ABSTRACT
Title of dissertation: ADJUNCT CONTROL:
SYNTAX AND PROCESSING
Jeffrey Jack Green, Doctor of Philosophy, 2018
Dissertation directed by: Professor Alexander Williams
Department of Linguistics
This dissertation analyzes the syntax and processing of adjunct control. Ad-
junct control is the referential relation between the implicit (PRO) subject of a
non-finite adjunct clause and its understood antecedent, as in the temporal adjunct
in ‘Holly1 went to bed [after PRO1 drinking milk] ’, or the rationale clause in ‘August1
sat on the couch [in order PRO1 to read library books] ’. Adjunct control is often as-
sumed to involve a syntactic ‘Obligatory Control’ (OC) dependency, but I show that
some adjuncts also permit what is referred to as ‘Non-Obligatory Control’ (NOC),
as in the sentences ‘The food tasted better [after PRO drinking milk] ’ and ‘The book
was checked out from the library [in order PRO to read it] ’, where PRO refers to
some unnamed entity. I argue that for some adjuncts, OC and NOC are not in
complementary distribution, contrary to assumptions of much prior literature, but
in agreement with Landau (2017). Contrary to implicit assumptions of Landau,
however, I also show that this OC/NOC duality does not extend to all adjuncts. I
outline assumptions that Landau’s theory would have to make in order to accom-
modate the wider distribution of OC and NOC in adjuncts, but argue that this is
better accomplished within the Movement Theory of Control (Hornstein, 1999) by
relaxing the assumption that all adjuncts are phases.
Even in adjuncts where both OC and NOC are possible, OC is often strongly
preferred. I argue that this is in large part due to interpretive biases in processing.
As a foundational step in examining what these processing biases are, the second
part of this dissertation uses visual-world eyetracking to compare the timecourse of
interpretation of subject-controlled PRO and overt pronouns in temporal adjuncts.
The results suggest that PRO can be interpreted just as quickly as overt pronouns
once the relevant bottom-up input is received. These experiments also provide
evidence that structural predictions can facilitate reference resolution independent
of next-mention predictions.
ADJUNCT CONTROL:
SYNTAX AND PROCESSING
by
Jeffrey Jack Green
Dissertation submitted to the Faculty of the Graduate School of the
University of Maryland, College Park in partial fulfillment
of the requirements for the degree of
Doctor of Philosophy
2018
Advisory Committee:
Professor Alexander Williams, Chair
Professor Ellen Lau
Professor Norbert Hornstein
Professor Maria Polinsky
Professor Jan Edwards
©c Copyright by
Jeffrey Jack Green
2018
Forward
Portions of Chapter 3 represent joint work with Alexander Williams and are
published in A. Williams and Green (2017). The experiments in Chapter 6 were com-
pleted in collaboration with Michael McCourt, Ellen Lau, and Alexander Williams.
ii
Dedication
To Alli, August, and Holly,
who are the reason I do all of this.
iii
Acknowledgments
As is the case with any major milestone, this dissertation is the culmination
of a long journey, along which there have been many people who have helped me
in more ways than I can express, and more than I think even I realize. I’d like to
thank just a few of those people.
First, thank you to my parents and siblings, who taught me to work hard and
who always believe in me.
Thank you to my fellow graduate students. To my cohort, with whom it has
been a pleasure to work and study. To Mike McCourt, who let me join his research
group on implicit control and who remained a helpful and supportive collaborator
for years after. And to others who helped me learn various research and analy-
sis methods, including Sol Lago, Dan Parker, Lara Ehrenhofer, Rachel Adler, Jon
Burnsky, and Anna Namyst. Thanks also to many other fellow students and friends
who supported me throughout my program.
Thank you to my committee. First, to my advisor and mentor Alexander
Williams, for always having insightful comments and questions, for his extensive
feedback on many stages of writing this dissertation, and for always encouraging
me. To Ellen Lau, for her mentorship on all my psycholinguistic work and for
iv
helping me see and focus on the broader implications of my research. To Norbert
Hornstein and Maria Polinsky, for helpful discussions and interesting insights. To
Jan Edwards, for teaching me to understand growth curve modeling. And to all of
them for being willing to read this dissertation.
Thank you to other faculty members who have helped me in my research,
including Howard Lasnik and Omer Preminger, and pretty much everyone else in
the department. To audiences at several conferences and at talks at UMD, and
to anonymous reviewers of abstracts and drafts of work from several parts of this
dissertation. To my teaching mentors at UMD: Valentine Hacquard, Tonia Bleam,
and Peggy Antonisse. And to my undergraduate research and teaching mentors:
Rachel Hayes-Harb, Darryl Kropf, Aniko Csirmaz, and Chenghua Hu.
To my funding sources: the Department of Linguistics, the Language Science
Center, Alexander Williams and Ellen Lau’s research funds, and the National Science
Foundation under Grant No. 1449815. To Kim Kwok to invaluable help in pretty
much all of my administrative questions over the last five years.
I would like to express deep gratitude to God, who directed me to try linguistics
and to whom I attribute the inspiration behind these ideas and the strength to
continue, and who put so many wonderful people in my life.
Finally, thank you to my wonderful family. Thank you to August, who always
cheers me on. And to Holly, who gives me more reasons to smile. And thank you to
Alli, who kept me working hard, believed in me when I got discouraged, and helped
me believe in myself. Her love and support have been the most important thing to
me in the past five years.
v
Contents
Forward ii
Dedication iii
Acknowledgments iv
Contents vi
List of Tables xi
List of Figures xii
List of Abbreviations xiii
I Theory 1
1 Introduction 2
1.1 Constraints on reference . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Obligatory and Non-obligatory Control . . . . . . . . . . . . . . . . . 6
1.3 Adjunct control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Organization of this dissertation . . . . . . . . . . . . . . . . . . . . . 12
2 Temporal adjuncts 15
2.1 General properties of temporal adjuncts . . . . . . . . . . . . . . . . 17
2.2 Temporal adjuncts classified as OC . . . . . . . . . . . . . . . . . . . 18
2.3 The distribution of NOC in temporal adjuncts . . . . . . . . . . . . . 22
2.3.1 Structural effects . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.1.1 Attachment site of the adjunct . . . . . . . . . . . . 23
2.3.1.2 Structural properties of the matrix clause . . . . . . 27
2.3.1.3 Structural properties of the temporal adjunct . . . . 37
2.3.2 Non-structural effects . . . . . . . . . . . . . . . . . . . . . . . 40
2.3.2.1 Availability of subject-control reading . . . . . . . . 41
2.3.2.2 “Initiator” status . . . . . . . . . . . . . . . . . . . . 46
2.3.2.3 Topichood . . . . . . . . . . . . . . . . . . . . . . . . 47
2.3.2.4 Logophoricity . . . . . . . . . . . . . . . . . . . . . . 51
2.4 Proposal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
vi
2.4.1 Grammar allows both OC and NOC . . . . . . . . . . . . . . 58
2.4.2 Preference for OC . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.4.2.1 Preference for local antecedent . . . . . . . . . . . . 60
2.4.2.2 Preference for syntactic relations . . . . . . . . . . . 62
2.4.2.3 Theory-internal preferences . . . . . . . . . . . . . . 63
2.4.3 Factors favoring NOC . . . . . . . . . . . . . . . . . . . . . . 65
2.4.3.1 Attachment site . . . . . . . . . . . . . . . . . . . . . 66
2.4.3.2 Unavailability of subject control . . . . . . . . . . . . 68
2.4.3.3 Implicit arguments . . . . . . . . . . . . . . . . . . . 69
2.4.3.4 Humanness of subject . . . . . . . . . . . . . . . . . 69
2.4.3.5 Logophoricity in context . . . . . . . . . . . . . . . . 70
2.5 Implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.5.1 Addressing original evidence of OC . . . . . . . . . . . . . . . 72
2.5.2 Other types of control . . . . . . . . . . . . . . . . . . . . . . 76
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3 Rationale clauses 81
3.1 Background on Rationale Clauses . . . . . . . . . . . . . . . . . . . . 82
3.1.1 Local RatCs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.1.2 Remote RatCs . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.2 OC accounts of RatCs . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.2.1 OC and local control in RatCs . . . . . . . . . . . . . . . . . . 93
3.2.1.1 Control by the subject, not by the object or non-c-
commanding arguments . . . . . . . . . . . . . . . . 94
3.2.1.2 Implicit control . . . . . . . . . . . . . . . . . . . . . 95
3.2.1.3 Superimplicit control . . . . . . . . . . . . . . . . . . 97
3.2.1.4 Fact control . . . . . . . . . . . . . . . . . . . . . . . 98
3.2.1.5 Summary of local RatCs . . . . . . . . . . . . . . . . 99
3.2.2 OC and remote control . . . . . . . . . . . . . . . . . . . . . . 99
3.2.2.1 Specifier ellipsis . . . . . . . . . . . . . . . . . . . . . 101
3.2.2.2 Relative clause ellipsis . . . . . . . . . . . . . . . . . 106
3.2.3 Summary of OC accounts of RatCs . . . . . . . . . . . . . . . 111
3.3 NOC accounts of RatCs . . . . . . . . . . . . . . . . . . . . . . . . . 112
3.3.1 Farkas 1988 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
3.3.2 Responsibility Theory . . . . . . . . . . . . . . . . . . . . . . 117
3.3.2.1 Responsibility Theory and subject control . . . . . . 118
3.3.2.2 Responsibility Theory and implicit control . . . . . . 121
3.3.2.3 Responsibility Theory and superimplicit control . . . 123
3.3.2.4 Responsibility Theory and fact control . . . . . . . . 125
3.3.2.5 Responsibility Theory and remote control . . . . . . 127
3.3.3 Challenges for Responsibility Theory . . . . . . . . . . . . . . 128
3.3.3.1 Ruling out object control . . . . . . . . . . . . . . . 129
3.3.3.2 Cases of OC in RatCs . . . . . . . . . . . . . . . . . 133
3.4 Both OC and NOC are possible for RatCs . . . . . . . . . . . . . . . 135
3.4.1 The distribution of OC and NOC in RatCs . . . . . . . . . . . 136
vii
3.4.1.1 Remote control is always NOC . . . . . . . . . . . . 136
3.4.1.2 Local subject control . . . . . . . . . . . . . . . . . . 137
3.4.1.3 Semantic anomaly and local NOC . . . . . . . . . . . 139
3.4.2 Ruling out local object control . . . . . . . . . . . . . . . . . . 141
3.4.2.1 OC required with passive RatCs? . . . . . . . . . . . 144
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
4 Control theories and other adjuncts 151
4.1 The two-tiered theory of control . . . . . . . . . . . . . . . . . . . . . 152
4.1.1 Background on the TTC . . . . . . . . . . . . . . . . . . . . . 152
4.1.2 Why OC is preferred over NOC . . . . . . . . . . . . . . . . . 158
4.1.3 Why complement control only allows OC . . . . . . . . . . . . 159
4.1.4 The interpretation of NOC . . . . . . . . . . . . . . . . . . . . 161
4.2 Other adjuncts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
4.2.1 Adjuncts allowing both OC and NOC . . . . . . . . . . . . . . 162
4.2.1.1 Response clauses . . . . . . . . . . . . . . . . . . . . 162
4.2.1.2 Object-gap purpose clauses . . . . . . . . . . . . . . 172
4.2.1.3 Absolutives . . . . . . . . . . . . . . . . . . . . . . . 173
4.2.2 Adjuncts with only NOC: Speaker-oriented adverbials . . . . . 174
4.2.3 Adjuncts with only OC . . . . . . . . . . . . . . . . . . . . . . 176
4.2.3.1 Goal clauses . . . . . . . . . . . . . . . . . . . . . . . 176
4.2.3.2 Result clauses and stimulus clauses . . . . . . . . . . 178
4.2.3.3 Subject-gap purpose clauses . . . . . . . . . . . . . . 180
4.2.3.4 Telic clauses . . . . . . . . . . . . . . . . . . . . . . . 181
4.3 The movement theory of control . . . . . . . . . . . . . . . . . . . . . 184
4.3.1 Background on the MTC . . . . . . . . . . . . . . . . . . . . . 184
4.3.1.1 Why OC is preferred over NOC, but not required . . 186
4.3.1.2 Why complement control differs . . . . . . . . . . . . 189
4.3.1.3 The interpretation of NOC . . . . . . . . . . . . . . 193
4.3.2 The MTC and other adjuncts . . . . . . . . . . . . . . . . . . 194
4.3.3 Problems for the MTC . . . . . . . . . . . . . . . . . . . . . . 199
4.3.3.1 Telic clauses . . . . . . . . . . . . . . . . . . . . . . . 199
4.3.3.2 Speaker-oriented adverbials . . . . . . . . . . . . . . 201
4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
II Processing 206
5 Processing adjunct control 207
5.1 Previous research on the processing of adjunct control . . . . . . . . . 208
5.1.1 Syntactic representation . . . . . . . . . . . . . . . . . . . . . 208
5.1.2 Speed of processing control structures . . . . . . . . . . . . . . 213
5.1.3 The quality of retrieval . . . . . . . . . . . . . . . . . . . . . . 215
5.2 Remaining questions . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
viii
6 Processing temporal adjuncts 220
6.1 Background: anaphora resolution generally . . . . . . . . . . . . . . . 221
6.2 Experiment 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
6.2.1 Materials and methods . . . . . . . . . . . . . . . . . . . . . . 228
6.2.1.1 Participants . . . . . . . . . . . . . . . . . . . . . . . 228
6.2.1.2 Materials . . . . . . . . . . . . . . . . . . . . . . . . 228
6.2.1.3 Procedure . . . . . . . . . . . . . . . . . . . . . . . . 232
6.2.1.4 Data analysis . . . . . . . . . . . . . . . . . . . . . . 233
6.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
6.2.2.1 Effect of focus . . . . . . . . . . . . . . . . . . . . . . 236
6.2.2.2 Estimating time of interpretation . . . . . . . . . . . 237
6.2.2.3 Effect of Cue . . . . . . . . . . . . . . . . . . . . . . 238
6.2.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
6.2.3.1 A general subject preference? . . . . . . . . . . . . . 241
6.2.3.2 Rapid use of structural information . . . . . . . . . . 242
6.2.3.3 Prediction in the processing of PRO? . . . . . . . . . 244
6.2.3.4 Learned strategy? . . . . . . . . . . . . . . . . . . . 246
6.3 Experiment 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
6.3.1 Materials and methods . . . . . . . . . . . . . . . . . . . . . . 247
6.3.1.1 Participants . . . . . . . . . . . . . . . . . . . . . . . 247
6.3.1.2 Materials . . . . . . . . . . . . . . . . . . . . . . . . 248
6.3.1.3 Procedure . . . . . . . . . . . . . . . . . . . . . . . . 248
6.3.1.4 Data analysis . . . . . . . . . . . . . . . . . . . . . . 249
6.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
6.3.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
6.4 Experiment 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
6.4.1 Materials and methods . . . . . . . . . . . . . . . . . . . . . . 255
6.4.1.1 Participants . . . . . . . . . . . . . . . . . . . . . . . 255
6.4.1.2 Materials . . . . . . . . . . . . . . . . . . . . . . . . 256
6.4.1.3 Procedure . . . . . . . . . . . . . . . . . . . . . . . . 257
6.4.1.4 Data analysis . . . . . . . . . . . . . . . . . . . . . . 257
6.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
6.4.2.1 Effect of focus . . . . . . . . . . . . . . . . . . . . . . 258
6.4.2.2 Estimating time of interpretation . . . . . . . . . . . 259
6.4.2.3 Effect of cue . . . . . . . . . . . . . . . . . . . . . . 259
6.4.2.4 Effect of cloze . . . . . . . . . . . . . . . . . . . . . . 260
6.4.2.5 Plotting from offset of critical word . . . . . . . . . . 261
6.4.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
6.4.3.1 Summary of findings . . . . . . . . . . . . . . . . . . 263
6.4.3.2 Effect of prediction . . . . . . . . . . . . . . . . . . . 264
6.4.3.3 Effect of cue length? . . . . . . . . . . . . . . . . . . 265
6.5 General discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
6.5.1 Prediction in the resolution of anaphora . . . . . . . . . . . . 269
6.5.2 Additional implications . . . . . . . . . . . . . . . . . . . . . . 271
6.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
ix
7 Conclusion 274
7.1 Implications of this research . . . . . . . . . . . . . . . . . . . . . . . 275
7.1.1 Adjuncts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
7.1.2 The OC/NOC distinction . . . . . . . . . . . . . . . . . . . . 277
7.1.3 Constraints on reference . . . . . . . . . . . . . . . . . . . . . 280
7.1.4 Prediction in reference resolution . . . . . . . . . . . . . . . . 283
7.2 Remaining questions and future work . . . . . . . . . . . . . . . . . . 284
7.2.1 Variability in adjunct control . . . . . . . . . . . . . . . . . . 284
7.2.2 Biases contributing to OC preference . . . . . . . . . . . . . . 285
7.2.3 Preference for upward movement . . . . . . . . . . . . . . . . 286
7.2.4 Interaction of different sources of information . . . . . . . . . 287
7.2.5 Processing and representation . . . . . . . . . . . . . . . . . . 288
7.2.6 The acquisition of adjunct control . . . . . . . . . . . . . . . . 289
7.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
References 291
x
List of Tables
1 Experiment 1 sample stimuli . . . . . . . . . . . . . . . . . . . . . . . 230
2 Experiment 2 sample cloze items . . . . . . . . . . . . . . . . . . . . 248
3 Experiment 3: Sample fillers . . . . . . . . . . . . . . . . . . . . . . . 256
xi
List of Figures
1 McCourt et al.’s results . . . . . . . . . . . . . . . . . . . . . . . . . . 211
2 Experiment 1: Sample visual stimulus and regions of interest . . . . . 232
3 Experiment 1: Effect of focus . . . . . . . . . . . . . . . . . . . . . . 236
4 Experiment 1: Divergence point . . . . . . . . . . . . . . . . . . . . . 237
5 Experiment 1: Effect of Cue . . . . . . . . . . . . . . . . . . . . . . . 239
6 Experiment 1: Results from offset of critical word . . . . . . . . . . . 245
7 Experiment 2: Responses by type . . . . . . . . . . . . . . . . . . . . 250
8 Experiment 2: Responses by referent . . . . . . . . . . . . . . . . . . 251
9 Effect of cloze on Experiment 1 . . . . . . . . . . . . . . . . . . . . . 252
10 Experiment 3: Sample filler visual stimuli . . . . . . . . . . . . . . . . 257
11 Experiment 3: Effect of focus . . . . . . . . . . . . . . . . . . . . . . 258
12 Experiment 3: Divergence point . . . . . . . . . . . . . . . . . . . . . 259
13 Experiment 3: Effect of Cue . . . . . . . . . . . . . . . . . . . . . . . 260
14 Effect of cloze on Experiment 3 . . . . . . . . . . . . . . . . . . . . . 261
15 Experiment 3: Results from offset of critical word . . . . . . . . . . . 262
xii
List of Abbreviations
act Active
AOI area of interest
CED Condition on Extraction Domain
do Direct Object
Fin Predication head in the TTC
gen Genetive
io Indirect Object
MTC Movement Theory of Control
NOC Non-Obligatory Control
NPI negative polarity item
Num Numeration
OC Obligatory Control
OPC object-gap purpose clause
pass Passive
PC purpose clause
PCC Principle of Controller Choice
PCCRC Principle of Controller Choice in RatCs
RatC rationale clause
RESP Responsibility relation
SPC subject-gap purpose clause
subj Subject
TTC Two-tiered Theory of Control
xiii
Part I
Theory
1
Chapter 1: Introduction
The study of referential dependencies has strongly influenced the development
of syntactic, semantic, and pragmatic theory. The study of how such dependencies
are resolved in online comprehension has similarly influenced our understanding of
the way we integrate information and the rapid deployment of memory structures
in discourse processing. This dissertation explores the syntax and processing of one
family of referential dependencies: the dependency between PRO1 in an adjunct
clause and its understood antecedent in sentences such as those in (1) and (2).
(1) a. Holly1 went to bed [after PRO1 drinking some milk].
b. August1 sat on the couch [in order PRO1 to read some library books].
(2) a. The food tasted better [after PRO drinking milk].
b. The book was checked out from the library [in order PRO to read it].
The first part of the dissertation discusses the distribution of Obligatory Control
1 Although I use the PRO label for clarity throughout the dissertation, the conclusions are not
dependent on a PRO analysis of the implicit subject of control structures. I will also not generally
distinguish between labels of the implicit subject in OC and NOC, as some theories do (such as
Hornstein (1999); see §4.3), although the nature of the implicit subject may indeed differ in the
two types of control. For a discussion on the debate over the representation of PRO, see Hornstein
(2003).
2
(OC) (1) and Non-Obligatory Control (NOC) (2) in adjuncts as well as what factors
may influence the choice of controller. It will be shown that OC and NOC are
both allowed by the grammar for some adjuncts, in contrast to the often-assumed
complementarity of OC and NOC. I argue that instead of sentence structure alone
determining the availability of NOC, extra-grammatical factors such as perspective
and responsibility are also involved.
The second part of the dissertation details a preliminary investigation of the
real-time processing and interpretation of adjunct control where it is shown that
PRO in temporal adjuncts can be interpreted as quickly as overt pronouns in similar
sentences.
1.1 Constraints on reference
Some constraints on reference are grammatical, in that reference is determined
by structural properties of the sentence itself. For example, in (3), the anaphor
himself must be understood as coreferential with Jon’s father ; it cannot refer to any
other entity, no matter what context the sentence is uttered in. The referent of the
reflexive is determined based on the structure of the host sentence. Because reflexives
require a local antecedent in a particular syntactic configuration, Chomsky (1980,
1981) posits a syntactic dependency (binding) between reflexives such as himself
and a c-commanding antecedent (Jon’s father in (3)) within a limited domain. This
syntactic dependency is then interpreted as coreference. Therefore, according to
what he calls Principle A of the Binding Theory (Chomsky, 1981), the referent of a
3
reflexive is completely determined by the grammatical structure of the sentence it
is found in.
(3) Bill1 said [Jon2’s father]3 likes himself3/*1/*2/*4.
A similar grammatical constraint (Principle B) applies to pronouns, but in-
stead of determining reference, the constraint rules out coreference with locally
c-commanding antecedents.2 For example, in (4), the pronoun him may not corefer
with the locally c-commanding Mike.3 However, him may refer to any other salient
male individual, including Jon’s father, in contrast to the reflexive in (3). When
him does corefer with Jon’s father in (4), this does not seem to involve a syntactic
link between the two, since structural position is not decisive. Jon’s father and
him will be understood as coreferential only if pragmatic context permits this, by
making it sufficiently clear that the speaker means to refer to Jon’s father in using
the pronoun. Given the proper context, however, him could just as easily refer to
someone else. Although a grammatical constraint rules out one possible referent for
the pronoun, non-grammatical, contextual constraints determine who a speaker can
ultimately refer to with the word him.
(4) Bill1 told [Jon2’s father]3 that Mike4 likes him1/2/3/5/*4.
2 Although see Heim and Kratzer (1998) and Reinhart (1983) for a discussion of examples such
as (i) from Evans (1980), in which the grammatical constraint on the referent of the pronoun seems
to be suspended.
(i) I know what Bill and Mary have in common. Mary likes Bill and Bill likes him too.
3 Note that by using coindexation in this and other examples, I am not necessarily making the
claim that the relation is syntactic. Instead, coindexation will simply be used to demonstrate
coreference possibilities.
4
Other times, there are no grammatical constraints on reference, and the effect
of non-grammatical constraints becomes more obvious. For example, in interpreting
(5), we are guided by whichever resolution of they results in the most coherent
discourse (Hobbs, 1979; Kehler, 2004; Kehler, Kertz, Rohde, & Elman, 2008). In
(5a) the pronoun is understood as referring to the city council, because it is more
plausible that the denial of a permit would result from their fear of violence than
from the demonstrators’ fears. The preference is reversed in (5b). Notice that the
interpretation of the pronoun here is not being determined grammatically, since it
is in the same structural position in each case. Instead, coherence is constraining
reference.
(5) The city council denied the demonstrators a permit because. . .
a. . . . they feared violence.
b. . . . they advocated violence.
Coherence and plausibility are not the only non-grammatical constraints on
reference, and there are some non-grammatical constraints for which structure mat-
ters, due to its effects on information structure. Kehler (2004) compares (5) to (6),
which differs only in that the first clause is passive. Importantly, the subject of a
passive is even more likely than the subject of an active to be mentioned in sur-
rounding discourse, and in this sense is more ‘topical’. Because of this, the pronouns
in both (6a,b) is resolved to the demonstrators, even though in (6a), resolving they
to the city council would make a more coherent discourse; we are left wondering why
5
the demonstrators’ fear of violence would lead to them being denied a permit. As
Kehler notes, in this case, coreference, determined by information structural factors,
is constraining coherence, rather than the other way around.
(6) The demonstrators were denied a permit by the city council because. . .
a. . . . they feared violence.
b. . . . they advocated violence.
We see, then, that the referent of a reflexive is determined by grammatical con-
straints, while the referent of a pronoun is determined by other factors, even though
syntactic structure can be relevant, even if its relevance is indirect, by way of infor-
mation structure.
1.2 Obligatory and Non-obligatory Control
This dissertation discusses PRO, the silent subject of a non-finite clause. A
similar distinction to the contrast between reflexives and pronouns has been applied
to PRO, under the headings of OC and NOC (Williams, 1980). In traditional defi-
nitions of OC, the referent of PRO is uniquely determined. Consider the examples
in (7), for example. Using reflexives as a diagnostic because they require a local
antecedent, we can see that the referent of PRO is fixed by the expression. In (7a),
PRO can only refer to Harry. It cannot be used to mean that Harry promised that
someone else would make themself invisible. Similarly, (7b) cannot be used to ex-
press that Ron’s mother promised that he would make himself invisible; PRO must
6
refer to Ron’s mother. In (7c), PRO again must refer to Harry; the sentence can-
not be used to report Hermione hearing that Harry promised that she would make
herself invisible.
(7) a. Harry1 promised PRO1 to make himself/*oneself invisible.
b. *Ron1’s mother promised PRO1 to make himself invisible.
c. *Hermione1 heard that Harry promised PRO1 to make herself invisible.
The interpretation of PRO in (7) differs from that of overt pronouns in similar
constructions. In each of the examples in (8), the pronouns can refer to any salient
individual consistent with the pronouns’ gender presuppositions and as determined
by context.
(8) a. Harry1 promised that he1/2 would make himself1/2 invisible.
b. [Ron1’s mother]2 promised that he1/3/she2/3 would make himself1/3/
herself2/3 invisible.
c. Hermione1 heard that Harry2 promised that she1/3/he2/3 would make
herself1/3/himself2/3 invisible.
Because of the interpretive restrictions on PRO in examples such as (7), OC is
often classified as a syntactic constraint on reference, with PRO entering a syntactic
dependency with a locally c-commanding antecedent (Chomsky, 1981; Manzini,
7
1983; Hornstein, 1999; Landau, 2000).4 In other words, the referent of PRO is said
to be determined by the sentence itself, just as was the case with reflexive pronouns.
In NOC, on the other hand, the referent of PRO is determined by the speaker
or hearer, based (at least) on the context of the utterance; the sentence itself is not
sufficient to determine the referent of PRO. In (9a), for example, PRO receives an
arbitrary interpretation; in (9b), it can be coreferential with either Fleur or Bill, or it
can receive a generic/arbitrary interpretation, referring to no individual specifically.
Because the referent of PRO is not strictly determined by the sentence itself, it
is believed not to enter a syntactic or semantic dependency with its understood
antecedent. Instead, its interpretation is similar to an overt pronoun (Chomsky,
1981; Manzini, 1983; Hornstein, 2001; Landau, 2015).
(9) a. [PROarb to park here] is illegal.
b. Fleur1 realized [that it would help Bill2 [PRO1/2/arb to behave herself1/
himself2/oneself in public]].
Control, then, can either result from syntactic binding on the one hand or
discourse resolution of something like a free pronoun on the other. OC occurs
when PRO is bound by a local antecedent in the proper syntactic configuration
(c-command); NOC occurs when PRO’s interpretation is not locally bound. More
generally, two noun phrases in a sentence or discourse can be construed as corefer-
4 There are also those who maintain that the dependency is not syntactic, but semantic, i.e.
that the dependency is mediated by the embedding verb, whose meaning requires that PRO in its
complement be coreferent with another one of its arguments (Farkas, 1988; Chierchia, 1989; Sag &
Pollard, 1991). In either case, the sentential context determines the interpretation of PRO, either
through the semantics of a verb, or through the semantics of a binding relation.
8
ential through either binding (by which I mean any syntactic/semantic dependency
between the two noun-phrases that determines coreference, not just the dependency
so labeled in Chomsky (1981)), or not by binding. In binding, coreference is enforced
by the grammar. It is determined based on syntactic or semantic properties of the
sentence that contains the phrases. Without binding, this is not so. Coreference is
decided not by the syntactic properties of the sentence, but only by its speaker or
hearer, based on contextual and other extra-grammatical constraints.
It has been argued (e.g. by Manzini, 1983) or at least implicitly assumed in
many theories of control, that OC and NOC are grammatically complementary. If
a given structure allows NOC, it is only because OC is blocked. Whether PRO
will be obligatorily or non-obligatorily controlled in examples such as (7) and (9),
for example, is dependent on the structural role of the clause containing PRO (i.e.
whether it is the complement of a control verb or a subject). Even some more
modern theories such as the Movement Theory of Control (MTC), while eschewing
grammatical complementarity between the two, still argue that “NOC is only licit
where OC is not” i.e. that “NOC will only occur where [OC] is prohibited” (Boeckx
& Hornstein, 2007, p. 252). McFadden and Sundaresan (2016) make a similar argu-
ment, stating that NOC only happens when establishing an OC dependency fails.
However, I will demonstrate that this is not the case. In some adjuncts, a single
structure can indeed support both OC and NOC resolutions of PRO.
For the purposes of this dissertation, I will follow Hornstein (1999) and Lan-
dau (2017) in my use of the terms OC and NOC. OC will be used to refer to cases
where the interpretation of PRO is determined by a syntactic dependency between
9
it and a local antecedent, even if such a syntactic dependency is not grammatically
obligatory, as I argue is sometimes the case. Where a syntactic (OC) dependency is
required by the grammar, I will label it “strict OC.” NOC will refer to any case of
control where PRO’s referent is not determined by a syntactic dependency. There
may also be cases of “strict NOC,” where a syntactic antecedent for PRO is gram-
matically impossible. Note that throughout this dissertation I will use “control” in
a somewhat nontraditional way simply to refer to understood coreference between
PRO and an antecedent; PRO will be said to be “controlled” by its antecedent
“controller,” regardless of whether there is any grammatical dependency between
the two.
1.3 Adjunct control
The study of control has most often focused on complement control, as in
(7), and theories of the control dependency have been built around its properties.
Adjunct control has been largely ignored, despite the fact that adjuncts and comple-
ment are quite different syntactically. For example, complements, but not adjuncts,
are selected by the main verb; and adjuncts, but not complements, are often islands
to wh-movement. Where adjunct control is discussed, it is often assumed to be
strictly OC, parallel to complement control (Mohanan, 1983; Clark, 1990; Horn-
stein, 1999; Pires, 2007). Because PRO in adjuncts such as those in (10) can only
be controlled by the matrix subject, it is argued that adjunct control must involve
a syntactic dependency between it and PRO, it being the closest c-commanding
10
potential antecedent to PRO. Even theories that do not rely on a syntactic binding
relation for complement control (e.g. Culicover & Jackendoff, 2001) do presume that
control into adjuncts is syntactically based.5 Assuming complementarity between
OC and NOC, this OC dependency should be obligatory.
(10) a. Ron1 talked to Harry2 [before PRO1/*2 leaving the room].
b. [PRO1/*2 Having washed himself/*herself/*oneself], Harry1 talked to
Ginny2.
c. Harry1 grew up [PRO1/*2 to be a famous wizard].
d. The ball1 fell, [only PRO1/*2 to be picked up again].
Landau (2017), however, argues that temporal and rationale adjuncts are not
restricted to OC. Although the control relation in the examples in (10a) may indeed
be OC, the examples in (11) demonstrate that NOC is also possible.
(11) a. All preparations were made [before PRO inviting the senator to the
hearing].
b. The painting was on the wall [in order PRO to check how it would be
received]. (Landau, 2017, p. 98)
But even Landau maintains something of the spirit of complementarity between
OC and NOC. He claims that NOC will only be possible if OC would make the
5 Culicover and Jackendoff (2001) note that for some adjuncts some semantic components may
be important, but their discussion is limited mainly to Rationale Clauses ([in order PRO to. . . ]).
See also Farkas (1988).
11
sentence denote a strange or unlikely situation. This distinction also does not hold
in all instances. I will argue that OC and NOC truly are non-complementary in
many adjuncts. NOC does not always require that OC be ruled out syntactically or
otherwise.
1.4 Organization of this dissertation
This dissertation explores the following questions: How do structural (OC)
and non-structural (NOC) referential dependencies interact? What is the repre-
sentation of OC and NOC? When the grammar allows both OC and NOC, how
do comprehenders select an interpretation? How quickly do comprehenders resolve
control dependencies in adjuncts? Is the processing of adjunct control distinct from
the processing of other forms of anaphora? These questions relate to broader issues
in the study of anaphora, how the processing of dependencies differs when there is an
overt versus an implicit cue to that dependency, whether syntactic and non-syntactic
anaphoric dependencies show distinct processing profiles, and the difference between
arguments and adjuncts in syntax and processing.
The dissertation is organized into two main parts. The remainder of Part I
focuses on the syntactic theory of adjunct control. Chapter 2 expands on Landau’s
(2017) arguments that temporal adjuncts allow both OC and NOC. I demonstrate
that several factors can influence the distribution of NOC, but none of these factors
on their own strictly determines when it will be available. For example, NOC is not
restricted to environments where OC would lead to strange interpretations, or to
12
adjuncts in active voice, contrary to claims made in Landau (2017), although each
of these can contribute to the likelihood of NOC. In Chapter 3, I provide evidence in
agreement with Landau’s (2017) argument that rationale adjuncts also allow both
OC and NOC, contrary to earlier theories claiming that they were either strictly
OC (e.g. Whelpton, 2002) or strictly NOC (e.g. A. Williams, 2015).
Chapter 4 evaluates the prediction implicit in Landau (2017) that all adjuncts
will similarly allow both OC and NOC. Although this prediction is confirmed in
several different adjuncts, others require either strict OC or strict NOC. I outline
the assumptions Landau’s theory would have to make about these adjuncts in order
to accommodate them, but argue that such assumptions are unfounded in some
cases. I propose an alternative theory couched within the MTC (Hornstein, 1999;
Boeckx, Hornstein, & Nunes, 2010) that is better able to handle most of the strict
OC adjuncts, although I note two cases where it struggles.
Part II is focused on the real-time processing of adjunct control. Chapter 5
reviews some of the previous literature, focusing on what types of questions pro-
cessing measures have been used to ask. In Chapter 6, I present two visual-world
eyetracking experiments investigating the timecourse of interpretation of PRO and
overt pronouns in temporal adjuncts. Results show that listeners are just as fast to
interpret PRO as an overt pronoun once the relevant bottom-up input is received.
This chapter serves as foundational work for future research on the interpretive bi-
ases involved in adjunct control. Its two main contributions to this line of research
are that it shows that the visual-world paradigm can be successfully used to track
the interpretation of PRO, and that it provides baseline evidence on how long it
13
takes to interpret PRO in the most basic constructions. This chapter also provides
evidence for the effect of structural predictions in the interpretation of PRO.
Chapter 7 concludes the dissertation with a summary and a look forward to
outstanding issues.
14
Chapter 2: Temporal adjuncts
Much of the evidence that has been used to classify adjunct control as OC
involves gerunds in temporal adjuncts such as the one in example (1) (Clark, 1990;
Hornstein, 1999, 2001; Pires, 2007), although it is far from clear that all instances
of adjunct control behave in the same way (Landau, 2013).
(1) Ron1 talked to Harry2 [before PRO1/*2 leaving the room].
In this chapter, I discuss temporal adjuncts as one construction that challenges the
traditional complementarity of OC and NOC. I will demonstrate, expanding on
previous evidence from Landau (2013, 2017) and others, that both OC and NOC
are allowed by the grammar for temporal adjuncts, the former illustrated in (2a)
and the latter in (2b). I argue that the choice for one or the other is influenced by
non-structural factors.
(2) a. The pizza1 tasted better [after PRO1 being heated in the microwave].
b. The pizza tasted better [after PRO drinking root beer].
After a detailed discussion of the distribution of non-subject control of tem-
15
poral adjuncts, I will propose that both OC and NOC are allowed by the grammar
for these adjuncts. OC will always involve control by the matrix subject, as it is
the only syntactically accessible potential antecedent PRO in temporal adjuncts.
Although NOC may in principle also include control by the subject, it will be illus-
trated by having PRO refer to an individual other than the referent of the matrix
subject in order to more clearly distinguish it from OC. I will also discuss evidence
that at least some instances of subject control in fact must be OC, and not just a
possible interpretation of PRO under NOC. I will argue that under NOC readings,
PRO is resolved to a logophoric perspective holder, and under OC to the poten-
tially non-logophoric referent of the matrix subject. Although both OC and NOC
are possible for temporal adjuncts, factors including potential performance biases
in comprehension lead to a preference for OC, with PRO referring to the subject of
the next higher clause.
The chapter is organized as follows. In §2.1, I outline some general properties
of temporal adjuncts. Then, in §2.2, I discuss why temporal adjuncts have generally
been classified as OC. In §2.3, I provide evidence on non-subject control of temporal
adjuncts, concluding that structural factors alone cannot predict its distribution.
Section 2.4 outlines details of the proposal that both OC and NOC are possible
for temporal adjuncts, as well as reasons for why OC is often strongly preferred.
Section 2.5 discusses the implications of this proposal for the data discussed and for
broader theories of control.
16
2.1 General properties of temporal adjuncts
Temporal adjuncts are those headed by prepositions1 such as before, after,
while, and when (Landau, 2013), as in (3). Adjuncts headed by without are also
sometimes included, and although they are not semantically temporal, I will also
include them here. Temporal adjuncts can appear sentence finally (3a), sentence
initially (3b), extraposed (as marked by an intonational break) (3c), or they can be
adjoined medially (3d), similar to the distribution of S-level adverbials (Jackendoff,
1972).
(3) a. Harry left the room [before/after/while/without talking to Ron].
b. [Before/after/while/without talking to Ron], Harry left the room.
c. Harry left the room—[before/after/while/without talking to Ron].
d. Harry, [before/after/while/without talking to Ron], left the room.
Temporal adjuncts modify the event time of the clause to which they attach,
with the exception of those headed by without, which describes how the event was
performed, not when it happened. Below the preposition heading the temporal
adjunct is a null operator that binds an empty category lower in the structure of the
adjunct (Larson, 1987; Williams, 1992). According to Haegeman (2010), the moving
operator is the preposition itself, as illustrated in (4). Whether the operator is the
preposition or a separate null element, where it moves from can result in ambiguity
of interpretation ((4a) vs. (4b)).
1 Or possibly complementizers (Hornstein, 1999).
17
(4) a. Harry left [after1 [Ron said [he would leave t1]]].
‘Harry left after the time of departure predicted by Ron.’
b. Harry left [after1 [Ron said [he would leave] t1]].
‘Harry left after Ron’s statement about his departure.’
2.2 Temporal adjuncts classified as OC
Temporal adjuncts have been classified as strictly OC in both theoretical
(Clark, 1990; Hornstein, 1999, 2001; Pires, 2007; Boeckx et al., 2010) and psycholin-
guistic (Parker, Lago, & Phillips, 2015; Gerard, 2016) literature. Several reasons
have been given to justify this classification. First, PRO in temporal adjuncts can
be controlled by the local subject of the matrix clause, but not the object (5a)
(Hornstein, 1999; Boeckx et al., 2010), non-c-commanding matrix arguments (5b)
(Pires, 2007; Boeckx et al., 2010), or non-local arguments (5c) (Clark, 1990; Pires,
2007; Boeckx et al., 2010).
(5) a. Harry1 saw Ron2 [before/while/after PRO1/*2 leaving the room].
b. *Ron1’s sister left [without PRO1 having shaved himself].
c. Harry1 said [that [Ginny2’s brother]3 left [after PRO3/*1/*2 eating a
bagel]].2
Second, because the reflexive oneself is not licensed in (6a), Clark (1990) used the
2 Note that Harry is only precluded from being the antecedent in (5c) when the adjunct adjoins
to the lower clause, i.e. when what Harry reported was that Ginny’s brother left after eating a
bagel. If the adjunct attaches high, i.e. if Ginny’s brother left, and Harry reported that after eating
a bagel, then Harry must be the antecedent.
18
example to argue that PRO in temporal adjuncts cannot have an arbitrary interpre-
tation (compare to (6b), where arbitrary interpretation and oneself are possible).
(6) a. Ron1 felt old after [PRO1/*arb seeing himself/*oneself in the mirror].
b. Ron1 said that [[PRO1/arb to hurt himself1/oneself] would be a bad idea].
Third, PRO in temporal adjuncts only allows a “sloppy” interpretation (Ross, 1967,
1969) under ellipsis (Pires, 2007; Boeckx et al., 2010). The elided PRO in the
second conjunct of (7a) does not receive an identical interpretation to PRO in the
first conjunct (i.e. it does not refer to Ron), but instead must be bound by the local
subject (Harry). If PRO were not locally bound, but instead was referring to the
referent of the first subject as a free pronoun, this would be unexpected (compare
(7a) to (7b)).
(7) a. Ron1 left [before PRO1 dancing], and Harry did too.
‘And Harry2 left before he2/*Ron danced.’
b. Ron1 left [before he1/2 danced], and Harry did too.
‘And Harry2 left before he1/2/3 danced.’
Fourth, PRO can only have a bound interpretation when controlled by only-DPs.
(8) Only Harry1 survived [after PRO1 being hit with the killing curse].
‘[Nobody else]1 survived after he1/*Harry was hit with the killing curse.’
