarian does not allow A-bar movement from purpose
clauses, even though A-bar movement from (nite) embedded clauses are possible
in general.
(68) Bulgarian
a. Dojdox [(za) da te pitam].
came.1SG in order PRT you ask.1SG
`I came in order to ask you.' (Stateva, 2005, p. 143, ex. 21)
b. Berlin1, vjarvax [?e ?te posteja edin den t1].
Berlin believed.1SG that will visit.1SG some day
`Berlin, I believed that I will visit some day' (ibid., p. 145, ex. 27b)
c. *Berlin1, zaminax na ekskurzija [(za) da posteja t1].
Berlin went.1SG on trip in order PRT visit.1SG
`Berlin, I went on a trip in order to visit.' (ibid. ex. 27a)
d. *Kakvo1 dojde v Berlin [(za) da poseti? t1]?
what came.2SG in Berlin in order PRT visit.2SG
`What did you come to Berlin (in order) to visit?' (ibid., p. 144, ex.
26)
e. *Kakvo1 otiva? v magazina [(za) da kupi? t1]?
what went in store in order PRT buy.2SG
`What did you go to the store to buy?' (Krapova, 2010b, p. 220, fn.
10, ex. 1, citing Pen?ev, 1998)
One might wonder whether there are other restrictions that independently
prevent A-bar movement out of these clauses in these languages. It is however
82
dicult to nd restrictions that cover enough empirical ground within and across
languages.
For instance, it is possible to think of Bulgarian purpose clauses as PPs:
Krapova (2010b) argues that Bulgarian za is a preposition. The ban on A-bar
movement from Bulgarian purpose clauses might follow from a general constraint
against any kind of movement from a PP. But this argument does not extend to
German, for instance; Abels (2003) observes that German allows A-bar movement
from a PP (69), as long as A-bar movement does not result in a stranded preposition.
(69) [?ber welches Thema]1 hast du mich noch mal [PP nach einem Buch t1]
about which topic have you me again after a book
gefragt?
asked
Which topic did you ask me about a book on again?' (Abels, 2003, p. 211,
ex. 284)
As an alternative, let us assume that German um and Bulgarian za are com-
plementizers (pace Giusti, 1986 and Krapova, 2010b, for example), and stipulate
that in these languages, A-bar movement from a purposive clause is blocked by the
presence of an overt complementizer. Such an account fails to make clear predic-
tions about adjunct clauses without um or za, much less predict the fact that A-bar
movement from such adjuncts is also degraded (70).
(70) ??Was1 bist du gekommen [(um) t1 anzusehen]?
what are you come COMP look-at
`What did you come to look at?' (M?ller, 1995, p. 87 ex. 134a, his judg-
ment) [German]
The above restriction is also unique to German and Bulgarian, and it is not
at all clear how it can be extended to other languages. For example, in Polish,
83
?eby-marked clauses allow A-bar movement of objects (71), so it is not possible to
attribute the unacceptability of (65) to the presence of ?eby.
(71) Co1 Maria chce, ?eby Janek kupi? t1?
what Maria wants COMP Janek bought
`What does Maria want Janek to buy?' (Lasnik and Saito, 1984, p. 279 ex.
158a)
2.3.2 Alternative hypotheses about indirect evidence
One might wonder whether there are other kinds of indirect evidence that
learners might be using. In this section, I consider three other potential sources of
indirect evidence involving embedded clauses. I give arguments from comparative
evidence and learning considerations against these alternative accounts.
2.3.2.1 Long-distance A-bar movement from (nite) complement clauses
The rst hypothesis is that the ban on A-bar movement from relative clauses
follows from a more general ban on A-bar movement from all embedded nite clauses.
German and Polish, for example, have been reported to impose restrictions on A-
bar movement from embedded nite clauses in general (for discussion of restrictions
on A-bar movement from complement clauses, see Chapter 4). Northern German
varieties, for instance, are claimed to ban long-distance A-bar movement from dass-
headed clauses (72a); alternative strategies, such as wh-copying, must be used in-
stead (Haider, 2010; Kvam, 1983; Salzmann, 2017, etc.). Likewise, Polish generally
disallow A-bar movement from embedded nite clauses, not even for a verb like
my?le? `think' (72b) (see Witko? (1995) and references therein), although for some
84
speakers, m?wi? say and powiedzie? tell exceptionally allow A-bar movement to
cross them (Cichocki, 1983).
(72) a. %Wen1 denkst du [dass Irina t1 einl?dt]?
who.ACC think you that Irina invites?
`Who do you think that Irina invites?' (Fanselow et al., 2005, p. 48 ex.
16; % = variation in acceptability) [German]
b. *Co1 Janek my?li [?e studenci czytaj? t1]?
what Jon think that students read
`What does John think that the students read?' (Witko?, 1995, p. 229,
ex. 39) [Polish]
There is, however, reason to be cautious about this line of reasoning. Long
distance A-bar movement of the kind in (72a) is reported to be acceptable in south-
ern German varieties (Fanselow et al., 2005; Featherston, 2004; Giusti, 1986; Haider,
2010), and there is some experimental evidence suggesting that the restriction might
be extragrammatical: Featherston (2004) nds no dierence in acceptability ratings
of long-distance A-bar movement by northern German speakers or southern Ger-
man speakers, while Fanselow et al. (2005) nd that exposing northern speakers
to long-distance A-bar movement increases their likelihood of producing the same
structure.
In addition, Bulgarian and French allow long-distance A-bar movement from
complement clauses, unlike (northern) German and Polish. In fact, Bulgarian fa-
mously freely allows multiple wh-phrases to move from an embedded clause, in-
cluding embedded wh-questions (73) (e.g. Rudin, 1988; Stateva, 2005). French
long-distance A-bar movement is well-known for triggering a que/qui alternation in
complementizers (74) (see Sportiche, 2011 for a recent analysis).
(73) Bulgarian
85
a. Berlin1, vjarvax [?e ?te posteja edin den t1].
Berlin believed.1SG that will visit.1SG some day
`Berlin, I believed that I will visit some day' (Stateva, 2005, p. 145, ex.
27b) (=(68b))
b. Koj1 kude2 misli? [?e t1 e oti?ul t2]?
who where think.2SG that has gone
`Who do you think (that) went where?' (Rudin, 1988, p. 450, p. 6a)
c. Vidjah edna kniga, kojato1 se ?udja [koj znae [koj prodava
saw.1SG a book which wonder.1SG who knows who sells
t1]].
`I saw a book which I wonder who knows who sells.' (ibid. p. 457, ex.
19)
(74) French (Sportiche, 2011, p. 83, exx. 2 and 3)
a. [Avec qui]1 tu crois [que Jean danse t1]?
with whom you think that Jean dances
`Who do you think that Jean dances with?'
b. Qui1 tu crois [qui t1 danse avec Jean]?
who you think that dances with Jean
`Who do you think dances with Jean?'
2.3.2.2 How relative clauses are marked
For the sake of comprehensiveness, I note that the marking of relative clauses
and complement clauses in principle could be a source of indirect positive evidence
(Cinque, 2010). However, as mentioned in Section 2.2.4, there are empirical and
conceptual problems with this hypothesis.
2.3.2.3 Raising vs. matching relative clauses
Another possible distinction, briey oated in a footnote by Sichel (2018) (p.
344, fn. 3) in a discussion of the relative clause marking analysis mooted by Cinque
(2010), is that the dierence reduces to a distinction between raising and matching
86
relative clauses. This distinction in part stems from Sichel's analysis of porous
relative clauses in Hebrew. Sichel argues that there are two distinct types of relative
clauses in Hebrew: raising and matching relative clauses. Porous relative clauses
are raising relative clauses, and not matching ones.
Broadly speaking, the distinction between these two analyses of relative clauses
is in the way the head NP of the relative clauses is represented, and consequently
in the overall structural analysis of these complex NPs. (I will defer more detailed
discussion of Sichel's raising analysis of relative clauses to Section 5.2.3.1; the details
do not bear on the discussion in the present section.)
In raising analyses (75a), the relative clause is the complement of the deter-
miner; the head NP is represented inside the relative clause, having moved from the
gap position to the periphery of the relative clause. In matching analyses (75b), the
determiner's complement is an NP, consisting of the head NP, which is represented
outside of the relative clause (for ease of reference, call this an external head), and
the relative clause, which adjoins to the head NP. In the relative clause, a phrase
consisting of an operator and a NP moves from the gap position to the periphery
of the relative clause. Because the head NP matches the NP inside the phrase, it
licenses the ellipsis of the latter NP. In other words, in raising analyses, the repre-
sentation of relative clauses involve only A-bar movement, but in matching analyses,
the representation of relative clauses involve both A-bar movement and ellipsis.
(75) a. Raising relative clauses
87
DP
D CP
the
NP C?1
book that John read t1
b. Matching relative clauses
DP
D NP
the
NP CP
book
DP ?1 C
Op NP
that John read t1
book
Sichel argues using binding evidence that Hebrew porous relative clauses are
raising relative clauses and not matching relative clauses. If this is the case, one way
to capture the cross-linguistic dierence in A-bar movement from relative clauses is
to claim that languages like German (Polish, . . . ) all lack raising relative clauses;
perhaps they are all matching relative clauses. If so, then there is no analysis of
these relative clauses in these languages that is compatible with A-bar movement,
in contrast to Hebrew.
In principle, this hypothesis about porous relative clauses should be easily
testable. To support this hypothesis, one might show that the raising analysis of
relative clauses is generally inappropriate for languages like German, Bulgarian, and
Polish.
At rst glance, it seems that this hypothesis is too strong, though. It has
88
been independently claimed that relative clauses in general require both raising and
matching analyses. For example, standard arguments for raising analyses involve
VP idioms and reconstruction eects: suppose that a NP must be found (or must
not be found) in a particular conguration inside a clause, such as the object of
a VP idiom (76). If this NP can be relativized, then there must be a level of
representation in which the NP is found inside the relative clause. This is a scenario
that straightforwardly follows from a raising analysis, as in (77a). The matching
analysis, on the other hand, posits the existence of an external NP, whose idiomatic
status is unclear. For example, in (77b), there are two tokens of strings. The rst
is found inside the relative clause, and it is reasonable to think that it receives an
idiomatic interpretation. The second one is the external head, outside the relative
clause. It is unclear how this strings is interpreted, since it is not part of an idiom.22
(76) a. The strings Mary pulled got John into college.
b. *The strings got John into college.
(77) a. The [RC strings1 Mary pulled t1] got John into college. [Raising
analysis]
b. The [NP strings [RC strings1 Mary pulled t1] got John into college.
[Matching analysis]
Reconstruction and VP idiom eects can be found in German, Bulgarian,
and Polish relative clauses, consistent with the idea that relative clauses in these
languages can be derived with a raising analysis. If so, then these datapoints pose
a problem for the hypothesis that relative clauses in these languages only have a
22For thoroughness, not all idioms relativize easily, as Jerey Lidz points out (p.c.). For example,
the cat in the idiom let the cat out of the bag cannot be relativized to produce a sentence like The
cat that Mary let out of the bag shocked everyone, even though the cat can be reasonably taken
to be a metaphor for some contextually-salient secret.
89
matching analysis.
(78) a. Der [Wesenszug von sich1]2, den Peter1 noch nicht t2 kannte,
the trait of self which Peter still not knew
st?rte niemanden.
annoyed no.one
`No one was annoyed by the side of himself that Peter did not know
yet.' (Salzmann, 2018, p. 191 ex. 4a) [German]
b. Natiskat1, deto Evropa okazva t1 na Balgarija zaradi OMO
the.pressure that Europe exerts on Bulgaria because.of OMO
Ilinden, e nepravomeren.
Ilinden is illegal.
`The pressure that Europe puts on Bulgaria because of OMO Ilinden
is illegal.' (Krapova, 2010a, p. 1247 ex. 17b) [Bulgarian]
c. S??w1 co on nie rzuca? t1 na wiatr . . .
words that he not throw on wind
`Empty promises that he did not make . . . ' (Szczegielniak, 2004, p.
377 ex. 12b) [Polish]
A further problem for testing this hypothesis is the fact that these eects
can be modeled under either analysis, once certain ancillary assumptions are made;
this means that it is dicult to locate a phenomenon in the recent literature that
might unambiguously adjudicate between either analysis. This is the approach that
Henderson (2007) explicitly takes in his argument for a raising-only analysis (for
English). For example, a traditional argument for the matching analysis involves
the absence of Principle C eects; a raising analysis incorrectly predicts a sentence
like The [picture of John1]2 that you you thought he1 liked t2 is on the mantel (from
Henderson, 2007) to be unacceptable. Henderson argues against this analysis by
assuming that vehicle change is possible under A-bar movement, following Sar
(1999). Salzmann (2018), arguing for a matching-only analysis, obviously does not
adopt such a proposal. Instead, he tries to account for phenomena that motivate the
raising analysis, such as the preservation of an idiomatic reading in relative clauses,
90
with the assumption that unlicensed external heads can exceptionally delete at LF.
It is also interesting to see that the relative clause literature is not concerned
about the cross-linguistic distribution of these relative clauses, i.e. whether one
type of relative clause is found in some languages and another in a dierent set of
languages. Instead, the literature is primarily concerned with a parsimony question:
whether both raising and matching analyses (or a combination of raising, matching
and head external analyses) are necessary cross-linguistically (Bhatt, 2002; Hulsey
and Sauerland, 2006; Sauerland, 1998), or whether one of these analysis is su-
cient for modelling relative clauses cross-linguistically (Henderson, 2007; Salzmann,
2018). Salzmann (2018), who strongly argues for the matching analysis and against
the raising analysis, bases his arguments on with evidence from English and German
relative clauses. If Salzmann's claim that English only has matching relative clauses
is essentially correct, then Sichel's analysis of porous relative clauses can only apply
to Hebrew but not to English, an asymmetry that we might nd curious. Likewise,
assuming again that Sichel's analysis of porous relative clauses is correct, Hender-
son's argument for a raising-only analysis extends beyond English to a language like
German, then one would expect that both Hebrew and German should allow A-bar
movement from relative clauses to the same degree, a prediction that is inconsistent
with existing descriptions.
91
2.4 Conclusion
In this chapter, I reviewed existing observations about cross-linguistic variation
on A-bar movement from relative clauses. I also presented experimental results that
conrm the claim that A-bar movement from nite relative clauses in English shows
an acceptability contrast, complementing previous research by Kush et al. (2013).
Corpus evidence suggests that English and Mandarin Chinese learners have minimal
direct evidence that A-bar movement from relative clauses is possible under excep-
tional circumstances in their languages. I argued that learners of these languages
might learn about this particular phenomenon indirectly, through A-bar movement
from purposive clauses, and showed that cross-linguistically, the acceptability of A-
bar movement from porous relative clauses covaries with the acceptability of A-bar
movement from purposive clauses, consistent with this indirect learning account.
92
Chapter 3: Learning with purposive clauses
The previous chapter described the following learning problem: A number of
languages, like English and Mandarin Chinese, allow A-bar movement from relative
clauses under exceptional circumstances. However, because A-bar movement from
these porous relative clauses is relatively infrequent in these languages, it is unlikely
that learners actually get to observe A-bar movement from any kind of relative
clause. Nevertheless, learners succeed in acquiring a grammar that allows A-bar
movement from porous relative clauses.
The solution proposed at the end of the chapter looked like this: In the case
of English and Mandarin Chinese (and possibly Hebrew, etc.), learners encounter
indirect positive evidence for an acceptability contrast within A-bar movement from
relative clauses. I proposed that the evidence might take the form of acceptable
A-bar movement from purposive clauses, and provided cross-linguistic evidence to
support this account. In contrast, speakers of a language like German, Bulgarian,
etc., do not observe A-bar movement from either purposive clauses or porous rela-
tive clauses, since these are generally unacceptable. Consequently, they do not get
evidence that their languages allow A-bar movement from porous relative clauses
(or purposive clauses).
93
The goal of this chapter is to develop this account further: why might pur-
posive clauses be the relevant evidence, and what reason is there to think that this
is a feasible learning account? I suggest that this is because purposive clauses and
porous relative clauses form a natural class. To make this argument, I show how the
two possible analyses of porous relative clauses presented in the previous chapter
can be extended or adapted to accommodate purposive clauses. If these analyses are
on the right track, then it suggests that among the many properties that purposive
clauses and porous relative clauses have, learners are particularly sensitive to their
tense and temporal properties or height within a main clause.
Of course, showing cross-linguistic covariation and a formal basis for it does
not prove with certainty that learners actually use A-bar movement from purposive
clauses to learn about A-bar movement from porous relative clauses. This account
also requires that A-bar movement from purposive clauses should be present in the
linguistic experience of English and Mandarin Chinese speakers, and absent in lan-
guages like German or Bulgarian. In the second part of the chapter (sections 3.3
and 3.4), I analyze three corpora: the CHILDES Treebank, the Penn Treebank, and
the Chinese Treebank. These datasets collectively represent child-ambient English
speech and the kind of English and Mandarin Chinese speech and texts that older
speakers have access to. Consistent with the learning account, I nd that A-bar
movement from purposive clauses occur naturally in the Penn Treebank and Chi-
nese Treebank, at least at a rate higher than A-bar movement from porous relative
clauses.
94
3.1 How purposive clauses, porous relative clauses might form a natural
class
My goal in this section is to argue that porous relative clauses and purposive
clauses form a natural class within the larger class of clausal adjuncts. Specically,
I argue that both porous relative clauses and purposive clauses show tense depen-
dence. I further argue that learners are sensitive to this particular property, and use
it to draw conclusions about A-bar movement from purposive clauses and porous
relative clauses.
In this section, I describe the two types of purposive clauses (in part following
the classication of Whelpton, 1995) and discuss tense dependence in purposive
clauses, showing how they form a natural class with porous relative clauses. I then
show how the analyses proposed for porous relative clauses in the previous chapter
can be adapted for purposive clauses.
3.1.1 Two types of purposive clauses
Following Faraci, 1974; Jones, 1991; Whelpton, 1995 (also Nissenbaum, 2005),
I will divide purposive clauses into two classes: purpose clauses and rationale clauses.
For comprehensiveness, in this section, I summarize some of the more salient syn-
tactic properties of these two types of clauses.
First, both types of clauses can be distinguished by the (optional) presence
of in order and the gaps that are allowed in these clauses (examples are based on
95
adapting examples and discussion in pp. 56 of Whelpton, 1995).
(1) a. Purpose clauses cannot be marked with in order, rationale clauses can.
(i) Mary brought a book along [(*in order) _ to read _]. [Purpose
clause]
(ii) David spoke Welsh [(in order) for Damian to feel more at home].
[Rationale clause]
b. If there is a gap in a rationale clause, it must be a subject gap.
(i) David1 spoke in Welsh [_1 to keep the gossip secret]. [Rationale
clause]
(ii) *David1 spoke in Welsh [in order for Damian to understand _-
1]. (unacceptable on the reading . . . for Damian to understand
David, assuming that Damian only understands Welsh.)
c. The theme argument of the main clause controls a gap (or the gap, if
there is only one) in the purpose clause, while the agent argument of
the main clause controls the subject gap (if there is one) in a rationale
clause.
(i) Mary brought a book1 along [for John to read _1].
(ii) David1 spoke Welsh to Damian2 [_{1/?2} to impress the visitors].
Structurally speaking, both types of clauses adjoin low in a clause, specically,
at a VP level (or lower). The key evidence comes from VP preposing: both types
of clauses can be preposed with a VP (2) (Whelpton, 1995, pp. 100 and 109).
Acceptability here is also sensitive to phonological weight: the more material there
is in a preposed VP, the less acceptable preposing becomes.
(2) a. Mary said she would bring a book along [to read], and [V P bring a book
along [to read]], she did.
b. David said he would speak in Welsh [(in order) to keep the gossip secret],
and [V P speak in Welsh [(in order) to keep the gossip secret]], he did.
A number of diagnostics also show that purpose clauses adjoin to a lower
position than rationale clauses do. To give one example, when a VP is preposed, a
purpose clause must be preposed along with the VP, while rationale clauses can be
stranded (3).1
1For other diagnostics, such as do so anaphora and linear precedence, see Whelpton, 1995, pp.
96
(3) a. Mary said she would bring a book along [to read], and *[V P bring a book
along], she did [to read].
b. David said he would speak in Welsh [(in order) to keep the gossip secret],
and [V P speak in Welsh], he did [(in order) to keep the gossip secret].
3.1.2 Tense dependence
In the previous chapter, I argued that porous relative clauses can be distin-
guished from other relative clauses in terms of their tense and temporal interpreta-
tion properties. More specically, porous relative clauses display what I call tense
dependence: their tense is semantically dependent on the tense of the matrix clause.
This dependence manifests itself either semantically, in that tense in porous relative
clauses lacks a later-than-matrix reading, and/or morphosyntactically, in the form
of tense harmony.
In this section, I point out that purposive clauses also exhibit tense depen-
dence. If so, one reason why learners would use purposive clauses to draw infer-
ences about A-bar movement and porous relative clauses is because of the tense
dependence property that is common to both purposive clauses and porous relative
clauses. In other words, even though purposive clauses and porous relative clauses
are morphosyntactically dierent, the tense dependence property is distinct enough
for learners to treat them as members of a natural class.
Temporally, the main event denoted by the predicate of a purposive clause can
only be located in the future of the main clause predicate in some possible world,
and not the other way around. In (4), the buying of chocolate or giving of wine
99100, 109, 121.
97
comes after the act of going to the store or bringing a bottle of wine.
(4) a. Mary went to the store [to buy chocolate].
b. Mary brought a bottle of wine along [to give to John].
One cannot use the purposive clause to describe a situation that temporally
preceded and is causally related to the main clause predicate (5), even though
broadly speaking, purposive clauses describe a causal relation between the main
clause predicate and the adjunct predicate. This temporal restriction is, of course,
closely related to the causal / teleological semantics of the construction, in which
the adjunct describes the goal or outcome that is intended to follow from the action
denoted by the predicate in the main clause.
(5) a. #Mary went to the store [to want to buy chocolate]. (Scenario: Mary
wanted to buy chocolate, and then went to the store to do so.)
b. #Mary brought a bottle of wine along [to be invited by John to the party].
(Scenario: Mary was invited by John to the party, and consequently
brought a bottle of a wine along.)
There is also a morphosyntactic dimension to the tense dependence of purpo-
sive clauses. A key characteristic of purposive clauses is non-nite tense. Generally,
non-nite tense is found in embedded contexts, but not main clauses.2 The general
absence of non-nite tense in main clause contexts is arguably consistent with the
notion that non-nite tense is dependent, requiring the presence of a structurally
higher tense morpheme. Because of this requirement, these tenses cannot appear in
main clause contexts, where there is, by denition, no higher tense morpheme.
2Exceptions in English include non-nite exclamatives (To think that they would do such a
thing! ) or certain non-nite wh-questions (How to write well? )
98
3.1.3 So (that), because, and only to clauses
I next discuss a number of dierent nite clausal adjuncts that appear to show
some kind of tense dependence and yet are incompatible with A-bar movement of
arguments. To preview the discussion, I argue that these do not necessarily challenge
the generalization that relates tense dependence and A-bar movement; these nite
adjuncts arguably have independent tenses; the tense dependence observed is instead
a semantic artifact of these constructions.
The rst type of adjunct clauses are so (that) clauses, which are semantically
very similar to purposive clauses (6). As Truswell (2007, 2011) points out, these
nite adjunct clauses do not allow A-bar movement, not even of NP objects.
(6) a. John1 left early [so (that) he1 could catch a ight]. (? John left early
to catch a ight.)
b. *What2/Which ight2 did John1 leave early [so (that) he1 could catch
t2]?
c. ?*That's the ight2 that John1 left early [so (that) he1 could catch t2].
In so (that) constructions, the tense must morphologically match that of the
main clause.
(7) a. John left early [so that he {could/??can} catch a ight].
b. John leaves at six tomorrow morning [so that he {??could/can} catch a
ight].
The second type of adjunct clauses are because clauses, which are also seman-
tically very similar to purposive clauses.
(8) a. John1 left early [because he1 had to catch a ight]. (? John left early
to catch a ight.)
b. *What2/Which ight2 did John1 leave early [because he1 had to catch
99
t2]?
c. ?*That's the ight2 that John1 left early [because he1 had to catch t2].
For because constructions, a present tense can appear in the because clause
when the main clause is in the past tense; this conguration produces a double
access reading (Ogihara, 1996, chapter 6): for (9a) to be felicitous, the obligation
for John to catch the ight must be true at the time of John's leaving and at the
time (9a) was uttered. In contrast, past tense in the adjunct clause cannot co-occur
with a present tense in the main clause.
(9) a. John left early [because he {had/has} to catch a ight].
b. John leaves early [because he {*had/has} to catch a ight].
The third type of clauses are only to telic adjuncts (Whelpton, 1995). Only
to adjuncts are non-nite  argued to be a hallmark of defective tense  and the
time of the embedded predicate is located in the future of the main clause predicate.
(10) a. John left early [only to run into Mary].
b. *Who1 did John leave [only to run into t1]?
c. *That's the person1 that John left [only to run into t1].
For these three types of clauses, I would like to suggest that these do not nec-
essarily counter-exemplify the generalization about tense dependence. More specif-
ically, I would like to suggest that there is no inherent dependence between the
embedded and main clause tenses; the tenses here are independent. The tense de-
pendence we observe is instead an artifact of the semantics of the constructions.3
For example, so that clauses describe outcomes intended at the time of the main
clause predicate. When one says that John left early so that he could catch a ight,
3Special thanks to Valentine Hacquard for discussion on these issues.
100
at the point of John's leaving (say, time t), he also had the intent (at time t) of
being able to catch a ight. As a result, the tense of the adjunct has to match
that of the main clause. Similarly, because clauses describe the cause(s) behind the
action denoted by the main clause predicate. When one says John left early because
he had to catch a ight, the cause of John's leaving early at time t was his desires
(at time t) to catch a ight. Telic clauses, on the other hand, appear to describe
unintended actual outcomes; John left early only to run into Mary is felicitous only
if the reason of John's leaving early was to avoid Mary and he actually ran into
Mary. Since outcomes are by denition temporally located in the relative future,
there has to be a temporal dependency between the embedded predicate and the
main clause predicate.
In the same vein, one might wish to argue that the future reading of purposive
clauses can be attributed to the construction. Put dierently, in this alternative
analysis, purposive clauses have independent tense; the tense dependence observed
for purposive clauses also has a semantic or pragmatic origin, and not a syntactic
one. This is, of course, not a desirable analysis in this context, and it strikes me that
the dierence between purposive clause and the adjunct clauses described in this
section might ultimately have a structural explanation: there is a case to be made
that purposive clauses adjoin to a position inside a VP, while so (that), because,
and only to adjuncts arguably adjoin to a higher position outside a VP; see Section
3.1.5.1 for evidence.
101
3.1.4 Other issues
3.1.4.1 Temporal adjunct clauses
Temporal adjunct clauses, such as those headed by before and after, also ap-
pear to pose a problem for the tense dependence generalization. In English, when
the clause in the temporal adjunct is simplex (without embedded clauses), the tense
in the temporal adjunct must match that of the main clause (11), a phenomenon
rst described by Geis (1970) with the label tense harmony.4 The matching thus
suggests a dependency and raises the possibility that these temporal adjuncts have
defective tense. However, these adjuncts are incompatible with overt A-bar move-
ment of arguments (12).5
(11) a. Mary leaves [after/before John *called/calls Sam].
b. Mary left [after/before John called/*calls Sam].
(12) a. ??Who did Mary leave [before/after John called t1]?
b. ??That's the person Mary left [before/after John called t1].
As I see it, there are three theoretical options for dealing with this problem. All
three are based on independent observations in the literature on temporal adjunct
clauses and temporal adjuncts in general.
The rst is to argue that the tenses in temporal adjuncts are actually inde-
4This sense of tense harmony should be kept distinct from the sense of tense harmony found
in the literature on specicational copular clauses. For more discussion about the tense harmony
rule and how it aects the interpretation of temporal adjuncts, I refer readers to Geis, 1970;
Kusumoto, 2017.
5But see Truswell (2007, 2011) for extensive discussion of A-bar movement from temporal
adjunct clauses whose complements are bare present participials, e.g. %What did John read the
manual before using t1?
102
pendent, and the restrictions exemplied in (11) are derived from semantic and
pragmatic considerations. There is thus no tense dependence in these adjuncts.
Here, I present an argument developed by Kusumoto (2017). Making the standard
assumption that the present tense in English has an indexical component, Kusumoto
argues, quite reasonably, that sentences like (13) are ruled out as a matter of contra-
diction. In (13a), Mary's leaving is in the past, while the phone call between John
and Sam is not, but it is also asserted that the phone call precedes Mary's leaving.
Following Stump (1985), she argues that sentences like (14) are unacceptable be-
cause they are semantically redundant (p. 204). In (14a), Mary's leaving is in the
past, and the phone call is not. Even without the temporal adjunct, it is already
clear that Mary's leaving precedes the phone call.
(13) a. *Mary left after John calls Sam.
b. *Mary leaves (tomorrow) before John called Sam.
(14) a. #Mary left before John calls Sam.
b. #Mary leaves (tomorrow) after John called Sam.
The second option is to argue that A-bar movement from temporal adjuncts
actually involves A-bar movement from an unpronounced denite description that
refers to a time interval that is presupposed to exist; such movement is known to be
unacceptable. To see this, consider the examples in (15), which are close paraphrases
of each other. What distinguishes (15a) from (15b) and (15c) is the complements of
before: in (15a), the complement is a clause, while in the other two examples, the
complement is a nominal, one that refers to a specic time interval (2 o'clock in the
afternoon or John's phone call time).
103
(15) Context: John called Sam at 2 p.m. Mary left at 1 pm.
a. Mary left before John called Sam.
b. Mary left before 2 p.m.
c. Mary left before the time of John's calling Sam.
To capture the semantic similarities, some have proposed that there is a silent
operator in the before clause in (15a). This operator takes the clause John called
Sam as an argument and returns a denite description, the earliest time interval of
this event denoted by the clause (for formal implementation of this idea, see Beaver
and Condoravdi, 2003, Sharvit, 2013, among others). An illustration of this general
idea is shown in (16), where the operator is notated as THE-EARLIEST; note that
this analysis abstracts away from the fact that before clauses are not veridical, in
that proposition described in a before clause is not necessarily true (see Beaver and
Condoravdi, 2003 for comments on how time and veridicality/modality might be
encoded on the operator).
(16) . . . before [THE-EARLIEST [John called Sam]].
Seen in this light, A-bar movement from these temporal adjuncts becomes a
case of A-bar movement from a denite description. Further, if I understand Beaver
and Condoravdi's analysis correctly, this description picks out a particular time
interval that exists in the timeline. In this regard, this denite description resembles
the one in (17a), which picks out a particular photograph. Since A-bar movement
from denite descriptions is known to be unacceptable ((17b), (17c)) (Chomsky,
1973; Davies and Dubinsky, 2003; Fiengo and Higginbotham, 1981; Gu?ron, 1980;
Huang, 2018a; Simonenko, 2015), it is not surprising that (18) is also unacceptable.
104
(17) a. Mary saw [this photo of John].
b. ??Who did Mary see this photo of t1?
c. ??That's the person who Mary saw this photo of t1.
(18) ??That's the person Mary left [before/after John called t1]. (=(12b))
The third option for ruling out A-bar movement from before clauses is an older,
but related, idea that a before clause contains A-bar movement of a phonologically
null time argument to form a relative clause, as schematized in (19), where capital-
ized elements are phonologically null (Geis, 1970; Larson, 1990; von Stechow and
Gr?nn, 2013b; also see Sharvit, 2013 paper for a critique).
(19) Mary left before [TIME1 [John called Sam t1]].
There are two arguments for this analysis. The rst is the observation that
there is a close paraphrase of sentences like (15a) that involves relative clauses (20),
and hence A-bar movement.
(20) Mary left before the time [at which John called Sam].
Combining the denite description and relative clause analyses yields an anal-
ysis of temporal adjuncts where a preposition like before or after takes a denite
NP that has a presuppositional reading and is modied by the clause. A-bar move-
ment from the before clause in this analysis becomes a case of A-bar movement
from relative clause inside a denite complex NP from another adjunct, which is
unacceptable (21). Seen in this light, it is unsurprising why A-bar movement from
before (or after) clauses is unacceptable.
(21) a. *Who1 did Mary leave before the time at which John called t1?
b. *That's the person1 that Mary left before the time at which John called
105
t1.
The relative clause/A-bar movement analysis receives further support from an
ambiguity that arises when before appears with a clause that has an attitude verb
with a complement clause. The ambiguity can be modeled as A-bar movement of
either a time adverbial in the higher clause or in the embedded clause.
(22) Joan left before Harry told her to leave. (Geis, 1970, p. 127, ex. 42)
a. Reading 1: Upstairs reading
(i) Joan left at 10 a.m. Harry was going to talk to her at 11 a.m. to
tell her to leave at 3 p.m.
(ii) . . . before [TIME1 Harry told her t1 [to leave]]
b. Reading 2: Downstairs reading
(i) Joan left at 10 a.m. Harry told her yesterday to tell her to leave
at 11 a.m., but she left early.
(ii) . . . before [TIME1 Harry told her [to leave t1]]
As evidence for a A-bar movement analysis, consider the fact that the down-
stairs reading disappears when the embedded clause is found in an island congu-
ration.
(23) Joan left before Harry told her of his desire for her to leave.
Only reading: Joan left at time t1 before Harry told her at time t1 that he
wanted her to leave. (ibid. p. 129, ex. 54)
3.1.4.2 Mandarin Chinese wei(le) adjuncts
As mentioned previously, Mandarin Chinese also has purposive clauses that
allow A-bar movement; these purposive clauses appear in a position after the main
VP (24) (= (57) in the previous chapter).
(24) a. [Zhangsan qu Bali [fangwen t1] de] geshou1 . . .
Zhangsan go Paris interview MOD singer
106
`The singer that Zhangsan went to Paris to interview . . . '
b. [Zhangsan mai-le yi-ping hong-jiu [(lai) queyue t1] de]
Zhangsan buy-PRF one-bottle red-wine come please MOD
jingli1 . . .
manager
`The manager who Zhangsan bought a bottle of red wine to please . . . '
However, Mandarin Chinese also has another purposive clause construction
where the adjunct clause is marked by wei(le) and precedes the main verb (25);
A-bar movement from this clause is not possible (26).
(25) a. Zhangsan [weile fangwen zhe-ming geshou] qu-le Bali.
Zhangsan WEILE interview this-CL singer go-PRF Paris.
`Zhangsan went to Paris to interview this singer.'
b. Zhangsan [weile shuaxin zhe-xiang jilu] fushi jin-yao.
Zhangsan WEILE break this-CL record consume banned-drug
`Zhangsan consumed performance-enhancing drugs in order to break
this record.'
(26) a. *?[Zhangsan [weile fangwen t1] qu-le Bali de] geshou1 . . .
Zhangsan WEILE interview go-PRF Paris MOD singer
`The singer that Zhangsan went to Paris to interview . . . '
b. *?[Zhangsan [weile shuaxin t1] fushi jin-yao de]
Zhangsan WEILE break consume banned-drugs MOD
zhe-xiang jilu1 . . .
this-CL record
`This record that Zhangsan consumed performance-enhancing drugs to
break . . . '
This dierence cannot be easily attributed to a dierence in how tense is (ab-
stractly) represented; in modal and temporal terms, there is no dierence between
wei(le) adjuncts and post-VP purposive clauses. Nor is there clear evidence that
wei(le) clauses are structurally dierent than post-VP purposive clauses, beyond
how they are lineally ordered with respect to the VP; for instance, both types of
clauses can be elided as part of a VP (27).
107
(27) a. Lisi hui [weile canjia bisai] qu Bali, Zhangsan ye hui.
Lisi will WEILE participate competition go Paris Zhangsan also will
`Lisi will go to Paris to participate in the competition. Zhangsan will
(go to Paris to participate in the competition), too.'
b. Lisi hui qu Bali [canjia bisai], Zhangsan ye hui.
Lisi will go Paris participate competition Zhangsan also will
`Lisi will go to Paris to participate in the competition. Zhangsan will
(go to Paris to participate in the competition, too), too.'
I would like to suggest that the low acceptability of (26) might follow from
independent properties of Mandarin Chinese, on the assumption that wei(le) is a
preposition. Mandarin Chinese PPs generally do not allow A-bar movement from
within, as shown in (28) and (29). Evidence for the prepositional nature of wei(le)
comes from the fact that it can take a nominal complement (28) and appears in a
pre-verbal position typical of Mandarin Chinese PPs.
(28) a. PPs with nominal complements
(i) Zhangsan [PP weile Lisi] cizhi.
Zhangsan for Lisi quit
`Zhangsan resigned because of Lisi.'
(ii) Zhangsan [PP dui Lisi] hen manyi.
Zhangsan about Lisi very satised
`Zhangsan is satised with Lisi.'
(iii) Zhangsan hui [PP gen Lisi] qu Deguo.
Zhangsan will with Lisi go Germany
`Zhangsan will go to Germany with Lisi.'
b. No preposition stranding
(i) *Lisi1, Zhangsan [weile t1] cizhi.
Lisi Zhangsan for quit
`Lisi, Zhangsan resigned because of him.'
(ii) *Lisi1, Zhangsan [dui t1] hen manyi.
(iii) Lisi1, Zhangsan hui [gen t1] quo Deguo.
(29) a. PPs with clausal complements
(i) Zhangsan [PP dui Lisi bu hui shuo deyu] hen
Zhangsan about Lisi NEG can speak German very
zaiyi.
be.bothered
`Zhangsan is very bothered that Lisi cannot speak German.'
108
(ii) Zhe-jian shi [PP gen Lisi bu hui shuo deyu]
this-CL matter with Lisi NEG can speak German
wanquan mei-you guanxi.
completely NEG-exist relation
`This matter has nothing to do with Lisi's not being able to speak
German.'
b. No A-bar movement from clausal complements of prepositions
(i) *[Zhangsan [PP dui Lisi bu hui shuo t1] hen zaiyi
Zhangsan about Lisi NEG can speak very be.bothered
de] yuyan1 . . .
MOD language
`The language that Zhangsan is very bothered that Lisi cannot
speak it . . . '
(ii) *[Zhe-jian shi [PP gen Lisi bu hui shuo t1] wanquan
this-CL matter with Lisi NEG can speak completely
mei-you guanxi de] yuyan1 . . .
NEG-exist relation MOD language
`The language that this matter has nothing to do with Lisi's not
being able to speak it . . . '
For transparency, I note that Lin and Liao (2019) analyze a lai -marked ratio-
nale clause not as an adjunct, but (somewhat unconventionally) as the complement
of the main verb, with a secondary predicate function. They argue against an ad-
junct analysis by pointing out that Chinese generally blocks right adjunction. In
their analysis, the object is assumed to be generated in a specier position, and the
verb raises to a position that precedes the object and the rationale clause. However,
their complement analysis has a number of weaknesses. First, Mandarin duration
and frequency phrases appear on the right of a VP. If these phrases are adjuncts,
they challenge the generalization that Chinese lacks right adjunction (a point that
Lin and Liao concede in their footnote 4). Second, it is dicult to evaluate their
proposal in the absence of more clearly articulated assumptions about the syntax of
secondary predication and how secondary predication is syntactically distinct from
109
adjunction. In fact, Lin and Liao's proposal seems to imply that secondary predi-
cates appear post-verbally, but this is not the case for all secondary predicates in
Mandarin (Zhang, 2001). Third, this analysis is silent on why these clauses cannot
contain a wh-adverb like weishenme why (30a), especially in light of the fact there
are clausal complements in Mandarin that can (30b). In contrast, the low accept-
ability of (30a) can be captured more easily under an adjunct analysis, since it is
known that wh-adverbs like weishenme cannot appear in adjuncts (and islands) in
Mandarin Chinese (30c) (see J.-W. Lin, 1992).6
(30) a. *Zhangsan mai shu [lai weishenme quyue Lisi]?
Zhangsan buy book come why please Lisi
`What is the reason such that Zhangsan bought books to please Lisi
for that reason?' (Plausible answer: Because a happy Lisi would help
Zhangsan get promoted.)
b. Zhangsan shuo [Lisi weishenme cizhi]?
Zhangsan say Lisi why resign
`Why1 did Zhangsan say [that Lisi resigned t1]?'
c. *Zhangsan [zai jingli weishenme kaichu Lisi zhihou] cizhi?
Zhangsan at manager how dismiss Lisi after resign
`What is the reason such that Zhangsan resigned after the manager
red Lisi for that reason?'
3.1.5 Extending the analysis for porous relative clauses to purposive
clauses
Given that the argument that purposive clauses also exhibit tense dependence,
it is interesting to also consider the question of how tense dependence is represented
in purposive clauses, and how this representation explains why A-bar movement
6Of course, there could well be some independent factor that explains why weishenme is unac-
ceptable in a clausal complement, but what this factor might be is not addressed in Lin and Liao's
proposal.
110
from purposive clauses is possible. In the previous chapter, I suggested two ways of
modeling tense dependence and A-bar movement in porous relative clauses. In this
section, I discuss ways to adapt these analyses to also account for A-bar movement
from purposive clauses.
3.1.5.1 Adapting a movement-based analysis
The rst analysis of porous relative clauses presented in the previous chapter
claims that they are porous because they do not undergo any movement and there-
fore do not freeze (echoing Sichel, 2018); this movement-based analysis accounts for
some of the interpretations associated with NPs modied by porous relative clauses
and can explain a number of restrictions on tense and temporal interpretations in
porous relative clauses.
Extending this approach to other adjuncts poses a problem. The null hypoth-
esis is that adjuncts, in general, do not move, except for e.g. stylistic reasons when
they are fronted to a clause-initial position. Consequently, all else being equal, any
(clausal) adjunct that remains in situ should be compatible with A-bar movement.
This prediction is too strong, as it allows because, so (that) clauses (among others)
to be compatible with A-bar movement.
In order to extend this movement-based analysis to purposive clauses but not
other adjuncts, other assumptions are necessary. As I see it, a relatively simple
option is to reframe this movement-based analysis as a height-based one (echoing a
part of the proposal in Truswell, 2007, 2011). Specically, one might hypothesize
111
that only clausal adjuncts that remain within a VP at LF are compatible with A-bar
movement. This formulation includes porous relative clauses, on the assumption
that these relative clauses modify VP-internal NPs that remain in situ. It also
captures purposive clauses, which adjoin to positions inside a VP (see VP preposing
data in Section 3.1.1), while excluding clausal adjuncts like so (that) and because
clauses. The VP preposing diagnostic (31) shows that both so (that) and because
clauses do not prepose as easily as a (non-nite) purposive clause; in (31c), I changed
the object from a ight to a seven a.m. ight to make the VP heavier and thus more
comparable to the examples with so that and because. The same can be said for only
to telic adjuncts, which attach to a relatively high position in a clause (Whelpton,
1995) (32).
(31) a. ??Mary thought that John should leave early [so that he could catch a
ight], and [leave early [so that he could catch a ight]] he did.
b. *Mary thought that John should leave early [because he had to catch a
ight], and [leave early [because he had to catch a ight]] he did.
c. Mary thought that John should leave early [(in order) to catch a seven
a.m. ight], and [leave early [(in order) to catch a seven a.m. ight]]
he did.
(32) ??Kate predicted that John would leave [only to run into Mary], and [leave
[only to run into Mary]] he did.
Strictly speaking, what the VP preposing data show is that purposive clauses
are found in a VP-internal position in overt syntax. This alone does not mean
that purposive clauses remain in this VP-internal position at LF. One thus requires
the additional assumption that purposive clauses do not move further at LF, which
strikes me as a fairly reasonable assumption, since there is no independent evidence
112
to think that they do.
This height-based reframing overgenerates, predicting that all VP-internal ad-
juncts are compatible with A-bar movement. As Truswell (2007, 2011) observes,
this prediction is only partially borne out. It is thus necessary to introduce other
constraints on A-bar movement from adjuncts to prevent overgeneration.
For English at least, A-bar movement can take place from structurally-low PP
adjuncts, resulting in preposition stranding (33), but there are notable exceptions.
For example, stranding is less acceptable for temporal prepositions like before or
during and for latinate ones like via (cf. from, which is similar semantically) (34).
Cross-linguistically, preposition stranding is rare, occurring robustly only in a small
number of Germanic languages. Additional constraints are thus needed to explain
the limited availability of P-stranding under A-bar movement.
(33) John said he would observe the leaf [with a microscope], and [observe the
leaf [with a microscope]] he did.
a. What1 did John observe the leaf [with t1]?
b. That's the microscope1 that John observed the leaf [with t1].
(34) a. John said he would leave [before 3 pm], and [leave [before 3 pm]] he
did.
(i) ??What time1 did John leave [before t1]?
(ii) *That's the time of the day1 that John left [before t1].
b. Mary said she would order the hat [via the website], and [order the hat
[via the website]] she did.
(i) *Which website1 did Mary order the hat [via t1]?
(ii) *That's the website1 that Mary ordered the hat [via t1].
As for gerundival adjuncts, which are also structurally low, Truswell (2007,
2011) argues that only those bear a particular event semantic relation with the
113
main VP may do so, as set out by his Single Event Grouping Condition7 (35).
(35) a. Mary said John would drive crazy [whistling Happy Birthday ], and
[drive her crazy [whistling Happy Birthday ]] he did.
(i) [What song]1 did John drive Mary crazy [whistling t1]?
(ii) That's the song1 that John drove Mary crazy [whistling t1].
b. John said he would arrive [wearing a suit], and [arrive [wearing a suit]]
he did.
(i) [What kind of clothing]1 did John arrive [wearing t1]?
(ii) That's the suit1 that John arrive [wearing t1].
c. John said he would work on his homework [whistling Happy Birthday ],
and [work on his homework [whistling Happy Birthday ]] he did.
(i) *[What song]1 did John work on his homework [whistling t1]?
(ii) *That's the song1 John worked on his homework [whistling t1].
(adapted from various examples in Truswell, 2011)
For thoroughness, temporal adjuncts are also VP-internal (36). but as men-
tioned in Section 3.1.4.1, there might be independent reasons why A-bar movement
from temporal adjuncts is unacceptable.8
(36) Mary said she would leave [{before/after} John called Sam], and [leave
[{before/after} John called Sam]] she did.
a. ??Who1 did Mary leave [before John called t1]?
b. *That's the person1 that Mary left [before John called t1].
Setting aside these instances of overgeneration, from a cross-linguistic varia-
tion perspective, this height-based reframing means that languages vary in terms of
whether A-bar movement is possible from clausal adjuncts that are found within a
VP at LF. While this analysis allows us to accommodate purposive clauses and a
7Truswell intends for the Single Event Grouping Condition to cover A-bar movement from pur-
posive clauses. It is logically possible, although perhaps a little redundant, that A-bar movement
from purposive clauses can be accounted for by both tense dependence (or more precisely, a height-
based analysis or defective tense analysis that derives tense dependence) and by the Single Event
Grouping Condition.
8But see Truswell, 2007, 2011 for discussion on relatively acceptable A-bar movement from
temporal adjuncts that have gerundival complements, like Which book did John design his garden
after reading?. Truswell however acknowledges that not all speakers accept such sentences, but
does not address this fact in detail.
114
range of other adjuncts, it leaves us with a somewhat unsatisfying account of why
regular relative clauses and certain adjuncts are incompatible with A-bar movement.
In the original movement-based analysis, the fact that regular relative clauses are
incompatible with A-bar movement is attributed to the idea that the NPs they
modify move and then freeze. The height-based reframing does away with freezing
eects, since freezing does not explain why many adjuncts that adjoin to VP are
incompatible with A-bar movement. Instead, it replaces freezing with a stipulation:
VP-internal adjuncts are compatible with A-bar movement; VP-external ones are
not.
In addition, this height-based analysis still predicts that A-bar movement is
not possible from adjuncts outside of a VP; this prediction is inconsistent with the
Mandarin Chinese examples where A-bar movement is observed with porous relative
clauses modifying a subject (37).
(37) [Zhe-ben shu]1 wo renwei [Subject [du-guo t1 de] ren] bu duo.
this-CL book I think read-EXP MOD person not many
`This book, I think there are not many people who have read it.' (Mandarin
Chinese; Xu and Langendoen, 1985 p. 14 ex. 61a) (= (41) in Chapter 2)
3.1.5.2 A defective tense analysis
A defective tense analysis avoids some of the stipulation and undergeneration
issues associated with the movement/height-based analysis. In this alternative anal-
ysis, tense dependence is the semantic reex of defective tense morphemes. At the
same time, this defectiveness neutralizes the locality domain (a phase that corre-
sponds to the clause), thus allowing A-bar movement from porous relative clauses
115
without incurring a violation of the Phase Impenetrability Condition.
This analysis can be easily extended to accommodate purposive clauses. As
mentioned above, purposive clauses are clauses with non-nite tense, which we might
take to be a morphological realization of defectiveness. Further, in purposive clauses,
the interpretation of this tense morpheme is necessarily in the future of the time
interval denoted by the tense in the main clause. From a cross-linguistic variation
perspective, what this analysis means is that languages vary in terms of whether
defective tense can neutralize phases or not.
As presented, this particular analysis is silent about the well-formedness of A-
bar movement from other non-clausal adjuncts, such as PPs and present participial
adjuncts. To make a clear prediction, one needs to take a position as to whether these
adjuncts are phases. If they are, one needs to further take a position on whether
they contain escape hatches on the left edge of the phase for A-bar movement.
Further, in the context of present participial adjuncts, because these adjuncts do
not contain a non-nite to and are not inected for tense, one must also take a
position on whether they contain (silent) tense morphemes.
The lack of clear predictions is perhaps not an undesirable empirical outcome,
because these adjuncts do not behave uniformly with respect to A-bar movement.
As noted above, it is not unreasonable to think that there are other conditions on
A-bar movement from these adjuncts.
116
3.2 Alternative characterizations of purposive clauses and porous rel-
ative clauses
In this section, I consider other treatments of purposive clauses and porous
relative clauses. I consider three distinct approaches: a discourse-based one, cen-
tered on the notion of backgroundedness, a semantic one, based on the notion of
veridicality, and a morphosyntactic one, based on mood. Generally, there are two
ways in which these characterizations are less ideal than the tense dependence one
put forward in the previous chapter and the above sections. First, as described
in greater detail below, these approaches have empirical shortcomings; they either
undergenerate, predicting that certain porous relative clauses are incompatible with
A-bar movement, or overgenerate, incorrectly predicting acceptable A-bar move-
ment from certain regular relative clauses. Second, it is not always clear how to tie
these characterizations to A-bar movement, short of introducing a stipulation like
no A-bar movement from backgrounded/veridical adjuncts. To the extent that
the tense dependence approach yields less stipulative analyses of A-bar movement,
it is more desirable.
3.2.1 Backgroundedness
There is a tradition of framing constraints on A-bar movement in discourse
terms. For example, structures (including adjuncts) that are incompatible with A-
bar movement are often described as being backgrounded (Goldberg, 2006) or not
117
dominant (Erteschik-Shir, 1973).
For the purpose of discussion, I will use as a case study the analysis of A-
bar movement from porous relative clauses and purposive clauses put forward by
Goldberg (2006), as this is the most recent worked out account in this literature, as
far as I am aware of. Goldberg suggests that both kinds of adjuncts form a natural
class in discourse terms: they are non-backgrounded, unlike other adjuncts.
One limitation of Goldberg's approach (and often of discourse-based approaches
in general) is that the notion of backgroundedness is dicult to make precise, be-
cause it partly requires one to reliably identify what a speaker intends to assert
when he/she produces a particular utterance. For example, Goldberg considers a
constituent to be backgrounded if it correspond[s] neither to the primary topic
nor to part of the potential focus domain [of the sentence], such that a topic serves
to contextualize other elements in the clause and a potential focus domain is that
part of a sentence that is interpretable as being asserted (p. 130; see references
therein for more detail).
Goldberg also oers a diagnostic for backgroundedness: a negation test, vari-
ants of which have been previously proposed in the literature, starting with Erteschik-
Shir, 1973. In this particular version, an embedded clause is backgrounded if sen-
tential negation (in the main clause) does not imply a negation of the proposition
expressed by the subordinate clause (p. 143). Although she does not fully explain
the test, it is quite straightforward to see from her examples to see how the negation
test is intended to work.
The negation test in the existential construction, as presented by Goldberg,
118
predicts correctly that the relative clause is not backgrounded and therefore com-
patible with A-bar movement (38). Applying it to a denite NP in a non-porous
context shows that it is backgrounded (39), and therefore blocks A-bar movement.
(38) There are many children [who like this kind of ice cream]. (Goldberg, 2006,
p. 146 ex. 59)
a. Sentential negation: There are not many children [who like this kind
of ice cream].
b. Proposition expressed by relative clause9: Many children like this kind
of ice cream.
c. Negation of proposition by relative clause: Many children don't like
this kind of ice cream.
d. (a) can imply (c): There are not many children [who like this kind of
ice cream] does imply Many children don't like this kind of ice cream.
(39) Mary saw the children [who like this kind of ice cream].
a. Sentential negation: Mary didn't see the children [who like this kind of
ice cream].
b. Proposition expressed by relative clause: The children like this kind
of ice cream.
c. Negation of proposition by relative clause: The children don't like this
kind of ice cream.
d. (a) does not imply (c): Mary didn't see the children [who like this kind
of ice cream] does not imply The children don't like this kind of ice
cream.
The negation test shows that purposive clauses pattern like porous relative
clauses.
(40) Mary went to the store [to buy apples].
a. Sentential negation: Mary didn't go to the store [to buy apples].
b. Proposition expressed by purposive clause: [Mary] bought apples
c. Negation of proposition by purposive clause: Mary didn't buy apples.
d. (a) can imply (c): Mary didn't go to the store to buy apples is compat-
ible with Mary didn't buy apples; for all we know, Mary went to the
store to buy only bananas.
9Note that in standard formal semantic analyses, relative clauses denote properties, not propo-
sitions. I will simply follow Goldberg's discussion in these examples.
119
However, Goldberg concedes that the test overgenerates. Specically, she
points out that it predicts that indenite NPs are always not backgrounded, giv-
ing an example like (41), and therefore should always be compatible with A-bar
movement, contrary to fact (42).
(41) She met a boy [who resembled her father]. (adapted from ibid. p. 147, ex.
63)
a. Sentential negation: She didn't meet a boy who resembled her father.
b. Proposition expressed by relative clause: A boy resembled her father
c. Negation of contents of relative clause: A boy did not resemble her
father
d. (a) can imply (c): She didn't meet a boy who resembled her father is
compatible with A boy did not resemble her father; for all we know,
no boys resemble the father in question.
(42) a. *Who1 did she meet a boy [who resembled t1]?
b. *That's the relative1 she met a boy [who resembled t1].
Even supposing that a more appropriate negation test can be found, this par-
ticular discourse-based approach is silent about cross-linguistic variation. For this
analysis to cover the facts for languages like German, one would need to assume that
either backgroundedness or negation somehow works dierently in German and lan-
guages like English. This would pose non-trivial questions about how learners might
acquire such cross-linguistic distinctions, since backgroundedness and negation are
highly abstract notions that are not explicitly taught.
3.2.2 Veridicality
Another potential approach, very similar to the intuition that underlies the
negation test and the backgroundedness analysis, is to analyze both porous relative
120
clauses and purposive clauses as non-veridical clausal adjuncts and to claim that
in languages like English, such adjuncts are exceptionally compatible with A-bar
movement. For the sake of argument, I will dene the notion of adjunct propo-
sition as the proposition denoted by the (highest) TP in a porous relative clause
or purposive clause, after existential closure of any unbound arguments. A clausal
adjunct is non-veridical if the adjunct proposition is not entailed by the proposition
denoted by the highest nite TP that contains the adjunct.
By these denitions, the porous relative clause in a non de re construction is
non-veridical (43a). In this construction, the complex NP receives a de dicto or third
reading, and does not denote any actual individual. Consequently, the adjunct is
a predicate that does not describe any actual individual, and the correpsonding
proposition is not necessarily true. A purposive clause is also non-veridical; in the
example in (43b), it could well be the case that there were no apples to be bought;
perhaps the store had run out of apples.
(43) a. Mary wants to nd a programmer [who can solve this problem]. ?6
someone can solve this problem [Non de re construction]
b. Mary went to the store to buy apples. ?6 someone bought apples
[Purposive clause]
It also arguably correctly predicts that because, so that, only to, and after
adjunct clauses are veridical, and therefore block A-bar movement.
(44) a. Mary went to the store [because she wanted to buy apples].? someone
wanted to buy apples (namely, Mary)
b. Mary went to the store [because she could buy apples]. ? someone
could buy apples (namely, Mary)
c. Mary left the room [only to run into John]. ? someone ran into John
(namely, Mary)
d. Mary left [after John called Sam]. ? John called Sam
121
However, this approach appears to under-generate; the denitions adopted
show that the other porous relative clauses are veridical, and so should block A-bar
movement (45). It also over-generates for before clauses, since these are not veridical
(46), although as observed above in Section 3.1.4.1, there are independent reasons
for why A-bar movement from before clauses should be unacceptable.
(45) a. Stevie is the only person [who wanted to record the song].? someone
wanted to record the song (namely, Stevie)
b. There are many people [who condone violence] in this town. ? some-
one condones violence
c. I know many people [who've been taken in by this trick]. ? someone
has been taken in by this trick
(46) Mozart died [before he completed his Requiem]. ?6 Mozart completed his
Requiem
3.2.3 Mood
Alternatively, one could posit the presence of mood morphology, and tie island-
hood to that, with a constraint like No A-bar movement from an adjunct marked
in the indicative mood. Some initial support for such a constraint comes from Ro-
mance languages, where a complex NP that receives a de dicto reading has a relative
clause that is marked in the subjunctive. In contrast, when the relative clause is in
the indicative, the complex NP receives a de re reading.
(47) Spanish (Romero, 2004, p. 272, exx. 1617, also Bleam, 1999 and references
therein)
a. En 1990, Juan estaba buscando a un profesor que pod?a
in 1990 Juan was looking-for PRT a professor that could.IND
hablar 7 idiomas.
speak 7 languages
122
`In 1990, Juan was looking for a professor that could.IND speak 7
languages. ' [Only de re reading of complex NP]
b. En 1990, Juan estaba buscando a un profesor que pudiese
In 1990 Juan was looking-for PRT a professor that could.SBJ
hablar 7 idiomas.
speak 7 languages
`In 1990, Juan was looking for a professor that could.SBJ speak 7 lan-
guages.' [Only de dicto
reading]
It is also necessary to assume some kind of covert mood morphology. Doing so
is necessary for purposive clauses, which are non-nite and do not show overt indica-
tive or subjunctive mood morphology, not even in Brazilian Portuguese; one might
hypothesize that purposive clauses have phonologically null subjunctive mood. In
a similar way, there are also languages in which nite clauses do not make a clear
morphological distinction, like English. One solution is to assume that these lan-
guages make the same mood distinction as Romance, except that mood morphology
is not directly observable.
For the sake of argument, suppose that purposive clauses and porous relative
clauses bear (abstract) subjunctive mood, even though their verbal morphology
does not suggest it. Learners observe A-bar movement from purposive clauses, and
conclude that all subjunctive-marked adjuncts, including porous relative clauses,
are compatible with A-bar movement.
However, this analysis overgenerates. Consider armative existential con-
structions and predicational constructions. In Romance, the relative clause of the
complex NP is marked in the indicative in these constructions (48). If mood in
English tracks mood in Romance, then the relative clause in English sentences like
123
There are many Americans who condone violence orMaria is the person who wanted
to record this song should bear indicative mood, which, by hypothesis, is incompat-
ible with A-bar movement. This prediction is not borne out: A-bar movement from
the relative clauses in these circumstances is relatively acceptable.
(48) Brazilian Portuguese (Gesoel Mendes, p.c.)
a. Tem muitos americanos que aceitam viol?ncia.
have many Americans that condone.IND violence
`There are many Americans who condone violence.
b. A Maria ? a pessoa que quis gravar essa m?sica.
the Maria is the person that wanted.IND record this song
`Maria is the person who wanted to record this song.'
3.3 Corpus study 2: A-bar movement from purposive clauses in the
input in English
In previous sections, I observed that porous relative clauses and purposive
clauses appear to pattern together with respect to their A-bar movement properties.
I argued that this is because they are members of a natural class: adjuncts that show
tense dependence. I also presented formal analyses of how tense dependence might
be derived and be connected to A-bar movement.
Of course, showing that porous relative clauses and purposive clauses form a
natural class does not immediately validate the proposed indirect learning account.
In fact, the account also requires that learners get to observe A-bar movement from
purposive clauses in the input. Using English and Mandarin Chinese corpus data,
I will argue that such structures are indeed available. I will then discuss what kind
of learning biases and knowledge learners need to make the appropriate inferences
124
to arrive at an adult-like grammar.
In principle, to really establish this point, one would ideally show that A-
bar movement from purposive clauses or porous relative clauses is absent in the
linguistic experience of German (Bulgarian, Polish, etc.) speakers, perhaps through
the analysis of comparable corpora. Unfortunately, this is not feasible, as I do not
have the relevant language skills or the relevant datasets to work with. However, as
noted previously, existing descriptions suggest that these structures are unacceptable
to native speakers. It is thus likely that such structures are either absent, or occur
only extremely rarely, such as in the event of a production error. It is thus plausible
to believe that A-bar movement from these kinds of clauses are much less likely to
be readily attested in the linguistic experience for speakers of these languages.
3.3.1 Datasets
Following Pearl and Sprouse (2013a,b), I chose two English language datasets
for analysis: the CHILDES Treebank (ibid.) and the Penn Treebank (Marcus et al.,
1999). As was the case for the rst corpus study (section 2.2.2), The ATIS-3 dataset
was excluded from the Penn Treebank analysis because it consists of queries about
air travel.
These two treebanks were chosen for several reasons.
First, the data in these treebanks can be thought to represent linguistic input
at dierent stages in the language acquisition process. The CHILDES Treebank is a
proxy of the input a monolingual learner of English encounters in the rst 5 years of
125
life (age of child = 0;65;1), while the Penn Treebank is intended as a proxy (albeit
an imperfect one) of the input encountered by an older monolingual English learner.
The Penn Treebank also makes up for certain limitations of the CHILDES
data. The CHILDES dataset studied is relatively small, pooling data from the
Brown, Soderstrom, Suppes, and Valian corpora (Brown, 1973; Soderstrom et al.,
2008; Suppes, 1974; Valian, 1991). For these four corpora, utterances of all types
were annotated, yielding a total of 173,058 utterances. Assuming that children hear
1 million utterances in the rst three years of learning a language, or approximately
333,000 utterances a year (Pearl and Sprouse, 2013a,b), these four corpora would
constitute only half a year of data.
A null result from a CHILDES corpus analysis  i.e. the absence of A-bar
movement from purposive clauses  might simply be an artifact of its size. The
slightly larger Penn Treebank (251,255 utterances, excluding the 572 utterances
from ATIS) compensates for this limitation to a certain extent.
Alternatively, learners might learn about A-bar movement from purposive
clauses and porous relative clauses at a relatively late age, after exposure to more
sophisticated syntax, which, for the purposes of this paper, is represented by the
Penn Treebank.
3.3.2 How much input does a learner get?
We need to normalize corpus counts so that we can estimate how frequently
a learner might observe such constructions. To that end, it is important to also
126
determine how much input a learner gets. Hart and Risley (1995) report that the
quantity of input varies widely across households, but generally increases with family
socioeconomic status (SES). I therefore estimated annual parent utterances for pro-
fessional families (i.e. high SES) and welfare families (low SES), using Hart and
Risley's average utterances per hour for these families (487 and 176, respectively)
and extrapolating this per-hour gure to a 5200-hour year (Hart and Risley, 2003,
p. 8). I note that they used the same extrapolation method to make the claim
that children from low SES backgrounds face a 30-million word gap. This method
produced a range of about 900,000 to 2.5 million utterances a year.
In contrast, other researchers have adopted more conservative estimates. Pearl
and Sprouse (2013b) (p. 40, also Pearl and Sprouse, 2013a) assume that children
hear only 1 million utterances in the rst three years of learning a language, or
approximately 333,000 utterances a year or about 64 utterances per hour. In a paper
on syntactic bootstrapping, White et al. (2018), adopting a conservative estimate,
assume that children hear at least 633 sentences a day (p. 10) or 231,000 sentences
a year (approximately 45 sentences an hour, given a 14-hour day).
One might be skeptical of Hart and Risley's assumption that children get
exposed to as many as 5,200 hours of linguistic input a year. I note, however, that
halving their estimates yields 450,000 to 1.25 million utterances a year, which are
still substantially higher than Pearl and Sprouse's or White et al.'s estimates.
127
Utterances encountered a year
White et al. estimate 230,936
Pearl and Sprouse estimate: 333,333
(49) Hart and Risley estimates for . . .
. . . Welfare (low SES) families 915,200
. . . Professional (high SES) families 2,532,400
Of course, quantity of input is not the only thing that varies for learners;
quality varies as well. For instance, there is evidence that children from low-SES
households are more likely to observe tokens with simpler syntax (e.g. Ho, 2003;
Huttenlocher et al., 2002, 2007). Put dierently, in the context of purposive clauses,
compared to their higher-SES peers, learners from low-SES backgrounds might ob-
serve purposive clauses less frequently in absolute terms (because they observe fewer
tokens in a given time period in general) and in relative terms (because purposive
clauses and other syntactically complex structures are under-represented in their
input). Consequently, they are also less likely to observe A-bar movement from pur-
posive clauses. While I do not have any reliable way to measure this dierence in
quality, this fact means that frequency estimates calculated below is likely to have a
large variance, i.e. it is likely to both overstate and understate the amount of input
available for sizeable parts of the population.
3.3.3 CHILDES Treebank study: Methods
As Pearl and Sprouse noted, the data in the CHILDES Treebank is much more
conversational in nature, containing more speech errors and fragments, which in turn
raises concerns about how an utterance might be assigned a tree representation. Out
of an abundance of caution, I extracted sentences of the surface forms ... V ...
128
to ... V ... with the Python Natural Language Toolkit (Bird et al., 2009).
This extraction algorithm generated many false positives. For instance, it is
intended to extract sentences containing purposive clauses like Mary left [to catch
the bus]. However, it also extracts sentences with non-nite arguments that resemble
the adjuncts of interest, e.g. Mary wanted [to catch the bus].
Most tokens can be straightforwardly identied as complements or adjuncts.
However, there are a small number of cases where it is less clear whether a clausal
constituent is a rationale clause (an adjunct) or a complement clause. In these
instances, the following tests were used; a clausal constituent needs to pass all tests
to be considered a rationale clause.
(50) Tests for rationale clauses
a. Fronting: an adjunct can typically be fronted.
b. Omission: an adjunct can typically be omitted (Ernst, 2001).
c. In order insertion for rationale clauses: in a VP of the schema V
...(for NP) to VP, the to VP is a rationale clause if in order can
be inserted immediately before (for NP) to (Whelpton, 1995).
d. Do so test for rationale clauses adjoined to the VP headed by an agen-
tive main verb: in a VP of the schema V ...to VP, the to VP is a
rationale clause if the preceding material can be replaced with do so
(Rissman and Rawlins, 2017).
e. Actuality entailment test for rationale clauses: The proposition denoted
by a rationale clause can be denied in a following sentence.
3.3.4 CHILDES Treebank: Results
The CHILDES Treebank script yielded 11,880 sentences containing at least
one non-nite clause. Half (5,907 sentences) were randomly selected for manual
coding. Of this set of 5,907 sentences, only one was a clear case of A-bar movement
129
from the adjunct (51).10
(51) because that 's what they were born to be *t*-1 . [Brown
corpus, Adam]
The results suggest that out of 173,000 sentences, there are approximately 2
sentences with A-bar movement from a purposive clause. In absolute terms, this is,
of course, very rare.
3.3.5 Penn Treebank study: Methods
I wrote a script using the Python Natural Language Toolkit to extract all A-bar
dependencies. For each A-bar dependency, following Pearl and Sprouse (2013a,b),
the script constructed a sequence of container nodes: the maximal projections that
dominate the gap but not the wh-phrase. The container nodes therefore represent
the path of movement of the wh-phrase.
I then selected A-bar dependencies whose container node sequences contain
maximal projections labeled as S-PRP and S-CLR. These projections are likely to
be adjuncts of verbal projections, according to the Penn Treebank documentation.
I also selected A-bar dependencies whose container node sequences contain an NP
that immediately dominates an SBAR or S, with the goal of identifying nite and
non-nite relative clauses. These sentences were then manually checked to ensure
that the A-bar dependencies involve A-bar movement from what are unambiguously
10(51) exemplies a purpose clause, not a rationale clause: the (subject) gap is bound by the
internal argument of the matrix verb; see e.g. Jones, 1991. There were also three examples of A-bar
movement from a to VP constituent following go, such as who did ya go to see ? (Brown
corpus, Sarah). Applying the criteria in (50), I do not count these examples as clear examples of
adjuncts. For example, this constituent cannot be freely omitted or preposed like adjuncts, e.g.
*To see Mary, John went.
130
adjuncts.
3.3.6 Penn Treebank: Results
The above process yielded 0 tokens of A-bar movement from a nite comple-
ment or adjunct of an NP, as mentioned in Section 2.2.1. The process did yield
16 tokens that contained A-bar movement from non-nite clausal constituents that
quite clearly have a purposive interpretation (52); I list all 16 examples in (52).
The majority of these tokens feature relativization; no wh-question example was
observed. A minority of these clausal adjuncts (5, to be specic) were analyzed as
adjoining to NPs in the Treebank, although a case for a VP adjunction analysis can
also be made, especially since they appear in the same position as an adjunct of VP
and there is no clear positive morphosyntactic evidence that they are adjoined to
NPs. Overall, the rate of occurrence of A-bar movement from a non-nite clausal
adjunct, 16 out of 251,000 utterances, is higher relative to the rate in the CHILDES
Treebank. In absolute terms, though, these A-bar movement examples are still fairly
uncommon in the Penn Treebank.
(52) Non-nite clauses that are unambiguously adjuncts in the Penn Treebank
a. A-bar movement from what the Penn Treebank analyzes as VP-level
adjuncts
(i) That is one thing 0 the experiments are designed *-2 *-
3 to find *T*-1 out .
(ii) What I am here * to do *T*-2 is * to report on the gy-
rations of the struggle  a struggle that *T*-1 amounts
to self-redefinition  to see if we can predict its fu-
ture course .
(iii) He is throttling the liberty 0 my father gave his life
* to win *T*-1  ! !
(iv) Beyond that misty gray of the rain , he saw the stretch-
131
ing hutment , low diminutive log cabins , chinked * with
mud , with doorways 0 a man would have *-2 to crouch * to
get through *T*-1 , with roofs of tenting laid * over boughs
or boards from hardtack boxes , or fence rails , with cranky
chimneys of sticks and dried mud .
(v) I think 0 the last movie that we went out *-1 to see *T*-2
was DANCES WITH WOLVES . E_S
(vi) and \[ it will , \+ depending on what topic you called
in *-1 to hear about *T*-2 whether it was the news or the
weather or a soap opera update , it will \] give you ,
um , updated information . E_S
(vii) We do have a lot of things \[ that , \+ you know , that
, \] you know , the taxes are there * to support *T*-1
, E_S
(viii) \[ he , \+ he \] said that 's what we were going there
*-3 to do *T*-1 E_S
(ix) And I think though that we need *-1 to have some liquid
*ICH*-3 around that you do n't have *-2 to go through the
company *-5 to get *T*-4 . E_S
(x) Not only does this come close to a violation of law , it
violates the trust 0 we have all worked *-1 to develop *T*-2 .

(xi) It also takes money that CNN has been reluctant *-1 to
spend *T*-2 *-3 to make programs and hire talent that view-
ers will tune in specially *-5 to see *T*-4 .
b. A-bar movement from what the Penn Treebank analyzes as NP-level
adjuncts
(i) It was all 0 Greg had time 0 * to see *T*-2 *T*-1 .
(ii) It has identified itself with the very tension and ter-
ror 0 it once did so much 0 *T*-1 to alleviate *T*-2 .
(iii) and \[ the one , \+ I have another one \] that *T*-1 's
three and a half and have all these clothes , and toys
and stuff that we 're just trying *-2 to find places
0 * to put *T*-4 *T*-3 , E_S
(iv) It seems 0 that 'd be a subject that \[ I , \+ I \] do
n't have much 0 *-1 to say *T*-2 on *T*-3 . E_S
(v) but \[ I , \+ uh , I \] basically never charge anything
0 I do n't have the money in the bank 0 *-2 to pay for
*T*-1 *T*-3 . E_S
132
3.3.7 General discussion
I rst summarize the CHILDES Treebank and Penn Treebank studies, giving
estimates for the number of utterances with A-bar movement from a purposive clause
that a learner might hear in a year (53).
(53) Summary of corpus study results
50% random Penn
sample of Treebank
CHILDES sample
Treebank
A-bar movement from purposive clauses 1 16
Extrapolated to . . .
- 231k utterances (White et al. estimate) 3 15
- 333k utterances (P&S estimate) 4 21
- 915k utterances (low-SES families) 11 58
- 2.5m utterances (high-SES families) 29 161
If A-bar movement from purposive clause is indeed the positive evidence learn-
ers use to infer that A-bar movement from porous relative clauses is possible, then
the CHILDES corpus data suggest that this evidence is relatively rare in the input
that a child receives in the rst three years of language acquisition. Exactly how rare
it is in absolute terms depends in part on one's assumptions about the amount of
input (setting aside questions about the quality of input or how accurately children
can parse their input, of course). Assuming Pearl and Sprouse's estimate of 333,000
utterances a year, a child comes across about 4 instances of A-bar movement from
rationale clauses a year, or about one example every three months. Adopting Hart
and Risley's estimates, however, one concludes that children from low-SES families
might hear as many as 11 such examples every year, or about one example a month;
children from high-SES families hear maybe 3 times as many examples. It is plausi-
133
ble, however, that these gures are still too infrequent for the learner, especially for
a child learner who might not be able to attend to all examples or assign a correct
parse to the ones that are attended to.
The frequency of positive evidence that A-bar movement is much higher in the
Penn Treebank than in the CHILDES dataset. Whether the rate of occurrence is
high enough in absolute terms depends in part on one's assumptions about the size
of the input, as well as whether older learners are more likely than younger learners
to assign the right parse to these utterances, and how they use positive evidence
to learn about their grammars. However, if Hart and Risley's methodology and
conclusions are sound, it appears that children from low SES backgrounds, who are
exposed to relatively little input, come across an average of 58 examples a year, or
about one example a week. The median English learner, who is exposed to even
more utterances, is therefore even more likely to come across examples like these,
parse them, recognize that their languages allow A-bar movement from purposive
clauses and therefore other (adjunct) clauses that exhibit tense dependence, such as
porous relative clauses.11
11More specically, if Hart and Risley's conclusions are sound, they predict that (i) the rate
of learning for young learners about the well-formedness of A-bar movement from porous relative
clauses might vary along SES lines and (ii) adult judgments about A-bar movement from porous
relative clauses might vary along SES lines.
134
3.4 Corpus study 3: A-bar movement from purposive clauses in the
input in Mandarin Chinese
One might be concerned about drawing conclusions about this indirect learning
strategy from English data alone. To address this concern, I repeat the same analysis
on the 51,774-sentence Chinese Treebank (Xue et al., 2010), whose range of registers
is comparable to that of the Penn Treebank. To preview the results, A-bar movement
from purposive clauses in Mandarin Chinese occur at a rate not dissimilar to that
of English. The Mandarin data thus provide independent support for this learning
account.
3.4.1 Methods
An inspection of the trees in the Chinese Treebank shows that purposive
clauses are annotated as VPs that are sisters to another VP. This annotation is too
coarse, as it does not indicate which VP contains the main predicate of the clause,
and groups purposive clause constructions with several others, such as manner and
coordination constructions (54).
(54) a. Nan Bao . . . [V P1 zhizao jia xinwen] [V P2 wumie
Southern Metropolis Paper create fake news slander
Wang Hongbin].
Wang Hongbin
The Southern Metropolis newspaper . . . created fake news to slander
Wang Hongbin.' [VP1 = main predicate; VP2 = purposive clause]
b. . . . zhiyao bu shi  [V P1 bu zou zheng tu] [V P2 moude]
as.long.as NEG be NEG walk correct path obtained
de liyi . . .
MOD advantage
135
`. . . as long as these were not advantages that were obtained (through)
improper means' (lit. `obtained by not walking along the correct path')
[VP1 = manner; VP2 = main predicate]
c. . . . yi-ge [V P1 zuo de jingmei] [V P2 dan meiyou duoshao
one-CL make PRT beautiful but NEG.exist how.many
ren zhidao] de wangzhan . . .
person know website
`. . . a website that is beautifully made but known by only a few people
. . . ' [Coordination]
Subsequently, I wrote a Python script that extracted these pairs of VPs, yield-
ing 7,744 VP pairs in total. Since the objective was to identify A-bar movement
from purposive clauses, which are the second of a sequence of two VPs, I selected for
tokens where the second VP ended with a trace, and then manually checked each
of the selected 150 sentences to ensure that the second VP was a purposive clause
and not the main VP. I used the following criteria, classifying the second VP as
a purposive clause if it meets all three of these criteria: (i) the rst VP describes
a situation necessary for the situation described in the second VP to take place,
(ii) there is some (implied or explicitly-expressed) agent shared between both VPs,
(iii) the rst VP cannot have use or expend -like semantics, just in case these verbs
all take the second VP as a complement, and (iv) the proposition denoted by the
second VP is not necessarily true, i.e. it can be negated with a follow-up clause.
I further excluded structures that resembled idioms or xed expressions and
structures where the second VP contained the focus particles cai and jiu, which are
usually found with main predicates, and hence likely to indicate that this VP is a
main predicate.
136
3.4.2 Results
This process yielded three relative clause examples, listed below. I have brack-
eted the relative clauses and underlined the purposive clause within each relative
clause.
(55) a. . . . [yue qian wo qu Huasha canjia] de Diyijie Shijie
month before I go Warsaw participate MOD inaugural world
Minzhu Luntan . . .
democracy forum
`. . . The Inaugural Forum on Global Democracy that I went to War-
saw a month ago to participate in . . . '
b. . . . [yumin chu-chun gandao Diaoyutai huo Dong Hai qu
sherman early-spring rush Diaoyutai or East Sea go
lanjie] de xiao xiaojuan . . .
intercept MOD small squid
`the juvenile squid that shermen rush to Diaoyutai or the East Sea in
early spring to intercept'
c. . . . hen duo ziliao-pian dou shi [Heiyoulong ziji dao
very many informational-lm all be Heiyoulong self go
guowai xunzhao] de . . .
overseas nd MOD
`. . . many of the informational videos were all such that Heiyoulong
went overseas to nd them . . . '
3.4.3 Discussion
For ease of comparison, I will assume that Mandarin-learning children are
exposed to the same amount of utterances as their English-learning peers. This
assumption allows us to estimate the number of A-bar movement from purposive
clauses observed by a Mandarin learner in a year, which I report in (56).
(56) Summary of corpus study results
137
Chinese Treebank
A-bar movement from purposive clauses 3
Extrapolated to . . .
- 231k utterances (White et al. estimate) 13
- 333k utterances (P&S estimate) 19
- 915k utterances (low-SES families) 53
- 2.5m utterances (high-SES families) 147
The frequency of direct positive evidence for the A-bar movement from pur-
posive clauses (and indirectly for A-bar movement from porous relative clauses) in
the Chinese Treebank is comparable to that for the Penn Treebank. For example,
assuming learners are exposed to 915,000 tokens a year (for low-SES families), a
Chinese learner might observe about 53 examples of A-bar movement from purposive
clauses a year, compared to the 58 examples observed by their English counterparts.
3.5 Corpus study 4: How much positive evidence is necessary? Evi-
dence from four other constructions in English
I argued above that speakers of English (Mandarin, etc.) might learn that
A-bar movement from porous relative clauses is acceptable in their languages by
observing A-bar movement from purposive clauses. The corpus studies showed that
while A-bar movement from purposive clauses do occur, they do so infrequently,
which raises valid questions about whether there are sucient positive examples for
the learner to revise the default assumption. As far as I know, there is no theory
or consensus in the literature on the specic relationship between the availability of
evidence and acquisition outcomes.
In the absence of a consensus or a theory that oers testable predictions, I
138
look at other constructions that have to be learned on the basis of sparse evidence.
Ideally, one would focus on constructions that are minimally dierent from A-bar
movement from porous relative clauses, namely, constructions whose syntactic prop-
erties are independently known to be learned indirectly from a supercially dierent
construction that occurs infrequently. However, I have not been able to identify a
construction that meet these criteria: certainly, there are constructions that have
been claimed to be learned indirectly (e.g. the compounding parameter (Snyder,
2001) or A-bar movement that does not exhibit the that-trace eect (Chac?n, 2015;
Rizzi, 1982)), but it is not clear whether the relevant evidence is as sparse as is
needed for the purpose of this analysis.
The next best alternative is to look at constructions that have to be learned
directly from sparse evidence. I identied four other constructions (57): (pseudo-
)gapping involving the subject and a non-subject (suggested to me by Howard
Lasnik, p.c.), pseudopassives, and two constructions related to A-bar movement,
namely, swiping (also known as sluice-stranding), and what I will call the verbless
no matter  construction, where no matter appears with a wh-question that lacks a
verb (see Culicover, 1999 for more discussion of the properties of the last two con-
structions). I will use these constructions to estimate how little positive evidence
(measured in terms of frequency of appearance in a corpus) is needed to learn a
particular construction by adulthood.
(57) a. John likes Coke, and Mary Pepsi. [Gapping]
b. This bed was slept in. [Pseudopassives]
c. Mary and John had a meeting, but I don't know what about. [Swiping]
d. We suggest that you say very little, no matter how strange the idea.
139
[Verbless no matter construction, Culicover, 1999, p. 111, ex. 30c]
Three criteria guide the selection of these constructions. First, they are likely
to have been learned from direct (positive) evidence. To the best of my knowledge,
these constructions are not uniformly attested across a wide range of languages; in
fact, they are typologically marked and in the case of no matter, idiosyncratic to
English.
Second, to the best of my knowledge, these phenomena do not appear to cor-
relate with another more easily-observable property of English, making it less likely
that they are learned indirectly. For example, one might imagine that the distri-
bution of pseudopassives and swiping is identical to that of preposition-stranding,
but as Truswell (2008) and Merchant (2002) show, this is not the case; there are
preposition-stranding languages that lack pseudopassives and/or swiping.
Third, impressionistically, these are low-frequency constructions, which ar-
guably pose learning problems not unlike that posed by porous relative clauses/purposive
clauses. A search of a corpus, in this case, the Penn Treebank, for these construc-
tions therefore yields an estimate of how much direct evidence is available in an
English speaker's linguistic experience.
Because the analysis is done on constructions that are learned directly, there
is an important caveat in interpreting the results of this analysis. For these four
constructions, the evidence that learners get is unambiguously direct  an instance
of pseudo-gapping constitutes evidence that pseudo-gapping is grammatical. The
same directness of evidence clearly holds for A-bar movement from purposive clauses,
140
but less so for A-bar movement from porous relative clauses, since purposive clauses
are morphosyntactically dierent from porous relative clauses in English (Mandarin
Chinese, Hebrew, and so on), diering at least in terms of niteness and/or tense
morphology, whether subjects are overtly expressed, and what they adjoin to.12
3.5.1 Methods
The Penn Treebank has its own annotation scheme for various constructions
that it considers to be gapping; the subject/non-subject gapping tokens were ex-
tracted by running a search for trees with the relevant annotation.
Pseudo-passives were identied by identifying tokens where a verb in the past
participle form (VBN) preceded a preposition that in turn immediately preceded
an A-trace (e.g. *-1).
Swiping tokens were identied manually after extracting strings that consisted
of a bare wh-phrase followed by a preposition; in this regard I follow Merchant (2002)
in assuming that the most acceptable tokens of swiping involve a bare wh-phrase,
pace Radford and Iwasaki (2015), who report naturally-occurring examples with
D-linked wh-phrases.
Verbless no matter constructions were identied manually by a search for
sentences containing the string no matter.
12Thanks to Jerey Lidz for discussion on this point.
141
3.5.2 Results
(58) Frequency of gapping, pseudo-passives, swiping, and verbless no matter
constructions in the Penn Treebank
Gapping Pseudo- Swiping Verbless
passives no
matter
In entire PTB dataset 235 94 7 2
Extrapolated to . . .
- 200k utterances (White et al. es- 216 86 6 2
timate)
- 333k utterances (P&S estimate) 312 125 9 3
- 915k utterances (low-SES fami- 856 342 25 7
lies)
- 2.5m utterances (high-SES fami- 2,369 947 71 20
lies)
The results bear out our intuition that these constructions occur infrequently,
i.e. positive evidence for the learner is sparse. Of the four constructions, the most
common is subject/non-subject gapping, which appears 235 times in the corpus.
In contrast, only 7 tokens were found for swiping and 2 tokens for the verbless no
matter construction.
3.5.3 Discussion: Further inferences a learner has to make
The results suggest that, rst, the sparsity of positive evidence is true not
only for A-bar movement from clausal adjuncts but also for at least several other
constructions. Second, sparsity itself is not necessarily a problem for learning a
particular rule or constraint. More specically, if the Penn Treebank frequency esti-
mates are a reasonable approximation of an (American) English speaker's linguistic
experience, English speakers can acquire constructions like swiping and verbless no
matter constructions with as few as 2071 tokens a year (assuming Hart and Risley
142
(1995)'s estimates) (58). It is thus not implausible that they can also draw con-
clusions about A-bar movement from porous relative clauses by observing A-bar
movement from purposive clauses, which occur at a higher rate.
One might object to this argument from swiping and the verbless no matter
construction. First, the argument assumes that there is no signicant dierence
in the nature of the rules and constraints that underlie swiping, verbless no mat-
ter construction, and A-bar movement from porous relative clauses (and purposive
clauses). It is logically possible that there is a stronger innate bias for the rules
and constraints involved in the rst two constructions, so that the sparsity of direct
evidence is less of an obstacle. Second, even if it were the case that there is no
signicant dierence in biases, it might be the case that the rules and constraints
for these two constructions are in some sense less complex, and therefore easier to
learn. For instance, as mentioned above, the inferences from evidence that learners
have to make about swiping, the verbless no matter construction, and A-bar move-
ment from purposive clauses are much more direct, and thus in a way easier, than
the inferences learners have to draw about A-bar movement from porous relative
clauses. I note that this dierence holds even when one assumes that at some ab-
stract level of representation, porous relative clauses and purposive clauses involve
the same structures.
I will not have much else to say about the rst assumption, since there is no
independent evidence for the biases learners have for any of these constructions.
However, it is possible to qualitatively evaluate the complexity of the learning task
for each construction, by spelling out more explicitly the inferences learners have to
143
make.
I assume that the learner can identify the appropriate predicate-argument
relations denoted by these construtions, even if they might be unfamiliar with the
construction. On the assumption that these relations are encoded at a D(eep)
Structure-like representation, the learner has to deduce the rules and conditions
that apply on this representation to generate the strings observed in the input.
3.5.3.1 Swiping
The learning task here turns out to be fairly challenging. Swiping is subject
to several conditions that are dicult to learn from a small number of examples.
First, as Merchant points out, swiping is most acceptable when the wh-phrase is not
D-linked (i.e. when the wh-phrase is monomorphemic, but see Radford and Iwasaki
for counterexamples). There are additional constraints on what prepositions may
appear; for example, the preposition must also be a simplex one, appearing without
modiers (Culicover, 1999). A learner might be able to infer these properties when
presented with a large enough set of examples, but it is not clear whether he/she
can do so on the basis of 671 tokens a year (58).
Because of these restrictions on the form of swiping, swiping tokens are short.
Consequently, learners might misanalyze instances of swiping as monomorphemic
lexical items and/or idiomatic expressions that happen to resemble the concatena-
tion of a bare wh-phrase and a preposition, rather than the result of a productive
phenomenon. Alternatively, the low frequency of swiping tokens might cause learn-
144
ers to incorrectly treat them as speech errors.
Further, even though a swiping token consists of only a wh-phrase and a
preposition, it is interpreted as if there were an entire clause that was elided.13 As
noted in the literature on swiping and ellipsis, this clause must be recovered from
the context. It is not at all clear whether the need to recover a clause is obvious
to the learner. Ideally, there should be cues that a clause is involved: perhaps
there would be a morphosyntactic sign that a clause has been elided  the clausal
equivalent of do-support, which applies when a VP is elided in the absence of a
modal auxiliary)  or perhaps at least the preposition in the swiping token might
be found prominently in a previous utterance.
Such cues are absent. There are of course no morphological signs of clausal
ellipsis in English. In the examples seen in the PTB, the preposition and wh-phrase
do not have any obvious linguistic antecedents, so there is no clear syntactic evidence
that the swiping token is related to some clause in the context. For example, in the
following dialogue taken from the Penn Treebank, What for? is interpreted along
the lines of What are you on your way to see the Jacobs woman for? but neither
for nor what are found in the previous sentence, I'm on my way to see the Jacobs
woman.
(59) I 'm on my way *-1 to see the Jacobs woman  .
 Gilborn 's secretary ? ? What for *T*-1 ? ? You do n't think
0 Gilborn is the   ? ?
Despite the obscurity of the restrictions on swiping and the relative sparsity of
13It is unclear whether there is syntactic structure present in the ellipsis site (Merchant, 2001);
swiping does not provide the kind of connectivity eects that are needed to substantiate such a
claim.
145
positive evidence, there is some (rather weak) evidence that swiping can be acquired
relatively early, although this seems to be the exception than the rule. In a study of
longitudinal CHILDES corpora for 20 children, Sugisaki (2008) found two children
who produced swiping constructions. Excluding one child that produced only what
for, the other child (Aran, from the Manchester Corpus (Theakston et al., 2001))
produced a variety of forms from 2 years 7 months of age, namely: who for (2
tokens), what for (2 tokens), what with (1 token), and who from (1 token).
Assuming that production is a reliable indicator of acquisition (admittedly a
assumption that many might question), the production data would mean that Aran
was able to deduce the syntax of swiping at from relatively few tokens. Note that
this would be true even if one assumes that Aran grew up in a linguistic environment
where swiping is atypically common, since he would have only at most about two
years of linguistic input to work with.14
3.5.3.2 Verbless no matter construction
As Culicover (1999) notes, there are three features unique to this construction:
(i) no matter appears to take a small clauselike complement, without any overt
verb, (ii) the wh-phrase and NP that follows it are interpreted as predicate and
subject respectively (60a), and (iii) the subject must be a denite generic (60b). The
sparsity of evidence makes all three features dicult to detect: learners need to infer
that the complement lacks a verb and thus is not, for instance, a nite clause that
14Although Aran used swiping relatively frequently, this does not seem to be the case for the
adult speaker in the corpus, his mother. In the entire corpus, Aran's mother produced only one
token of swiping: what for.