19
Fifth, Pires (2007) argues that because PRO in (9) does not allow split antecedents,
it must be obligatorily controlled by the matrix subject alone.
(9) *Peter1 expected Susan2 to break up with him [without PRO1+2 hurting them-
selves/each other].
Finally, it has been argued based on examples such as (10a) that NOC requires a
[+human] interpretation (Clark, 1990; Landau, 2013).3 Because PRO in temporal
adjuncts can take an inanimate antecedent (10b–c), it must not be NOC, at least
in those cases.4
(10) a. *[PRO Raining] would be a disaster right now.
b. It1 never snows [after PRO1 raining in this part of the world].
c. The dishwasher1 finally broke [after PRO1 working well for many years].
Under the assumption that OC and NOC are grammatically complementary,
3 Landau (2001, 2017) explains this restriction by assuming that NOC is logophoric in nature.
Because only humans can be logophoric centers, it follows that NOC would require a [+human]
antecedent. A similar effect is discussed by Charnavel and Sportiche (2016), who argue that
animate reflexives may sometimes actually be logophors, while inanimate reflexives are always
true anaphors subject to Condition A of the binding theory. Logophoricity effects in temporal
adjunct control will be discussed in §2.3.2.4.
4 This argument only follows, of course, if the premise is accepted that NOC must be [+human].
Norbert Hornstein (p.c.) gives the following example with apparent remote control (see §3.1.2) of
an inanimate PRO.
(i) Jon framed the portrait1 in a black frame. [PRO1 Being framed in a light color] dulled its1
natural colors.
But it is not clear that is actually NOC. Hornstein and Kiguchi (2003) and Boeckx and Hornstein
(2007) argue that (ii) involves OC (see §4.3). Perhaps PRO in (i) is also in an OC dependency
with its.
(ii) John said that [[PRO1 washing herself1] delighted Mary1]
20
if these temporal adjuncts have properties of OC, then all adjuncts with the same
syntactic structure would also have to be OC. However, there are cases where PRO
appears to refer to a non-local or non-c-commanding antecedent. A few examples
from previous literature are given in (11). In each of these, PRO is understood as
referring to someone other than the subject of the next higher clause. If control
of temporal adjuncts were always OC, this would be unexpected, as the controller
should always be the matrix subject, being the only local and syntactically accessible
antecedent, as was seen in the examples above.
(11) a. John1’s fears always go out of control, [when PRO1 listening to Larouche].
(Williams, 1992, p. 307)
b. Potatoes are tastier [after PRO boiling them]
(Ackema & Schoorlemmer, 1995, p. 182)
c. That oasis was a vision [after PRO dragging ourselves through the
desert all day]. (Landau, 2017, p. 7)
d. The stairs were washed [before PRO entering the basement].
(Landau, 2013, p. 228)
e. The president was elected [without PRO considering his competence].
(Roeper, 1987, p. 297)
f. There will be no progress [without PRO investing economic and human
resources] (Landau, 2013, p. 232)
g. [After PRO collecting some money], a bank account was opened.
(Kawasaki, 1993, p. 165)
21
h. [Before PRO entering the basement], the stairs were washed.
(Landau, 2000, p. 22)
i. [After PRO pitching the tents], darkness fell quickly.
(Kawasaki, 1993, p. 173)
One might argue, as did Roeper (1987) for (11e), that these types of examples
still constitute OC. The matrix clauses in (b-e,g-h), for instance, all arguably con-
tain some sort of implicit argument that could be serving as an obligatory controller
for PRO: an implicit agent in (d,e,g,h) and an implicit experiencer in (b,c). How-
ever, this cannot be the case in (a,f,i). In addition, one might argue that because
the adjuncts in (g-i) are fronted, their apparent NOC should not be compared with
the OC behavior of right-adjoined adjuncts. In the next sections, however, I will
demonstrate why any account that relies on implicit argumenthood or on struc-
tural differences between the examples in (11) and the OC examples above cannot
adequately predict when NOC readings of temporal adjuncts will be available.
2.3 The distribution of NOC in temporal adjuncts
The accounts above that classify control of temporal adjuncts as OC and
that assume grammatical complementarity between OC and NOC predict that non-
subject control should be available only if there is some structural feature that
rules out the normal OC reading. This section explores the distribution of non-
subject control of temporal adjuncts and demonstrates that it cannot be reduced
22
to structural properties alone. Instead, discourse factors, especially logophoricity,
affect when NOC is possible.
2.3.1 Structural effects
I will argue that structural features alone cannot delineate the availability of
OC by the matrix subject and NOC by non-subjects, although they can have an
affect on how PRO in temporal adjuncts is interpreted. Here I discuss the effect of
the adjunction site, of structural properties of the matrix clause, and of the structure
of the adjunct.
2.3.1.1 Attachment site of the adjunct
Several authors have noted that the attachment site of a temporal adjunct
affects the likelihood of its receiving an OC or NOC interpretation. For example, an
adjunct is more likely to have an NOC reading when it is adjoined to the left than
the right (Williams, 1992; Lyngfelt, 2009a, 2009b; Landau, 2000, 2013). Although
a non-subject NOC reading is available for each of the preposed adjuncts in (12),
when the adjunct is not preposed (and when there is no intonational break before
the adjunct), only subject control appears to be available (13).
(12) a. [After PROarb pitching the tents], darkness fell quickly.
(Kawasaki, 1993, p. 173)
b. [Before PRO1/arb entering the basement], the rain1 washed the stairs.
(Landau, 2000, p. 22)
23
c. [After PRO1/2 being sick in bed all day], Ginny1 brought me2 a nice
surprise.
d. [After PROarb bewitching his broom], Harry1 broke his arm.
(13) a. #Darkness1 fell quickly [after PRO1/*2 pitching the tents].
(Landau, 2000, p. 178)
b. The rain1 washed the stairs [before PRO1/*2 entering the basement].
(Landau, 2017, p. 4)
c. Ginny1 brought me2 a nice surprise [after PRO1/*2 being sick in bed all
day].
d. Harry1 broke his arm [after PRO1/*2 bewitching his broom].
Landau (2000) and Lyngfelt (2001, 2009b) both suggest that the restrictions
on interpretation seen in examples such as (13) as compared to (12) are due to
configurational properties of preposed adjuncts. They suggest that clause-initial
adjuncts are attached higher than clause-final ones, and that they lie outside the
syntactic domain of the matrix subject. If OC requires c-command, as is generally
assumed (Williams, 1980; Chomsky, 1981; Landau, 2000, 2015), then the matrix
subject would not able to bind PRO in a high clause-initial adjunct. Similarly,
when the adjunct is extraposed, NOC is more available (Williams, 1992; Landau,
2013).
(14) a. Darkness1 fell quickly—after PRO1/arb pitching the tents.
24
b. The rain1 washed the stairs—before PRO1/?arb entering the basement
(Landau, 2013, p. 252)
c. Ginny1 brought me2 a nice surprise—after PRO1/2 being sick in bed all
day.
d. Harry1 broke his arm—after PRO1/arb bewitching his broom.
This data suggests a first attempt to reduce non-subject control of temporal
adjuncts to something structural. Perhaps NOC in temporal adjuncts is available if
and only if the adjunct is attached outside the c-command domain of the subject.
This explanation would allow us to keep structural complementarity between OC
and NOC. However, examples that counter this explanation can easily be found.
Landau (2013) notes, in fact, that the preference for clause-initial temporal adjuncts
to be NOC and for clause-final ones to be OC is just that: a preference. Again under
the assumption that NOC requires a human antecedent, the examples in (15) and
(16) show that both sentence initial and extraposed adjuncts allow OC.
(15) a. [After PRO1 causing a lot of trouble], the dishwasher1 finally broke
down. (Kawasaki, 1993, p. 31)
b. [After PRO1 snowing a lot in February], it1 never rains heavily in March
in this area. (Kawasaki, 1993, p. 31)
(16) a. The dishwasher1 finally broke down—after PRO1 causing a lot of trou-
ble.
25
b. It1 never rains heavily in March in this area—after PRO1 snowing a lot
in February.
Perhaps in these cases, the adjunct reconstructs to a position c-commanded
by the matrix subject, and it is in fact in a location amenable to OC. If so, one
could still argue that NOC only obtains when the adjunct is not c-commanded
by the subject. However, even with a low, local adjunct, NOC is still sometimes
possible. The most salient reading of (17), indeed the only reading possible if the
pronoun them is taken to refer to the potatoes, is one in which PRO has an arbitrary
interpretation; PRO cannot corefer with the matrix clause subject.
(17) Potatoes are tastier [after PRO boiling them1].
(Ackema & Schoorlemmer, 1995, p. 182)
Still, one could argue that in (17) it is still the case that the adjunct is attached
higher than the matrix subject, and it is this that allows NOC. One final piece
of evidence that attachment site does not strictly determine control type is (18).
Here the adjunct is attached medially, which arguably must be below the matrix
subject. Yet even in this case, OC is possible in both (18a,b), again assuming that
NOC requires a [+human] antecedent, and that an inanimate controller is therefore
indicative of OC.
(18) a. Lemons1, [after PRO1 being left on the counter for too long], tend to
go moldy.
26
b. Lemons, [after PROarb eating sugar], taste especially sour.
Because both OC and NOC are possible in at least some temporal adjuncts,
regardless of where they are attached, the distribution of obligatory subject control
and NOC by some other entity cannot be reduced to this structural feature. With
respect to attachment site, OC and NOC temporal adjuncts appear in overlapping
distribution, which means that other factors (as will be discussed below) determine
which reading will be most salient. This is not to say that attachment site does not
play a role; as several authors have observed, temporal adjuncts adjoined clause-
finally are less likely to allow NOC readings. But, I will argue, this is not due to
syntactic requirements, but instead to performance biases. This will be discussed in
more detail in §2.4.
2.3.1.2 Structural properties of the matrix clause
Since attachment site cannot adequately predict the availability of OC and
NOC in temporal adjuncts, another place to look for a structural distinction between
the two may be in the properties of the matrix clause. Lyngfelt (2002), as cited in
Lyngfelt (2009a), suggests that NOC is more likely to be available when the matrix
clause is either a passive (19a) or is headed by an expletive (19b).5 This is, he
5 A reviewer suggests that in (19b) the adjunct is actually adjoined to the embedded clause,
and its PRO is controlled by a higher PRO, as in (i). However, it seems at least possible to me for
the temporal adjunct to modify the matrix clause, as in (ii), with the meaning that it was hard
for a while after finishing the book to eat meat. And even in this case, PRO may be understood
to refer to the speaker, which interpretation can only be NOC.
(i) It was hard [PRO1 to eat any kind of meat for a while [after PRO1 finishing this book]].
(ii) It was hard [PRO to eat any kind of meat] for a while [after PRO finishing this book].
27
claims, especially true for sentence-final adjuncts; in all cases of non-subject control
of clause-final adjuncts found in a corpus study of Swedish infinitives, the matrix
clause had at least one of these two properties. I will examine the effect of the
passive here, and the effect of expletive subjects in §2.3.2.
(19) a. The president was elected [without PRO considering his competence].
(Roeper, 1987, p. 297)
b. It was hard (for me1) to eat any type of meat for a while [after PRO1
finishing this book]. (Modified from Lyngfelt, 2009a, p. 39)
The most natural interpretation of (19a) is that the people who elected the president
are the ones who did not consider his competence. Therefore, as mentioned above,
it has been argued that (19a) does not actually involve NOC, but that PRO is
controlled by an unexpressed but syntactically present agent of the matrix clause.
This explanation could also apply to each of the examples in (20), which allow PRO
to refer to the understood matrix agent.
(20) a. The bank was robbed [while PRO masquerading as security guards].
(Landau, 2017, p. 6)
b. All preparations were made [before PRO inviting the senator to the
hearing]. (Landau, 2017, p. 6)
c. All the preparations were made [before PRO arriving to court].
(Landau, 2017, p. 7)
d. The pizza was eaten [after PRO ordering it].
28
e. This dissertation was written [without PRO drinking any coffee].
f. The war was won [without PRO firing a shot].
g. The ship was sunk [before PRO collecting the insurance].
Furthermore, this explanation could also potentially extend to any case of non-
subject control of a temporal adjunct where the matrix clause involves some sort
of implicit argument. Ackema and Schoorlemmer (1995) claim, for example, that
(21a) involves OC by the implicit taster. Similarly, (21b) could involve control by an
implicit experiencer, since the most salient and perhaps only possible interpretation
is that the oasis was a vision to the same people who dragged themselves through
the desert.
(21) a. Potatoes1 are tastier [after PRO boiling them1].
(Ackema & Schoorlemmer, 1995, p. 182)
b. That oasis was a vision [after PRO dragging ourselves through the
desert all day.] (Landau, 2017, p. 7)
However, there are at least three challenges for this explanation. First, Landau
(2017), following Manzini (1986) and Kawasaki (1993), has argued that control by
the implicit agent of a passive is NOC based on the examples in (22) and (23). In
(22c), if the hitter is understood to be the one rolling down the hill (an interpretation
that already is marginal), then that hitter must be human. This is true even though
an inanimate argument can control PRO when it is the subject of an active matrix
clause (22a), or can be the passive agent without control (22b). The same is true
29
for (23). If control of PRO in a temporal adjunct by a passive agent was due to
syntactic binding, a [+human] requirement would be surprising.
(22) a. The avalanche1 hit the house [before PRO1 rolling down the hill].
b. The house was hit.
c. Mary said the house was hit [before PRO rolling down the hill].
(23) a. The rain1 washed the stairs [before PRO1 entering the basement].
b. The stairs were washed.
c. The stairs were washed [before PRO entering the basement].
Second, there are cases of non-subject control where PRO can be interpreted
as someone other than the agent of the passive matrix clause or other implicit
arguments, as is the case for the examples in (23c) and (24). Example (23c) can
easily mean that some person entered the basement after some other person or
thing washed the stairs. This is especially true if we understand the sentence as
representing the perspective of the person entering the basement. Example (24a)
would be naturally uttered in a situation where sailors undocked the ship, and then
were surprised by someone else sinking it. Example (24b) is a more subtle case.
Although it does seem to mean that the winners of the war did not fire a shot,
it would sound false in a situation where the losers fired lots of shots, even if the
winners fired none. What (24b) implies, then, is that the war was won without
anyone firing a shot; in this case, PRO does not refer exclusively to the understood
agent of the matrix verb. In (24c), PRO still refers to some person boiling the
30
potatoes, even though that person cannot be represented as an implicit experiencer,
in contrast to (21a). But even (21a) does not require the boiler and the taster to be
the same individual (Landau, 2013).
(24) a. The ship was sunk [immediately after PRO undocking it].
b. The war was won [without PRO firing a shot].
c. The potatoes1 got softer [after PROarb boiling them1].
This point is even more clear for the passive cases when the agent is represented
in a by-phrase, removing the possibility of an implicit agent (Lasnik, 1988). In each
of the examples in (25), PRO still can refer to some person that no argument in
the matrix clause refers to. Similarly, in (26), PRO can still refer to the chef who
prepared the food, even though the tasters are his guests.
(25) a. The stairs were washed by the rain [before PRO entering the basement].
b. The ship was sunk by enemy torpedoes [immediately after PRO un-
docking it].
c. The ship was sunk by a hired crook [before PRO collecting the insur-
ance].
(26) The potatoes tasted better to the guests [after PRO boiling them].
Finally, having a passive matrix clause does not guarantee that control by the
unexpressed agent will be possible. In each of the examples in (27), the only possible
reading seems to involve control by the matrix subject. Interpretations where PRO
31
refers to any other entity, including the matrix agent, are unavailable. Combined
with the previous two points, this suggests that other factors, not just the presence
of an implicit argument, are what is important.6
(27) a. My wallet1 was stolen [while PRO1/*2 distracting me].
(McCourt, Green, Lau, & Williams, 2015, p. 3)
b. The tents1 were set up [after PRO1/*2 being locked in the trunk all
day].
c. The bank1 was robbed [while PRO1/*2 being caught on camera].
(Landau, 2017, p. 7)
d. #The bank1 was robbed [while PRO1/*2 being smart not to raise the
alarm]. (Landau, 2017, p. 7)
Non-subject control in temporal adjuncts cannot be reduced to OC by an
implicit argument. Their presence or absence cannot account for the availability
of non-subject control. It is also not the case that the passive transformation itself
blocks OC and thus allows NOC. The examples in (27) clearly demonstrate that OC
is not only possible with a passive matrix clause, but is at least sometimes required.
Why, then, did Lyngfelt (2002, 2009a) find NOC to be more likely with a
passive matrix clause, if passives themselves do not block OC? Additional evidence
can potentially be found in other constructions that share properties of the passive.
For example, perhaps something about having a derived matrix subject makes NOC
6 Landau (2017) states that one of those factors is the voice of the adjunct, claiming that passive
or similar adjuncts do not allow non-subject control. That hypothesis would apply to (b-d), but
not (a). I will return to this idea in §2.3.1.3
32
more likely. If that were the case, then we would expect not only passives, but all
matrix clauses with derived subjects to allow NOC of temporal adjuncts adjoined to
them. This prediction seems to be supported by comparing the availability of non-
subject control in unaccusatives, which have a derived subject, and unergatives,
which do not (Levin & Rappaport Hovav, 1995). Each of the examples in (28)
contains an unaccusative predicate in the matrix clause, and NOC is available. This
is especially evident for (a-d), in which it is easy to interpret the pronoun in the
embedded clause as coreferential with the subject of the matrix clause. If PRO
is controlled by the matrix subject, this would result in a Condition B violation
(Chomsky, 1981). But NOC is also possible in (e-f), in which control by the matrix
subject would not result in a binding condition violation. The examples in (29), on
the other hand, have unergatives in the matrix clause, and the only possible reading
is OC.
(28) Unaccusatives:
a. The clock broke [after PRO dropping it].
b. The cheese melted [after PRO putting it in the microwave].
c. The ship sank [after PRO hitting it with a torpedo].
d. Aladdin sank to the bottom of the sea [after PRO weighing him down
with a cannon ball].
e. The glass broke [after PRO singing the right high note].
f. The box closed [after PRO letting go of the lid].
33
(29) Unergatives:
a. *Peter1 dined at the café every morning [after PRO pointing it out to
him1].
b. *Mary1 resigned [after PRO finding the scandal about her1].
c. *The shepherds1 rejoiced [after PRO telling them1 about Jesus].
However, the difference in acceptability of the examples in (28) and (29) may
be confounded by other features of the subject. Aside from (28d), the examples
in (28) all have non-human subjects and (29) human ones. Indeed, unergatives in
general often have animate or human subjects and unaccusatives often have inani-
mate, non-human ones. This confound is highlighted by the contrasts in (30), which
contains unaccusative main-clause predicates, and (31), which has unergatives. In
both cases, non-subject control is more likely with an inanimate subject than a hu-
man one. For the unergative in (31) specifically, when the baby is referred to with
a gendered pronoun in (31a), non-subject control is not available. But the NOC
reading becomes much better when the baby is referred to with the neuter [-human]
pronoun in (31b). The effect of humanness will be discussed more in §2.3.2.
(30) a. The clock1 fell down [after PRO kicking it1/the table].
b. Joseph1 fell down [after PRO1/*2 kicking *him1/the table].
(31) a. ??The baby1 slept well [after PRO feeding her1].
b. The baby1 slept well [after PRO feeding it1].
34
There are many other cases of OC in temporal adjuncts where the matrix
clause has a derived subject. The examples in (32) contain passive matrix clauses
and allow OC by the derived subject, and (33) has additional unaccusatives with
the same property. That this is in fact OC can be diagnosed by the inanimacy of
the controlling subject. These examples demonstrate that the cases of NOC above
cannot be due to OC being unavailable, since having a derived subject does not
block OC.
(32) Passives:
a. The tomatoes1 were eaten [only after PRO1 fermenting/being pickled
in brine for a few days.]
b. ?My phone1 was stolen [while PRO1 sitting in my locker].
c. The cake1 was destroyed [after PRO1 being sat on].
(33) Unaccusatives:
a. The ship1 sank [after PRO1 running into an iceberg].
b. The snow1 thankfully stopped [after PRO1 falling non-stop for a week].
c. The glass1 broke [after PRO1 falling off the table].
d. The train1 departed [after PRO1 filling up with passengers].
e. The ball1 dropped [after PRO1 hanging in the air for 10 seconds].
Finally, not only is having a derived subject in the matrix clause not sufficient
for allowing NOC, it is also not necessary. Examples (34) and (35) provide exam-
35
ples of transitives and unergatives, which do not have derived subjects, but which
nevertheless allow NOC.
(34) Transitives:
a. My arrow finally hit the target [after PRO practicing for hours and
hours].
b. The rain ruined my bike [after PRO leaving it outside overnight].
(35) Unergatives:
a. The clock worked well [after PRO putting in new batteries].
b. Peter talked for hours [after PRO giving him the microphone].
c. Strangely the candidates talked avidly when we asked them where they
were from, but they hesitated [after PRO asking them about their
work].
d. The doorbell buzzed [after PRO pushing a little button].
Both implicit argumenthood and derived subjecthood fail to sufficiently delin-
eate the distribution of OC and NOC in temporal adjuncts. Note that an interaction
between the structure of the adjunct and its attachment site is also insufficient. As
was demonstrated in §2.3.1.1, right-adjoined adjuncts with no prosodic break were
the least likely to allow NOC. Each of the examples in this section, both those that
allowed NOC and those that did not, used right-adjoined adjuncts. Therefore, we
cannot even form a disjunctive generalization stating that NOC is only allowed if the
36
adjunct is attached high or the matrix clause has an implicit argument or derived
subject.
2.3.1.3 Structural properties of the temporal adjunct
We have just seen that the structure of the matrix clause cannot sufficiently
determine the distribution of OC and NOC in temporal adjuncts. It has also been
proposed that the structure of the adjunct itself plays a role. Landau (2017, pp. 6-7)
provides evidence suggesting that when the adjunct clause is in the passive voice,
as with the (b) and (c) examples in (36)–(38), only OC by the matrix subject is
possible. NOC is blocked, even though it is available in similar sentences with active
adjuncts, as seen in the (a) examples.
(36) a. All preparations were made [before PRO inviting the senator to the
hearing].
b. The senator1 was warned [before PRO1 being invited to the hearing].
c. *All preparations were made [before PRO being invited to the hearing].
(37) a. That oasis was a vision [after PRO dragging ourselves through the
desert all day].
b. We1 finally found water [after PRO1 being dragged through the desert
all day].
c. *That oasis was a vision [after PRO being dragged through the desert
all day].
37
(38) a. There won’t be any progress [without PRO insisting on guidance from
the outside].
b. They1 will make no progress [without PRO1 being guided from the
outside].
c. *There won’t be any progress [without PRO being guided from the out-
side].
Landau notes that the generalization that passive adjuncts only allow OC also
seems to be true of copular adjuncts (39), but not of unaccusatives (40). Only the
latter allows both OC and NOC. Although he does not provide a an explanation of
this generalization, he hints that it might be due to the syntactic structure of the
adjunct, suggesting that perhaps the generalization does not apply to unaccusatives
because they can sometimes have agentive readings and perhaps agentive syntax.
(39) a. The robbers1 filled their bags [while PRO1 being smart not to raise the
alarm].
b. *The bank was robbed [while PRO being smart not to raise the alarm].
(40) a. My lawyer1 made all the preparations [before PRO1 arriving to court].
b. All the preparations were made [before PRO arriving to court].
Again, this generalization is insufficient. Although it does appear that having
a passive or copular adjunct makes it less likely that NOC will occur, this is not a
hard restriction. Example (41a), which is quite similar to (37c), allows a logophoric
38
interpretation of PRO, despite its being the subject of a passive adjunct. The same
is true for (41b), in which PRO can easily refer to the speaker (suppose the speaker
dislikes mushrooms, but hates celery even more), and for (41c), which allows PRO
to refer to a non-c-commanding argument embedded within the matrix subject.
Similarly, Example (41d), which contains a copular clause in the adjunct, also allows
non-subject control, at least for some speakers.
(41) a. That Gatorade was a vision [after PRO being worked to exhaustion by
my coach today].
b. Mushrooms aren’t too bad [after PRO being forced to eat celery].
c. ?The prisoner1’s sentence was finally released [after PRO1 being forced
to wait for over week while the judge made his decision].
d. ?The house stayed dry during the storm [after PRO being smart not to
leave the windows open].
Even if it were true that passive adjuncts only allowed OC, that would still not
be sufficient to determine the full distribution of OC and NOC. Something more
would still need to be said about adjuncts in active voice, which allow both types
of control. It is of course possible that OC and NOC adjuncts differ structurally
due to inherent syntactic differences in derivation of OC and NOC themselves, as is
the case in the Two-tiered Theory of Control (TTC) (Landau, 2017, see §4.1). But
other structural features of the adjunct cannot determine whether OC or NOC will
be possible. That the distribution of OC and NOC in temporal adjuncts cannot
39
be reduced to structural properties of either the embedded or matrix clause is best
illustrated by examples such as the following. Each of the examples in (42), (43),
and (44b) can be interpreted as either OC by the local subject or NOC by a non-
local antecedent without changing the structure of the sentence at all (again, modulo
changes in the representation of OC and NOC themselves).
(42) Chickens1 taste better [after PRO1/2 being forced to eat celery].
(43) The pool1 was the perfect temperature [after PRO1/2 being in the hot sun
all day].
(44) a. (i) At some point yesterday, the dishwasher broke. Also yesterday,
Peter came home for a 2-week visit. Mom thought that Peter
broke the dishwasher.
(ii) Peter ordered a new dishwasher, and the store had it delivered.
It is not working, and the store manager doesn’t want to replace
it. He said that Peter must have broken it.
b. Peter1 says [the dishwasher2 broke [before PRO1/2 even arriving at the
house]].7
2.3.2 Non-structural effects
We have just seen that structural properties can influence whether NOC is
more likely to be available, but they cannot be used to fully delineate the distribu-
7 This sentence has a potential structural ambiguity similar to the one in (5c), discussed in
Footnote 2. The key is that PRO may refer to Peter even when the adjunct attaches low. That
is, PRO may refer to Peter even if Peter is saying this long after he arrived.
40
tion of OC and NOC in temporal adjuncts. This section examines non-structural
properties of the sentences and discourses in which these temporal adjuncts occur
and the effect they have on control.
2.3.2.1 Availability of subject-control reading
At least three authors (Williams, 1992; Lyngfelt, 2001, 2009a, 2009b; Landau,
2017) have noted that non-subject control into adjuncts is more available when
subject control would make the sentence denote a bizarre or impossible situation, or
in the terms of Landau (2017), a semantic anomaly. In (45a), for example, John’s
fears, being inanimate, cannot listen to anything. And in (45b), it seems odd for
stairs to enter the basement after being washed. Williams (1992) and Landau (2017)
argue that NOC interpretations of these sentences are only possible because of the
anomaly that OC would have produced. Although not relying on a grammatical
distinction between the two, this maintains a spirit of OC/NOC complementarity;
NOC is only available because OC is not.
(45) a. John1’s fears2 always go out of control, [when PRO1/#2 listening to
Larouche]. (Williams, 1992, p. 307)
b. The stairs1 were washed [before PRO#1 entering the basement].
(Landau, 2013, p. 228)
The same is true of the examples from Lyngfelt in (46). Although these have
the confounding attribute that the adjunct is fronted, which we have already seen
41
makes NOC more likely, the same NOC reading seems to persist even when the
adjuncts are moved to the right in (47).
(46) a. [After PRO1/#2 finishing her1 studies], an overseas job2 was a welcome
change. (Lyngfelt, 2009a, p. 39)
b. [After PRO#1 flying from Europe to California], the jetlag1 was tough,
but not unbearable. (Lyngfelt, 2009b, p. 155)
(47) a. An overseas job2 was a welcome change [after PRO1/#2 finishing her1
studies].
b. The jetlag1 was tough [after PRO#1 flying from Europe to California],
but not unbearable.
Lyngfelt (2009a) makes the similar observation that NOC is more likely when
the matrix clause has an expletive subject, as in (48). For the expletive to control
PRO in these examples would require it to take the external θ-role of the embedded
verb. Since expletives cannot hold any thematic relation, this would be impossible.
When OC is blocked for this reason, NOC becomes available.
(48) a. It was hard (for me1) to eat any type of meat for a while [after PRO1
finishing this book]. (Modified from Lyngfelt, 2009a, p. 39)
b. There1 will be no progress [without PRO*1 investing economic and
human resources] (Landau, 2013, p. 232)
This may explain why non-subject control is more available when the matrix
42
clause is a passive or unaccusative, as noted above. Because passive and unac-
cusative subjects are often inanimate, if the verb in the adjunct requires an animate
subject, OC by the matrix subject would be ruled out. In the examples in (49)
and (50), repeated from (20) and (28), the external θ-role of the embedded verb
requires an animate argument. Because the matrix subject is inanimate, OC by it
is semantically anomalous, and NOC becomes more readily available.
(49) a. The bank was robbed [while PRO masquerading as security guards].
(Landau, 2017, p. 6)
b. All preparations were made [before PRO inviting the senator to the
hearing]. (Landau, 2017, p. 6)
c. This dissertation was written [without PRO drinking any coffee].
d. The ship was sunk [before PRO collecting the insurance].
(50) a. The clock broke [after PRO dropping it].
b. The glass broke [after PRO singing the right high note].
c. The cheese melted [after PRO putting it in the microwave].
d. The ship sank [after PRO hitting it with a torpedo].
e. The box closed [after PRO letting go of the lid].
Subject control is further discouraged in many of the examples seen here by
the binding conditions (Chomsky, 1981). In (51a), for example, subject control
would induce a Condition A violation at the reflexive ourselves. Similarly, in (51b),
43
if the pronoun them in the embedded clause is understood to refer to the potatoes,
then subject control of PRO would cause a Condition B violation.
(51) a. That oasis1 was a vision [after PRO*1 dragging ourselves through the
desert all day]. (Landau, 2017, p. 7)
b. Potatoes1 are tastier [after PRO*1 boiling them1]
(Ackema & Schoorlemmer, 1995, p. 182)
However, although I have used these violations to more clearly demonstrate that
non-subject control is taking place, they are not necessary. Non-subject control is
just as possible in the similar examples in (52), in which subject control would not
cause binding condition issues, but would still lead to semantic anomaly.
(52) a. The potatoes1 were tastier [after PRO#1 eating carrots].
b. That oasis1 was a vision [after PRO#1 dragging our luggage through
the desert all day].
But once again, although the unavailability of subject control may contribute
to the possibility of non-subject control, it is neither necessary, as Landau (2017)
claims, nor sufficient. Several of the examples above, some of which are repeated in
(53), only allow subject control, even though this reading leads to semantic anomaly
or binding violations. This is further illustrated by (54)
(53) a. *Ron1’s sister left [without PRO1 having shaved himself].
44
b. Darkness1 fell quickly [after PRO1/*2 pitching the tents].
(Landau, 2000, p. 178)
c. The bank1 was robbed [while PRO1/*2 being smart not to raise the
alarm]. (Landau, 2017, p. 7)
(54) The toy1 broke [before PRO1/*2 feeling very sad].
That semantic anomaly of the subject-control reading is not required is demon-
strated by the examples repeated in (55). In each of these sentences, the subject of
the matrix clause makes a perfect, non-anomalous OC controller, and yet PRO can
also be understood to refer to someone else through NOC.
(55) a. The president1 was elected [without PRO1/2 considering his compe-
tence].
(Roeper, 1987, p. 297)
b. Chickens1 taste better [after PRO1/2 being forced to eat celery].
c. The pool1 was the perfect temperature [after PRO1/2 being in the hot
sun all day].
d. At some point yesterday, the dishwasher broke. Also yesterday, Peter
came home for a 2-week visit. Mom thought that Peter broke the
dishwasher.
Peter1 says [the dishwasher2 broke [before PRO1/2 even arriving at the
house]].
45
2.3.2.2 “Initiator” status
As discussed above, Landau (2017) claimed that NOC in adjunct control is only
available if the adjunct is in active voice and if OC would lead to “semantic anomaly.”
We have seen that neither of these generalizations is entirely correct. In a recent talk,
Landau (2018) attempted to reduce these effects to a single generalization: NOC
is only available if PRO and its potential OC controller mismatch in [init] features,
where an argument is [+init] if it is the “initiator” who intentionally brings about a
situation (see Farkas, 1988). In (56), for example, NOC is only possible because the
potential OC controller (the matrix subject) and PRO mismatch in [init] features.
(56) The president[-init] was elected [without PRO[+init] considering his compe-
tence].
In (57), repeated from (54), on the other hand, both the potential OC controller
and PRO are [-init]. Because these values match, only OC is possible, despite the
fact that OC results in the sentence denoting an unlikely situation. Landau’s (2018)
account thus fares better than Landau’s (2017) semantic anomaly generalization
described in the previous section.
(57) #The toy[-init] broke [before PRO[-init] feeling very sad].
However, the [init] account is still insufficient. In the sentences in (58), for
46
example, PRO and its potential OC controller mismatch in [init] values, and yet
NOC still appears to be possible.
(58) a. Mushrooms[-init] aren’t too bad [after PRO[-init] being forced to eat a
whole plate of celery].
b. The pool[-init] was the perfect temperature [after PRO[-init] being in the
hot sun all day].
2.3.2.3 Topichood
Kawasaki (1993) argued that NOC of PRO in adjuncts is really control by the
topic of the sentence.8 This means that non-subject control into adjuncts should
be available just in those cases where the sentence can be understood as having
as its topic someone other than its subject. Kawasaki’s account is situated in the
framework of File Change Semantics (Heim, 1983). When PRO in an adjunct refers
to someone other than the matrix subject, the sentence is understood to be making a
comment on that individual, and the proposition expressed by the sentence is stored
on their file card. The adjunct is understood to be a predicate of the topic under
which it is filed. For example, the sentence in (59) is in some sense about the person
working on the problem sets and the ease with which they finish them. Because of
8 This is similar to aspects of Centering Theory (Grosz, Weinstein, & Joshi, 1995), which argues
that if anything is going to be referred to with a pronoun, then it will be the “Backward Looking
Center”, which is the discourse referent evoked by an utterance that is the most “highly ranked”
of those also evoked by the prior utterance, according to a ranking of accessibility to anaphora.
Thus the Backward Looking Center is a kind of topic, in that it is the referent most likely to have
been continued from the prior utterance.
47
that, the proposition is filed on the file card corresponding to that person, and PRO
can refer to them instead of to Problem Set 2, the referent of the matrix subject.
(59) [After PRO1 working on Problem Set 1], Problem Set 2 shouldn’t take (you1)
much time. (Kawasaki, 1993, p. 163)
Kawasaki bases her claim on three main observations. First, PRO in adjuncts
never introduces a new referent in the discourse. Instead, it often is coreferential
with the DP subject of either the current or previous sentence. This is expected
if the adjunct is a predicate serving to comment on a previously-established topic.
For example, in (60a), the topic of the first sentence is easily understood to be the
women who were harassed, and PRO in the adjunct can easily refer to them. In
(60b), on the other hand, the first sentence favors John as its topic, and so PRO is
more likely to refer to him.
(60) a. Many women were harassed by John. [After PRO1 talking to the man-
ager], complaints were filed.
b. ??John harassed many women. [After PRO talking to the manager],
complaints were filed. (Kawasaki, 1993, p. 177)
Second, Kawasaki observes that in adjunct control sentences, the matrix clause
must provide some information on the understood referent of PRO, such that the
sentence can be understood as having that referent as its topic. For example, in
(61a), the final clause can be understood to be a comment about the pirates in
48
virtue of their being the passive agent, and PRO in the adjunct can refer to them.
When the intransitive sank is substituted for the passive verb in (61b), however, this
is not the case. Kawasaki claims that in this case, PRO can only refer to the pirates
if it is their ship that sank, such that the matrix clause can still be understood as a
comment about the pirates.
(61) a. Suddenly the pirates1 showed up from behind the rocks. [After PRO1
robbing the passengers], the ship was sunk.
b. Suddenly the pirates1 showed up from behind the rocks. [After PRO1
robbing the passengers], the ship sank. (Kawasaki, 1993, p. 172)
Kawasaki’s third observation is that the acceptability of the sentence decreases
where the referent of the missing subject is not the best candidate for the sentence
topic. For example, in (62a), PRO cannot refer to the passive agent of the matrix
clause, because the topic of the sentence is much more likely to be the human
subject, despite the fact that PRO can refer to a passive agent in other instances
(62b). According to Kawasaki, this is partly due to the fact that subjects, especially
when human, are more likely to be sentence topics (Reinhart, 1981; Comrie, 1989).
(62) a. *[Before PRO1 leaving for New York], Mary was criticized by John1.
(Kawasaki, 1993, p. 165)
b. [After PRO1 sinking the ships], many houses were destroyed by the
soldiers1. (Kawasaki, 1993, p. 163)
49
Kawasaki’s examples all contain fronted adjuncts, which we have seen are more
permissive with respect to non-subject control. But her account may also shed light
on why NOC is less likely for right-adjoined adjuncts. NOC is available in (63a),
according to Kawasaki, if the matrix clause is taken to be about the people who
pitched the tents (i.e. if it talks about what happened to them, not what happened
somewhere else). Although she does not discuss the right-adjoined counterpart in
(63b), it seems likely that when the matrix clause is introduced first, the sentence
will be understood to be about the darkness itself; for this reason, the most likely
interpretation is one where darkness controls PRO.
(63) a. [After PRO pitching the tents], darkness fell quickly.
(Kawasaki, 1993, p. 173)
b. #Darkness fell quickly [after PRO pitching the tents].
However, Kawasaki’s control by sentence topic cannot account for all the data.
One example is (64), repeated from (24c). This sentence allows NOC control by some
unnamed person, but the strongest candidate for the topic of the sentence is the
potatoes. Kawasaki’s account may apply if one assumed that the potatoes had to
belong to the boiler, similar to the way the sinking ship had to belong to the pirates
in (61b), but this is problematic. In (61b) it is possible to imagine that the sentence
is describing what happened to the pirates, but it seems like a stretch to say that
(64) describes what happened to the person who boiled the potatoes.
(64) The potatoes got softer [after PRO boiling them].
50
2.3.2.4 Logophoricity
The effect of logophoricity on adjunct control was noted at least as early as
Williams (1992), and was also observed by Landau (2000). Williams argued that
there are two types of adjunct control: predicative and logophoric. Logophoric
control was said to occur any time the antecedent of PRO was not in the local
syntactic domain of PRO, in which case PRO must refer to some perspective holder
or source of information, i.e. the logophoric center of the sentence (Sells, 1987; Huang
& Liu, 2001). An example from Williams is given in (65). Here PRO can refer to
John because the sentence is expressing his emotional state. Similar arguments have
been made by Landau (2015, 2017), who states that any case of non-predicative
control must be logophoric.
(65) John1’s fears2 always go out of control, [when PRO1/#2 listening to Larouche].
This observation seems to be correct. In many if not all of the examples in this
chapter, when NOC is available, PRO is understood as referring to some perspective
holder. This can often be the speaker, as they are, in the terms of Sells (1987), the
source of the report. In (66), for example, we see that PRO can have a logophoric
interpretation in that it refers to a set of individuals containing the speaker.
(66) That oasis was a vision [after PRO dragging ourselves through the desert
all day]. (Landau, 2017, p. 7)
51
NOC is more difficult in (67a), in which PRO cannot refer to the speaker due to
the third person reflexive in the adjunct. But, when a new source is given in the
preceding context, NOC is possible, with PRO referring to that source, as in (67b).
Note that in (67b) the tense has been changed in order to ensure that the adjunct
is understood to be adjoined to the embedded clause.
(67) a. ??That oasis was a vision [after PRO dragging himself through the desert
all day].
b. Sam1 thinks [that oasis will be a vision [after PRO1 dragging himself
through the desert all day]].
This is further demonstrated in the contrast in (68). When PRO in this
adjunct is understood to refer to the speaker, it is not necessary to explicitly state
the perspective holder; in (68a), the PP for me is optional. In order for PRO to refer
to someone other than the speaker, however, the PP is no longer optional, since it
is necessary to establish the intended referent of PRO as a logophoric center, unless
the sentence is embedded in a larger discourse establishing the perspective holder
(68c).
(68) a. It was hard (for me1) to eat any kind of meat for a while [after PRO1
finishing this book].
b. It was hard *(for Jim2) to eat any kind of meat for a while [after PRO2
finishing this book].
52
c. Jim1 used to love steak, but he1 said that [it was hard (for him1) to eat
any kind of meat for a while [after PRO1 finishing this book]].
Additional evidence comes from the example repeated in (69). As already
discussed, PRO in this example can refer to someone other than the passive agent
(the sinker of the ship). At least to me, this sentence best describes a situation seen
from the eyes of the people who undocked the ship, i.e. from the perspective of the
referent of PRO. It would be strange for the sinkers of the ship to utter (69) unless
they were also the undockers.
(69) The ship was sunk [immediately after PRO undocking it].
The same point can be made with (70). PRO in this example may refer to the
tasters of the potatoes, in which case their perspective is at least implied by the
predicate “tastier”. But as was already noted, the person boiling the potatoes does
not necessarily need to be the same as the person tasting the potatoes in order for
(70) to be acceptable. However, using (70) to describe such a situation only seems
natural if it is the boilers who are making the report, not the tasters.
(70) The potatoes were tastier [after PRO boiling them].
Although PRO in temporal adjuncts can in many cases refer to the speaker
as a logophoric source, other factors may restrict this interpretation. We have
seen, for example, that NOC is possible for the fronted adjunct in (71a), but not
53
for the right-adjoined adjunct in (71b). However, it is not the case that (71b) never
allows NOC. When additional context is given to explicitly establish a perspective
holder, logophoric control becomes available (72). This suggests that NOC is widely
and generally available for temporal adjuncts; for some, though, logophoric control
requires additional contextual support.
(71) a. [Before PRO pitching the tents], darkness fell.
b. #Darkness fell [before PRO pitching the tents].
(72) Peter1 and Jon2 were arguing about when they finally got their tents pitched.
Peter1 thought [that darkness fell [before PRO1+2 pitching the tents]].
The same is seen in (73). Given the context in (73a), PRO in (73b(i)) can refer to
Peter, as the source. Note that the context on its own, however, is not enough.