146
just happens to occur without a verb due to some speech error or perception error.
Similarly, it is dicult to tell from the evidence whether the presence of denite
generic and the fact that the wh-phrase is always a predicate are simply coincidences
due to the small sample or reect a restriction specic to the construction.
(60) a. Everyone has certain basic rights, no matter who *(is / might be)
president. [wh-phrase is the subject]
b. We suggest that you say as little as possible, no matter how upset
{Mary and John / she and he / clients} *(might be). [Subject is not
denite generic]
One possibility is that learners do eventually come across enough tokens of the
verbless no matter construction to support the hypothesis that there are actually
two, and not one, no matter constructions. At this point, for a learning account
to work, one would need to assume that the learner compares the distribution of
wh-phrase predicates and denite generic subjects in these verbless no matter con-
structions with the equivalent distribution in regular no matter constructions; it
seems likely that there will be clear dierences in the distributions to suggest that
the verbless no matter construction is subject to additional restrictions.
3.6 What kind of learner do we need?
In this section, I describe how learners of a language like English might use
observations of A-bar movement from purposive clauses to infer that A-bar move-
ment from porous relative clauses is possible. There are several distinct problems
here.
The rst problem relates to the observation and representation of the evidence:
147
learners must be able to observe and correctly analyze A-bar movement from pur-
posive clauses and recognize it as such, as opposed to, for instance, treating these
tokens as mere production errors or as well-formed tokens that do not involve A-bar
movement. It is arguably simple for a learner to observe that A-bar movement has
taken place; the tokens in the input are mostly in the form of relative clauses, which
are prototypical A-bar movement constructions. I will also assume that learners are
exposed elsewhere to suciently many instances of adjuncts and have a familiarity
with the argument structure of lexical items, so that they can tell that these exam-
ples involve A-bar movement from an adjunct (specically, a purposive clause), as
opposed to A-bar movement from a specier or complement.
Further, learners must analyze purposive clauses and porous relative clauses
in a particular manner so that they can be treated as members of a natural class. At
the very least, they need to recognize that both purposive clauses and porous relative
clauses are distinguished by their tense dependence. I hypothesized in Chapter 2
(Section 2.2) that learners might use the distribution of tense morphemes in porous
relative clauses; they might use the presence of non-nite to to determine that pur-
posive clauses have dependent tense. Only then can they use an observed instance
of A-bar movement from purposive clauses to conclude that A-bar movement from
another type of adjunct with tense dependence (i.e. porous relative clauses) is also
possible.
The second problem, an important one theoretically, relates to how learners
interpret the positive evidence. This problem lies in the fact that purposive clauses
are ultimately dierent from porous relative clauses, even though purposive clauses
148
and porous relative clauses form a natural class. Consequently, there are many
other natural classes that purposive clauses belong to, and logically speaking, the
observation of A-bar movement from purposive clauses does not automatically invite
the inference that A-bar movement from porous relative clauses is possible.
In other words, in principle, there are many alternative conclusions a learner
could draw from A-bar movement from purposive clauses. (61) presents some of
these possible conclusions, obtained via a supercial introspection of the syntax,
semantics, and pragmatics of the tokens involving A-bar movement from purposive
clauses in the Penn Treebank dataset. Note that this set of conclusions here is
stated with respect to adjuncts; one can imagine other conclusions that are stated
in non-adjunct terms.
(61) a. Syntactic conclusions: Allow A-bar movement from . . .
(i) . . . any adjunct
(ii) . . . an adjunct whose subject is bound by the matrix subject.
(iii) . . . a non-nite adjunct.
b. Conclusions based on truth-conditional semantics: allow A-bar move-
ment from . . .
(i) . . . adjuncts whose propositions are not always true of the actual
world
(ii) . . . an adjunct that is temporally located to the future of the event
of the main predicate, since these clauses describe goals.
(iii) . . . an adjunct that is temporally located to the future of utterance
time, since these clauses describe goals.
c. Conclusions based on event semantics: allow A-bar movement from an
adjunct that describes a result where the main predicate is a necessary
condition for that result.
d. Conclusions in construction/idiom terms: allow A-bar movement from
. . .
(i) . . . an adjunct that has some combination of the above syntactic
and semanctic properties, e.g. a non-nite adjunct that is tem-
porally located to the future of the main predicate . . .
(ii) . . . an adjunct that is lexically and structurally identical to the
counterexamples.
149
Despite the range of options, learners of English (Mandarin, Hebrew, etc.)
seem to invariably draw the right conclusion about their languages, allowing A-
bar movement from purposive clauses and porous relative clauses but not from the
adjuncts listed above. This raises the question of why they converge and how they
rule out the other options.
The most likely solution involves a very restricted hypothesis space about
A-bar movement: the range of hypotheses about their grammars that learners ac-
tually consider is small. The size of this hypothesis space and its contents (i.e. the
hypotheses) are presumably innately determined.
More specically, in this context, there are at least two distinct hypotheses,
one that is responsible for generating languages like Polish and German, where A-
bar movement from relative clauses and purposive clauses is not acceptable, and
another that generates languages like English and Mandarin Chinese, where A-
bar movement from these adjuncts is acceptable. For example, supposing that the
defective tense analysis of porous relative clauses and purposive clauses is correct,
learners consider at least the following two hypotheses about their languages: (i)
Defective tense morphemes neutralize phases, (ii) Defective tense morphemes do not
neutralize phases.
Because there are only a small number of hypotheses to begin with, the ob-
servation of A-bar movement from purposive clauses counts as evidence for an even
smaller number of hypotheses (possibly only one, e.g. the defective tense analysis
or the movement/height-based analysis), which explains convergence.15
15Of course, as mentioned in the start of Chapter 2, there appears to be more ne-grained
150
We may contrast this conclusion with an alternative in which learners entertain
a much larger set of hypotheses, as in (61), and arrive at the right conclusion by trial
and error. In the latter scenario, convergence is only possible if A-bar movement
from purposive clauses is abundant in the input, so that learners get enough tokens
to rule out inappropriate conclusions. However, corpus analyses show that in at least
English and Mandarin Chinese, A-bar movement from purposive clauses occurs only
infrequently.
This particular conclusion recalls key ideas from the Principles and Parame-
ters framework, in which languages were hypothesized to dier along a number of
innately-dened dimensions (parameters; ideally a very small number), such that
there are only a few ways in which each parameter might vary. These limits on pa-
rameters and parameter values are also restrictions on a learner's hypothesis space.
A third problem pertains to the mechanism by which the evidence is used to
determine one's grammar. There is a wide range of proposals in the literature that
can be applied straightforwardly to the case of porous relative clauses and purposive
clauses; these adjuncts do not pose a major theoretical problem in this domain.
In the classical Principles and Parameters framework, this task is accomplished
by assuming that the hypotheses are ranked, such that the most restrictive hypoth-
esis is the default parameter setting. Where necessary, learners switch to the next
cross-linguistic variation even within languages that allow A-bar movement from porous relative
clauses, e.g. whether adjuncts can move from porous relative clauses. Given current gaps in our
knowledge of this particular kind of variation, it is unclear to me whether this kind of variation
reects dierences in the specic condition(s) that govern A-bar movement from porous relative
clauses or dierences in other aspects of A-bar movement. For example, perhaps this particular
dierence relating to adjuncts actually reects dierences in terms of what kind of adjuncts may
undergo A-bar movement.
151
most conservative alternative setting that can accommodate positive evidence that
is incompatible with the current setting (Berwick, 1985; Dell, 1981; Manzini and
Wexler, 1987). Alternatively, it is also possible to produce similar outcomes through
more probabilistic approaches. To briey summarize some of these approaches: For
example, Gibson and Wexler (1994) present a non-deterministic analysis where pos-
itive evidence causes learners to consider an alternative grammar, by resetting a
parameter chosen at random. Yang (2002) suggest that learners assign probabilities
to multiple grammars. They increase the probability associated with grammars that
can best accommodate observed linguistic evidence, while decreasing the probabil-
ity associated with grammars that cannot do so. Pearl (2007) suggests a Bayesian
updating procedure and shows how this procedure works on dierent kinds of hy-
pothesis spaces.
3.7 Conclusion
This chapter started with the proposal that learners of languages like English
might use A-bar movement from purposive clauses to learn that their languages allow
A-bar movement from porous relative clauses. I discussed various hypotheses about
why purposive clauses might be the relevant evidence in this context, suggesting
that this is because both purposive clauses and porous relative clauses exhibit tense
dependence. I then presented corpus studies to argue that A-bar movement from
purposive clauses occurs relatively frequently (compared to A-bar movement from
porous relative clauses) in natural contexts in both English and Mandarin Chinese,
152
a necessary condition if this learning account is to work.
That said, positive evidence is not sucient to explain why learners of these
language succeed. In particular, I argued that in this case, circumstances suggest
that there must be restrictions on the hypothesis space. In absolute terms, A-bar
movement from purposive clauses occurs infrequently in English and Mandarin Chi-
nese, so the evidence for A-bar movement from porous relative clauses is sparse,
in addition to being indirect. Without restrictions on the hypothesis space, there
would be too many logically coherent conclusions for learners to draw about their
languages by just observing A-bar movement from purposive clauses; there is no
guarantee that they use it to infer that their languages allow A-bar movement from
porous relative clauses. These restrictions thus allow learners to overcome the ev-
idence sparsity problem and ensures that learners exposed to the same language
converge on the same kind of grammar.
153
Chapter 4: Variation in A-bar movement from complement clauses
4.1 Introduction
While researchers following Ross's footsteps have proposed constraints on
overt A-bar movement from a variety of embedded clauses, one class of embedded
clauses has remained relatively exempt from these proposals: declarative comple-
ment clauses of verbs, especially verbs that describe mental states and communica-
tion, such as think and say. The absence of constraints on A-bar movement from
these clauses is perhaps not surprising, since that makes it possible to straightfor-
wardly explain why an English sentence like (1) is acceptable.
(1) Who1 does John {think/say} [that Mary likes t1]?
However, within English and many other languages, not all verbs freely allow
(long-distance) A-bar movement from their declarative clausal complements. In
descriptions of long-distance A-bar movement, it is common to split the class of
verbs into two sub-classes, typically labeled as bridge verbs and non-bridge verbs,
such that A-bar movement across the rst class is acceptable (taken to indicate well-
formedness) and A-bar movement across the second is not (2).
0For discussions about some of the issues presented in this chapter, I would like to thank Amy
Rose Deal, Sandhya Sundaresan, Aaron Steven White, and Charles Yang. Errors are mine alone,
of course.
154
(2) *Who1 does John {shout/quip} [that Mary likes t1]? [Non-bridge verb]
A question that naturally arises is which verbs are bridge verbs  whether
they vary across languages  and why these verbs have such an impact on A-bar
movement. Supposing that there is clear reason to think there is cross-linguistic
variation, one might ask about how children learn these points of variation.
For ease of reference in this chapter, I will dene the following terms:
(3) a. Representational attitude verbs, or attitude verbs for short, refer to
clause-embedding verbs that describe communication and representa-
tional mental states. These include verbs like say, think, believe,
predict, and so on, where the speaker asserts a proposition or believes
in the truth of the proposition denoted by the verb's clausal comple-
ment. This denition excludes attitude verbs with preferential or control
semantics, like want or try.
b. Bridge verbs, as a descriptive term, refers to the set of attitude verbs
that allow A-bar movement. Because there is no widely accepted list of
bridge verbs for English (or for any other language, as far as I can tell), I
will mostly rely on informal intuitions to determine whether to describe
a verb as a bridge verb or not.
c. Bridge eects refers to the acceptability (or lack thereof) of A-bar
movement from the clausal complements of verbs.
d. Bridge contexts, or long-distance A-bar movement, refers to syntactic
congurations where a phrase, such as a wh-phrase, moves from the
clausal complement of a verb to an A-bar position, whether the verb is
a bridge verb or not.
This chapter is organized as follows. First, in Section 4.2, I discuss complica-
tions with the bridge/non-bridge distinction and point out a general data quality
issue in existing discussion of this distinction. With these caveats in mind, I discuss
existing competing generalizations about bridge verbs in Section 4.3. Specically, I
review a number of hypotheses about bridge eects.
Second, I evaluate existing hypotheses in Section 4.4. I describe a larger scale
acceptability judgment study for 100 verbs in English, showing that existing propos-
155
als about bridge verbs provide relatively little coverage of the range of acceptability.
In Section 4.5, I present a post-hoc analysis of the data, showing that the relevant
dimensions are likely more ne-grained and specic than those suggested in existing
proposals.
Third, I discuss consequences of this study for these generalizations and our
theories of the learning of long-distance A-bar movement. In Section 4.6, I dis-
cuss the extent to which bridge eects might be pragmatic or syntactic in nature;
the former suggestion is often made in the context of long-distance wh-questions
(Ambridge and Goldberg, 2008; Erteschik-Shir, 1973; Erteschik-Shir and Lappin,
1979; M?ller, 2015b). In Section 4.7, I present data from a smaller scale, informal
survey of English and Dutch native speakers that examines the eect of the pres-
ence or absence of a suitable context on long-distance A-bar movement. To preview
the results, speakers across both languages seem to nd that long-distance A-bar
movement improves (somewhat) in acceptability with the presence of a suitable
context for manner of speaking verbs. The survey also provides some evidence that
long-distance A-bar movement for cognitive factive verbs has dierent acceptability
proles in both languages. This fact suggests that in at least some languages, some
aspect of bridge phenomena is learned from linguistic experience.
In Section 4.8, I discuss the possibility that the dierences between English
and Dutch for cognitive factive verbs might be derived from English learners getting
direct evidence that long-distance A-bar movement is compatible with these verbs.
I compare two accounts about the kinds of generalizations English learners might
make on the basis of the evidence they get.
156
This chapter also seeks to make an empirical contribution to the literature; the
data collection reported in this chapter is aimed at addressing the data quality issue
mentioned in Section 4.2. The formal experiment in Section 4.4 seeks to expand
our understanding of bridge eects across a larger set of clause-embedding verbs
in English, while the informal survey in Section 4.7 seeks to evaluate the extent of
cross-linguistic variation between English and Dutch and the eect of context on
acceptability of long-distance A-bar movement.
4.2 An overview of the literature on bridge verbs
It is important to note that the bridge/non-bridge distinction, while useful and
inuential in the literature on A-bar movement, understates the complexity of A-
bar movement from complement clauses. In this section, I enumerate several issues
pointed out in the literature, which I also intend as a summary of the state of aairs
in the literature. The takeaway is that there are a range of factors contributing to
acceptability. The bridge/non-bridge distinction arguably captures only one par-
ticular dimension, which manifests itself in the acceptability of A-bar movement of
arguments (such as objects). A lack of data means that it is harder to tell whether
this particular dimension aects A-bar movement of other kinds of wh-phrases.
4.2.1 Restrictions on what moves
The rst issue involves restrictions on the items that move. A number of verbs,
such as admit, are said to allow long-distance A-bar movement of (wh-)arguments
157
(typically objects) but not (wh-)adjuncts (4), while another set of verbs, such as
whisper, are held to block long-distance A-bar movement of both wh-arguments and
wh-adjuncts. (In other words, the complements of the rst set of verbs are weak
islands, while those of the second set are strong islands, in the sense of Szabolcsi
and Zwarts, 1993, Honcoop, 1998, and Abrus?n, 2014, among others. I am not
aware of any verbs that allow long-distance A-bar movement of (wh-)adjuncts but
not (wh-)arguments.) The nature of this argument/adjunct asymmetry is less well-
studied. It is unclear whether there is systematic cross-linguistic variation for the
A-bar movement of adjuncts, or whether there is some universal semantic or dis-
course property that characterizes the verbs that allow A-bar movement of adjuncts
(see Cattell, 1978; Hegarty, 1992; Honcoop, 1998). More basic cross-linguistic data
collection is necessary, and I leave this to future research.
(4) a. Who1 did John admit [that Mary likes t1]?
b. *Why1 did John admit [that Mary likes Ben t1]?
(5) a. ??Who1 did John whisper [that Mary likes t1]?
b. *Why1 did John whisper [that Mary likes Ben t1]?
A related complication involves A-bar movement of subjects, which is some-
times said to be degraded relative to A-bar movement of objects for at least certain
attitude verbs (typically factives). While I do not wish to challenge the validity
of the judgments, I would like to point out that the datapoints in English might
require additional support for the point to be valid, due to a confound involving the
that-trace eect. As Grimshaw (2015) notes, a number of these verbs, especially
low-frequency, high-register ones, might require their complement clause to contain
158
the complementizer that. However, in English, a subject gap in a complement clause
cannot appear immediately following a that. In other words, it is hard to determine
whether examples like (6) show that A-bar movement of subjects is degraded with-
out evidence showing that these complement clauses allow a null complementizer.
Ideally, one could address the confound by including the overt complementizer that
and an adverb that comes between that and the subject gap, which neutralizes the
that-trace eect (an observation due to Bresnan, 1977 and more recently elaborated
by Kandybowicz (2006) and Douglas (2017)). (7), for instance, does not seem too
bad to me, compared to (6a).
(6) a. *Who do you regret [t1 stole the cookies]? (Basse, 2008, p. 54, ex. 3)
b. *Who do you remember/deny [(that) stole the cookies]? (Kastner, 2015,
p. 165, ex. 30)
(7) Who do you regret [that, during the blackout last night, t1 stole the cookies]?
Finally, I note that there appears to be a slight improvement in acceptabil-
ity when the wh-phrase is discourse-linked (D-linked, Pesetsky, 1987), at least for
certain non-bridge verbs in Dutch, although existing descriptions are sketchy. Specif-
ically, Broekhuis and Corver (2016) report that in Dutch a factive verb like weten
`know' typically blocks A-bar movement of a bare wh-argument like wat `what,'
while a verb like denken `think' does not. Making the wh-argument D-linked im-
proves acceptability for long-distance A-bar movement across weten, at least for a
subset of speakers (8).
(8) Dutch (a and b examples from Broekhuis and Corver, 2016, p. 1397, ex.
214).
159
a. Wat {dacht/*wist} Marie [dat Marie t1 gekocht had]?
what thought/knew Jan that Marie bought had
`What did Jan think/know that Marie had bought?'
b. Welk boek1 {dacht/%wist} Marie [dat Marie t1 gekocht had]?
which book thought/knew Jan that Marie bought had
`What did Jan think/know that Marie had bought?' (% = some speakers
only)
4.2.2 Construction-specic eects
A second issue involves the type of A-bar movement construction. Typically,
examples intended to show whether a verb is a bridge verb or not involve wh-
questions: if a wh-phrase can move from the complement clause of an attitude verb,
then the verb is a bridge verb; if the wh-phrase cannot, the verb is not a bridge
verb.
However, as various researchers have pointed out, wh-questions come with
their own pragmatics and felicity conditions (Abrus?n, 2011; Gibson, 2018; Liu et al.,
2019; M?ller, 2015b; Oshima, 2006; Schwarz and Simonenko, 2018; Simonenko, 2013,
2015). If this point holds, an unacceptable wh-question formed by long-distance A-
bar movement might not be clear evidence that the main verb blocks A-bar move-
ment from the complement clause. Instead, the low acceptability might just be an
artifact that arises jointly from the properties of wh-questions and the verb. For
example, a number of proposals report that wh-questions involving A-bar movement
from factive verb complements are relatively unacceptable (Abrus?n, 2011; M?ller,
2015b; Oshima, 2006; Schwarz and Simonenko, 2018). Although technical details
and background assumptions vary, these proposals tend to attribute the low ac-
160
ceptability to a conspiracy between the semantics of wh-questions and the fact that
the complements of these verbs are presupposed to be true or part of the common
ground.
A further consequence of this line of reasoning is that it predicts that there are
other A-bar movement constructions in which A-bar movement from complement
clauses is more acceptable; a claim most recently put forward by Gibson (2018)/Liu
et al. (2019) with it-clefts, which feature A-bar movement (9).1
(9) a. What1 did Susan know that Anthony liked t1? [Wh-question] <
b. It was the pie1 that Susan knew that Anthony liked t1. [It-cleft]
(`A < B' = A is less acceptable than B)
4.2.3 Cross-linguistic variation
Cross-linguistically, bridge eects appear to vary across languages, a fact
that further complicates eorts to derive an empirically adequate generalization.
I did a cross-linguistic literature survey, looking at informal descriptions of bridge
verbs and A-bar movement from nite (indicative) complement clauses in the liter-
ature. The languages covered were: English (Abrus?n, 2014; Ambridge and Gold-
berg, 2008; D?browska, 2008, 2013; Erteschik-Shir, 1973; Fodor, 1992; Gibson, 2018;
Goldberg, 2006; Kastner, 2015; Truswell, 2011), Danish (Erteschik-Shir, 1973), Ger-
man (Fanselow et al., 2005; Featherston, 2004; Haider, 2010, also Aaron Doliana,
p.c.), Dutch (Broekhuis and Corver, 2016, 2017, also Annemarie van Dooren, p.c.),
Kashmiri (Bhatt, 1999; Manetta, 2011), Polish (Cichocki, 1983; Witko?, 1995), Ar-
1A bibliographic note: Gibson, 2018 is a report of various projects, including the one indepen-
dently presented as Liu et al., 2019.
161
menian (Anyadi and Tamrazian, 1993), Russian (Antonenko, 2010; Khomitsevich,
2008; Ross, 1967), Hebrew (Kastner, 2015; Preminger, 2014), Hungarian (Kiss, 2002;
Mar?cz, 1991), Georgian (Harris, 1981), and Sundanese (Davies, 2003).
Of these languages, a number are claimed to generally lack A-bar movement
from indicative complement clauses: these include Sundanese, Armenian, and Geor-
gian; Georgian only allows A-bar movement from the complement clauses of three
modal verbs: unda `he wants,' ?euZlia `he can/is able to,' s?. irdeba `he needs it' (Har-
ris, 1981, p. 18). Certain varieties of Polish have a uniform ban on A-bar movement
from indicative complements, while other varieties reported by Cichocki (1983) ap-
parently allow A-bar movement of co `what' (but not other kinds of wh-phrases) from
the indicative complements of m?wi? `say' and powiedzie? `tell.'2 Witko? (1995) re-
ports that Polish also blocks A-bar movement from subjunctive complements in gen-
eral, unless the verb that takes a subjunctive complement can also take a non-nite
complement. Russian imposes a ban on A-bar movement from indicative comple-
ments, while A-bar movement from subjunctive complements is more acceptable,
although the degree of amelioration appears to vary by native speakers.
The other languages surveyed allow A-bar movement from (indicative) com-
plements, although to varying degrees, as the following set of examples show. For
instance, Kashmiri, Dutch, and certain varieties of German are said to allow long-
distance A-bar movement from verbs like `say' and `think,' but not for factive verbs,
verbs such as `know' or `realize.' Finally, English, Hebrew, Scandinavian languages,
2It is unclear to me why there is such a distinction. Cichocki (1983) does not discuss this further
in the paper, and I am not aware of subsequent work that deals with it. Since `what' appears to
be the exception, I take that this restriction is likely to reect idiosyncratic properties of `what,'
such as its syntax or pragmatics.
162
and some varieties of Hungarian have a fairly liberal denition: A-bar movement
from the complements of these verbs is usually reported to be acceptable. There
is also what appears to be at least interspeaker, if not inter-dialectal dierences.
For example, it has been suggested that speakers of northern German dialects do
not allow long-distance A-bar movement at all (Haider, 2010, citing Kvam (1983)),
although Featherston (2004) and Fanselow et al. (2005) report experimental results
suggesting that the dierences might be not as clear-cut as it is usually presented.
For instance, Featherston (2004) presents experimental results showing that speak-
ers of northern and southern German varieties rate long-distance A-bar movement in
similar ways, contrary to generalizations reported in the literature. Mar?cz (1991)
also reports that there are Hungarian speakers who do not accept any kind of long-
distance A-bar movement, although I am not aware of any follow-up work on that,
as in the case of German.
(10) Verbs of saying
a. What1 did Mary say [that John saw t1]? [English]
b. Wen1 hat sie prophezeit [dass er t1 heiraten w?rde]?
whom has she prophesied that he marry would
`Who has she prophesied that he would marry?' [German, from
Haider, 2010 p. 91 ex. 13a, his judgment]
c. Wat zei Jan [dat Marie t1 gelezen had]?
what said Jan that Marie read had
`What did Jan say that Marie had read?' [Dutch, Broekhuis and
Corver, 2016, p. 758, ex. 327a]
d. ?Co1 Janek powiedzia? [?e studenci czytaj? t1]?
what Jan said/told that students read
`What did Jan say/tell that students read?' [Polish, Witko?, 1995 p.
229 ex. 41, his judgment]
(11) Verbs of thinking
a. What1 did Mary think [that John saw t1]?
163
b. Was1 glaubst Du den [dass t1 verantwortlich ist f?r die
what thinks you PRT that responsible is for the
Abweichung]?
deviance
`What do you think is responsible for the deviance?' (Haider, 2010 p.
88, ex. 6a)
c. *Wat denk je [dat Peter t1 gekocht heeft]?
what think you that Peter bought has
`What do you think that Peter has bought?' (Broekhuis and Corver,
2016, p. 1124, ex. ib)
d. *Co1 Janek my?li [?e studenci czytaj? t1]?
what Jan thought that students read
`What did Jan think that students read?' (Witko?, 1995 p. 229, ex.
39, his judgment)
(12) Cognitive factive verbs
a. What1 did Mary realize [that John saw t1]?
b. *Wen1 hat sie eingesehen [dass er t1 heiraten w?rde]?
whom has she realized that he marry would
`Who has she realized that he would marry?' (Haider, 2010, p. 91, ex.
13a, his judgment)
c. *Wat weet je [dat Peter t1 gekocht heeft]?
what knows you that Peter bought has
`What do you know that Peter has bought?' (Broekhuis and Corver,
2016, p. 1124, ex. ib)
d. *?Co1 Iwona wie [?e Tomek je t1 zach?annie]?
what Iwona know that Tom eats greedily
`What does Iwona know that Tom eats greedily?' (Witko?, 1995 p.
229, ex. 38, his judgment)
Even though the literature reports variation in terms of what verbs allow
long-distance A-bar movement, the survey also suggests that the variation is not
completely random. For example, no description of languages I looked at included
manner of speaking verbs in their list of bridge verbs or examples of long-distance
A-bar movement.
I further note that Hebrew, Scandinavian languages, Romance languages, Bul-
garian, and Hungarian (among others) are also reported to allow A-bar movement
164
from embedded wh-questions, although the acceptability and well-formedness of A-
bar movement from embedded wh-questions in Romance and Scandinavian have
been questioned in recent work (Almeida, 2014; Kush et al., 2018; Sprouse et al.,
2016). I will not have much to say about this phenomenon here. To the extent that
these languages freely allow A-bar movement from embedded wh-questions in these
languages (contrary to the core intuition in proposals such as Relativized Minimality
or Shortest Move (Chomsky, 1995; Preminger, 2014; Rizzi, 1990)), this phenomenon
provides yet another example of cross-linguistic variation in bridge verbs. Alterna-
tively, to the extent that it is ill-formed, there are several conclusions one could
draw. First, there are grammatical principles that block A-bar movement from
wh-questions, as envisioned under standard theories of A-bar movement. Alterna-
tively, if one believes that there are no such constraints in these languages, then one
might attribute the low acceptability either to the notion that these verbs are not
bridge verbs or perhaps to some pragmatic principle barring certain movement from
questions.
4.2.4 Concerns about empirical generalizations
The issues mentioned above hint at a larger data quality issue in the literature
on bridge verbs. There is no denitive list of bridge verbs, although there are a
number of proposals on the market about which classes of verbs are bridge verbs
(or are compatible with A-bar movement). However, as far as I can tell, the data
and hypotheses reported in the literature are based on informal judgments, often
165
for a small number of speakers, for a small number of verbs relative to the class of
clause-embedding verbs. In English, and presumably in other languages, there are
likely to be easily dozens of clause-embedding verbs, if not hundreds of them, in the
lexicon (cf. White and Rawlins, 2018, White and Rawlins, 2016, Anand et al., to
appear, among others). However, the largest lists of bridge verbs I have seen are a
list of 21 Hungarian predicates found in Mar?cz, 1991 (p. 159, ex. 12)3 and a list
of 23 Dutch predicates in Broekhuis and Corver, 2016 (p. 760, ex. 331, which is in
turn based on data reported by Hoeksema (2006)).
This is not an issue that is conned to papers that use informal judgments;
papers that report formal experiments also usually look at a fairly small number of
verbs, even though they get judgments from a relatively larger pool of native speak-
ers. For example, Featherston (2004) looked at eight German clause-embedding
verbs, while Ambridge and Goldberg (2008) and D?browska (2013) look at about a
dozen verbs. The one exception that I am aware of is Gibson, 2018/Liu et al., 2019,
which looked at 50 verbs in English.
To compound the problem, existing discussions do not always provide clarity
about what the appropriate generalizations might be. To give an example from
English that I will return to later, the bridge verb status of factive verbs, such
as know, realize, and regret, is disputed. Impressionistically, there is a majority
opinion in the formal linguitics literature that long-distance A-bar movement is
compatible with these verbs; recent proposals that have adopted this position include
3To be precise, the 21 predicates are predicates that are compatible with long-distance A-bar
movement; Mar?cz (1991) (p. 159, fn 5) mentions another 29 predicates that are compatible with
long-distance left dislocation but not with long-distance A-bar movement.
166
de Cuba, 2007, Abrus?n, 2014, Kastner, 2015, among others. However, there are also
dissenting views: Truswell (2011) nds them to be unacceptable. In the experimental
literature, Ambridge and Goldberg (2008) report that native speakers nd them
unacceptable, but Gibson (2018)/Liu et al. (2019) dispute this nding. Broekhuis
and Corver (2016, pp. 13971398) also report some interspeaker variation for the
factive verbs betreuren `regret' and weten `know' in Dutch.
Setting aside for now the dispute in the literature about the acceptability of
A-bar movement from the complement clauses of factive verbs, I note that instances
of A-bar movement from these clauses do occur naturally in English. For example, a
search of the Penn Treebank yielded examples that seem perfectly acceptable, such
as the following:
(13)  Unless the market goes to 19 million units  which we all know
it 's not going to do  we have the inescapable fact that the trans-
plants are adding capacity ,  Mr. Lutz said last month .
4.3 Hypotheses about bridge eects
In this section, I turn my attention to existing proposals about what property
(or properties) of verbs distinguish bridge verbs; in other words, what properties
are responsible for certain verbs being compatible with A-bar movement. I will
abstract away from the theoretical question about what the ultimate nature of these
properties might be. The primary goal of this section is to identify the dimensions
that aect the acceptability of long-distance A-bar movement, and so what is more
important for the discussion in this section is the higher-level generalizations.
167
To avoid misrepresenting the reasoning behind these proposals, I will try to
use examples that are provided in the literature. One should keep in mind that
some of these examples might actually not be sucient for illustrating the relevant
point. For example, it is common to see an unacceptable wh-question used as an
example to show that a verb is incompatible with long-distance A-bar movement. As
mentioned earlier, the low acceptability of this wh-question might have an alternative
explanation: the verb could well be compatible with A-bar movement, but just not
with wh-questions, due to the pragmatics of questions.
4.3.1 Factivity
The rst hypothesis is that factive verbs block long-distance A-bar movement.
These verbs presuppose the truth of their complements; a classic diagnostic checks
if the proposition denoted by the complement is perceived to be true regardless of
whether the verb is negated.
(14) a. Factive verb: regret
(i) John regrets that he saw the movie. ? John saw the movie.
(ii) John doesn't regret that he saw the movie. ? John saw the
movie.
b. Non-factive verb: believe
(i) John believes that he saw the movie. 6? John saw the movie.
(ii) John doesn't believe that he saw the movie. 6? John saw the
movie.
A strong version of the claim about factive verbs is found in descriptions of
German (Haider, 2010) (p. 91) and Dutch (Broekhuis and Corver, 2016) (p. 759), in
which A-bar movement of both arguments and adjuncts is reported to be degraded.
As pointed out above, there is some disagreement about the validity of this claim
168
for English: A-bar movement of arguments (but not adjuncts) is usually claimed
to be acceptable (Abrus?n, 2014; de Cuba, 2007; Gibson, 2018; Honcoop, 1998;
Kastner, 2015; Kiparsky and Kiparsky, 1970; Liu et al., 2019; Postal, 1972); that
said, Truswell (2011), Ambridge and Goldberg (2008), and D?browska (2013) report
that A-bar movement of arguments are degraded in the context of factive verbs; the
latter two providing evidence from formal acceptability judgment experiments.
(15) a. Who1 did John {say/think/(*)regret} [that Mary kissed t1]? (adapted
from Truswell, 2011, p. 178 and 180, exx. 5 and 7, his judgment)
b. ??What1 did she realize [that he saw t1]? (Ambridge and Goldberg, 2008,
p. 351, Table 1, their judgment)
In addition to a classic binary factivenon-factive distinction, more ne-grained
distinctions can be drawn. More specically, verbs can be sorted into four distinct
classes using the following entailment tests (Karttunen, 1971, see also White and
Rawlins, 2018 and references therein):
(16) Factive verb, e.g. regret
a. VERB p entails p.
Mary regrets that she oended John. ? Mary oended John.
b. NEG VERB p entails p.
Mary does not regret that she oended John. ? Mary oended John.
c. if VERB p, q entails p.
If I regret that I have made a mistake, I will let you know. ? The
speaker has made a mistake.
(17) Semifactive verb, e.g. realize
a. VERB p entails p.
Mary realized that the earth is at ? The earth is at.
b. NEG VERB p also entails p.
Mary did not realize that the earth is at ? The earth is at.
c. if VERB p, q does not entail p.
If I realize that I have made a mistake, I will let you know. ?6 The
speaker has made a mistake.
169
(18) Veridical verb, e.g. prove
a. VERB p entails p.
Mary proved that the earth is at ? The earth is at.
b. NEG VERB p does not entail p.
Mary did not prove that the earth is at 6? The earth is at.
c. if VERB p, q does not entail p.
If I prove that the earth is at, I will let you know. ?6 The earth is
at.
(19) Non-factive verb, e.g. think, say
a. VERB p does not entail p.
Mary thinks/said that the earth is at 6? The earth is at.
b. NEG VERB p does not entail p.
Mary didn't think/say that the earth is at ?6 The earth is at.
c. if VERB p, q does not entail p.
If I think that I have made a mistake, I will let you know. 6? The
speaker has made a mistake.
We can then construct hierarchies based on how many of these entailments
a verb might trigger. The most ne-grained is a four-way hierarchy (20), but it is
possible to collapse the factive and semifactive classes (and possibly the veridical
verbs) to produce a more classic binary classication.
(20) Factivity hierarchy
Non-factive > Veridical > Semifactive > Factive
4.3.2 An indexical shift hierarchy
Another theme that surfaces in the literature is that bridge verbs can be dened
in terms of non-entailment semantic properties. For example, as mentioned in the
introduction, it is not uncommon to read that verbs of saying (Mar?cz, 1991, p.
159) or verbs of believing (Haider, 2010, p. 91) are bridge verbs.
Here, I would like to build on this intuition to point out what looks like an
170
implicational hierarchy. The survey of the literature suggests the following implica-
tions: If a language allows A-bar movement across a verb of knowledge, it will allow
A-bar movement across a verb of thought. If a language allows A-bar movement
across a verb of thought, it will allow A-bar movement across a verb of saying. The
reverse implications, however, do not seem to hold true. Just because A-bar move-
ment across a verb of thought is acceptable does not entail that A-bar movement
across a verb of knowledge is. Likewise, there is some limited evidence from Polish
varieties suggesting that just because A-bar movement is acceptable across a verb of
saying does not necessarily entail it is acceptable across a verb of thought. Cichocki
(1983) observes that for these varieties, m?wi? `say' and powiedzie? `tell' allow A-bar
movement of co `what' (but not other kinds of wh-phrases), but verbs like `think'
do not.
(21) a. Cross-linguistic variation in bridge verbs (% = only some varieties of a
language)
Verbs of Verbs of Verbs of
Languages saying thought knowledge
English, Hebrew, Hungarian%, Y Y Y
Scandinavian
English%, Dutch, German%, Kashmiri Y Y *
Polish% Y * *
Armenian, Georgian, German%, * * *
Hungarian%, Polish%, Sundanese
b. (i) Verbs of saying: say, claim, announce, tell, explain, allege, deny,
admit, conrm, agree
(ii) Verbs of thought: think, expect, believe, feel, guess,
(iii) Verbs of knowledge: know, realize, nd out, discover, notice, see
This hierarchy oers an alternative perspective on the status of factive verbs.
In this view, long-distance A-bar movement is not sensitive to factivity per se.
171
Rather, it is sensitive to verb class, and the verbs of knowledge happen to have
factivity (among other properties) as a distinguishing characteristic.4
This hierarchy of verbs might strike one as arbitrary. However, there appears
to be some independent empirical support for it: very similar hierarchies have been
proposed independently for other kinds of phenomena, such as indexical shift and
logophoricity (Culy, 1994; Deal, 2017; Sundaresan, 2018). A natural hypothesis to
pursue is that the same hierarchy is found in various cross-clausal phenomena like
indexical shift and long-distance A-bar movement. To the best of my knowledge,
this is a novel hypothesis about bridge verbs.
Further, for all its arbitrariness, I would like to note that there have been
proposals to derive such a hierarchy from more abstract properties. Speas (2004),
for example, argues that the hierarchy reects how pragmatic features are encoded
on syntactic categories. Since my goal is not to build a case for this particular thesis,
I will refer interested readers to Speas' paper for details and additional references.
For the sake of discussion, I will provide a quick review of the indexical shift
literature.
Generally, indexical shift refers to the fact that in some languages, indexical
lexical items, such as pronouns like I, me, and you do not always have to refer to the
speaker or the listener. In complement clauses of attitude verbs, these indexicals can
co-refer with arguments in a higher clause. An embedded `I' or `me,' for example,
4From my reading of the literature, it is dicult to say with certainty what the membership of
this class might be, again reecting gaps in our data. Cross-linguistically, many have noted that
cognitive factives like `know' and `realize' are not compatible with long-distance A-bar movement.
It is less clear if emotive factives like regret or hate belong to this class. Conceptually, they do not
seem incompatible: intuitively, someone regrets that p appears to presuppose someone knows that
p.
172
can co-refer with the subject of the attitude verb, for instance. An example of this is
illustrated in (22), for the Indo-Iranian language Zazaki (Anand and Nevins, 2004).
(22) HEseni1 va kE Ez1/2 dEwletia.
Hesen.OBL said that I rich.be-PRES
`Hesen said that {I am, Hesen is} rich.' (Anand and Nevins, 2004, p. 21,
ex. 4)
However, in Zazaki, not all clause-embedding verbs shift. As Anand and
Nevins observe, [o]ther attitude verbs (e.g., believe, think, dream), including those
of verbal discourse (e.g., hear, yell) do not allow Ez [the rst person indexical] to
shift (ibid. p. 36, fn. 3).
(23) HEsen1 teRmine kEno kE Ez?1/2 nEwesha.
Hesen believe does that I sick.be-PRES
`Hesen believes that {I am, *Hesen is} sick.' (ibid. p. 36, note 3, ex. i)
Synthesizing descriptions of indexical shift, Sundaresan and Deal argue for an
implicational hierarchy (24) that bears a striking resemblance to the one described
above for long-distance A-bar movement.
(24) Cross-linguistic variation in indexical shift (adapted from Sundaresan, 2018,
p. 28, ex. 54)
Verbs of Verbs of Verbs of
Languages saying thought knowledge
Nez Perce, Japanese% Y Y Y
Slave, Laz Y Y *
Tamil, Telugu, Zazaki, Amharic, Kurmanji Y * *
English, French * * *
It is important to point out that the same caveat about data quality that I
made about bridge verbs also applies here. Like bridge verbs, the data on indexical
shift are based on judgments for a small number of verbs, elicited from what is likely
a small pool of native speakers. It is unclear to what extent the generalizations
173
provided above are robust. For example, contrary to descriptions in the indexical
shift literature, it turns out that indexical shift in Japanese is not widely accepted
(Masato Nakamura, Yohei Oseki, Hisao Kurokami, Yasutada Sudo, Yoshiki Fujiwara
p.c.). Nor is it immediately clear from the literature whether manner-of-speaking
verbs shift. Anand and Nevins are clear that they do not do so in Zazaki, while
Yasu Sudo judged indexical shift to be acceptable for the Japanese verb sasayaku
`whisper, murmur' (p.c.).
Setting aside these issues for now, to the extent that the parallels between
long-distance A-bar movement, indexical shift, and logophoricity are valid, this par-
ticular analysis of bridge verbs can be taken as novel independent evidence for the
hierarchy.5
4.3.3 A comment on manner of speaking verbs
As presented, the two hypotheses presented might create the impression that
the manner of speaking verbs like shout or whisper pattern together with verbs
like claim and say in allowing long-distance A-bar movement, since both types of
verbs are non-factive verbs of verbal communication. This is an undesirable outcome,
since, as far as I can tell, there is no language in which manner of speaking verbs
freely allow long-distance A-bar movement. For these two hypotheses to be viable
5In fact, to the extent that the A-bar movement generalizations are robust, they further suggest
that this hierarchy might have an important role in our theories of syntax and semantics, especially
for theories of cross-clausal phenomena like logophoricity, indexical shift, and long-distance A-bar
movement. After all, there is no a priori reason that attitude verbs might be ordered in this
particular fashion. Likewise, there is also no a priori reason to expect that these cross-clausal
phenomena, which have very distinct morphosyntactic and semantic proles, might be subject to
the same hierarchy.
174
hypotheses about bridge verbs, we require some independent principle that renders
manner of speaking verbs incompatible with long-distance A-bar movement.
As one might expect, there are several options for implementing this. One
straightforward way of doing so is to stipulate that there is some general syntactic
constraint against A-bar movement from the complement clause of a manner of
speaking verb. This would provide an explanation why there appears to be no
language in which manner of speaking verbs are compatible with long-distance A-
bar movement. Alternatively, one might posit some kind of discourse or pragmatic
principle.
Whatever the correct implementation might be, it should be clear from the
discussion that the restriction is likely to be related to the fact that manner of
speaking verbs encode manner semantics in a way that the other verbs do not
(Beavers and Koontz-Garboden, 2012; Rappaport Hovav and Levin, 2008). In this
section, I present diagnostics proposed by Beavers and Koontz-Garboden (2012)
that suggest that manner of speaking verbs form a distinct class from the other
non-manner of speaking verbs: specically, the distinction between these two classes
parallels that between Beavers and Koontz-Garboden's manner verbs and result
verbs. I have adjusted the descriptions of these diagnostics slightly to make them
more precise.
(25) Diagnostics for result and manner verbs (Beavers and Koontz-Garboden,
2012)
a. Result verbs cannot be followed with a denial of results.
(i) #Shane just broke the vase, but it is not broken. (p. 337, ex. 9a)
(ii) Tracy just swept the oor, but nothing is dierent about it. (ex.
11a)
175
b. Result verbs do not allow object deletion
(i) Intransitive use
1. *All last night, Kim broke. (ex. 16b)
2. All last night, Kim scrubbed. (p. 339, ex. 14b)
(ii) Out-prexation
1. ??Kim outbroke the other vase-smasher. (ex. 19a)
2. Cinderella outscrubbed her stepsisters. (ex. 15b)
c. Result verbs impose restrictions on resultatives (unsubcategorized ob-
jects)
(i) Cinderella scrubbed her knees sore.
(ii) *Kim dimmed her eyes sore.
d. Manner verbs impose selectional restrictions (on subjects)
(i) {John/The hammer} broke the vase with a hammer. (based on
p. 344, ex. 31)
(ii) {John/*The sti brush} scrubbed the oor. (based on ex. 32)
e. Manner verbs do not allow denial of action (by the subject) (NH: I nd
the contrast less clear; to the extent that the contrast its hard to get,
this is a less convincing diagnostic)
(i) Jim destroyed his car, but didn't move a musclerather, after he
bought it he just let it sit on his neighbor's lawn on cinder blocks,
untouched, until it disintegrated! (p. 346, ex. 36)
(ii) #Jim ran/jogged/blinked, but didn't move a muscle. (p. 345 ex.
35)
f. Manner verbs that denote complex actions are durative. (NH: this fails
to identify manner verbs that denote simple actions)
(i) It took Jim ve minutes to break the window. (window breaking
must occur at the end of ve minutes; p. 348, ex. 42a)
(ii) It took Jim ve minutes to clean the table. (cleaning can take
place throughout the ve minutes; ex. 42b)
These diagnostics can be applied to verbs of saying and manner of speaking
verbs; I will use claim and shout as examples of the respective classes. As the
examples in (26) show, some of these diagnostics, such as the denial of result and the
non-punctuality diagnostics are inconclusive. However, verbs of saying and manner-
of-speaking verbs behave quite dierently with respect to the object diagnostic,
the resultative diagnostic, and the subject selectional restriction diagnostic. For
comparison, I also apply the same diagnostics to verbs of thought and knowledge,
176
taking believe and realize as examples.