Given the same context in (73a), NOC is not available in (73b(ii)). Rather than
explicitly representing Peter’s perspective, thoughts, or feelings, this sentence states
the perspective of the speaker. Hence, PRO cannot refer to Peter.
(73) a. At some point yesterday, the dishwasher broke. Also yesterday, Peter
came home for a 2-week visit. Mom thought that Peter broke the
dishwasher.
b. (i) Peter1 says [the dishwasher2 broke [before PRO1/2 even arriving
at the house]].
54
(ii) *But it turned out that [the dishwasher2 broke [before PRO1 even
arriving at the house]]
Effects of logophoricity can also subsume at least some of the effects of top-
ichood noted in Kawasaki (1993). For example, Kawasaki claimed that (74) is
only acceptable when the ship belongs to the pirates, and that this is because only
through ownership of the ship could the pirates become the topic of the matrix
clause. But it could also be that the ship needs to belong to the pirates in order
to make the pirates the logophoric pivot of the sentence, such that the situation is
described from their physical point of view (Sells, 1987).
(74) Suddenly the pirates1 showed up from behind the rocks. [After PRO1 rob-
bing the passengers], the ship sank.
(Kawasaki, 1993, p. 172)
And in (75), PRO can refer to the discourse addressee, which Landau (2015)
argues can also be a logophoric center.
(75) [After PRO1 working on Problem Set 1], Problem Set 2 shouldn’t take (you1)
much time. (Kawasaki, 1993, p. 163)
These examples all demonstrate the importance of logophoricity in NOC of
temporal adjuncts. In each case, PRO refers either to the speaker or some other
established perspective holder or logophoric center. This is only true, though, for
55
NOC. In OC, the controller need not be a human logophoric center, or even animate,
as has already been shown with examples such as (76).
(76) a. It1 never snows [after PRO1 raining in this part of the world].
b. The dishwasher1 finally broke [after PRO1 working well for many years].
The importance of logophoricity has much more to do with the interpretation
of PRO in NOC, and less to do with its distribution. Although NOC becomes more
available with an established logophoric center, logophoricity tells us nothing about
the structural tendencies seen above, such as why right-adjoined adjuncts are less
likely to allow NOC than fronted ones.
2.4 Proposal
The preceding sections have demonstrated that the distribution of OC and
NOC in temporal adjuncts cannot be determined based solely on the attachment
site of the adjunct or structural properties of either the adjunct or the matrix clause.
Instead, it was shown that discourse factors such as logophoricity are also impor-
tant. This suggests that NOC and OC are not strictly complementary, contrary
to common views (Williams, 1980; Chomsky, 1981; Manzini, 1983; McFadden &
Sundaresan, 2016) on the distribution of the two types of control. I propose that
both OC and NOC are available in temporal adjuncts (other types of adjuncts will
be discussed in Chapters 3 and 4). OC readings are the result of a syntactic de-
pendency between PRO and its antecedent, while the NOC readings are generally
56
logophoric, with PRO referring to some perspective holder in the discourse (Sells,
1987; Landau, 2015, 2017), although generic (‘arbitrary PRO’) readings are also pos-
sible. Note that it cannot be the case that control of temporal adjuncts is always
NOC or logophoric. OC is required at least in cases such as (77) where an inanimate
subject controls PRO, since NOC and all cases of logophoric control require human
antecedents. Other tests also confirm that these and other cases of subject control
are OC. This will be discussed in §2.5.1.
(77) a. It1 never snows [after PRO1 raining in this part of the world].
b. The dishwasher1 finally broke [after PRO1 working well for many years].
The idea that both OC and NOC are available for control of temporal adjuncts
is similar to claims made by Williams (1992) and developed by Landau (2015, 2017),
who stated that adjunct control can either be predicative (what I have been calling
OC) or logophoric (my NOC cases). But instead of claiming that NOC is only
available when OC is not, either because of the structural relation of the adjunct to
the matrix clause or because of other factors, I maintain that both OC and NOC
are truly always allowed by the grammar. This is not to say that OC and NOC are
equally available. Indeed, we have seen much evidence that OC is often strongly
preferred. I will argue that this is due to interpretive biases in processing.
This section will outline the details of this proposal. In §2.4.1, I discuss in
further detail the grammatical distinctions between OC and NOC. The preference
for OC in temporal adjuncts and its sources will be discussed in §2.4.2, and the
57
factors that may lead to NOC in §2.4.3. The following section, §2.5, will discuss
further implications of the proposal.
2.4.1 Grammar allows both OC and NOC
Previous theories have claimed that OC and NOC are syntactically comple-
mentary. What this meant is that it was the structure of the clause of which PRO
was a subject or its attachment site that determined control type. For example,
when the controlled clause is a verbal complement, as in (78a), OC is required.
When it is a clausal subject, as in (78b), it is NOC.
(78) a. Harry1 managed [PRO1 to win the tournament].
b. [PROarb To win the tournament] would be very impressive.
I have argued that temporal adjuncts cannot be classified as OC or NOC
based on structural properties in the same way. The type of control obtained is
not determined by attachment site or internal structural properties of the adjunct
or the matrix clause. Depending on which theory of control is taken to be correct,
though, there may still be syntactic differences between OC and NOC. Under the
MTC (Hornstein, 1999; Boeckx et al., 2010), for example, OC entails that PRO is
in reality the trace of movement, while in NOC it is a null pronominal. According
to the TTC (Landau, 2015), NOC has an additional layer in the clausal structure
of the controlled constituent compared to OC.9 But these differences simply reflect
9 The TTC does not actually distinguish between OC and NOC. Instead, the readings that I
have been calling OC would be classified as predicative control, and those that I have called NOC
would be logophoric control. Although the distinction is important, as predicative and logophoric
58
the definition of OC and NOC in those theories. What I mean when I say that the
two types of control are not grammatically complementary in temporal adjuncts is
that structural factors other than those that are definitional to OC and NOC do
not determine whether OC or NOC will occur. In other words, both a trace and a
pronominal, in terms of the MTC, or both predicative and logophoric structures, in
terms of the TTC, can appear in the same place. In this sense, OC and NOC are
in overlapping distribution syntactically.
This distinction is an important step in explaining the data above. To repeat
examples given above, (79)–(81) can be interpreted as OC or NOC, without, ar-
guably, changing the structure (modulo possible differences between OC and NOC
themselves), but based on context. This is only possible if structural factors alone
do not determine the type of control a given construction will have. Other factors
must be involved.
(79) Chickens1 taste better [after PRO1/2 being forced to eat celery].
(80) The pool1 was the perfect temperature [after PRO1/2 being in the hot sun
all day].
(81) a. (i) At some point yesterday, the dishwasher broke. Also yesterday,
Peter came home for a 2-week visit. Mom thought that Peter
broke the dishwasher.
(ii) Peter ordered a new dishwasher, and the store had it delivered.
control do not map directly to OC and NOC in all cases, it is sufficient at this point to assume
that the terms are roughly equivalent, at least for the cases of adjunct control discussed here. This
will be discussed in greater detail in Chapter 4.
59
It is not working, and the store manager doesn’t want to replace
it. He said that Peter must have broken it.
b. Peter1 says [the dishwasher2 broke [before PRO1/2 even arriving at the
house]].
2.4.2 Preference for OC
Although both OC and NOC in temporal adjuncts are allowed by the syntax,
this does not mean that both are equally available. Especially for right-adjoined ad-
juncts, NOC readings are acceptable only in particular contexts. On its own, (81b),
for example, only allows PRO to refer to the dishwasher. Having a logophoric con-
troller requires a strong supporting context like (81a(i)), and even then, the reading
is difficult or marginal for some speakers. This suggests, as has been noted by Lan-
dau (2017), that there is a strong preference for OC. This preference exists despite
the fact that both OC and NOC are allowed by the grammar. It is this preference
for obligatory subject control that led previous researchers to classify adjunct con-
trol as strictly OC. There are several potential reasons for this preference. As will
be seen, most if not all of these factors necessarily involve processing biases, rather
than grammatical preferences. I discuss a few here.
2.4.2.1 Preference for local antecedent
First, the preference for OC over NOC may reflect a general preference for local
antecedents in online processing. When a comprehender looks for an antecedent
to PRO, the OC reading requires a search of the local syntactic context. Perhaps
60
searching this context is less difficult than searching the surrounding discourse. This
preference for local antecedents would not be unique to adjunct control. There
is evidence that it also exists for picture NPs. Cunnings and Sturt (2014) used
eyetracking while reading to investigate the processing of sentences such as (82).
They found slower reading times when the reflexive matched the gender features
of the more distant antecedent compared to when it matched the more local one,
suggesting that it was more difficult to establish the non-local anaphora. This was
true despite the fact that in offline measures, participants readily accepted both
local and non-local antecedents for the reflexives.
(82) Mary heard about the soldier’s picture of herself/himself.
With PRO in temporal adjuncts, the picture is somewhat different. Unlike the
picture NPs in Cunnings and Sturt’s experiments, which allowed both readings in
offline judgments, the strong preference for the local antecedent (OC) reading seen in
many examples in this chapter does not seem like a temporary processing effect, since
the judgments reported are for offline interpretations. Thus, although the preference
for a local antecedent may be a contributing factor for the strong predominance of
OC, it is likely not the whole story. Still, it is possible that establishing a non-local
logophoric dependency is more difficult than forming a local syntactic one, adding
to the preference for OC over NOC in most cases.
61
2.4.2.2 Preference for syntactic relations
As a similar, and perhaps not unrelated, possibility, Landau (2017) cites a
common observation that “wired-in” operations take priority over grammar-external
ones. This means that when a syntactic control relation (OC) is possible, it will be
preferred over non-syntactic control by a discourse entity (NOC) (see also Lebeaux,
1984; Williams, 1992; Kawasaki, 1993).
This idea is similar to the proposal from Reinhart (1983) discussed in Heim
and Kratzer (1998), in which it is stated that if the same meaning can be expressed
by two minimally different LFs, one of which has a bound variable and the other
a free pronoun ((83a) and (83b) respectively), then the form with a bound vari-
able (83b) will be preferred by both speakers and hearers.10 For adjunct control,
this translates to OC being preferred over NOC because the former relies on PRO
being syntactically bound by its antecedent, while in the latter, PRO receives its
interpretation through the discourse.
(83) a. Felix voted for him1. gc = [1 → Felix]
b. Felix 1 [t1 voted for himself1] (Heim & Kratzer, 1998, p. 271)
Under such proposals, OC would be preferred over NOC because of a pref-
erence for syntactic relations over non-syntactic ones at an interface outside of the
syntax proper.11 But contrary to what Landau (2017) claims, it is not always nec-
10 Although see Lasnik (1991), Fox (2000), Reinhart (2006), Roelofsen (2008), Heim (1998, 2009),
Drummond (2011), a.o., for alternative theories and discussions on problems with Reinhart’s (1983)
proposal.
11 These proposals are often discussed in such a way that the comparison of different LFs is at an
62
essary to rule out the bound variable reading (e.g. through semantic anomaly) in
order for NOC to occur, as has been illustrated already with the examples in (55)
(see §2.3.2.1).
2.4.2.3 Theory-internal preferences
The previous preferences could apply regardless of what theory of control one
subscribes to. Within individual control frameworks, there may be additional factors
that would favor OC over NOC. According to the TTC, for example, predicative
control, which corresponds to what I have labeled OC, is structurally simpler than
logophoric control (NOC).12 As illustrated in (84), adjuncts with NOC would have
an additional clausal projection above that which is found in OC adjuncts (Landau,
2017).
(84) Clausal structure for OC and NOC under the TTC
a. [before [FinP Fin [TP PRO. . . ]]] (OC)
b. [before [CP C+log [FinP Fin [TP PRO. . . ]]]] (NOC)
Based on this difference, Landau (2017) argues (following Grimshaw, 1994;
Bošković, 1996; Speas, 2006) that OC is preferred to NOC due to a preference
to assign strings the simplest possible structure. Because the NOC reading would
require an additional layer of structure in the clausal projection of the adjunct, it
is disfavored when the simpler OC structure is possible. Note that this would be a
interface between two components of the grammar, but it may also be possible for the comparison
to happen outside the grammar, at an interface with the processor.
12 See Footnote 9.
63
processing preference, rather than one enforced by the grammar. In order for the
grammar to distinguish between the two structures in (84), it would have to compare
derivations that can lead to different interpretations. This would be problematic,
as it would be unclear why the grammar would not similarly prefer (85a) to (85b),
since the former has a simpler structure than the latter.
(85) a. John left.
b. John left yesterday.
Under the MTC, instead of OC and NOC differing in clausal structure, they
differ in the representation of PRO. In OC structures, PRO is a deleted copy
in a chain, i.e. it is the trace of movement of the antecedent from the embedded
clause to the matrix clause. In NOC, PRO is simply the null pronominal pro, and
its interpretation is more free. According to Hornstein (1999), using this pro is a
“costly” operation akin to do-support, and is therefore dispreferred to OC through
movement. As such, NOC via pronominalization should only be available when OC
via movement is not.13
For temporal adjuncts, this would mean that OC is preferred over NOC be-
cause the operation to introduce the null pronoun allowing NOC readings is “costly.”
NOC should only be possible when other factors either reduce the relative “cost”
of pronominalization, or when OC via movement is blocked. What is not clear is
what this “cost” is. Is it another processing preference, or is there something in
13 It was later argued that this generalization is a bit too strong, and it was clarified that NOC
should be ruled out in cases where OC could have established the same interpretation (Boeckx &
Hornstein, 2007; Boeckx et al., 2010). See Chapter 4.
64
the derivation of NOC that makes it dispreferred in the syntax? Formulations of
do-support (e.g. Chomsky, 1991) place the preference squarely in the grammar by
stating that do is only inserted to rescue a derivation that would otherwise crash. If
NOC via pronominalization is akin to do-support, then OC would also be preferred
in the grammar. However, this results in complementarity of OC and NOC, which
we have established is not the case for temporal adjuncts. Therefore, it may be more
likely that the preference for movement over pronominalization is due to processing
biases, and may even be reducible to other factors, such as a processing preference
for syntactic versus discourse dependencies more generally.
I will return to this issue in Chapter 4, which includes a more detailed discus-
sion of both the MTC and the TTC.
2.4.3 Factors favoring NOC
There are many reasons that OC may be preferred over NOC, some of which
depend on which broader theory of control is assumed. The fact that preferences
of this type have been proposed in both the MTC and the TTC, as well as in
other domains such as variable binding, suggest that some sort of OC preference
is on the right track. But this is only a preference, not a strict requirement—
otherwise, no NOC reading would ever be possible for temporal adjuncts or any
other sentence. This section returns to several factors introduced in §2.3 that may
lead to an increased availability of NOC readings and discusses why these factors
may help overcome the OC preference. Although none of these factors on their own
65
can guarantee that NOC will be available in a given sentence for any given speaker,
each can contribute to that possibility.
2.4.3.1 Attachment site
As noted above, the preference for OC is much stronger for right-adjoined
and medially-adjoined adjuncts than for fronted or extraposed adjuncts. There are
at least two possible explanations for this effect. First, temporal adjuncts may
optionally be able to base-generate in the fronted or extraposed position. In such
a derivation, the matrix subject would never c-command PRO in the adjunct, and
could therefore not bind it. OC would thus not be available, and NOC would not
need to compete with it. Note that high base-generation would be an option, but
not the only possible derivation, as even fronted and extraposed adjuncts can exhibit
OC, as the examples repeated in (86) and (87) demonstrate.
(86) a. [After PRO1 causing a lot of trouble], the dishwasher1 finally broke
down. (Kawasaki, 1993, p. 31)
b. [After PRO1 snowing a lot in February], it1 never rains heavily in March
in this area. (Kawasaki, 1993, p. 31)
(87) a. The dishwasher1 finally broke down—after PRO1 causing a lot of trou-
ble.
b. It1 never rains heavily in March in this area—after PRO1 snowing a lot
in February.
66
If c-command is required for OC, as most theories of control agree,14 then
these adjuncts must have originated low, and it is plausible that OC would still be
preferred. However, this is not necessarily the case. Suppose that the OC readings
in (86) and (87) require reconstruction of the adjunct from its surface position to a
lower position c-commanded by the subject, as proposed by Borer (1989). If there
is a processing cost associated with reconstruction, then it may balance the extra
cost associated with NOC, negating the basic preference for OC. Artstein (2005)
provided evidence that when temporal adjuncts are preposed, they tend to take
wide scope over the matrix clause. While (88a) is ambiguous with respect to the
scope of the two quantifiers, (88b) only has a wide-scope reading for the adjunct.
This is compatible with reconstruction being dispreferred.
(88) a. A secretary cried [after each executive resigned].
b. After each executive resigned, a secretary cried.
Anderson (2004) similarly notes that surface scope is generally preferred over inverse
scope in the online comprehension of sentences such as (89). This cannot be a direct
comparison, since it involves quantifier raising instead of reconstruction, but it does
give further support that surface scope is preferred in online processing.
(89) A climber scaled every cliff.
If there is indeed a cost for OC in fronted and extraposed adjuncts, then even
14 Hornstein and Kiguchi (2003) argue that c-command is not actually required for OC under
the MTC. See §4.3.3.2.
67
a derivation where the adjunct starts low and moves to the fronted or extraposed
position would not favor OC as strongly as with right-adjoined adjuncts remaining
in the lower position. Therefore, even if temporal adjuncts cannot base-generate
high, there would be reason to expect the preference for OC to be reduced and the
potential for NOC to be correspondingly strengthened.
A separate possible source of the greater availability of NOC at least in left-
adjoined adjuncts is mentioned in Boeckx and Hornstein (2007), though to different
effect. It is likely that comprehenders prefer to resolve referential expressions as
soon as possible upon encountering them (see, e.g. Nicol & Swinney, 1989; Badecker
& Straub, 2002). In right- and medially-adjoined adjuncts, this would be possible
through syntactic OC immediately upon encountering PRO. But in left-adjoined
adjuncts, PRO is encountered before any relevant potential OC controller. How-
ever, if NOC is assumed, PRO may in many instances be immediately resolved to
a relevant logophoric antecedent, if one has been previously established in the dis-
course or to the speaker as the source. The preference for immediate resolution of
reference would thus be in competition with the OC preference, which would lead to
greater availability of NOC. However, a separate explanation would still be needed
for extraposed adjuncts.
2.4.3.2 Unavailability of subject control
Another factor that can decrease the likelihood of OC is the unavailability of
subject control. As many authors have noted, and as discussed above, NOC readings
often occur when OC would produce a sentence that describes a bizarre situation.
68
Since listeners are generally charitable and try to maximize the coherence of what
their interlocutors say (Davidson, 1967; Hobbs, 1979), they may be more likely to
ignore their preference for OC in such cases. However, as noted above, having a
sentence where OC would result in a reading describing a strange situation is not
always sufficient to allow NOC. Evidently, even when the interpretation is unlikely
to be true in the real world, the preference for OC is sometimes strong enough to
rule out NOC.
2.4.3.3 Implicit arguments
It was also seen that temporal adjuncts adjoined to passive matrix clauses and
other clauses that have been argued to contain implicit arguments are more likely to
permit NOC. This may be because often the implicit arguments, especially implicit
experiencers as in (90), can serve as logophoric perspective holders.
(90) Those lemons tasted especially sour [after PRO eating sugar].
Since the matrix clause in (90) facilitates the establishment of an implicit logophoric
center, it may reduce the overall processing cost associated with establishing an NOC
referential dependency. This reduction would help explain the greater availability
of NOC in at least some cases involving implicit arguments.
2.4.3.4 Humanness of subject
We have also seen that when the subject of the matrix clause is human, non-
subject control is less likely. This was true for many of the examples with unergatives
69
in their matrix clauses seen above, which often have human subjects and are less
likely to exhibit NOC than examples with unaccusatives, which often have inanimate
subjects. In addition, studies in the processing literature have found that sentences
containing temporal adjuncts are generally more acceptable when the matrix-clause
controller is animate (see, e.g. Parker et al., 2015).15 One possible explanation
for this is that since NOC interpretations are logophoric, they require a human
subject. When the matrix clause is human, it is possible for its referent to serve as
a logophoric center. Therefore, if NOC were to occur, the matrix subject could still
be chosen as a logophoric center. Because of this, even if NOC were to be allowed,
it would not always be detected, since it would result in the same interpretation as
OC.
2.4.3.5 Logophoricity in context
If, as I have argued, NOC in temporal adjuncts is logophoric in nature, then it
is unsurprising that aspects of logophoricity in the surrounding context would affect
whether NOC is likely to occur. When the context establishes a salient perspective
holder, the potential difficulty with forming logophoric dependencies would be di-
minished, which in turn would reduce the preference for OC. This was seen in the
contrast between (91) and (92). On its own, (91) for many speakers seems to only
have the strange reading resulting from OC where darkness itself pitched the tents.
15 Parker et al. (2015) classify control of temporal adjuncts as OC. If this were always the case,
then the preference for animate subjects that they found would be somewhat surprising. However,
under the current proposal, because temporal adjuncts can also have logophoric interpretations,
the preference for an animate controller is more expected, since logophoric interpretations must
involve human, and therefore animate, antecedents.
70
But NOC is more available if the same sentence is set in a context that establishes
a potential logophoric antecedent (92).
(91) #Darkness fell [before PRO pitching the tents].
(92) Peter1 and Jon2 were arguing about when they finally got their tents pitched.
Peter1 thought [that darkness fell [before PRO1+2 pitching the tents]].
2.5 Implications
Classifying control of temporal adjuncts as strictly OC only accounts for in-
stances of control by the matrix subject. But non-subject control of temporal ad-
juncts is also possible, although it is less common and its acceptability may vary
from speaker to speaker (Kawasaki, 1993). The above proposal provides a way to
account for these facts. First, it is able to account for non-subject control examples,
as it allows for the possibility of logophoric NOC. The fact that NOC is less common
may stem from a number of different factors favoring OC. The account also leaves
room for speaker variation. NOC in temporal adjuncts is dependent on a number
of interpretive biases, and individual speakers may assign different weights to those
biases. Finally, in allowing for both NOC and OC, the proposal can still account for
the data that was previously used to classify control of temporal adjuncts as OC.
This section first returns to the evidence for OC in temporal adjuncts outlined
in §2.2. I will then briefly discuss the implications of the above proposal for other
types of adjunct control and for complement control.
71
2.5.1 Addressing original evidence of OC
Although we have seen much evidence that NOC is possible in temporal ad-
juncts, there were good reasons for classifying it as OC. In many instances, temporal
adjuncts display properties known to be associated with OC. This is completely con-
sistent with the current proposal. If temporal adjuncts can be either OC or NOC,
it is predicted that the properties associated with OC will hold, but only for OC
adjuncts, not for those that have logophoric NOC. I will now show that this is the
case for each of the OC properties discussed in §2.2.
The first piece of evidence that was used to argue for OC in temporal adjuncts
is that often, only the local subject can control PRO, not the object, other non-
c-commanding elements, or non-local arguments. This was seen in the examples
repeated in (93).
(93) a. Harry1 saw Ron2 [before/while/after PRO1/*2 leaving the room].
b. *Ron1’s sister left [without PRO1 having shaved himself].
c. Harry1 said [that [Ginny2’s brother]3 left [after PRO3/*1/*2 eating a
bagel]]
Beside the fact that one could potentially rule out non-c-commanding con-
trollers in (93) for independent reasons (e.g. perhaps for (93a), Harry but not Ron
is a good logophoric antecedent), we have seen several examples where PRO does
corefer with a non-c-commanding element (94a) or with a non-local argument (94b).
72
In the example in (94c), suggested by a reviewer, it appears that control by the ob-
ject may also be possible.16
(94) a. John1’s fears always go out of control, [when PRO1 listening to Larouche].
(Williams, 1992, p. 307)
b. Peter1 said [the dishwasher broke [before PRO1 getting home]].
c. The security guard stopped the woman1 [before PRO1 boarding the
plane].17
Second, OC adjuncts do not license arbitrary interpretations. In (95), for
example, oneself is not licensed. However, when the sentence allows an NOC inter-
pretation, arbitrary control is possible, and oneself is licensed (96).
(95) Ron1 felt old after [PRO1/*arb seeing himself/*oneself in the mirror].
(96) Adrenalin always kicks in [after PROarb putting oneself in danger].
Third, OC only allows sloppy interpretation under ellipsis, as in the OC exam-
ple in (97). However, if the antecedent to the ellipsis site contains an NOC adjunct,
16 Nunes (2014) discusses examples of object control of temporal adjuncts in Portuguese such
as (i). In contrast to (94c), object control occurs in Portuguese only when the object undergoes
wh-movement, and Nunes argues that such cases are still OC.
(i) [Que mulher]2 é que [o João]1 cumprimentou t2 depois de [ec1/2 entrar na sala]
which woman is that the João greeted after of enter in-the room
‘Which woman did João greet after he/she entered the room?’
17 Although this example seems acceptable, Norbert Hornstein (p.c.) notes that when a reflexive
is placed in the adjunct, the example seems to get worse (i). This could suggest that object control
is not really possible, but it could also be that (i) is unacceptable for other reasons.
(i) *The security guard stopped the woman1 [before PRO1 printing herself1 a boarding pass].
73
strict identity should be possible. In (98a), the sentence can be understood to mean
in part, that the chef will be the one to salt both the potatoes and the carrots.
And in in (98b), insofar as the logophoric control reading is acceptable in the first
conjunct, strict identity in the second conjunct is also possible.
(97) Ron says [Ginny1 left [before PRO1 dancing], and Harry thinks [Hermione
did, too].
‘And Harry thinks Hermione2 left before she2/*Ginny danced.’
(98) a. The chef1 thinks [the potatoes will sell better [after PRO1 adding more
salt]], and his wife2 thinks [the carrots will, too].
‘. . . the carrots will sell better after he1/she2 adds more salt.’
b. Peter1 says [his dishwasher broke [before PRO1 getting home]], and
Sara thinks [his1 bed did, too].
‘And Sara thinks Peter1’s bed broke before he1 got home.’
Fourth, PRO in OC can only have a bound interpretation when controlled by
only-DPs. This property holds when the adjunct is OC (99), but as predicted, it is
not present for NOC adjuncts (100).
(99) Only Harry1 survived [after PRO1 being hit with the killing curse].
‘[Nobody else]1 survived after he1/*Harry was hit with the killing curse.’
(100) ?Only Peter said [his dishwasher broke [before PRO getting home]]. (Ev-
74
eryone else thinks that he broke it.)
‘Nobody else said that Peter1’s dishwasher broke before he1 got home.’
Finally, it was demonstrated that PRO in OC contexts does not allow split
antecedents (101); it must controlled by the matrix subject alone. If this is a prop-
erty of OC, then it may not hold in NOC contexts, although split antecedents may
still be ruled out in NOC for other reasons. The examples in (102) are acceptable
to some speakers, but are marginal to others. However, the speakers who do not
accept them may only have allowed OC readings in these contexts. When the NOC
reading is facilitated by preposing the adjunct (103), it becomes more acceptable
for PRO to have a split antecedent and to license each other or themselves.
(101) *Peter1 expected Susan2 to break up with him [without PRO1+2 hurting
themselves/each other].
(102) a. ??Peter1 told his pen pal2 [he thought [the world might end [after PRO1+2
meeting each other1+2]]].
b. ??Peter1 promised Susan2 [that the alarm would go off [after PRO1+2
killing themselves1+2]].
(103) a. ?Peter1 told his pen pal2 [he thought [that [after PRO1+2 meeting each
other1+2], the world might end]].
b. ?Peter1 promised Susan2 [that [after PRO1+2 killing themselves1+2], the
alarm would go off].
75
The authors who classified control of temporal adjuncts as OC were correct
for the majority of the data. When the adjunct is controlled by the subject of the
next higher clause, as is the case for most temporal adjuncts, all the properties of
OC are displayed. However, in other examples, where PRO in temporal adjuncts
may refer to someone other than the referent of the next highest subject, none of
these properties hold. This is consistent with the proposal that the grammar allows
both OC and NOC dependencies in the same type of adjunct.
2.5.2 Other types of control
Although this chapter has only dealt with temporal adjuncts, this proposal
may also apply to other types of adjuncts. For example, although rationale clauses
(RatCs) such as those in (104) have been argued to involve OC (Roeper, 1987;
Mauner, Tanenhaus, & Carlson, 1995), (105) gives examples where NOC has also
been argued to be possible (A. Williams, 2015; A. Williams & Green, 2017; Landau,
2017). It is possible, as Landau (2017) has argued, that cases such as (104) are
indeed OC and (105) NOC. RatCs will be discussed more in Chapter 3.
(104) Harry1 went into the shop [in order PRO1 to purchase a broom].
(105) a. Harry1 went into the shop. The goal was [PRO1 to purchase a broom].
b. A young girl cut the ribbon, [just PRO to acquire the votes of local
women]. (A. Williams & Green, 2017)
There are many other types of adjunct control, some of which can involve con-
76
trol by a non-subject. A broader discussion of these adjuncts is found in Chapter 4;
it will be shown that many of them, but not all, also allow both OC and NOC.
Besides its potential application to other types of adjunct control, this proposal
must be situated within some broader theory of control. I have argued that both OC
and NOC are possible in at least some types of adjuncts. But the same is not true
for complement control, illustrated in (106), which always requires a local syntactic
controller (Chomsky, 1981; Hornstein, 1999, 2001; Boeckx et al., 2010; Landau,
2000, 2015). In contrast to adjunct control, only OC is possible for complement
control.
(106) a. Harry1 managed [PRO1/*2 to win the tournament].
b. Ron2 persuaded Harry1 [PRO1/*2/*3 to win the tournament].
c. Harry1 promised Ron2 [PRO1/*2/*3 to win the tournament].
There are several possible reasons for this difference, dependent on the broader
theory of control chosen. In semantic theories of control, complement control is
OC because controller choice is an entailment of the verb whose arguments are
the controller and the controlled clause. According to Sag and Pollard (1991), for
example, the meaning of the verb promise entails that the implicit subject of the
promise clause will refer to the promiser (or the “committor” in Sag and Pollard’s
terms), resulting in OC by the argument corresponding to the promiser. The same
kind of controller choice cannot be applied to adjuncts, though, since they are not
77
coarguments with their controller. Chapter 4 will discuss why complement and
adjunct control differ within the MTC and the TTC.
2.6 Conclusion
In this chapter, I have argued that the grammar allows both OC and NOC into
temporal adjuncts. The strong preference for OC by the matrix subject that previ-
ously led researchers to classify it as OC is not due (solely) to syntactic constraints
against NOC, but to performance biases in comprehension and/or production, which
may include preferences for bound over free variables, for movement over pronomi-
nalization, and/or for assigning the simplest possible structure to a string.
This proposal is an important step to understanding syntactic and pragmatic
constraints on adjunct control, but it does leave several important questions. First,
I have sketched several possible interpretive biases leading to a preference for OC.
These biases all have at least some grounds in previous research, but it is unclear
which of them are actually at play for the interpretation of temporal adjuncts.
Especially useful would be studies on the online processing of adjunct control. Part II
of this dissertation is an important first step in that direction.
A second remaining question involves the constraints on the interpretation of
NOC. I have assumed, following Landau (2017), that NOC of temporal adjuncts
is generally logophoric, meaning that PRO must refer to some logophoric center
or perspective holder in the discourse. One argument for this assumption is the
requirement for NOC to involve a [+human] antecedent. This feature comes for
78
free in a logophoric account, as only humans can be logophoric perspective holders.
If the interpretation of NOC is more free, and can refer to the discourse topic, for
example, as Kawasaki (1993) argued, the [+human] restriction is left without a
complete explanation.
However, Lyngfelt (2000) provides (107) in Swedish, which he argues is an
example of control by the discourse topic. It would be somewhat difficult to re-
duce this example to logophoric control, since the understood antecedent to PRO
is not clearly the perspective holder. The English counterpart has the same NOC
interpretation.
(107) Tiger Woods1 var i praktiken bortr̊aknad fr̊an segerstriden, efter tv̊a “måns-
kliga” inledningsronder. Men efter att PRO1 ig̊ar ha tangerat banrekordet
v̊agar ingen2 r̊akna bort golfens nye “Golden Boy”1.
‘Tiger Woods1 was in practice dismissed from the winning competition,
after two “human” starting rounds. But after PRO1 having touched the
record for the course yesterday, no one2 dares to disregard golf’s new
“Golden Boy”1.’ (Lyngfelt, 2000, p. 32)
In addition, examples with a generic statement favor arbitrary interpretations
for PRO, and these interpretations do not require a specific perspective holder. It
is not clear that (108), for example, represents the perspective of the boilers of
the potatoes. But generic interpretations may be a principled exception to the
perspective requirement (see Bhatt & Izvorski, 1998).
79
(108) Potatoes get softer [after PRO boiling them].
One final remaining question is whether the preference for OC holds equally
across different types of temporal adjuncts. For example, NOC may be more avail-
able for adjuncts headed by while compared to those headed by before or after (Maria
Polinsky, p.c.). The source of such a difference, if robust, remains unexplained.
(109) a. ?The dishwasher1 finally broke [while PRO2 entertaining a big party
for Thanksgiving].
b. *The dishwasher1 finally broke [before PRO2 entertaining a big party
for Thanksgiving]
Although the analysis in this chapter is preliminary, it highlights some of the
difficulties associated with adjunct control, and is an important step to understand-
ing how syntactic and pragmatic constraints on reference are deployed in compre-
hension. In addition, it provides support for theories of control that do not rely on
strict complementarity between syntactic and non-syntactic forms of control.
80
Chapter 3: Rationale clauses
The previous chapter argued that both OC and NOC are available for temporal
adjuncts. This chapter argues, following Landau (2017), that the same is true for
rationale clauses (RatCs), alternatively called In-Order Clauses or a Reason Clauses
(Faraci, 1974; Jones, 1985; Whelpton, 2002; A. Williams, 2015), as illustrated in (1).
(1) Rita1 interviewed Harry2 [in order PRO1/*2/*3 to feel better about herself/
*himself/*oneself].
In this example, PRO is understood to be coreferential with Rita; it cannot
be coreferential with the matrix object Harry, and it cannot receive an arbitrary
interpretation. As was the case for temporal adjuncts, some authors have claimed
that this is due to a syntactic dependency between the two (Roeper, 1987; Whelpton,
2002), or in other words, that control in RatCs is OC. Others reject grammatical
accounts of control in RatCs, arguing for a discourse-based NOC approach (Farkas,
1988; Landau, 2000; A. Williams, 2015; A. Williams & Green, 2017). Continuing
in the traditional assumption that OC and NOC are complementary, these authors
have found evidence that at least some RatCs display properties of one type of
control, and have extrapolated that all RatCs must therefore also be of that type.
81
However, at least two authors have previously argued that both OC and NOC
are possible in RatCs: Español-Echevarŕıa (2000) and Landau (2017). I provide
evidence in this chapter for that view, just as was argued to be the case for temporal
adjuncts in Chapter 2. However, in contrast to NOC in temporal adjuncts, which
required PRO’s understood antecedent to be a logophoric perspective holder, I argue,
following A. Williams (2015), that PRO’s interpretation in RatCs is sensitive to
notions of responsibility (see also A. Williams & Green, 2017).
This chapter is organized as follows. In §3.1, I discuss general properties of
RatCs. In §3.2 I discuss whether control in RatCs may be strict OC, noting both
the virtues of such grammatical accounts and their failings. I do the same for NOC
accounts in §3.3. I argue in §3.4 that, just like temporal adjuncts, RatCs allow
both OC and NOC, and illustrate how this may solve the challenges faced by strict
OC or NOC accounts. Section 3.5 concludes the chapter, including a note on the
distribution of NOC in RatCs and other remaining questions.
3.1 Background on Rationale Clauses
RatCs provide a teleological explanation for the fact or proposition expressed
by their target clause. In (2), for example, the RatC β gives the reason or goal
behind the target, i.e. the fact expressed by the target clause α.
(2) [Harry1 went into the forest]α [in order PRO1 to confront Voldemort]β.
This section will outline the types of examples that have been used in previous
82
literature to classify control in RatCs as either OC or NOC. This discussion will
largely be expository, with analysis of the data coming in later sections. I will
demonstrate that PRO in a RatC can be controlled by the explicit subject of the
target clause, but not by the object. It can also be controlled by unexpressed entities
who can be seen as responsible for the fact expressed by the target clause, or by the
target fact itself.
3.1.1 Local RatCs
The majority of the literature on RatCs has focused on what A. Williams
and Green (2017) term “local control,” cases where the RatC is adjoined to the
target clause, and where PRO has an antecedent in that clause. In (2), PRO is
controlled by the subject of the target clause Harry, meaning that Harry and PRO
are coreferential; we understand that Harry is both the person who went into the
forest, and also the one who is to confront Voldemort. PRO in a RatC can be
controlled, then, by the explicit subject of the matrix/target clause.
Control by the matrix object, on the other hand, is impossible. Given a context
where sharks cover themselves with parasites, only (3a) can be used to mean that
the sharks are having their gills kept clean. Example (3b) can only mean that the
parasites themselves have gills that they intend to be kept clean (A. Williams, 2015;
A. Williams & Green, 2017).
(3) a. These sharks1 cover themselves with parasites2 [PRO1/*2 to have their
gills kept clean.
83
b. Parasites2 cover these sharks1 [PRO2/*1 to have their gills kept clean].
The same effect is seen in (4), where the intended escaper must be understood
as Dobby, and that protecting Harry would make it possible for Dobby to escape.
Example (4) cannot be used to mean that Dobby protected Harry in order for Harry
to escape. This is in contrast to when there is an overt pronoun as the subject of
the complement clause. When PRO in (4) is replaced with him in (5), it is much
more likely to refer to the object Harry than to Dobby.
(4) Dobby1 protected Harry2 [in order PRO1/*2 to escape from the Malfoys].
(5) Dobby1 protected Harry2 [in order for him2/??1 to escape from the Malfoys].
The restriction against control by the object also extends to other non-c-
commanding arguments. In the examples from Whelpton (2002, p. 187) in (6),
the subject DP can control PRO, but a possessor embedded within the subject
cannot.
(6) a. *David1’s wife2 brought some articles on reference along (in order) PRO1
to discuss semantics with her2.
b. David1’s wife2 brought some articles on reference along (in order) PRO2
to discuss semantics with him1.
In at least some cases, when the object is promoted to subject position, as in
a passive, it can control PRO. In other words, surface subjects can control PRO,
84
even if that subject was underlyingly an object. In the passive (7a), one possible
reading is that Sam got himself arrested so that he (Sam) could get off the streets.
This reading, with Sam controlling PRO, is only possible when Sam is seen as
intending to be arrested and to be somehow responsible for the arrest. In the words
of Zubizarreta (1982), he must be viewed as the “secondary agent” of the arrest.
Note, though, that the secondary agent reading is only available for the subject of a
passive. It is impossible for the object Sam to be understood as a secondary agent
in the active version in (7b).
(7) a. Sam2 was arrested (by the cops1) [just PRO1/2 to get him/himself off
the streets].
b. The cops1 arrested Sam2 [just PRO1/*2 to get him/*himself off the
streets].
PRO in (7a) can refer to Sam, but it can also refer to the agents of the arrest,
i.e. the cops. As A. Williams (2015) puts it, PRO can be controlled by a subject of
the target clause, whether the surface subject, or what he calls the deep subject, i.e.
the entity that would have been the surface subject had the sentence been active.
Control by a surface object is impossible.1
In addition to control by the explicit subject of the target clause, PRO in a
RatC allows implicit control. In (7a), it is not necessary to include by the cops
1 Compare to the seemingly similar object-oriented purpose clause, which allows object control
(Faraci, 1974; Jones, 1991; Whelpton, 2002):
(i) We brought John1 along [PRO1 to talk to Mary].
85
in order to get the reading that whoever arrested Sam was doing it so that they
might get him off the streets. Similarly in (8), PRO is controlled by the entity
corresponding to the deep subject role of the predicate sank the boat, i.e. the sinker.
This control relation is implicit because PRO is bound by no audible dependent.2
Thus, even when the deep subject role is unrealized on the surface, it is available as
a controller for PRO in a RatC.
(8) The boat was sunk [PRO to collect the insurance]. (Roeper, 1987)
Furthermore, RatCs allow what I will call super -implicit control, where PRO
is understood as referring to an entity not named by any arguments of the target
predicate, whether explicit or implicit. In these cases, the controller is not syntacti-
cally related to the target clause at any point. For example, the passive in (9a) can
be used to describe a situation where a certain political group had a young girl cut
a ribbon so that that group could gain support. PRO can refer to the organizers of
the event even when they can in no way have been represented syntactically as part
of the sentence (A. Williams, 2015; A. Williams & Green, 2017). This is also true
for the active form if stress is placed on a young girl (9b).
(9) a. The ribbon was cut by a young girl [just PRO to acquire the support of
female voters].
2 Whether the controller is still syntactically present is a matter of much debate (Roeper, 1987;
Mauner et al., 1995; Lasnik, 1988; Landau, 2000, 2017; McCourt et al., 2015).
86
b. A young girl cut the ribbon [just PRO to acquire the support of female
voters].
However, although superimplicit control is possible in some instances, it is
not always. Suppose that Jon hired a crook to burn his house down so that he
(Jon) could later collect on the insurance policy. In such a situation, even when
Jon is understood as the organizer/responsible party behind the crook’s actions, it
is difficult or impossible to understand a speaker, in using (10), as meaning that the
intended collector of the insurance is Jon, and not the crook he hired.
(10) A hired crook1 burned down the house (for Jon2) [in order PRO1/*2 to collect
the insurance].
To the same point, Whelpton (2002, pp. 190–191) gives the examples in (11). Even
if it is understood that the soldiers in these examples are only following orders, and
that their commanding officer is responsible for their actions, PRO cannot refer to
the commanding officer; it seems only to allow control by the local subject, the
soldiers.
(11) a. *The soldiers1 unwittingly turned on the water [(in order) PRO2 to
drown the prisoners].
b. The officer1 knew that if the prisoners were drowning, the approaching
enemy troops would have to stop and save them. . .
87
(i) The soldiers2 therefore turned on the water [(in order) PRO2 to
give themselves2 time to escape].
(ii) *The soldiers2 therefore turned on the water [(in order) PRO1 to
give himself1 time to escape].
Finally, PRO in a RatC can also refer to the target fact itself. For example,
in the famous example from E. Williams (1974) given in (12) (and also discussed
in E. Williams, 1985; Lasnik, 1988, a.o.), it is not the grass that is promoting
photosynthesis. Instead, the promoter of photosynthesis is the fact itself that grass
is green.
(12) [Grass is green]1 [PRO1 to promote photosynthesis].
In sum, PRO in a local RatC can be controlled by (i) the surface subject of
the target clause, but not the object or other non-c-commanding arguments, (ii) the
understood demoted subject of a passive, (iii) other agents who can be understood
as responsible for the fact expressed by the target clause, or (iv) the fact expressed
by the target clause.
3.1.2 Remote RatCs
RatCs need not be adjuncts to their target clause. They may also occur as
the object of specificational copular sentences (Higgins, 1973) with subjects alluding
to the target fact such as the goal, the reason, or a pronoun, as in (13). In these
types of sentences, control is remote, since the antecedent to PRO is outside of
88
PRO’s syntactic domain (Dowty, 1989; Sag & Pollard, 1991; A. Williams, 2015; A.
Williams & Green, 2017).
(13) a. Harry1 went into the forest. The reason was [PRO1 to confront Volde-
mort].
b. I don’t know why Harry1 went into the forest. Maybe it was [PRO1 to
confront Voldemort].
In the remote RatCs in (13), PRO is understood as coreferential with Harry,
just as it would be in a local RatC. Indeed, the same restrictions on reference seen
for PRO in local RatCs also seem to apply to the referent of PRO in remote RatCs.
Just like in local RatCs, PRO in a remote RatC can be controlled by the subject
of the target clause, but not the object or other non-c-commanding arguments, as
illustrated in (14a–c). Remote RatCs also allow control by the implicit demoted
subject of a passive (14d) and superimplicit control (14e). When superimplicit
control is blocked in local RatCs, it also is in remote RatCs (14f). Finally, control
by the target fact is still available (14g).
(14) a. Dobby1 protected Harry2. The goal was [PRO1/*2 to escape from the
Malfoys].
b. Sam1 was arrested by the police. The reason was [just PRO1 to get
himself off the streets].
c. *David1’s wife2 brought some articles on reference along. The goal was
PRO1 to discuss semantics with her2.
89
d. The ship was sunk. The reason was [PRO to collect the insurance].
e. The ribbon was cut by a young girl. The goal was [PRO to acquire the
support of female voters].
f. A hired crook1 burned down the house. The reason was [PRO1/*2 to
collect the insurance].
g. [Grass is green]1. The reason is [PRO1 to promote photosynthesis].
However, although the referential constraints on PRO in a remote RatC seem
to parallel those in local RatCs, in several other ways, remote RatCs do not behave
as if they were local. Elements embedded within the target clause cannot scope
over remote RatCs like they do in local RatCs. As demonstrated for local RatCs
in (15), negation in the target clause can license a negative polarity item (NPI) in
the RatC (15a), and both negative and positive quantifiers in object position of the
target clause can bind a variable in the RatC ((15b,c)).
(15) a. The coach did not approach the team in order to give them any advice.
b. The coach approached [no member of the team]1 in order to give him1
advice.
c. The coach approached [every member of the team]1 in order to give
him1 advice.
This is only possible for readings where negation scopes over the RatC. In a situation
where one is asking why the coach did not approach the team, it is infelicitous to
answer with (15a), which can only be used to answer the question “What didn’t the
90
coach do?” or as a contradiction to a claim that the coach did approach the team to
give them some advice. Similarly, (15b) cannot be used to answer a question about
why the coach approached no member of the team, in which case variable binding
into the RatC is impossible (16).
(16) *The coach approached [no member of the team]1 in order not to intimidate
him1.
Remote RatCs do not demonstrate the same scope ambiguities. An NPI in a
remote RatC cannot be licensed by negation in the target clause under any context
(17a), and negative quantifiers in object position in the target clause cannot bind
variables in the RatC (17b). In other words, negation in the target fact can scope
over local RatCs, but not remote RatCs. We do see apparent variable binding by
every into remote RatCs (17c), but this is also possible in local RatCs when the
RatC scopes high ((15c) can be used felicitously even in answer to the question
“Why did the coach approach every member of the team?”). It is therefore likely
that the apparent variable binding by every member of the team in (15c) when the
RatC scopes high and in (17c) is due to special properties of every, rather than
something to due with RatCs or remote control.
(17) a. *The coach did not approach the team. The reason was to give them
any advice.
b. *The coach approached [no member of the team]1. The reason was to
give him1 advice.
91
c. The coach approached [every member of the team]1. The reason was
to give him1 advice.
Geurts (2011) discusses examples with every such as (18) from Karttunen
(1976) to introduce what he calls “piggyback anaphora” (these effects have also
been called “telescoping” of quantifiers (Roberts, 1987)). One reading of (18) is
that Harvey courts a different girl at every convention, and each of them went to
a ball with him. She behaves like a variable bound by a girl within the scope of
the universal quantifier. Piggyback anaphora in (17c) would also explain the lack of
binding effects in remote RatCs with negative quantifiers (17b). Negative quantifiers
do not allow piggyback anaphora, as can be seen in in (19).
(18) Harvey courts [a girl]1 at every convention. She1 always comes to the ball
with him.
(19) a. *Harvey courts [no girl]1 at every convention. She1 always comes to the
ball with him.
b. *Harvey courts [a girl]1 at no convention. She1 always comes to the ball
with him.
If (17c) is an instance of piggyback anaphora, then it is an exception to the otherwise
correct generalization that there is no grammatical binding into remote RatCs from
elements within the target clause.
92
3.2 OC accounts of RatCs
Because PRO in RatCs, even in remote RatCs, can be coreferent with the sub-
ject, but not the object or other non-c-commanding arguments of the target clause,
a grammatical OC account seems particularly suited. A subject/object asymmetry
and c-command restrictions are prime candidates for syntactic explanation. This
section will discuss what such an OC account would look like. I begin with local
RatCs, where syntactic OC accounts have the most success, although I will demon-
strate that even here additional stipulations are required. I then outline what would
be necessary for an OC account of remote control, demonstrating that remote RatCs
present an insurmountable challenge to purely OC theories.
3.2.1 OC and local control in RatCs
Recall that PRO in a RatC can refer to (i) the surface subject of an active
or passive target clause — but not the object (20a–c), (ii) what is understood as
the demoted subject of a passive (20d), (iii) other agents who can be understood as
responsible for the fact expressed by the matrix clause (20e), but only some of the
time (20f), or (iv) the target fact itself (20g).
(20) a. Dobby1 protected Harry2 [in order PRO1/*2 to escape from the Malfoys].
b. Sam1 was arrested by the police [just PRO1 to get himself off the
streets].
93
c. *David1’s wife2 brought some articles on reference along (in order) PRO1
to discuss semantics with her2.
d. The ship was sunk [in order PRO to collect the insurance].
e. The ribbon was cut by a young girl [in order PRO to acquire the support
of female voters].
f. A hired crook1 burned down the house [in order PRO1/*2 to collect the
insurance].
g. [Grass is green]1 [in order PRO1 to promote photosynthesis].
Theories positing a grammatical OC dependency in local RatCs can easily han-
dle cases of control by the subject of the matrix clause, and are able to rule out
object control and control by non-subject arguments with a simple c-command re-
quirement. However, as will be shown, superimplicit control cannot receive an OC
account, and allowing an NOC account of it creates problems for positing strict OC
in other RatCs, even in describing control by the subject versus the object. Because
of this, even local RatCs cannot be classified as strictly OC. Such a classification
would both under- and over-generate without additional stipulations such as those
that are already required for NOC accounts.
3.2.1.1 Control by the subject, not by the object or non-c-commanding arguments
Syntactic theories of OC maintain that the grammatical dependency between
PRO and its antecedent requires that the latter c-command the former (Williams,
1980; Chomsky, 1981; Landau, 2004, 2015). According to Whelpton (2002), a local
94
RatC adjoins to the matrix clause at the uppermost V′ (v′/Voice′/Asp′). Because of
this, the subject of the matrix clause would c-command the RatC and the PRO it
contains, but the object would not. Under most theories of OC, the position of the
adjunct rules out control by any argument other than the subject because the pred-
ication/control relation involved itself must involve the most local c-commanding
argument (e.g. Williams, 1980; Chomsky, 1981; Landau, 2015). Under the MTC
(Hornstein, 1999; Boeckx et al., 2010), the same effect is achieved without an explicit
c-command requirement, but because OC by the object or other non-c-commanding
arguments would incur a Merge-over-Move violation (Chomsky, 1995), causing the
derivation to crash.
3.2.1.2 Implicit control
In addition to control by the explicit subject of the matrix clause, RatCs allow
control by the understood demoted subject of a passive. In order to maintain a strict
OC account of control of RatCs, it is necessary to assume that the demoted subject
is syntactically represented and is available as a syntactic controller (Chomsky, 1986;
Jaeggli, 1986; Roeper, 1987; Collins, 2005). Baker, Johnson, and Roberts (1989)
instantiate such a theory by claiming that the passive morpheme -en receives the
agent θ-role and is able to bind the PRO of a RatC. A similar theory would place
something like a null pro in the spec-vP, as in (21).
(21) The boat [vP pro1 [was sunk [RatC PRO1 to collect the insurance]]]
95
Assuming this structure, OC of PRO in a RatC by the implicit demoted subject
of a passive clause would be quite similar to control by the subject of an active clause.
In either case, the external argument of the vP, whether explicit or implicit, serves
as the syntactic antecedent to PRO (Whelpton, 2002).
However, such grammatical accounts of the passive do not have much to say
about why implicit control is not possible in more contexts. In (22), we see that
implicit predicative control is impossible (Landau, 2015). But if implicit control of
RatCs is by a silent demoted subject in passives, then silent arguments (at least in
the passive in (22c)) should be able to control PRO in other constructions.
(22) a. John ate *(the meat) raw.
b. I am now hiring *(people) [for John to work with].
c. The room was left (*angry at the guests).
Because of this and other issues, the presence of a null pronoun which receives
the agent θ-role in passives is controversial. Lasnik (1988), E. Williams (1985),
Landau (2000), and A. Williams (2015) give strong arguments that there is no such
null argument in the syntax of passives. If this is the case, then there can be no
syntactic OC dependency between it and PRO.
Furthermore, implicit control by the understood agent of a passive must be
interpreted as [+human] only when it controls PRO (Manzini, 1986). Thus (23a)
cannot be interpreted to mean that the rain washed the stairs and then entered the
basement, even thought that interpretation is possible with an active matrix clause
96
(Landau, 2017, p. 4). This restriction on interpretation is only expected if control
by the implicit agent is really NOC.
(23) a. The stairs were washed [before PRO entering the basement].
b. The rain1 washed the stairs [before PRO1 entering the basement].
3.2.1.3 Superimplicit control
RatCs can also have a controller that is completely absent from the sentence,
as was seen in (20e), repeated in (24). Because there is no antecedent for PRO in
this sentence, it must be NOC.
(24) The ribbon was cut by a young girl [in order PRO to acquire the support
of female voters].
One potential explanation for the interpretation in (24) is that this really is
OC by the agent of the passive, and that the young girl is acquiring support for or
in behalf of the organizers of the event, or that the young girl is understood to be
in a part-whole relationship with the organizers (Tom Roeper, p.c.). However, this
kind of relation is not necessary for superimplicit control to occur. In (25a), PRO
more clearly refers to the organizers of the event, as inanimate signs cannot acquire
support, even on behalf of the organizers (25b). The same point is illustrated by
(26). Superimplicit control in RatCs cannot be OC.
97
(25) a. A pink sign greeted the guests [just PRO to acquire the support of
female voters].
b. #A pink sign acquired the support of female voters (for the organizers).
(26) The subtitles were in several different languages [in order PRO to acquire
the support of immigrant voters].
3.2.1.4 Fact control
We will now turn to control by the fact expressed by the matrix clause. In
(27), repeated from (20g) above, it is not the grass that promotes photosynthesis.
It is the fact that grass is green. Thus, the fact expressed by the matrix proposition
controls the PRO of the RatC.
(27) [Grass is green]1 [in order PRO1 to promote photosynthesis].
E. Williams (1985) and Lasnik (1988) label these as cases of “S-control,” stating
that the sentence itself is controlling PRO. One could attempt to instantiate this
idea syntactically by stating that the controller is the event variable. But not all
cases of fact control involve eventualities. For example, in (28), there is no single
event described by the proposition ewes kill easily, and yet it is this fact that is
understood to ensure the survival of local predators.
(28) Ewes kill easily, to ensure the survival of the local predators.
(A. Williams, 2015, p. 296)
98
Therefore, under an OC account of control of RatCs, it must be the case that PRO
can be controlled either by the matrix subject, the event variable, or when what is
explained is not an event but a fact about one or more events, by a syntactically-
represented situation variable (see, e.g. Percus, 2000).
3.2.1.5 Summary of local RatCs
Strict OC theories can account for the structural restrictions on control of
PRO in RatCs. The preference for subject control and the restriction against object
control and control by other structurally inaccessible arguments is obtained through
the c-command requirement common to syntactic theories of control. However, even
if subject control in RatCs is OC, at least implicit control and superimplicit control
must be instances of NOC. Thus, even local control cannot be fully explained by a
strict OC account.
3.2.2 OC and remote control3
Although the restriction against object control and non-c-commanding argu-
ments in local RatCs could be the result of an OC dependency, the same restrictions
appear to occur in the interpretation of remote RatCs. PRO in a remote RatC can
refer to the individual named by the subject but not the object of the target clause.
But in remote control, PRO and its understood antecedent are in separate sentences.
This is a major problem for OC accounts of RatCs, since, by definition, syntactic
dependencies must be within a single sentence. Because of this, A. Williams (2015)
3 Portions of the arguments in §3.2.2 are also published in A. Williams and Green (2017).
99
gives remote control of RatCs as evidence against the need for a grammatical de-
pendency between PRO in a RatC and its understood antecedent. If remote RatCs
must be an instance of NOC, then whatever explains the restriction against object
control in remote RatCs could also explain local RatCs.
To maintain an OC account of remote RatCs, and to explain the structural
restrictions on controller choice, would require that the remoteness of remote con-
trol is an illusion, and PRO in a remote RatC is bound by some sort of silent
constituent. The underlying structure of a remote RatC such as (29), for example,
could be something like one of the options in (30). Although this would still not
allow the antecedent to c-command PRO, as is generally required for syntactic con-
trol (Williams, 1980; Chomsky, 1981; Landau, 2004), it simplifies the problem by at
least putting the two in the same sentence.
(29) Harry1 broke into the ministry. The reason was [PRO1 to save Sirius].
(30) a. [DP Harry1 [D′ The reason]] was PRO1 to save Sirius.
b. The reason [Harry1 broke into the ministry] was PRO1 to save Sirius.
Unless there is some covert binder in the remote RatC, as in (30b), a possibility
I will examine in some detail and reject below, PRO is not in the same syntactic
domain as its controller and there can be no syntactic or semantic dependency
between the two. Since, as I will argue, there is no hidden binder in a remote RatC,
OC accounts of remote RatCs must be rejected. Instead, this must be a case of NOC.
PRO must be something like a free pronoun whose domain of reference is restricted
100
by non-grammatical factors such that Harry is PRO’s only possible antecedent in
(29). This will serve as additional evidence that at least some RatCs are NOC.
3.2.2.1 Specifier ellipsis
One way to attempt an OC account of control of remote RatCs is to posit elided
material in the specifier of the subject of the specificational sentence containing the
remote RatC as in (31).
(31) Jon sank the boat. [DP Jon1 [D′ The [NP reason]]] was PRO1 to collect the
insurance.
This is similar to Roeper’s (1987) theory of PRO in the subject of NP, which
he used to account for control out of nominals such as in (32). In a DP framework,
PRO would be the specifier of D. PRO being in this position is consistent with
Roeper’s judgment for (33); when the specifier of D is occupied by another element,
control out of the nominal is impossible.4
(32) the PRO1 destruction of the city [PRO1 to prove a point].
(Roeper, 1987, p. 280)
(33) *[DP [DP the city]’s [NP destruction [RatC PRO to prove a point]]
(adapted from Roeper, 1987, p. 280)
4 I disagree with this judgment. It is especially improved when placed in a sentence such as (i).
(i) [The city’s destruction (just) to prove a point] was cruel.
Whatever the status of (33) and (i), it does not directly effect whether there could be an elided
specifier of D.
101
The specifier ellipsis account would need to take this one step further, assuming
that instead of PRO, DPs can have other silent specifiers, such as the elided DP
in (31). The specifier-ellipsis account would put PRO in the same sentence as its
antecedent, which is crucial for any binding account of control (although I will show
that they are still not in the same syntactic domain). It also would potentially
provide a unified account of other instances of remote control. In the examples in
(34), the pronoun it could have an elided specifier as well. According to Elbourne
(2013), it is a determiner with a silent NP complement. Also including an elided
specifier would give the it of (34a), for example, the structure in (35).
(34) a. Jon1 made a promise. It was PRO1 to shave himself. (Dowty, 1989)
b. Sandy1 wanted something. It was PRO1 to leave the party.
c. Jon commanded Sara1 to do something. It was PRO1 to go home.
(35) [DP Jon [D′ It [NP promise]]]
However, there are at least three reasons to reject this specifier-ellipsis account.
The biggest problem is that it is unable to independently rule out control by the
object. As demonstrated above in (14a), repeated in (36), remote RatCs generally
do not allow control by the object of the target clause.
(36) Dobby1 protected Harry2. The goal was [PRO1/*2 to escape from the Mal-
foys].
But if all that is required for ellipsis of the specifier is identity with a DP antecedent,
102
the account provides no reason why the example in (36) should not be able to have
the structure in (37), with an elided specifier in identity with the object of the
preceding clause.
(37) *Dobby protected Harry1. [DP [DP Harry]1 The reason] was PRO1 to escape
from the Malfoys.
In order to explain the restrictions on which DPs can serve as antecedents for ellipsis,
this account would need to simply stipulate that the elided DP can only have as
its antecedent the (deep) subject of the target clause or the party explanatorily
responsible for the fact it expresses. Thus, while it would account for the referent of
PRO syntactically, it is really only moving the problem to a higher null category (the
elided specifier). Some other restrictions on which DPs could appear as antecedents
would be necessary, and it is not clear that specifier ellipsis would add anything
significant.
A second problem for the specifier ellipsis account is one of c-command. Even
though PRO and its antecedent are in the same sentence, the elided material in a
remote RatC such as (31) is embedded within the subject DP, as illustrated in (38),
and hence does not c-command PRO.
103
(38) XP
DP
Jon D′ was
The reason PRO to collect. . .
It is not clear that any syntactic binding relation can hold between a specifier
in the relevant position and anything in the complement of the copula. In (39), the
specifier Victor is unable to bind the reflexive himself. Even a reflexive in a picture-
NP, which generally can be good even without a c-commanding antecedent (Postal,
1971) cannot take the specifier of the subject as an antecedent (40) (although see
Pollard and Sag (1992) for arguments that picture-NPs do not require a syntactic
dependency in the first place).
(39) Victor’s heart attack was a benefit to him/??himself.
(40) Lin’s1 downfall was that image of him1/??himself1 on the wall.
Finally, the third problem is that ellipsis of a specifier is not obviously licensed.
The literature has discussed two requirements for ellipsis: an identity requirement
(Chomsky, 1965; Sag, 1976, a.o.), and a licensing head (Lobeck, 1995; Saito &
Murasugi, 1990). Identity with an antecedent is easily met (although as discussed
there must be additional restrictions on what can serve as the antecedent). In (31),
for example, the specifier is identical with the subject of the preceding clause. The
requirement for a licensing head runs into problems, though. According to Lobeck
(1995), the ellipsis site “must be properly head-governed, and governed by an X-
104
0 specified for strong agreement” (Lobeck, 1995, p. 41), where ‘strong agreement’
is defined as agreement that is morphologically realized in a productive number
of cases. Lobeck relies on Chomsky’s 1986 definitions for government and proper
government given in (41).
(41) a. Government
a governs b iff a m-commands b and there is no y, y a barrier for b,
such that y excludes a.
b. Proper Government
b is properly governed by a iff a theta-governs or antecedent governs
b
(i) Theta-Government
a theta-governs b if a is a zero-level category that theta-marks
b, and a, b are sisters
(ii) Antecedent-Government
a antecedent-governs b iff a is coindexed with b and no barrier
intervenes between a and b.
In order for the specifier to be elided under this approach, it would need to be prop-
erly head governed, i.e. it must be either theta-governed or antecedent-governed
by a head specified for strong agreement. There is no syntactically accessible an-
tecedent that could govern the specifier, and although D may be specified for strong
agreement with N, it head-governs but does not theta-mark its specifier, and the
105
two are not sisters. Therefore, the specifier is not properly head-governed, and
Lobeck’s requirements for ellipsis are not met. Even removing government as part
of the licensing definition does not help. Lobeck (1995) and Saito and Murasugi
(1990) agree that the agreeing functional head can only license the ellipsis of its
complement ; proper head government is only one way to accomplish this. Since the
proposed elided constituent is the specifier of D, the ellipsis is not licensed.
For these reasons, a specifier ellipsis account of remote control must be rejected.
3.2.2.2 Relative clause ellipsis
We have just seen that specifier ellipsis cannot account for remote control. Our
next attempt at a syntactic OC account will be relative clause ellipsis, illustrated in
(42). However, this attempt will also prove insufficient.
(42) Jon sank the boat. [DP The reason [RelC Jon1 sank the boat]] was PRO1 to
collect the insurance.
In (42), ellipsis of the relative clause would be under identity with the target clause
(although as will be discussed below, it is not clear that their surface identity is suf-
ficient). Through ellipsis, the remoteness of remote control would be removed. This
is true for both remote RatCs and other cases of remote control. Again assuming an
Elbourne-style account of it, but without the elided specifier discussed above, the
remote clause in (34a) would have the structure in (43).
(43) [DP It [NP promise Jon1 made]] was PRO1 to shave himself.
106
On the surface, this account seems to suffer from the same c-command problem
as specifier-ellipsis did. In (42), the elided antecedent is embedded in a relative clause
that does not c-command PRO. This is a challenge for the relative clause ellipsis
account, but it is one that is general to all specificational sentences (Higgins, 1973;
Heycock & Kroch, 1999). Although in each of the examples in (44) the reflexive is
not, on the surface, c-commanded by its antecedent, it still seems to be licensed, and
a non-reflexive pronoun in the same position is ungrammatical. Therefore, whatever
allows binding to occur in these examples may also explain the c-command problem
in (42).
(44) a. What Jon1 is is proud of himself1/*him1.
(Heycock & Kroch, 1999, p. 369)
b. The way Lee1 disturbs people is with images of herself1/*her1.
c. The thing Jon1 likes is that picture of himself1/*him1
There are several reasons to doubt a relative clause ellipsis account. The first
is that there are cases of remote RatCs where relative clause ellipsis will not help.
For example, when the reason is replaced with the goal, the relevant relative clause
cannot occur.
(45) a. Jon1 sank the boat. The goal was PRO1 to collect the insurance.
b. *The goal [(that) Jon sank the boat] was to collect the insurance.
It is possible that the underlying structure is something more like (46a), or perhaps
107
like (46b) (Maria Polinsky, p.c., based on ideas from Rizzi (1990)), but then the
material to be elided is no longer identical to its antecedent, and it is not clear that
its elision would be licensed, if identity with an antecedent is required. Even if we
suppose that this relative clause is close enough to identical with the target clause
to license ellipsis (see Ross (1969) for similar examples), binding out of this specifier
position is impossible, as (47) and (48) illustrate.
(46) a. Jon sank the boat. The goal [of Jon’s sinking the boat] was PRO to
collect the insurance.
b. Jon sank the boat. The goal [that Jon pursued/had in sinking the boat]
was PRO to collect the insurance.
(47) a. What Paul’s sinking the boat did was make him/*himself happy.
b. Paul’s sinking the boat made him/*himself happy.
(48) The goal [that Paul pursued/had in sinking the boat] seemed reasonable to
him/*himself.
In addition, a pronoun can take the place of the ellipsis site in examples such
as (49).
(49) Jon1 sank the boat. The reason/goal behind it was PRO1 to collect the
insurance.
Again, even assuming that pronouns can have unpronounced subjects (and other
108
content), binding out of the complement of a preposition in this position is impossible
(50). Therefore, there seems to be no place to put elided content to get the proper
configuration for syntactically controlling PRO.
(50) The best evidence for Lin’s improvement is that image of him/*himself on
the wall.
Furthermore, remote RatCs, just like local RatCs, do not require a syntacti-
cally expressed antecedent. In (14e), repeated in (51a), PRO may be understood as
whoever organized the event. In this and other cases of superimplicit control, the
relative clause ellipsis illustrated in (51b) will do nothing to allow the understood
antecedent to syntactically bind PRO.
(51) a. The ribbon was cut by a young girl. The goal was [PRO to acquire the
support of female voters].
b. The goal [of the ribbon’s being cut by a young girl] was PRO to acquire
the support of the female voters.
Finally, it is not clear that the relative clause ellipsis in (42) is licensed. Al-
though the elided relative clause is, on the surface, identical with a preceding clause,
they do differ structurally. Howard Lasnik (p.c.) notes that the structure of the
relative clause is likely more complicated than what is shown in (42). Following a
proposal by Larson (1987) dealing with temporal adjuncts, the relative clause would
109
be headed by a null operator that binds an empty category in the structure where
a RatC would be, as in (52).
(52) The reason [RelC Op1 [Jon [sank the boat] 1]]. . .
This is intended to capture the fact that (53) can either express the reason why
Mary said something, as with the continuation in (53a), or the reason behind the
fact she is reporting, as in (53b).
(53) The reason [CP1 Mary1 said [CP2 Jon2 sank the boat] was
a. PRO1 to get him arrested.
b. PRO2 to collect the insurance.
When the variable bound by the null operator is in the higher clause, the former
reading results, and when it is in the lower, it is the latter. Due to the presence
of the null operator and the variable in its structure, the elided relative clause is
not identical to its antecedent clause, and ellipsis may not be licensed. However,
as there is some debate over how strict the identity requirement for ellipsis is (see,
for example, Ross, 1969; Wyngaerd & Zwart, 1999; Merchant, 2001), the differences
here may be insignificant enough not to interfere with ellipsis.
Even so, the requirement for a licensing functional head is also not clearly met.
Ellipsis of a relative clause could possibly be licensed by N, which strongly agrees in
φ-features with (at least) the D head, and which may properly govern the relative
clause if we assume that N θ-marks it. However, this would be a strange assumption
110
to make, and ellipsis is generally only licensed by agreeing functional heads, which
N is not. Because of this, even though there are cases where relative clause ellipsis
could potentially explain remote control, such ellipsis is not licensed, and therefore
should be impossible.
The major objections outlined in this section against a relative clause ellip-
sis account have focused on the fact that there are instances of remote control of
RatCs that it cannot explain, such as (45a) and (49). Therefore, an independent,
non-grammatical account is already needed to explain at least these examples. Be-
cause this non-grammatical explanation is independently required, a simpler theory
would be to rely on it for the interpretation of PRO in all remote RatCs. Further-
more, even in instances where material in a relative clause could in principle be
syntactically connected to PRO, ellipsis of this relative clause does not appear to
be licensed according to established requirements (Lobeck, 1995; Saito & Murasugi,
1990). Therefore, an OC account of remote RatCs is untenable.
3.2.3 Summary of OC accounts of RatCs
This section has demonstrated that a strict OC theory of control is unable
to account for many RatCs. Although OC may explain structural restrictions on
PRO’s antecedent in local control, these same structural restrictions are apparent in
remote RatCs. In order to maintain an OC account of control in remote RatCs, it
would be necessary to posit some sort of elided material in order for the controller to
be in the same syntactic domain as PRO. But this is impossible. Even if such ellipsis
were licensed, the elided material would not be in a position to bind PRO. Remote
111
control of RatCs, then, cannot be OC. Instead, remote control must be NOC.
Furthermore, even local RatCs exhibit NOC in cases of implicit and superimplicit
control.
3.3 NOC accounts of RatCs
The previous section demonstrated that at least some RatCs, including remote
RatCs, must be NOC. Because the same structural restrictions on interpretation
seen in local RatCs appear to be active in remote RatCs, might it be the case that
all RatCs are NOC? In this section I introduce A. Williams’s (2015) Responsibility
Theory, which is in turn based on ideas from Farkas (1988). This theory attempts
to provide a unified NOC account of both local and remote control of RatCs. I will
begin with a summary of Farkas’s theory and its limitations. This will be followed
by a summary of Responsibility Theory and how it attempts to explain the possible
referents of PRO in RatCs, detailing the correct predictions it makes where Farkas’s
theory did not, and highlighting how it handles cases where OC accounts fail.
3.3.1 Farkas 1988
Farkas (1988) is one of several papers arguing against theories of control that
rely on syntactic binding. Instead, Farkas proposes that controller choice in OC
constructions is based on a small number of what she calls “semantic relations”.
The one that will be relevant here is the Responsibility relation (RESP). Farkas
takes as a premise the claim that certain situations in the world have what she
112
terms initiators, where an initiator is defined as an entity who brings the situation
about. More specifically, a situation may be the result of some act performed by
an initiator with the intention of bringing that situation about. Farkas’s basic
claim is that certain OC predicates have as part of their meaning a RESP-relation
that relates one of the arguments of the predicate to the situation described by
the embedded clause. That argument is then understood as the initiator of the
situation (the referent of the initiator must be left to the pragmatics). Farkas labels
these predicates as RESP-inducing. For example, the verb promise has as part of
its meaning the fact that the promiser is in the RESP-relation to the thing being
promised. In reporting Jon’s promise, (54a) states that Jon made himself responsible
for a situation of leaving the room (i.e. he makes himself a potential initiator of such
a situation).5 The verb persuade, on the other hand, carries as part of its meaning
the fact that the persuadee is in the RESP-relation with the embedded situation. In
reporting Jon’s persuasion of Bill, (54b) states that Bill has accepted responsibility
for a situation of leaving the room.
(54) a. Jon1 promised Bill2 [PRO1 to leave the room].
b. Jon1 persuaded Bill2 [PRO2 to leave the room].
PRO in OC constructions is linked to the initiator by a stipulated constraint
on reference which Farkas calls the Principle of Controller Choice (PCC). This
states that for RESP-inducing Vs, the controller of the infinitival complement is
5 See Searle (1965) on how to promise.
113
the argument linked to the initiator of the situation described by the complement.
Because of the RESP-relation in (54a), Jon is understood as the initiator of the
embedded situation, so PRO is understood to refer to him. In (54b), Bill is the
initiator, and so PRO refers to him.
In some marked cases, the PCC does not apply.6 In well-known cases of
control shift (Comrie, 1984) such as (55b), PRO does not refer to the initiator of
the situation, which is understood to be Jon, as it does in (55a). Instead, PRO
can be understood as Mary. Farkas claims that in (55b), a kind of coercion takes
place; the controller can be an individual with whom the initiator is in the “A-
relation,” where “A-relation” is defined by Farkas as existing between two entities
x and y, when x brings about the actions of y. Thus, although in general PRO in
the complement of a RESP-inducing predicate will be the initiator of the embedded
situation, it may also be an individual over whose (relevant) actions the initiator
has some control. Although Farkas does not explicitly state this, it seems to be the
case that the A-relation can only hold between co-arguments of the RESP-inducing
predicate. Were this not the case, (55b) could have a meaning where Jon promised
Mary that some other person would be allowed to leave.
(55) a. Jon1 promised Mary [PRO1 to leave early].
b. Jon promised Mary2 [PRO2/*3 to be allowed to leave early].
Finally, Farkas claims that if a situation has an intentional agent, that agent
6 Farkas does not elaborate on why these are marked.
114
will be understood either as the initiator of the situation, or as someone with whom
the initiator is in the A-relation. If a situation has no intentional agent, then it
will only be viewed as an intentional situation if it can be understood that an
individual brought the situation about intentionally. Because of this, and because
the RESP-relation requires an initiator in the embedded situation, complements of
RESP-inducing predicates such as promise and persuade must describe situations
that one could intentionally bring about. Because people generally do not have much
control over what they look like, (56) will ordinarily sound like it cannot possibly
be true.  promised 

(56) #Jon  persuaded Pete to resemble Bill.
Controller choice under this account is based on a constraint on PRO’s refer-
ent (the PCC) that states that PRO refers to the situation’s initiator. The initiator
is in turn determined by the semantics of the embedding predicate. Farkas intends
this account to also apply to RatCs. However, it is not immediately clear how this
is to be done. For complement control, the account states that the embedding pred-
icate determines which of its arguments is in the RESP-relation with the embedded
situation, i.e. which will be understood as the initiator. For RatCs such as in (57),
in order would induce a RESP-relation that specifies that there is an initiator of
the situation in the RatC, and the PCC would determine that PRO refers to it.
However, because the arguments of the matrix clause are not arguments of in order,
there would be no way for the initiator of the embedded situation to be determined
115
semantically the way it is in complement control. Farkas does not go into details
about how this would be accomplished, but it is clear that there would need to
be some sort of pragmatic restriction on the the potential referent of the initiator.
Farkas defines an initiator of a situation as someone that completed some action
with the intention of bringing about the situation. In (57), for example, that action
is given in the matrix clause. Because Dobby is the one who brought about that
action, he is understood as the initiator of the intended resulting situation, given in
the RatC.
(57) Dobby1 protected Harry2 in order PRO1/*2 to escape from the Malfoys.
This account runs into problems when the understood initiator of the situation
described in the RatC is not the one who performed the action in the target clause,
as is the case with the superimplicit control illustrated in (9a), repeated in (58).
(58) The ribbon was cut by a young girl [just PRO to acquire the support of
female voters].
Here, PRO would be tied to the initiator of the embedded situation. Using similar
reasoning as described for (57) above, the initiator should be whoever performed the
action described in the target clause, i.e. the young girl. However, as was discussed
earlier, PRO in (58) can be understood as whoever organized the ribbon-cutting
event. In other words, the initiator of the situation in the RatC is some unnamed
entity, and it is not clear how Farkas’s account would reach the correct interpretation.
116
3.3.2 Responsibility Theory
Responsibility Theory (A. Williams, 2015; see also A. Williams & Green,
2017), is a narrow adaptation of Farkas’s theory into a purely pragmatic account.
Instead of attempting to account for complement control, Williams proposes a con-
straint on the referent of PRO similar to Farkas’s PCC, but that only applies to
PRO in RatCs. According to Williams, PRO in a RatC will refer to the party
responsible for the target fact that the RatC is meant to explain (or to the target
fact itself). Williams does not require reference to the initiator of any situation; the
domain of reference for PRO is limited to the responsible party. As we will see, this
makes a correct prediction in (58), where Farkas’s theory struggled.
Let us now turn to how Responsibility Theory captures the control possibilities
discussed above. I will focus my discussion on local RatCs, but the same analysis
will apply to remote RatCs, since, as we will see, syntactic locality is irrelevant
to Responsibility Theory. Recall again that PRO in a RatC can refer to (i) the
surface subject of an active or passive target clause — but not the object or non-
c-commanding arguments (59a–c), (ii) what is understood as the demoted subject
of a passive (59d), (iii) other agents who can be understood as responsible for the
fact expressed by the matrix clause (59e) but only some of the time (59f), or (iv)
the target fact itself (59g).
(59) a. Dobby1 protected Harry2 [in order PRO1/*2 to escape from the Malfoys].
117
b. Sam1 was arrested by the police [just PRO1 to get himself off the
streets].
c. *David1’s wife2 brought some articles on reference along (in order) PRO1
to discuss semantics with her2.
d. The ship was sunk [in order PRO to collect the insurance].
e. The ribbon was cut by a young girl [in order PRO to acquire the support
of female voters].
f. A hired crook1 burned down the house [in order PRO1/*2 to collect the
insurance].
g. [Grass is green]1 [in order PRO1 to promote photosynthesis].
In the following, I will elaborate on how Responsibility Theory is able to explain
this data, even where OC accounts struggled. I then discuss two challenges Respon-
sibility Theory faces that provide evidence against any such NOC-exclusive account
of control of RatCs.
3.3.2.1 Responsibility Theory and subject control
In (59a), Dobby controls PRO. On the Responsibility Theory account, this is
because he is viewed as responsible for the target fact, namely that he protected
Harry. Other surface subjects can also be seen as responsible for the fact expressed
by the matrix clause. The patient of the arrest in (59b) is available as a controller
just in case Sam can be seen as being arrested intentionally. Putting it the other
way around, understanding Sam as the controller requires one to infer that he was
118
responsible for his being arrested, and that he intended his arrest to result in his
getting off the streets. If Sam did not aim to get off the streets by means of getting
himself arrested, then this use of (59b) would be false. Of course one might instead
view the police as primarily responsible for their arresting Sam. Accordingly, if the
police intended by performing the arrest to get Sam off the streets, PRO can also
refer to them, as in (60). For this reason Zubizarreta (1982) would refer to Sam as
a secondary agent.
(60) Sam was arrested by the police2 [just PRO2 to get him off the streets].
The ability to become a secondary agent is limited. Roeper (1987) noted that
secondary agents are limited to subject position. Even if (59b) is true, and Sam
intended to get off the streets by means of getting arrested, (61), where Sam is
named by the surface object rather than the subject, cannot be used to describe
this situation. If Sam’s ability to become a secondary agent is what allows PRO to
refer to him in (59b), it must be the case that the grammatical position of a DP
has an effect on how its referent is represented in relation to the situation described,
beyond its thematic relation to the event. It must be the case that non-subjects are
represented as not responsible (or at least not directly responsible). Because Sam
is named by the surface object in (61), A. Williams and Green (2017) claim that he
is represented as not responsible for what happened. When he is no longer referred
to in object position, as in (59b), the sentence can have an interpretation where
he is responsible. This is consistent with Zubizarreta’s observations about the use
119
of subject-oriented adverbs like intentionally. Compare the passive (62a), in which
intentionally can refer either to Mary’s intentions or to Joe’s, to the active (62b),
in which it can refer to Joe’s, but not Mary’s, who is in object position.
(61) * The police arrested Sam1 [just PRO1 to get himself off the streets].
(62) a. Mary was seduced intentionally (by Joe).
b. Joe seduced Mary intentionally. (Zubizarreta, 1982, p. 43)
Because the referents of objects are viewed as not responsible under Respon-
sibility Theory, object control is also ruled out in (59a). Harry cannot be seen as
responsible for the target fact when referred to in surface object position. Therefore,
in order for PRO to refer to Harry, he would have to be represented in a way that
allows him to be responsible for the fact expressed. As we have just seen, one way
to do this is to move Harry out of object position in a passive, as in (63). Although
a reading of the passive (63) with Harry as the controller is difficult, it is possible
to the extent that Harry can be viewed as responsible for his being protected by
Dobby.
(63) ?Harry2 was protected (by Dobby) [in order PRO2 to escape from the Mal-
foys].
According to A. Williams and Green (2017), Harry and Sam in (59a) and
(61), respectively, cannot be seen as responsible secondary agents because they are
represented in surface object position and thereby seen as not responsible. Because
120
of this, PRO cannot refer to them. The same reasoning can be used to rule-out
control by other non-c-commanding arguments. In (59c), David’s wife is represented
as more directly responsible for bringing the articles, and so PRO can only be
understood to refer to her, not to David. However, A. Williams and Green (2017)
point out that if it is explicitly stated that the referent of a non-subject argument
such as the object is responsible, then control by the object may become possible,
at least in remote control. This is illustrated in (64).
(64) Dobby was protecting Harry1, and this was an idea Harry had after he
learned about elf-magic from Hermione. The goal was PRO1 to escape from
the Malfoys.
In sum, Responsibility Theory can easily account for subject control RatCs.
PRO refers to the party responsible for the target clause, which in most cases will
be named by the subject rather than the object in actives. In passives, subjects
can be the controller because their referents can be viewed as secondary agents of
the target fact. Control by the object is ruled out in local control by postulating
that “the referent of a surface object is defeasibly depicted as nonresponsible” (A.
Williams & Green, 2017, p. 242).
3.3.2.2 Responsibility Theory and implicit control
The assumptions required for explicit control also allow for control by the
demoted subject of a passive (59d), repeated in (65). Here, PRO is understood as
121
whoever was responsible for the sinking of the ship. This account does not rely on
the syntactic presence of the demoted subject, although it is compatible with such
theories. Because the principles of Responsibility Theory will select as controller the
party that is explanatorily responsible for the fact expressed in the matrix clause,
it follows that the controller can be the agent of the matrix event, even if it is not
present in the syntax of the matrix clause.
(65) The ship was sunk [in order PRO to collect the insurance].
There are cases where control by the demoted subject of a passive is impossible.
For example, (66a) cannot mean that the ship was sunk so that the sinker would
become a hero (compare to (66b), in which that interpretation is available). This
example presents somewhat of a problem for Responsibility Theory; if the sinkers
are understood to be responsible for the target fact, PRO should be able to refer
to them. However, this is an even bigger problem for OC accounts, which would
need to explain why a structural relation such as the binding of PRO by the implicit
demoted subject is possible in (65) but not (66a). A. Williams (2015) notes that
for Responsibility Theory, it is sufficient to state that the unacceptability in (66a) is
due to conditions on what sorts of predicates make good RatCs (see Clark, 1990), or
to matters of perspective (see Sundaresan, 2016, on perspective and long-distance
anaphora), rather than just the situation described. Note that slightly changing the
predicate in the RatC improves it (67).
(66) a. #The ship was sunk in order to become a hero. (Lasnik, 1988, p. 12)
122
b. Jon1 sank the ship in order PRO1 to become a hero.
(67) The king’s ship was sunk just to acquire notoriety.
(A. Williams, 2015, p. 299)
Furthermore, A. Williams (2015) contrasts (68a) with (68b), arguing that the
latter is infelicitous because the killers are no longer responsible for the target fact.