(26) Diagnostics for result and manner verbs, as applied to various attitude verbs
a. Result verbs cannot be followed with a denial of results:
(i) Drew realized / believed / claimed that the earth was at, but
#there was no such realization / no such belief / no such claim
on his part.
(ii) Drew shouted that the earth was at, but #there was no shout.
b. Result verbs do not allow object deletion
(i) Intransitive use
1. *Drew realized / believed / claimed. (excluding null comple-
ment anaphora, which have an anaphoric reading.)
2. Drew shouted. (no anaphoric reading)
(ii) Out-prexation
1. *Drew out-knew / out-believed / out-claimed Mary.
2. Drew out-shouted Mary.
c. Result verbs impose restrictions on resultatives (unsubcategorized ob-
jects)
(i) *Drew claimed himself hoarse. (e.g. he made so many claims that
he lost his voice)
(ii) Drew shouted himself hoarse.
d. Manner verbs impose selectional restrictions (on subjects)
(i) {John/The paper} claimed that the earth was at. (cf. Anand
et al., to appear; Anand and Hacquard, 2009, 2014)
(ii) {Drew/*Drew's vocal tract} shouted.
e. Manner verbs do not allow denial of action by the subject
(i) The paper claimed that the earth was at. (since papers are
inanimate, there is no action by the subject to be denied. But
this test is problematic for mental state verbs like think, since
these verbs arguably do not describe any action.)
(ii) #Drew shouted, but didn't move his articulatory muscles.
f. Manner verbs that denote complex actions are durative: n/a; doxastics
are stative or achievements (e.g. realize), while communicative verbs
are
4.3.4 Bridge eects are sensitive to ne-grained semantic distinctions
There is of course no a priori reason to think that long-distance A-bar move-
ment must be sensitive to relatively coarse-grained lexical semantic distinctions like
177
the ones described above. There are proposals on the market that divide the set
of attitude verbs in a much more nuanced way. Here, I will discuss two distinct
classication systems: Anand et al., to appear and Levin, 1993, which both cover a
relatively large set of attitude verbs in English. While these researchers do not dis-
cuss bridge eects at all, a number of their proposed verb classes correspond rather
neatly to verb classes that have been conventionally held to be incompatible with
long-distance A-bar movement, such as factive verbs or manner of speaking verbs.
In their survey of more than 400 clause-embedding verbs, Anand et al. pro-
posed three dimensions for classifying attitude verbs: their general lexical semantics
(doxastic (mental states), communicative, . . . ), whether they are factive or not,
and whether they allow non-agentive repository of information subjects. The last
property refers to the fact that a number of these verbs have the option of having an
inanimate subject, as long as that subject denotes a repository of information and
it is understood that there is an agent responsible for the creation of that repository
of information. For example, book and article are both agentive repository of infor-
mation nouns, and can appear as the subject of a verb like state (27a). In contrast,
they cannot occur as the subject of shout (27b).
(27) a. {Mary / The book} stated that the earth was at. [Agentive
repository of information subject possible]
b. {Mary / *The book} shouted that the earth was at. [Agentive
repository of information subject impossible]
In this classication system, factive verbs form a natural class. While there is
no dedicated class for manner of speaking verbs, I note that these verbs form the bulk
of non-factive communicative verbs that are incompatible with agentive repository
178
of information subjects (the remainder, as noted in their footnote 3, being a small
but semantically diverse set of verbs, including phrasal verbs like add in, put in,
verbs of publishing (for lack of a better term) such as record, publish, editorialize,
and the high-frequency tell.).
Levin proposes an even ner classication system, based on a verb's seman-
tics and syntactic properties. (To the extent that there is a principled correlation
between a verb's syntax and its semantics, Levin's system is ultimately organized
along semantic lines.) In this system, attitude verbs appear in seven major semantic
classes, such as verbs with predicative complements, verbs of perception, verbs
of desire, verbs of social interaction and so on. Verbs within each semantic class
are then sorted into smaller subclasses in part based on the syntactic frames that
they can appear in.
In this system, manner of speaking verbs form a distinct class of their own
(Section 37.3, pp. 204206). However, there is what appears to be a major omission
in this system: Levin (1993) does not mention factive verbs at all. Nor does she
discuss verbs that take embedded question as complements. This particular subcat-
egorization property is often hypothesized to be a feature that characterizes factive
verbs (see e.g. Spector and Egr?, 2015; White and Rawlins, 2018).
4.3.5 Bridge eects reect backgroundedness eects
Another approach, often found in functionalist or discourse-based discussions,
claims that bridge eects follow from two independent properties: (i) certain verbs
179
foreground or draw attention to the proposition denoted by their complement,
and (ii) there is a general constraint against A-bar movement from backgrounded
complements (Ambridge and Goldberg, 2008; Deane, 1991; Erteschik-Shir, 1973;
Goldberg, 2006).
Despite its conceptual simplicity, this approach faces a major problem: it has
always been dicult to make precise in theoretical terms what it means for some-
thing to be foregrounded (or backgrounded). There are also few clear diagnostics
for backgroundedness. The preferred diagnostic that Goldberg and colleagues use
is negation, where an embedded clause is judged to be foregrounded if negation
of the main clause entails that the proposition denoted by the embedded clause is
negated (28). One problem with this diagnostic, especially in the domain of attitude
verbs, is that it introduces a confound, since a proper subset of verbs like think are
neg-raising, and so easily imply the negation of the embedded proposition. Conse-
quently, it is unclear whether a verb passes this diagnostic because it foregrounds
its complement or because it is neg-raising.
(28) V foregrounds p . . . to the extent that NEG V p ? ? p
a. She didn't think that he left. ? ? He didn't leave. [that he left is
foregrounded]
b. She didn't notice that he left. ?6 He didn't leave. [that he left is
backgrounded.]
(adapted from Ambridge and Goldberg (2008), p. 368, exx. 2627)
Another potential issue is about the range of data this approach is intended to
cover. As I read it, Ambridge and Goldberg, 2008 is an account of the circumstances
in which long-distance A-bar movement to form wh-questions is acceptable: . . .
180
even when questions appear to be semantically appropriate, there are contraints on
what can count as a question (p. 358); their examples and critical sentences all
involve wh-questions. However, Goldberg (2006), which also put forward essentially
the same backgroundedness proposal, discusses examples featuring wh-questions and
relativization. For concreteness, I will assume that the proposal is intended to cover
a broader range of A-bar dependencies, as implied in Goldberg, 2006.
4.3.6 Exemplar eects
In this hypothesis, the acceptability of long-distance A-bar movement for a
given verb is ultimately related to how semantically similar that verb is to some
of the most common bridge verbs like think and say. The intuition is as follows:
like most verbs, bridge verbs follow a power law distribution in naturally-occurring
context (D?browska, 2013). Hence, when asked to judge (or produce or process)
novel long-distance A-bar movement structures, speakers implicitly compare them
to high-frequency forms involving think and say, which are the most familiar.
This idea is easily modeled as a processing eect, as it is usually described in
the usage-based literature (D?browska, 2004, 2008, 2013; Verhagen, 2006): Speakers
create templates based on these high-frequency forms and use them for production
and comprehension, and processing any other kind of structure necessarily requires
modication and incurs a cost.6
6This is not to say that exemplar eects inuence acceptability only through processing factors,
such as production or comprehension, although it is certainly relatively straightforward to model
these eects that way. For example, it is logically possible that the degree of well-formedness of
long-distance A-bar movement exhibits exemplar eects.
181
4.3.7 A frequency-based reductionist account
In recent work, Gibson (2018)/Liu et al. (2019) put forward a reductionist
argument, in which the dierences in acceptability in A-bar movement associated
with bridge eects simply reect dierences in the frequency of verbs appearing with
clausal complements.
This account is built around the following observation: A-bar movement is
the most acceptable when it occurs with bridge verbs and degraded when it ap-
pears with manner of speaking verbs, but this correlation happens to be confounded
with frequency: bridge verbs appear frequently and often take nite clausal com-
plements, while a prototypical non-bridge verb, such as a manner-of-speaking verb,
is less frequent and can appear without clausal complements. Liu et al. (2019) pro-
vide experimental evidence for a set of 48 English verbs to support this argument.
Featherston (2004) makes a similar observation for German, although with a much
smaller set of verbs.
Conceptually, for English at least, this frequency-based analysis is not en-
tirely unmotivated, since frequency is widely known to inuence language process-
ing, which in turn impacts acceptability. This idea manifests itself in several ways
within psycholinguistics, in the form of priming and satiation, where exposure in the
very short term leads to more frequent production and ease of processing, and/or
greater acceptability. This idea can further be seen as a lexicalized version of Pearl
and Sprouse's account (2013a; 2013b), which models acceptability of long-distance
A-bar movement as a function of frequency of observing long-distance A-bar move-
182
ment structures in one's linguistic experience.
That said, there are three sets of issues associated with this frequency-based
treatment.
First, it ties the acceptability of a wh-dependency to two elements within it:
a verb and its clausal complement. One obvious question is why these two elements
should matter more than other elements for the same dependency. For example,
why does the aspect of the verb, or the tense of the embedded clause, seem to not
matter? Another question is why other kinds of dependencies do not seem to be
sensitive to these elements. For example, the dependency between a pronoun and its
antecedent does not seem to be sensitive to these elements (29a). The dependency
between a reexive and its antecedent, on the other hand, is sensitive to the mere
presence of a clausal complement; the frequency of the verb that takes the clause as
a complement does not matter (29b).
(29) a. Every girl1 {thought/said/whispered/yelled/regretted} that John liked
her1.
b. *Every girl1 {thought/said/whispered/yelled/regretted} that John liked
herself1.
Second, to rephrase an argument of Featherston's (2004), correlation is not
causation. There might be some other linguistic property that is simultaneously
responsible for the acceptability of long-distance A-bar movement and frequency.
Featherston considers two candidates: phonological weight (as measured in number
of syllables) and semantic weight (although he does not provide a denition of it).
For the sake of discussion, suppose that being semantically light (or phonologically
light) might mean that a verb can be used in more contexts than a heavier verb,
183
which contributes to the frequency of use. That said, Featherston is also careful to
note that proving a causal relation between weight and frequency is dicult.
4.4 Acceptability judgments for 100 attitude verbs
The proposals make distinct predictions that one can test empirically. To
do so, I ran an acceptability judgment experiment on Amazon Mechanical Turk to
collect ratings for 100 verbs. I recruited a total of 83 participants, all self-identied
native speakers of American English based in the United States.
There were two main goals in conducting this study.
The rst goal is to remedy an empirical gap in our understanding of bridge
verbs and bridge eects. Although the eld has long recognized that long-distance
A-bar movement is not equally acceptable across all clause-embedding verbs, we
know relatively little about the actual range of variation in acceptability.
Second, as reviewed above, there are a number of existing claims about the
nature of bridge eects that are made on the basis of relatively few verbs. The large
number of attitude verbs in a language like English means that it is not feasible to
test these claims through informal introspection, one verb at a time. A larger scale
formal study, as described here, allows a more comprehensive evaluation of these
claims.
More specically, to the extent that one of these claims is a better account of
bridge eects, a statistical model of acceptability judgments that is built around this
claim should outperform statistical models based on alternative claims. Adopting
184
this quantitative approach thus allows us to adjudicate between various claims and
to identify the dimension(s) that truly have an impact on the acceptability of long-
distance A-bar movement. Doing so will help clarify the kind of learning a native
speaker might have to do (or not have to do) regarding the syntax of long-distance
A-bar movement.
4.4.1 Methods
4.4.1.1 Verb selection
First, I identied a set of verbs that take declarative complement clauses in
naturally occurring contexts; this was so that there would be some assurance that
the relevant verbclause frame would not be completely unfamiliar to participants.
To do so, I wrote a Python script that identied declarative nite clausal sisters
(annotated with the label SBAR) of (phrasal) verbs in the Penn Treebank (approx-
imately 251,000 tokens).
To minimize the likelihood of false positives in the identication of comple-
ment clauses in the Penn Treebank, I manually checked verb lemmas that appeared
with a declarative nite sister nine times or fewer and corrected the annotation if
necessary. I assume that all verb lemmas that appeared at a higher frequency take
a complement clause.
The semi-automated process yielded 20,756 instances of (phrasal) verbs ap-
pearing with a complement clause-like frame, distributed across 279 unique verb
lemmas. These verbs have a power law-like distribution. The most common 20
185
verbs, listed in (30), appear a total of 16,817 times (81%), while 87 verbs appear
only once.
(30) Most common 20 verbs appearing with nite declarative clausal comple-
ments
say, think, guess, know, believe, tell, feel, nd, hope, note, mean, add, show,
suggest, wish, realize, see, indicate, seem, report
One of the goals of this study is to replicate claims about frequency and bridge
eects made by Gibson (2018)/Liu et al. (2019). To that end, I selected 40 of their
48 verbs; this set includes some of the most frequent verbs, such as say and think,
as well as most of their manner of speaking verbs. These 40 verbs also appeared in
the Penn Treebank sample.
The remaining 60 verbs were selected from the Penn Treebank sample in a
pseudo-random fashion, to exclude raising verbs like seem.
Finally, I checked to make sure that the 100 verbs chosen also occur with a
complement clause in the British National Corpus, a corpus that is 24 times the size
of the Penn Treebank (see below for rationale).
The full set of 100 verbs is listed in (31).
(31) accept, admit, advise, arm, alert, allow, argue, assert, assume, attest, be-
lieve, boast, brag, claim, comment, communicate, complain, concede, con-
cur, conde, conrm, counter, declare, decree, deny, disagree, disclose, dis-
cover, doubt, ensure, expect, expound, fear, feel, gure, forecast, foresee,
forget, fret, grumble, guarantee, guess, hate, hear, hint, hope, hypothesize,
indicate, infer, inform, insinuate, insist, intimate, know, laugh, learn, like,
maintain, mean, muse, mutter, notice, notify, observe, pledge, presume,
proclaim, promise, propose, protest, prove, read, realize, regret, remark,
remember, repeat, report, respond, reveal, say, scream, shout, sigh, snap,
specify, state, suspect, swear, testify, theorize, think, urge, venture, verify,
vow, warn, whisper, worry, write
186
4.4.1.2 Obtaining better frequency estimates
As mentioned above, a frequency-based reductionist account posits that ac-
ceptability varies directly with how frequently a verb occurs with a complement
clause  for ease of reference, I will also call this the frequency of the verb-S
frame. To obtain better estimates of this frequency, I used the freely download-
able British National Corpus (BNC, BNC Consortium, 2007) to obtain frequencies
for the verbs that were found to occur with at least one nite complement clause
in the Penn Treebank (PTB). Although verb-S frequencies can be computed easily
from the PTB, it is a relatively small corpus, raising concerns that estimates for
low-frequency items might be too coarse. Using the BNC, which consists of about 6
million sentence units and is thus about 24 times the size of the PTB (about 251,000
sentence units) reduces the analytical problem posed by coarse estimates.
The BNC was chosen over other large, free-to-use corpora like the Corpus of
Contemporary American English and the Google Books corpus for two reasons (note:
Liu et al. (2019) used the Google Books corpus to obtain frequency estimates). First,
the BNC is available for download and hence is available for customized analysis;
the other two datasets must be accessed via online interfaces. Because all three
corpora are collections of sentences with lemma and part of speech information but
no other structural information, there is no easy way to use the online interface to
obtain good estimates of a verb taking a complement clause. The second reason
is that the BNC contains both spoken and written registers (10% and 90% of all
words, respectively). This makes it more balanced than the Google Books corpus.
187
These two features were judged to outweigh one drawback of using the BNC:
the fact that the BNC is a corpus of British English varieties, while the PTB data
are based on American English. In my judgment, the syntax of clause-embedding
verb in these varieties is likely to be very similar, and so dialectal dierences are
likely to be small.
To obtain estimates of frequency of complement clauses, for each attitude verb,
I wrote a script to randomly extract a sample of sentences containing the verb.
Each sentence was then automatically parsed using the Stanford Parser (Klein and
Manning, 2003; Manning et al., 2014) into machine-readable syntax trees. A second
script then analyzed each tree to identify the sisters of the verb, and if one of them
was a clause, identify certain syntactic characteristics of the clause: whether it was
a complement clause or an adjunct clause, whether it was nite or not, whether it
was an embedded question or a declarative, and so on. Granted, the Stanford Parser
is not perfectly accurate. I will, however, make the simplifying assumption that the
estimates are reliable and broadly reect how frqeuently these verbs appear with
complement clauses.
To speed up the sampling, parsing, and analysis, I divided the set of verbs into
four smaller sets based on their base frequencies in the BNC. Verbs whose lemmas
are recorded to appear at least 100,000 times in the list created by Adam Kilgarri
(https://www.kilgarriff.co.uk/bnc-readme.html) were randomly sampled at
a 5% rate, those whose lemmas appear 10,00099,999 times at 10%, those whose
lemmas appear 8009,999 times at 25%, and verbs less than 800 times at 100%
(i.e. there was no random sampling for these very low frequency verbs). The fre-
188
Figure 4.1: Frequencies of verbs with complement clauses in the Penn Treebank and
the British National Corpus
9
6
3
0
0.0 2.5 5.0 7.5
Log of verb count in Penn Treebank
quency gures obtained through this process were then scaled to obtain corpus-level
estimates.
One might (rightly) be concerned that, because of dialectal dierences, fre-
quency estimates might diverge signicantly for less frequent items in both datasets.
An analysis nds no cause for concern: there is a high degree of correlation between
both sets of estimates (Figure 4.1); log frequencies for one corpus is strongly corre-
lated with log frequencies in the other (Spearman ? = 0.836, p < 0.001).
4.4.1.3 Identifying the other independent variables
Factivity measures I used the diagnostics discussed in Section 4.3.1 to code a verb
as factive, semifactive, veridical, or non-factive. Since the particular implementation
discussed above excluded manner of speaking verbs, all manner of speaking verbs
were treated as a separate manner of speaking class, instead of being coded as
189
Log of verb count in British National Corpus
non-factive.
Anand et al.'s classication system I maintained their binary factivity distinction,
while making a few minor corrections. (Specically, I reclassied fret as an emotive
verb, and not a communicative verb, and treat admit and hear as non-factive.)
Note that in this system, a verb can appear in multiple categories: guarantee, for
instance, is treated as both an inferential verb and a communicative verb.
Indexical shift hierarchy Anand et al.'s system was also used to determine what
categories a verb might belong to in the indexical shift hierarchy. First, I coded
all factive verbs (emotive factives like regret and cognitive factives like notice) as
verbs of knowledge. All non-factive doxastics were coded as verbs of thinking, and
all non-manner of speaking communicative verbs were coded as verbs of saying. All
other verbs  non-factive emotives, inferential, and manner of speaking verbs 
were coded separately as emotive, inferential, and manner of speaking verbs.
Levin's classication system I adapted the classication system in Levin, 1993
(organized as chapters and sections in the book) to produce the following verb
classes (32); note that a verb can appear in multiple classes due to polysemy or by
virtue of their subcategorization frames:
(32) Verb classes based on Levin's classication system
a. Mental state verbs: Chapter 29 Verbs with predicate complements;
these are verbs that can appear with small clause complements or non-
nite complements.
b. Realize verbs: This refers to a set of mental state verbs that are distin-
guished by being able to appear with an embedded question, like realize
and know. Levin does not use embedded questions as a criterion.
190
c. Perception verbs: Chapter 30
d. Emotive verbs: Chapter 31 Psych-Verbs and 32 Verbs of desire;
these two chapters are combined due to their similar semantics
e. Judgment verbs: Chapter 33
f. Social interaction verbs: Chapter 36
g. Various classes of communication verbs: Chapter 37
(i) Manner (37.3): This includes all manner of speaking verbs, as
well as write, which describes an instrument of communication.
(ii) Chitchat verbs (37.6)
(iii) Say verbs (37.7)
(iv) Complain verbs (37.8)
(v) Advise verbs (37.9)
(vi) Subjunctive verbs: This refers to a set of communicative verbs
that describe commands and orders, such as insist ; syntactically
they can occur with a subjunctive clause. Levin does not distin-
guish this class or use subjunctive clauses as a criterion.
h. Non-verbal expression verbs: Chapter 38 Verbs of sounds made by
animals
i. Verbs of appearance: Chapter 48 Verbs of appearance, disappearance,
and occurrence
j. Other verbs: this includes a small set of miscellaneous verbs such as
Measure verbs which are scattered across a relatively diverse set of
verb classes. I consolidated these to avoid spuriously creating verb
classes for the statistical analysis.
Backgroundedness measures I use an existing set of entailment metrics for mea-
suring backgroundedness, as dened by Goldberg (2006) and colleagues. In a large
scale study of the factive and veridical properties of over 300 English attitude verbs
by White and Rawlins (2018), participants were presented with a set of sentences
like (33a). 10 participants rated each sentence for acceptability on a 7-point Likert
scale and answered a question that was based on the complement clause, as in (33b).
(33) a. Someone {thought/didn't think} that a particular thing happened.
b. Did that thing happen? [Options: Yes, Maybe or maybe not, No]
Of interest are the sentences in which there was matrix negation  Someone
191
didn't think that a particular thing happened  and the responses to the associated
question. To the extent that participants responded `No,' we may conclude, fol-
lowing Ambridge and Goldberg (2008), that the verb foregrounds its complement
clause. In the data that White and Rawlins make available, there are raw responses
(Yes, Maybe or maybe not, and No) for each participant and each verb, as well
as an average response for each verb that has been normalized to adjust for individ-
ual dierences, such that the more likely participants gave a negative response, the
more negative the normalized score is.
To estimate the backgroundedness of a verb, I did the following: First, I
excluded participants that White and Rawlins indicated should be excluded from
analysis based on their native languages. Next, I extracted only the test sentences
like the one in (33a), where the matrix verb was in the active voice, was negated,
and took a nite complement clause. For each verb, I calculated the percentage
of `no' responses and extracted the normalized responses. The greater the extent
of backgrounding (and hence by hypothesis the lower the acceptability rating), the
lower the rate of `no' responses, and the higher the normalized score.
Verb similarity measures The exemplar eects hypothesis predicts that the accept-
ability of long-distance A-bar movement for a verb depends on how semantically
similar the verb is to the most common bridge verbs. Following Ambridge and
Goldberg (2008); D?browska (2013), I operationalize this by measuring the simi-
larity of a given verb to two high-frequency bridge verbs: think and say. There
are many options on the market for measuring verb similarity, and I selected the
192
following three approaches for doing so.
The rst approach is to use WordNet (Fellbaum, 1998; Miller, 1995). The
Python Natural Language Toolkit (NLTK) (Bird et al., 2009) interface generates a
suite of similarity measures for any given pair of verbs (or more specically, verb
readings, since most verbs are polysemous and so have multiple readings). To iden-
tify the appropriate reading for a given verb, I checked the denitions, examples,
and sentence frames given by WordNet for each reading of a verb. I consider a
sense relevant if (i) the sentence frames listed for that reading include one where
the verb appears with a that CLAUSE (i.e. a complement clause), or (ii) if the
examples listed for that reading include one with a complement clause, or (iii) if
the denition was judged to be compatible with a complement clause. This process
yielded between 1 to 5 readings for each verb.
From looking at the denitions given for say and think, I identied two relevant
readings for each verb; call these four readings the baseline readings.
I then wrote a Python script to generate similarity scores for each reading of
a verb and each of the baseline readings. From this set of similarity scores, I took
the highest similarity score.
From the set of six similarity scores available in the NLTK WordNet pack-
age, I selected two: a path similarity score, which measures similarity in terms of
WordNet's hierarchy of concepts, and the Jiang-Conrath similarity score (Jiang and
Conrath, 1997). The denition of the Jiang-Conrath score makes reference to the
`information content' of each reading relativized to a corpus (Resnik, 1995). For the
purpose of calculating this score, I used the pre-compiled information content data
193
for the Penn Treebank provided in the WordNet package.
The second approach is to use word vectors (more precisely, lemma vectors)
as a proxy for semantic representations; I calculate similarity using cosine similarity
between these vectors. To do so, I used the Penn Treebank as a corpus; each sen-
tence was converted to lower case, stripped of null syntactic elements (e.g. traces),
punctuation, and stop words as dened by the NLTK package, and then lemma-
tized. I trained a Word2Vec model (Mikolov et al., 2013; ?eh??ek and Sojka, 2010)
on this lemmatized corpus. I included all lemmas in the training process, instead of
dropping very low-frequency items, because doing so would have eliminated some
attitude verbs of interest. The training process generates a word vector for every
lemma in the lemmatized corpus.
For each of the verbs, I calculated cosine similarity values for that verb's vector
relative to the vectors of say and think. The higher cosine similarity value was taken
to be the verb's similarity score.
The third approach is to use syntactic frames as a proxy for semantic represen-
tations. Specically, I used the MegaAttitude dataset (White and Rawlins, 2016),
which consists of acceptability ratings assigned to 50 distinct syntactic frames for
1000 clause-embedding verb verbs. Adapting the normalization procedure described
by White and Rawlins, I normalized the ratings for each participant using a Ridit
transformation. Mean Ridit-transformed ratings were then calculated for each frame
for each verb, yielding a vector of ratings for each verb.
Again, for each of verb of interest, a cosine similarity value was computed for
that verb's vector and the vector for say and think. The higher similarity value was
194
taken to be the verb's similarity score.
4.4.1.4 Collecting acceptability ratings
Minimal pairs for each verb were constructed. To ensure that ratings are
broadly comparable, the minimal pairs were all based on the following template:
(34) a. Matrix A-bar movement (baseline): This is the NOUN that VERBed
that {no one/someone/everyone} (AUX) VERB(ed/en) the ADJEC-
TIVE NOUN.
b. Long-distance A-bar movement: This is the ADJECTIVE NOUN that
the NOUNVERBed that {no one/someone/everyone} (AUX) VERB(ed/en).
I chose to construct sentences containing relative clauses rather than wh-
questions to avoid confounds with the pragmatics of questions, following Gibson
(2018)/Liu et al. (2019). Questions are usually posed to acquire information from
one's conversation partner, and in the absence of context, it might be odd to ask
that person about a third individual's attitudes or statements (see Gibson, 2018
for similar comments). The contents of a relative clause, in contrast, is usually
perceived to be common knowledge.
Subjects of the attitude verbs were uniformly human: the lawyer, the teacher.
The exception is mean, which had a non-human subject (the outage). The reason
for making an exception for mean is because when it appears with a human sub-
ject, it triggers a presupposition that there is another proposition under discussion
introduced by the referent of the subject. For example, Mary meant that John is
late is felicitous only if Mary had made another statement previously. To avoid this
complication, I decided to use a non-human subject, which does not trigger such a
195
presupposition.
To ensure that any low acceptability can be reasonably attributed to long-
distance A-bar movement and not due to the choice of lexical items, stimuli sentences
were constructed to maximize plausibility in the matrix A-bar movement condition.
For example, since a verb like testify is typically used in legalistic situations, the
relevant minimal pair for such a verb might feature lexical items like lawyer and
witness. Similarly, since one usually whispers something that is not intended to be
publicly known, the matrix A-bar movement sentence for whisper was deliberately
constructed to suggest something conspiratorial: This is the actor who whispered
that everyone hated the arrogant director.
Each participant rated 10 minimal pairs along a 7-point Likert scale, where
1 indicates lowest acceptability and 7 the greatest. In addition, they also rated 20
sentences that include preposition stranding in implausible contexts (35a), ungram-
matical preposition stranding (35b), and A-bar movement from a complex NP island
(35c), which served as llers and control sentences.
(35) a. This is the key witness the cook washed a meal for. [Preposition
stranding in implausible context]
b. This is the war for independence the writer wrote novels after. [Un-
grammatical preposition stranding]
c. This is the dicult client the manager regretted the fact that no one
contacted. [Complex NP Island Constraint violation]
Before starting the experiment proper, each participant did 6 practice sen-
tences for familiarization. Participants were paid $4.
196
4.4.2 Data analysis
For graphing purposes, I normalized the acceptability ratings for each partic-
ipant using a z-score transformation.
For statistical analysis, I also used z-scores. I used the `LME4' package in R
to t linear mixed-eects models for the acceptability ratings, using a maximum
likelihood estimation procedure (Bates et al., 2015; R Core Team, 2019). The xed
eects in these models were the factors outlined above: verb classes, frequency
measures, verb similarity measures, and so on. I included random intercepts for
participant and for verb. (I also tried more complex models with random intercepts
and slopes for participant and for verb, but they were either too complex, failed to
converge, or ran into boundary t issues that suggest over-tting problems.)
One important reason for using a linear mixed eects model is because one
can use mixed eects models, but not other models, to calculate two useful values:
a marginal R2 and a conditional R2 (Nakagawa and Schielzeth, 2013). The former
value measures the variance that is explained by the xed factors in the model, in
this case, factors such as semantic class, frequency, verb similarity, and so on. The
latter value measures the variance explained by both xed and random factors, and
so provides a sense of the overall quality of t. For this analysis, I used the `MuMIn'
package in R (Barton, 2019).
197
Figure 4.2: Normalized acceptability of matrix A-bar movement vs. long-distance
A-bar movement
Normalized acceptability
of wh?movement
1.0
0.5
0.0
?0.5
?1.0
Matrix wh?movement Long?distance wh?movement
Wh?movement
4.4.3 Analysis 1: Absolute acceptability of long-distance A-bar move-
ment for the fuller set of verbs
Figure 4.2 shows that normalized acceptability ratings for the baseline ma-
trix A-bar movement condition is generally higher than the long-distance A-bar
movement condition, and clustered more closely. Both properties suggest that the
lexicalizations designed were largely successful at maximizing sentence plausibility
and hence acceptability for this condition.
198
Acceptability (z?score)
Average z-score acceptability judgments long-distance A-bar movement for all
100 verbs are given in (36):
(36) Verbs sorted by average absolute acceptability of long-distance A-bar move-
ment (normalized by participant)
Verb z-score Verb z-score Verb z-score
mean -1.12 concede 0.06 guess 0.37
intimate -1.06 disclose 0.06 declare 0.39
snap -0.81 concur 0.07 propose 0.41
scream -0.78 reveal 0.09 know 0.42
allow -0.76 discover 0.09 advise 0.42
respond -0.7 alert 0.1 verify 0.42
counter -0.56 hear 0.12 ensure 0.43
sigh -0.5 promise 0.12 realize 0.43
attest -0.3 learn 0.13 arm 0.44
shout -0.28 insinuate 0.14 argue 0.45
inform -0.28 accept 0.14 indicate 0.45
repeat -0.26 notice 0.14 gure 0.47
like -0.25 protest 0.17 conrm 0.47
hate -0.24 read 0.17 report 0.51
disagree -0.23 forecast 0.18 urge 0.52
venture -0.17 doubt 0.18 guarantee 0.53
testify -0.1 state 0.2 worry 0.53
brag -0.1 grumble 0.2 expect 0.54
whisper -0.09 proclaim 0.21 specify 0.58
laugh -0.09 communicate 0.23 forget 0.59
remember -0.09 write 0.25 suspect 0.59
observe -0.05 warn 0.26 think 0.6
conde -0.04 foresee 0.27 vow 0.62
pledge -0.03 fret 0.29 feel 0.63
admit -0.03 complain 0.29 remark 0.68
deny -0.03 mutter 0.29 decree 0.69
expound -0.02 muse 0.3 assume 0.72
regret -0.01 assert 0.3 believe 0.74
comment 0.01 theorize 0.31 fear 0.76
notify 0.02 hint 0.32 presume 0.8
maintain 0.03 hypothesize 0.33 insist 0.83
prove 0.04 say 0.34 hope 0.95
boast 0.05 swear 0.37
infer 0.06 claim 0.37
Weighted mean z-score for all 100 verbs = 0.16
So that the models could be compared with each other, data for 15 verbs had
199
to be dropped from the ordinal regressions reported in (37): these 15 verbs were not
analyzed by White and Rawlins (2016) in their MegaAttitude dataset.
(37) Comparisons of models of absolute acceptability of long-distance A-bar
movement
Marginal Conditional
Model AIC logLik R2 R2
Accounts based on linguistic distinctions
Factivity 1243 -613 0.0602 0.3324
Indexical shift hierarchy 1240 -610 0.0778 0.3326
Anand et al. classication 1236 -606 0.0970 0.3324
Levin, simplied 1258 -612 0.0685 0.3352
Backgroundedness
Backgroundedness, percentage score 1253 -622 0.0092 0.3346
Normalized backgroundedness 1256 -623 0.0001 0.3344
Verb similarity accounts
Path similarity 1253 -621 0.0097 0.3340
Jiang-Corvath/PTB similarity 1251 -621 0.0145 0.3331
Cosine similarity, Word2Vec/PTB 1251 -621 0.0153 0.3358
Cosine similarity, MegaAttitude 1256 -623 0.0002 0.3346
Non-linguistic account
Frequency 1239 -614 0.0537 0.3305
Models were t with this formula: raw acceptability ? independent vari-
able(s) + (1 | verb) + (1 | participant). Interaction eects were specied for
the indexical shift hierarchy and Anand et al. models but not in the Levin
model. A maximum likelihood procedure was used.
Several measures of quality are reported here. The rst is the Akaike Infor-
mation Criterion (AIC), which measures relative goodness of t adjusted for model
complexity, to facilitate the identication of more parsimonious models. The lower
the AIC, the better the model. The second measure is the log likelihood (logLik):
the closer the log likelihood to zero, the better the model. The third and fourth are
marginal and conditional R2 values, which respectively measure the contribution of
the xed factors and both xed and random factors.
There are several interesting points to note about these results. First, the AIC
200
and log likelihood values favor the frequency, factivity, Anand et al., and indexical
shift models: they account for much of variance in judgments while being relatively
parsimonious.
Second, as measured with conditional R2s, all models capture about a third of
the variance. However, the contribution of the xed factors  the factors of interest
 varies signicantly depending on the model. The better-performing models are
those listed above; in particular, the Anand et al. yields the highest marginal R2,
about 0.097.
Overall, a frequency-based reductionist account does not fare too well under
this analysis. The analysis reveals that frequency has a statistically signicant eect
(? = 0.071, p < 0.001), similar to what Gibson (2018) found. The frequency-only
model is also one of the most parsimonious, as measured by the AIC. However, the
marginal R2 values suggest that frequency does not capture the range of eects as
well as models that are based on semantic factors.
The results provide little support for backgroundedness and verb similarity ac-
counts; their marginal R2s are even lower. This nding aligns with Gibson, 2018/Liu
et al., 2019, which also report a lack of correlation between backgroundedness and
acceptability.
I take a closer look at two models below. The rst is the Anand et al. model,
which has the best log likelihood and marginal R2 values; the eect sizes for various
dimensions are reported in (38). This model eectively shows that manner of speak-
ing and emotive verbs have a negative eect on acceptability via what is essentially
a statistical conspiracy: There is a negative eect on acceptability for non-agentive
201
repository of information (ROI) (? = ?0.397), i.e. if a verb obligatorily takes a
subject that is human and not an agentive repository of information. This set of
verbs include manner of speaking, emotive, and doxastic verbs. However, there is a
positive interaction eect of the same magnitude for doxastic verbs and non-agentive
ROI (? = 0.370).
This model oers little evidence that factivity matters for acceptability. While
non-factivity has a positive eect on acceptability (? = 0.389), this eect is not sig-
nicant. This result is consistent with existing claims in the formal syntax literature
that factive verbs are compatible with long-distance A-bar movement in English.7
(38) Results of Anand et al. (to appear) model for absolute acceptability of long-
distance A-bar movement
Variable Estimate Std. err. t p
(Intercept = Inference verbs) -0.0604 0.2893 -0.209 0.8352
Doxastic verbs 0.3077 0.3403 0.904 0.3685
Communicative verbs -0.1011 0.1667 -0.606 0.5460
Emotive verbs 0.3273 0.2159 1.516 0.1335
Non-factive 0.3894 0.2364 1.647 0.1035
Non-agentive ROI -0.3966 0.0968 -4.097 <0.001 ***
Doxastic * Communicative -0.1229 0.3426 -0.359 0.7210
Doxastic * Non-factive -0.2222 0.2697 -0.824 0.4126
Doxastic * Non-agentive ROI 0.3702 0.1835 2.018 0.0467 *
The second model is the factivity model, in which verbs were sorted into ve
discrete categories, including one for manner of speaking verbs. In numerical terms,
compared to non-factive verbs, veridical verbs have a slightly higher acceptability,
while semifactive, factive, and manner of speaking verbs have a lower acceptability.
However, only the eect for manner of speaking verbs is signicant. The eect for
7Running the same analysis on the full set of 100 verbs yields similar results: there is a sig-
nicant negative eect associated with the non-agentive ROI property, and there is a positive
but marginally signicant eect associated with non-factivity (p = 0.072). The interaction eect
between doxastic and non-agentive ROI verbs is now insignicant.
202
factive verbs is only marginally signicant (p = 0.0634), suggesting that there might
be no dierence between non-factive and factive verbs.
(39) Results of factivity model for absolute acceptability of long-distance A-bar
movement
Variable Estimate Std. err. t p
(Intercept = non-factive) 0.2864 0.0460 6.220 <0.001 ***
Veridical 0.0084 0.1448 0.058 0.9538
Semifactive -0.0740 0.1327 -0.557 0.5787
Factive -0.2693 0.1430 -1.883 0.0634 .
Manner of speaking -0.5075 0.1173 -4.327 < 0.001 ***
4.4.4 Analysis 2: Relative acceptability of long-distance A-bar move-
ment for the full dataset
In this section, I consider the extent to which the various hypotheses can ac-
count for the relative acceptability long-distance A-bar movement: the dierence in
raw acceptability ratings compared to the acceptability of matrix A-bar movement.
Fitting the same regression models to relative acceptability shows that the
quality of the t declines substantially across the board. The xed and random
eects account for only about a quarter of the variance, and the contribution of the
xed eects, as approximated by marginal R2, is much lower, accounting for only up
to about 5% of the variance. We might thus conclude that the various hypotheses
are not particularly successful at modeling the variation in relative acceptability.
Although we can still use the AIC and log likelihood to identify which models
provide the best t, it is unclear, given the R2 values, to what extent doing so is a
meaningful exercise in model selection.
This conclusion is somewhat unexpected: we might expect relative acceptabil-
203
ity to closely track absolute acceptability, and so to the extent that a model can
account for variance in absolute acceptability, we expect it to be similarly successful
for relative acceptability. One post hoc explanation is that there is actually variation
in acceptability in the matrix A-bar movement condition that osets or obscures the
variation for the long-distance A-bar movement condition. Consequently, the eect
size is much smaller and harder to detect with the current sample size.
As D?browska (2013) points out, that strictly speaking, that the verb similar-
ity accounts of D?browska (2004, 2008, 2013) and Verhagen (2006) actually make
no predictions about whether relative acceptability should be inuenced by verb
similarity. I show the results of verb similarity model here for the sake of compre-
hensiveness. Their lexical template hypothesis merely predicts that verb similarity
aects the comprehension or production of long-distance A-bar movement (in fact,
more specically, long-distance wh-questions). A prediction that verb similarity af-
fects relative acceptability requires the ancillary assumption that verb similarity has
no eect on the processing of matrix A-bar movement.
(40) Comparisons of models of relative acceptability of long-distance A-bar move-
ment
204
Marginal Conditional
Model AIC logLik R2 R2
Accounts based on linguistic distinctions
Factivity 1488 -737 0.0166 0.2203
Indexical shift hierarchy 1471 -727 0.0396 0.2452
Anand et al. classication 1468 -723 0.0479 0.2582
Levin classication 1491 -730 0.0330 0.2486
Backgroundedness
Backgroundedness, percentage score 1493 -743 0.0036 0.2038
Normalized backgroundedness 1495 -743 0.0019 0.2082
Verb similarity
Path similarity 1492 -742 0.0046 0.2068
Jiang-Corvath/PTB similarity 1494 -743 0.0024 0.2015
Cosine similarity, Word2Vec/PTB 1484 -738 0.0150 0.2324
Cosine similarity, MegaAttitude 1496 -744 0.0000 0.2024
Non-linguistic account
Frequency 1477 -734 0.0225 0.2251
4.4.5 General discussion
The statistical analysis described show that factors such as frequency and
verb similarity explain relatively little of the variance in absolute acceptability of
long-distance A-bar movement. Similarly, the results also provide little support for
backgroundedness as an explanation of bridge eects.
These results also provide at best weak support for claims that bridge eects
reect sensitivity to certain verb classes or factivity. Although they perform rela-
tively better, the fact is that they too explain only a small degree of the variance.
For example, the Anand et al. model captures about a tenth of the total variance
(marginal R2 = 0.09).
To sum up, the results suggest that existing proposals on bridge eects have
limited explanatory power. It appears that we have yet to fully identify the relevant
dimensions that long-distance A-bar movement in English is sensitive to. This
205
might reect limitations in our methods: perhaps there are issues or limitations in
how example sentences are constructed or judged or the number of verbs that are
investigated in a given study. I return to this issue in Section 4.5.4.
4.5 Just which dimensions matter for long-distance A-bar movement
in English? A post-hoc analysis
In this section, I describe a post-hoc analysis of the acceptability judgments
to try to identify the semantic properties that might matter for long-distance A-bar
movement.
4.5.1 Methods
I calculate average z-score values of absolute acceptability of long-distance A-
bar movement for each verb, and then sort the verbs from lowest acceptability to
highest.8 I focus on the verbs with low acceptability ratings, and try to identify
what properties (semantic, syntactic, discourse, etc.) they might have in common
with each other. The verbs that appear in the bottom half are listed in (41).
(41) Bottom 50 verbs, sorted by absolute acceptability of long-distance A-bar
movement
8Using raw absolute acceptability produces more or less the same ordering.
206
Verb Average z-score Verb Average z-score
mean -1.12 deny -0.03
intimate -1.06 expound -0.02
snap -0.81 regret -0.01
scream -0.78 comment 0.01
allow -0.76 notify 0.02
respond -0.70 maintain 0.03
counter -0.56 prove 0.04
sigh -0.50 boast 0.05
attest -0.30 infer 0.06
shout -0.28 concede 0.06
inform -0.28 disclose 0.06
repeat -0.26 concur 0.07
like -0.25 reveal 0.09
hate -0.24 discover 0.09
disagree -0.23 alert 0.10
venture -0.17 hear 0.12
testify -0.10 promise 0.12
brag -0.10 learn 0.13
whisper -0.09 insinuate 0.14
laugh -0.09 accept 0.14
remember -0.09 notice 0.14
observe -0.05 protest 0.17
conde -0.04 read 0.17
pledge -0.03 forecast 0.18
admit -0.03 doubt 0.18
4.5.2 Verb classes
I was able to sort many of these verbs into the following classes based on
semantic content. These classes might also include verbs for which long-distance
A-bar movement is relatively acceptable; I have listed such verbs in parentheses.
(42) a. Complain verbs (cf. Levin, 1993)
(i) Examples: brag, boast
(ii) Characteristics: Semantically, these verbs convey some kind of
subjective judgment on the part of the attitude holder (i.e. the
bragger or the boaster) with respect to the embedded propo-
sition. These verbs can appear intransitively, with a nite clause,
but not with a nominal object, e.g. *Mary bragged her achieve-
207
ments.
b. Disclose verbs
(i) Examples: disclose, reveal
(ii) Characteristics: Semantically, these verbs describe the annoucne-
ment of some secret or little-known material.
c. Emotive verbs
(i) Examples: like, hate, regret (also fear)
(ii) Characteristics: These describe emotions the attitude holder has
towards a proposition. The ones that are least acceptable are
factive, like Mary {regrets/hates} that John is always late.
d. Hint verbs
(i) Examples: conde, hint, insinuate, intimate
(ii) Characteristics: These verbs all describe the communication of a
matter in a non-open fashion: either the contents are deliberately
obscured, as in hint or insinuate or the communication is done in
a low-key, private fashion, as conde or intimate suggest.
e. Inform verbs
(i) Examples: inform, notify, alert
(ii) Characteristics: These verbs can appear only with a recipient ob-
ject, e.g. Mary informed John. If they take recipient and proposi-
tional arguments, both arguments must appear in a double object
frame, e.g. Mary informed *(John) that he was going to be late.
f. Legalistic verbs (cf. Levin, 1993)
(i) Examples: testify, arm, attest
(ii) Characteristics: These verbs have a performative quality and are
often used in legal contexts. These verbs appear with a nite com-
plement clause, but not with a non-nite one, e.g. Mary armed
that she will go but not *Mary armed to go, and therefore can
be distinguished from similarly performative pledge verbs like
swear or vow, which are also more acceptable in the context of
long-distance A-bar movement.
g. Manner of speaking (cf. Levin, 1993; Zwicky, 1971)
(i) Examples: snap, scream, sigh, shout, laugh
(ii) Characteristics: Semantically, these verbs describe the physical
manner of utterance. As dened here, this class consists of several
distinct subclasses, corresponding in large part to Levin's manner
of speaking and non-verbal expression verbs.
h. Pledge verbs
(i) Examples: pledge, promise, swear, vow
(ii) Characteristics: These verbs have a performative quality and can
be used to express future commitments. Syntactically, they can
appear with both nite and non-nite clauses, in contrast to per-
formative verbs like attest or testify.
i. Reactive verbs (cf. response stance verbs (Cattell, 1978; Honcoop,
1998))
208
(i) Examples: allow, respond, counter, deny, maintain, concede, ac-
cept
(ii) Characteristics: These verbs describe how the attitude holder
reacts to a particular proposition that present in the discourse.