At the very least, properties of the sheep are also in part responsible for the fact that
they are easily killed. Under grammatical theories, this contrast is surprising. There
is no reason why an adjunct like easily should block a syntactic relation between
PRO and the demoted subject.
(68) a. Several ewes were killed. And the reason was to survive the winter.
b. #Ewes are easily killed. And the reason is to survive the winter.
‘So that the killers might survive the winter’
(A. Williams, 2015, p. 299)
3.3.2.3 Responsibility Theory and superimplicit control
Because NOC accounts do not rely on the syntactic presence of the controller
of PRO, superimplicit control requires no new assumptions. In (59e), repeated in
(69), PRO can be understood as the organizer of the ribbon cutting because that
person can be seen as responsible for the target fact. This is not limited to passives.
In the active (70), a similar reading is more difficult, but still possible, especially
with stress placed on young girl.
123
(69) The ribbon was cut by a young girl [in order PRO to acquire the support
of female voters].
(70) A young girl cut the ribbon [in order PRO to acquire the support of female
voters].
However, the availability of superimplicit readings needs to be constrained. In
(59f), repeated in (71), PRO must be understood as the hired crook. It cannot refer
to the people who hired the crook, even though they are, in a sense, responsible for
the fact that the hired crook did what he did. The problem is further illustrated
in (72), which cannot mean that whoever was responsible for the fire was to collect
the insurance.
(71) A hired crook1 burned down the house [in order PRO1/*2 to collect the
insurance].
(72) #A fire burned down the house [in order PRO to collect the insurance].
Interestingly, when (71) is passivized, superimplicit control becomes available.
The passive in (73) can have a reading that the intended insurance collector is
someone who hired the crook to burn down the house.
(73) The house was burned down (by a hired crook) in order PRO to collect the
insurance.
Recall that A. Williams and Green (2017) explain the unavailability of ob-
124
ject control by postulating that surface objects are viewed as not responsible, while
subjects are likely to be seen as such. In (71), perhaps superimplicit control is
unavailable because its active form highlights the crook’s responsibility and down-
plays the responsibility of his employer. Because of this, the crook is the most
likely controller for PRO in the RatC under Responsibility Theory. The passive in
(73) deemphasizes his responsibility, and superimplicit control becomes available.
In (69) and (70), the little girl is more likely that the crook in (71) to be viewed as
an instrument instead of the responsible party, and superimplicit control is possible.
As for (72), superimplicit control is unavailable because the most natural reading is
simply one in which the fire is responsible for the house burning down, no person
being implicated.
3.3.2.4 Responsibility Theory and fact control
Finally, we have seen that PRO in a RatC can be controlled by the fact
expressed by the target clause. In (74a), repeated from (59g), one possible reading
is that it is the fact that grass is green that promotes photosynthesis. This seems to
be a problem for Responsibility Theory, since facts cannot be intentional, responsible
parties. For this reason, A. Williams (2015) simply stipulates that in a RatC, PRO
can be controlled either by the responsible party, or by the fact expressed by the
target clause, with no explanation of the difference between the two. However
stipulative it may be, this bifurcation seems necessary. According to (74a) for
example, why is grass green? Because its greenness promotes photosynthesis. An
alternative explanation could be that (74a) is just another example of superimplicit
125
control, where PRO refers to the entity responsible for grass being green, e.g. God,
giving the paraphrase in (74b). This, though, would leave us wondering why God’s
promotion of photosynthesis explains the fact that grass is green. One answer, of
course, is that the greenness of grass promotes photosynthesis. However, Alexander
Williams (p.c.) notes that (74a) does not leave that conclusion to our imagination.
With (74a) we say that photosynthesis is promoted by the greenness of grass.
(74) a. Grass is green in order PRO to promote photosynthesis.
b. 9 Grass is green so that God/the designer can promote photosynthesis.
Similarly, (75a) does not seem to have the meaning in (75b), even though the
people who designed the car are responsible for the fact that its sides are sloped.
Instead, it is the fact that the sides are sloped that reduces drag.
(75) a. The sides of the car are sloped in order PRO to reduce drag.
b. 9 The sides are sloped so that the designers might reduce drag.
Furthermore, suppose that the sinkers of the ship in (76) are responsible for
the target fact. Howard Lasnik (p.c.) notes that control by the responsible party in
this case would lead to a Condition C violation. Admittedly Condition C violations
can be weak (see (77), although this still seems worse than (76)). Still (76) does not
seem to mean that the ship was sunk so that the sinkers might enrich themselves.
Instead, it conveys the meaning that the ship was sunk so that the sinking might
enrich the sinkers.
126
(76) The ship was sunk in order PRO to enrich the sinkers.
(77) ?The sinkers1 just want PRO1 to enrich the sinkers1.
Inasmuch as fact control is possible, it requires an explanation. Notice that
it is still the case that the RatC expresses the teleological explanation of the target
fact. When there is a responsible party, as in (75a) and (76), it must have been the
intention of the responsible party that the target fact lead to the situation described
by the RatC. In this way, fact control is similar to control through the Farkas’s A-
relation. Fact control is possible if the party responsible for that fact, if there is one,
intended the result described in the RatC. However, this only holds when there is
a responsible party. In uttering (74a), a speaker conveys a teleological explanation
for the greenness of grass, but does not necessarily have to be committed to there
being an intentional party responsible for that fact.7
Although Responsibility Theory as described here requires some stipulation to
account for fact control, grammatical theories of control are also unable to account
for it without difficulty, as demonstrated above.
3.3.2.5 Responsibility Theory and remote control
Responsibility Theory as outlined for local RatCs applies equally well to re-
mote RatCs. Because no structural dependency is required between PRO and its
understood antecedent, it makes no difference whether the two are in the same
7 It is possible that a speaker of (74a) does imply that there is a responsible party, possibly
speaking in a metaphorical sense, without committing to the belief in one. If this is the case, then
even fact control of RatCs would entail a responsible party.
127
sentence; the same interpretive restrictions can be posited for both Remote and lo-
cal RatCs. Also, although Responsibility Theory does not make specific predictions
about whether there will be other connectivity effects between elements of the target
clause and the RatC in remote control, it is at least consistent with a lack thereof,
since it does not require positing any syntactic link between the two clauses.
3.3.3 Challenges for Responsibility Theory
Responsibility Theory posits that PRO in RatCs is NOC, and that its inter-
pretation is restricted by notions of responsibility, as summarized in (78).
(78) PRO in a RatC will refer to
i. the entity responsible for the target fact, or
ii. the target fact.
Subject control is predicted because subjects are likely to be seen as responsible
for the fact described by their host clause. This seems especially true when the
subject is an agent. The subject of a passive may also control PRO in a RatC
just in case it assumes a secondary agent role, in that they are understood to be
responsible for the fact expressed. In a similar manner, implicit, superimplicit, and
remote control are possible just in case the entity to whom PRO refers is understood
to be (directly) responsible for the target fact and the target fact does not highlight
the responsibility of some other argument. Responsibility Theory thus is able to
cover more of the data than an exclusively OC account.
128
However, Responsibility Theory, and all NOC accounts, face at least two chal-
lenges: how to rule out control by a surface object, and how to account for apparent
cases of OC by an inanimate subject.
3.3.3.1 Ruling out object control
In A. Williams and Green (2017), object control is ruled out only under the
assumption that surface objects are viewed as not (directly) responsible for the
fact that contains them. This is a necessary stipulation, but, as A. Williams and
Green (2017) acknowledge, it is unattractive because it posits a conventional effect
of sentence structure beyond those of information structure, and independent of
truth conditions. In (79), for example, Responsibility Theory requires that Sam
must be understood as not responsible for his arrest. What I mean here is not that
Sam did not cause his arrest. Instead, I mean that Sam must be understood as not
responsible for the arresting itself, although this difference is subtle.
(79) The police arrested Sam (in order to get him/*himself off the streets).
Although raising the object to subject position in a passive allows Sam to be
understood as responsible for the arresting (80), this is only possible when Sam is the
subject of a passive. In the active form in (79), Sam is understood as not (primarily)
responsible for the arrest. Even if it is explicitly stated that Sam initiated the arrest
and was responsible for it, as in (81), Responsibility Theory must state that Sam
cannot be understood as a responsible party in the relevant sense.
129
(80) Sam was arrested by the police (in order to get him/himself off the streets).
(81) *On Sam1’s order, the police arrested Sam1 [in order PRO1 to get himself
off the streets]. (Norbert Hornstein, p.c.)
There are at least two reasons why (81) should not allow control by the object under
Responsibility Theory. First, perhaps lexical agency (in active clauses) is more
important for RatCs looking for the responsible party. Because the police receives
an agent θ-role in (81), the sentence is highlighting the police’s responsibility, and
PRO therefore refers to them. But this same reasoning does not seem to apply
in (80), where it could be argued that the police is still receiving an agent θ-role.
However, recall the contrast between (71) and (73), repeated in (82). In (82b), but
not (82a), superimplicit control is available. This was argued to be the case because
the passive deemphasizes the crook’s responsibility, while the active highlights it.
Similarly in (80), even though the police may receive an agent θ-role, the passive
form of the sentence deemphasizes the police’s responsibility, and control by the
secondary agent Sam becomes available.
(82) a. A hired crook1 burned down the house [in order PRO1/*2 to collect the
insurance].
b. The house was burned down (by a hired crook) in order PRO to collect
the insurance.
Second, even without a preference to choose the grammatical agent as the
130
responsible party, it is still not clear that Sam in (81) can be seen as directly respon-
sible for the target fact, despite explicitly stating that he ordered the police to arrest
him and is thereby indirectly responsible. A similar phenomenon has been observed
for verbs of causation (see Pietroski, 2005; A. Williams, 2015, for discussion). For
example, assume that Jon puts butter in the microwave and presses start, and does
nothing else. Although Jon does not directly cause the butter to melt (it is the mi-
crowave that does), (83) is still considered true. However, if Jon forces some other
person to put the butter in the microwave and push start, (83) becomes infelicitous.
Because there is some other (human) agent who is more directly responsible for
what caused the butter to melt, their role interferes with notions of Jon’s causation.
Similarly in (81), it may be true that Sam caused the police to arrest him, but their
responsibility for the arrest is more direct. Because of this, it is less surprising that
PRO in the RatC would refer to the police rather than Sam under Responsibility
Theory.
(83) Jon melted some butter.
As seen in (64), repeated in (84), control by the object may be more acceptable
in remote control. A. Williams and Green (2017) explain that this is because objects
are represented as defeasibly nonresponsible. When further elaboration is given,
object control becomes available.
(84) Dobby was protecting Harry1, and this was an idea Harry had after he
131
learned about elf-magic from Hermione. The goal was PRO1 to escape from
the Malfoys.
However this explanation does not fully predict the available contrasts between
local and remote control. As seen in (85) and (86), even when the elaboration
establishing the responsibility of the object is given prior to the target fact, control
by the object is only available in remote control, not in local control.8
(85) Because of an idea Harry had after learning about elf magic, Dobby pro-
tected Harry1. . .
a. *. . . in order PRO1 to escape the Malfoys.
b. The goal was PRO1 to escape the Malfoys.
(86) On his order, the police arrested Sam1. . .
a. *. . . in order PRO1 to get himself1 off the streets.
b. The goal was PRO1 to get himself off the streets.
Note that it is still necessary to maintain that agentive subjects are depicted as
more responsible than patient objects, since even in remote control, subject control
is preferred without supporting context establishing the object as responsible. How-
8 The remote control examples are less acceptable when the subject of the copular sentence is the
reason instead of the goal (i). This is an interesting puzzle, but an exploration of the distinctions
between the various kinds of sentences that can host a remote RatC is outside the scope of this
dissertation.
(i) a. *Because of an idea Harry had after learning about elf magic, Dobby protected Harry1.
The reason was PRO1 to escape from the Malfoys.
b. *On his order, the police arrested Sam1. The reason was PRO1 to get himself off the
streets.
132
ever, responsibility fails to explain why this preference is defeasible only in remote
control.
3.3.3.2 Cases of OC in RatCs
The second challenge to Responsibility Theory and any other NOC theory of
control in RatCs is that it is possible for an inanimate, non-responsible subject in
the matrix clause to serve as controller (Landau, 2017). Because NOC requires PRO
to have a [+human] interpretation, the examples in (87) must be instances of OC.
(87) a. This book1 was written [in order PRO1 to be read]. (Williams, 1992)
b. The house1 was emptied [in order PRO1 to be demolished].
(Español-Echevarŕıa, 2000)
c. The poem1 was written by Hughes [in order PRO1 to be admired by
silly girls]. (Hirschbühler, 1988)
d. A test wire chamber module1 was built [in order PRO1 to be inserted
into the cryostat]. (Landau, 2017)
e. The car1 is streamlined [in order PRO1 to accelerate faster].
(Whelpton, 2002)
f. The battleship1 is near the coast [in order PRO1 to shell the harbor].
(Whelpton, 2002)
Although he rejects NOC accounts of RatCs, Whelpton (2002) proposes an
explanation for examples such as (87) that may negate their problematic status for
133
Responsibility Theory. He suggests that the adjuncts in (87) are not actually RatCs,
but subject-gap purpose clauses (SPCs), similar to the one in (88). Although Jones
(1991) uses the presence of in order as a diagnostic for RatCs, Whelpton argues
that in order is a general marker of subject-oriented control related to readings of
intention and necessity.
(88) They sent the journalist1 into the arena [e1 to photograph the crowds].
(Whelpton, 2002, p. 193)
In an SPC, the empty subject of the adjunct is bound by the object of a
transitive clause, as illustrated in (88) (Jones, 1991; Whelpton, 2002). However,
Whelpton argues that it can also be bound by the subject of a copular clause. Part
of the evidence he gives involves the ambiguity of (89). Under one reading, the
journalist took the independent decision to be in the arena for the purpose stated.
Under the other, the journalist was sent to the arena for that purpose. The former
reading would be given a RatC analysis and the latter an SPC one.
(89) The journalist1 was in the arena [(in order) e1 to photograph the crowds].
(Whelpton, 2002, p. 194)
However, although this explanation is tempting, it faces problems. Whelpton
(2002) argues that in order may only preface an SPC when the clause is adjoined to
a copular sentence. If this is correct, then the adjuncts in (87) adjoined to passive
matrix clauses cannot be SPCs. Even when the sentence structure is changed to
134
pattern more like canonical SPCs, as in (90), some of the examples still do not allow
SPCs, giving further evidence that the corresponding examples in (87) do contain
RatCs controlled by an inanimate subject, not SPCs.9
(90) a. *Harry wrote this book1 [e1 to be read].
b. *Workers emptied the house [e1 to be demolished].
c. *Hughes wrote the poem1 [e1 to be admired by silly girls].
Because there is good evidence that OC by an inanimate subject is possible
in at least some RatCs, any NOC-exclusive theory, including Responsibility Theory,
will fail to account for all the data.
3.4 Both OC and NOC are possible for RatCs
The previous two sections have demonstrated that neither strict OC nor strict
NOC theories of control can adequately account for control in RatCs. The natural
conclusion is that Landau (2017) is correct in stating that RatCs, just like temporal
adjuncts, allow both OC and NOC. This solves many of the problems that strict
OC and strict NOC accounts face. OC accounts were unable to account for remote,
implicit, or superimplicit control, and NOC accounts were unable to explain control
by inanimate subjects. If both OC and NOC are possible, then this is not a problem.
9 This does not preclude the possibility that some SPCs may be headed by in order. For example,
putting (87d) into a more canonical SPC structure is acceptable (i). This suggests that at least
(87d) may contain an SPC.
(i) The engineer built a test wire chamber module1 [e1 to be inserted into the cryostat].
135
However, as was the case for temporal adjuncts, it is not the case that both types of
control will always be available for a given sentence. The remainder of this section
will first discuss the distribution of OC and NOC in general. This will be followed
by a proposal on how to rule out object control in local RatCs while allowing it in
remote control.
3.4.1 The distribution of OC and NOC in RatCs
Both OC and NOC are possible for RatCs, but just as with temporal adjuncts,
it is not the case that both are equally available in all contexts. Here I discuss some
factors that may influence or determine which variety of control will occur.
3.4.1.1 Remote control is always NOC
I argued in §3.2.2 that remote control cannot involve a syntactic OC depen-
dency. If this is true, then all remote RatCs should exhibit NOC. This leads to the
prediction that remote versions of the examples in (87) with control by an inani-
mate subject should be ungrammatical. This prediction is borne out. In each of the
examples in (91), PRO cannot be understood to corefer with the inanimate subject
of the target clause, despite the fact that control by such a subject is possible in
local RatCs.10
(91) a. *This book1 was written. The reason/goal/purpose was PRO1 to be
read.
10 Examples (91c,e,f) may be acceptable under an implicit or superimplicit control reading, with
PRO referring to Hughes, the driver of the car, or the people manning the ship, respectively.
Crucially, though, they do not allow control by the inanimate subject of the target clause.
136
b. *The house1 was emptied. The reason/goal/purpose was PRO1 to be
demolished.
c. *The poem1 was written by Hughes. The reason/goal/purpose was
PRO1 to be admired by silly girls.
d. *A test wire chamber module1 was built. The reason/goal/purpose was
PRO1 to be inserted into the cryostat.
e. *The car1 is streamlined. The reason/goal/purpose is PRO1 to acceler-
ate faster.11
f. *The battleship1 is near the coast. The reason/goal/purpose is PRO1
to shell the harbor.
3.4.1.2 Local subject control
The case is not as clear with local RatCs, which can allow either OC or NOC.
Control by the matrix subject in local RatCs could often be either OC, its interpre-
tation being determined by syntactic binding, or NOC, the interpretation arising
based on principles such as those in Responsibility Theory. For example, in (92),
OC would result in Harry being the controller, since it is in the correct position
to syntactically bind PRO. NOC would result in the same interpretation under
Responsibility Theory, since Harry is the agent and will be understood as the one
most directly responsible for the target fact.
11 Interestingly, fact control appears to be an exception. Although facts are not animate, remote
fact control is possible, as with (i) (Alexander Williams, p.c.). See also §3.5.
(i) The car is streamlined. The point is PRO to reduce drag.
137
(92) Harry1 stayed up late with Hermione in order [PRO1 to master the sum-
moning charm].
Two tests for OC suggest that local subject control is OC. First, only sloppy
identity under ellipsis seems possible.
(93) Harry1 stayed up late in order PRO1 to master the summoning charm, and
Hermione2 did too.
‘And Hermione stayed up late in order PRO2/*1 to master the summoning
charm.’
However, this test does not provide sure evidence of OC, since Responsibility Theory
may also predict sloppy identity for NOC of RatCs. Under Responsibility Theory,
PRO will refer to the party responsible for the target fact. In the second conjunct,
the clause to which the elided RatC adjoins is ‘Hermione worked hard every day,’
and this is the most likely candidate for the target fact. Therefore, PRO will still
refer to Hermione, because as the agent of the second conjunct, she is the one who
will be understood as most directly responsible for that fact.
The second test may be more reliable. PRO in a local RatC can only have
a bound interpretation when controlled by only-DPs. Example (94) remains true
in a situation where Hermione stayed up late in order to help Harry master the
summoning charm, but it is false if Ron also stayed up in order to master it himself.
(94) Only Harry1 stayed up late in order PRO1 to master the summoning charm.
138
‘[Nobody else]2 stayed up late in order PRO2/*1 to master the summoning
charm.’
In addition, an OC analysis of local subject control in RatCs would be con-
sistent with the proposed preference for bound versus free variables from Reinhart
(1983) discussed in §2.4.2.2.
3.4.1.3 Semantic anomaly and local NOC
It is clear that cases of implicit and superimplicit control must be NOC, but
there are restrictions on when they can occur. For example, as was the case for
temporal adjuncts, it appears that when both an OC and an NOC derivation are
possible, OC is preferred. Landau (2017) states this even more strongly, arguing
NOC in RatCs, will only be available when OC by the matrix subject is blocked,
such as when it would lead to “semantic anomaly.” For example, we have seen that
superimplicit control is available in (95). According to Landau, this is only possible
because OC by the matrix subject would make the sentence denote an unlikely
situation: ribbons cannot normally acquire support.
(95) The ribbon was cut by a young girl [in order PRO to acquire the support
of female voters].
In contrast, in (96), it is perfectly plausible that a hired crook could collect insurance.
Therefore OC by the matrix subject is preferred, and superimplicit NOC is blocked.
139
(96) A hired crook1 burned down the house [in order PRO1/*2 to collect the
insurance].
However, this explanation is probably not completely correct. After all, as was
seen in Chapter 1, Kehler et al. (2008) demonstrated that people are much more
likely to resolve the pronoun in (97) to the subject of a passive, even when it would
lead to a very improbable story. Given this strong preference in addition to the
overall preference for OC over NOC, it seems that comprehenders should strongly
prefer subject control in (95), with the interpretation that the ribbon is the one
winning votes, even though this interpretation results in a very strange story.
(97) The demonstrators were denied a permit by the city council because. . .
a. . . . they feared violence.
b. . . . they advocated violence.
Even ignoring this problem, there are cases of implicit control where OC by
the matrix subject would not lead to semantic anomaly. For example, in (98), even
though there is nothing semantically or pragmatically anomalous about young girls
acquiring support, (98) still allows a superimplicit control reading where it is the
organizers of the event acquiring support. But perhaps here, the young girl is either
acquiring support for the organizers, or she can be seen as an instrument of the
organizers unable to acquire support on her own.
140
(98) A young girl cut the ribbon [in order PRO to acquire the support of female
voters].
A stronger example is (99). Here, the guests would make a perfectly good
controller. It is not at all implausible that the guests would acquire the support of
female voters, and this reading is in fact available if the guests are seen as responsible
for how they were greeted. It is less likely that the guests would be seen as acquiring
support on behalf of the event organizers. If semantic anomaly of OC were required
for NOC to be possible, then the preferred interpretation of (99) should be that the
guests intended to acquire support. Instead, (99) has an interpretation parallel to
(95), where the intended support-acquirers are the organizers of the event.
(99) The guests were greeted by a young girl in order PRO to acquire the support
of female voters. (Maria Polinsky (p.c.))
As was the case with temporal adjuncts, it appears that subject control leading
to an unlikely situation may reduce the preference for OC, but this is not strictly
required for NOC to occur.
3.4.2 Ruling out local object control
One of the biggest challenges for Responsibility Theory is ruling out object
control in local RatCs. A. Williams and Green’s (2017) attempt was to suggest that
surface objects are depicted as defeasibly nonresponsible. Although it may be true
that subjects (especially agents) are seen as more responsible than object patients
141
given the preference for subject control even in remote RatCs, this explanation fails
to predict differences in the availability of object control in local and remote RatCs,
as was seen in §3.3.3.1 with the examples repeated in (100) and (101). Whereas
object control never seems to be available in local RatCs, it is possible in remote
RatCs if it is explicitly stated that the object is responsible.
(100) Because of an idea Harry had after learning about elf magic, Dobby pro-
tected Harry1. . .
a. *. . . in order PRO1 to escape the Malfoys.
b. The goal was PRO1 to escape the Malfoys.
(101) On his order, the police arrested Sam1. . .
a. *. . . in order PRO1 to get himself1 off the streets.
b. The goal was PRO1 to get himself off the streets.
Why is the constraint against object control so strong in local RatCs? Object
control is ruled out in a strict OC account, because the object does not c-command
the RatC and can therefore not syntactically bind its PRO. But given that NOC is
also possible, what prevents NOC by the matrix object in local RatCs, given that it
is possible in remote control? One possibility is that object control in local RatCs
is ruled out simply due to the general preference for OC over NOC. In (100a), for
example, Dobby is in a position to syntactically control PRO, and it is perfectly
plausible that he would be the one escaping the Malfoys and that his helping Harry
142
could lead to that result. Because this OC reading is available and plausible, NOC
by the object is blocked. In remote control, the OC preference is not at play, since
OC is impossible. Therefore, object control is possible in (100b).
This explanation predicts that not only object control, but all NOC readings
will be ruled out in (100a,101a). This seems to be correct. Even if the speaker
highlights someone else’s responsibility in the matter, the only available reading
seems to be OC by the subject in a local RatC (102a). But in a remote RatC,
control by that third party is possible (102b).
(102) Because his owner1 compelled him to, Dobby2 protected Harry3. . .
a. . . . in order PRO2/*1/*3 to escape the Malfoys.
b. The goal was PRO1/2/*3 to escape from the Malfoys.
This explanation further predicts that if a speaker will accept superimplicit
control in a local RatC adjoined to an active, transitive target clause, then NOC
by the matrix object should also be possible. However, superimplicit control with
transitive matrix clauses is difficult, given that the subject of the matrix clause is
usually an agent and will therefore be understood as responsible. One example that
may validate the prediction is (103). Inasmuch as a speaker allows the superimplicit
control in (103a), they should also allow NOC by the object in (103b).
(103) a. At Trump1’s request, a multilingual sign greeted the guests each morn-
ing in order PRO1 to acquire the support of the immigrant population.
143
b. At his request, a multilingual sign greeted Trump1 each morning in
order PRO1 to practice connecting with the immigrant population.
Object control is therefore ruled out any time NOC is. When the RatC is
obligatorily controlled, then only the subject of the matrix clause is in a position to
bind it. Even when NOC is possible, as is always the case for remote control, object
control may be dispreferred if the subject is an intentional agent that is understood
to be more directly responsible for the target fact. But, if it is made clear that the
object is responsible, then object control will be possible anywhere NOC is.
3.4.2.1 OC required with passive RatCs?
Landau (2017) presents what he terms the Jaeggli-Roeper Generalization,
given in (104).
(104) The Jaeggli-Roeper Generalization (Landau, 2017)
Active RatC allows OC or NOC; passive RatC imposes OC.
This generalization is based on the fact that inanimate subjects can control PRO
in a passive RatC (see (87)), but in examples such as (105)–(107), implicit control,
which has been demonstrated to be NOC, is impossible.
(105) *The report was carefully prepared [PRO to be congratulated by the board
of directors]. (Jaeggli, 1986)
(106) a. A vote was taken [PRO to elect a president].
144
b. *A vote was taken [PRO to be elected president].
c. John1 took a vote [PRO1 to be elected president]. (Roeper, 1987)
(107) a. John1 sank the ship [PRO1 to be promoted].
b. *The ship was sunk [PRO to be promoted]. (Lasnik, 1988)
Landau argued that this same generalization applies to temporal adjunct con-
trol. Although I argued in Chapter 2 that the generalization does not always hold
for temporal adjuncts, I believe that it holds for RatCs. Although Landau leaves the
explanation of this generalization as an unsolved puzzle, Español-Echevarŕıa (2000)
attempts to provide an explanation. Español-Echevarŕıa, who was the first to my
knowledge to propose a hybrid OC/NOC account of RatCs, claims that a passive
RatC disallows NOC due to principles governing controller choice in NOC. Fol-
lowing Farkas (1988), Español-Echevarŕıa states that under NOC interpretations,
the understood antecedent to PRO in a RatC is chosen based on a Principle of
Controller Choice in RatCs (PCCRC), as given in (108).
(108) PCCRC : (Español-Echevarŕıa, 2000, p. 107)
For any s and s′, such that s corresponds to the matrix situation and s′ to
the purposive one in a purpose construction, the controller of the purposive
infinitival clause is an individual i such that:
i. both RESP(i, s) and RESP(i, s′) hold, and
ii. i must bear an agentive role from the rationale predicate.
145
For the examples in (105), (106b), and (107b), because the RatC is a passive,
PRO neither receives an agentive θ-role, nor can it viewed as responsible for the
embedded situation. Because of this, NOC is blocked, and OC is the only option.12
Additional evidence for this explanation is given in (109). Even though the RatC is
not in passive voice, NOC by the implicit agent is still blocked, because the predicate
become rich does not assign an agentive role to its subject.
(109) a. John wrote this book [in order PRO to become rich].
b. *This book was written [in order PRO to become rich].
(Español-Echevarŕıa, 2000)
This explanation is similar to Landau’s (2018) [init] theory discussed in §2.3.2.2,
since the [init] feature is closely tied, though not identical, to agency. Under Lan-
dau’s account, (109b) requires OC because the matrix subject (the potential OC
controller) and PRO match in [-init] features.
But there are at least three reasons why such explanations cannot be correct.
First, an agent θ-role for PRO is not required for NOC. Nor is it necessary for there
to be a mismatch in [init] values. In (110), for example, PRO does not receive an
agentive θ-role, and both it and the matrix subject are [-init], and yet NOC appears
to be possible.
(110) The king’s ship[-init] was sunk just PRO[-init] to acquire notoriety.
12 Español-Echevarŕıa gives a movement account of OC, following Hornstein (1999).
146
Second, it is not clear why PRO in a passive RatC could not be understood as
responsible for the situation as a secondary agent, thereby becoming [+init], akin
to Sam in the example repeated in (111). In (106b), for example, why could the
person who took the vote not be understood as somehow responsible for their being
elected president, thus becoming [+init]? If that were the case, then PRO in (112)
would mismatch in [init] values with the potential OC controller, as illustrated in
(112). Contrary to fact, NOC should be possible in such a situation.
(111) Sam[+init] was arrested by the police [just PRO[+init] to get himself off the
streets].
(112) *A vote[-init] was taken [PRO[+init] to be elected president].
Finally, if the restriction against NOC in passive RatCs was due to the re-
quirements of NOC interpretation alone, as Español-Echevarŕıa (2000), though not
Landau (2018), maintains, then we would expect remote control for passive RatCs
to be impossible. However, NOC is much more acceptable in remote control.
(113) a. The report was carefully prepared. The goal was [PRO to be congrat-
ulated by the board of directors].
b. A vote was taken. The goal was [PRO to be elected president].
c. The ship was sunk. The goal was [PRO to be promoted].
Perhaps the restriction against NOC in passive local RatCs is not due to
a single factor, but because of a combination of separate interpretive preferences.
147
Among these could be a preference for NOC PRO in RatCs to have an agentive role
as well as a general preference for OC over NOC. In addition, it was seen that a
passive temporal adjunct also is less likely to allow NOC, and there may be some
other interpretive constraint disfavoring NOC in passive adjuncts generally. These
preferences are compounded in local RatCs to give a strong restriction against NOC
when the RatC is passive. However, in remote control, OC is impossible, and
the preference for it is void. In such cases, because NOC is the only option, the
preference for PRO in NOC to be an agent can be ignored.
This type of explanation would also explain why NOC with passive temporal
adjuncts is possible, even if passive temporal adjuncts more frequently exhibit OC.
Because there is no responsibility requirement, the constraint against NOC is less
than with RatCs, and therefore is more easily overridden.
3.5 Conclusion
This chapter has provided evidence that control in RatCs can be either OC
or NOC, as was argued to be the case for temporal adjuncts in Chapter 2. In OC,
which can only apply in local RatCs, the controller is the main clause subject in
most cases, but may also possibly be an event or situation variable in cases of fact
control if it is in fact OC. NOC occurs in all instances of remote control, where there
can be no syntactic dependency between PRO and arguments of the target clause.
In addition, NOC may occur in local RatCs. This is most frequently seen in cases of
implicit control, when PRO is understood to refer to the implicit agent of a passive
148
target clause. NOC is also possible in cases of what I have called superimplicit
control, where PRO is understood to refer to some person who is not an argument
of the target clause, whether explicit or implicit.
As is the case for temporal adjuncts, grammatical constraints and syntactic
structure alone cannot fully account for when NOC will be possible in local RatCs.
Instead, the distribution of NOC must be due at least in part to processing or
interpretive preferences or biases. These are likely to be similar to the preferences
responsible for the distribution of NOC in temporal adjuncts.
The referent of PRO in NOC with RatCs is determined based on who is un-
derstood to be explanatorily responsible for the fact expressed by the target clause.
This differs from NOC PRO in temporal adjuncts, where PRO had to refer to a
logophoric perspective holder, but did not induce notions of responsibility. This
difference likely arises from the meaning of the Rat head (sometimes pronounced in
order in local RatCs), which may impose additional restrictions on the domain of
PRO’s reference (A. Williams & Green, 2017). This seems likely, given that even
RatCs exhibiting OC seem to implicate some sort of responsible party or “initiator”
(Landau, 2017, 2018).
One question that remains is how to deal with fact control. As discussed
above, in local RatCs, one might initially think that fact control is OC by either an
event or a situation variable in the syntax. However, fact control is also possible
in remote RatCs, such as in (114).13 Given that remote control can only be NOC,
syntactic control by an event or situation variable is impossible in such cases.
13 Also see Footnote 11.
149
(114) [Grass is green]1. The reason is [PRO1 to promote photosynthesis].
Responsibility Theory has little to say about how fact control arises, other to say
that PRO in a RatC can refer either to the party responsible for the target fact, or
to the target fact itself. Although this kind of disjunctive observation is somewhat
unsatisfactory, I leave further explanation to later research.
150
Chapter 4: Control theories and other adjuncts
The previous two chapters have presented evidence that adjunct control can
neither be classified as strictly OC nor as strictly NOC. Instead, I have agreed with
Landau (2017) that at least temporal and rationale adjuncts are hybrid environ-
ments, where both OC and NOC may occur. In contrast to Landau, though, I have
argued that NOC does not require that OC be ruled out semantically or otherwise;
NOC is not simply a last-resort option arising when OC fails. Instead, the grammar
freely allows both OC and NOC in these adjuncts. Still, OC is strongly preferred,
and this is likely due to performance biases. Factors contributing to the choice of
OC and NOC may include conceptual plausibility and structural features of the
sentence. In addition, NOC interpretation requires that the understood referent of
PRO be a logophoric perspective holder, and in the case of RatCs, that this referent
be understood as explanatorily responsible for the fact the RatC is meant to explain.
This chapter investigates some of the broader implications of the previous two
chapters. In §4.1 I review and expand on the way Landau (2017) implements this
duality of control for temporal and rationale adjuncts in the Two-tiered Theory of
Control (TTC) (Landau, 2015). As will be seen, his account leads to the prediction
that all cases of adjunct control, not just temporal and rationale adjuncts, should
151
allow both OC and NOC given the right context. Section 4.2 therefore briefly
discusses several other types of adjuncts, illustrating that this is true for some,
but not others. I then discuss possible explanations for why certain adjuncts are
exceptions, highlighting problem cases. In §4.3, I propose an alternative explanation
within the framework of the Movement Theory of Control (MTC) (Hornstein, 1999;
Boeckx et al., 2010) that can better account for many strict OC adjuncts that the
TTC struggles with. Section 4.4 concludes the theory portion of the dissertation.
4.1 The two-tiered theory of control
In order to accommodate the data described in the previous chapters, any
theory of control should be able to explain at least three things: (i) why OC is
preferred to NOC in these adjuncts, but is not strictly required; (ii) why complement
control does not allow the same duality, being instead strictly OC; and (iii) how the
domain of reference for PRO is restricted in NOC contexts (i.e. why perspective or
responsibility, etc., are necessary). After a brief summary of the mechanics of the
TTC, I will discuss how this theory answers these questions.
4.1.1 Background on the TTC
The TTC rejects traditional formulations of OC and NOC. Instead, it classifies
control as either predicative or logophoric, similar to the way Williams (1992) does.
In complements, logophoric control occurs with attitude predicates, and predicative
with non-attitude predicates. The term “predicative control” is used to describe the
152
situation when a verb denotes a relation R between an individual x and a property p
that entails R∗(x, p(x)), where R∗ expresses essentially the same concept as R. For
example, “x managed p” is true iff MANAGE*(x, p(x)). This kind of predicative
relation is said to apply in both complements and adjuncts, resulting in OC for
both, although as will be discussed, Landau does not make it clear how exactly
this works for adjuncts. Logophoric control, on the other hand, which appears only
with attitude predicates, involves a relation between x and p(i), where i is a variable
restricted to the author or addressee of the context of evaluation. In complement
control, this context is restricted to include only the matrix attitude holder of and
their addressee, and the variable is associated with either of them, resulting in OC.
In adjuncts, the variable i is not bound any argument of the matrix clause, and the
context of evaluation is not restricted to any specific context, which results in NOC.
These ideas are implemented as follows. In both predicative and logophoric
control, PRO is held to be a “minimal pronoun” with unvalued φ-features that,
depending on the derivation, can surface as controlled PRO, a reflexive, a bound
lexical pronoun, a resumptive pronoun, or other types of pronouns. Syntactically,
this minimal pronoun will inherit whatever φ-features its binder is specified for. And
semantically, it is vacuous. That is, even relative to an assignment function it would
not by itself be assigned a value. In both predicative and logophoric control, a non-
finite TP containing this minimal pronoun is the complement of a predication head
(Fin), as illustrated in (1). This Fin has an uninterpretable [uD] feature, which
attracts PRO to its specifier position. Because movement automatically causes
λ-abstraction (Heim & Kratzer, 1998), as PRO raises to spec-FinP, it creates a
153
λ-predicate. But, due to its radical impoverishment and lack of semantic content,
Landau states that upon movement, PRO does not saturate the lambda-predicate.
The embedded FinP thus itself denotes a property; in (1), it is the predicate λx. x
stay healthy.
(1) a. PRO to stay healthy
b. FinP = λx. x stay healthy
PRO ′2 Fin
Fin[uD] TP
{D,φ:}= PRO2 T′
to vP
stay healthy
In predicative complement control, illustrated for (2) in (3), the FinP merges
directly with the matrix verb. Landau states that the property denoted by the
FinP is then “applied” directly to an argument of the matrix clause. But Landau
cannot mean this literally, since the FinP is not sister to any argument in the matrix
clause. So he may mean for the λ to be ‘passed up the tree’ and then ‘identified’
with the subject argument of the matrix VP, so that JVPK = λx. V(x, JFinPK(x)).
This could be achieved by repeated Function Composition, or by raising PRO to
the level of the matrix VP, plus some rule that identifies the ‘raised’ argument with
one from the matrix. But then Landau’s theory becomes nearly identical to the
Movement Theory of Control. So perhaps Landau is instead using “applied” in a
154
less mechanical sense. Perhaps he means only that the relation expressed by the
matrix verb, between the referents of its object and its subject, property p and
individual s, entails a further relation between s and the proposition that p(s). For
example, perhaps Ron managed to be polite expresses a manage-relation between
Ron and being polite, but then this relation holds only if Ron also has a certain
relation to the proposition that he is polite. In that case, Landau’s “f applies to x”
would just mean that the sentence entails Q(fx) for a certain Q. And this is itself
is a variant of what is sometimes called the semantic theory of complement control,
articulated in Sag and Pollard (1991) and its precedents, and briefly discussed in
§2.5.2. Predicative object control is similar, with the FinP being “applied” as a
small clause predicate to the matrix object.
(2) John managed to stay healthy.
(3) Derivation of predicative subject control (see Landau, 2015, p. 26)
TP
John1 T
′
T vP
John1 v
′
v VP
managed v tV FinP
PRO to stay healthy
In logophoric control, which occurs with attitude verbs, the predicate denoted
155
by the FinP applies—this time in the literal sense, with the predicate combining with
its sister as a function applying to an argument—to an intermediate null pronoun
that is bound by the matrix controller. There is also additional structure in the CP
of the embedded clause that is intended to capture the de se properties of OC PRO
in logophoric control. This is illustrated for the sentence (4) in (5).
(4) John intends to visit Athens.
(5) Derivation of logophoric control (modified from Landau, 2015, p. 44)
TP
John T′
λx T′
T vP
John v′x
v VP
intends CPQ
λG7 CP
GP C′
G′ i′ C[uD]-i′ FinP
G7 pro PRO to visit Athensx[D,φ:]
presup:G7=Gself
variable binding
Following Schlenker (2003) and others, Landau (2015) claims that attitude
verbs act as quantifiers over sets of contexts, where a context is a tuple of coordinates
156
i = <x,y,TIME(i),WORLD(i)>, where x and y are variables corresponding to the
author and addressee of the context. Landau refers to these coordinates of the
context as “arguments” of C, in an extension of the term’s usual sense. Furthermore,
C is said to have an uninterpretable [uD] feature, which is satisfied by the presence
of the null pronoun in its specifier. G7 in (5) is a “concept generator” (Percus &
Sauerland, 2003) that takes pro as an argument and entails that the attitude holder
is in some acquaintance relation with the “res” (individual) that will be the referent
of pro. The meaning of the embedding verb determines that the relevant attitude
holder is the one in the matrix clause, but does not specify what the acquaintance
relation is. This meaning comes as a presupposition triggered by the embedded C
that the acquaintance relation G7 is either the self relation for verbs like intend,
which have an obligatory de se interpretation, or the thou relation for verbs like
ask which exhibit obligatory de te. This presupposition guarantees that pro will
correspond either to the author or addressee of the matrix attitude predicate,
leading to either subject or object control. In this indirect way, the sentence in (5)
comes to entail that John as known to himself has the property of visiting Athens
in all worlds compatible with John’s intentions in the actual world.
In sum, complement control consists of a FinP denoting a property being
associated either with the matrix controller in predicative control, or with an inter-
mediate pronoun which is bound by a matrix argument in logophoric control. Which
structure the embedded clause takes is determined based on the specific embedding
verb.
Landau (2015, 2017) states that adjunct control also falls into these two cat-
157
egories. But because adjuncts are not selected, they are free to adjoin with either
a predicative structure or with a logophoric one. Predicative structures give rise to
what I have been calling OC readings. But if Landau is assuming a theory of pred-
ication similar to semantic theories of control, where the embedding verb mediates
the predication, then it is unclear how this would apply to adjuncts. Assuming that
the predication relation does somehow obtain, then the antecedent of PRO will be
limited to the closest c-commanding DP. If the adjunct attaches with a logophoric
structure, on the other hand, then there is no restriction on the G7 relation or the
embedded context, and PRO will refer to the author or addressee of any relevant
context, resulting in logophoric NOC.
4.1.2 Why OC is preferred over NOC
As illustrated succinctly in (6), predicative (OC) adjuncts have a simpler struc-
ture than logophoric (NOC) ones.