They have fairly diverse syntactic properties: some of these verbs
can take nominal complements, others must appear with certain
prepositions. The diversity in the syntax of these verbs suggests
that this class consists of several smaller classes of verbs.
j. Learn verbs
(i) Examples: discover, hear, learn, notice, observe
(ii) Characteristics: These verbs describe the acquisition of certain
beliefs or knowledge from external sources; the source can be lin-
guistically expressed with the preposition from, e.g. Mary dis-
covered (from the newspaper) that John won the lottery. Most
members of this class are factive; the exception is hear.
Another property that appears to be relevant is whether the verb in an af-
rmative frame entails the truth of the declarative complement: this includes the
veridical, semifactive, and factive verbs. For ease of reference, I will call these the
veridical+ verbs. Note that this property is orthogonal to the verb classes de-
scribed above.
The relevance of veridicality is in a way surprising, since the best-performing
model from the previous analysis, the Anand et al. model, found no statistically
signicant eect for the closely-related property of factivity.
(43) Veridical+ verbs
a. Examples: mean (with a non-human subject), like, hate, remember,
discover, hear, learn (also know, indicate, conrm, forget)
b. Characteristics: These verbs all display the following entailment: Sub-
ject Verb p? p. This set of verbs consists of several distinct subclasses,
such as emotive factives like hate and like, cognitive factives like realize
know, discover, and notice.
209
4.5.3 A statistical analysis
I test the validity of this classication scheme by applying it to the same kind
of mixed eects model. The reason for doing so is not so much as to conrm the
validity of the classication, but to assess the relative improvement of using this
classication system over existing proposals. Compared to other proposed verb
classication systems (except for the Levin system), this particular classication
system makes more ne-grained distinctions, at the risk of overtting the data. The
Akaike Information Criterion is particularly well-suited for this assessment task. It
is dened in a way that balances empirical coverage with model complexity, so that
all else being equal, a more parsimonious model has a lower AIC value.
The results of the statistical analysis are shown in (44), with results from pre-
vious analyses included below for ease of comparison. Excluding the slight decrease
in overall model t, as measured by the conditional R2, this post hoc classication
scheme accounts for about 14% of the variance, half again as much as the Anand et
al. model and almost three times as much as the frequency-only model. The AIC
value is also the lowest compared to all previous models, suggesting that this model
does not overt.
(44) Statistical assessment of post-hoc classication system
210
Marginal Conditional
Model AIC logLik R2 R2
Post hoc classication 1226 -596 0.1411 0.3289
Accounts based on linguistic distinctions
Factivity 1243 -613 0.0602 0.3324
Indexical shift hierarchy 1240 -610 0.0778 0.3326
Anand et al. classication 1236 -606 0.0970 0.3324
Levin, simplied 1258 -612 0.0685 0.3352
Backgroundedness
Backgroundedness, percentage score 1253 -622 0.0092 0.3346
Normalized backgroundedness 1256 -623 0.0001 0.3344
Verb similarity account
Path similarity 1253 -621 0.0097 0.3340
Jiang-Corvath/PTB similarity 1251 -621 0.0145 0.3331
Cosine similarity, Word2Vec/PTB 1251 -621 0.0153 0.3358
Cosine similarity, MegaAttitude 1256 -623 0.0002 0.3346
Non-linguistic account
Frequency 1239 -614 0.0537 0.3305
Models were t with this formula: raw acceptability ? independent vari-
able(s) + (1 | verb) + (1 | participant). Interaction eects were specied for
the indexical shift hierarchy and Anand et al. models but not in the Levin
model. A maximum likelihood procedure was used.
I present estimates of the eects for each verb class in (45). This analysis
shows that most of eects are negative, as predicted, although not every eect size
is statistically signicant. Signicant eects were observed for complain verbs,
legalistic verbs, manner of speaking verbs, reactive verbs, and emotive veridical+
(for all intents and purposes, emotive factive) verbs. A marginally signicant ef-
fect was observed for hint verbs. The signicance of eect sizes for manner of
speaking verbs and emotive factive verbs corroborates previous analyses, giving us
more reason to think that these particular verb classes are incompatible with A-bar
movement.
The analysis also shows again that there is no statistically signicant main
eect for veridicality+: veridicality, semifactivity, and factivity. This is perhaps
211
not surprising since there is overlap between this class and the emotive factive verbs.9
In other words, the post-hoc analysis seems to be consistent with claims in the
literature that factivity per se has little impact on A-bar movement in English. How
might we reconcile this nding with reports by Ambridge and Goldberg (2008) and
Truswell (2011), in which factivity appears to degrade A-bar movement? While I do
not have a denitive answer, it strikes me that the structures that are reported to
be unacceptable typically involve wh-questions. The A-bar dependencies of interest
in the experiment are instances of relativization, not instances of movement of a
wh-phrase to form wh-questions. One plausible account is that the unacceptability
in this instance is related to the pragmatics of wh-questions (Abrus?n, 2011; Gibson,
2018; Liu et al., 2019; M?ller, 2015b; Oshima, 2006; Schwarz and Simonenko, 2018;
Simonenko, 2013, 2015). This would then explain why little or no negative eect
was found for factive verbs in this study and also in Liu et al., 2019, who also used
a mix of relative clause and wh-question stimuli.
(45) Results of model for post hoc classication
9Running the same analysis on the full set of 100 verbs yielded similar results.
212
Variable Estimate Std. err. t p
(Intercept = non-veridical) 0.3915 0.0469 8.352 <0.0001 ***
Complain verbs -0.2741 0.1321 -2.076 0.0412 *
Disclose verbs -0.3002 0.2410 -1.246 0.2175
Emotive verbs 0.3302 0.2602 1.269 0.2078
Hint verbs -0.2788 0.1532 -1.819 0.0724 .
Learn verbs -0.2552 0.2597 -0.982 0.3287
Legalistic verbs -0.3788 0.1563 -2.424 0.0173 *
Manner of speaking verbs -0.7616 0.1117 -6.819 <0.0001 ***
Pledge verbs -0.0666 0.1540 -0.432 0.6667
Reactive verbs -0.3997 0.1003 -3.984 0.0001 ***
Veridical+ -0.0322 0.0985 -0.3270 0.7449
Disclose * Veridical+ 0.0167 0.3650 0.046 0.9636
Emotive * Veridical+ -0.8242 0.3096 -2.662 0.0093 **
Learn * Veridical+ -0.0284 0.3005 -0.094 0.9250
4.5.4 Interim summary
There are a number of points that one can take away from this quantitative
evaluation of bridge eects.
The rst is a negative point: the initial set of analyses show that existing
proposals do not fully capture the range of acceptability of long-distance A-bar
movement, especially once we test these proposals with a much larger set of verbs
that better captures the diversity of clause-embedding verb in English. In some
cases, e.g. factivity, the relevant eect turns out to be very small, or is statistically
insignicant. In other cases, such as verb frequency, the factor has a statistically
signicant eect on acceptability. However, the factor only accounts for a relatively
small amount of the variance in acceptability  in other words, some unknown
factor(s) might more eectively explain the overall acceptability patterns.
The second is a more positive point: The relative success of the post-hoc
analysis suggests that the semantic distinctions that aect the acceptability of long-
213
distance A-bar movement are very ne-grained; these distinctions are unlikely to be
fully captured by existing claims about bridge eects. Of course, this is not to claim
that the post hoc classication system sketched above is a denitive description
of bridge verbs and non-bridge verbs in American English or even generally across
languages.
More generally, I take these analyses to make a more important methodological
point: a full description of bridge eects requires a more careful and systematic
study of attitude verbs. While it is certainly scientically desirable to give as simple
an analysis as possible, given the large number of verbs involved, one has to be
careful when making claims about the empirical coverage of a particular analysis.
A quantitative approach can be helpful in this regard, as it makes it possible to
compare ratings across dierent verb classes, and to adjudicate between competing
proposals. However, informal introspection still has an important role to play: it
can help us formulate novel hypotheses, as the post hoc analysis demonstrates.
While the data reported are based on American English speaker judgments,
they also bear on how we should evaluate descriptions or proposals about bridge
verbs in other languages. A-bar movement in English is relatively well-studied com-
pared to many other languages. For example, I mentioned that there is an argu-
ment/adjunct asymmetry in overt A-bar movement; most work on this topic is built
around English data.10 Similarly, research on A-bar movement from an experimen-
tal syntax perspective is also often based on English or explicit comparisons with
10This is not to deny the fact that other non-English languages, such as East Asian languages
like Chinese, Japanese, Korean, etc., have also played signicant roles in advancing our knowledge
of argument/adjunct asymmetries in covert A-bar movement.
214
English; some exceptions include e.g. Featherston, 2004 and Fanselow et al., 2005
for German, Almeida, 2014 for Brazilian Portuguese, Sprouse et al., 2016 and Kush
et al., 2018 for Norwegian. The fact that existing proposals do not fully capture
the richness of bridge eects in English then suggests that they are likely to also
understate the richness of bridge eects in other languages. Again, this should not
be understood as dismissing the value of existing descriptions or proposals; rather,
it is a reminder that we need to be careful about how we construct examples and
use acceptability judgments to support a proposal.
4.6 The source of bridge eects
4.6.1 Three hypotheses
We saw in the above section that there are a number of ne-grained seman-
tic dimensions that seem to reduce the acceptability of long-distance A-bar move-
ment. Logically, there are (at least) three distinct positions compatible with this
evidence that one could take about the nature of these eects and the syntactic
well-formedness of long-distance A-bar movement.
First, one could treat these eects as purely pragmatic in nature, and take long-
distance A-bar movement to be syntactically well-formed, regardless of the choice
of clause-embedding verb. For illustration, consider the restriction on long-distance
A-bar movement imposed by manner of speaking verbs. Given how prominently
the mode of speaking is featured, it is not dicult to imagine that there could be
general pragmatic or discourse principles that restrict the circumstances in which
215
they can be used (Ambridge and Goldberg, 2008; Erteschik-Shir and Lappin, 1979;
M?ller, 2015b). For instance, consider the following sentence, which is based on an
example discussed by M?ller (2015b) and Erteschik-Shir and Lappin (1979).
(46) What1 did he lisp that they'd do t1?
a. Degraded if he refers to some generic person.
b. More acceptable if he refers to Truman Capote, known for lisping.
The Gricean-based approach advocated by (M?ller, 2015b, pp. 43-44) goes
roughly as follows: The point of asking the question (46) is to nd out about the
contents of a speech act, and a general communication verb that is relatively neutral
about the manner of speech, like say, should suce. Using lisp is too informative,
which reduces the sentence's acceptability. However, if the context is about Truman
Capote, for whom lisping is a distinctive way of speaking, then the choice of using lisp
becomes clearer to a listener (or reader), and the sentence becomes more acceptable,
which reects its syntactic well-formedness.
It seems reasonable to also apply a pragmatic explanation to the other classes
identied in the post-hoc analysis, such as reactive verbs like deny. As mentioned,
these verbs presuppose a proposition that the attitude holder reacts to. However,
in a formal experiment where sentences are presented with little or no context,
such a presupposition might not readily available, which might explain the lower
acceptability.
Alternatively, one might take these eects to show that syntactically, long-
distance A-bar movement is restricted to certain verb classes. There are several
ways of formally implementing this idea: one could simply stipulate representa-
216
tional / derivational constraints that rule out long-distance A-bar movement from
complement clauses of certain classes of verbs. Alternatively, assuming that A-bar
movement takes place successive cyclically through the left periphery of embedded
clauses, one could also do so via subcategorization: certain verbs might take com-
plement clauses that block A-bar movement through their left periphery; I review
these possibilities more thoroughly in Chapter 5.
All else being equal, this account predicts that long-distance A-bar movement
should always be unacceptable with certain verb classes, which seems contrary to
existing reports about the ameliorative eect of context. One possibility is to explain
this as a repair eect. For the sake of illustration, I will return to the lisping
example. Here, when the context makes clear that the lisping is done by Capote,
the listener/reader might draw upon their knowledge of Capote's distinctive way of
speaking, observing that the verb lisp is a fair characterization of a regular speech act
by Capote, just like how say is used to describe a regular speech act by the average
person. This observation enables them to repair the representation of (46), by
replacing lisp with say. Successful repair would yield a well-formed representation,
which explains why (46) is relatively acceptable.
A third approach is a hybrid one: there is no single source of bridge eects.
For some verbs, the reduced acceptability is due to pragmatics, but for other verbs,
the reduced acceptability reects the violation of some syntactic constraint.
217
4.6.2 Evaluating these hypotheses from a conceptual perspective
From a theoretical perspective, the rst approach  treating bridge eects
as having a pragmatic origin  is conceptually perhaps the most parsimonious. It
ts in neatly with a simple theory of long-distance A-bar movement: long-distance
A-bar movement from complement clauses of verbs is always syntactically well-
formed. From the perspective of learning syntax, this scenario is also not unwel-
come. It means that what learners need to do is to learn that their language allows
A-bar movement from complement clauses of verbs, and we know independently
that evidence for this conclusion exists: see Pearl and Sprouse, 2013a,b for data
on naturally occurring instances of long-distance A-bar movement in English, and
Verhagen, 2006, Schippers, 2012, Broekhuis and Corver, 2018 for Dutch data.
In contrast, the other two approaches seem to be more stipulative, even though
they are also logically coherent. For instance, one might wonder whether there is
any insight to be gained by claiming that there is some constraint referring to
certain verb classes that block long-distance A-bar movement; it appears to be
simply restating the facts. The hybrid approach, by claiming that both pragmatic
and syntactic approaches are necessary, also comes across as encoding a certain
amount of redundancy.
Having said all these, the fact is that far more work is needed to determine
the exact role of context in aecting acceptability of long-distance A-bar movement.
While there is no doubt that context inuences acceptability judgments, the fact is
that there is no widely accepted theory about the mechanism(s) by which it aects
218
acceptability or about the size of the eect.
More specically, I have pointed out that accounts in which bridge eects are
syntactic will also require an account of amelioration. The repair scenario that I
sketched above, while perhaps plausible, obviously requires more work, as there are
still open questions about what factors trigger repair and aect the likelihood of
repair.11
As for pragmatic accounts of bridge eects, as mentioned, there are worked
out accounts about how pragmatic or discourse principles can aect acceptability
for certain verb classes, such as manner of speaking verbs. It is less clear whether
these accounts can be reasonably extended in a way that explains the range of ac-
ceptability judgments for a more representative set of verbs, such as those reported
by Gibson (2018)/Liu et al. (2019) or in the acceptability judgment experiment in
Section 4.4. Although it seems feasible, I am not aware of any specic theory on
the market that achieves this kind of empirical coverage, in large part because the
acceptability proles of many of these verbs are poorly understood. The experi-
mental results discussed above thus present new challenges  although perhaps not
unsurmountable ones  for pragmatic accounts of bridge eects.
4.6.3 Evaluating these hypotheses from an empirical perspective
Empirically, one prediction that the rst (pragmatics) approach might make is
cross-linguistic uniformity in the acceptability proles of long-distance A-bar move-
11Put dierently, we do not want a general theory in which repair applies too easily, such that
it predicts that speakers readily accept ungrammatical structure.
219
ment. This prediction follows from two ancillary assumptions: (i) verbs with similar
lexical semantics and syntax should have similar pragmatics, (ii) pragmatic princi-
ples are likely to be universal, especially considering the fact that they are attested
cross-linguistically and are dicult to learn directly. As an example, consider the
fact that clause-embedding verbs like English shout and Dutch schreeuwen `yell,
shout' have very similar lexical semantics and syntax. From assumption (i), it fol-
lows that they have similar pragmatics. If so, by assumption (ii), long-distance
A-bar movement for shout and schreeuwen (and other pairs in English and Dutch,
or in other languages) is subject to the same pragmatic restrictions, and therefore
should be similar in acceptability.
In contrast, the other two accounts, in which (some) bridge eects have a
syntactic origin, do not automatically make this prediction. One could imagine a
scenario where languages vary in terms of which verb classes block long-distance
A-bar movement, although this raises questions about how these distinctions might
be learned.
Factive verbs appear to bear on this particular issue. In English, factivity
per se (or the closely-related property of veridicality) seems to have little or no
signicant eect on the acceptability of long-distance A-bar movement, as suggested
by the models reviewed in preceding sections  the Anand et al. model and the
post-hoc classication model.12
12As noted previously, this outcome is inconsistent with certain claims in the literature, e.g.
Ambridge and Goldberg, 2008; Truswell, 2011. Of course, factivity might give rise to an interaction
eect with other semantic properties, as in the case of emotive factives. It is also not unreasonable
that certain types of A-bar dependencies, such as wh-movement in the context of wh-questions,
might interact with factivity to lower acceptability.
220
For the sake of exposition, suppose that these results do show that in English,
long-distance A-bar movement itself is not sensitive to factivity. In that case, English
provides an interesting contrast with certain varieties of Dutch. As mentioned in the
introduction of this chapter, Broekhuis and Corver (2016) report that in Dutch a
factive verb like weten `know' typically blocks A-bar movement of a bare wh-phrase
like wat `what,' while a verb like denken `think' does not. Making the wh-phrase
D-linked improves acceptability, but not uniformly  there are speakers who reject
the D-linked sentence (47).
(47) Dutch (a and b examples from Broekhuis and Corver, 2016, p. 1397, ex.
214).
a. Wat {dacht/*weet} Jan dat Marie t1 gekocht had]?
what thought/knew Jan that Marie bought had
`What did Jan think/know that Marie had bought?'
b. Welk boek1 {dacht/%wist} Marie [dat Marie t1 gekocht had]?
which book thought/knew Jan that Marie bought had
`What did Jan think/know that Marie had bought?' (% = some speak-
ers only)
c. Welke lm {*herinnerde/*realiseerde} Marie zich dat Jan zag?
which movie remembered/realized Marie self that Jan saw
Intended: `Which movie did Marie remember/realize that Jan saw?'
(Annemarie van Dooren, p.c.)
It is tricky to explain away the low acceptability as an idiosyncratic fact about
wh-questions, as a similar unacceptability can be observed with relativization (48).
(48) a. Dit is de lm die Marie dacht dat Jan heeft gezien.
this is the movie that Marie thought that Jan has seen
Intended: `This is the movie that Marie thought that Jan has seen.
(Annemarie van Dooren, p.c.)
b. ??Dit is de lm die Marie zich {herinnerde/realiseerde} dat Jan
this is the movie that Marie self remembered/realized that Jan
heeft gezien.
has seen
Intended: `This is the movie that Marie remembered/realized that Jan
221
has seen. (ibid.)
However, the reports of interspeaker variation in Dutch might cause one to
wonder to what extent this cross-linguistic variation reects a data quality issue.
More specically, for the sake of argument, perhaps some Dutch speakers were re-
jecting long-distance A-bar movement with factive verbs, not because long-distance
A-bar movement in these environments is necessarily worse than it is in English, but
because the sentences in question were constructed in a way that required a specic
context that was not as salient. If an appropriate context had been supplied, these
Dutch speakers would nd long-distance A-bar movement with (cognitive) factive
verbs as acceptable as their English-speaking counterparts did. Such a scenario
would allow one to claim that there are no systematic dierences between English
and Dutch, consistent with the position that bridge eects are pragmatic and not
syntactic.
Finally, it is also important to make two points: First, this discussion is not
intended to cast doubt on how existing data points were collected or how the data
were interpreted. Rather, the goal is to reconsider the factors that might be driving
dierences in judgments. Second, the issue of cross-linguistic variation is not neces-
sarily an issue that is restricted to factivity or to English and Dutch. It is logically
possible that other languages and/or other verb classes might show similar kinds of
variation. This discussion is centered on factive verbs in these languages in large
part because these issues are relatively well-studied and we have some evidence to
believe that cross-linguistic variation in this dimension exists.
222
4.7 A survey of English and Dutch speakers
To investigate this question on pragmatics and long-distance A-bar movement,
I designed a short survey for native speakers of English and Dutch. The main goal
was to determine whether long-distance A-bar movement improves in acceptability,
perhaps to the extent of becoming perfectly acceptable, when a suitable context is
present; a secondary goal was to validate existing claims about interspeaker variation
regarding cognitive factive verbs in Dutch.
Practical constraints meant that this survey could not be run as a formal
experiment with suitable controls: doing so would entail creating additional control
sentences in Dutch and recruiting enough Dutch speakers to participate. The task
of reading sentences with context also called for a certain amount of attention and
processing that might be dicult to guarantee in a formal experiment setting where
participants were anonymous. Consequently, I opted to conduct a more informal
survey over email, which also allowed me to get judgments and any comments that
native speaker consultants might have about the sentences. It also allowed me to
follow up with them to clarify what their intuitions were.
4.7.1 Survey materials
This survey had a 3x2 design. There are three types of clause-embedding verbs
(or adverbverb combinations): manner of speaking verbs, say or claim modied
with a manner-like adverb, and cognitive factive verbs. Manner of speaking and
cognitive factive verbs were chosen in part because there are explicit discussion and
223
examples in the literature of whether A-bar movement from their complements is
acceptable or not (Ambridge and Goldberg, 2008; D?browska, 2008, 2013; M?ller,
2015b; Oshima, 2006). The adverbsay condition was added as a control for man-
ner of speaking verbs, to see if the presence of an adverbial-like element that provided
additional information about the event denoted by a bridge-like clause-embedding
verb would reduce acceptability.
Sentences were intended to be presented without and with an appropriate con-
text, but were constructed in a way so that they should be the most acceptable when
presented with a suitable context. This goal also determined the choice of verbs;
it was easier to create sentences and contexts for certain verbs. For the manner of
speaking verbs, sentences took the form of wh-questions, although some of them for-
mally feature relativization, and not movement of a wh-phrase. This was consciously
modeled after observations by Ambridge and Goldberg (2008), M?ller (2015b), and
Erteschik-Shir and Lappin (1979), that such questions become acceptable when the
event denoted by the verb is contextually salient.13
For cognitive factive verbs, the A-bar dependencies in question took the form
of relativization. These sentences were modeled after naturally-occurring examples
in the Penn Treebank that I judged to be the most acceptable.
The full set of contexts and sentences in English are presented in (49). An-
nemarie van Dooren kindly provided the Dutch translations (50), which also featured
long-distance A-bar movement (instead of, for instance, resumptive pronouns). For
13Special thanks to Alexander Williams, Omer Preminger, Valentine Hacquard, Annemarie van
Dooren, and Aaron Doliana for discussion and comments on these contexts.
224
ease of presentation here, the verbs are bolded and the complement clauses put
between square brackets.
(49) English sentences
a. Manner of speaking verbs
(i) Context: You remember how Mary was whispering to us during
the movie that John liked some girl?
Critical sentence: Who's the girl that Mary whispered [that
John liked]?
(ii) Context: You know how Mary shouted during the protest that
the police had shot a teenager?
Critical sentence: Who's the teenager that Mary shouted [that
the police shot]?
(iii) Context: You know how Marie mumbled to the client that the
CEO wanted to buy some company?
Critical sentence: Which company did Marie mumble [that the
CEO wanted to buy]?
(iv) Context: You know how the suspect stammered that the terrorists
were going to attack an airport?
Critical sentence: Which airport did the suspect stammer [the
terrorists were going to attack]?
b. (Manner) adverbsay combination
(i) Context: You remember howMary was quietly saying to us during
the movie that John met some actor?
Critical sentence: Who's the actor that Mary quietly said [that
John met]?
(ii) Context: You know how Mary was loudly claiming during the
protest that the police had shot a teenager?
Critical sentence: Who's the teenager that Mary loudly claimed
[that the police shot]?
(iii) Context: You know how Mary secretly told the client that the
CEO wanted to buy some company?
Critical sentence: Which company did Mary secretly say [the
CEO wanted to buy]?
(iv) Context: You know how the suspect reluctantly said that the
terrorists were going to attack an airport?
Critical sentence: Which airport did the suspect reluctantly say
[the terrorists were going to attack]?
c. Cognitive factive verbs
(i) Context: As a smart sales manager, Mary goes to meals with her
clients and quickly learns what they like.
Critical sentence: Last week, she gave a client a kind of wine that
she had noticed [he often ordered at dinner].
225
(ii) Context: Sally is an art student who is planning a trip to New
York. She has told John about all the museums that she plans to
visit.
Critical sentence: These are the museums that she found out
[that students can visit for free].
(iii) Context: John lost his job recently.
Critical sentence: Mary went to visit him to give him $300, which
she knew [he needed for rent].
(iv) Context: When you invest in a business, it is important to un-
derstand the risks.
Critical sentence: People always lose money on investments that
they don't realize [are risky].
(50) Dutch sentences
a. Manner of speaking verbs
(i) Context: Je herinnert je toch hoe Marie tijdens de lm ons in-
uisterde dat Jan een bepaald meisje leuk vindt?
Critical sentence: Wie is het meisje dat Marie ons inuisterde
[dat Jan leuk vindt]?
(ii) Context: Je weet toch hoe Marie tijdens het protest schreeuwde
dat de politie een tiener had neergeschoten?
Critical sentence: Wie is de tiener die Marie schreeuwde [dat de
politie had neergeschoten]?
(iii) Context: Je weet toch hoe Marie tegen de klant mompelde dat
de directeur een bepaald bedrijf wilde overnemen?
Critical sentence: Welk bedrijf mompelde Marie [dat de di-
recteur over wilde nemen]?
(iv) Context: Je weet toch hoe de verdachte tijdens de ondervraging
stamelde dat de terroristen een bepaald vliegveld zouden gaan
aanvallen?
Critical sentence: Welk vliegveld stamelde de verdachte [dat de
terroristen zouden gaan aanvallen]?
b. Adverbsay combination
(i) Context: Je herinnert je toch hoe Marie tijdens de lm ons stil-
letjes zei dat Jan een bepaalde acteur heeft ontmoet?
Critical sentence: Wie is de acteur die Marie stilletjes zei [dat
Jan heeft ontmoet]?
(ii) Context: Je weet toch hoe Marie tijdens het protest luidkeels
claimde dat de politie een tiener had neergeschoten?
Critical sentence: Wie is de tiener die Marie luidkeels claimde
[dat de politie had neergeschoten]?
(iii) Context: Je weet toch hoe Marie de klant in het geheim vertelde
dat de directeur een bepaald bedrijf wilde overnemen?
226
Critical sentence: Welk bedrijf vertelde Marie de klant in het
geheim [dat de directeur over wilde nemen]?
(iv) Context: Je weet toch hoe de verdachte aarzelend zei dat de ter-
roristen een bepaald vliegveld zouden gaan aanvallen?
Critical sentence: Welk vliegveld zei de verdachte aarzelend [dat
de terroristen zouden gaan aanvallen]?
c. Cognitive factive verbs
(i) Context: Een slimme verkoper zoals Marie gaat uit eten met haar
klanten om er snel achter te komen wat ze lekker vinden.
Critical sentence: Vorige week gaf ze een klant een wijn die ze
had opgemerkt [dat hij vaak bestelde tijdens het eten].
(ii) Context: Marie studeert aan de kunstacademie en is bezig een
reis naar New York voor te bereiden.
Critical sentence: Ze heeft Jan verteld over alle musea die ze van
plan is te bezoeken. Dit zijn de musea die ze heeft ontdekt [dat
studenten gratis kunnen bezoeken].
(iii) Context: Jan is onlangs ontslagen.
Critical sentence: Marie ging bij hem op bezoek om hem 300 euro
te geven, die ze wist [dat hij nodig had om de huur te betalen].
(iv) Context: Als je in zaken investeert is het belangrijk om de risico's
te kennen.
Critical sentence: Mensen draaien altijd verlies op investeringen
die ze zich niet hadden gerealiseerd [dat riskant waren].14
4.7.2 Collecting and evaluating native speaker judgments
I surveyed 7 native speakers of English and 7 Dutch speakers. The Dutch
speakers were all speakers of Netherlands dialects.15
Native speakers rst read and rated the critical sentences without context.
They were then presented with a suitable context preceding each sentence, and told
14In this particular example, the extracted element corresponds to the subject of the embedded
clause that is marked with the complementizer dat, i.e. the subject of riskant waren. In English,
the same conguration is unacceptable, due to a violation of the that-trace eect: *. . . investments
that they don't realize [that are risky]. In standard Dutch at least, there is no that(dat)-trace
constraint (see Broekhuis and Corver, 2016, p. 1386 and references therein). However, a Dutch
native speaker consultant commented that this sentence could become even more acceptable if
there were a subject pronoun ze they after dat.
15My sincere thanks to Maxime Tulling and Anne Giskes for putting me in touch with native
speakers of Dutch.
227
to rate them taking the context into consideration. Most native speakers did not
have a syntax background, and therefore were not familiar with the issues being
investigated. To help native speakers unfamiliar with this kind of judgment task,
I explicitly asked all of them to rate sentences along a 7-point Likert scale, with 1
being the least acceptable, and 7 the most acceptable. Doing so also made it easier
to measure the eect of context.
4.7.3 Summary of ratings
For ease of comparison, for each speaker, I report median ratings for each of
the three verb conditions, without and with contexts.
(51) Median ratings for English native speakers (1 = completely unacceptable,
7 = completely acceptable)
Speaker #1 #2 #3 #4 #5 #6 #7
Manner of speaking verbs, no con- 3 2.5 5.5 6 4 4 4.5
text
Manner of speaking verbs, with 7 4 6.5 6.5 5 6 4
context
Adverbsay, no context 5 3.5 5.5 6 5.5 5.5 5
Adverbsay, with context 7 4 6.5 5 6 6 3.5
Cognitive factive verbs, no context 7 7 7 7 7 7 7
Cognitive factive verbs, with con- 7 7 7 7 6.5 7 7
text
(52) Median ratings for Dutch native speakers (1 = completely unacceptable, 7
= completely acceptable)
228
Speaker #1 #2 #3 #4 #5 #6 #7
Manner of speaking verbs, no con- 3.5 2 4 2.5 3 2 5.5
text
Manner of speaking verbs, with 4 5 6 6.5 4 2 2
context
Adverbsay, no context 4.5 3 5.5 3 3.5 2 5
Adverbsay, with context 4.5 3.5 6.5 4 5 2.5 2
Cognitive factive verbs, no context 6 1.5 6 4 4 2 2.5
Cognitive factive verbs, with con- 7 3.5 6.5 4.5 6 3 4.5
text
Separately, I sought to verify with Dutch speakers whether they allowed long-
distance A-bar movement  specically, to form wh-questions and relative clauses
 without a context, across the prototypical verb denken `think,' in eect using
denken as a baseline. For transparency, some speakers found relativization to be of
intermediate acceptability. Others, however, found relativization perfectly accept-
able. Some speakers also mentioned a preference for paraphrases with resumptive
pronouns, even if they found the baseline sentences to be fairly acceptable. Over-
all, speaker ratings for these baseline sentences did not seem to correlate with their
ratings for the critical sentences, except for Speaker #6, who consistently assigned
low ratings.
This informal survey yields a few preliminary generalizations.
First, for a majority of English speakers, long-distance A-bar movement from
the complements of manner speaking verbs and adverbsay combinations is rela-
tively degraded without context. In such environments, most speakers assigned low
to intermediate ratings on a 7-point scale. With context  in these cases, an intro-
ductory sentence that made salient the event denoted by these verbs (and adverbs)
 the same sentences generally became more acceptable. However, the degree of
229
improvement varied by speaker; an inspection of ratings also shows that there is
variation at the verb level. Similar observations apply for Dutch.
A linguistically-trained Dutch speaker (but naive to the issues studied here)
volunteered the observation that with context, the improvement was striking for
some of the verbs;16 a linguistically-naive English speaker independently echoed the
same sentiment.
For one English speaker (#7) and one Dutch speaker (#7), however, the pres-
ence of context actually caused long-distance A-bar movement to become less ac-
ceptable. In these examples, the verb was rst presented in the context sentence
and then repeated in the critical sentence featuring long-distance A-bar movement.
Comments from the English speaker seem to suggest that the repetition of the verb
actually made the critical sentence less acceptable.
An interesting case of variation surfaced with the cognitive factive verbs. For
English speakers, these examples were consistently highly acceptable, even without
an introductory sentence providing context. This outcome is unsurprising, since
the sentences were deliberately modeled upon naturally-occurring examples in an
English corpus that were chosen for their naturalness.
However, the equivalent sentences for Dutch did not show the same uniformity
in ratings, even though these sentences express the same sentiments as their English
counterparts. While there were Dutch speakers who found these to be acceptable
(#1, #3), others assigned what appeared to be much lower ratings, e.g. Dutch
16However, an inspection of that speaker's ratings showed that the improvement does not apply
across the board, which means it is unlikely that there is a general satiation eect.
230
speakers #2, #6, #7. Setting aside speaker #6, I note that speakers #2 and #7
accepted A-bar movement across the baseline denken `think,' so their low ratings
do not reect some general bias against long-distance A-bar movement.
In fact, for these sentences, some speakers suggested that the sentences could
be improved by changing the relative clauses from instances of long-distance A-bar
movement (with the relative pronoun die as the head of the dependency, and a
gap in the tail) to a related construction featuring waarvan (lit. where of) and a
resumptive pronoun (or a demonstrative); the suggested revisions are indicated in
(53b) and (54b). Another native speaker suggested that the equivalent sentences
can be expressed using what appears to be a wh-copying construction, where a wh-
element appears in the left periphery of the complement clause of the factive verb
(54c).
(53) a. Marie ging bij Jan op bezoek om hem 300 euro te geven,
Marie went to Jan on visit COMP him 300 euro to give
die1 ze wist dat hij t1 nodig had om de huur te
REL.PRN she knew that he need had COMP the rent to
betalen.
pay
`Marie went to visit Jan to give him 300 euros, which she knew he
needed to pay rent.' (original sentence)
b. . . . 300 euro te geven, waarvan ze wist dat hij ze/die nodig
300 euros to give where.of she knew that he it/that need
had om de huur te betalen.
have to the rent to pay.
`. . . to give [him] 300 euros, which she knew that he needed to pay
rent.' (suggested change by native speaker consultants)
(54) a. Dit zijn de musea die1 Marie heeft ontdekt dat
this are the museums REL.PRN Marie has discovered that
studenten t1 gratis kunnen bezoeken.
students free can visit
231
`These are the museums that Marie discovered that students can visit
for free.' (original sentence)
b. Dit zijn de musea waarvan Marie heeft ontdekt dat
this are the museums where.of Marie has discovered that
studenten ze gratis kunnen bezoeken.
students them free can visit
`These are the museums that Marie discovered that students can visit
for free.' (suggested change by native speaker consultant)
c. Dit zijn de musea die Marie heeft ontdekt waar
this are the museums REL.PRN Marie has discovered where
studenten gratis op bezoek kunnen
students free on visit can
`These are the museums that Marie discovered that students can visit
for free.' (suggested change by native speaker consultant)
4.7.4 Discussion
There are a few takeaways from the data collected from this survey.
First, with a suitable context, A-bar movement from complement clauses of
manner of speaking verbs can become fairly acceptable, although the improvement
is not consistent across all verbs or all speakers. This variation might simply be due
to interspeaker dierences in perceiving how appropriate the context is for a given
verb. If so, this amelioration eect is consistent with the claim that the canonical low
acceptability associated with these constructions is pragmatic in nature (Ambridge
and Goldberg, 2008; Erteschik-Shir and Lappin, 1979; M?ller, 2015b), as well as
being consistent with a scenario in which these structures are ungrammatical but
repairable if a suitable context is available.
Second, the results reveal dierent acceptability proles for long-distance A-
bar movement with cognitive factive verbs for English and Dutch, suggesting deeper
dierences between these two languages and interspeaker variation within Dutch.
232
Specically, some Dutch speakers were willing to accept long-distance A-bar move-
ment across a verb like denken `think' but not across cognitive factive verbs. The
interspeaker variation observed here is consistent with existing reports in the Dutch
syntax literature that not all speakers nd A-bar movement from complement clauses
of (cognitive) factive verbs to be acceptable (Broekhuis and Corver, 2016 and ref-
erences therein, also comments by Annemarie van Dooren, p.c.). It is unlikely that
this variation can be blamed on the absence of an appropriate linguistic context,
since the equivalent English sentences, which are at least syntactically and semanti-
cally/contextually identical, receive much higher ratings by English native speakers.
Comments from Dutch native speakers also validate existing reports in the
literature that in such instances, speakers might prefer alternative constructions,
such as the waarvanresumptive pronoun construction. An alternative construction
described by Broekhuis and Corver (2016) and Salzmann (2017) is the syntactically
similar resumptive prolepsis construction, where the head of the wh-dependency is
a preposition (like van) followed by a wh-phrase, and the tail of the dependency a
resumptive pronoun (55).
(55) Van welk boek wist Jan niet dat Els het gekocht had?
of which book knew Jan not that Else it bought had
`Of which book didn't Jan know that Els had bought [it]?' (Resumptive
prolepsis, Broekhuis and Corver, 2016, p. 1404 ex. 232b)
Third, the English data collected in informal survey echoed the ndings of
the formal acceptability judgment experiment reported in Section 4.4: compared to
manner of speaking verbs, there is little evidence to suggest that cognitive factive
verbs block A-bar movement from their complement clauses.
233
These results, if they are on the right track, have implications for our under-
standing of bridge eects. Recall that a primary goal for this survey was to evaluate
the hypothesis in which bridge eects are pragmatic in nature. The specic hypoth-
esis tested was that long-distance A-bar movement can become acceptable when a
suitable context is present, on the linking hypothesis that verbs with similar lexi-
cal semantics should have similar pragmatics. The data for cognitive factive verbs
appear to pose a problem for this hypothesis, as the acceptability proles for these
verbs turn out to be dierent between English and Dutch, when context is controlled
for.
To maintain this general hypothesis in light of the survey ndings, one might
posit a more nuanced view of these verbs, namely, that their pragmatics or semantics
vary subtly in English and Dutch. Consequently, the contexts that justify the use
of English know vs. Dutch weten are dierent, contrary to how this particular
survey was set up. In fact, interspeaker variation in Dutch would suggest that the
pragmatics and semantics of these verbs vary across Dutch speakers (or varieties),
and correspondingly so do the contexts that justify their use.
4.8 Directly learning about the acceptability of long-distance A-bar
movement with cognitive factive verbs / Corpus study 5
How might this cross-linguistic dierence for cognitive factive verbs arise? One
straightforward hypothesis is that English speakers get to observe more reliable
direct evidence that cognitive factive verbs are compatible with long-distance A-bar
234
movement, compared to Dutch speakers. Observing these tokens then leads English
learners to learn either (i) the pragmatics (or semantics) of these verbs are such that
they are compatible with long-distance A-bar movement, even when relatively little
context is available, or (ii) that long-distance A-bar movement from the complements
of these verbs is syntactically well-formed in English.
As I see it, there are two requirements for this direct learning scenario to work:
English speakers must be able to (i) identify a class of cognitive factive verbs and (ii)
observe A-bar movement from the complements of factive verbs in their linguistic
experience, at least a rate higher than Dutch speakers. I address these in turn.
As a class, cognitive factive verbs are distinctive in their syntax and semantics.
As verbs of cognition, these verbs are subject to animacy restrictions: when these
verbs appear in an active voice frame, the subject cannot be non-human, since non-
human entities are not capable of thinking; Anand et al. (to appear) further observe
that the subject cannot even be an agentive repository of information, the product
of some human entity (56); this fact sets it apart from other clause-embedding verbs
that can appear with non-human subjects, such as communicative verbs.
(56) {*The table/*The report/Mary} {knew/realized/forgot/found out} that
John was responsible.
Further, a number of these verbs are distinguished from non-factive verbs like
think in that they can take embedded questions as complements (57a) (see also
Spector and Egr?, 2015; White and Rawlins, 2018) or, at least in some cases, omit
them entirely (57b) if the verb and context allows it.17 These subcategorization
17In other words, some of these verbs allow null complement anaphora. This is not a distin-
guishing characteristic for cognitive factive verbs, as there are some cognitive factive verbs that
235
properties set the cognitive factive verbs apart from the emotive factive verbs (57c),
(57d).
(57) a. Mary {knew/realized/forgot/found out} who was late.
b. Mary {knew/realized/forgot/found out}.
c. *Mary {regretted/hated} who was late.
d. *Mary {regretted/hated}.
Semantically, these verbs are distinct in presupposing the truth of their com-
plements, if the complements are declarative clauses. This property might corre-
spond with relatively easy-to-observe cues: to the extent that there is a correlation
between presupposed propositions and discourse-old propositions (i.e. propositions
that are previously mentioned and present in the common ground) and learners are
able to track propositions in the discourse, these verbs might be more likely than
non-factive attitude verbs like think to take complements whose propositions are
discourse old (that said, Dudley et al. (2017) did not nd a clear dierence in a
corpus study of child-directed speech).
The second requirement is also not as far-fetched as it might rst appear to
be. Impressionistically, there are factive verbs that are relatively high frequency in
English: for example, know, remember, forget. Although these verbs appear in a
variety of frames, given their frequency, it is not implausible that these verbs might
appear naturally in bridge contexts.18
require overt clausal complements, such as discover (cf. nd out, which allows null complement
anaphora).
18I note that there are actually two versions of this hypothesis. In the weak version of this
hypothesis, English speakers observe A-bar movement from complements of some cognitive factive
verbs, and learn that these verbs  and only these verbs  are compatible with A-bar move-
ment. In the strong version, English speakers observe these tokens and generalize across the set of
cognitive factive verbs. I return to this point later in Section 4.8.3.
236
In contrast, given what we know about interspeaker variation about long-
distance A-bar movement and (cognitive) factive verbs in Dutch, it is possible that
the equivalent evidence is less robust for learners. Some speakers of Dutch do not
allow long-distance A-bar movement with cognitive factive verbs and presumably
do not produce these structures. All else being equal, these structures are less likely
to occur naturally in Dutch.
To preview the rest of this section: I assess this hypothesis with an analysis of
an English corpus, the Penn Treebank. I show that there is limited direct evidence
that cognitive factive verbs can appear in long-distance A-bar movement congura-
tions: most cognitive factive verbs appear infrequently in such congurations, if at
all. However, as data from the previous sections show, native speakers of English
seem to possess intuitions that long-distance A-bar movement from the complements
of these verbs are (relatively) acceptable.
I argue that one solution that might let us reconcile the sparsity of direct
evidence with native speaker intuitions is that speakers form generalizations about
long-distance A-bar movement from cognitive factive verbs. I discuss two general-
ization accounts, testing them with the Penn Treebank data.
4.8.1 Methods
Ideally, evaluating this hypothesis would require looking at the linguistic ex-
perience of Dutch and English speakers, for instance, through the lens of corpora.
While it is not feasible for me to look at Dutch data, since I do not have access to
237
such a dataset nor the Dutch skills to analyze such data adequately, such an analysis
is possible for English. Specically, I turned to the Penn Treebank again. In Section
4.4.1.1, I described extracting all instances of a verb occurring with a declarative
complement clause from the Penn Treebank.
I wrote a Python script that analyzed this set of complement clauses. The
script counted the number of traces within each complement clause, and for each
trace, determined whether its antecedent was outside of the clause. If the antecedent
was, then one concludes that A-bar movement has taken place from the embedded
clause, and the verb appears in a bridge context. The script also automatically
agged for exclusion instances of direct quotation (e.g. John said Mary is a ge-
nius.) and S-lifting (e.g. Mary is a genius, John said.).
To minimize the likelihood of false positives in the identication of complement
clauses, I manually checked verb lemmas that appeared with a declarative nite
sister nine times or fewer and corrected the annotation if necessary. I assume that all
verb lemmas that appeared at a higher frequency take a complement clause. Within
this set of complement clauses, the script agged about 640 tokens of potential long-
distance A-bar movement. To rule out false positives in this process, I manually
checked each token.
4.8.2 Results
I identied 593 instances of long-distance A-bar movement, associated with
57 unique attitude verb lemmas. Crucially, there are seven factive verbs in this set,
238
appearing a total of 44 times: know, which appears the most frequently at 34 times,
remember, see, learn, realize (two times each), nd out, and understand (1 time
each).
For transparency, I further note that these 57 verbs were distributed according
to what looks like a power law distribution. The top 10 verbs appeared a total of
516 times (87% of 593). The remaining 47 verbs appeared a total of 77 times, with
31 verbs appearing only once.
(58) Clause-embedding verbs appearing in long-distance A-bar movement con-
texts in the Penn Treebank dataset
Frequency of Number of
bridge contexts Verbs unique verbs
More than 100x think (260x), say (117x) 2
Between 10x and 99x know (34x), feel (33x), believe (25x), 5
guess (17x), hope (11x)
Between 2x and 9x seem (7x), claim, nd, wish (6x each), 19
estimate, fear (4x each), agree, doubt,
gure, imagine, suspect, tell (3x each),
contend, decide, learn, realize, remember,
see, suppose (2x each)
Only 1x allege, announce, anticipate, argue, ask, 31
assert, assume, bet, calculate, concede,
confess, conrm, consider, deem, expect,
fancy, nd out, gather, hear,
imply, indicate, pretend, promise,
propose, read, recognize, recommend,
suggest, swear, take it, understand
Total n/a 57
4.8.3 Discussion
There are two takeaways from the Penn Treebank analysis: rst, factive verbs
occur naturally in bridge contexts, even though it is sometimes remarked that A-bar
movement from the complement clauses of these verbs is unacceptable.
239
Second, overall, the rate of occurrence is relatively high. The Penn Treebank
contains about a quarter million tokens. Assuming that learners are exposed to
915k2.5m utterances a year (Hart and Risley, 2003, 1995), this works out to about
168464 tokens, or from about one token every other day to just over one token a
day. To the extent that the range of sentences in the Penn Treebank is representative
of the linguistic experience for older speakers and learners of English, the results
suggest that learners have access to evidence that at least some factive verbs are
compatible with long-distance A-bar movement.
As I see it, the main issue that this study reveals concerns the robustness
of the evidence. Most of the tokens that English learners observe likely involve
know, and it is presumably easy for them to conclude that know is compatible with
long-distance A-bar movement. The other verbs occur much less frequently, if they
appear at all.
This result sits somewhat uneasily with the data collected in the formal ac-
ceptability judgment experiment and the informal survey. For example, the informal
survey showed that English native speakers can accept long-distance A-bar move-
ment with cognitive factive verbs like discover and notice, which do not occur with
long-distance A-bar movement in the Penn Treebank dataset.
One hypothesis that might explain how native speakers have these intuitions
despite sparse direct evidence is that they form generalizations about cognitive fac-
tive verbs as a class, namely, that long-distance A-bar movement is compatible with
this class of verbs. This particular claim awaits further empirical conrmation, as
the acceptability judgment study in Section 4.4 and the informal survey in Section
240
4.7 tested a relatively small number of these verbs. But it is consistent with the
data obtained so far, and also consistent with a number of descriptions of English
that claim that long-distance A-bar movement is compatible with these verbs (e.g.
Kastner, 2015, Kiparsky and Kiparsky, 1970, Abrus?n, 2014), and for the sake of
discussion in the following sections, I will tentatively adopt it.
4.8.4 Generalizing despite sparse input
In this section, I review two dierent approaches about how learners might
form generalizations from sparse input. I will rst consider the Suciency Princi-
ple proposed by Yang (2016, 2017), which has the advantage of making very clear
quantitative predictions about the evidence needed for forming generalizations. I
will show that the cognitive factive data in the Penn Treebank poses a problem for
this particular proposal. I then sketch an alternative Bayesian proposal.
4.8.4.1 The Suciency Principle
Yang presents a theory of computational eciency that explains how learners
might learn generalizations and exceptions. The core intuition underlying Yang's
proposal is this: for a rule to generalize, or for a particular morphological form to
become the default form, it needs to rst be observed to have applied in suciently
many instances; a mathematical formula, which he calls the Suciency Principle
(59), determines whether generalization occurs or not. When a generalization could
in principle apply either a superset or a subset of items, learners always attempt
241
to generalize to the superset. Only when generalization at the superset level is
impossible does the learner consider the possibility of generalization at the subset
level.
(59) Suciency Principle: Let R be a generalization over a set S with N candi-
dates, of which M items are attested to follow R. R can be extended to all
N items in S i: M ? N ? N ; otherwise, assume that R only applies for
ln(N)
the subset of M items. (adapted from Yang, 2017 ex. 12)
Yang discusses the limited productivity of the dative alternation in English:
the double object construction is only applicable to some, but not all, ditransitive
verbs (Ambridge et al., 2014; Baker, 1979; Bowerman, 1988; Perfors et al., 2010;
Pinker, 1984, 1989, etc.). But it is clearly intended to be applied to other kinds of
phenomena, and here I would like to evaluate it using the Penn Treebank dataset.
In the previous section, I reported that there were 57 attitude verbs that appeared
in long-distance A-bar movement contexts; in the context of this proposal, M = 57.
To determine what N is, I used the script described in Section 4.8.1 to estimate
the number of attitude verbs that appear with a complement clause. This yielded
an estimate of N = 279, after excluding idiomatic expressions like keep in mind or
have it (that) and the neologism grok.
Yang's formula yields a threshold of 279 ? 279 = 229.4: long-distance A-
ln(279)
bar movement must be attested with at least 230 verbs before it generalizes to the
broader set of attitude verbs. This threshold is clearly not met. This predicts
correctly that long-distance A-bar movement is not compatible with all attitude
verbs.
If long-distance A-bar movement from complement clauses of cognitive factive
242
verbs is generally acceptable to English speakers, and if Yang's proposal is correct
and the Penn Treebank is representative of an English speaker's linguistic experience,
then one expects that enough cognitive factive verbs appear with long-distance A-
bar movement in the Penn Treebank to satisfy the Suciency Principle.
Reviewing the set of 230 attitude verbs in the Penn Treebank, I was able to
identify 17 cognitive factive verbs (60), i.e. N = 17, which appear with complement
clauses for a total of 1,706 times. Applying the Suciency Principle formula yields
a threshold of 11, i.e. 11 verbs must appear in long-distance A-bar movement before
a learner generalizes long-distance A-bar movement to all cognitive factive verbs.
However, M = 7: only 7 verbs do so in the Penn Treebank dataset. According to
the Suciency Principle, there should be no generalization for the class of cognitive
factive verbs.
(60) Cognitive factive verbs appearing with complement clauses in Penn Tree-
bank dataset:
know* (988 times), realize* (154), see* (151), remember* (74), understand*
(60), learn* (52), notice (52), acknowledge (42), discover (42), nd out*
(30), forget (25), observe (15), recall (15), gure out (3), ascertain (1), rec-
ollect (1), grasp (1)
* = verb also occurs naturally with long-distance A-bar movement in the
Penn Treebank
While this outcome is clearly not ideal for the Suciency Principle, technically
speaking, this is a null result. One might therefore interpret this outcome as an
indicator that at least some linking assumption is faulty: perhaps the Penn Treebank
is not perfectly representative of the linguistic experience of the participants of the
acceptability judgment experiment or the informal survey, or if it is, perhaps the
dataset is too small a sample to yield useful counts for the purposes of the Suciency
243
Principle. Future research could, for example, apply the same counting methodology
to a larger dataset that might be a better approximation of the linguistic experience
of English speakers.
4.8.4.2 An alternative Bayesian-inspired approach
That said, just because long-distance A-bar movement fails to naturally occur
with a large set of verbs does not mean that all these verbs are incompatible with
long-distance A-bar movement. Because many of these verbs appear infrequently
with complement clauses in the rst place, it is perhaps not surprising to the learner
that she fails to observe long-distance A-bar movement co-occurring with these
verbs.
I would like to suggest an alternative Bayesian-inspired perspective on the
data that allows us to reconcile the idea that cognitive factives can generally appear
in long-distance A-bar movement contexts, even though only a very small number
of cognitive factive verbs naturally appear these contexts. In this approach, the
learner considers two hypotheses: the rst is the null hypothesis that long-distance
A-bar movement is not fully productive for a given class of verbs, while the second
is the alternative hypothesis that it is productive but infrequent. The learner then
considers the extent to which these hypotheses can explain the actual instances
of long-distance A-bar movement observed in her linguistic experience. It is in this
respect that this account is Bayesian (for a sample of applications of Bayesian models
on various domains of language acquisition, see, among others, Dillon et al., 2013;
244
Feldman et al., 2013; Pearl, 2007; Perfors et al., 2011a,b; Perkins et al., 2017; Xu
and Tenenbaum, 2007; Yang, 2016, 2017.).19
More specically, suppose that the learner generates a prediction about how
often long-distance A-bar movement occurs with each verb. The learner then com-
pares whether the actual attested frequency of long-distance A-bar movement for
that verb is comparable to the predicted value. We may say that long-distance
A-bar movement is conspicuously absent for this verb if it does not appear in
long-distance A-bar movement, contrary to the learner's predictions. The absence of
long-distance A-bar movement for this verb then serves as indirect negative evidence
against the alternative hypothesis that long-distance A-bar movement is productive.
To give an example: suppose a verb occurs with a complement clause 1,000
times, but of these 1,000 tokens there are no instances of long-distance A-bar move-
ment. Suppose further that the learner predicts that there should be 30 tokens
of long-distance A-bar movement. Because the observed number of long-distance
A-bar movement is 0, which is less than the predicted value, long-distance A-bar
movement is said to be conspicuously absent for this verb.
Following Yang's proposal, I assume that, if generalization is not possible for
a given class, learners will check whether generalization is possible for each of its
subclasses. Learners apply this process recursively until there are no subclasses left.
In broad terms, this analysis is similar to Yang's proposal in at least three ways:
19Note that as stated, the null hypothesis will always oer a perfect t for the data. In a Bayesian
framework, this means that the likelihood P (data|Hnull) = 1. Even if the alternative hypothesis
also oers a perfect t for the data, all else being equal, the learner should not be able to decide
in favor of one hypothesis over the other. To the extent that learners do decide in favor of the
alternative hypothesis on productivity (i.e. P (Halt|data) > P (Hnull|data)), one would need to
posit a prior bias for the alternative (i.e. P (Halt) > P (Hnull) ).
245
(i) it frames the learning problem as a matter of choosing between two (implicit)
hypotheses about whether the phenomenon in question is productive or not, (ii) it
assumes that the learner is sensitive to verb syntax and semantics, which is necessary
for identifying which verb classes to formulate a generalization for, (iii) it assumes
that learners proceed from superset to subset when formulating generalizations.
Importantly, it diers in terms of what learners are tracking. In the case of the
Suciency Principle, the learner is sensitive to the number of distinct verbs that
appear in long-distance A-bar movement; how frequently a verb appears with long-
distance A-bar movement does not matter at all. In contrast, this analysis assumes
that the learner can track a number of frequencies: how frequently a verb appears
with complement clauses and long-distance A-bar movement, which enables her to
make predictions about how frequently a verb should appear with long-distance
A-bar movement if that were a productive process.
4.8.4.3 Methods
The above learning scenario can be implemented as a simple Monte Carlo
simulation.
For concreteness, suppose that N distinct attitude verbs naturally occur p
times with complement clauses; call this set S. Within the set S, there are M
distinct verbs that naturally occur q times with long-distance A-bar movement.
From S, one can generate a random sample (with replacement) of q items. The
attitude verbs in this random sample can be understood as the attitude verbs that
246
appear with long-distance A-bar movement if long-distance A-bar movement were
fully productive (but infrequent) for this class of attitude verbs. How frequently
each of the N attitude verbs appears in the random sample can be calculated.
This random sampling process can then be repeated a fairly large number of
times in order to produce a range of how often a given verb should appear in a
long-distance A-bar movement context. Specically, I calculate a 95% condence
interval. I consider long-distance A-bar movement conspicuously absent for a verb
if the verb appears zero times in bridge contexts in the original Penn Treebank
sample and if the condence interval is above zero.
I apply this method on two sets of verbs: the rst is the full set of 279 attitude
verbs; for this set, N = 279, p = 20,756, M = 58, q = 596. The second set are the
cognitive factives: N = 17, p = 1,706, M = 7, q = 44.
4.8.4.4 Results and discussion
For the full set of attitude verbs, there were four verbs for which long-distance
A-bar movement was conspicuously absent. For space reasons, I will only report
the data for these four verbs (61).
(61) Verbs for which long-distance A-bar movement is conspicuously absent
Predicted number of
Tokens with complement long-distance A-bar movement
Verb clause instances (95% conf. int.)
note 225 26
mean 195 25
add 169 15
show 168 15
Although four out of 279 verbs might not seem to be a lot, these four verbs
247
are relatively frequent. Show, for example, is the 13th most common verb that
appears with a nite declarative complement clause. It is precisely this frequency
that makes the absence of long-distance A-bar movement conspicuous in the rst
place. An English learner who takes seriously these existence of such verbs might
thus conclude that long-distance A-bar movement is not compatible with all attitude
verbs.
As for the cognitive factive verbs, there were no verbs for which long-distance
A-bar movement was conspicuously absent, an outcome which is consistent with the
hypothesis that long-distance A-bar movement is productive for this class of verbs
and the results of the acceptability judgment experiment. Certainly, as mentioned
previously, there were ten verbs in the Penn Treebank for which there was no long-
distance A-bar movement attested: acknowledge, ascertain, discover, gure out,
forget, grasp, notice, observe, recall, recollect. However, these verbs are part of a
subset of cognitive factive verbs that are so low-frequency in the rst place that a
non-occurrence of long-distance A-bar movement was expected (62).
(62) Predicted instances of long-distance A-bar movement for cognitive factive
verbs in the Penn Treebank
248
Predicted number of
Tokens with complement long-distance A-bar movement
Verb clause instances (95% conf. int.)
know 988 2034
realize 154 19
see 151 18
remember 74 05
understand 60 05
learn 52 04
notice 52 04
acknowledge 42 04
discover 42 04
nd out 30 03
forget 25 03
observe 15 02
recall 15 02
gure out 3 01
ascertain 1 01
grasp 1 01
recollect 1 01
4.8.5 Prospects of applying these learning accounts to other instances
of bridge eects
One might wonder how eective this generalization mechanism (or the Su-
ciency Principle) might be when applied to other verbs or verb classes with similar
distributional and acceptability proles. For example, there are other English verbs
outside the class of cognitive factives for which long-distance A-bar movement is
relatively acceptable but occur infrequently in a corpus like the Penn Treebank. For
the sake of illustration, consider the verbs specify and vow, for which long-distance
A-bar movement is relatively acceptable: the z-score of long-distance A-bar move-
ment for specify is 0.58, while that for vow is 0.62 (36). However, both verbs do not
249
appear in long-distance A-bar movement congurations in the Penn Treebank, so
one might argue that learners lack direct evidence that these verbs are compatible
with long-distance A-bar movement.
I see two main challenges in applying these proposals to other instances. The
rst challenge is that we do not know for certain whether there is any learning with
respect to bridge eects for other verbs. The reason for applying these proposals to
the class of cognitive factive verbs is because we had good evidence (from English
and Dutch) to think that there might some learning that takes place for these
verbs. We lack the same kind of evidence for verbs like vow and specify : for all we
know, there is some innate discourse or pragmatic principle which ensures that long-
distance A-bar movement is cross-linguistic variation acceptable for verbs with vow
or specify semantics. If so, applying the Suciency Principle or the Bayesian-style
generalization proposal to these verbs would be beside the point.
Even if we had good reason to think that there is some learning implicated
for vow and specify (e.g. clear evidence of cross-linguistic variation), we might still
need to overcome another challenge, namely, identifying the relevant verb classes for
these verbs. Both the Suciency Principle and the above Bayesian-style approach
presuppose that learners try to generalize over verb classes. Sometimes, it is easy to
determine what the verb classes are: the discussion above looked at attitude verbs
and cognitive factive verbs, both of which are verb classes that are relatively easy
to describe and well-studied in the literature.
But sometimes identifying verb classes is not as straightforward, as is the case
for specify and vow : they are quite clearly non-manner of speaking communication
250
verbs, for sure, but what subclasses do they belong to? What other verbs are
in the same subclass? These questions might be answered through further cross-
linguistic investigation in bridge eects: perhaps we will see clearly which other
verbs pattern like specify or vow, which will let us identify the appropriate verb
class(es). But for now, the lack of a consensus on verb classes or a ne-grained
typology of attitude verbs constitute another challenge for empirically testing these
generalization proposals.
4.8.6 Additional comments on the results of the corpus analysis
Finally, the results show that a relatively large variety of clause-embedding
verbs can appear in bridge contexts. I would like to point out here that this nd-
ing contrasts with what D?browska and Verhagen reports. For example, D?browska
(2004) looked at wh-questions involving long-distance A-bar movement in the Manch-
ester child language corpus (Theakston et al., 2001). Of the 325 questions identied,
she reports that think and say appear in 96% of the questions (280 and 33 respec-
tively); the verbs in the remaining 4% of questions are tell (6), pretend (3), see (1),
hope (1), reckon (1), suppose (1) (p. 197).20.
This dierence is almost certainly attributable to the dierences in the size of
corpora used, the registers (child-ambient speech have simpler lexicons and syntax),
and the fact that this corpus study considered a wider range of long-distance A-bar
movement constructions. That said, the results replicate their nding that the vast
20For reasons unknown to me, the subtotals add up to 326 and not the 325 total reported by
D?browska (2004).
251
majority of tokens are centered around a very small number of verbs.
4.9 Conclusion
This chapter was divided into three distinct parts. First, I started with a
review of cross-linguistic variation in bridge verbs, including existing generalizations
about bridge eects. Second, I presented an acceptability judgment experiment
that was designed to address gaps in our empirical understanding of bridge eects
in English and to test existing claims. Analyses showed that current claims are
likely too coarse-grained to model the range of bridge eects. Subsequently, on the
basis of a post hoc analysis, I proposed a more ne-grained classication of verbs
that oers better empirical coverage. Third, I discussed whether the ne-grained
semantic distinctions that aect the acceptability of long-distance A-bar movement
might be more simply captured by general pragmatic principles, contrasting that
with a hypothesis where long-distance A-bar movement as a syntactic operation /
dependency is sensitive to the presence of certain verb classes. An informal survey
of native speakers of English and Dutch showed that the low acceptability of long-
distance A-bar movement with manner of speaking verbs can be ameliorated with
context, although the degree of amelioration varies. Survey results also provide some
evidence that cognitive factive verbs have dierent acceptability judgment proles
in English and Dutch.
To the extent there is a case of cross-linguistic dierence in cognitive factive
verbs, it suggests that, even if bridge eects are generally pragmatic in nature, there
252
are still some aspects that can be learned from one's linguistic experience. Taking
these verbs as a case study, I suggested that English speakers might have more access
to direct evidence that long-distance A-bar movement is compatible with cognitive
factive verbs, drawing evidence from an English corpus (the Penn Treebank). The
study also shows that the direct evidence might be sparse for many of these verbs,
which implies the existence of generalization. I discussed how learners might form
generalizations on the basis of sparse evidence, contrasting a Suciency Principle
approach with a Bayesian-inspired approach.
253
Chapter 5: Cross-linguistic variation and consequences for formal the-
ories of A-bar movement constraints
In previous chapters, I observed that cross-linguistic variation in A-bar move-
ment constraints poses challenges for the language learner. In this chapter, I turn
my attention to discussing the challenges and implications cross-linguistic variation
has for how we model these constraints formally, by reviewing existing proposals
about these constraints. To preview the discussion: much theorizing in the litera-
ture about these constraints is based on careful descriptive work on a small number
of languages and the reasonable null hypothesis that the same generalizations hold
cross-linguistically. This process yields very elegant descriptions of constraints, at
least for certain languages.
However, these theories tend to struggle to accommodate cross-linguistic vari-
ation. I would like to suggest that we should instead be comfortable with less elegant
descriptions that have better cross-linguistic coverage, on the condition that we can
give plausible (or better yet, elegant) learning accounts that can explain the range
of cross-linguistic variation. Let me be clear that I am not advocating that our
theories of syntax should just restate cross-linguistic dierences. Rather, what I
am proposing is that to the extent that cross-linguistic variation obliges us to forgo
254
elegant descriptions of the knowledge that adults have about their languages, we
should instead seek principled descriptions of the process by which they come to
possess that knowledge.
5.1 A critique of unied analyses of A-bar movement from relative
clauses and complement clauses
Since the two case studies reviewed in this dissertation both involve constraints
on A-bar movement from embedded clauses, it is not implausible to think that they
are amenable to a unied analysis. In this section, I review several such proposals
that have been put forward recently, showing that they do not adequately capture
the range of variation in constraints on A-bar movement from embedded clauses
within and across languages. I then consider how these proposals can be adapted
to improve their empirical coverage.
5.1.1 An event semantic approach: Truswell 2007, 2011
The core of Truswell's proposal is that A-bar movement is subject to an event
semantic condition, the Single Event Grouping Condition (among other conditions).
Empirically, Truswell intends for this event semantic condition to cover A-bar move-
ment from a set of adjuncts in which the verb is morphologically not nite (including
purposive clauses) and extends it to complement clauses for a select set of clause-
embedding verbs.
Truswell proposes that in congurations where A-bar movement is acceptable,
255
the event denoted by the predicate in the adjunct forms a single event grouping
with the event denoted by the matrix predicate. As an example, consider (1). The
matrix predicate denotes an event of going to the store, while the adjunct predicate
denotes an event of buying chocolate. Since it is a necessary condition for the
adjunct event, the matrix event is said to enable the adjunct event. In Truswell's
theory of event structure, both events then form an event grouping.
(1) a. John went to the store to buy chocolate.
b. What1 did John go to the store to buy t1?
c. That's the chocolate1 that John went to the store to buy t1.
Although Truswell treats tensed (nite) adjuncts to be uniformly islands (Truswell,
2011, pp. 175177), his account might work to a certain extent with A-bar movement
from nite porous relative clauses, a phenomenon not covered in his dissertation or
book. For example, in some contexts, the matrix predicate might be understood as
denoting an event that is an enablement relation with the event in the nite relative
clause. A-bar movement from the relative clause thus does not violate the Single
Event Grouping Condition. For instance, to the extent that (2) is acceptable, one
might argue that the president's desire to appoint someone enables that person to
lead the department in question. To the extent that an enablement relation exists,
A-bar movement satises the Single Event Grouping Condition.
(2) a. The president wants to appoint someone who can lead this department.
b. That's the department1 that the president wants to appoint someone
who can lead t1.
However, there are other acceptable examples where the matrix and adjunct
events might not stand in such a relation. Consider an example in the literature on
256
porous relative clauses, (3). In this event, the event denoted by matrix predicate
know[ing] a lot of people . . . is not a necessary condition for the event in the relative
clause, which is a state of being taken in by some trick: it is hard to see how such an
acquaintance relation is related to getting deceived. To the extent that they do not
form a single extended event, the event structure of this sentence in this context fails
to satisfy the Single Event Grouping Condition. We therefore predict that A-bar
movement out of the relative clause should be unacceptable, contrary to fact (3b).
(3) a. I've known a lot of people who've been taken in by this trick.
b. That's the trick1 that I've known a lot of people who've been taken in
by t1. (Chung and McCloskey, 1983, p. 708, ex. 9a)
The porous relative clause datapoints thus suggest that the Single Event
Grouping Condition does not fully handle all restrictions on A-bar movement from
adjuncts, pace Truswell. That said, this does not necessarily mean that the Single
Event Grouping Condition (more precisely, the intuition underlying the condition)
is invalid; it could well be the case that there are other conditions that determine
when A-bar movement from adjuncts is possible.
Truswell also extends his Single Event Grouping Condition to A-bar move-
ment from factive complements, e.g. the complement of a verb like regret. Although
Truswell acknowledges that there are conicting judgments about A-bar movement
from factive complements in English, he makes the assumption that these structures
are ungrammatical and attempts to derive the ill-formedness from the condition.
The intuition pursued is that in factive complements, the speaker presupposes the
existence of the event denoted by the embedded predicate. Consequently, this event
257
does not form a single event grouping with the event denoted by the matrix pred-
icate. In contrast, for non-factive complements (e.g. the complement of think), no
such presupposition exists; the event denoted by the embedded clause is an argu-
ment of the event described by the non-factive verb, and therefore forms a single
event grouping with it.
While ambitious in scope and oering many interesting observations, Truswell's
proposal seems to struggle more with A-bar movement from factive complements.
For one, it does not predict the fact that A-bar movement from factive predicates
like know and realize occurs naturally and is relatively acceptable to English speak-
ers. In addition, it appears to predict that A-bar movement from the complement
of manner of speaking verbs should be acceptable, since these verbs are non-factive
and so the events denoted in their complement clauses are not presupposed to ex-
ist. (To be fair, Truswell explicitly says in his book that he will not discuss A-bar
movement from the complements of manner of speaking verbs (p. 179 fn. 4).)
Even if we set aside these concerns with the Single Event Grouping Condition,
there is still a cross-linguistic problem that it presents. To the extent that the
Single Event Grouping Condition is universal, it predicts that all languages will
behave more or less like English, allowing A-bar movement from clausal adjuncts,
gerundival adjuncts, and even PPs. As Truswell himself recognized in his book, this
prediction is not borne out.
258
5.1.2 A discourse-based approach: Goldberg 2006 and Ambridge and
Goldberg 2008
As mentioned previously in Chapters 3 and 4, Goldberg (and colleagues) pro-
vides a discourse-based analysis for A-bar movement from relative clauses and pur-
posive clauses, centered around the notion of backgroundedness. In Goldberg, 2006,
she argues that rationale clauses and presentational relative clauses (i.e. porous
relative clauses) and complement clauses of certain non-factive verbs are all not
backgrounded. The evidence provided in the book on islands is almost all taken
from English. To the extent that this backgroundedness condition, motivated us-
ing English examples, is intended to be universal, it overgenerates, as it predicts
incorrectly that all languages should behave like English. (To be fair, Ambridge
and Goldberg (2008) also acknowledge that the backgroundedness proposal does
not fully address cross-linguistic variation, at least in the case of bridge verbs. They
observe, for example, that say is compatible with long-distance A-bar movement
in English, but not in Polish, although they do not say exactly how this dierence
can be derived (p. 384).)
Setting aside issues related to cross-linguistic variation, one limitation of this
proposal is the absence of a formal denition for backgroundedness. Instead, senten-
tial negation is presented as a diagnostic: a subordinate clause is taken to be back-
grounded if sentential negation impl[ies] a negation of the proposition expressed by
the subordinate clause (Goldberg, 2006 p. 143). The absence of formal denitions
or a thorough discussion of how sentential negation works, or why sentential nega-
259
tion (and not some other kind of negation or sentence-level operation) is particularly
relevant to backgroundedness, makes it sometimes dicult to see how island eects
can be attributed to backgroundedness.
The backgrounded island proposal runs into empirical diculties with complex
NPs modied by relative clauses, as Goldberg acknowledges. I discussed this in
greater detail in Chapter 3 (Section 3.2.1). To summarize the issue, Goldberg (2006)
considers the sentence in (4), and concludes that sentential negation can imply the
negation of the contents of the relative clause. The relative clause is therefore
considered as not backgrounded. A-bar movement from this relative clause is thus
predicted to be possible.
(4) She met a boy [who resembled her father]. (ibid. p. 147, ex. 63)
a. Sentential negation: She didn't meet a boy who resembled her father.
b. Content of relative clause: A boy resembled her father
c. Negation of contents of relative clause: A boy did not resemble her
father1
d. Sentential negation implies negation of relative clause, so the relative
clause is not backgrounded: She didn't meet a boy [who resembled her
father]. ? A boy did not resemble her father
Goldberg correctly observes that this prediction is not borne out, presenting
(5) as an example. Recognizing that this is a problem, she notes that an explanation
of this fact remains elusive (p. 148).
(5) *Who1 did she meet a boy [who resembled t1]?
Ambridge and Goldberg (2008) extend the backgroundedness hypothesis to
1Technically, the content of the relative clause could also be Some person x does not resemble
her father. In addition, negation could also scope high, yielding no boy resembled her father or
no person x resembles her father. Goldberg does not discuss these issues, and I will not do so
either.
260
bridge eects. Once again, they employ a negation test to measure background-
edness, where the degree to which a clause C is considered backgrounded varies
inversely with the extent to which main clause negation implies that C itself is
negated (p. 360).
(6) a. She didn't think that he left. ? He didn't leave.
b. She didn't shout that he left. 6? He didn't leave / He left.
(adapted from ibid. p. 360 exx. 26-27)
Used in this context, the test is especially problematic, as it seems to equate
non-backgroundedness with neg-raising. For illustration, consider a verb like think.
In Ambridge and Goldberg's framework, the negation test implies that the comple-
ment of think is not backgrounded. However, it is unclear whether this is a valid
conclusion to draw. One could imagine a scenario in which the complement of think
is actually backgrounded, and the entailment facts only obtain because think is a
neg-raising verb.
A further problem for this backgroundedness proposal is that it predicts that
A-bar movement from complements of factive verbs should be unacceptable. While
Goldberg and colleagues have provided support for this hypothesis, I note that this
prediction is not always borne out. For one, A-bar movement from factive verb com-
plements occur naturally in English. Second, Gibson (2018)/Liu et al. (2019) were
unsuccessful at reproducing these eects, and the acceptability judgment experi-
ment reported in Chapter 4 also failed to elicit the same eects. Statistical analysis
of the acceptability judgments also showed that backgroundedness explained only a
small portion of the variance in acceptability in long-distance A-bar movement.
261
5.1.3 A distributional learning approach: Pearl and Sprouse 2013a,
2013b
Pearl and Sprouse (2013a,b) model island eects as a consequence of distri-
butional learning: learners use certain statistical facts about A-bar movement to
deduce what the relevant constraints might be for their language. Pearl and Sprouse
assume that learners can parse all wh-dependencies in an adult-like fashion. In par-
ticular, they assume that learners represent each wh-dependency as a sequence of
container nodes: the maximal projections that dominate the gap but not the wh-
phrase, with the addition of start and end symbols. The learner then analyzes
these sequences as trigrams of container nodes, whose frequencies are tracked in the
course of language acquisition.
(7) a. [CP Who1 did [IP she [V P like t1]]]?
b. Container nodes sequence: start-IP-VP-end
c. Container node trigrams: start-IP-VP, IP-VP-end
(Pearl and Sprouse, 2013b p. 36 ex. 12)
The acceptability judgment native speakers assign to a wh-dependency is fur-
ther modeled as the probability of observing the container node sequence that rep-
resents that dependency, which they dene as the product of trigram probabilities.
In other words, the acceptability associated with the wh-dependency in a novel ex-
pression like What spaceship1 did the nurse paint t1, which has an identical set of
container node trigrams as (7a), is estimated by the product of the frequencies of
start-IP-VP and IP-VP-end.
Based on a study of wh-dependencies in naturally occurring contexts and com-
262
putational modeling, Pearl and Sprouse show that the container node sequence and
trigrams associated with island-violating A-bar dependencies  like those that in-
volve A-bar movement from a relative clause  are of very low probability, even
after applying a smoothing constant to raise the frequencies of these trigrams to
just above zero. This fact correctly predicts the low acceptability associated with
these dependencies. In contrast, because A-bar movement from nite complement
clauses is attested in the input (although rare), the container node approach is
able to obtain non-zero frequencies for the relevant trigrams, which explains why
long-distance A-bar movement is relatively more acceptable.
The container node trigram approach is elegant, in that it ideally requires the
learner to track only syntactic information in the form of container nodes and the
frequencies of these trigrams. This simplicity poses a problem for the modeling of
the acceptability of A-bar movement from clausal adjuncts and bridge eects. The
argument goes as follows. Underlying Pearl and Sprouse's model is the intuition that
the acceptability of an A-bar dependency varies directly with two factors: the length
of the dependency and the frequency of each container node trigram associated with
the dependency.
For illustration purposes, consider the following minimal pair.
(8) a. This is the paper1 we need to nd some person [who understands t1].
b. *This is the paper1 we need to nd the person [who understands t1].
I rst argue that the length of the dependency in (8a) (a porous relative clause)
is as long as the one in (8b) (a regular relative clause), if not longer. All else being
equal, the acceptability of (8a) should be lower than that of (8b).
263
Suppose that there is some kind of covert movement of the complex NP in
(8b) and A-bar movement applies after covert movement, as proposed in Sichel,
2018 and in the movement/height-based analysis in Chapter 2. The LF position
of the complex NP in (8b) is thus higher than the LF position of the complex
NP in (8a), and therefore A-bar movement passes through fewer projections in
(8b). Alternatively, suppose that there is no covert movement of the complex NP,
pace Sichel. In that case, A-bar movement passes through the same number of
projections.
Next, for the sake of argument, suppose that the length of the dependencies
are identical. In that case, the relative acceptability of (8a) requires at least some
of the container node trigrams in (8a) to be more frequent than that in (8b). Pearl
and Sprouse appear to assume that the learner tracks coarse-grained, relatively
universal syntactic categories (e.g. IP, VP, but not whether the IP is nite or not,
nor whether the VP is passive or not; p. 35). This assumption means that there
is no dierence in terms of the frequencies of the container node trigrams. (The
only exception they make is for material in the CP layer, in order to model wh-
island eects and the that-trace eect. Specically, they assume that learners track
whether there is a wh-phrase in the specier position or whether it is headed by an
overt complementizer, such as that.)
As I see it, for the container nodes proposal to plausibly model the dierence
in acceptability, one could assume that learners are tracking more abstract, ne-
grained information. To see why this is so, for the sake of argument, suppose that
both dependencies are equally long, and that the distinction separating the examples
264
is ultimately due to the relative clause in (8a) containing defective tense, while the
relative clause in (8b) contains independent tense. Under these assumptions, we
might be able to derive the dierence in acceptability by supposing that there are
more instances of A-bar movement that cross the projection of a defective tense head
than the projection of an independent tense head. To do so, though, one would have
to assume that learners make a distinction between defective tense (both nite and
non-nite forms, which are both dependent) and non-defective (independent) tenses,
and encode defectiveness in their representation of container nodes.
A similar issue occurs for bridge eects, if these are to be modeled syntactically:
in container node terms, A-bar movement from the clausal complement of a bridge
verb is identical to A-bar movement from the clausal complement of a non-bridge
verb. To the extent that bridge eects have a syntactic explanation, one must
either introduce additional conditions on A-bar movement, or assume that learners
are sensitive to and track what verbs subcategorize for a clause, or deny the premise
that the complements of bridge verbs and non-bridge verbs are identical, contrary
to appearances.
More specically, to model bridge eects, one could stipulate that the learner
lexicalizes a trigram if it involves a verb and its nite complement. Learners then
track the frequency of these lexicalized trigrams and generalize accordingly. This
analysis would be a straightforward extension of Pearl and Sprouse's proposal. As
mentioned, they already assume that learners can track lexical items in a CP, for
example, whether there is an overt complementizer like that. While easily imple-
mented, this solution might strike one as contrary to the spirit of Pearl and Sprouse's
265
proposal, which at its heart assumes that learners use simple, coarse-grained syn-
tactic categories (which do not contain lexical information) to learn about island
constraints.
5.2 A-bar movement from relative clauses
In this section, I discuss consequences specic to the observation that A-bar
movement from relative clauses is possible under exceptional circumstances. I orga-
nize my discussion to start with broader architectural issues.
5.2.1 Implications for derivational approaches to adjunct islandhood
As Truswell (2007, 2011) points out, the proposal that A-bar movement from
adjuncts is possible in limited has consequences on a set of approaches found in
the syntax literature on adjunct islandhood, such as Hunter, 2015; Stepanov, 2007;
Uriagereka, 1999. These approaches elegantly attribute island properties of adjuncts
to some basic property of the derivational system.
Abstracting away from technical details, these derivational approaches claim
that some basic requirement of the grammar causes an adjunction operation to
become incompatible with A-bar movement from inside the adjunct. For example,
for Uriagereka (1999), this requirement is related to linearization, which forces an
adjunct to be Spelled-out and become opaque to A-bar movement; for Stepanov
(2007), the requirement is a Least Tampering Condition that causes adjuncts to
be merged postcyclically, after A-bar movement operations have applied.
266
In general, these proposals, to quote Boeckx (2012), are not modiable enough
for the linguist. But importantly, they are also unlikely to be modiable enough
for the learner. To accommodate the fact that A-bar movement from an adjunct
is possible, the basic grammatical requirement needs to be waived in certain con-
texts. While such an assumption would capture the facts, it is an ad hoc solution
to the problem. This solution implies that learners can alter basic properties of the
derivational system, which is conceptually unappealing.
For illustration, consider the inuential Multiple Spell-out proposal of Uriagereka,
1999. In this proposal, the linear ordering of any two lexical items A and B is entirely
dependent on whether A asymmetrically c-commands B or not. This linearization
algorithm runs into a problem when dealing with adjuncts and speciers of a con-
stituent XP (9). A lexical item inside the adjunct YP, A, cannot be ordered with
a lexical item outside of the adjunct, B, as neither asymmetrically c-commands the
other.
(9) XP
YP X?
. . . A . . . . . . B . . .
Uriagereka thus proposes that when two complex structures are to be merged
together, one must be Spelled-out, so that it is no longer syntactically complex but
is instead linearized and behaves like a lexical compound (p. 115). He assumes
that Spell-out applies to adjuncts and speciers, and so these constituents become
inaccessible to further syntactic operations like A-bar movement. The asymmet-
ric c-command requirement of the linearization algorithm thus derives the desired
267
generalization that adjuncts (and speciers) are islands.
This analysis however needs to be reconciled with the fact that adult native
speakers of languages like English allow A-bar movement from certain adjuncts.
Technically speaking, the most direct way is to claim that this linearization al-
gorithm is relaxed for these adjuncts. However, such a solution seems ad hoc:
asymmetric c-command is a fundamental property of the (universal) linearization
algorithm, and is therefore not something that we want to be able to relax or waive
under language-specic circumstances. The same ad hoc problem exists from a
learnability perspective: it is dicult to imagine that a learner can learn to make
changes to the linearization algorithm to accommodate counter-examples in the in-
put.
5.2.2 Remarks on syntactic parameters and restricted hypothesis spaces
In Section 3.6, I argued that there are (presumably innate) restrictions on
a learner's hypothesis space that ensure that learners end up drawing conclusions
about A-bar movement from porous relative clauses from their observations of A-bar
movement from purposive clauses, even though these two adjuncts are supercally
very dierent. This argument goes as follows: In principle, there are many logically
coherent conclusions one might be able to draw about one's grammar from observ-
ing A-bar movement from purposive clauses; for instance, one might take these
observations to mean that A-bar movement from all adjuncts is possible. However,
in the case of English and Mandarin Chinese at least, learners seem to converge
268
on a more conservative conclusion: A-bar movement from porous relative clauses
is also possible. Convergence can either be attributed to learners successfully rul-
ing out alternative hypotheses on observing tokens that are inconsistent with these
hypotheses, or to some innate bias that prevents learners from even considering
these hypotheses in the rst place. Since A-bar movement from purposive clauses
is infrequent in languages like English and Mandarin Chinese, the rst scenario is
unlikely.
This kind of restriction on hypothesis space is core to the notion of syntactic
parameters, and so it is useful to consider this argument in light of current views on
syntactic parameters. Parameters have been criticized on conceptual and empirical
grounds. Depending on how they are implemented, they can be stipulative, which is
inconsistent with the goals of Minimalist theories. Some proposals are too powerful,
predicting (incorrectly) that certain phenomena should be absent (e.g. Gilligan's
(1989) typological work on Rizzi's proposals about null subjects). At the same time,
languages appear to vary along too many morphosyntactic dimensions to be easily
captured with a small number of parameters.
While innate parameters/restrictions on hypothesis spaces are perhaps not ef-
fective for capturing the full range of cross-linguistic variation, I concur with Chac?n
(2015) that this limitation does not constitute an argument for eliminating them
from our theories of syntax. They are especially crucial in accounts in which learn-
ers use indirect evidence, in which the evidence is always (by denition) dierent in
some way from the phenomenon to be learned. Consequently, if there is no restric-
tion, there would be no way to guarantee that learners will draw the appropriate
269
inference about the target phenomenon. From a theoretical perspective, what this
limitation means is that we ought to be much more judicious about when we in-
voke notions such as parameters/restrictions on hypothesis spaces to account for
cross-linguistic variation.
5.2.3 A critique of two recent approaches on A-bar movement from
relative clauses
In this section, I summarize two recent proposals that attempt to explain A-
bar movement from relative clauses and related clausal adjuncts, comparing their
ideas and predictions with the ones presented in Chapters 2 and 3.
5.2.3.1 A syntactic approach: Sichel 2018
Sichel's goal is to account for A-bar movement from nite relative clauses in
Hebrew. Adapting her proposal to account for the English facts is straightforward,
and I have done so in Chapter 2. Here, I describe her account in greater detail.
There are two components to her account. First, she assumes that the rel-
evant nite relative clauses are derived via a raising analysis, adapting propos-
als of Bianchi, 2000; Brame, 1968; Kayne, 1994; Preminger, 2010; Schachter, 1973;
Vergnaud, 1974, etc. Specically, she assumes that there are (at least) three poten-
tial landing sites for a wh-phrase in the left periphery of the clause: the speciers
of CP, FP, and Foc(us)P.
The raising analysis she proposes is illustrated in (10). The complement of
270
the indenite determiner a is a CP; the relative clause does not adjoin to a nominal
projection, contrary to more conventional analyses of relative clauses. The highest
wh-phrase in the CP raises to only Spec,FocP, and the NP inside this wh-phrase
moves to the specier of a higher FP. Importantly, nothing occupies the edge of CP.
(10) [DP a [CP . . . [FP person2 . . . [FocP [which t2]1 . . . t1 can understand this
paper]]]]
(adapted from Sichel, 2018, pp 353354, ex. 34)2
In the case of A-bar movement from a complex NP like (10), the other wh-
phrase moves directly to the Spec,CP of the complex NP. It is then free to move
from this position to a higher position (11).
(11) [Which paper]3 do we need to nd [ ?DP a [CP t3 [FP person2 . . . [FocP [which
t2]1 . . . t1 can understand t3]]]]?
On its own, this rst assumption overgenerates, as it predicts, incorrectly,
all kinds of (raising) relative clauses permit A-bar movement from within. Sichel
thus supplements this assumption with a freezing condition (in the spirit of the one
proposed in Wexler and Culicover, 1980): constituents in derived A-positions are
opaque to subextraction (her ex. 38, p. 356).
This part of Sichel's analysis is adopted in Chapter 2, and I will present it
again, here, for ease of reference. The freezing condition is connected directly to
Sichel's observation that A-bar movement is only possible from relative clauses that
are part of complex NPs that are not presuppositional, which includes NPs that
2The raising analysis itself does not account for the fact that the wh-element in such an example
surfaces as who and not which, i.e. the relative clause is realized as . . . a person {who/*which} can
understand this paper. Presumably the who / which alternation takes place at a post-syntactic
level.
271
receive a non de re reading or NPs in an existential construction.
Sichel's movement-based analysis is elegant, and reduces the nuances in the
island constraint associated with relative clauses to an independently-motivated con-
straint on A-bar movement, namely, freezing. That said, there are several limitations
with this proposal.
The rst relates to why NPs move, which determines whether it freezes and
thus whether A-bar movement is possible. As far as I can tell, Sichel does not present
a full theory of A-movement, only observing that her proposed [movement of NPs]
is closer to movement in the domain of Dierential Object Marking, although she
rules out QR and/or type conict resolution as possible explanations of A-movement
(her footnote 20, p. 355). In the absence of a more specic theory of NP movement,
this component of the proposal is dicult to evaluate.
The second kind of limitation relates to empirical coverage. Sichel's paper is
concerned with A-bar movement from Hebrew relative clauses, is therefore mostly
silent about cross-linguistic variation and cross-linguistic predictions. For this rea-
son, it would be unfair to hold issues related to cross-linguistic variation against
Sichel's proposal, since there is no intent on her part to make claims about A-bar
movement from relative clauses across languages, even though, as mentioned in
Section 2.3.2.3, Sichel is aware of cross-linguistic variation.
With this all said, I would like to point out a few issues related to cross-
linguistic variation that Sichel's paper does not address. These happen to also be
issues that I address in this dissertation.
Because Sichel's paper is silent about cross-linguistic variation, the paper is
272
also silent about how speakers might come to acquire the relevant constraint for their
language. In contrast, in Chapters 2 and 3, I consider cross-linguistic dierences,
observe that the relevant direct experience is very sparse in English and Mandarin
Chinese, and develop a learning account based on indirect positive evidence.
Second, to the extent that subjects are in derived A-positions (i.e. subjects
move from some VP-internal position and freeze), relative clauses that are part
of subjects should not allow A-bar movement. As pointed out in Chapter 2, this
prediction turns out to be too strong for Mandarin Chinese. It is in response to
datapoints like this that I also consider a defective tense analysis in Chapters 2 and
3.
5.2.3.2 A processing approach: Kush et al., 2013
Kush et al. (2013), building on observations by Allwood (1982) and Eng-
dahl (1997) for Scandinavian languages, observe that A-bar movement from relative
clauses in English is sensitive to the type of verb that selects the complex NP. In
this paper, they consider the hypothesis that the verbs that allow A-bar movement
from a relative clause are those that select small clause complements, i.e. verbs
like see, nd, and know. With a series of formal experiments, they show that A-bar
movement from relative clauses is signicantly more acceptable with main verbs that
select small clauses than with main verbs that do not, providing support for their
hypothesis. To restate their ideas in term of porousness, a porous relative clause (in
English) is one modifying the nominal complement of a verb that can also take a
273
small clause.
However, in absolute terms, A-bar movement from their relative clauses was
rated as marginal at best (p. 254), not unlike the experimental results reported
in Section 5.3. To explain this marginal acceptability, Kush et al. propose the idea
of partial amelioration. Consider the following example (12).
(12) That is the bill that he saw many senators who supported at the congress.
(ibid. p. 257, ex. 20a)
The parser rst encounters a ller, the bill, and the complementizer that, and
concludes that there is a A-bar movement gap in the sentence. Upon encountering
see, the parser realizes that see can take either a noun phrase or a small clause com-
plement. Encountering many senators causes the parser to favor a clausal analysis
where many senators is the subject; this is because English allows a A-bar move-
ment gap in the predicate of a small clause (cf. This is the bill1 that he saw [SC
many senators support t1].). English also allows gaps inside a noun phrase, but the
noun phrase has to be headed by a picture noun or the nominalized form of a verb
(e.g. Davies and Dubinsky, 2003), and many senators is not such a noun phrase.
However, the sentence as presented is incompatible with the small clause analy-
sis. Assuming that A-bar movement from relative clauses is actually ungrammatical
in English, Kush et al. argue that at this point, the parser starts to repair the
representation, converting it from a subject relative clause to a small clause, so that
it would be well-formed for the representation to contain a gap. Although they do
not provide an example of the repaired representation, it is presumably something
like (13); note that the tense of the verb in the relative clause must change for the
274
small clause analysis to be well-formed. Kush et al. suggest that using a discarded
analysis to repair the representation is costly, which degrades the acceptability of
the sentence.
(13) After repair: That is the bill that he saw many senators support at the
congress.