(6) a. OC adjunct: [PP before/in order [FinP Fin [TP. . . ]]]
b. NOC adjunct: [PP before/in order [CP C+log [FinP Fin [TP. . . ]]]]
(Landau, 2017, p. 100)
Following Grimshaw (1994), Bošković (1996), and Speas (2006), Landau (2017)
proposes a principle of economy of projection, which favors the derivation with a
less complex structure, all else being equal, resulting in the more complex NOC
structure in (6b) only being licensed if the OC structure in (6a) produces “semantic
158
deviance.” As was seen above, semantic anomaly of the OC derivation is not always
necessary for NOC to be available. Therefore, although the preference for simpler
structure may be correct, it must be the case that factors other than semantic
anomaly can override that preference.
4.1.3 Why complement control only allows OC
Under the TTC, adjunct control can either be predicative or logophoric. In
complement control, which type of control structure is used is determined based on
selection by the embedding predicate. OC is ensured in non-attitude complements
because the embedded clause is simply a predicate whose argument is instantiated by
the closest c-commanding argument of the matrix clause. NOC is ruled out because
the predication relation must be satisfied locally. Logophoric NOC is also ruled
out in non-attitude contexts because of the selectional restrictions of the embedding
verb.1
Attitude predicates, on the other hand, select a logophoric complement and
set the context of evaluation of the clausal complement as the attitude context of
the matrix event. This ensures that the controller will be either the author or the
addressee of the matrix event. It is important for Landau that the C is required
to project either of these variables to its specifier to check its [uD] feature, but
1 NOC is actually possible in complements with a wh-phrase in their spec-CP (see (i)). Note
that (i) involves an attitude predicate, and therefore would involve logophoric control. Although
Landau (2015) does not address this question, perhaps this has to do with the fact that wh-
infinitives always have a modal interpretation, which in turn leads to generic interpretation of
PRO (Bhatt, 1999). How this is instantiated in the TTC, though, is not clear.
(i) Jon asked [*(how) PRO1 to shave oneself1].
159
which variable is projected is not lexically specified. This ensures that the property
denoted by the FinP can apply to either the author or the addressee of the matrix
verb, allowing for well-known cases of control shift such as in (7). In (7a), pro is
bound by the addressee of the matrix attitude, and in (7b), by the author.
However, this is still an instance of OC, since PRO in either case must be controlled
by an argument of the matrix clause.2
(7) a. Jim asked Maryy [CP proy [FinP PRO to leave]].
b. Jimx asked Mary [CP prox [FinP PRO to be allowed to leave]].
Adjuncts differ from complements only in that they are not selected. Because
of this, they are free to have either a predicative or a logophoric structure. With
predicative adjuncts, PRO’s λ is “applied” to the closest c-commanding argument,
which for temporal and rationale clauses is the matrix subject. However, there is
still a remaining question of how this is done. Here there is no mediating verb,
and a semantic approach where the predication is mediated by the relation between
the two clauses is unlikely, as there is nothing at all in the meaning of while, for
example, that suggests that the two concurrent events should share a participant,
much less one named by the subjects of the clauses.3 So it must be the case that
the λ is passed up the tree, either through movement or through a complex series
2 Although control shift is possible, there generally is a default controller based on the control
verb. For example, persuade generally exhibits object control, and promise subject control, and
it takes extra context to get control shift. Landau’s (2015) theory provides no explanation of the
default.
3 Rothstein (2004) stipulates of depictive secondary predicates, which have a semantics like
while-adjuncts, that they combine with their hosts by a semantic rule that enforces exactly this
sort of sharing. Nonetheless, this stipulation is not one that follows from the concept of temporal
concurrence.
160
of function composition. Putting aside this question, and assuming that the λ does
somehow “apply” to the matrix subject, it will result in OC due to the nature of
predication. Logophoric adjuncts will be NOC because there is no restriction on the
context of evaluation of the embedded clause. The null pronoun in the specifier of
C that binds PRO is therefore free to refer to the author or addressee of any
relevant context.
4.1.4 The interpretation of NOC
NOC in adjuncts is always logophoric under the TTC; predicative control nec-
essarily results in OC. This provides a direct explanation of why PRO’s understood
referent must be some sort of perspective holder in the discourse. PRO will be bound
by either the author or the addressee of the embedded context. But again, since
this context is not selected by anything in the matrix clause, what context of evalu-
ation these author and addressee pertain to may vary given the right discourse
context.
4.2 Other adjuncts
The previous chapters discussed temporal and rationale adjuncts, using them
as evidence that both OC and NOC are possible in adjunct control. However, there
are several different types of adjuncts that exhibit control, each with their own dis-
tinct properties (see Huettner, 1989; Landau, 2013). Therefore, although temporal
and rationale adjuncts may allow both OC and NOC, this does not immediately
161
mean that all adjuncts will likewise. However, the TTC would seem to predict that
this duality of control should extend to all adjuncts, although Landau (2015, 2017)
does not explicitly make this claim. If temporal and rationale adjuncts allow both
OC and NOC because their structure is unselected by a matrix verb, then the same
should be true of all unselected adjuncts.
This section briefly discusses several types of adjuncts. Some of these will be
shown to follow the pattern of rationale and temporal adjuncts, allowing both OC
and NOC. Others seem to be restricted to one or the other. For some of these, I
will propose a possible explanation for why they differ from other adjuncts within
the framework of the TTC, but others present a genuine problem. The following
section, §4.3, proposes an alternative account within the framework of the MTC.
4.2.1 Adjuncts allowing both OC and NOC
The TTC predicts that all adjuncts should exhibit the same duality of control
as temporal and rationale adjuncts, allowing both predicative OC and logophoric
NOC. Here I discuss three adjuncts that confirm this prediction: the response
clause, the telic clause, and the absolutive clause. Although I discuss the first in
some detail, I will only give preliminary evidence for the others.
4.2.1.1 Response clauses
One adjunct that has been largely ignored in the literature is illustrated in (8).
These adjuncts have been called reason clauses (Betancort, Carreiras, & Acuña-
Fariña, 2006), rationale clauses (Hornstein, 1999), justification clauses (Stromdahl,
162
2018), or simply gerundival complements to for (A. Williams & Green, 2017). For
clarity, and to distinguish them from RatCs, which have also been called reason
clauses (A. Williams, 2015), I will call the adjuncts in (8) “response clauses.” The
reason is that these adjuncts entail that the act of the main clause was meant as a
response to what is described by the adjunct (A. Williams & Green, 2017; see also
Fillmore, 1971; Fabricius-Hansen & Sæbø, 2011).
(8) a. Harry1 got a trophy [for PRO1 winning the tournament].
b. Mickey baked Minnie1 a cake [for PRO1 doing so well on the test].
c. Mickey hugged Minnie1 [for PRO1 doing so well on the test].
Some responses clauses may be understood as conceptual “arguments” of what
I will call blame-class verbs, such as blame, praise, and thank, as in (9). The meaning
of the verb blame, for example, requires that there is an either explicit or implicit
offense associated with the blame (Fillmore, 1971). This is not the case with the
response clauses in (8). Although their conceptual relationship to the main clause
predicate may differ, I will assume for now that the response clauses in either case
have the same syntactic status.
(9) a. Harry blamed Snape1 [for PRO1 breaking in].
b. Hermione praised Harry1 [for PRO1 being such a great wizard].
c. Harry thanked Ron1 [for PRO1 teaching him to play chess].
PRO in response clauses can be coreferential with the matrix subject, as in
163
(8a), or with the indirect or direct object, as in (8b–c) and each of the examples
in (9). Response clauses also allow control shift, similar to what was seen for com-
plements in (7). This is illustrated in (10) and (11). Example (10a) prefers object
control, while (10b) exhibits subject control. And in (11), either subject or object
control is possible, as long as Harry views Malfoy as ultimately responsible.
(10) a. The principal thanked Sara1 [for PRO1 chaperoning the dance].
b. Sara1 thanked the principal [for PRO1 being allowed to chaperone the
dance].
(11) Harry1 blamed Malfoy2 [for PRO1/2 losing the match].
Control shift is also possible for response clauses not associated with a blame-class
verb (A. Williams & Green, 2017). In (12), PRO may be coreferential with either
the main clause subject or object, depending on whether the flowers are viewed as
an apology, in which case the subject controls PRO, or as a reward, in which case
the controller is the object.
(12) Jack1 brought Sharon2 flowers [PRO1/2 for working so late].
The fact that PRO is able to be controlled by either argument of the matrix
clause gives our first piece of evidence that NOC is possible. Under the TTC, control
shift is only possible in logophoric control, which allows pro to correspond either to
the author or to the addressee of the relevant context. In predicative control for
both complements and adjuncts, PRO’s controller will always be the next highest
164
c-commanding DP. Variable or shifting control should be impossible in predicative
control, unless, of course, the adjunct has variable attachment sites.
There is, in fact, some evidence that response clauses have multiple possible
attachment sites, and that attachment height can interact with controller choice.
First, in (13), if the adjunct provides an explanation for the baking of the cake,
then Minnie is the controller. But if it is an explanation for why Mickey had/got
to do some activity, then Mickey is the controller. Therefore, it could be the case
that response clauses exhibit OC by the object when the adjunct attaches low, and
OC by the subject when it attaches higher.
(13) Mickey1 had/got to bake Minnie2 a cake for PRO1/2 doing so well on the
test.
It is standardly assumed that the phrase do so replaces a VP, and that ele-
ments that cannot be stranded after do so are selected arguments of the verb (see,
e.g. Whelpton, 1995; DeArmond & Hedberg, 1998). For example, in (14), ‘in the
living/bedroom’ is an argument of the verb put, and therefore cannot be stranded;
it must be included in the interpretation of did so (underlining marks the intended
content of do so).
(14) a. *Mildred put three books in the living room, and Dwight did so in the
bedroom.
b. Mildred put three books in the living room, and Dwight did so, too.
165
VP adjuncts, on the other hand, can optionally be stranded. Since they Chomsky-
adjoin to the VP, as in (15), do so can either replace the lower VP, leaving the
adjunct stranded, or the higher VP, in which case the adjunct will be included in
the elided content. In (16), ‘in the living room/bedroom’ is a VP adjunct and can
therefore optionally strand.
(15) Chomsky-adjunction
VP
VP Adjunct
V (DP)
(16) a. Mildred read three books in the living room, and Dwight did so in the
bedroom.
b. Mildred read three books in the living room, and Dwight did so, too.
Based on the do so test, response clauses sometimes appear to behave as verbal
arguments. In (17), do so seems to obligatorily include the response clause.
(17) a. ??Jill sent Justin to his room for spilling water everywhere, and later Pe-
ter did so for refusing to clean up.
b. Jill sent Justin to his room for spilling water everywhere, and later Pe-
ter did so, too.
Response clauses in other instances appear to act as VP adjuncts. In contrast to
(17), which indicated that the response clause was within the minimal VP (i.e. that
166
it was behaving like a verbal argument), in (18) the do so test passes whether the
response clause is included in the elided content or not.
(18) a. Jack1 brought Sharon flowers [for PRO1 working so late], and David2
did so [for PRO2 leaving work early].
b. Jack1 brought Sharon flowers [for PRO1 working so late], and David2
did so, too.
It is tempting to conclude from (13), (17), and (18) that object-controlled re-
sponse clauses are always situated lower in the tree than subject-controlled response
clauses, and that therefore place of attachment can determine controller choice. Such
a hypothesis would be consistent with response clauses being predicative OC. How-
ever, such a conclusion would be premature. As (19) and (20) make plain, even
object-controlled response clauses sometimes behave as VP adjuncts with respect
to the do so-test.
(19) Jack brought Sharon1 flowers [for PRO1 working so late],
a. and David did so [for PRO1 finishing the project early].
b. and David did so, too.
(20) The car was always getting dirty and it was Billy’s job to wash it. Last
month his mom washed it for him1 [for PRO1 working so hard in school]
a. and yesterday his dad did so, too.
b. and yesterday his dad did so [for PRO1 passing his biology exam].
167
Even response clauses associated with blame-class verbs sometimes behave
as VP adjuncts, not as arguments, whether PRO is controlled by the main clause
object, as in (21), or whether control shifts to the subject, as in (22).
(21) Peter thanked/praised Justin1 [for PRO1 eating all of his dinner],
a. and Jill did so [for PRO1 cleaning up after himself].
b. and Jill did so, too.
(22) Sara1 thanked the principal [for PRO1 being allowed to go to the dance],
a. and Tyler2 did so [for PRO2 being allowed to read in the library in-
stead].
b. and Tyler2 did so, too.
Interestingly, both (18) and (22) require sloppy identity under ellipsis, which is
suggestive of OC. Landau (2017) claims that OC in adjuncts is always predicative.
Therefore, at least these response clauses must be exhibiting predicative OC. But
once again, if response clauses were strictly predicative OC, then we would expect
attachment site to strictly determine controller, which is not the case.
Further evidence that response clauses cannot be strictly predicative OC is
the fact that controller choice is not limited to c-commanding arguments. First,
response clauses allow implicit control, which we have seen is NOC. In (23a), PRO
may refer to the implicit recipient of the medal. Example (23b) shows the same
168
thing, but uses a pronoun to remove the possibility that the response clause adjoins
within the object DP.4
(23) a. They awarded a medal [for PRO winning the contest].
b. They awarded it [for winning the contest].
Second, in at least some contexts, response clauses allow non-commanding con-
trollers, as (24) demonstrates. If control of response clauses were strictly predicative,
then there would be no way for the embedded possessor to control PRO.
(24) a. I left them in [her1 locker] [for PRO1 being so kind to me].
(Stromdahl, 2018)
b. I put roses on the front porch of her1 house [for PRO1 being so kind to
me].
The fact that sloppy identity was required under ellipsis (see (18) and (22))
is evidence for OC in at least some response clauses, but we have seen that this
cannot be strict predicative OC. It is also in principle possible under the TTC
4 Note that implicit control is not always possible for response clauses. In (i), PRO requires an
explicit antecedent (Alexander Williams, p.c.).
(i) The dog barked *(at someone1) [for PRO1 approaching the house too quickly].
I will not attempt an explanation of when implicit control will be available, but note that implicit
control was also not always available for RatCs (compare (ii) to (iii)) or for temporal adjuncts
((iv) vs. (v)).
(ii) The ship was sunk [PRO to collect the insurance].
(iii) *The ship was sunk [PRO to become a hero].
(iv) Those lemons tasted especially sour [after PRO eating sugar].
(v) *My iPad was stolen while [PRO distracting me].
169
for logophoric control to be OC, but even if control in response clauses were strict
logophoric OC, the account would face several problems. First, this would require
the adjuncts to be selected by some element of the matrix clause. Under the TTC,
this is the only way for the author and addressee of the embedded context to be
fixed as the matrix attitude holder and their audience. Although one could argue
that this is the case for blame-class predicates, it cannot be the case that all response
clauses are selected. Furthermore, (21) and (22) provide evidence that even for the
blame-class, response clauses may behave syntactically as adjuncts.
A second major problem for positing strict logophoric OC in response clauses
is illustrated in (25). PRO in response clauses allows inanimate controllers. Under
the TTC, this can only be predicative control; logophoric control requires animate
controllers, since only animate entities can be logophoric centers of the context.
(25) a. The book1 was praised [for PRO1 showing how, and under what cir-
cumstances, a religion grows].5
b. I included this book1 in the book fair [for PRO1 being so well written].
Because there is evidence that control in response clauses must sometimes be
predicative and sometimes logophoric (assuming a TTC framework), we are left with
the conclusion that response clauses allow either, consistent with the generalization
that adjuncts allow both OC and NOC.
I conclude my discussion of response clauses with a brief note about the effects
5 https://en.wikipedia.org/wiki/Jane Shaw
170
of attachment site. We saw above that the height of the response clause can affect
controller choice. This may be in part because it determines which predicative
controllers are possible, but it also may have an effect on the interpretation of
logophoric control. It may be the case that attachment site is not determining a
syntactic controller, but what the coherence relation between the two clauses will be.
If the adjunct in (13), repeated in (26) is attached high, then it will be understood
to be giving an explanation for why Mickey got to do something. If it attaches
low, then it will be an explanation for why the cake was baked for Minnie. These
coherence relations would then determine who the understood referent of PRO is,
similar to the way coherence can determine interpretation of overt pronouns (Hobbs,
1979; Kehler et al., 2008).
(26) Mickey1 had/got to bake Minnie2 a cake for PRO1/2 doing so well on the
test.
This explanation is even more likely given the fact that overt pronouns have the
same interpretive restrictions as PRO in response clauses (A. Williams & Green,
2017), as seen in (27).
(27) a. #Abe1 married Sue for PRO1/his1 writing nice poems.
(A. Williams & Green, 2017, p. 242)
b. Sue married Abe1 for PRO1/his1 writing nice poems.
(A. Williams & Green, 2017, p. 242)
c. Jack1 brought Sharon2 flowers for PRO1/2/his1/her2 working so late.
171
4.2.1.2 Object-gap purpose clauses
Another adjunct that allows both OC and NOC is the object-gap purpose
clause (OPC). Purpose clauses (PCs) contain one obligatory gap that is coreferen-
tial with the theme argument of the matrix clause (Jones, 1991; Kawasaki, 1993;
Whelpton, 1995). In OPCs, the obligatory gap is in the object position of the
embedded clause. The subject may contain an additional empty category (PRO),
which is generally coreferential with the indirect object of the main clause, if one is
present (28a), or otherwise with the main clause subject (28b). Lexical subjects are
also possible (28c).6
(28) a. Chandler bought Monica1 a new frying pan2 [PRO1 to cook with t2].
b. Monica1 bought a new frying pan2 [PRO1 to cook with t2].
c. Chandler bought a new frying pan2 [for Monica to cook with t2].
Several authors have claimed that the dependency between the PRO subject of
an OPC and its antecedent is OC (e.g. Nishigauchi, 1984; Kawasaki, 1993; Landau,
2013; Hallman, 2015). Evidence in support of OC being possible is that only sloppy
identity is available under ellipsis in (29a), which can only be used to say that
Chandler got his new frying pan with the intent of cooking with it himself. Also,
PRO can only have a bound interpretation when controlled by an only-DP. Example
6 Many authors agree that the object gap in an OPC is the result of empty operator movement
to the CP, illustrated in (i) (Chomsky, 1980; Williams, 1980, 1992; Whelpton, 2002).
(i) [CP Op1 [TP PRO to. . . t1. . . ]]
For simplicity, I do not include this operator in the examples in the main text.
172
(29b) entails that no one else bought a frying pan to cook with themselves; it is still
true in a situation where both Monica and Chandler bought frying pans intended
for Monica’s cooking.
(29) a. Monica1 bought a new frying pan [PRO1 to cook with t ], and Chandler3
did, too.
b. Only Monica1 bought a new frying pan [PRO1 to cook with t ].
But there is also strong evidence that NOC is sometimes possible. For example,
the PRO subject of an OPC can have a non-c-commanding or syntactically absent
controller ((30a) and (30b), respectively).
(30) a. I left it1 in [her2 mailbox] [PRO2 to look over t1 once she returned from
the Bahamas]. (modified from Stromdahl, 2018)
b. They1’re kept in the overhead compartment [PRO to use t1 in case of
an emergency]. (modified from Stromdahl, 2018)
OPCs therefore allow both OC and NOC, similar to other adjuncts seen.
4.2.1.3 Absolutives
A final adjunct that appears to allow both OC and NOC is the absolutive
clause. These adjuncts adjoin as bare participial phrases, as in (31). Pires (2007)
argues that absolutive adjuncts are OC, based on many of the tests used to classify
173
temporal adjuncts as such (see §2.2). For example, in (31a), PRO’s controller must
c-command it.
(31) a. [Peter2’s daughter]1 went on to college, [PRO1/*2 being the best student
in the class]. (Pires, 2007, p. 197)
b. [PRO1 Having come this far], I1 can’t go back.
However, non-c-commanding controllers are possible given the right context.
In (32), PRO can refer to the speaker. This is especially true when the adjunct
is left-adjoined, but it is also possible, if slightly more marginal, when it is right-
adjoined. Absolutive clauses therefore also appear to conform to the generalization
that adjunct control allows both OC and NOC.
(32) a. [PRO Sitting in class the day after the party], my eyes refused to stay
open.
b. ?My eyes refused to stay open, [PRO sitting in class the day after the
party].
4.2.2 Adjuncts with only NOC: Speaker-oriented adverbials
The previous adjuncts pattern with temporal and rationale adjuncts, but there
may be some that differ, in seeming contrast to the predictions of the TTC. Here I
discuss speaker-oriented adverbials, which seem to be restricted to NOC. Speaker-
oriented adverbials represent the speaker’s evaluation of the proposition denoted by
the matrix clause or the speaker’s degree of certainty about that proposition’s truth
174
(Ernst, 2009). These adverbials can surface as an adverb, as in (33a), an infinitive
verb phrase, as in (33b), or a participial phrase, as in (33c).
(33) a. Honestly, Jon would be better off without Mary.
b. [PRO To be honest], Jon would be better off without Mary.
c. [PRO Judging from experience], Jon will be better off without Mary.
PRO in (33b,c) refers to the speaker of the utterance. It cannot be coreferential
with the main clause subject, unless that subject also refers to the speaker. These
adjuncts must therefore be NOC. The same interpretation obtains if the adjuncts
are adjoined to the right.
(34) a. Jon would be better off without Mary, honestly.
b. Jon would be better off without Mary, PRO to be honest.
c. Jon will be better off without Mary, PRO judging from experience..
According to Cinque (1999), evaluative and evidential adverbs attach high
in the structure, to functional projections well above the TP. If speaker-oriented
adverbials containing control structures also attach high, then no argument of the
sentence would be in a c-commanding position to bind PRO, and OC by an argu-
ment of the main clause would be impossible. Under the TTC, the adjunct could
in principle adjoin with either a predicative or a logophoric structure, but only lo-
gophoric NOC would lead to a convergent derivation. With a predicative structure,
the predicate would have no argument to be applied to and would remain unsatu-
175
rated. These adjuncts therefore do not pose a serious challenge to the predictions
of the TTC.
4.2.3 Adjuncts with only OC
There also appear to be some adjuncts that are strictly OC. This again seems
to contrast with the predictions of the TTC. Under the TTC, strict OC is only
predicted if the structure and/or context of evaluation of the adjunct is selected by
some element of the matrix clause. Otherwise, it should be able to freely adjoin with
either a predicative or a logophoric structure with a free context of evaluation. If,
however, what appears to be an adjunct is actually a(n optionally) selected argument
of the matrix verb, then strict OC would be possible. This indeed seems to be the
case for some OC “adjuncts.” Others, however, present a genuine challenge to the
TTC, as they appear to display strict OC, and yet it is highly unlikely that they
are selected by any element of the matrix clause.
4.2.3.1 Goal clauses
One adjunct that may be strictly OC is the goal clause, illustrated in (35)
(Landau, 2013). To the best of my knowledge, these adjuncts do not allow NOC.
(35) a. Max1 works hard [PRO1 to stay out of jail]. (Landau, 2013, p. 221)
b. Max1 works hard [PRO1 to avoid doing any work].
There is some evidence that goal clauses adjoin within the VP. In (36), the
goal clause must be included in a do so anaphor. The intended interpretation of
176
(36) is that Max puts great effort into avoiding confrontation, and Billy puts great
effort into avoiding exerting himself. Note that this is distinct from the similar (37),
which contains a RatC and has the meaning that Max literally works hard, and so
does Billy. Unlike the goal clause in (36), the RatC in (37) adjoins outside the VP,
and can therefore be excluded from a do so anaphor’s interpretation.
(36) *Max works hard [to avoid confrontation] and Billy does so [to avoid exerting
himself].
(37) Max works hard (in order) to earn a living, and Billy does so (in order) to
use up excess energy.
It is not implausible that the goal clause adjoins within the VP because it is
selected as an (optional) argument of the matrix verb. Goal clauses are compatible
with only a small number of matrix predicates (such as work hard and labor dili-
gently), and the interpretation is similar to cases generally analyzed as complement
control, such as (38).
(38) Max1 tried hard [PRO1 to stay out of jail].
If goal clauses are arguments of the matrix verb, merged within the VP, then the
TTC could allow strict OC. The main clause predicate would select the goal clause
with a predicative structure, being a non-attitude predicate. This is in contrast to
unselected adjuncts adjoining higher in the structure, like temporal and rationale
adjuncts, which can adjoin with either a predicative or logophoric structure..
177
4.2.3.2 Result clauses and stimulus clauses
Two other adjuncts that may be strictly OC are result clauses (39) and stim-
ulus clauses (40) (Huettner, 1989; Landau, 2013).
(39) a. The damp1 seeped in [PRO1 to chill our bones].
(Huettner, 1989, p. 26)
b. Harry1 grew up [PRO1 to be a famous wizard].
c. Ron1 awoke [PRO1 to find the fire had gone out].
(40) a. Ginny1 smiled/shuddered [PRO1 to think what a fool she had been].
b. Hermione1 blushed [PRO1 to recall her flirtations with Ron].
c. Fudge1 gnashed his teeth [PRO1 to think of Sirius’s escape].
If these really are strictly OC, the TTC would have to say that they, too, are not
really adjuncts, but (optional) selected arguments of the main verb. Intuitively, this
seems less likely for these clauses than for goal clauses. But like goal clauses, result
and stimulus clauses must be included in the interpretation of a do so anaphor,
which generally is taken as an indication of argumenthood.
(41) a. *Harry grew up [to be a famous wizard], and Dudley did so [to be a nice
guy].
b. *Ginny smiled [to think what a fool she had been], and Hermione did so
[to think how smart she had been].
178
c. *Ginny shuddered [to think what a fool she had been], and Hermione
did so [to think how silly she must have looked].
However, Whelpton (1995) suggests that do so does not target the verb and its
arguments directly, but simply the VP node, and furthermore, that some adjuncts
may Chomsky-adjoin to the VP, as in (42a), but, crucially, that other adjuncts may
adjoin to a V′ level, as in (42b). Adjuncts Chomsky-adjoined to the VP may be
optionally included in the interpretation of do so, as the elided constituent could
be either the lower or higher VP. V′-adjoined adjuncts, on the other hand, would
obligatorily be included in do so, despite their being adjuncts.7
(42) a. Chomsky adjunction b. V′ adjunction
VP VP
VP Adjunct Spec V′
Spec V′ V′ Adjunct
V V
If result or stimulus clauses are true adjuncts adjoined with the structure in
(42b), then the TTC would have no explanation for why they exhibit strict OC.
Despite their being located within the VP, they would be unselected by the verb,
and there should be no restriction on their adjoining with either a predicative or a
logophoric structure.
7 A similar distinction could also be made even in Bare Phrase Structure (Chomsky, 1995; see
also Hornstein, 2008), since in (42a), the adjunct adjoins higher than the specifier, and in (42b),
it is lower.
179
4.2.3.3 Subject-gap purpose clauses
Subject-gap purpose clauses (SPCs), such as the one in (43), are similar to
OPCs, but the obligatory gap coreferring with the theme of the matrix clause is in
subject position, and there are no other gaps.8
(43) Harry brought Hermione1 along [PRO1 to deal with the security guards].
SPCs allow inanimate controllers, in contrast to OPCs, which is indicative of OC.
(44) a. I bought the shelf1 [PRO1 to hold books]. (Williams, 1992, p. 317)
b. I bought this blender1 [PRO1 to help me make split pea soup].
(Landau, 2015, p. 22)
Like the OC adjuncts discussed above, SPCs adjoin within the minimal VP (Jones,
1991; Whelpton, 1995), as illustrated in (45).
(45) VP (Whelpton, 1995, p. 107)
Spec V′
V′ Purpose Clause
V NP
Once again, it seems unlikely that all SPCs are selected by some element of
8 Many discussions of SPCs (e.g. Jones, 1991; Whelpton, 2002) maintain that the empty category
in an SPC is an operator that raises to Spec-CP, just like the object gap in an OPC (see Footnote 6).
I continue to use PRO here for convenience, but neither the MTC nor the TTC posit that PRO is
a distinct syntactic entity. Under the TTC, it is in fact something like an operator-variable pair,
once it raises to Spec-FinP. Under the MTC, OC PRO will be a trace of movement, and according
to Nunes (2001), the same is true of variables bound by a null operators.
180
the matrix clause, and control in SPCs under the TTC is predicted to be either
OC if the adjunct is adjoined with a predicative structure or NOC if a logophoric
structure is used instead. It is clear that OC is possible, but I have no evidence that
NOC is as well. If they are indeed strictly OC, then the TTC has no satisfactory
explanation of why that would be so.
4.2.3.4 Telic clauses
A final adjunct that is restricted to OC is the telic clause. Telic clauses,
illustrated in (46), express the outcome of the event denoted by the modified verb.9
(46) a. Harry1 nervously opened the letter, [only PRO1 to discover it wasn’t
actually for him].
b. It1 was sunny outside all morning, [only PRO1 to start raining right
when we decided to go to the park].
c. Due to our diligent firefighters, the historic mansion1 remained un-
touched by the wildfire, [only PRO1 to be destroyed by an earthquake
three months later].
These adjuncts, also known as “outcome” clauses, have received little attention in
the literature. Where they have been discussed, they have been classified as OC by
the matrix subject (Huettner, 1989; Whelpton, 1995; Landau, 2013). That OC is
possible is made evident by (46b,c), which have inanimate controllers. In addition,
9 See Whelpton (2001) for a detailed discussion of telic clauses and their semantic properties.
181
(47) demonstrates that implicit control is not possible, which provides further, albeit
not conclusive, evidence for strict OC.10
(47) a. *The meal was devoured, only PRO to discover it was poisoned.
b. The side-door on the plane was opened [only *(for me) to realize that
my parachute wasn’t fastened properly].
But, to reiterate, strict OC is unexpected under the TTC, unless the structure of
the adjunct is selected by some element in the matrix clause, which seems unlikely.
If the adjunct is not selected, the TTC predicts that NOC be should possible given
the right context.
The challenge that telic clauses present for the TTC is even greater than it was
for result, stimulus, and subject-gap purpose clauses. These adjuncts were shown
to adjoin within the VP, behaving syntactically as pseudo-arguments and making
the claim more feasible that their structure is selected by the matrix verb, as would
be required by the TTC for strict OC. However, telic clauses can adjoin to a huge
variety of predicates, and are unlikely to be included as an optional argument of
each. Furthermore, they adjoin too high in the structure to be considered verbal
arguments. Example (48) demonstrates that temporal PPs such as on Monday
are Chomsky-adjoined to the VP; they can either be stranded outside of a do so
anaphor, or included in its interpretation. Whelpton, 1995 uses VP fronting as
10 This kind of evidence cannot be taken as conclusive because the implicit control in these
examples could potentially be ruled out for other reasons. As noted above, even adjuncts that
allow implicit control in some instances do not always. It may be the case that other examples
diagnostic of NOC are eluding me.
182
an additional diagnostic for elements being associated with the VP. In (49), the
temporal modifier can be fronted along with the VP and its arguments.
(48) a. Rachel [VP [VP prepared some pasta] on Monday], and Monica did so
on Tuesday.
b. Rachel [VP [VP prepared some pasta] on Monday], and Monica did so,
too.
(49) Rachel asked Monica to prepare her some salmon on Monday, so prepare
her some salmon on Monday, Monica did.
Telic clauses, however, cannot front with the VP, as (50c) makes plain. Based on
this kind of evidence, Whelpton argues that telic clauses adjoin outside the VP, at
the IP level, well beyond the selectional domain of the verb.
(50) a. Chandler asked Monica to prepare him some spaghetti, so she made
him spaghetti, only to realize that he actually wanted linguine.
b. . . . so make him some spaghetti, she did, only to realize that he actually
wanted linguine.
c. *. . . so make him some spaghetti, only to realize that he actually wanted
linguine, she did.
Because telic clauses cannot be feasibly selected by the matrix verb, they
remain a striking problem case. Although the TTC predicts that they should allow
both OC and NOC, only OC appears to be possible.
183
4.3 The movement theory of control
Landau’s (2017) analysis of control of temporal and rationale adjuncts, situ-
ated within the TTC, predicts that all adjuncts should allow both OC and NOC.
This account can therefore, of course, be easily extended to adjuncts where this is
true. The TTC struggled, however, with certain adjuncts that appear to exhibit
only OC, but are unlikely to be selected as arguments of the matrix verb. This
section demonstrates that the duality of adjuncts with respect to OC and NOC
distribution is not contingent on Landau’s (2015, 2017) implementation of control.
With minimal assumptions, the Movement Theory of Control (MTC) does just as
well as the TTC. In fact, when phase theory is taken into account, the MTC does
better than the TTC at explaining some cases of strict OC.11
4.3.1 Background on the MTC
The MTC (Hornstein, 1999; Boeckx et al., 2010, a.o.) is prototypical of many
syntactic theories of control in that it establishes a specific syntactic dependency (in
this case movement) between PRO and a its antecedent in OC structures. In short,
OC differs from NOC in that in the former, what is traditionally labeled PRO is the
trace of movement of the controller from the embedded to the matrix clause, while
in the latter, PRO is the null pronominal pro. This is implemented as follows.
The MTC allows a DP chain to receive more than one θ-role, in contrast
to previous theories assuming a traditional θ-criterion (Chomsky, 1981) limiting
11 See Boeckx et al. (2010, ch. 7) for a broader discussion on problems with selectional theories
of control.
184
DP chains to a single role. OC is the result of movement from one θ-position to
another. A DP-trace is called ‘PRO’ when it occurs in the lower θ-position, but
the term ‘PRO’ has no other theoretical significance. Example (51) illustrates how
this works. Here, Jon moves from the subject position of the embedded buy to the
subject position of the matrix want, receiving the θ-feature of both verbs. In this
way, Jon is understood as both the wanter and the buyer in the sentence.
(51) Jon1 wants [Jon
12
1 to buy a car].
OC in adjuncts is handled with the same mechanism, i.e. the controlled element
and the controller are two links in an A-chain. The only difference is that adjunct
control requires what has been called sideward movement (Nunes, 1995), in which
an element can move from one syntactic object to the root of another within a single
derivation, as illustrated in (52).13
(52) Derivation of adjunct control (simplified from Boeckx et al., 2010, p. 88)
a. Applications of select, merge, and copy:
Num = John , Tφ+0 1, saw0, Mary0, after0, T
φ−
0, eating0, lunch0
PP = [after John Tφ+ eating lunch]
VP = [saw Mary]
b. Copying of John:
12 This derivation has been greatly simplified. Jon would be merged as the specifier of a little v
in the embedded clause, receiving the external θ-role of buy. It would then move to the embedded
spec-TP for EPP reasons, then to the matrix v, receiving the external θ-role of want, and finally
to spec-TP of the matrix clause, for Case and EPP.
13 These derivations use machinery of the Minimalist Program. ‘Num’ refers to the Numeration
(Chomsky, 1995).
185
PP = [after John Tφ+ eating lunch]
VP = [saw Mary]
N = John
c. Merger of John and VP:
PP = [after John Tφ+ eating lunch]
VP = [John saw Mary]
d. Applications of select, merge, and copy, and deletion in phonological
component:
Num = John0, T
φ+
0, saw0, Mary0, after
φ−
0, T 0, eating0, lunch0
TP = [John [Tφ+ [VP [VP John saw Mary][PP after John T
φ− eating
lunch]]]]
NOC interpretations arise under the MTC when the null subject of the con-
trolled clause is not a trace of movement, but is instead the null pronoun pro more
commonly seen in pro-drop languages.
4.3.1.1 Why OC is preferred over NOC, but not required
Under the MTC, movement resulting in OC and pronominalization resulting
in NOC are constrained by principles of economy. According to Hornstein (1999,
2001, 2003), pronominalization is more “costly” than movement, and is only possible
as a last-resort option when movement is blocked. However, the last-resort status
of pronominalization as originally formulated is too strong, as it would rule out any
case of NOC in adjuncts in favor of OC by the matrix subject, which could arise
186
through sideward movement. Boeckx and Hornstein (2007) acknowledge a similar
problem. If pronominalization were strictly last-resort, then (53a) should be ruled
out due to the availability of (53b), which Hornstein and Kiguchi (2003) argue arises
through movement.
(53) a. John1 said that [pro1 washing himself1 delighted Mary].
b. John said that [PRO1 washing herself1 delighted Mary1].
In order to allow both (53a,b), Boeckx and Hornstein (2007) propose that
pronominalization is only ruled out when movement could not establish the same
interpretation (see also Boeckx et al., 2010). Thus, NOC is not always ruled out
by the grammar when OC is available, as long as NOC results in an interpretation
that OC could not have.14 However, OC interpretations may also be preferred more
generally. Boeckx and Hornstein assume that parsers are transparent with respect to
grammars, meaning that if a grammar has a preference, then the parser will respect
that preference. When a comprehender reaches a gap in the incremental processing
of a sentence with a control structure, they will prefer to posit a trace of movement
if possible rather than a null pronoun, all things being equal. The pronominalization
and interpretation in (54b) is not ruled out by the grammar, but because parsers
would prefer to posit a trace as the null subject of the adjunct, resulting in (54a).
(54) a. John1 kissed Mary2 without PRO1/*2 getting embarrassed.
14 This idea is similar to what Reinhart (1983) proposed for bound versus free variables, as
discussed in §2.4.2.2.
187
b. John1 kissed Mary2 without pro*1/2 getting embarrassed.
The strong preference for OC interpretations in adjunct control is therefore
due to in part to a grammatical constraint requiring OC if it is possible for a given
interpretation,15 and in part due to a parsing preference (resulting from the gram-
matical constraint) for positing traces over null pronouns in incremental processing.
This combination results in a general preference for OC over NOC.
Boeckx and Hornstein (2007) and Boeckx et al. (2010) argue that parsers are
completely transparent to grammatical constraints, and that if positing a trace in
a given position is possible, parsers must do so. This would explain why examples
such as (55) only seem to allow OC by the matrix subject, even when it results in
a strange interpretation. Upon encountering the empty category in the adjunct, it
is possible to posit a trace linked to the matrix subject, and so parsers must do so.
(55) #The bank1 was robbed [while PRO1/*2 being smart not to raise the alarm].
(Landau, 2017, p. 7)
However, this does not explain why the same is not true in either of the examples
in (56). Evidently, parsers are able to override the preference for a trace in at least
some circumstances, even when a trace in the relevant position would be possible.
15 This explanation has the consequence that all cases of subject control in adjuncts must be OC,
inasmuch as sideward movement from the adjunct to the subject position of the matrix clause is
licit, but that any non-subject control interpretation must be NOC.
188
(56) a. The president1 was elected [without PRO1/2 considering his compe-
tence]. (Roeper, 1987, p. 297)
b. The potatoes1 were tastier [after PRO#1 eating carrots].
The idea that the preference for OC in general is due to a parsing preference,
and not a grammatical constraint, fits well with the data given above demonstrating
that structural properties of the sentence cannot fully determine when NOC will be
possible. Structural, discourse, and processing factors may all compete in deter-
mining whether NOC is possible. An example of the interaction of such constraints
is given in Boeckx and Hornstein (2007). It was argued that NOC is available in
(53a) despite the availability of OC in (53b) because at the point where the empty
category is encountered, if parsers posited a trace, they would not yet be able to
assign it a referent; the only positions that something could have moved to from
that position come later in the sentence. The preference to quickly assign reference
to dependent elements (see, e.g. Nicol & Swinney, 1989; Badecker & Straub, 2002)
is in competition with the preference for a trace. This competition results in the
possibility of either a trace or a pronominal being posited, depending on which pref-
erence is favored. Similar competition may be involved in NOC cases of adjunct
control.
4.3.1.2 Why complement control differs
The MTC assumes that OC is preferred over NOC in complements for the
same reason that it is in adjuncts. A grammatical constraint rules out NOC for
189
any interpretation that could have arisen through movement, and since parsers are
transparent to the grammar, a parsing preference will rule out any NOC interpreta-
tion when an OC one is available. However, for adjuncts, but not for complements,
this preference can be overridden in the proper context. It is not clear why comple-
ments and adjuncts should differ in this respect under the MTC, but one possibility
is that sideward movement, which is involved in the derivation of adjuncts, may for
some reason be more “costly” than upward movement, which takes place in com-
plement control. If that is true, then the additional cost of sideward movement may
be enough to reduce the parser’s preference for traces over pros in adjuncts. This
would lead to the preference scale for the derivation of control represented in (57).
(57) Preference scale for derivations of control:
upward OC  sideward OC  NOC via pronominalization
But what is the source of the preferences in (57)? Why would sideward move-
ment be more costly than upward movement? As discussed above, the preference for
OC over NOC is in large part due to processing preferences. This accounts for why
NOC is lowest on the scale. Might the preference for upward versus sideward move-
ment also be the result of processing biases? Perhaps sentences involving sideward
movement are simply more difficult to process than sentences with upward move-
ment. One reason why this might be the case begins with the fact that adjuncts are
optional. Because they are optional, perhaps they are less predictable. And if they
are less predictable, a comprehender will have less of an expectation, at the matrix
190
verb, of needing to form a control dependency, than when the verb selects a control
complement. But then again, even if this is true, why should online predictability
have anything to do with what interpretations are licit or available offline? There
is no logical connection here, so the effect would simply have to be stipulated. Un-
like the preference for OC over NOC, this effect cannot be reduced to other known
parsing preferences, such as the preference for local antecedents over distant ones or
the preference for syntactic versus discourse dependencies. Therefore, stating that
(57) is due exclusively to processing preferences is tenuous.
It is also not immediately clear what grammatical constraints would result in
(57). In terms of the Minimalist Program, both upward and sideward movement
involve copy and re-merger of an element of the growing syntactic structure. Neither
involves more steps or even different underlying operations that could lead to a
greater cost. There is only one real difference between the two, so therein must
be found the cost difference: sideward movement involves copying an element from
one tree and moving it to the root of another, as in (58a), while upward movement
involves copying from and re-merging to the same tree, as in (58b).
(58) a. vP
Harry v′
v VP Adjunct
celebrated after Harry winning
Sideward movement
191
b. vP
Harry v′
v VP
U
pw promised Complementard
mov Harry to winement
Perhaps sideward movement incurs a derivational penalty akin to the one that
has been posited for moving out of islands (Chomsky, 1972; Lasnik, 2001). Although
I do not speculate here on the exact nature or source of this penalty, let us simply
suppose by hypothesis that there is a penalty for moving an element from within
one tree to the root of another, unconnected tree. Suppose, furthermore, that this
penalty is weak,16 and does not cause the derivation to crash, but instead results in
a lower “degree of grammaticality” (Epstein, 1990; see also Chomsky, 1965, 1986;
Lasnik & Saito, 1984). Because both upward and sideward movement can lead
to convergent derivations, both may result in grammatical strings; those involving
sideward movement would simply be lower on the scale between fully grammatical
and fully ungrammatical.