This can be contrasted with (14), which is less acceptable. Observing thatmeet
does not take a small clause complement, they argue that there is no discarded small
clause analysis that the parser can use. As a result, it is not possible for the parser
to repair the representation to get a well-formed one with a gap.
(14) *That is the bill that he met many senators who support at the congress.
(ibid. p. 257 ex. 20b)
Kush et al.'s proposal is notable in that it is the rst published empirical
study of the acceptability of A-bar movement from relative clauses in English, and
also the rst to present a concrete proposal about the marginal acceptability of
A-bar movement from porous relative clauses. However, this proposal has several
weaknesses.
The rst is that the repair hypothesis over-generates. For example, consider
a sharply unacceptable sentence like (15a) where A-bar movement has taken place
from a denite complex NP. All else being equal, the parser should be able to use a
small clause analysis to repair the representation, yielding a structure like (15b). As
far as I can tell, the repair steps needed are identical to the ones for (12). However,
there appears to be no partial amelioration for (15a).
(15) a. *That was the bill1 that he saw [NP the senators who supported t1 at
275
the congress].
b. After repair: That was the bill1 that he saw [SC the senators support
t1 at the congress].
A second problem is that it does not generalize easily to other languages.
Consider Mandarin Chinese. For one, the main predicates in the Mandarin examples
reviewed above do not take small clauses. For example, zhaodao to nd in (16a)
is the main verb of a sentence that allows A-bar movement from a relative clause.
However, zhaodao does not take a small clause complement; it does not have the to
perceive reading associated with English to nd. In fact, it only has the sense of to
locate (something/someone) (16b). If the parser repairs (16a), it does not do so by
using a discarded small clause analysis.
(16) a. [Zheme wanpi de haizi]1, wo zhao-bu-dao [yuanyi shouyang t1
so naughty MOD child I -can't-nd willing adopt
de] ren.
MOD person
`Such a naughty child, I cannot nd someone who is willing to adopt
him/her.' (Zhang, 2002 ex. 5c) )
b. *Lisi zhaodao [ta shagua].
Lisi nd 3sg fool
Intended: `Lisi nds him/her a fool.' (compare with dang to treat,
Lisi dang ta shagua Lisi treats him/her as a fool)
Further, even if zhaodao did take a small clause complement, it is dicult
to see how the parser might repair the representation in a way that minimizes
cost. Mandarin subject relative clauses precede the noun they modify, producing
predicatesubject word order: in (16a), the predicate yuanyi shouyang . . . willing
to adopt . . .  precedes the noun that serves as its subject, ren person. However,
for all other clauses (setting aside right dislocation constructions where the subject
appears at the end of the sentence), the word order is reversed: the subject precedes
276
the predicate. If (16a) is acceptable because of repair, one needs to assume that
the repair process here involves a drastic revision of the word order so that zhaodao
would have a small clause complement. Such a change should be extremely costly.
However, there is no sharp degradation in acceptability.
5.3 Comments on the acceptability of A-bar movement from porous
relative clauses and purposive clauses
As mentioned, Kush et al. (2013) report a formal acceptability judgment ex-
periment where A-bar movement from porous relative clauses in existential and
evidential existential environments in English were judged to be of marginal ac-
ceptability. In this section, I describe an acceptability judgment experiment that
evaluates the acceptability of A-bar movement from two types of adjuncts in English:
porous relative clauses in non de re environments and from purposive clauses.
To preview the results, the judgment data show quite clearly that A-bar move-
ment from purposive clauses is acceptable, validating existing reports (e.g. Truswell,
2007, 2011). They also conrm the nding by Kush et al. (2013) that A-bar move-
ment from porous relative clauses (in this case, the non de re construction) is more
acceptable compared to canonical A-bar movement from relative clauses, although
in absolute terms, A-bar movement from porous relative clauses is not perfectly
acceptable. Pace Kush et al., I argue that the intermediate acceptability does not
necessarily show that A-bar movement from porous relative clauses is ill-formed,
pointing to similar lowered acceptability of dependencies involving non-island nite
277
clauses and reported cross-linguistic variation. In addition, the reduced acceptabil-
ity of A-bar movement from porous relative clauses can be plausibly connected to
parsing diculty.
5.3.1 Methods
5.3.1.1 Participants
34 adults based in the United States who passed an American English native
speaker prociency test were recruited via Amazon Mechanical Turk. They received
$3 in monetary compensation.
5.3.1.2 Materials
The experiment had a 2x2 factorial design, combining adjunct type (rela-
tive clause and purposive clause) and whether A-bar movement had taken place
from the adjunct. In these contexts, the sentences with purposive clauses are near-
paraphrases of the sentences with relative clauses. Any dierence in acceptability
is thus most likely attributable to syntactic dierences between these two construc-
tions.
(17) Porous relative clauses: relative clauses in a non de re construction
a. This is the CEO who really has to hire a senior executive who can
oversee the troubled business. [A-bar movement from outside relative
clause]
b. This is the troubled business that the CEO really has to hire a senior
executive who can oversee. [A-bar movement from relative clause]
(18) Purposive clauses
278
a. This is the CEO who really has to hire a senior executive to oversee
the troubled business. [A-bar movement from outside purposive
clause]
b. This is the troubled business that the CEO really has to hire a senior
executive to oversee. [A-bar movement from purposive clause]
16 test sentences were created with dierent lexicalizations. The test sentences
were then distributed across four lists in a Latin Square fashion, so that each partic-
ipant read four test sentences per condition and saw each lexicalization only once.
The test sentences were then combined with 12 acceptable llers and 20 unaccept-
able llers; assuming the four sentences with A-bar movement from relative clauses
to be relatively unacceptable, this distribution of llers and sentences ensured that
participants saw 24 acceptable sentences and an equivalent number of unaccept-
able sentences. Among the llers were four sentences with A-bar movement from a
relative clause adjoined to a denite complex noun phrase; an example is given in
(19).
(19) The spokesperson conrmed that these were the children that the mayor
rewarded the reghters who rescued. [A-bar movement from regular
relative clause modifying denite NP]
5.3.1.3 Instructions
Participants were instructed to rate each sentence along a 7-point Likert scale
based on how natural they sounded or if they would say such a sentence, with 1
being the least natural, and 7 being the most natural. Before starting the experiment
proper, participants were presented with six practice sentences. Participants were
free to take as much as time as they needed to give a judgment.
279
5.3.2 Results
To control for individual biases and variability in assigning ratings, accept-
ability ratings were normalized at the participant level by calculating z-scores: after
calculating a mean and a standard deviation for each participant, the mean value
was then subtracted from each rating and the result divided by the standard devi-
ation to obtain the z-score.
Data from one participant were excluded. Across all items, this participant
took only an average of 1.3 seconds to give a rating, when the median participant
took about 10 seconds to do so. This participant's fast response suggests that he/she
might not have completed the task as instructed.
The results show a three-way split in acceptability: A-bar movement from
a purposive clause is signicantly more acceptable than A-bar movement from a
porous relative clause (mean z-score: 0.30 vs. ?0.40, p < 0.01), which is in turn
also signicantly more acceptable than A-bar movement from a relative clause that
is part of a denite complex noun phrase (19) (mean z-score: ?0.91, p < 0.01)
(Figure 5.1).
The dierence between porous relative clauses and purposive clauses cannot
be entirely attributed to the adjunct type alone. An ANOVA analysis shows that
there are signicant eects for gap position, adjunct type, and the interaction of the
two factors (all three p-values < 0.001). In other words, there is a superadditivity
eect for A-bar movement from porous relative clauses relative to A-bar movement
from purposive clauses (Figure 5.2).
280
Figure 5.1: A-bar movement from adjuncts
Purposive clause
Porous RC
RC in definite NP
?1.0 ?0.5 0.0
Z?score
Figure 5.2: Superadditivity in A-bar movement from porous relative clauses
0.50
0.25
Adjunct
Porous RC
0.00 Purposive clause
-0.25
-0.50
Gap outside adjunct Gap inside adjunct
Gap location
281
Z-score
Adjunct type
Figure 5.3: Distribution of z-scores for A-bar movement from porous relative clauses
9
6
3
0
?0.50 ?0.25 0.00
Mean z?score
An analysis of the judgments for A-bar movement from porous relative clauses
also shows that the lower acceptability is not a statistical eect due to the presence
of two distinct groups of participants, one that judges these sentences to be perfectly
acceptable and one that does not. Figure 5.3 shows that participants' z-scores for
these sentences have a unimodal distribution.
5.3.3 Discussion
The two-way contrast for A-bar movement from relative clauses corroborates
claims in the literature, which present examples like (17) as exceptions to the rela-
tive clause constraint. The acceptability of purposive clauses is also consistent with
existing claims in the literature (Truswell, 2007, 2011), even though this accept-
ability is, perhaps, unexpected, given that purposive clauses are adjuncts and the
hypothesis that adjuncts are islands.
Complicating the picture, however, is the fact that A-bar movement from
282
Number of participants
porous relative clauses is not fully acceptable, especially when compared to A-bar
movement from purposive clauses. This split is perhaps unexpected: informal judg-
ments in the literature have tended to present these examples in a way that gives
the impression that they are both completely acceptable; example sentences are of-
ten presented without asterisks or question marks that indicate degradedness. That
said, for some researchers, the absence of these diacritics might merely indicate
acceptability relative to some baseline that is salient in the context, e.g. A-bar
movement from regular relative clauses, which is much less acceptable.
Nevertheless, to the extent that the intended claim is that A-bar movement
from porous relative clauses and purposive clauses is perfectly acceptable, the judg-
ment data presented above are inconsistent with these reports. However, it might
be hasty to expect the purposive clause examples to be directly comparable to the
porous relative clause examples in acceptability. After all, these constructions are
not perfect minimal pairs, and their dierences might contribute to the acceptability
dierence.
First, they dier in niteness. There is independent experimental evidence
that all else being equal, dependencies are more acceptable when they cross the
boundary of a non-nite clause than that of a nite clause. For example, Michel
and Goodall (2013) demonstrate this contrast for A-bar movement from various is-
lands in English. Another study is reported in Grano and Lasnik, 2018, pp. 474480.
As mentioned above, Grano and Lasnik looked at how certain clause-bound depen-
dencies can exceptionally cross into nite clauses when the clauses contain bound
pronoun subjects (20). They measured the acceptability of these constructions with
283
a formal experiment. They found that relative to their non-nite counterparts (20a),
the nite clause examples (20c) were less acceptable.
(20) Too/enough construction (ibid.) (=(44) in Chapter 2)
a. This book is too valuable for James to lend _ to Bill.
b. *This book is too valuable for James to claim [that Mark lent _ to Bill].
c. ?This book is too valuable for James1 to claim [that he1 lent _ to Bill].
The contrast they observed thus parallels the niteness contrast found above
for porous relative clauses and purposive clauses. Further, to the extent that the
porous relative clause examples are the tense analogues of the bound pronoun sub-
ject phenomena, as set out in the defective tense account (Section 2.1.4.2), the
intermediate acceptability of A-bar movement from porous relative clauses and the
bound pronoun subject phenomena is not unexpected.
I will have to leave for future research the questions of why niteness has
such an eect on acceptability, and to what extent this eect holds true cross-
linguistically. Researchers like Kluender (2004) and Hofmeister (2007) have at-
tempted to cast this dierence as a processing eect, although the specics of such
an account appears to be unclear to me. One conclusion that we can draw from
these two sets of results is that we cannot use this nite/non-nite contrast to exclu-
sively diagnose islandhood, even though it is tempting to do so on the basis of the
results in Michel and Goodall, 2013. The three dependencies experimentally tested
by Grano and Lasnik (2018) cross the nite and non-nite clausal complements of
verbs like claim and promise, which are not islands in English by any conventional
denition.
284
Another dierence between the purposive clause and porous relative clause
examples, which we might more easily connect with the acceptability dierence, is
that relative clauses are characterized by A-bar movement, and so A-bar movement
from a relative clause involves the structural conguration in (21).
(21) . . . wh1 . . . [RC wh2 . . . t2 . . . t1 ]
In Chapter 2, I suggested that the conguration might be unacceptable because
it is relatively dicult to process, and not because it is ill-formed per se. At the
point wh2 is processed, the parser has to resolve two A-bar dependencies within
the same relative clause. This task might give rise to similarity-based interference
(Lewis, 1996; see also Hofmeister, 2007 and Kluender and Kutas, 1993 for similar
comments), thus reducing acceptability.
In contrast, there is no independent evidence that the purposive clauses (more
precisely, rationale clauses) are characterized by A-bar movement. Consequently,
A-bar movement from a purposive clause does not involve the same structural con-
guration. There is thus no equivalent similarity-based interference, which explains
the greater acceptability.
Of course, one might argue that A-bar movement from relative clauses is actu-
ally ill-formed (ungrammatical). More specically, one might see this conguration
as violating the principle of Relativized Minimality (Rizzi, 1990), since the A-bar
movement of wh1 skips an A-bar position (the one occupied by wh2, the relative
pronoun or operator that characterizes the relative clause). In contrast, to the ex-
tent that rationale clauses are not characterized by A-bar movement, there is no
285
violation of Relativized Minimality.
In a framework where grammaticality is binary, i.e. representations are either
well-formed (grammatical) or ill-formed (ungrammatical), this analysis means that
there is no grammaticality distinction between A-bar movement from porous relative
clauses and A-bar movement from regular relative clauses. To accommodate the
fact that there is an acceptability dierence, one could posit some kind of repair
mechanism that succeeds with porous relative clauses but not with regular relative
clauses, as Kush et al. (2013) do. Alternatively, one might claim that grammaticality
is actually gradient (see work as early as Chomsky, 1965; for the notion of gradient
well-formedness in the context of A-bar movement, see e.g. Chomsky, 1986): in
contrast to A-bar movement from regular relative clauses, A-bar movement from
porous relative clauses is more acceptable because these relative clauses do not
freeze / are structually low enough (under the movement/height-based analysis) or
are not phases (under the defective tense analwysis). A-bar movement from porous
relative clauses in this regard is thus relatively well-formed.
Clearly, it is not a trivial matter to tease apart these dierent accounts.
All things considered, though, the rst account, in which A-bar movement from
porous relative clauses is ill-formed but more acceptable because of processing ef-
fects, seems to me to be the most attractive: it allows us to maintain a simpler
binary model of grammaticality, while deriving the marginal acceptability as an in-
stance of similarity-based interference, a relatively well-motivated psycholinguistic
eect. In contrast, developing a plausible repair account is not as straightforward,
as pointed out in the above section on Kush et al.'s proposal (Section 5.2.3.2), while
286
the gradient grammaticality account requires us to posit a more complex theory of
grammaticality.
5.4 Bridge eects
To recap, the main issue, inherited from Ross's dissertation, is the basic as-
sumption that long-distance A-bar movement from the complement clauses of verbs
is well-formed. This assumption is largely inherited in the subjacency, CED, Bar-
riers, and phase frameworks that have dominated generative discussions on long-
distance A-bar movement (Chomsky, 1973, 1977, 1986, 2000, 2001; Huang, 1982)
(although Chomsky, 1977 mentions briey the eect of bridge verbs, as Charles Yang
(quoting Julie Legate) points out to me). In the subjacency and phase frameworks,
VPs do not block A-bar movement, while the CED and Barriers frameworks do not
ban A-bar movement from complement clauses. Consequently, they all predict that
the sentences in (22) should be uniformly acceptable, contrary to fact.
(22) a. *Who does John {shout/quip} that Mary likes?
b. *Co1 Janek my?li [?e studenci czytaj? t1]?
what Jan thought that students read
`What did Jan think that students read?' (Polish, Witko?, 1995, p. 229
ex. 39, his judgment)
c. Who does John think that Mary likes?
The results from the acceptabilty judgment experiment and post hoc anal-
ysis in Chapter 4 suggest that long-distance A-bar movement is sensitive to rel-
atively ne-grained semantic distinctions, rather than broad, coarse-grained cate-
gories. This sensitivity poses an interesting challenge from the perspective of formal
287
description: how should these ne-grained eects be modeled?
To the extent that the eects are ultimately pragmatic or discourse in na-
ture (Ambridge and Goldberg, 2008; Erteschik-Shir and Lappin, 1979; M?ller, 1995;
Oshima, 2006), positing some kind of syntactic constraint would be irrelevant. In
Chapter 4, I observed that from a syntactic theory perspective, this analysis means
that we can maintain a relatively simple syntactic account of long-distance A-bar
movement: A-bar movement from the complement clauses of verbs is always syn-
tactically well-formed.
The existence of cross-linguistic variation presents complications for this ap-
proach. In Chapter 4, I discussed the case of (cognitive) factives in English and
Dutch: English appears to allow long-distance A-bar movement from the comple-
ment clauses of these verbs more freely than Dutch. To maintain the position that
bridge eects reect pragmatic or discourse constraints, one would have to claim
that the pragmatic or discourse properties of verbs with similar lexical semantics
dier cross-linguistically, which raises questions about what these dierences might
be and how they might be learned. In contrast, cross-linguistic variation can be
more easily accommodated in an analysis where there are language-specic syntac-
tic constraints on A-bar movement that make reference to specic verbs or verb
classes.
In fact, there are various technical options that can be easily used to implement
this latter idea. One option is to implement the constraint via subcategorization
and successive cyclic movement, assuming that long-distance A-bar movement must
take place successive cyclically. Suppose that complement clauses are headed by
288
C, and C in principle comes in two dierent types: one with a wh-feature that
facilitates successive cyclic movement, and one without. What learners need to do
is to determine (from linguistic experience) which verbs can take CP complements
headed by a C with a wh-feature; these verbs allow long-distance A-bar movement.
Another solution would be to invoke a classical generate-and-lter approach.
We might state a bridge verb constraint as follows:
(23) a. Constraint: *wh1 . . . V [CP . . . t1 . . . ,
b. such that V 6? BRV, where BRV is the set of bridge verbs
We might imagine this constraint applying throughout a derivation, whenever
A-bar movement takes place, or, assuming a Y-model, this constraint might apply
at the end of a syntactic derivation, as a constraint on syntactic representations.
To the extent that there is evidence from ellipsis showing that A-bar movement
constraints can be violated as long as the resulting wh-dependency is not overtly
realized (see Merchant, 2001, Lasnik, 2001, and others on rescue of island violations
by ellipsis), the latter implementation is empirically more desirable.
One additional implication of this lter account is that it requires one to
assume that learners begin with knowledge of an abstract category of bridge verbs,
even though they do not necessarily know which verbs belong to that category.
5.5 Evaluation of existing proposals about bridge eects
In the following sections, I review and evaluate formal analyses of bridge eects
presented in the literature. Because of the breadth of the literature, I cannot present
289
every single proposal in detail. Instead, I seek to provide a high-level overview of
the key ideas, discussing specic proposals only when necessary. For the most part,
I will restrict my discussion to English bridge eects, which have received the most
attention. Readers should keep in mind that cross-linguistic variation means that
a theory that works well for English does not necessarily extend naturally to other
languages.
5.5.1 Syntactic approaches
Most syntactic proposals focus on accounting for acceptability contrasts be-
tween selected classes of clause-embedding verbs, typically within a single language.
While these are often elegant, they do not necessarily extend easily to other contrasts
within the same language. As a result, they do not always come close to capturing
the range of variation within a language (i.e. between various verb classes) and
across languages.
To account for dierences between bridge and non-bridge verbs within a lan-
guage, an intuition frequently pursued is that bridge verbs take clausal complements,
while non-bridge verbs do not, contrary to appearances.
For example, Snyder (1992) and de Cuba (2018) claim that the embedded
clauses that appear with manner of speaking verbs are not complements, but are
appositives or adjuncts. This appositive / adjunct analysis explains why the clauses
are optional, i.e. why manner-of-speaking verbs can appear in intransitive frames.
The restriction on A-bar movement from embedded clauses associated with these
290
verbs follows from this analysis as a violation of an adjunct island condition.
A close variant of this structural approach suggests that factive verbs actually
contain nominal structure, so A-bar movement from the clauses eectively violate
the Complex NP Island Constraint. This idea can be partially traced to Kiparsky
and Kiparsky, 1970, who claim that factive verbs take a complex NP complement;
however, they are careful to point out that A-bar movement from factive comple-
ments is acceptable, suggesting that A-bar movement takes after an operation that
deletes nominal structure (p. 162, fn. 9). More recently, Kastner (2015) argues
that these verbs subcategorize for DPs, where the D head takes a clausal comple-
ment. The same idea has been extended to manner-of speaking verbs by Berrebi
and Bassel (2017). Their proposal also bears a strong resemblance to the proposals
of Snyder (1992) and de Cuba (2018). In addition to arguing that embedded clauses
with manner of-speaking verbs are appositive, they also hypothesize that manner-
of-speaking VPs contain nominal structure, which reects the fact that these verbs
have zero-derived nominal forms (also Zwicky, 1971, Levin, 1993).
(24) a. regret [NP [CP [Kiparsky and Kiparsky, 1970]
b. regret [DP [CP [Kastner, 2015]
c. [V P make [NP a grunt ] CP ] ? grunt CP [adapted from Snyder,
1992, p. 3 ex. 8]
d. [V P whisper [NP (a whisper) ] CP ] ? whisper CP [adapted from
de Cuba, 2018, p. 8 ex. 30]
The above proposals attribute restrictions on long-distance A-bar movement
to the presence of additional structure intervening between a clause and a verb. In
a series of papers, de Cuba (2007, et seq.) reverses this intuition, suggesting that
non-factive verbs like think take clausal complements that have an extra functional
291
projection, cP, that causes the clause to receive a non-factive interpretation. In de
Cuba's analysis, the extra functional projection, combined with the independently
motivated Adjunction Prohibition (25a) (McCloskey, 2006), exceptionally makes
available an escape hatch in an embedded clause, so that wh-phrases may move out
of it.
(25) a. Adjunction Prohibition (McCloskey, 2006, p. 93, ex. 30): Adjunc-
tion to a phrase which is s-selected by a lexical (open class) head is
ungrammatical (cf. Chomsky, 1986, which assumes that adjunction is
only possible only to a maximal projection . . . that is a nonargument
(p. 6 ex. 6)).
b. think [cP [CP [A-bar movement to adjoin to CP possible]
c. regret [CP [A-bar movement to adjoin to CP impossible]
Finally, a third approach claims that verbs take clausal complements of the
same size: there is no additional functional projection that blocks or facilitates
long-distance A-bar movement. To explain dierences between the complements of
dierent verb classes, one posits dierences within these CPs. For example, Hegarty
(1992) argues that complement CPs of factive and non-factive verbs have dierent
kinds of C heads: factive Cs [discharge] the event position of the clause at IPwhile
non-factive (propositional') ones do not (pp. 3233). Haegeman and ?r?gdi (2010)
suggests that factive CPs are referential due to the presence of a null operator
on their left periphery. These hypothesized structural dierences are then linked
directly to the availability of long-distance A-bar movement.
292
5.5.2 Non-syntactic approaches
Non-syntactic approaches, as the name suggests, strip out the syntactic mid-
dleman, by restricting A-bar movement to contexts that bear a particular semantic
or discourse property. This property has been labeled in various ways: dominant
in Erteschik-Shir, 1973, Erteschik-Shir and Lappin, 1979, and Erteschik-Shir, 2006,
non-backgrounded in Goldberg, 2006 and Ambridge and Goldberg, 2008. In this
tradition, bridge verbs are compatible with long-distance A-bar movement because
their clausal complements have this property, while this is not the case for other
verbs.
Looking at long-distance A-bar movement of wh-adjuncts (a phenomenon that
is not discussed in this dissertation), Cattell (1978) suggests that long-distance A-
bar movement of wh-adjuncts is possible only from the complements of volunteer
stance verbs, which introduce to the common ground the proposition denoted by
the complement clause (Cattell does not specify whether these verbs vary in terms
of A-bar movement of wh-arguments, nor discuss manner-of-speaking verbs.) Cat-
tell's typology has been inuential, although most adopters (Hegarty, de Cuba and
colleagues, Kastner, etc.) have chosen to implement his intuition in syntactic terms.
5.5.3 Evaluation of these approaches
Syntactic approaches tend to capture within-language variation with a widely-
accepted and independently-motivated mechanism: subcategorization / c-selection.
The weakness, however, is that independent morphosyntactic evidence in favor of
293
the proposed subcategorization dierences is at best mixed. For example, consider
de Cuba's proposal that non-factive verbs take a cP complement. The arguments for
cP are entirely semantic in nature, as far as I can tell. Likewise, while it is the case
that a factive verb like regret and remember can take nominal complements, it does
not follow that when it appears with an embedded clause, the clause must be found
inside a nominal complement, pace Kiparsky and Kiparsky, 1970 and Kastner, 2015.
There is also empirical evidence against the hypothesis that embedded clauses
appearing with non-bridge verbs are embedded in some silent nominal projection. I
present three arguments from English.
The rst argument is from conjunction. Two conjoined nominals can be op-
tionally modied with the adverb both, but it seems worse to conjoin an embedded
clause and a(n overt) nominal in the context of a factive verb (26a) and a manner-
of-speaking verb (26b).
(26) a. Factive verb
(i) The organizers regret both [the fact that some panellists were
unprepared] and [the fact that the panel started late].
(ii) ??The organizers regret both [that some panellists were unprepared]
and [the fact that the panel started late].
b. Manner-of-speaking verb
(i) The child whispered (to Mary) both [the news that it was snowing]
and [the rumor that schools would be closed the next day].
(ii) ??The child whispered (to Mary) both [that it was snowing] and
[the rumor that schools would be closed the next day].
The second argument is based on whether these clauses can appear in sub-
ject and object positions (A-positions). Nominal complements of the vast majority
of verbs can appear in the subject position when the verb is in the passive voice
(27a). To the extent that embedded clauses in the context of manner-of-speaking
294
and factive verbs are contained in silent nominal projections, they should appear
felicitously in subject positions. This prediction is borne out for factive verbs (27b)
but not for manner-of-speaking verbs (27c) (see also Baltin, 1982, p. 11).
(27) a. [The fact that there is so much socioeconomic inequality] is regretted
by many.
b. [That there is so much socioeconomic inequality] is regretted by many.
c. *[That John is an idiot] was shouted by Mary.
However, the fact that clauses can appear in subject positions for passivized
factive verbs is not a watertight argument that these clauses are contained within
covert DPs or NPs. There is no a priori requirement for subjects to be nominal.
Another distributional test is to embed a passivized factive construction as the
complement of an Exceptional Case Marking (ECM) verb like expect or believe (28a).
If the passivized clausal subject is contained within a nominal expression, then this
nominal expression should be able to appear in a position that gets accusative case,
along the lines of (28a). This prediction is not borne out (28b).
(28) a. I expect [the fact that there is so much socioeconomic inequality] to be
regretted by many.
b. *I expect [that there is so much socioeconomic inequality] to be regretted
by many.
The third argument is specic to a recent proposal by Kastner (2015). Kastner
assumes explicitly that verbs like regret or deny, which have a presuppositional
lexical semantics, subcategorize for a denite DP . . .  (p. 161).
(29) [V P regret [DP ? [CP . . . ]]] (regret takes a DP complement headed by ?,
which is a covert denite determiner) (Kastner, 2015, p. 161, exx. 17 and
18)
This assumption seems too strong, since these verbs can take indenite com-
295
plex NPs as complements (30).
(30) Mary {regretted / denied / remembered} a statement that her employees
made. [Indenite nominal complement]
Setting this subcategorization fact aside, if the embedded clauses appearing
with these verbs are part of a denite DP, then they should topicalize easily in the
same way that denite DPs topicalize. This fact is not borne out, as observed by
Baltin (1982).
(31) a. John remembered {that Mary liked apples / the story}.
b. {*That Mary liked apples/The story}, John remembered.
c. John whined that he was hungry. (Baltin, 1982, p. 11, ex. 44a)
d. *That he was hungry, John whined. (ibid. ex. 44b)
Non-syntactic denitions of bridge verbs sidestep the above problems by obvi-
ating the need for morphosyntactic evidence. Broadly speaking, there are two issues
with these proposals.
First, even though the intuitions behind these proposals are quite plausible,
it can be challenging to identify independent evidence for these proposals. For
example, Erteschik-Shir, Goldberg, and colleagues rely heavily on negation tests.
Erteschik-Shir (1973) uses a lie test: bridge verbs introduce propositions whose
truth can be challenged with a That's a lie! or, less accusatorily, That's not
true. This test works well for many high-frequency non-factive verbs (32), factive
verbs (33), and some manner-of-speaking verbs (34a), but seems to predicts that a
manner-of-speaking verb like whisper should allow long-distance A-bar movement
when the context is right (35a).
(32) A: John {said/thinks} that the earth is at.
296
B: That's a lie. We have photos of the earth from the International Space
Station.
(33) A: John regrets that he missed the party yesterday.
B: That's not true. #He was there at the party. (OK: . . .He is not sorry
that he missed it.)
(34) a. A: John mumbled that the earth rotated around the moon.
B: That's not true. #We have photos of the earth and the moon
from space.
(35) a. A: John whispered (to me) that Mary killed Sam.
B: That's a lie. It's John who killed Sam! (cf. ??It was Sam who
John whispered that Mary killed.)
Cattell's typology of stance verbs is more ne-grained, but suers from similar
weaknesses. For one, he does not oer clear tests for determining what kind of
stance a verb conveys, nor how it relates to other kinds of A-bar movement. Nor
does his discussion extend immediately or easily to manner-of-speaking verbs, as
Erteschik-Shir (2006) points out.
A second issue is in terms of data coverage. One attractive property of these
proposals is that in principle, they can describe and capture relatively many classes
of verbs with a single abstract property. However, it is unclear to what extent there
is a property that can characterize the range of acceptability of long-distance A-bar
movement; existing proposals are often built on a small number of case studies, such
as long-distance A-bar movement from the complement clauses of manner of speak-
ing verbs and factive verbs. The formal acceptability judgment study in Chapter
4 (Section 4.4), however, suggests that there are more verb classes for which long-
distance A-bar movement is not fully acceptable. As far as I can tell, there is no
297
existing discourse-based or pragmatics-based account that claims to cover all these
dierent verb classes, nor, as far as I can tell, is there an account that can be easily
extended to do so.
5.5.4 Modelling variation within and across languages
Even supposing that a parsimonious syntactic or semantic characterization of
bridge verbs or bridge eects in English is available, this characterization will still
be inadequate as a general theory of bridge verbs and bridge eects, because not
all languages behave like English in allowing long-distance A-bar movement across
verbs like say and think, and realize.
Cross-linguistic variation thus poses a serious challenge for accounts that dene
bridge eects in purely semantic and/or discourse terms, a point that Ambridge and
Goldberg acknowledge, citing again the case of Polish as an example (p. 376). As
mentioned previously, the only way these accounts can accommodate cross-linguistic
dierences is by positing that the semantics or discourse properties of these verbs
vary across languages. Consequently, one implicit assumption that these accounts
make is that learners can deduce very subtle facts about the semantics and discourse
properties of these verbs. This is by no means a trivial assumption, as many of these
verbs describe highly abstract mental states, the specics of which are dicult to
deduce from physical context alone (Hacquard and Lidz, 2019; Harrigan et al., 2016;
Huang et al., 2018).
Syntactic accounts in principle could fare better, especially if they explain
298
bridge eects as a product of lexically idiosyncratic subcategorization dierences.
More specically, we might imagine that there are dierent types of complement
clauses, of which only some allow long-distance A-bar movement, e.g. because they
contain the right features and/or an escape hatch in the left periphery. Only some
verbs subcategorize for these complement clauses that allow long-distance A-bar
movement; others take complements that do not.
The idea being exploited here is that subcategorization requirements vary
across languages, even when we control for lexical semantics (to the extent that
it is possible to do so). For example, French croire `believe' can take a non-nite
complement without an overt subject (i.e. it has control syntax) while English
`believe' does not (e.g. Bo?kovi?, 1996).
(36) Pierre croit [PRO avoir convaincu son auditoire].
Pierre believes to.have convinced his audience
`Pierre believes that he has convinced his audience.' (cf. *Pierre believes to
have convinced his audience.; Bo?kovi?, 1996, p. 286, ex. 20a)
Taking this subcategorization idea seriously requires providing a theory of
clausal syntax that provides a typology of complement clauses that attitude verbs
can take in principle and a theory of acquisition that explains the process by which
learners determine which verbs take what kind of complements; one such account is
presented above in Section 5.4.
Another not-unpopular subcategorization-based approach derives bridge ef-
fects by assuming that non-bridge verbs only appear with adjunct clauses, not com-
plement clauses, thus reducing bridge eects to a classic adjunct/complement dis-
tinction. This approach assumes that adjuncts are islands, a generalization that I
299
have claimed is problematic, as much as it is elegant. Setting that concern aside,
using this approach to model cross-linguistic variation presents a problem in that it
requires assuming that subcategorization properties vary in much deeper ways cross-
linguistically. For example, long-distance A-bar movement across cognitive factive
verbs is possible in English but not for certain Dutch varieties. In this particular
subcategorization approach, one would need to postulate that in these Dutch vari-
eties, cognitive factive verbs are obligatorily intransitive, so any subordinate clauses
that appear with them are adjuncts, not complements. This strikes me as an odd
position to take: while subcategorization requirements do vary for a given lexical
item (and their counterparts) across languages (cf. French croir and English be-
lieve), it seems unlikely that these requirements would vary at the level of a verb
class.
Alternatively, one could maintain that complement clauses have the same in-
ternal syntax cross-linguistically, and derive the facts by stipulating that comple-
ment clauses undergo certain structural operations in some languages that render
them opaque for long-distance A-bar movement. An example is found in Witko?,
1995, section 7, which deals with the problem of why Polish does not seem to allow
long-distance A-bar movement. Witko? rightly observes that this fact is not in line
with a Barriers framework (p. 252), which allows long-distance A-bar movement
from complement clauses. (In fact, the core assumptions in Barriers allow long-
distance A-bar movement from complement clauses of nouns, obliging Chomsky to
stipulate that nouns assign oblique Case and that this imposes an inherent barrier
to government (Chomsky, 1986 p. 36).) Witko?, following Cinque (1990), stip-
300
ulates that clause-embedding verbs in Polish take complement clauses, but these
clauses extrapose, presumably obligatorily, so these clauses are not L-marked. An
adjunct undergoing A-bar movement will fail to adjoin to these clauses and cannot
move out of them. As for arguments, Witko? attributes their failure to move to the
presence of tense in the embedded clause, following proposals by Manzini (1992).3
As far as I can tell, there are two challenges for this extraposition and tense
analysis. First, there is no independent evidence that clauses extrapose, not least
because such an extraposition operation is string vacuous: Polish is head-initial, so
a complement and an extraposed complement both appear on the right of a head.
Second, it is dicult to see how that kind of distinction might be learned: to derive
the dierences between Polish and English, one must assume that the Polish learner
gures out that complement clauses always extrapose, while the English learner
gures out that extraposition is not always obligatory. Since complement clauses
supercially appear in the same post-verbal position in both languages, it is unclear
how learners might draw these conclusions.
3As far as I can tell, Witko? could have also derived the desired blocking eects by simply
appealing to the Freezing Condition (Wexler and Culicover, 1980), which blocks A-bar movement
from extraposed elements. But this option is technically not available under the Barriers frame-
work that he is operating in.
301
Chapter 6: Future directions and conclusion
6.1 Future work
In previous chapters, I surveyed existing work on A-bar movement from rela-
tive clauses and complement clauses, clarifying generalizations as well as contribut-
ing new descriptions to the literature. I also observed that cross-linguistic variation
in these phenomena raise learning problems, and oered proposals on how learn-
ers might overcome them with evidence in their linguistic experience and learning
mechanisms and biases. However, there are still many open questions that are not
fully understood, and they await further investigation. I list a few of them here.
First, the account presented in Chapters 2 and 3 is intended to cover A-
bar movement from clausal adjuncts, such as porous relative clauses, purposive
clauses, and temporal adjuncts, among others. However, it does not deal with other
subclasses of adjuncts that systematically show within-language and cross-linguistic
variation, including the present participial adjuncts (i.e. gerundival adjuncts in
English) (1) studied by Truswell (2007, 2008, 2011).
(1) A-bar movement from present participial adjuncts
a. What1 did Mary drive John crazy [trying to x t1]?
b. What kind of hat1 did John come back [wearing t1]?
302
Truswell shows that this phenomenon has a very limited cross-linguistic dis-
tribution, suggesting that this phenomenon is learned. As far as I can tell, these
structures seem to be very rare in child-ambient speech and adult-directed texts and
speech in English. Further, I am not aware of another A-bar movement phenomenon
that covaries with A-bar movement from these participial adjuncts. Supposing that
this phenomenon is grammatical (pace Kohrt et al. (submitted)), it is not clear to
me what the relevant learning account might be.
Second, even though I have shown that A-bar movement from certain adjuncts
is possible in some languages, I observed in Chapter 2 and 4 that not all phrases
may freely move from them. For example, porous relative clauses do not permit
overt movement of adjuncts (2), nor do complement clauses of many verbs (3).
(2) a. *This is the studio where1 Paul and Stevie were the only ones2 [who2 t2
wanted to record the song t1].
b. *[How thoroughly]1 do we need to nd someone2 [who2 t2 understands
this paper t1]?
c. *[Under what circumstances]1 are there many Americans2 [who2 t2 con-
done violence t1]?
(3) a. *That is how1 Mary realized [John cleaned the kitchen t1]: with her
toothbrush.
b. *That is the reason why1 John noticed [that Mary left the window open
t1]: to let in the breeze.
One reasonable interpretation of these facts is that the conditions on overt A-
bar movement of adjuncts are more restrictive than those on overt A-bar movement
of arguments, as Cattell (1978), Hegarty (1992), and Abrus?n (2014) (among others)
suggest. However, this is only an intuition: I am not aware of a general theory
that addresses all conditions on A-bar movement of adjuncts. The state-of-the-art
303
appears to be Abrus?n, 2014, but this proposal deals with only A-bar movement
from a subset of complement clauses; it does not address in detail, for instance, the
observations of Cattell (1978) (see discussion in pp. 7980 in Abrus?n, 2014).
Finally, there is much more descriptive work that needs to be done to under-
stand the nature of bridge verbs and long-distance A-bar movement. The results
of the acceptability judgment experiment and post hoc classication presented in
Chapter 4 shows that many existing generalizations about bridge verbs are not ne-
grained enough to capture the range of judgments. There are several next steps for
this topic. For example, the post hoc classication is specic enough to make fairly
clear predictions about which other verbs might be incompatible with long-distance
A-bar movement; these predictions can be tested with a follow-up acceptability
judgment study.
A closely-related issue is the extent to which these ne-grained distinctions can
be more succinctly captured under a general pragmatic or discourse principle. I sug-
gested that this calls for further theorizing, as existing pragmatics or discourse-based
accounts tend to cover explicitly only a small set of verb classes, such as manner of
speaking verbs. Further comparative research is also necessary to understand how
exactly the pragmatics of these verbs dier cross-linguistically, and to what extent
this predicts cross-linguistic variation in the acceptability of long-distance A-bar
movement for these verbs.
304
6.2 Conclusion
One of the key goals in modern linguistics is to understand the nature of
the language faculty. Constraints on A-bar movement have played an important
role in this research program. They are negative in nature, highly abstract, and
dicult to observe directly. Nevertheless, speakers come to possess knowledge of
these constraints.
A standard solution to this puzzle is to claim that the knowledge of these
constraints is innately specied, and so minimal learning is required on the part of
the learner. In this dissertation, I discussed two case studies involving constraints
on A-bar movement that pose a problem to this particular proposal. I pointed
out that A-bar movement from relative clauses and complement clauses of verbs is
sensitive to ne-grained distinctions within a language, and these distinctions vary
to a certain extent across languages.
This kind of variation appears to pose major learning problems: given how
abstract these constraints are in the rst place, how might learners identify the
appropriate variant for their languages?
In the case of A-bar movement from porous relative clauses, I observed that
porous relative clauses can be distinguished from other relative clauses using certain
semantic properties. Of interest was a property I called tense dependence. I
presented two distinct syntactic hypotheses to capture this property, building on
work by researchers such as Sichel (2018), Ogihara (1996), and Grano and Lasnik
(2018), and showed how these treatments of tense dependence can be linked to A-bar
305
movement.
Further, not all languages have A-bar movement from porous relative clauses
(Cinque, 2010), which suggests that some aspect of this phenomenon is learned from
one's linguistic environment. However, an analysis of corpora in two dierent lan-
guages  English and Mandarin Chinese  shows that learners of these languages
are very unlikely to observe A-bar movement from porous relative clauses (cf. Pearl
and Sprouse, 2013a,b). Building on the observation of tense dependence, I suggested
that learners of these languages resolve the learning problem by using evidence from
a supercially-unrelated source: A-bar movement from non-nite purposive clauses.
I argued that despite appearances, purposive clauses, like porous relative clauses,
also have dependent tense. I presented evidence from cross-linguistic co-variation
and corpus analyses to show that learners of English and Mandarin Chinese corpora
can observe A-bar movement from purposive clauses.
The fact that English and Mandarin learners can observe A-bar movement
from purposive clauses, however, does not in principle guarantee that they will
conclude that their languages allow A-bar movement from porous relative clauses. I
argued that the fact that these learners draw such a conclusion implies that there are
restrictions on a learner's hypothesis space. As a result, for learners, the observation
of A-bar movement from purposive clauses provides clear evidence for a very small
number of hypotheses (possibly only one).
I also reviewed variation in the acceptability of A-bar movement from the com-
plement clauses of verbs. I noted that existing generalizations about these eects are
often made on the basis of a small number of verbs. I presented a 100-verb accept-
306
ability judgment experiment that allowed for a more systematic assessment of these
claims for English. The experiment results suggest that existing proposals about
bridge verbs are too coarse-grained: English speakers appear to be sensitive to more
ne-grained distinctions than those usually described in the literature. However, the
experiment results also suggest that in English, the factivity of a verb has a limited
impact on long-distance A-bar movement, which provides an interesting contrast
with languages like Dutch, where for some speakers, long-distance A-bar movement
across (at least) cognitive factive verbs is more marked. While English learners have
access to evidence that long-distance A-bar movement is compatible with cognitive
factive verbs, the evidence is not robust for many such verbs. I argue that it is
likely that English learners generalize across the class of cognitive factive verbs. To
that end, I looked at what kind of learning mechanism might let them form these
generalizations despite spare evidence, comparing the Suciency Principle (Yang,
2016, 2017) and an alternative Bayesian-inspired approach.
Finally, the fact that these constraints can be shaped by linguistic experience
also has implications for our theories of syntax. I argued, for instance, that existing
accounts of these phenomena, as much as they are elegant, tend to lack sucient
exibility to capture the range of variation in these constraints. The variation we
observe suggests that A-bar movement from porous relative clauses and complement
clauses do not have a single source. Following Truswell (2007, 2011), I further noted
that the case of porous relative clauses and purposive clauses also presents challenges
for proposals that derive adjunct islandhood from more basic syntactic principles.
Overall, the picture that emerges from these case studies is that despite ap-
307
pearances, learners might have access to evidence in their linguistic experience that
allows them to learn the appropriate points of variation. However, the presence of
evidence alone is insucient, because the evidence might be sparse and/or look su-
percially very dierent from the phenomenon to be learned. It appears that some
kind of learning bias is needed, so that learners can draw the right conclusions
on the basis of limited evidence. More generally, the study of dicult-to-observe
phenomena through the lens of variation and actual linguistic experience can help
shed light on the learning biases and linguistic dimensions that matter for language
acquisition.
308
Bibliography
Abels, Klaus. 2003. Successive cyclicity, anti-locality, and adposition stranding.
Doctoral Dissertation, University of Connecticut.
Abrus?n, M?rta. 2011. Presuppositional and negative islands: A semantic account.
Natural Language Semantics 19:257321.
Abrus?n, M?rta. 2014. Weak island semantics . Oxford: Oxford University Press.
Abusch, Dorit. 1988. Sequence of tense, intensionality and scope. In Proceedings of
WCCFL 7 , ed. Hagit Borer, 114.
Akmajian, Adrian. 1970. Aspects of the grammar of focus in English. Doctoral
Dissertation, Massachusetts Institute of Technology.
Alexopoulou, Theodora, and Frank Keller. 2007. Locality, cyclicity, and resump-
tion: At the interface between the grammar and the human sentence processor.
Language 110160.
Allwood, Jens. 1976. The complex NP constraint as a non-universal rule and some
semantic factors inuencing the acceptability of Swedish sentences which violate
the CNPC. University of Massachusetts Occasional Papers in Linguistics 2:120.
Allwood, Jens. 1982. The complex NP constraint in Swedish. In Readings on
unbounded dependencies in Scandinavian languages , ed. Elisabet Engdahl and
Eva Ejerhed, 1532. Stockholm: Almqvist and Wiksell.
Almeida, Diogo. 2014. Subliminal wh-islands in Brazilian Portuguese and the con-
sequences for syntactic theory. Revista da ABRALIN 13:5593.
Altshuler, Daniel Gordon, Valentine Hacquard, Thomas Roberts, and Aaron Steven
White. 2015. On double access, cessation and parentheticality. In Proceedings of
Semantics and Linguistic Theory 25 , ed. Sarah D'Antonio, Mary Moroney, and
Carol Rose Little, 1837.
Ambridge, Ben, and Adele E. Goldberg. 2008. The island status of clausal com-
plements: Evidence in favor of an information structure explanation. Cognitive
Linguistics 19:349381.
309
Ambridge, Ben, Julian M Pine, Caroline F Rowland, Daniel Freudenthal, and
Franklin Chang. 2014. Avoiding dative overgeneralisation errors: semantics, statis-
tics or both? Language, Cognition and Neuroscience 29:218243.
Anand, Pranav, Jane Grimshaw, and Valentine Hacquard. to appear. Sentence
embedding predicates, factivity and subjects. In A festschrift for Lauri Karttunen,
ed. Cleo Condoravdi.
Anand, Pranav, and Valentine Hacquard. 2009. Epistemics with attitude. In Pro-
ceedings of Semantics and Linguistic Theory 18 , ed. Tova Friedman and Satoshi
Ito, 3754.
Anand, Pranav, and Valentine Hacquard. 2014. Factivity, belief and discourse. In
The art and craft of semantics: A festschrift for Irene Heim, ed. Lukas Crnic and
Uli Sauerland, 6990. Cambridge: MIT Working Papers in Linguistics.
Anand, Pranav, and Andrew Nevins. 2004. Shifty operators in changing contexts. In
Proceedings of Semantics and Linguistic Theory 14 , ed. Robert B. Young, 2037.
Antonenko, Andrei. 2010. Puzzles of Russian subjunctives. University of Pennsyl-
vania Working Papers in Linguistics 16.
Anyadi, Stefanie, and Armine Tamrazian. 1993. Wh-movement in Armenian and
Ruhr German. UCL Working Papers in Linguistics 5:122.
Baker, C. L. 1979. Syntactic theory and the projection problem. Linguistic Inquiry
10:533581.
Baltin, Mark R. 1982. A landing site theory of movement rules. Linguistic Inquiry
13:138.
Barros, Matthew, and Robert Frank. 2017. Shifty subjects and clause-mate restric-
tions. Presented at the 91st meeting of the Linguistic Society of America (LSA),
Austin, Texas.
Barton, Kamil. 2019. MuMIn: Multi-model inference. URL https://CRAN.
R-project.org/package=MuMIn, R package version 1.43.6.
Basse, Galen. 2008. Factive complements as defective phases. In Proceedings of the
27th West Coast Conference on Formal Linguistics , ed. Natasha Abner and Jason
Bishop, 5462. Cascadilla Proceedings Project Massachussetts.
Bates, Douglas, Martin M?chler, Ben Bolker, and Steve Walker. 2015. Fitting linear
mixed-eects models using lme4. Journal of Statistical Software 67:148.
Beaver, David, and Cleo Condoravdi. 2003. A uniform analysis of Before and After.
In Proceedings of Semantics and Linguistic Theory 13 , ed. Robert B. Young and
Yuping Zhou, 3754.
310
Beavers, John, and Andrew Koontz-Garboden. 2012. Manner and result in the roots
of verbal meaning. Linguistic Inquiry 43:331369.
Berrebi, Si, and Noa Bassel. 2017. Selectional split across speech verbs: non bridge
islands revisited. Talk given at SelectionFest, Nov 2017. Retrieved November 26,
2017 from website.
Berwick, Robert C. 1985. The acquisition of syntactic knowledge. Cambridge, MA:
MIT Press.
Bhatt, Rajesh. 2002. The raising analysis of relative clauses: Evidence from adjec-
tival modication. Natural Language Semantics 10:4390.
Bhatt, Rakesh Mohan. 1999. Verb movement and the syntax of Kashmiri . Dordrecht:
Kluwer.
Bianchi, Valentina. 2000. The raising analysis of relative clauses: A reply to Borsley.
Linguistic Inquiry 31:123140.
Bird, Steven, Ewan Klein, and Edward Loper. 2009. Natural language processing
with Python: analyzing text with the Natural Language Toolkit . O'Reilly Media,
Inc.
Bleam, Tonia M. 1999. Le?sta spanish and the syntax of clitic doubling. Doctoral
Dissertation, University of Delaware.
BNC Consortium. 2007. The British National Corpus, version 3 (BNC XML Edi-
tion).
Boeckx, Cedric. 2012. Syntactic islands . Cambridge: Cambridge University Press.
Bo?kovi?, ?eljko. 1996. Selection and the categorial status of innitival complements.
Natural Language & Linguistic Theory 14:269304.
Bowerman, Melissa. 1988. The `no negative evidence' problem: How do children
avoid constructing an overly general grammar? In Explaining language universals ,
ed. John Hawkins, 73101. Oxford: Basil Blackwell.
Boyd, Jeremy K, and Adele E Goldberg. 2011. Learning what not to say: The role
of statistical preemption and categorization in a-adjective production. Language
87:5583.
Brame, Michael. 1968. A new analysis of the relative clause: Evidence for an inter-
pretive theory. Unpublished manuscript, Massachusetts Institute of Technology.
Bresnan, Joan. 1977. Variables in the theory of transformations. In Formal syntax ,
ed. Peter Culicover, Thomas Wasow, and Adrian Akmajian, 157196. New York:
Academic Press.
311
Broekhuis, Hans, and Norbert Corver. 2016. Syntax of Dutch: Verbs and verb
phrases, volume 3 . Amsterdam: Amsterdam University Press.
Broekhuis, Hans, and Norbert Corver. 2017. 11.3.1.2. Wh-extraction from em-
bedded clauses (long wh-movement). URL http://www.taalportaal.org/
taalportaal/topic/link/syntax__Dutch__vp__V11_Word_order_Clause_
initial_position__V11_Word_order_Clause_initial_position.11.3.1.2.
xml, retrieved November 26, 2017 from website.
Broekhuis, Hans, and Norbert Corver. 2018. 5.1.6. Wh-extraction from argument
clauses. URL http://taalportaal.org/taalportaal/topic/link/syntax_
___Dutch__vp__V11_Word_order_Clause_initial_position__..__V5_42_
argument_clauses__V5_42_argument_clauses.5.1.6.xml, retrieved July 17,
2019 from website.
Brown, Roger. 1973. A rst language: The early stages . Cambridge, MA: Harvard
University Press.
Cattell, Ray. 1976. Constraints on movement rules. Language 52:1850.
Cattell, Ray. 1978. On the source of interrogative adverbs. Language 54:6177.
Chac?n, Dustin Alfonso. 2015. Comparative psychosyntax. Doctoral Dissertation,
University of Maryland, College Park, MD.
Chen, Sherry Yong. 2017. Processing tenses for the living and the dead: A psy-
cholinguistic investigation of lifetime eects in tensed and tenseless languages.
Master's thesis, University of Oxford.
Chomsky, Noam. 1957. Syntactic structures . The Hague: Mouton.
Chomsky, Noam. 1964. The logical basis of linguistic theory. In Proceedings of the
9th International Congress of Linguists , ed. H. G. Lunt, 9141008. The Hague:
Mouton.
Chomsky, Noam. 1965. Aspects of the theory of syntax . Cambridge, MA: MIT Press.
Chomsky, Noam. 1971. Problems of knowledge and freedom: The Russell lectures .
New York: Pantheon Books.
Chomsky, Noam. 1973. Conditions on transformations. In A festschrift for Mor-
ris Halle, ed. Stephen Anderson and Paul Kiparsky, 232286. New York: Holt,
Rinehart & Winston.
Chomsky, Noam. 1977. On wh-movement. In Formal syntax , ed. Peter W. Culicover,
Thomas Wasow, and Adrian Akmajian, 71132. New York: Academic Press.
Chomsky, Noam. 1981. Lectures on government and binding: The Pisa lectures .
Dordrecht: Foris.
312
Chomsky, Noam. 1986. Barriers . Cambridge, MA: MIT Press.
Chomsky, Noam. 1995. The Minimalist Program. Cambridge, MA: MIT Press.
Chomsky, Noam. 2000. Minimalist inquiries: The framework. In Step by step:
Essays on minimalist syntax in honor of Howard Lasnik , ed. Roger Martin, David
Michaels, and Juan Uriagereka, 89155. Cambridge, MA: MIT Press.
Chomsky, Noam. 2001. Derivation by phase. In Ken Hale: A life in language, ed.
Michael Kenstowicz, 152. Cambridge, MA: MIT Press.
Christensen, Ken Ramsh?j, and Anne Mette Nyvad. 2014. On the nature of es-
capable relative islands. Nordic Journal of Linguistics 37:2945.
Chung, Sandra, and William A Ladusaw. 2003. Restriction and saturation. Cam-
bridge, MA: MIT Press.
Chung, Sandra, and James McCloskey. 1983. On the interpretation of certain island
facts in GPSG. Linguistic Inquiry 14:704713.
Cichocki, Wladyslaw. 1983. Multiple wh-questions in Polish: A two-comp analysis.
Toronto Working Papers in Linguistics 4:5371.
Cinque, Guglielmo. 1990. Types of A	-dependencies . MIT Press.
Cinque, Guglielmo. 2010. On a selective violation of the Complex NP Constraint.
In Structure preserved: Studies in syntax for Jan Koster , ed. C. Jan-Wouter Zwart
and Mark de Vries, 8189. Amsterdam: John Benjamins.
de Cuba, Carlos. 2007. On (non) factivity, clausal complementation and the cp-eld.
Doctoral Dissertation, Stony Brook University.
de Cuba, Carlos. 2018. Manner-of-speaking that-complements as close apposition
structures. Proceedings of the Linguistic Society of America 3:321.
Culicover, Peter W. 1999. Syntactic nuts: Hard cases, syntactic theory, and language
acquisition. Oxford: Oxford University Press.
Culy, Christopher. 1994. Aspects of logophoric marking. Linguistics 32:10551094.
D?browska, Ewa. 2004. Language, mind and brain: Some psychological and neu-
rological constraints on theories of grammar . Edinburgh: Edinburgh University
Press.
D?browska, Ewa. 2008. Questions with long-distance dependencies: A usage-based
perspective. Cognitive Linguistics 19:391425.
D?browska, Ewa. 2013. Functional constraints, usage, and mental grammars: A
study of speakers' intuitions about questions with long-distance dependencies.
Cognitive Linguistics 24:633665.
313
Davies, William D. 2003. Extreme locality in Madurese wh-questions. Syntax 6:237
259.
Davies, William D, and Stanley Dubinsky. 2003. On extraction from NPs. Natural
Language & Linguistic Theory 21:137.
Deal, Amy Rose. 2017. Shifty asymmetries: universals and variation in shifty in-
dexicality. Manuscript.
Deane, Paul. 1991. Limits to attention: a cognitive theory of island phenomena.
Cognitive Linguistics 2:163.
Dell, Fran?ois. 1981. On the learnability of optional phonological rules. Linguistic
Inquiry 12:3137.
den Dikken, Marcel. 2007. Phase extension: contours of a theory of the role of head
movement in phrasal extraction. Theoretical Linguistics 33:141.
Dillon, Brian, Ewan Dunbar, and William Idsardi. 2013. A single-stage approach
to learning phonological categories: Insights from Inuktitut. Cognitive Science
37:344377.
Doherty, Cathal. 1993. Clauses without that : the case for bare sentential comple-
mentation in English. Doctoral Dissertation, University of California, Santa Cruz,
Santa Cruz, CA.
Doherty, Cathal. 2000. Clauses without that: The case for bare sentential comple-
mentation in English. New York: Garland.
Douglas, Jamie. 2017. Unifying the that-trace and anti-that-trace eects. Glossa: a
journal of general linguistics 2.
Dudley, Rachel, Meredith Rowe, Valentine Hacquard, and Jerey Lidz. 2017. Dis-
covering the factivity of know. In Proceedings of Semantics and Linguistic The-
ory 27 , ed. Dan Burgdorf, Jacob Collard, Sireemas Maspong, and Brynhildur
Stef?nsd?ttir, 600619.
Dunbar, Ewan. 2013. Statistical knowledge and learning in phonology. Doctoral
Dissertation, University of Maryland.
En?, M?rvet. 1987. Anchoring conditions for tense. Linguistic Inquiry 18:633658.
Engdahl, Elisabet. 1980. The syntax and semantics of questions in Swedish. Doctoral
Dissertation, University of Massachusetts, Amherst.
Engdahl, Elisabet. 1997. Relative clause extractions in context. Working papers in
Scandinavian syntax 60:5179.
Ernst, Thomas. 2001. The syntax of adjuncts . Cambridge: Cambridge University
Press.
314
Erteschik-Shir, Nomi. 1973. On the nature of island constraints. Doctoral Disserta-
tion, Massachusetts Institute of Technology.
Erteschik-Shir, Nomi. 2006. Bridge phenomena. In The Blackwell Companion to
Syntax , ed. Martin Everaert and Henk van Riemsdijk, 284294. Oxford: Blackwell
Publishing.
Erteschik-Shir, Nomi, and Shalom Lappin. 1979. Dominance and the functional
explanation of island phenomena. Theoretical Linguistics 6:4186.
Fanselow, Gisbert, Reinhold Kliegl, and Matthias Schlesewsky. 2005. Syntactic
variation in German wh-questions: Empirical investigations of weak crossover
violations and long wh-movement. In Linguistic variation yearbook , ed. Pierre
Pica, Johan Rooryck, and Jeroen van Craenenbroeck, volume 5, 3763. John
Benjamins.
Faraci, Robert Angelo. 1974. Aspects of the grammar of innitives and for -phrases.
Doctoral Dissertation, Massachusetts Institute of Technology.
Featherston, Samuel. 2004. Bridge verbs and V2 verbsthe same thing in spades?
Zeitschrift f?r Sprachwissenschaft 23:181209.
Feldman, Naomi H, Thomas L Griths, Sharon Goldwater, and James L Morgan.
2013. A role for the developing lexicon in phonetic category acquisition. Psycho-
logical Review 120:751778.
Fellbaum, Christiane, ed. 1998. Wordnet: An electronic lexical database. Cambridge,
MA: MIT Press.
Fiengo, Robert, and James Higginbotham. 1981. Opacity in NP. Linguistic Analysis
7:395421.
Fodor, Janet Dean. 1970. The linguistic description of opaque contexts. Doctoral
Dissertation, Massachusetts Institute of Technology.
Fodor, Janet Dean. 1992. Islands, learnability and the lexicon. In Island constraints:
Theory, acquisition and processing , ed. Helen Goodluck and Michael Rochemont,
109180. Dordrecht: Kluwer.
Geis, Michael Lorenz. 1970. Adverbial subordinate clauses in English. Doctoral
Dissertation, Massachusetts Institute of Technology.
Gennari, Silvia P. 2003. Tense meanings and temporal interpretation. Journal of
Semantics 20:3571.
Gibson, Edward. 2018. Meaning explanations of two syntactic islands: Subject is-
lands and embedded-clause islands. Presented at the University of Massachussetts,
Amherst. PDF available on author's website.
315
Gibson, Edward, and Kenneth Wexler. 1994. Triggers. Linguistic Inquiry 25:407
454.
Gilligan, Gary Martin. 1989. A cross-linguistic approach to the pro-drop parameter.
Doctoral Dissertation, University of Southern California.
Giusti, Giuliana. 1986. On the lack of wh-innitives with zu and the projection of
COMP in German. Groninger Arbeiten zur germanistischen Linguistik 28:115
169.
Goldberg, Adele E. 2006. Constructions at work: The nature of generalization in
language. Oxford: Oxford University Press.
Grano, Thomas. 2017. Finiteness contrasts without Tense? a view from Mandarin
Chinese. Journal of East Asian Linguistics 26:259299.
Grano, Thomas, and Howard Lasnik. 2018. How to neutralize a nite clause bound-
ary: Phase theory and the grammar of bound pronouns. Linguistic Inquiry
49:465499.
Grimshaw, Jane. 2015. Clausal complementation. Tutorial notes for Semantics and
Philosophy in Europe 8, Cambridge; posted on http://friederike-moltmann.
com/news/tutorial-clausal-complements-at-spe8-cambridge/.
Grosu, Alexander. 1994. Three studies in locality and case. London: Routledge.
Gu?ron, Jacqueline. 1980. On the syntax and semantics of PP extraposition. Lin-
guistic Inquiry 11:637678.
Hacquard, Valentine, and Jerey Lidz. 2019. Children's attitude problems: Boot-
strapping verb meaning from syntax and pragmatics. Mind & Language 34:7396.
Haegeman, Liliane, and Barbara ?r?gdi. 2010. Referential CPs and DPs: An oper-
ator movement account. Theoretical Linguistics 36:111152.
Haider, Hubert. 2010. The syntax of German. Cambridge: Cambridge University
Press.
Harrigan, Kaitlyn, Valentine Hacquard, and Jerey Lidz. 2016. Syntactic bootstrap-
ping in the acquisition of attitude verbs: think, want and hope. In Proceedings of
the 33rd West Coast Conference on Formal Linguistics , ed. Kyeong min Kim et
al. Cascadilla Proceedings Project.
Harris, Alice C. 1981. Georgian syntax: A study in relational grammar . Cambridge:
Cambridge University Press.
Hart, Betty, and Todd Risley. 2003. The early catastrophe. American Educator
27:69.
316
Hart, Betty, and Todd R Risley. 1995. Meaningful dierences in the everyday expe-
rience of young American children. Baltimore, MD: Paul H Brookes Publishing.
Hegarty, Michael Vincent. 1992. Adjunct extraction and chain congurations. Doc-
toral Dissertation, Massachusetts Institute of Technology.
Henderson, Brent. 2007. Matching and raising unied. Lingua 117:202220.
Higgins, Francis Roger. 1973. The pseudo-cleft construction in English. Doctoral
Dissertation, Massachusetts Institute of Technology.
Hoeksema, Jack. 2006. Hij zei van niet, maar knikte van ja: Distributie en di-
achronie van bijwoorden van polariteit ingeleid door van. Tabu 35:135158.
Ho, Erika. 2003. The specicity of environmental inuence: Socioeconomic status
aects early vocabulary development via maternal speech. Child Development
74:13681378.
Hofmeister, Philip. 2007. Representational complexity and memory retrieval in
language comprehension. Doctoral Dissertation, Stanford University.
Hofmeister, Philip, and Ivan A. Sag. 2010. Cognitive constraints and island eects.
Language 86:366415.
Honcoop, Martin. 1998. Dynamic excursions on weak islands. Doctoral Dissertation,
Leiden University.
Huang, C.-T. James. 1982. Logical relations in Chinese and the theory of grammar.
Doctoral Dissertation, Massachusetts Institute of Technology.
Huang, C.-T. James. 1989. Pro-drop in Chinese: A generalized control theory. In
The null subject parameter , ed. Osvaldo A. Jaeggli and Kenneth J. Sar, 185214.
Springer.
Huang, C-T James. 1993. Reconstruction and the structure of VP: Some theoretical
consequences. Linguistic Inquiry 24:103138.
Huang, Nick. 2015. On syntactic tense in Mandarin Chinese. In Proceedings of the
27th North American Conference on Chinese Linguistics , ed. Hongyin Tao et al.,
volume 2, 406423. Los Angeles: University of California, Los Angeles.
Huang, Nick. 2018a. The bound possessor eect: a new argument for the phasehood
of denite DPs. In Proceedings of NELS 48 , ed. Sherry Hucklebridge and Max
Nelson, volume 2, 3952.
Huang, Nick. 2018b. Control complements in Mandarin Chinese: implications for
restructuring and the Chinese niteness debate. Journal of East Asian Linguistics
27:347376.
317
Huang, Nick, Chia-Hsuan Liao, Valentine Hacquard, and Jerey Lidz. 2018. Learn-
ing attitude verb meanings in a morphosyntactically-poor language. In Proceed-
ings of the 42nd Proceedings of the 42nd annual Boston University Conference on
Language Development , ed. Anne B. Bertolini and Maxwell J. Kaplan, 359373.
Somerville Massachussetts: Cascadilla Proceedings Project.
Hulsey, Sarah, and Uli Sauerland. 2006. Sorting out relative clauses. Natural Lan-
guage Semantics 14:111137.
Hunter, Tim. 2015. Deconstructing merge and move to make room for adjunction.
Syntax 18:266319.
Huttenlocher, Janellen, Marina Vasilyeva, Elina Cymerman, and Susan Levine.
2002. Language input and child syntax. Cognitive Psychology 45:337374.
Huttenlocher, Janellen, Marina Vasilyeva, Heidi R Waterfall, Jack L Vevea, and
Larry V Hedges. 2007. The varieties of speech to young children. Developmental
Psychology 43:1062.
Hyams, Nina. 1986. Language acquisition and the theory of parameters . Dordrecht:
D. Reidel.
Jensen, Anne. 2002. S?tningsknuder i dansk [Knot sentences in Danish]. NyS 
Nydanske Studier & Almen kommunikationsteori 29:105124.
Jiang, Jay J, and David W Conrath. 1997. Semantic similarity based on corpus
statistics and lexical taxonomy. In Proceedings of the 10th Research on Compu-
tational Linguistics International Conference, 1933.
Jones, Charles. 1991. Purpose clauses: Syntax, thematics, and semantics of English
purpose constructions . Dordrecht: Kluwer.
Kandybowicz, Jason. 2006. Comp-trace eects explained away. In Proceedings of the
25th West Coast Conference on Formal Linguistics , ed. Donald Baumer, David
Montero, and Michael Scanlon, 220228. Somerville, MA: Cascadilla Proceedings
Project.
Karttunen, Lauri. 1971. Some observations on factivity. Papers in Linguistics 4:55
69.
Kastner, Itamar. 2015. Factivity mirrors interpretation: The selectional require-
ments of presuppositional verbs. Lingua 164:156188.
Kayne, Richard S. 1994. The antisymmetry of syntax . Cambridge, MA: MIT Press.
Keshet, Ezra. 2008. Only the strong: Restricting situation variables. In Proceedings
of Semantics and Linguistic Theory 18 , ed. Tova Friedman and Satoshi Ito, 483
495.
318
Khomitsevich, Olga. 2008. Dependencies across phases: From sequence of tense to
restrictions on movement. Doctoral Dissertation, University of Utrecht.
Kiparsky, Paul, and Carol Kiparsky. 1970. Fact. In Semantics: an interdisciplinary
reader in philosophy, linguistics and psychology , ed. Danny Steinberg and Leon
Jacobovits, 34569. Cambridge: Cambridge University Press.
Kiss, Katalin ?. 2002. The syntax of Hungarian. Cambridge: Cambridge University
Press.
Klecha, Peter. 2016. Modality and embedded temporal operators. Semantics and
Pragmatics 9:155.
Klein, Dan, and Christopher D. Manning. 2003. Accurate unlexicalized parsing. In
Proceedings of the 41st Meeting of the Association for Computational Linguistics ,
423430.
Kluender, Robert. 2004. Are subject islands subject to a processing account? In
Proceedings of the West Coast Conference on Formal Linguistics , ed. Vineeta
Chand, Ann Kelleher, Angelo J. Rodr?guez, and Benjamin Schmeiser, volume 23,
475499.
Kluender, Robert, and Marta Kutas. 1993. Subjacency as a processing phenomenon.
Language and Cognitive Processes 8:573633.
Kohrt, Annika, Trey Sorensen, and Dustin A. Chac?n. submitted. The real-time
status of semantic exceptions to the adjunct island constraint. In Proceedings of
the 54th Chicago Linguistics Society . Chicago: University of Chicago.
Krapova, Iliyana. 2010a. Bulgarian relative and factive clauses with an invariant
complementizer. Lingua 120:12401272.
Krapova, Iliyana. 2010b. Review of Referential CPs and DPs: An operator move-
ment account by Liliane Haegeman and Barbara ?r?gdi. Theoretical Linguistics
36:209223.
Kratzer, Angelika. 1998. More structural analogies between pronouns and tenses.
In Proceedings of Semantics and Linguistic Theory 8 , ed. Devon Strolovitch and
Aaron Lawson, 92110.
Kush, Dave, and Filippa Lindahl. 2011. On the escapability of islands in Scandina-
vian. Presented at the 85th meeting of the Linguistic Society of America (LSA),
Pittsburgh, Pennsylvania.
Kush, Dave, Terje Lohndal, and Jon Sprouse. 2018. Investigating variation in island
eects. Natural Language & Linguistic Theory 36:743779.
Kush, Dave, Akira Omaki, and Norbert Hornstein. 2013. Microvariation in islands?
In Experimental syntax and island eects , ed. Jon Sprouse and Norbert Hornstein,
chapter 11, 239264. Cambridge: Cambridge University Press.
319
Kusumoto, Kiyomi. 2005. On the quantication over times in natural language.
Natural Language Semantics 13:317357.
Kusumoto, Kiyomi. 2017. Geis-ambiguity and tense harmony. In A schrift to fest
Kyle Johnson, ed. Nicholas LaCara, Keir Moulton, and Anne-Michelle Tessier,
203205. Amherst, MA: University of Massachusetts Linguistics Open Access
Publications.
Kvam, Sigmund. 1983. Linksverschachtelung im Deutschen und Norwegischen.
T?bingen: Niemeyer.
Larson, Richard. 1990. Extraction and multiple selection in PP. The Linguistic
Review 7:169182.
Lasnik, Howard. 2001. When can you save a structure by destroying it? In Pro-
ceedings of NELS 31 , ed. Minjoo Kim and Uri Strauss, volume 2, 301320.
Lasnik, Howard, and Mamoru Saito. 1984. On the nature of proper government.
Linguistic Inquiry 15:235289.
Lasnik, Howard, and Mamoru Saito. 1992. Move alpha: Conditions on its application
and output . Cambridge, MA: MIT Press.
Legate, Julie Anne, and Charles Yang. 2007. Morphosyntactic learning and the
development of tense. Language Acquisition 14:315344.
Legate, Julie Anne, and Charles D Yang. 2002. Empirical re-assessment of stimulus
poverty arguments. The Linguistic Review 18:151162.
Levin, Beth. 1993. English verb classes and alternations: A preliminary investiga-
tion. University of Chicago Press.
Lewis, Richard. 1996. Interference in short-term memory: The magical number two
(or three) in sentence processing. Journal of Psycholinguistic Research 25:93115.
Li, Audrey Y.-H. 1990. Order and constituency in Mandarin Chinese. Dordrecht:
Kluwer.
Lin, Jo-Wang. 1992. The syntax of zenmeyang `how' and weishenme `why' in Man-
darin Chinese. Journal of East Asian Linguistics 1:293331.
Lin, T.-H. Jonah, and W.-W. Roger Liao. 2019. Syntactic structures of Mandarin
purposives. Linguistics 57:87126.
Lin, Tzong-Hong Jonah. 2015. Tense in Mandarin Chinese sentences. Syntax 18:320
342.
Lindahl, Filippa. 2017. Extraction from relative clauses in Swedish. Doctoral Dis-
sertation, University of Gothenburg.
320
Liu, Yingtong, Rachel Ryskin, Richard Futrell, and Edward Gibson. 2019. Factive
and manner-of-speaking islands are an artifact of nonlinearity in the acceptability
judgment task. Poster presented at the 32nd CUNY Conference on Sentence
Processing.
MacWhinney, Brian. 2000. The CHILDES project: Tools for analyzing talk. Third
Edition. Mahwah, New Jersey: Lawrence Erlbaum Associates.
Manetta, Emily. 2011. Peripheries in Kashmiri and Hindi-Urdu: The syntax of
discourse-driven movement . Amsterdam/Philadelphia: John Benjamins Publish-
ing.
Manning, Christopher D., Mihai Surdeanu, John Bauer, Steven J. Bethard
Jenny Finkel, and David McClosky. 2014. The Stanford CoreNLP Natural Lan-
guage Processing Toolkit. In Proceedings of the 52nd Annual Meeting of the
Association for Computational Linguistics: System Demonstrations , 5560.
Manzini, M. Rita, and Kenneth Wexler. 1987. Parameters, binding theory, and
learnability. Linguistic Inquiry 18:413444.
Manzini, Maria Rita. 1992. Locality: A theory and some of its empirical conse-
quences . Cambridge, MA: MIT Press.
Mar?cz, L?szl?. 1991. Asymmetries in Hungarian (iii). Anuario del Seminario de
Filolog?a Vasca Julio de Urquijo" 25:153253.
Marcus, Mitchell P., Beatrice Santorini, Mary Ann Marcinkiewicz, and Ann Taylor.
1999. Treebank-3 LDC99T42.
McCawley, James D. 1981. The syntax and semantics of English relative clauses.
Lingua 53:99149.
McCawley, James D. 1998. The syntactic phenomena of English. Chicago: Univer-
sity of Chicago Press.
McCloskey, James. 2006. Questions and questioning in a local English. In Crosslin-
guistic research in syntax and semantics: Negation, tense, and clausal architec-
ture, ed. Raaella Zanuttini, H?ctor Campos, Elena Herburger, and Paul Portner,
87126. Georgetown University Press Washington, DC.
McNally, Louise. 1997. A semantics for the English existential construction. New
York: Garland.
McNally, Louise. 2009. Properties, entity correlates of properties, and existentials.
In Quantication, deniteness, and nominalization, ed. Anastasia Giannakidou
and Monika Rathert, 163187. Oxford: Oxford University Press.
Merchant, Jason. 2001. The syntax of silence. Oxford: Oxford University Press.
321
Merchant, Jason. 2002. Swiping in Germanic. In Studies in Comparative Germanic
Syntax: Proceedings from the 15th Workshop on Comparative Germanic Syntax ,
ed. C. Jan-Wouter Zwart and Werner Abraham, volume 53, 289315. Amsterdam:
John Benjamins.
Michel, Dan, and Grant Goodall. 2013. Finiteness and the nature of island con-
straints. In Proceedings of GLOW in Asia IX: The Main Session, ed. Nobu Goto,
Koichi Otaki, Atsushi Sato, and Kensuke Takita, 188197. Mie, Japan: Mie Uni-
versity.
Mikolov, Tomas, Ilya Sutskever, Kai Chen, Greg S Corrado, and Je Dean. 2013.
Distributed representations of words and phrases and their compositionality. In
Advances in neural information processing systems , 31113119.
Miller, George A. 1995. Wordnet: a lexical database for English. Communications
of the ACM 38:3941.
M?ller, Gereon. 1995. A-bar syntax: A study in movement types . Berlin/New York:
Mouton de Gruyter.
M?ller, Christiane. 2015a. Against the small clause hypothesis: evidence from
Swedish relative clause extractions. Nordic Journal of Linguistics 38:6792.
M?ller, Sonja. 2015b. Deriving island constraints with Searle and Grice. a pragmatic
account of bridge verbs. Studia Linguistica 69:157.
Nakagawa, Shinichi, and Holger Schielzeth. 2013. A general and simple method for
obtaining R2 from generalized linear mixed-eects models. Methods in Ecology
and Evolution 4:133142.
Nissenbaum, Jon. 2005. States, events and VP structure: evidence from purposive
adjuncts. Presented at the 36th meeting of the North East Linguistic Society,
Amherst, MA.
Ogihara, Toshiyuki. 1996. Tense, attitudes, and scope. Dordrecht: Kluwer Academic
Publishers.
Oshima, David Y. 2006. On factive islands: pragmatic anomaly vs. pragmatic
infelicity. In New Frontiers in Articial Intelligence, ed. Takashi Washio, Ken
Satoh, Hideaki Takeda, and Akihiro Inokuchi, 147161. Springer.
Partee, Barbara Hall. 1974. Some structural analogies between tenses and pronouns
in English. The Journal of Philosophy 70:601609.
Paul, Waltraud. 2008. The serial verb construction in Chinese: A tenacious myth
and a gordian knot. The Linguistic Review 25:367411.
Pearl, Lisa. 2007. Necessary bias in natural language learning. Doctoral Dissertation,
University of Maryland.
322
Pearl, Lisa, and Jon Sprouse. 2013a. Computational models of acquisition for is-
lands. In Experimental syntax and islands eects , ed. Jon Sprouse and Norbert
Hornstein, 109131. Cambridge: Cambridge University Press.
Pearl, Lisa, and Jon Sprouse. 2013b. Syntactic islands and learning biases: Combin-
ing experimental syntax and computational modeling to investigate the language
acquisition problem. Language Acquisition 20:2368.
Pearl, Lisa S, and Benjamin Mis. 2016. The role of indirect positive evidence in
syntactic acquisition: A look at anaphoric one. Language 92:130.
Pen?ev, Iordan. 1998. Sintaksis na s?vremennija b?lgarski kni?oven ezik . Plovdiv.
Perfors, Amy, Joshua B Tenenbaum, Thomas L Griths, and Fei Xu. 2011a. A
tutorial introduction to Bayesian models of cognitive development. Cognition
120:302321.
Perfors, Amy, Joshua B Tenenbaum, and Terry Regier. 2011b. The learnability of
abstract syntactic principles. Cognition 118:306338.
Perfors, Amy, Joshua B Tenenbaum, and Elizabeth Wonnacott. 2010. Variability,
negative evidence, and the acquisition of verb argument constructions. Journal
of Child Language 37:607642.
Perkins, Laurel, Naomi Feldman, and Jerey Lidz. 2017. Learning an input lter for
argument structure acquisition. In Proceedings of the 7th workshop on Cognitive
Modeling and Computational Linguistics , 1119.
Pesetsky, David. 1987. Wh-in-situ: Movement and unselective binding. In The
representation of (in)deniteness , ed. Eric J. Reuland and Alice G. B. ter Meulen,
98129. Cambridge, MA: MIT Press.
Pesetsky, David Michael. 1982. Paths and categories. Doctoral Dissertation, Mas-
sachusetts Institute of Technology, Cambridge, MA.
Pinker, Steven. 1984. Language learnability and language development . Cambridge,
MA: Harvard University Press.
Pinker, Steven. 1989. Learnability and cognition: The acquisition of argument struc-
ture. Cambridge, MA: MIT Press.
Postal, Paul M. 1972. A few factive facts. Linguistic Inquiry 3:396400.
Postal, Paul M. 1998. Three investigations of extraction. Cambridge, MA: MIT
Press.
Preminger, Omer. 2010. Nested interrogatives and the locus of wh. In The com-
plementizer phase: Subjects and operators , ed. E. Phoevos Panagiotidis, 200235.
Oxford University Press.
323
Preminger, Omer. 2014. Agreement and its failures . Cambridge, MA: MIT Press.
Pullum, Georey K, and Barbara C Scholz. 2002. Empirical assessment of stimulus
poverty arguments. The Linguistic Review 18:950.
R Core Team. 2019. R: A language and environment for statistical computing .
R Foundation for Statistical Computing, Vienna, Austria. URL https://www.
R-project.org/.
Radford, Andrew, and Eiichi Iwasaki. 2015. On swiping in English. Natural Lan-
guage & Linguistic Theory 33:703744.
Rappaport Hovav, Malka, and Beth Levin. 2008. The English dative alternation:
The case for verb sensitivity. Journal of Linguistics 44:129167.
Reali, Florencia, and Morten H Christiansen. 2005. Uncovering the richness of
the stimulus: Structure dependence and indirect statistical evidence. Cognitive
Science 29:10071028.
?eh??ek, Radim, and Petr Sojka. 2010. Software Framework for Topic Mod-
elling with Large Corpora. In Proceedings of the LREC 2010 Workshop on
New Challenges for NLP Frameworks , 4550. Valletta, Malta: ELRA. http:
//is.muni.cz/publication/884893/en.
Resnik, Philip. 1995. Using information content to evaluate semantic similarity. In
Proceedings of the 14th International Joint Conference on Articial Intelligence,
448453.
Rissman, Lilia, and Kyle Rawlins. 2017. Ingredients of instrumental meaning. Jour-
nal of Semantics 34:507537.
Rizzi, Luigi. 1982. Issues in Italian syntax . Dordrecht: Foris.
Rizzi, Luigi. 1990. Relativized minimality . Cambridge, MA: MIT Press.
Romero, Maribel. 2004. Tense and intensionality in specicational copular sentences.
In Proceedings of Semantics and Linguistic Theory 14 , ed. Robert B. Young, 0,
271288.
Romero, Maribel. 2005. Concealed questions and specicational subjects. Linguistics
and Philosophy 28:687737.
Ross, John R. 1967. Constraints on variables in syntax. Doctoral Dissertation,
Massachusetts Institute of Technology.
Rubovitz-Mann, Talia. 2000. Extraction from relative clauses: An information struc-
ture account. Doctoral Dissertation, Hebrew University of Jerusalem.
Rudin, Catherine. 1988. On multiple questions and multiple wh fronting. Natural
Language & Linguistic Theory 6:445501.
324
Sar, Ken. 1999. Vehicle change and reconstruction in A	-chains. Linguistic Inquiry
30:587620.
Salzmann, Martin. 2017. Reconstruction and resumption in indirect A?-
dependencies . Berlin/Boston: Walter de Gruyter GmbH & Co KG.
Salzmann, Martin. 2018. A new version of the matching analysis of relative clauses.
In Reconstruction eects in relative clauses , ed. Manfred Krifka and Mathias
Schenner, 187223. Berlin: Walter de Gruyter.
Sauerland, Ulrich. 1998. On the making and meaning of chains. Doctoral Disserta-
tion, Massachusetts Institute of Technology.
Schachter, Paul. 1973. Focus and relativization. Language 49:1946.
Schippers, Ankelien. 2012. Variation and change in Germanic long-distance depen-
dencies. Doctoral Dissertation, University of Groningen.
Schwager, Magdalena. 2009. Speaking of qualities. In Proceedings of Semantics and
Linguistic Theory 19 , ed. Ed Cormany, Satoshi Ito, , and David Lutz, 0, 395412.
Schwarz, Bernhard, and Alexandra Simonenko. 2018. Factive islands and meaning-
driven unacceptability. Natural Language Semantics 26:253279.
Schwarz, Florian. 2012. Situation pronouns in determiner phrases. Natural Language
Semantics 20:431475.
Sharvit, Yael. 2003. Tense and identity in copular constructions. Natural Language
Semantics 11:363393.
Sharvit, Yael. 2013. On the universal principles of tense embedding: The lesson
from before. Journal of Semantics 31:263313.
Sichel, Ivy. 2018. Anatomy of a counterexample: Extraction from relative clauses.
Linguistic Inquiry 49:335378.
Simonenko, Alexandra. 2013. Grammatical ingredients of deniteness. Doctoral
Dissertation, McGill University.
Simonenko, Alexandra. 2015. Semantics of DP islands: The case of questions.
Journal of Semantics 33:661702.
Snyder, William. 1992. Wh-extraction and the lexical representation of verbs. Un-
published manuscript available from author's website.
Snyder, William. 2001. On the nature of syntactic variation: Evidence from complex
predicates and complex word-formation. Language 77:324342.
Soderstrom, Melanie, Megan Blossom, Rina Foygel, and James L Morgan. 2008.
Acoustical cues and grammatical units in speech to two preverbal infants. Journal
of Child Language 35:869902.
325
Speas, Margaret. 2004. Evidentiality, logophoricity and the syntactic representation
of pragmatic features. Lingua 114:255276.
Spector, Benjamin, and Paul Egr?. 2015. A uniform semantics for embedded inter-
rogatives: An answer, not necessarily the answer. Synthese 192:17291784.
Sportiche, Dominique. 2011. French relative qui. Linguistic Inquiry 42:83124.
Sprouse, Jon, Ivano Caponigro, Ciro Greco, and Carlo Cecchetto. 2016. Experi-
mental syntax and the variation of island eects in English and Italian. Natural
Language & Linguistic Theory 34:307344.
Stateva, Penka. 2005. On the status of parasitic gaps in Bulgarian. Journal of Slavic
Linguistics 137155.
von Stechow, Arnim, and Atle Gr?nn. 2013a. Tense in adjuncts part 1: Relative
clauses. Language and Linguistics Compass 7:295310.
von Stechow, Arnim, and Atle Gr?nn. 2013b. Tense in adjuncts part 2: Temporal
adverbial clauses. Language and Linguistics Compass 7:311327.
Stepanov, Arthur. 2001. Late adjunction and minimalist phrase structure. Syntax
4:94125.
Stepanov, Arthur. 2007. The end of CED? Minimalism and extraction domains.
Syntax 10:80126.
Stowell, Tim. 1993. Syntax of tense. Manuscript, University of California, Los
Angeles.
Stump, Gregory Thomas. 1985. The semantic variability of absolute constructions .
Dordrecht: Reidel.
Sugisaki, Koji. 2008. Swiping in child english. In Proceedings of the 27th West Coast
Conference on Formal Linguistics , ed. Kevin Ryan, 7177.
Sundaresan, Sandhya. 2018. An alternative model of indexical shift: Variation and
selection without context-overwriting. Manuscript.
Suppes, Patrick. 1974. The semantics of children's language. American Psychologist
29:103114.
Sybesma, Rint. 2007. Whether we tense-agree overtly or not. Linguistic Inquiry
38:580587.
Szabolcsi, Anna, and Frans Zwarts. 1993. Weak islands and an algebraic semantics
for scope taking. Natural Language Semantics 1:235284.
326
Szczegielniak, Adam. 2004. Two types of relative clauses in Slavic  evidence from
reconstruction and ellipsis. In Proceedings of the Thirty-Third Western Conference
on Linguistics , ed. Michal Temkin Mart?nez, Asier Alc?zar, and Roberto Mayoral
Hern?ndez, 373384. Fresno: California State University.
Taraldsen, Knut T. 1982. Extraction from relative clauses in Norwegian, 205221.
Stockholm: Almqvist and Wiksell.
Theakston, Anna L., Elena V. M. Lieven, Julian M. Pine, and Caroline F. Rowland.
2001. The role of performance limitations in the acquisition of verb-argument
structure: An alternative account. Journal of Child Language 28:127152.
Truswell, Robert. 2007. Extraction from adjuncts and the structure of events. Lingua
117:13551377.
Truswell, Robert. 2008. Preposition stranding, passivisation, and extraction from
adjuncts in Germanic. In Linguistic variation yearbook , ed. Jeroen van Craenen-
broeck, volume 8, 131178. John Benjamins Publishing Company.
Truswell, Robert. 2011. Events, phrases, and questions . Oxford: Oxford University
Press.
Tsai, Wei-Tien Dylan. 1995. Visibility, complement selection and the case require-
ment of CP. Journal of East Asian Linguistics 4:281312.
Uriagereka, Juan. 1999. Multiple spell-out. In Working minimalism, ed. Samuel
Epstein and Norbert Hornstein, 251282. MIT Press.
Valian, Virginia. 1990. Null subjects: A problem for parameter-setting models of
language acquisition. Cognition 35:105122.
Valian, Virginia. 1991. Syntactic subjects in the early speech of American and Italian
children. Cognition 40:2181.
Vergnaud, Jean-Roger. 1974. French relative clauses. Doctoral Dissertation, Mas-
sachusetts Institute of Technology.
Verhagen, Arie. 2006. On subjectivity and long distance wh-movement. In Sub-
jectication: Various paths to subjectivity , ed. Angeliki Athanasiadou, Costas
Canakis, and Bert Cornillie, 323346. Berlin: Mouton de Gruyter.
Viau, Joshua, and Jerey Lidz. 2011. Selective learning in the acquisition of kannada
ditransitives. Language 87:679714.
Wagner, Laura. 1998. The semantics and acquisition of time in language. Doctoral
Dissertation, University of Pennsylvania.
Wexler, Kenneth, and Peter Culicover. 1980. Formal principles of language acquisi-
tion. Cambridge, MA: MIT Press.
327
Whelpton, Matthew James. 1995. The syntax and semantics of innitives of result
in English. Doctoral Dissertation, University of Oxford.
White, Aaron Steven, Valentine Hacquard, and Jerey Lidz. 2018. The labeling
problem in syntactic bootstrapping. In Semantics in language acquisition, ed.
Kristen Syrett and Sudha Arunachalam, 198220. Amsterdam: John Benjamins.
White, Aaron Steven, and Kyle Rawlins. 2016. A computational model of S-
selection. In Proceedings of Semantics and Linguistic Theory 26 , ed. Mary Mo-
roney, Carol-Rose Little, Jacob Collard, and Dan Burgdorf, volume 26, 641663.
White, Aaron Steven, and Kyle Rawlins. 2018. The role of veridicality and factivity
in clause selection. In Proceedings of NELS 48 , ed. Sherry Hucklebridge and Max
Nelson, 221235.
Witko?, Jacek. 1995. Wh-extraction from clausal complements in Polish: A mini-
mality/locality account. Folia Linguistica 29:223264.
Xu, Fei, and Joshua B Tenenbaum. 2007. Word learning as Bayesian inference.
Psychological Review 114:245272.
Xu, Liejiong, and D. Terence Langendoen. 1985. Topic structures in Chinese. Lan-
guage 61:127.
Xue, Nianwen, Zixin Jiang, Xiuhong Zhong, Martha Palmer, Fei Xia, Fu-Dong
Chiou, and Meiyu Chang. 2010. Chinese Treebank 7.0 LDC2010T07.
Yang, Charles. 2002. Knowledge and learning in natural language. Oxford: Oxford
University Press.
Yang, Charles. 2016. The price of linguistic productivity: How children learn to
break the rules of language. Cambridge, MA: MIT Press.
Yang, Charles. 2017. Rage against the machine: Evaluation metrics in the 21st
century. Language Acquisition 24:100125.
Zhang, Niina. 2001. The structures of depictive and resultative constructions in
Chinese. In ZAS Papers in Linguistics 26 (Proceedings of the workshop on the
syntax of Predication). Berlin: ZAS.
Zhang, Niina. 2002. Island eects and episodic eventualities in Chinese topicaliza-
tion. In Linguistics by heart: Webfest for Horst-Dieter Gasde, ed. Daniel Hole,
Paul Law, and Niina Zhang. Berlin: Zentrum Allgemeine Sprachwissenschaft.
Zwicky, Arnold M. 1971. In a manner of speaking. Linguistic Inquiry 2:223233.
328