If this hypothesis is on the right track, the preference for upward OC versus
sideward OC is due to a weak grammatical constraint against movement to a non-
commanding position. If there is some performance penalty associated with this
grammatical violation, then the lower grammatical status of sideward OC would
16 See Chomsky (1965), Epstein (1990) for a discussion on different types of violations resulting
in penalties of differing strength.
192
make its appeal relative to NOC weaker than upward movement’s. This in turn
would lead to a greater probability of both OC and NOC being accepted in adjuncts
involving sideward movement than in complements, in which movement is upward.
Admittedly, positing that sideward movement is grammatically more costly
than upward movement is somewhat stipulative. But something like it as part of
the scale in (57) would be required in order to capture the differing availability of
NOC in complement and adjunct control under the MTC. Interestingly, sideward
movement is also involved in the examples from Boeckx and Hornstein (2007) in
(53), repeated in (59), that allow both OC and NOC derivations. This is consistent
with the idea that OC derived from sideward movement is less strongly preferred to
NOC than OC derived from upward movement is.
(59) a. John1 said that [pro1 washing himself1 delighted Mary].
b. John said that [PRO1 washing herself1 delighted Mary1].
4.3.1.3 The interpretation of NOC
Under the MTC, the interpretation of NOC pro is not constrained by the
syntax in the way the deleted copy of OC is. Instead, it is free to refer in the same
way an overt pronoun would. Although this does not explain why NOC requires
the understood antecedent to be a perspective holder in adjuncts, it is also not
incompatible with that fact. Boeckx and Hornstein (2004) tentatively suggest that
pro in NOC may just be a null logophoric pronoun, and as such, it requires its
193
referent to be a logophoric center in the discourse. Why a logophoric pro must
appear in NOC, rather than a regular pro, however, is left without explanation.
4.3.2 The MTC and other adjuncts
Hornstein (2001) assumes an Uriagereka (1999)-style adjunction in which ad-
juncts are linearized and therefore become islands before being merged with the
main clause (see also Nunes, 2001). If this is the case for all adjuncts, then the
MTC makes the same prediction that the TTC made: namely, that all adjuncts
should allow both OC and NOC. For the MTC, this will be because all cases of OC
of adjuncts will necessarily involve sideward movement. Because this is more costly
than upward movement, NOC should be available for all adjuncts given the right
context. The MTC will therefore do just as well as the TTC on adjuncts that allow
both OC and NOC. And just like the TTC, the MTC as outlined so far struggles
with adjuncts that allow only OC. In this section, I outline a proposal for how the
MTC can deal with strictly OC adjuncts that adjoin low. Section 4.3.3 will discuss
problem cases for the MTC.
Let us first turn to low OC adjuncts. Above it was seen that goal clauses
(60a), result clauses (60b), stimulus clauses (60c), and SPCs (60d) all exhibit strict
OC.
(60) a. Max1 works hard [PRO1 to stay out of jail].
b. Harry1 grew up [PRO1 to be a famous wizard].
c. Ginny1 shuddered/smiled [PRO1 to think what a fool she had been].
194
d. I bought this blender1 [PRO1 to help me make split pea soup].
Under the TTC, strict OC would have to be explained by positing that these clauses
are not actually adjuncts, but optionally selected arguments of the matrix verb. If
they are actually complements, then the MTC need not say anything more. But
this seems unlikely, especially for SPCs.
Under the version of the MTC outlined above, strict OC is only expected if
upward movement from the ‘PRO’ position to the controller position is possible.
There are two requirements that must be satisfied for this to be possible. First, the
adjunct must adjoin lower than the base position of the controller, and second, the
adjunct must not (yet) be an island to movement. We will look at each of these in
turn.
Each of the adjuncts in (60) was demonstrated above to adjoin to the VP. For
the adjuncts in (60a,c), this places the adjunct clearly under the merge position of
the subject in spec-vP, as illustrated in (61).
(61) vP
DP v′
v VP
V Adjunct
DP to. . .
195
But what if the controller is not base-generated in the vP, but as an object of V,
as is the case with (60b), which has an unaccusative subject controller, and (60d),
which has control by the object? Kratzer (1996), following Larson (1988), proposes
that direct objects of verbs are base-generated as the specifier of V. If that is the
case, and if adjuncts may adjoin to the V′ level as Whelpton (1995) argues, then
these adjuncts too may adjoin in a position lower than the base position of the
controller, as in (62).
(62) VP
DP V′
V′ Adjunct
V DP. . .
If these adjuncts adjoin with the structures in (61) and (62), they are in a po-
sition where upward movement from the controllee to the controller positions should
be possible. However, we are still left with the problem that these are adjuncts, and,
according to Hornstein (2001), Nunes (2001), and Uriagereka (1999), they should
become islands prior to or upon adjunction. But in phase theory (Chomsky, 2001),
this is not necessarily the case. Let us make the standard assumption that C and v
are phase heads. If an adjunct is smaller than a CP, then it should remain active,
even after being adjoined, until the next highest phase. Its subject should then be
196
able to upward move out of the adjunct into the matrix clause before the adjunct
becomes an island.
I assume that this is the case for the adjuncts in (60). Goal clauses, for
example, are smaller than a CP, as evidenced by the fact that a lexical subject is
impossible, in contrast to temporal and rationale adjuncts (see (63)).
(63) a. *Harry grew up [for him to be a great wizard].
b. Harry went to school [before he became a great wizard].
c. Harry had to go to school [in order for him to become a great wizard].
Therefore, the subject of a goal clause can upward move after adjunction. And
because the goal clause adjoins lower than the base position of the matrix subject
(see (61)), OC by the matrix subject is possible.17
This proposal may also work in a convergence-based approach to phasehood
(Felser, 2004; Grano & Lasnik, in press), in which a constituent is not spelled-
out until it has no more unvalued/uninterpretable features. Under the MTC, the
subject DP in the adjuncts in (60) have unvalued case features, which motivates their
movement to the matrix clause. Therefore, the adjunct will not become an island
directly upon adjunction, since it has not yet converged, and upward movement will
still be possible.
17 Note that Hornstein (1999) follows Nunes (1995) in assuming that the Extension Condition
(Chomsky, 1995) applies to adjuncts. If this is true, then an alternative derivation, whereby the
subject sideward moves out of the adjunct into the matrix clause followed by low, late adjunction,
is impossible.
197
Additional evidence that these adjuncts do not immediately become islands is
the fact that CED island effects (Huang, 1982) in them are weak or nonexistent.
(64) a. What1 did Max work hard [to do t1]?
b. What1 did Harry grow up [to be t1]?
c. Who1 did Ginny smile/shudder [to think the poison would kill t1]?
d. What1 did you buy this blender [to help you make t1]?
If, as I propose, upward movement is allowed in these adjuncts, leading to
strict OC, they must be distinguished from adjuncts in which both OC and NOC
are possible. There are two ways this could be done. First, if the adjunct adjoins
higher than the base position of the subject, sideward movement will be required,
regardless of the when the adjunct is spelled-out, since the adjunct will never be in
a position in the tree low enough to allow upward movement. Second, low adjuncts
with structures at least as big as a CP will also not allow upward movement to
the matrix subject position, since the CP adjunct will be spelled-out upon merger
of the matrix v.18 It must therefore be the case that all of the adjuncts discussed
above that allow both OC and NOC have one of these two properties. This is at
least true for OPCs, which have a CP structure (Chomsky, 1980; Williams, 1980,
1992; Whelpton, 2002), as illustrated in the examples repeated in (65). I leave it to
future research to verify that this is the case for other adjuncts.
18 Although under a convergence-based theory of phasehood, even CPs may allow upward move-
ment if they still contain elements with unvalued features.
198
(65) a. Chandler bought Monica1 a new frying pan2 [CP Op2 [PRO1 to work
with t2]].
b. I left it2 in [her1 mailbox] [CP Op2 [PRO1 to look over t2 once she
returned from the Bahamas]].
4.3.3 Problems for the MTC
The MTC is able to account for adjuncts allowing both OC and NOC, as well
as for low adjuncts allowing strict OC—something that that TTC struggled with.
However, there are at least two problem cases for the version of the MTC proposed
here: telic clauses, which display strict OC, but may adjoin higher than would be
required for upward movement, and speaker-oriented adverbials, which display strict
NOC.
4.3.3.1 Telic clauses
Strict OC is unexpected under the MTC, unless the adjunct attaches lower
than the base position of the controller and is small enough to be included in
the same phase, thus allowing control via upward movement. But as discussed
in §4.2.3.4, Whelpton (1995) provides evidence from VP fronting that telic clauses
adjoin to the IP level, well outside the merge position of the subject within the vP.
The relevant examples are repeated in (66).
(66) a. Chandler asked Monica to prepare him some spaghetti, so she made
him spaghetti, only to realize that he actually wanted linguine.
199
b. . . . so make him some spaghetti, she did, only to realize that he actually
wanted linguine.
c. *. . . so make him some spaghetti, only to realize that he actually wanted
linguine, she did.
Whelpton argues that because telic clauses cannot be fronted with the VP, they must
be higher than the (extended) verbal projection, i.e. at the IP level, as illustrated
in (67). Because of this, the only way for OC to occur would be through sideward
movement.
(67) TP
Monica T′1
T′ Telic Clause
T vP only Monica1 to realize. . .
Monica1 made spaghetti
nt
rd mo
veme
Sidew
a
Because sideward movement is required, the account presented here predicts
that NOC should be possible in telic clauses given the right context. In order to
account for telic clauses, the MTC would either have to say that they attach lower
than Whelpton (1995) thought, or provide evidence that NOC is possible. Again, I
leave further investigation of telic clauses to future research.
200
4.3.3.2 Speaker-oriented adverbials
Finally, speaker-oriented adverbials like those in (68) present something of a
challenge for the MTC. It was noted above that PRO in these adjuncts appears
only to be able to refer to the speaker, which is evidence for their being strictly
logophoric NOC.
(68) a. [PRO To be honest], Jon would be better off without Mary.
b. [PRO Judging from experience], Jon will be better off without Mary.
The TTC can explain this if speaker-oriented adverbials adjoin high in the clause,
above the position of the matrix subject, as in (69). In this position, the adjunct
is too high for a sentential argument to predicatively control PRO, and therefore,
logophoric control is the only option.
(69) XP
SOA . . .
PRO To be honest . . . TP
Jon will be. . .
Intuitively, it seems that a similar explanation should be able to apply to
the MTC. With the structure in (69), PRO is not c-commanded by any other
element of the sentence. Movement out of that position to any position in the
main clause would then result in an ill-formed chain, the head of the chain not c-
commanding the tail (Chomsky, 1995). Although the same is true initially for all
201
cases of sideward movement, here the problem is more serious, since even at the end
of the derivation, the final position of the moving element does not c-command the
trace in the adjunct. However, Hornstein and Kiguchi (2003) argue that c-command
is not required for sideward movement. They demonstrate that in examples like (70),
which involve sideward movement, PRO’s controller does not c-command it, even
on the surface, and yet the example demonstrates all the characteristics of OC. And
this is not limited to psych-verbs, as might be assumed based on theories such as
that of Belletti and Rizzi (1988), since PRO in (71) is also not c-commanded by its
OC antecedent (Norbert Hornstein, p.c.).
(70) John said [that [PRO1 washing herself] delighted Mary1].
(71) [PRO1 Brushing her teeth] made Mary1 late for the movie.
It is therefore not clear why in speaker-oriented adverbials the matrix subject should
not be able to sideward move from the adjunct into the main clause.
(72) * XP
SOA . . .
PRO To be honest . . . TP
S Jon will be. . .ideward movement
202
4.4 Conclusion
In the previous chapters, I argued, agreeing with Landau (2017), that tempo-
ral and rationale adjuncts allow both OC and NOC. Many early and some more
modern theories of control assumed that OC and NOC were in complementary dis-
tribution, and significant theorizing was done to explain that complementarity. For
some authors, this distribution was based on the structure of the sentence contain-
ing PRO. For Manzini (1983), for example, the distinction was quite general: OC
obtained when PRO was in an object clause, and NOC when it was in a subject
clause. Similarly, Landau (2000) claimed that NOC was structurally conditioned,
only obtaining for VP-external infinitives. For other authors (e.g. McFadden & Sun-
daresan, 2016; Hornstein, 1999; Landau, 2017, 2018), even if strict complementarity
was not part of the theory, it was argued that NOC is only possible where OC is
not.
I have shown that the distribution of the two types of control is not so simple.
If one defines OC as a syntactic (or predicative) dependency between PRO and
its antecedent, and NOC as a non-syntactic one, as Hornstein (1999) and Landau
(2017) do, the two types of control are not in complementary distribution; temporal,
rationale, and some other adjuncts allow both. However, it is still the case that
complements and some adjuncts are strictly OC. Labeling the syntactic version of
control as OC is therefore something of a misnomer, as a construction may exhibit
syntactic control without syntactic control being obligatory.19
19 Landau (2015) makes a related observation. In the TTC, OC need not always involve the same
syntactic dependency. In complements, both predicative and logophoric control are OC.
203
If temporal and rationale adjuncts allow both OC and NOC, a simple hy-
pothesis would be that the same is true for all adjuncts. This is in fact what the
discussion in Landau (2017) implies, though the claim is not made directly. This
chapter has shown that this hypothesis is too general. Some adjuncts seem to only
allow one type of control. Certain adjuncts appear to be strictly OC, while very
high ones may be strictly NOC. For the TTC, the strict OC adjuncts can only be
accounted for by claiming that they are in fact not adjuncts, but complements. If
such is the case, then the matrix verb may select a predicative structure for the
adjunct, resulting in strict predicative OC. This claim has some merit for adjuncts
like goal clauses, but it is far less plausible for SPCs and telic clauses.
The MTC, on the other hand, combined with phase theory, is able to handle
most of the strict OC adjuncts without having to claim that they are complements.
If the adjunct adjoins low enough and is not a phase, then it will allow upward
movement from the controlled to the controller position. If upward movement makes
OC more strongly preferred than sideward movement does, possibly due to sideward
movement incurring a weak grammatical penalty that does not ensue with upward
movement, then these adjuncts would be expected to require OC just as strongly
as complement control does. Despite these successes, the MTC struggles with telic
clauses, just as the TTC did. Since telic clauses adjoin higher than the merge
position of the matrix subject, upward movement out of the adjunct to the relevant
position is impossible. The MTC also has difficulty explaining the interpretation of
speaker-oriented adverbials, which are strictly NOC.
Finally, when an adjunct allows both OC and NOC, it is clear that the bound
204
variable reading of OC is often strongly preferred, although that preference may
have different sources based in part on which specific theory of control is espoused.
Regardless of what theory of control is chosen, I have provided initial evidence that
the preference for OC may be overridden in favor of NOC by (a combination of)
various factors, including attachment site and conceptual plausibility.
Unlike the distinction between upward and sideward movement, which I argue
has a grammatical basis, the preference for OC is more likely due to performance
biases. As a fist step in investigating the processing constraints involved, the re-
mainder of this dissertation investigates the real-time processing of adjunct control
structures. Although specific interpretive biases are not yet directly examined, the
experiments presented do provide foundational knowledge about the timing of the
interpretation of PRO. This knowledge will be crucial for future investigations of
how different factors and sources of information affect the processing of control.
These experiments will also give initial insights into the role of prediction in the the
interpretation of control.
205
Part II
Processing
206
Chapter 5: Processing adjunct control
The bulk of this dissertation has explored questions about the grammatical
constraints governing adjunct control and its syntactic representation. In the next
chapters, I investigate questions about how adjunct control is recognized and inter-
preted in real time. After reviewing some of the questions about adjunct control
that have been asked in previous literature in this chapter, Chapter 6 asks about the
timing of and the role of prediction in the resolution of PRO in subject-controlled
temporal adjuncts. A better understanding of the processing of adjunct control
may ultimately yield new diagnostics for questions about the grammar. For exam-
ple, it may provide new measures for distinguishing syntactic and discourse repre-
sentations. Or it may provide evidence on the nature of interpretive constraints,
including those leading to a preference for OC over NOC. This evidence may show
that some constraints that were assumed to be the result of processing biases are
not tenable as such, or that other constraints that appeared to be grammatical are
actually underlyingly due to processing.
207
5.1 Previous research on the processing of adjunct control
Much of the prior work on the processing of adjunct control structures has
focused on one of the following questions: (i) what can processing measures tell
us about the syntactic representation of sentences containing adjunct control? (ii)
how fast are control structures processed? and (iii) how accurately is the correct
antecedent to PRO retrieved in online processing? This section briefly reviews
the literature for each of these questions. Section 5.2 then discusses some of the
remaining questions.
5.1.1 Syntactic representation
One goal of previous processing research on adjunct control has been to ad-
judicate between competing theories of the syntactic representation of passives and
implicit control. In a series of papers, Mauner et al. (1995) and Mauner and Koenig
(1999, 2000) explored the processing of RatCs adjoined to target clauses with active,
long passive, short passive, or intransitive structures, as in (1).
(1) a. The contestant spun the game show’s wheel [to win a prize].
b. The game show’s wheel was spun by the contestant [to win a prize].
c. The game show’s wheel was spun [to win a prize].
d. The game show’s wheel spun [to win a prize].
Using a stop-making-sense task combined with self-paced reading (Mauner et al.,
1995; Mauner & Koenig, 2000) or eyetracking-while-reading (Mauner & Koenig,
208
1999), these studies consistently found that RatCs took significantly longer to pro-
cess and, where it was measured, that they were judged to make sense significantly
less frequently, when adjoined to an intransitive matrix clause than in any other
condition. There was no significant difference in the processing or judgments about
sentences containing short passives, long passives, or active sentences.
Mauner and colleagues took these processing facts as evidence in support of
the presence of a syntactically active implicit agent in short passives. Since there is
no difference in the processing of a RatC following either an active or passive, the
reasoning goes, it is likely to be the case that both involve the same interpretive
constraints, namely grammatical binding of PRO by the agent of the matrix clause.
However, in Chapter 3, it was argued that implicit control in RatCs is actually NOC
(see §3.2.1.2, specifically). If that is the case, then whether the so-called “implicit
agent” is syntactically represented is irrelevant, as it does not enter a syntactic
control relation with PRO. These studies also do not provide direct evidence for the
interpretation of PRO. It could therefore be the case that the slowdowns in reading
and the greater rejection rate of RatCs after intransitives were due not to difficulty
in finding an antecedent to PRO, but to other reasons. For example, (1d) may be
unacceptable simply because the RatC does not provide a reasonable explanation
for the fact expressed by the intransitive clause (see A. Williams & Green, 2017).
McCourt et al. (2015) provide additional experimental evidence that casts
doubts on Mauner and colleagues’ conclusions. Using self-paced reading, McCourt
et al. compare explicit and implicit control not only in local RatCs, but in remote
ones as well, as in (2).
209
(2) a. The committee interviewed the candidates [just in order to find the best
person for the job].
b. The candidates were interviewed [just in order to find the best person
for the job].
c. The committee interviewed the candidates. The reason was [to find the
best person for the job].
d. The candidates were interviewed. The reason was [to find the best person
for the job].
Remote control cannot involve a grammatical OC dependency between PRO and
its understood antecedent (see Chapter 3). Because of this, its interpretation must
be mediated pragmatically. Therefore, if processing measures such as self-paced
reading are able to distinguish between the resolution of syntactic and non-syntactic
dependencies, as Mauner and colleagues suggest, then one might expect there to be
a difference between the interpretation of remote and (explicit) local control. If, on
the other hand, there is no processing difference between local and remote control,
then, McCourt et al. argue, the lack of a processing difference between explicit
and implicit control in Mauner and colleagues’ studies cannot be used to argue for
grammatical binding of PRO in the latter.
This is indeed what McCourt et al. find (see Figure 1, their Figure 4). In
their Experiment 4, they found similar reading times at the RatC in sentences with
explicit and implicit control. Reading times in RatCs in a local configuration trended
210
Figure 1: Mean logged reading times (ms) for Experiment 4 in McCourt,
Green, Lau, and Williams (2015).
slower than those in a remote configuration at the non-finite verb, but the difference
was not significant (p = 0.07).1
McCourt et al. argue that if similar processing profiles reflect anaphoric de-
pendencies of the same type, then control in RatCs must be non-syntactic across
the board. Since remote control cannot involve OC, local control must not either.
However, the authors acknowledge that this argument relies on self-paced reading
times being sensitive to the difference between OC and NOC. It is possible that
OC and NOC are not distinguished by processing cost (see Cunnings, Patterson, &
1 Because there is good independent evidence that implicit control must be NOC, even if the
marginal difference between reading times in the local and remote control conditions in McCourt
et al. (2015) were meaningful, it still would not reflect a processing difference for OC and NOC.
If such an difference were shown, then we would expect local implicit control to pattern with the
remote control conditions.
211
Felser, 2014). Therefore, McCourt et al. conclude that their results either provide
evidence that RatCs are always NOC, which I argue in Chapter 3 is not the case,
or that they at least undermine Mauner et al.’s arguments for OC in implicit con-
trol. A third possibility, once again, is that PRO was not actually being interpreted
in these experiments. It could have been that comprehenders were at least initially
only shallowly interpreting the RatCs, being content to leave final resolution of PRO
until later, and only if it should prove to be necessary.
One other effect that McCourt et al. (2015) briefly note is the role of prediction
in the processing of RatCs. In their Experiment 3, they found that when the local
RatCs were prefaced not by just in order, but by a neutral temporal modifier, as
in (3), reading times in the local conditions were significantly slower than in the
remote conditions. This was taken to reflect a speed-up in the remote conditions
due to the subject of the remote clause being predictive of a RatC. Although
this experiment did not specifically manipulate predictiveness, it does suggest that
prediction of an upcoming structure led to facilitation in processing, supporting a
large body of literature on the role of prediction in sentence processing (see, e.g.
Kamide, Altmann, & Haywood, 2003; Yoshida, Dickey, & Sturt, 2013; Kehler &
Rohde, 2013).
(3) a. The committee interviewed the candidates [three days ago to find the
best person for the job].
b. The candidates were interviewed [three days ago to find the best person
for the job].
212
5.1.2 Speed of processing control structures
Another body of work has focused on the speed of processing control struc-
tures. In contrast to pronouns, there has been relatively little work on the timing of
the resolution of PRO. And the majority of what we do know comes from studies
of complement control as in (4), rather than adjunct control (5).
(4) a. Mickey1 managed [PRO1 to eat].
b. Mickey1 promised Minnie [PRO1 to eat].
c. Mickey persuaded Minnie1 [PRO1 to eat].
(5) Mickey1 talked to Minnie [before PRO1 eating].
Studies on complement control (e.g. Boland, Tanenhaus, & Garnsey, 1990; Demestre,
Meltzer, Garcia-Albea, & Vigil, 1999) have generally found that the anaphora can
be processed quickly. However, complement control (4) differs from adjunct control
(5) in important ways. First, although interpreting complement control structures
may still require the use of structural information (Hornstein, 1999; Landau, 2000,
2015), the matrix verb gives early cues that a control structure is coming as well as
what the referent of PRO will be. It is therefore plausible that the rapid processing
seen in complement control structures is due at least in part to information conveyed
by the main-clause verb.
With adjunct control, the participial clause is not selected by any element
of the matrix clause, and the choice of antecedent is often based strictly on the
213
attachment site of the adjunct (Clark, 1990).2 Because comprehenders do not have
advanced help from the embedding verb to determine who the speaker intends to
refer to, resolution of the implicit subject may be more difficult.
Betancort et al. (2006) found that at least some cases of adjunct control did
show greater processing difficulty than complement control. These authors investi-
gated the processing of RatCs in Spanish headed by para (6a) in comparison with
response clauses headed by por (6b) using eyetracking while reading. As Betancort
et al. explain, adjuncts headed with para are strongly biased to be interpreted with
subject control, and those with por with object control. These were compared with
complement control structures with similar biases towards either subject or object
control.
(6) a. Yolanda1 se-casó con Jorge2 [para PRO1 tener dinero]
Yolanda1 married with George [to PRO1 have money]
‘Yolanda married George in order to have money.’
b. Yolanda se-casó con Jorge1 [por PRO1 tener dinero]
‘Yolanda married with George1 [for PRO1 have money]
‘Yolanda married George because he had money.’
The results of this experiment showed that for adjuncts, in contrast to com-
plements, there were late effects of recency of the antecedent. Although there were
no reliable effects in first-pass reading times, both regression path and second-pass
reading times were significantly faster in object control adjuncts headed by por than
in subject control adjuncts headed by para. Betancort et al. state that this effect
2 This is only true when OC is involved. See Chapter 2.
214
supports earlier theories (Frazier, Clifton, & Randall, 1983; Nicol & Swinney, 1989)
that posited a Garden Path effect in the processing of control structures, such that
comprehenders initially take the most recent potential antecedent (here, the object
of the matrix clause) as the antecedent to PRO, and that this initial assumption
must later be revised for subject control sentences. This is argued to be the case
despite the fact that the preposition introducing the adjunct (para) strongly biases
toward subject control readings; the comprehender’s initial reaction goes contrary to
statistical probabilities of interpretation. In complement control, the authors found
no evidence for a Garden Path effect, in contrast to the earlier studies.
Although this conclusion provides an explanation for the slower processing of
the subject-controlled RatC, it is also largely stipulative. The experiments used pro-
vide no direct evidence on the timecourse of interpretation or on what antecedents
are being considered. The slowdowns in processing RatCs compared to response
clauses could therefore have been due to other differences between the two adjuncts,
not necessarily due to readers initially assigning an incorrect object control inter-
pretation for PRO in the RatC.
5.1.3 The quality of retrieval
Although previous experiments provided evidence about the speed of process-
ing of control structures, none used measures that that provided direct evidence on
when or whether PRO was actually interpreted. Parker et al. (2015) more directly
detected the activation, but their questions were less about the timing of PRO’s
resolution as about the accuracy of retrieval of PRO’s antecedent in memory. In
215
two experiments using a facilitatory interference paradigm (Pearlmutter, Garnsey,
& Bock, 1999), Parker et al. measured self-paced reading times in sentences contain-
ing temporal adjuncts such as (7). The first experiment manipulated the animacy
of the controller ((7a–b) vs. (7c–d)) and of an intervening distractor (the subject of
a relative clause modifying the matrix subject) ((7a,c) vs. (7b,d)).
(7) a. The doctor that the researcher evaluated extensively was commended
after disproving the controversial theory himself at the research institute
in Europe.
b. The doctor that the report evaluated extensively was commended after
disproving the controversial theory himself at the research institute in
Europe.
c. The experiment that the researcher evaluated extensively was commended
after disproving the controversial theory himself at the research institute
in Europe.
d. The experiment that the report evaluated extensively was commended
after disproving the controversial theory himself at the research institute
in Europe.
The first measure of the interpretation of PRO was at the gerundival verb.
Because PRO in temporal adjuncts prefers an animate controller, as Parker et al.
confirm, animacy may serve as a cue used in the search for an antecedent.3 It
3 This study assumes a Lewis and Vasishth (2005)-based model of memory and sentence pro-
cessing.
216
was found that reading times at the gerundival verb were faster for sentences with
inanimate main-clause subjects when the distractor was animate, suggesting that a
search for an antecedent to PRO had already begun at the verb.4
The second self-paced reading experiment provided further support that a
search for PRO’s antecedent was initiated by the adjunct’s verb. In items such as
(8), where both the target and the intervening distractor were animate, reading
times at the gerundival verb were slower when both characters were of the same
stereotypical gender, as in (8a). Parker et al. suggest that this could be due to a
“fan” effect, which can arise in grammatical contexts when multiple items match
the retrieval cue, and may reflect interference at the encoding stage. If this is true,
then it provides further evidence that participants were at least initiating a retrieval
for PRO’s antecedent at the embedded verb.
(8) a. The harpist that the diva liked very much was congratulated after play-
ing the beautiful song herself at the brand new recording studio.
b. The harpist that the guitarist liked very much was congratulated after
playing the beautiful song herself at the brand new recording studio.
c. The drummer that the diva liked very much was congratulated after
playing the beautiful song herself at the brand new recording studio.
4 Facilitation was also seen in the region following the reflexive, suggesting that at least some
of the time, the distractor was activated as PRO’s antecedent, or at least as a candidate, licensing
the reflexive. However, it is also possible that the effect from the reflexive was due to searching
the main clause arguments in memory, not PRO. See Parker and Phillips (2017) for discussion on
interference effects in the processing of reflexives.
217
d. The drummer that the guitarist liked very much was congratulated after
playing the beautiful song herself at the brand new recording studio.
Although self-paced reading times are a fairly indirect measure for anaphora
resolution, these experiments provide initial evidence for the rapid interpretation of
PRO in temporal adjuncts.
5.2 Remaining questions
The previous research on adjunct control leaves several remaining questions.
The simplest of these involves the timing of the interpretation of PRO in online
processing. Although Parker et al. (2015) provides some evidence that a search for
PRO’s antecedent is initiated early in the adjunct, it still does not provide direct
evidence for when that search is completed. None of the other studies on adjunct
control can distinguish the interpretation of PRO from the processing of the adjunct
more generally. Therefore, it is still not clear when PRO in these adjuncts is fully
interpreted, or whether listeners allow shallow, underspecified representations until
later in discourse, if full resolution is completed at all.
Second, these studies do not fully resolve the question of the path PRO’s
interpretation takes. Is the most recent potential filler considered first, even when
the syntax biases you towards more distant antecedents, as Betancort et al. (2006)
suggest? This could help explain some of the facilitatory interference effects in
Parker et al. (2015). But these results are also consistent with a simultaneous search
218
of all potential antecedents, as in the ACT-R model of sentence comprehension
(Lewis & Vasishth, 2005).
The experiments in Chapter 6 begin to address these questions for temporal
adjuncts using visual-world eyetracking, which can provide a more direct measure
of online interpretation of referential expressions. Previous studies have shown that
upon hearing a pronoun, for example, listeners tend to look at the character on the
screen to whom they assume the pronoun refers (Arnold, 1998; Arnold, Eisenband,
Brown-Schmidt, & Trueswell, 2000; Arnold & Lao, 2015). These results form a
foundation needed for successfully addressing other questions, such as the role of
prediction in the resolution of adjunct control, what interpretive and processing
biases are at play in the competition between OC and NOC, and how the process
of interpretation may vary for different types of adjunct control.
219
Chapter 6: Processing temporal adjuncts
This chapter1 explores the online interpretation of PRO in right-adjoined tem-
poral adjuncts, as in (1a), which are strongly biased towards subject control inter-
pretations, especially when the main clause subject is animate. This is compared
to the processing of overt pronouns, as in (1b).
(1) Mickey1 ran into Daisy2 in front of the school. . .
a. . . . [before PRO1 picking up a blue ball].
b. . . . [before he1/she2 picking up a blue ball].
Using visual-world eyetracking, it is shown that when listeners may believe a control
structure to be a likely continuation, they are just as quick to resolve reference when
they hear the non-finite verb “eating” as they are when they hear the pronoun “he.”
This is the case even though there is no overt expression in (1a) whose semantic
function is just to evoke a specific individual in the comprehenders memory, here in
the role of eater. When the experimental context is manipulated to reduce the like-
lihood of control structures, participants are slower to resolve PRO, but only insofar
as the cue to anaphora (the non-finite verb) is longer. This provides evidence that,
1 This chapter reports joint work with Michael McCourt, Ellen Lau, and Alexander Williams.
220
in at least some contexts, PRO is identified quite rapidly, despite being inaudible,
and is furthermore resolved online as part of the incremental interpretation, and
that listeners are able to make rapid use of structural information in the computa-
tions involved in the resolution of anaphora once the relevant bottom-up input is
received.
6.1 Background: anaphora resolution generally
Within the research on the interpretation and processing of referential expres-
sions, pronouns have received significant attention. For example, a speaker of (2)
can use both “Mickey Mouse” and “he” to refer to Mickey Mouse.
(2) Mickey Mouse talked to Minnie Mouse before he ate.
To understand this use of the name “Mickey Mouse,” it might be enough for a
listener to access its representation in his or her mental lexicon, as this might be
linked to the singular concept mickey mouse. But understanding the use of “he”
requires more than lexical access, since the pronoun is surely not itself linked to
mickey mouse. Instead, it requires consultation of other representations in mem-
ory, such as a representation of the discourse in which the pronoun occurs, of the
syntax of the local sentence, of the speaker’s interests, and so on.
Much work in psycholinguistics has been directed at the online processing of
pronouns and other anaphoric expressions. Their study promises to illuminate the
mechanisms by which disparate information sources are integrated in language com-
221
prehension (Ehrlich & Rayner, 1983; Blanchard, 1987; Arnold et al., 2000; Stewart,
Pickering, & Sanford, 2000; Kehler & Rohde, 2013). Given the potential complexity
of this task, one set of questions revolves around timing. How long does it take to
resolve reference? And in general, is resolution initiated immediately upon percep-
tion of the anaphoric expression? Or do we, often enough, resolve reference only
when our practical goals demand it? There is active pursuit of these questions in
the literature (see, e.g. Stewart, Holler, & Kidd, 2007; Karimi & Ferreira, 2016; see
also Ferreira, Bailey, & Ferraro, 2002, for a discussion on other instances of shallow
processing).
Seminal work by Arnold et al. (2000) using the visual world paradigm has
shown that pronouns can be very rapidly interpreted to infer the referent intended
by the speaker, especially when the pronoun’s “φ-features” (person, gender, and
number) agree with a single entity in the discourse. In their experiments, par-
ticipants viewed a picture with two characters and heard descriptions in which a
pronoun referred to one of them. Results showed that when the two characters
mismatched in gender, participants looked to the correct character in the picture
within approximately 200 ms after the offset of the pronoun, suggesting that they
were able to use the gender features of the pronoun to successfully resolve reference
by that time.
However, pronominal dependencies in English have a particular profile of prop-
erties that do not generalize across all referential dependencies. First, they are cued
by an overt anaphoric expression (the pronoun). This might seem unremarkable,
but there exist many common cases of referential dependencies that do not share
222
this property, such as with the null pro found in Spanish and many other languages,
and, of course, PRO. Second, their interpretation is strongly influenced by discourse
factors and less so by structural information. When a pronoun’s φ-features are com-
patible with multiple entities, comprehenders are guided in their identification of a
referent by factors such as information structure (Crawley, Stevenson, & Kleinman,
1990; Grosz et al., 1995; Kehler & Rohde, 2013) and discourse coherence (Hobbs,
1979; Kehler et al., 2008; Kehler & Rohde, 2013). They are also influenced at least
initially by psychological factors such as attention (Arnold & Lao, 2015). However,
at least for overt pronouns like English ‘him’ or ‘her,’2 structural information about
the antecedent is not directly relevant, beyond the constraints of Principle B, which
prohibits antecedents that are syntactically too local (Chomsky, 1981). Structural
information is relevant only indirectly, in any effects it may have on information
structure and coherence. Things are arguably different for PRO, however. Here the
dominant information relevant to reference resolution is the syntactic position of the
antecedent. So we have no reason to expect that its processing will proceed in the
same way as an overt pronoun’s.
This chapter asks a specific question in this area: whether it matters to the
timing of anaphora resolution if the surface cue to anaphora is an audible pronoun,
2 There are other sorts of pronouns that do seem to be targeted at grammatical relations, namely
subject-oriented anaphors (Katada, 1991; De Vos, 2012). For example, in Japanese, the subject-
oriented anaphor zibun must be coreferent with a subject. In contrast with the reflexive zibun-zisin,
though, this subject need not be local{. This is illustrated}in (i).
zibun
(i) Johni-ga [Billj-ga Mikek-ni
i/j/*k -no koto-o hanasita to] itta.
zibun-zisin
John-subj Bill-subj Mike-io ?*i/j/*k -gen matter-do told that said
‘John said that Bill told Mike about self.’ (Katada, 1991, p. 289)
223
like “he” in (2), or instead a non-finite participial verb in a control structure, like
“eating” in (3).
(3) Mickey Mouse talked to Minnie Mouse before eating.
A speaker can use (3) just like (2), to say that Mickey talked to Minnie before he,
Mickey, ate. But now the process of understanding the speaker will be different in
two important ways. First, the understanding that it was Mickey who ate depends
on different sorts of information in the two cases. For (3) but not (2) the sentence
itself determines the interpretation, due to its structure and meaning. Only (2)
can be used to say something else—for example, that Mickey talked to Minnie
before Donald ate—since “he” may be used to refer to any male salient in the
discourse. Second, only (2) contains an overt and unambiguous sign of reference to
a particular eater, namely “he.” For (3), in contrast, the signal of such reference
is both implicit and temporarily ambiguous. It is implicit, because it consists not
in the lexical semantic properties of any one word, but in the fact that “eating” is
here the predicate of a non-finite clause; and such clauses, when lacking an audible
subject, generally have their entailed subject role filled anaphorically. And it is
temporarily ambiguous, because the signal of anaphora is in principle uncertain
until after the word “eating,” as (4) makes plain. Here “eating” is not the predicate
of any clause, but rather the gerundival subject of one, and on this use it does not
signal anaphoric reference to any particular eater.
(4) Mickey Mouse talked to Minnie Mouse before eating was forbidden.
224
Determining whether these differences in understanding synonymous uses of
(2) versus (3), or similar pairs with temporal or other adjuncts with control de-
pendencies, impact the timing of reference resolution is an important step towards
better a understanding of what aspects of processing generalize across all referential
dependencies, and conversely, a better understanding of what drives the aspects of
processing that appear to depend on properties of the referential expression and
the relevant information sources. It may also provide evidence about when PRO is
identified and whether it is actually interpreted in online comprehension.
6.2 Experiment 1
Our first experiment examines the timecourse of interpretation of the implicit
anaphora in adjunct control structures to that of overt pronouns in similar sentences.
With a design similar to that of Arnold et al. (2000), we used visual-world eyetrack-
ing to measure the timecourse of interpretation of temporal adjuncts such as those
in (1). In several experiments (e.g. Arnold, 1998; Arnold et al., 2000; Arnold & Lao,
2015), it has been shown that upon encountering a pronoun, comprehenders consis-
tently look to the image of the character in a visual-world scene that they believe
the speaker is using the pronoun to refer to. This is taken as evidence for the timing
of resolution of the pronoun, especially when the task encourages participants to
look to what the sentence is talking about by having them verify the sentence they
hear against the picture it describes. We similarly expect that upon encountering
the non-finite verb in a sentence like (1a), comprehenders will look to the referent
225
evoked by the implicit subject if it is being interpreted immediately. Looks to the
correct referent will therefore be taken as evidence that reference resolution has
been successfully completed. We measure the timecourse of looks to that referent,
comparing it to the timecourse of the interpretation of overt pronouns as a baseline.
Despite the evidence that many language comprehension processes are rapid
and incremental, there are several reasons one might expect the processing and in-
terpretation of the implicit subject in adjunct control structures to proceed slowly.
Two such reasons were mentioned in the previous section. In contrast to when
there is an overt pronoun, control structures have no overt morpheme dedicated
to reference in their subject position. Instead, the first indication that a referent
is needed is given indirectly by the non-finite morphology (-ing) on the embedded
verb.3 Therefore, if processing is taking place incrementally, then at the embedded
verb, comprehenders must identify the verb and the event it corresponds to, no-
tice that the verb is missing a subject, and determine which character the speaker
intended to refer to with that missing subject. Furthermore, as illustrated in (5),
whether a control relation is even necessary may not become clear until several words
after the verb. If (5a) is spoken on its own, then an anaphoric dependency for the
implicit subject is necessary; it must be understood that Mickey talked to Minnie
before he, Mickey, ate pizza at the park. But if (5a) is continued with (5b), then the
nonfinite clause becomes the gerundival subject of the adjunct, and no anaphoric
dependency is necessary. Because the structure of the adjunct will not be clear to
3 Note that this also differs from complement control, in which the cue for control is given one
word earlier, at to.
226
the incremental processor at the point of the verb, it is possible that comprehenders
will wait to interpret the implicit subject until the structure is disambiguated.
(5) a. Mickey talked to Minnie before eating pizza at the park on Sunday. . .
b. . . . was forbidden.
To examine the timecourse of interpretation of the implicit subject in adjunct
control, the current experiment will compare it to the interpretation of a pronoun
intended to refer to the same character, as in (6).
(6) Mickey was talking to Minnie in front of a huge tree before eating/he ate
pizza.
Both of these expressions involve coreference with the main clause subject. In
general, there is often a strong bias towards pronominal coreference with a prior
subject rather than with an object (e.g. Arnold et al., 2000), and this may be in
part due to the discourse focus of subjects (Grosz et al., 1995). In order to reduce
a priori biases towards coreference with either argument of the main clause, the
experiment included a manipulation of which character was in focus by adding an
additional sentence prior to the one containing the adjunct, as in (7).
(7) a. Hey there’s Minnie! Mickey was talking to her in front of a huge tree
before eating/he ate pizza.
227
b. Hey there’s Mickey! He was talking to Minnie in front of a huge tree
before eating/he ate pizza.
Finally, as a measure of whether a subject bias was still in effect, we included a
control condition with a pronoun coreferring with the main clause object. If, for
example, participants are faster at resolving both subject-oriented pronouns and
PRO than object-oriented pronouns, this may be due simply to a bias to look at the
character corresponding to the subject. If, on the other hand, participants are just
at quick to resolve both types of pronouns, then any difference in the resolution of
PRO is less likely to be purely due to such biases.
6.2.1 Materials and methods
6.2.1.1 Participants
Thirty-three participants (22 female) were recruited at the University of Mary-
land campus, and prior written consent was obtained. Participants were all native
speakers of American English, and were at least 18 years of age (mean age = 22.5).
Each received $10 for their time. Two were excluded for low accuracy on compre-
hension questions (less than 80%), and one for equipment failure, for a total of 30
participants in the analysis.
6.2.1.2 Materials
Stimuli consisted of auditory descriptions of scenes involving two main partici-
pants selected out of a set of four well-known characters: Mickey and Minnie Mouse,
228
and Donald and Daisy Duck. The auditory stimuli each were composed of two sen-
tences: an initial sentence focusing on one of the two characters in the scene (e.g.
Hey, there’s Mickey! ), and a second sentence describing the content of the picture.
The second sentence began with a main clause predicate, with the two characters
as arguments (the subject and the object/indirect object). This was followed by
a prepositional phrase describing the third element of the image in order to draw
attention away from the two characters immediately before the critical region. The
description concluded with a temporal adjunct headed by before, after, or while.
In a 2×3 design, stimuli varied with respect to which character was focused on
initially (the subject or the object of the main clause), and with respect to the ref-
erential expression used as the subject of the temporal adjunct. In what we call the
“PRO” conditions (PRO being a common label for the implicit subject in control
structures (Chomsky, 1981)), the temporal adjunct was non-finite, including an im-
plicit “PRO” subject coreferential with the subject of the main clause. In the other
conditions the temporal adjunct was finite and had an overt pronominal subject
used to refer either the referent of the main clause subject (“pronsubj” conditions)
or the main clause object (“pronobj” conditions). In the critical stimuli and in all
but one of the fillers, the two characters were of opposite gender. The pronominal
subject (he/she) was therefore unambiguous. For the PRO conditions, because the
subject was implicit, the first indication that a referential dependency was required
was the non-finite verb. Within the experiment this cue was also unambiguous,
identifying the main clause subject as the antecedent of the implicit subject of the
temporal adjunct, although a non-finite verb in general is ambiguous in that it can
229
Table 1: Experiment 1 sample stimuli
Condition Preamble
FocObj Look there’s Mickey! Minnie was talking to him in front of a huge
tree
FocSubj Look there’s Minnie! She was talking to Mickey in front of a huge
tree
Temporal adjunct
PRO after putting on a nice new bow, and they seem to be having a good
time.
pronsubj after she put on a nice new bow, and they seem to be having a good
time.
pronobj after he put on a nice new hat, and they seem to be having a good
time.
be followed by continuations that do not require an OC dependency. Initial focus
was manipulated as an attempt to reduce any preferential looking to a single char-
acter based on their salience in the preceding discourse. A sample set of stimuli are
given in Table 1. Each set was paired with a single image.
Auditory stimuli were recorded in a sound-attenuated room by a single male
speaker of American English. Each stimulus set was recorded in parts, with each
part in a carrier phrase. After normalizing for intensity, the two recordings were
spliced together at the point indicated in (8) by the vertical bar. This was done
in order to minimize potentially confounding differences in the acoustics of the
stimuli before the critical word. The splice point immediately preceded the non-
finite verb in the PRO conditions, and the pronoun otherwise, so that there would
be no coarticulatory clues about the content of the upcoming critical word. After
splicing, the stimuli were filtered to remove noise. All recording and processing of
230
the auditory stimuli was done in Praat (Boersma, 2001; Boersma & Weenink, 2017).
The average duration of the critical word in the PRO conditions (the non-finite verb)
was 314 ms. In the pronsubj condition, the average critical word duration was 153
ms, and in the pronobj, it was 152 ms.
(8) Look there’s Mickey! Minnie was talking to him in front of a huge tree after
| Tom left.
Thirty sets of auditory stimuli were distributed across six lists using a Latin
square design, such that each participant saw five trials from each condition. Twenty
fillers were also included in each list, for a total of 50 items. The fillers were similar
in form to the critical items, and contained an equal number of PRO, pronsubj,
and pronobj sentences. The only difference was that the filler stimuli contained
discrepancies between the picture and the auditory description of it; the location
of the discrepancy was balanced throughout all parts of the description across filler
items. This was done to encourage participants to attend the visual stimulus at the
corresponding point in the description. Two practice items were also created: one
where the description matched the image, and one where it did not.
Visual stimuli consisted of scenes containing two characters located on the
bottom right and bottom left corners of the picture, equally distant from the center
and of roughly the same size. The stimuli were balanced with respect to which two
characters were seen together, as well as which character appeared on the left or
right side of the screen. The picture also contained a third prominent, inanimate
231
(a) Sample visual stimulus (b) Areas of interest (blue)
Figure 2: Experiment 1: Sample visual stimulus and regions of interest.
Figure (a) is the image displayed for the sample auditory stimuli from
Table 1. In Figure (b), the dotted black rectangles give the regions in
which the two characters were contained, and the dotted blue rectangles
show the regions of interest used in analyzing the eyetracking data. Any
fixations not in those regions were considered looks to “other.”
element that was placed at the top center. A sample image is given in Figure 2(a).
The two main regions of interest in the eyetracking analysis were fixed across trials,
and included the entire image of the character and a small surrounding area to allow
for gaze drift. This is illustrated in Figure 2(b). Any gazes not in these two areas
of interest were classified as looks to “other.”
6.2.1.3 Procedure
Eye movements were recorded using an EyeLink 1000 tower mounted eye-
tracker (S.R. Research, Mississauga, Ontario, Canada), interfaced with a PC, with
a sampling rate of 1000 Hz. Visual stimuli were displayed on a 23-in. LCD moni-
tor, approximately 104 cm away from where participants were seated. Viewing was
binocular, but only the right eye was recorded. Auditory stimuli were presented at
a comfortable volume using stand-alone speakers situated next to the monitor.
232
The experiment was implemented using the Experiment Builder software (S.R.
Research, http://www.sr-research.com/eb.html). At the beginning of the exper-
iment, participants were introduced to the four characters to ensure familiarity.
This was followed by a nine-point calibration procedure, and presentation of the
two practice items. Stimuli were then presented in five blocks of 10 items each,
with recalibration at the beginning of each block (and otherwise as needed) and a
drift-correction fixation point at the center of the screen to begin each trial. Vi-
sual and auditory stimuli were presented simultaneously, with the image remaining
on the screen for the duration of the auditory description. At the conclusion of
the auditory description, the image disappeared, and participants were asked to
indicate whether the description matched the image. The comprehension question
was always the same: participants indicated whether the description and the image
matched by pressing a button on a remote control.
Each participant was presented with the items from one of the six lists. The
order of experimental and filler items within that list was randomized for each
participant.
6.2.1.4 Data analysis
The minimum accuracy for participant inclusion on the comprehension ques-
tions was 80%. Two participants were excluded for low accuracy. One additional
participant was excluded for equipment malfunction. For the remaining participants
(n=30), eye movements were analyzed in the window from the onset of the critical
word using the eyetrackingR package (Dink & Ferguson, 2015) in R (R Core Team,
233
2014). Twenty-two trials were excluded due to high trackloss (greater than 25%),
resulting in a loss of 2.4% of the data. The mean trackloss per trial in the remaining
data was 2.2%.
In order to estimate when the correct interpretation was achieved in each
condition, we compared the divergence in looks to the characters corresponding to
the subject and the object within a single condition using pairwise cluster-based
permutation analyses with t-tests (Maris & Oostenveld, 2007) on data binned into
5 ms bins. The point at which looks to the two characters diverged was taken to be
evidence that successful interpretation was achieved. However, this analysis did not
allow for a direct comparison to measure significance of differences in timing across
conditions.
In order to compare conditions, mixed-effects growth curve models (Barr, 2008;
Mirman, Dixon, & Magnuson, 2008) were used. These analyses were used to deter-
mine whether our manipulations affected the pattern of looks to the target character
or the competitor in order to better understand any differences in timing seen from
the cluster-based analyses. The window of analysis in the growth-curve models ex-
tended from 100 to 1025 ms after the onset of the critical word (the non-finite verb
in the PRO conditions and the pronoun otherwise). The beginning of this window
was chosen empirically by plotting the grand mean of looks to the target and iden-
tifying the first consistent upward inflection point in the grand mean curve (Barr,
2008; Law II, Mahr, Schneeberg, & Edwards, 2016). The end of the window was
similarly chosen empirically, by identifying the first consistent downward inflection
point in the grand mean curve of looks to the target. Prior literature has found
234
that roughly 200–250 ms is required to plan an eye movement in response to stim-
ulus onset (see Matin, Shao, & Boff, 1993, for discussion), which is well within our
window of analysis.
Models were generated using the lmerTest package (Kuznetsova, Brockhoff, &
Christensen, 2014), which uses Satterwaithe approximations to calculate degrees of
freedom in determining significance. The pronsubj condition was used as the referent
condition, and the PRO and pronobj conditions were compared pairwise against it.
The empirical logit of the proportion of looks to each area of interest (AOI) was
combined into 25 ms time bins. Initial models indicated no significant effect of
the focus manipulation. Therefore, in the models reported here, data across the
focus manipulation were combined. These models included fixed effects Cue (type
of referential expression), Time, Time2, and when it resulted in a better fitting
model, Time3, as well as interactions between Cue and the time variables. The time
variables were coded as orthogonal polynomials in order to eliminate correlations
among the parameter estimates. Participants and items were included as random
effects, including the time variables as intercepts. In the results of these analyses,
an effect of Cue indicates a difference in the intercept of the model. Because the
other variables are coded as orthogonal polynomials, an effect of cue reflects an
overall difference in proportion of looks based on cue across the region of interest.
An effect of the Time variable reflects a difference in the overall increase or decrease
of looks over time. A Time2 effect captures differences in how much the curve is
bent upwards or downwards, and Time3 reflects additional curvature. See Winter
and Wieling (2016).
235
Target Competitor Target Competitor
1.00 Focus 1.00 Focus
competitor competitor
target target
0.75 0.75
0.50 0.50
0.25 0.25
0.00 0.00
−500 0 500 1000−500 0 500 1000 −500 0 500 1000−500 0 500 1000
Time relative to onset of critical word (ms) Time relative to onset of critical word (ms)
(a) Effect of focus on PRO items (b) Effect of focus on pronsubj items
Figure 3: Experiment 1: Effect of focus manipulation on PRO (Figure
(a)) and pronsubj (Figure (b)) items. The non-focused character received
more looks in the region prior to the critical word, but this effect disap-
pears before the onset of the non-finite verb in the PRO condition. In
the pronsubj, the effect lingers marginally, but disappears before other
effects. The x-axis represents time in milliseconds from 500 ms before
the onset of the critical word to 1000 ms after.
6.2.2 Results
6.2.2.1 Effect of focus
The effect of focus was measured on looks to the target and competitor for each
of the three cues using growth curve analyses. Initial focus did cause significantly
more looks to the character who was not initially focused immediately preceding
the critical word, but this effect did not continue into the critical region of interest,
where the focus manipulation had no significant effect, as illustrated for the PRO
condition in Figure 3(a). The only possible exception was for looks to the target
in the pronsubj conditions, as illustrated in Figure 3(b). Here there were marginally
more looks to the subject between 50 and 225 ms after the onset of the pronoun
236
Proportion of looks
Proportion of looks
PRO Pronsubj Pronobj
1.00
0.75
Character
0.50 Object
Subject
0.25
0.00
0 250 500 750 1000 0 250 500 750 1000 0 250 500 750 1000
Time from onset of critical word (ms)
Figure 4: Experiment 1: Divergence point. Proportion of looks to the
character corresponding to the main clause subject (red) or object (blue)
for the PRO, pronsubj, and pronobj conditions. The x-axis represents time
in milliseconds from the onset of the critical word.
when the initial focus was on the character corresponding to the matrix object
(p = 0.08 post-clustering in a cluster-based permutation analysis). This marginal
effect was likely spillover from effects in the pre-critical region, and it disappears by
250 ms after the onset of the pronoun. Because this was earlier than any effect of
the referential expression, as discussed below, and because there was no significant
effect of initial focus in any other condition, the data across the focus manipulation
were combined in further analyses.
6.2.2.2 Estimating time of interpretation
Looks to the subject or object at the critical word based on condition (PRO,
pronsubj, or pronobj) are plotted in Figure 4. Cluster-based analyses revealed that
looks to the subject and object diverged significantly in the PRO condition at 295
237
Proportion of looks
ms after the onset of the non-finite verb, in the pronsubj condition at 345 ms after
the onset of the pronoun, and in the pronobj condition at 370 ms (p < 0.05 pre-
clustering, p < 0.001 clustering from this time to the end of the window, in each
case). This is consistent with the results in Arnold et al. (2000) for the interpretation
of overt pronouns whose gender features unambiguously identify a referent. Once
again, however, this analysis does not allow for a direct comparison between the
conditions. Although the cluster-based analysis gives us rough timing estimates, it
does not tell us whether the differences in timing across conditions is significant.
6.2.2.3 Effect of Cue
Looks to the target or competitor across the three cue conditions are given in
Figure 5. Looks to the target began to increase and looks to the competitor began
to decrease by around 250 ms in each condition, continuing through 1025 ms post
stimulus onset. A mixed-effects growth curve analysis indicated a significantly lower
increase in looks to te target in the PRO condition overall compared to the pronsubj
condition, represented by an interaction of cue type and Time (t = −2.616, p <
0.01), reflecting the fact that the peak proportion of looks to the target in the PRO
condition was lower. But although the total increase in looks was lower, looks to the
target in the PRO condition appear numerically to begin to increase at an earlier
point than in either pronoun condition. There was no significant difference in the
total proportion of looks across conditions, nor were there any differences in the time
terms between the pronobj and pronsubj conditions in looks to the target. A mixed
effects growth curve model on looks to the competitor revealed a marginal difference
238
Target Competitor
1.00 Cue
PRO
Pronsubj
0.75 Pronobj
0.50
0.25
0.00
0 250 500 750 1000 0 250 500 750 1000
Time from onset of critical word (ms)
Figure 5: Experiment 1: Effect of Cue. Proportion of looks to the
target or competitor characters. In the PRO (red) and pronsubj (green)
conditions, the target corresponds to the main clause subject, and the
competitor to the main clause object; in the pronobj (blue) condition,
this is reversed. The x-axis represents time in milliseconds from the
onset of the critical word (the non-finite verb in PRO conditions, and
the pronoun in the other conditions). Shaded area in this and other
figures represents standard error.
in the Time variable between the PRO and pronsubj conditions (t = 1.94, p = 0.054),
with a higher rate of looks to the competitor in the pronsubj condition, as well as a
significant difference between the Time3 variables of the PRO and pronsubj conditions
due to early differences between the two (t = −2.119, p < 0.05). There were no
other significant effects.
6.2.3 Discussion
Experiment 1 investigated the process of reference resolution for the implicit
(PRO) subject of non-finite temporal adjuncts in comparison to overt pronouns
used to refer to the character corresponding to either the subject or the object
239
Proportion of looks
of the main clause. For the implicit subject in the temporal adjuncts used here,
reference is determined based on structural features of the sentence. For overt
pronouns, on the other hand, reference resolution is influenced by discourse factors,
guided by the morphological features of the pronoun; structural features are only
important insofar as they affect information structure. As the focus manipulation
had no significant effect on looks in the critical window, reference resolution in this
experiment appeared to be possible without regards to information structure.
Our results showed a numerical difference in the timing of reference resolution,
as demonstrated by divergence points in looks to either character (Figure 4), with
the divergence in the PRO condition occurring roughly 50 ms earlier than in the
pronsubj condition. Although the analysis of these divergence points in each of the
three conditions provides an estimated point of interpretation, it does not directly
compare timing differences between conditions. However, the growth-curve analysis
indicated significantly more looks to the competitor in the pronsubj condition than
in the PRO condition, indicating a relative lack of competition between the two
characters in the PRO condition. This difference is likely to have contributed to the
difference in timing seen. Still, although the significant difference in looks to the
competitor is suggestive that the timing differences are real, the strongest claim we
can make about timing is that the interpretation of PRO appears to be at least as
fast, if not faster, than the resolution of a pronoun referring to the same character.
This fast interpretation of the implicit subject is somewhat surprising consid-
ering the differences in which features are important for reference resolution between
PRO and pronouns, and despite large differences in the cue in terms of semantics,
240
morphology, and duration (duration of the non-finite verb in the PRO condition was
twice as long as the duration of the overt pronouns in the other conditions, with an
average of 314 ms vs. 153 ms).
6.2.3.1 A general subject preference?
The results of Experiment 1 suggest that the interpretation of the implicit
subject in adjunct control is just as fast, if not faster, than the interpretation of
overt pronouns. Importantly, this cannot be due solely to a general subject bias in
the materials. It is true that two-thirds of our items required participants to look
at the subject at the critical word, since that was the case for both the PRO and
the pronsubj conditions. There might also have been a general subject bias in the
critical region if these sentences led to the subject’s referent being predicted as the
likeliest next mentioned character (Kehler & Rohde, 2013). But if a subject bias
were strongly contributing the speed of interpretation of the implicit subject, then
we would expect the PRO condition to be identical to the pronsubj condition, which
was not the case; there were significantly more early looks to the competitor in the
pronsubj condition.
There are at least two additional reasons not to think the results of this ex-
periment stem solely from a subject bias. First, the focus manipulation was in-
cluded in order to reduce the overall preference for looking at one character over the
other. This manipulation succeeded in that the proportion of looks to the character
that was initially focused was lower than looks to the unfocused character immedi-
ately preceding the critical word. Importantly, this effect disappears either before
241
or rapidly after the onset of the critical word, as illustrated in Figure 3, suggest-
ing that participants were not relying solely on general looking preferences at that
point. Instead, the gender cues on the pronouns and the structural information for
the implicit subject’s interpretation were sufficient for participants to quickly and
unambiguously select a referent without regard to information structural biases for
either character.
Second, there was no significant difference in looks to either character in the
two pronoun conditions. In other words, participants were just as fast to look
at the correct character after hearing a pronoun, whether that character was the
subject or the object. In the pronobj condition, this could not have been due to a
subject preference, and yet it was no slower than looks to the subject in the pronsubj
conditions. Participants were also just as likely to look at the competitor upon
hearing a pronoun, whether that competitor corresponded to the subject or object.
6.2.3.2 Rapid use of structural information
Significantly fewer looks to the competitor in the PRO condition seems to have
contributed to fast resolution of the implicit subject. But why would the interpre-
tation of PRO be so fast? One possible explanation might be that in searching for
the intended antecedent of a pronoun, both the subject and the object are initially
considered, but for PRO, only the subject is. This would be consistent with the
idea that participants use only structural information in searching for a referent for
PRO, as the main clause subject was the only structurally accessible antecedent.
For pronouns, the search could at least initially include both characters. Accord-
242
ing to a family of two-staged models of pronoun interpretation, comprehenders first
search the discourse for all possible referents, and then after check them against
the features of the pronoun (but see Kehler, 2008, for criticisms of such models).
In our experiment, this would mean that participants would hear the pronoun and
activate both discourse characters, and only then filter out the character that did
not match the gender of the pronoun. This does not mean that both characters
in the discourse will have equal consideration. Pronouns are often more likely to
be used to refer to a preceding subject (Crawley et al., 1990; Grosz et al., 1995),
at least when the discourse seems to tell a story as a narration (Kehler & Rohde,
2013). Therefore, it could be that the referent corresponding to the subject would
receive greater consideration during the filtering stage than the one corresponding
to the object, although the pronsubj condition did not differ significantly from the
pronobj condition.
However, if this two-stage model were to account for the slightly faster res-
olution of PRO, it would need to somehow be stated so that all the referents are
not automatically retrieved in the first stage, but only the structurally licit referent
described by the subject. Nothing in the two stage model independently motivates
this. Additionally, having separate strategies for pronouns and PRO is different
from what cue-based retrieval accounts (e.g. Lewis & Vasishth, 2005) would predict,
in which the retrieval process would happen in a single step for both involving search
of all antecedents based on available cues.
243
6.2.3.3 Prediction in the processing of PRO?
Another possible explanation for why interpretation was so fast in the PRO
condition is that participants may have been actively predicting a control structure
in the adjunct. There are two potential sources for such a prediction in this exper-
iment. Perhaps the sentence contexts in the main clause of the items, possibly in
combination with the scene depicted on the screen, had some intrinsic property or
set of properties that made them likely to be continued with a control structure.
This possibility is tested in Experiment 2. Or, it may have been the case that
the experimental context had a large enough proportion of control structures that
participants correspondingly began predicting PRO. Under either possibility, the
speed of looks to the correct character in the PRO condition would not necessarily
reflect the speed of bottom-up interpretation of PRO. Experiment 1 leaves open
the possibility that the process of resolving the reference of PRO takes longer than
resolution of a pronoun if prediction of a control structure led participants to begin
that process earlier in the PRO condition.
There are at least two reasons why this seems to be a likely possibility. First,
looks to the target in the PRO condition began to increase around 125 ms after
the onset of the critical word. Given that saccade planning is generally assumed
to take 200–250 ms (Matin et al., 1993), this increase appears to be too early.
As noted above, this cannot only be due to a general preference to look at the
subject, since a similar early increase was not seen in the pronsubj condition. Instead,
it is possible that participants initially predicted a control structure and started
244
Target Competitor
1.00 Cue
PRO
Pronsubj
0.75 Pronobj
Cue
0.50
0.25
0.00
0 250 500 750 1000 0 250 500 750 1000
Time from offset of critical word (ms)
Figure 6: Experiment 1: Results from offset of critical word. Proportion
of looks to the target or competitor characters. The x-axis represents
time in milliseconds from the offset of the critical word.
planning a saccade towards the target character before the onset of the critical
word. When they heard the beginning of the critical word, if it was anything other
than a pronoun, they would have had no reason to doubt their prediction, and
would have continued to look at the subject. However, if they heard the onset of a
pronoun, then they would have reason to discard their prediction and cancel that
saccade.4
Second, when the results are plotted from the offset of the critical word instead
of the onset (Figure 6), it becomes especially clear that looks to the target in the
PRO condition began surprisingly fast. For pronouns, looks to the target did not
begin to rise until shortly after the pronoun offset, diverging significantly from looks
to the competitor within 225 ms. Because the non-finite verbs were on average
4 See Curtis, Cole, Rao, and D’Esposito (2004) on the speed of canceling a saccade.
245
Proportion of looks
around 160 ms longer than the pronouns, if reference resolution were occurring
at the same rate, we would expect looks to the target in the PRO condition to
diverge from looks to the competitor perhaps 160 ms faster when plotted from
critical word offset. And this is assuming that the same amount of auditory input
would be required in each case, which would be surprising given the differences in
the complexity of the cues. However, looks to the target in the PRO condition were
even faster than that, diverging from looks to the competitor within 10 ms of the
offset of the critical word—215 ms faster than in the pronoun condition, and roughly
55 ms earlier than could possibly be expected, unless the control structure received
expedited processing due to prediction. Experiments 2 and 3 were designed to assess
and reduce the likelihood that participants would predict a control structure.
Importantly, if prediction did lead to faster looks in the PRO condition, then
this prediction evidently did not result in an overall prediction of likely reference to
the character associated with the main clause subject. As already noted, if partic-
ipants were predicting subject reference, then we would expect looks in the PRO
and pronsubj conditions to be matched. If prediction is involved, then facilitation
in reference processing appears to be strictly tied to successful prediction about the
surface form.
6.2.3.4 Learned strategy?
A final factor that may have increased the speed of looks in the PRO condition
is that participants could have developed a strategy within the experiment to look
to the subject any time they did not hear “he” or “she.” If this were the case,
246
participants would not need to resolve the anaphoric control dependency at all in
the critical window. Although participants would still have to interpret the pronoun
in other conditions, hearing anything besides the pronoun would be enough to tell
them to look at the subject. This possibility is removed in Experiment 3. Such a
strategy could have independently caused artificially fast looks in the PRO condition,
or in tandem with prediction of a control structure, in which case participants would
actively predict a control structure, and only abandon that prediction if the onset
of a pronoun was heard.
6.3 Experiment 2
Looks to the character corresponding to the subject in the PRO conditions in
Experiment 1 were initially faster than looks to the correct character in the pronoun
conditions. Experiment 2 consisted of a cloze task designed to test whether the
critical items of Experiment 1 were likely to induce prediction of a control structure
after the preposition in the adjunct. Such a prediction may have facilitated looks
to the correct character when anything other than a pronoun was encountered in
Experiment 1, resulting in unexpectedly fast looking times in the PRO condition.
6.3.1 Materials and methods
6.3.1.1 Participants
Sixty participants (39 female) were recruited via Amazon MechanicalTurk, and
electronic consent was obtained. Participants were all native speakers of American
247
Table 2: Experiment 2 sample cloze items
Condition Text
FocObj Look there’s Mickey! Minnie was talking to him in front of a huge
tree after .
FocSubj Look there’s Minnie! She was talking to Mickey in front of a huge
tree after .
English, and had passed a native speaker proficiency test. Participants were at least
18 years of age (mean age = 37.5), and received $4 in compensation.
6.3.1.2 Materials
The materials for this experiment were based on the 30 critical items from
Experiment 1. For each item, participants were presented with the visual image and
the text of the description up to the preposition in the adjunct clause (see Figure 2(a)
and Table 2). Following the manipulation in Experiment 1, items varied within
subjects based on which character was initially focused. Items were randomized in
presentation.
6.3.1.3 Procedure
The experiment was presented using the Ibex platform (Alex Drummond:
http://spellout.net/ibexfarm). For each item, participants were presented with the
visual stimulus, with the text to be completed directly below. Participants were
asked to complete the sentence as quickly as possible with whatever first came to
mind. They were instructed that the sentence was intended to describe the image,
248
and that they could use the image as they decided how to complete the sentence.
Participants were familiarized with the task with two practice items.
6.3.1.4 Data analysis
Responses were coded for type (whether the response began with a non-finite
verb in a control structure (‘PRO’ responses), a pronoun, or other) and referent
(whether the subject of the adjunct, implicit or explicit, appeared to corefer with
the subject of the matrix clause only, the object only, both the subject and the
object, or some other entity). Intended referent was determined based on the entire
content of the response. If the response did not disambiguate the intended referent,
it was coded as unclear reference. The effect of focus on response type and referent
was tested with multinomial log-linear regression models using the nnet package in R
(Venables & Ripley, 2002). P-values were obtained with ANOVAs on the regression
models. Significance of differences in the proportions of type or referent of responses
within a single focus condition or across focus conditions was measured using linear
mixed-effects models (Bates, Mächler, Bolker, & Walker, 2015), with participants
as random effects. P-values for the linear mixed-effects models were obtained using
the lmerTest package (Kuznetsova et al., 2014).
In order to test whether predictability of control structures led to faster inter-
pretation in the PRO condition of Experiment 1, we fit mixed-effects growth curve
models to the empirical logit of the proportion of looks to the target and competitor
within the PRO condition, with the cloze probability for control structures in each
item obtained in Experiment 2 as a fixed effect.
249
1.00
0.75
Response
PRO
0.50 Pronsubj
Pronobj
other
0.25
0.00
FocObj FocSubj
Focus condition
Figure 7: Experiment 2: Responses by type. Proportion of responses
beginning with a control structure (PRO responses), a pronoun coref-
erential with either the subject or object or the main clause, or other,
based on whether the prompt had initial focus on the subject or the
object.
6.3.2 Results
The proportion of responses of each type are given in Figure 7. Responses be-
gan with control structures in 52% of responses, with a pronoun coreferential with
the main clause subject in 17% of responses, and with a pronoun coreferential with
the main clause object in 15% of responses. A linear mixed-effects model revealed
that control structures (PRO) were significantly more frequent than responses be-
ginning with either type of pronoun (PRO vs. pronsubj: t = 12.77, p < 0.001; PRO
vs. pronobj: t = −14.57, p < 0.001. Multinomial log-linear regression also showed a
significant effect of focus condition on response type (χ2 = 14.04, p < 0.001). Focus
also had a significant effect on the proportion of responses beginning with a control
250
Cloze probability
1.00
0.75
Referent
subj
obj
0.50
both
other
unclear
0.25
0.00
FocObj FocSubj
Condition
Figure 8: Experiment 2: Responses by referent. Proportion of responses
with clear initial coreference with the main clause subject, object, both,
or other, based on whether the prompt had initial focus on the subject
or the object.
structure (χ2 = 13.75, p < 0.001), with fewer PRO responses given in the FocObj
condition (45% vs. 56%).
Responses were also coded for which character was the intended subject of
the response. For overt pronouns, the intended referent was inferred based on the
gender and number of the pronoun. For responses containing a control structure,
intended referent was inferred strictly based on disambiguating information in the
rest of the sentence. For example, some responses were consistent with both char-
acters performing the action together. When the rest of the context did not make
it clear who the intended referent was, the response was coded as such, despite
the high probability that coreference with the main clause subject was intended.
This was done in order to not presuppose that participants were always intending
OC interpretations. Figure 8 gives the proportion of responses whose subject cor-
251
Proportion responses
Target Competitor
1.00 Cloze of 
V-ing
high
0.75 low
0.50
0.25
0.00
0 250 500 750 1000 0 250 500 750 1000
Time from onset of critical word
Figure 9: Effect of cloze on Experiment 1. Proportion of looks to the
target and competitor in the PRO condition of Experiment 1 based on
the cloze probability of a control structure in each item in Experiment 2.
The high cloze group consists of all items for which control structures
had a cloze probability of > 0.5 (mean = 0.619, sd = 0.078). The low
cloze group consists of all items with ≤ 0.5 cloze (mean = 0.421, sd =
0.069).
responded to each referent. A multinomial regression indicated no effect of focus
condition on intended referent (χ2 = 5.18, p > 0.1). Overall, responses indicat-
ing clear coreference with the main clause subject were significantly more frequent
than responses with clear coreference with the main clause object (59% vs. 19%,
t = −17.44, p < 0.001). This was largely due to responses with control structures.
When these responses are removed from the analysis, coreference with the main
clause subject was no more frequent than coreference with the main clause object
(39% vs. 35%, t = −0.93, p > 0.1).
The proportion of control (PRO) responses varied across individual items from
28% to 76%. Figure 9 illustrates the effect of the cloze probability of a control
252
Proportion of looks
structure on looks in the PRO condition of Experiment 1 (although the data in the
figure is split into two groups—high and low cloze—cloze was a continuous variable
in the growth curve model). The cloze probability of a control structure had only
a marginal main effect on total looks to the target (t = 1.717, p = 0.091), which
were at least numerically higher for items corresponding to those with high cloze for
control structures. There was a significant main effect of cloze on looks to the com-
petitor (t = −2.367, p < 0.05); higher cloze for control structures in Experiment 2
led to fewer looks to the competitor in the PRO condition in Experiment 1.
6.3.3 Discussion
This experiment tested one possible explanation for why looks to the target
referent were so fast in the PRO condition of Experiment 1. If the preambles of the
critical items led to prediction of an upcoming control structure, then participants
may have been able to use that prediction to aid their interpretation, leading to faster
looking times than when a less-predicted pronominal subject was heard. The overall
preference for control structure continuations compared to continuations beginning
with a pronoun in Experiment 2 is consistent with this interpretation. Because
control structures were more predicted than other continuations, looks to the subject
may have been faster in Experiment 1 when that expectation was realized.
In contrast to Experiment 1, the focus manipulation in Experiment 2 had a
significant effect on PRO responses. Because PRO continuations were significantly
more frequent in the FocSubj condition, it would be reasonable to expect looks in
Experiment 1 to have been even faster in PRO conditions when initial focus was on
253
the subject. This was not the case. However, although control continuations were
fewer in the FocObj condition, they were still significantly more frequent than con-
tinuations beginning with a subject-oriented pronoun or an object-oriented pronoun.
It may therefore be the case that control structures were still predicted enough to
cause facilitation in reference resolution in the PRO condition of Experiment 1.
The effect of cloze on looks in Experiment 1 was directly tested using a growth-
curve analysis. As illustrated in Figure 9, cloze had a significant main effect on
the proportion of looks to the competitor: higher cloze meant fewer looks overall.
This suggests that individual item contexts may have led to prediction of control
structures, which then may have facilitated rapid resolution of reference in the PRO
condition. Looks to the target were also numerically affected by cloze, but this effect
was not significant. One possible reason for this is that even the relatively low cloze
may have been high enough in comparison to other continuations to cause facilitated
looks to the target. However, it is also possible that cloze did not significantly affect
looks to the target because participants were already using the strategy mentioned
above of looking to the target whenever anything other than a pronoun onset was
heard in the critical region, and that prediction based on cloze did not add any
additional speed.
Experiment 2 also tested whether participants were more likely to continue the
adjunct with reference to the character corresponding to the main clause subject
or object. Because of the large proportion of PRO continuations, there was a cor-
respondingly large amount of coreference with the subject. However, when control
structures are removed, we did not find any independent subject preference. When
254
they did not use a control structure, people were just as likely to refer to either
character. This is consistent with the lack of a significant subject preference effect
seen in the pronoun conditions of Experiment 1.
6.4 Experiment 3
Experiment 3 was designed to further investigate why resolution of PRO in
Experiment 1 appeared to be faster than resolution of a pronoun. This experiment
included all the items from Experiment 1, in addition to extra fillers designed to to
remove the possibility of assuming that every non-pronoun onset in the temporal
adjunct was the onset of a non-finite verb and to reduce the overall proportion of
control structures. Additional fillers were also added to reduce within-experiment
bias to refer to the subject character. Although these changes do not affect the
predictability of control for any individual critical item, they do change the overall
biases within the experiment, which in turn may affect participants’ interpretive
strategies.
6.4.1 Materials and methods
6.4.1.1 Participants
Thirty-one participants (20 female) were recruited at the University of Mary-
land campus, and prior written consent was obtained. Participants were all native
speakers of English, and were at least 18 years of age (mean age = 21). Thirty of
the participants were compensated with course credit, and one with $5. One partic-
255
Table 3: Experiment 3: Sample fillers
Condition Stimulus
Filler Look there’s Donald! Minnie found him outside of Daisy’s house
(true) after Daisy kicked him out for being rude.
Filler Look there’s Donald! He brought Daisy to the pool after summer
(false) had already ended and it had closed.
ipant was excluded for high trackloss (>25%), for a total of 30 participants in the
analysis.
6.4.1.2 Materials
All stimuli from Experiment 1 were included. Forty new filler items were
included In addition to the 20 from Experiment 1. These consisted of 10 items
similar to the original fillers, but all with a pronoun coreferring with the main
clause object as the subject of the temporal adjunct clause. This was included in
order to reduce the overall proportion of items with subject reference in the adjunct.
The remaining 30 new fillers had a subject in the adjunct referring to something
or someone other than the two main characters; often this was the third prominent
element in the image (see Table 3 and Figure 10). Half of these 30 fillers were true
(i.e. the image matched the description), and half were false, as were the other 10
new fillers. This brought the total number of items each participant saw to 90, 45 of
which should have been judged true. The recording of the auditory stimuli followed
the same procedure as in Experiment 1.
256
(a) Sample true filler image (b) Sample false filler image
Figure 10: Experiment 3: Sample filler visual stimuli.
6.4.1.3 Procedure
Stimuli were presented in nine blocks of 10 items each. The procedure followed
that of Experiment 1 in all other respects.
6.4.1.4 Data analysis
The procedure for data analysis was also the same as in Experiment 1. Mean
proportion of looks to the target across conditions began to increase at 100 ms after
the onset of the critical word and to decrease at 1100 ms, so the window of analysis
was 100–1100 ms. Three trials were excluded for high trackloss (greater than 25%),
resulting in a loss of 0.33% of the data. The mean trackloss per trial in the remaining
data was 0.048%.
257
Target Competitor
1.00 Focus
competitor
target
0.75
0.50
0.25
0.00
−500 0 500 1000−500 0 500 1000
Time relative to onset of critical word (ms)
Figure 11: Experiment 3: Effect of focus manipulation on PRO con-
ditions. The non-focused character received more looks in the region
prior to the critical word, and this effect carries over into critical region.
Similar effects were seen for the other conditions, and there was no inter-
action between focus and cue. The x-axis represents time in milliseconds
from 500 ms before the onset of the critical word to 1000 ms after.
6.4.2 Results
6.4.2.1 Effect of focus
Experiment 3 exhibited a similar effect of focus as in Experiment 1, with a
greater proportion of looks to the non-focused character in the region prior to the
critical word. Unlike in Experiment 1, this effect spilled over into the critical region.
This is illustrated for the PRO condition in Figure 11. The effect was similar in the
pronoun conditions, and there was no interaction between cue focus. Because of this,
and in order to allow a more direct comparison with the results of Experiment 1,
the focus manipulation was again removed from further analysis.
258
Proportion of looks
PRO Pronsubj Pronobj
1.00
0.75
Character
0.50 object
subject
0.25
0.00
0 250 500 750 1000 0 250 500 750 1000 0 250 500 750 1000
Time from onset of critical word (ms)
Figure 12: Experiment 3: Divergence point. Proportion of looks to the
character corresponding to the main clause subject (red) or object (blue)
for the PRO, pronsubj, and pronobj conditions. The x-axis represents time
in milliseconds from the onset of the critical word.
6.4.2.2 Estimating time of interpretation
Looks to the subject or object at the critical word based on condition are
plotted in Figure 12. Cluster-based analyses revealed that looks to the subject and
object diverged significantly in the PRO condition at 410 ms after the onset of the
non-finite verb, in the pronsubj condition at 240 ms after the onset of the pronoun,
and in the pronobj condition at 365 ms (p < 0.05 pre-clustering, p < 0.001 clustering
from this time to the end of the window, in each case).
6.4.2.3 Effect of cue
The effect of cue is illustrated in Figure 13. A mixed-effects growth curve
analysis revealed a significant effect of cue, with the PRO condition receiving signif-
259
Proportion of looks
Target Competitor
1.00 PRO
Pronsubj
Pronobj
0.75
0.50
0.25
0.00
0 250 500 750 1000 0 250 500 750 1000
Time from onset of critical word (ms)
Figure 13: Experiment 3: Effect of Cue. Proportion of looks to the
target or competitor characters. In the PRO (red) and pronsubj (green)
conditions, the target corresponds to the main clause subject, and the
competitor to the main clause object; in the pronobj (blue) condition,
this is reversed. The x-axis represents time in milliseconds from the
onset of the critical word.
icantly fewer looks to the target than the pronsubj condition (t = −3.577, p < 0.001)
and significantly more looks to the competitor (t = 3.112, p < 0.01). There were
also marginally more looks to the target in the pronsubj condition than the pronobj
condition (t = −1.672, p = 0.096). Additionally, there was a significant differ-
ence in the Time variable between the PRO condition and the pronsubj condition
(t = −2.368, p < 0.05), with looks to the target having less overall increase in the
PRO condition.
6.4.2.4 Effect of cloze
The effect of predictability of a control structure based on the cloze results of
Experiment 2 on looks in the PRO condition of Experiment 3 are given in Figure 14.
260
Proportion of looks
Target Competitor
1.00 Cloze of 
V-ing
high
0.75 low
0.50
0.25
0.00
0 250 500 750 1000 0 250 500 750 1000
Time from onset of critical word
Figure 14: Effect of cloze on Experiment 3. Proportion of looks to the
target character in the PRO condition of Experiment 3 based on the
cloze probability of a control structure in each item in Experiment 2.
Cloze had no significant effect on looks to the target or the competitor in the PRO
condition. The only marginal effect was an interaction between the cloze value and
the Time variable in looks to the competitor (t = 1.704, p = 0.094).
6.4.2.5 Plotting from offset of critical word
Additional analyses were performed plotting the data from the offset of the
critical word until the grand mean across conditions of looks to the target began to
decrease (950 ms after the offset), as illustrated in Figure 15. Looks to the subject
and object characters diverge significantly in the PRO condition 115 ms after the
offset of the non-finite verb. In the pronsubj condition, they diverge at 120 ms after
the offset of the pronoun, and in the pronobj condition, at 195 ms.
Growth curve models revealed a significant difference in the Time variable in
261
Proportion of looks
Target Competitor
1.00 Cue
PRO
PronSubj
0.75 PronObj
0.50
0.25
0.00
0 250 500 750 1000 0 250 500 750 1000
Time from offset of critical word (ms)
Figure 15: Experiment 3: Results from offset of critical word. Proportion
of looks to the target or competitor characters. The x-axis represents
time in milliseconds from the offset of the critical word.
looks to the target between the PRO and pronsubj conditions (t = −4.166, p <
0.001). In looks to the competitor, there was a significant main effect of cue, with
more looks in the PRO condition than the pronsubj condition (t = 2.336, p < 0.05),
as well as a significant difference between the PRO and pronsubj conditions in the
Time variable (t = 2.008, p < 0.05) and a marginal difference in Time3 (t =
−1.667, p = 0.097). However, these differences were largely due to the leveling off
of looks in the PRO condition by 500 ms after critical word offset. If the data are
limited to a 500 ms window, there is no significant difference between conditions in
looks to the target, but looks to the competitor still show a difference between PRO
and pronsubj in the Time
2 variable (t = 2.142, p < 0.05).
262
Proportion of looks
6.4.3 Discussion
6.4.3.1 Summary of findings
The results of Experiment 3 show slower anaphora resolution in the PRO
condition than the pronoun conditions (see Figure 12). To reiterate, however, dif-
ferences across conditions in divergence points in looks to the target and competitor
were not tested directly. However, the results of the growth curve analyses are also
consistent with slower resolution of the implicit subject. The PRO condition saw
fewer looks to the target and more looks to the competitor overall, and the increase
in looks to the target was lower for PRO (see Figure 13). This pattern of results
contrasts with the pattern seen in Experiment 1, where there were fewer looks to
the competitor in the PRO condition than the pronoun conditions, and where the
divergence in looks was earlier. This suggests that participants were waiting longer
to resolve the anaphora in the PRO condition of Experiment 3. We attribute this
to the reduced probability of control structure continuations in this experiment and
its potential affect on participants’ predictions. The difference between divergence
points in the two pronoun conditions may have been due to a remaining preference
to interpret the pronoun as referring to the subject, despite the greater number of
items with object-oriented pronouns within the experiment, as evidenced by the
marginal difference in looks to the target in the pronsubj and pronobj conditions.
Although PRO now appears to be interpreted more slowly than an overt pro-
noun, despite a possible preference for subject reference, this difference may be
largely due to differences in the length of the cue to anaphora (the non-finite verb
263
vs. the overt pronoun). When plotting from the offset of the critical word to control
for cue length, differences between conditions disappear.
6.4.3.2 Effect of prediction
Prediction of a control structure seemed to contribute to the rapid interpreta-
tion of PRO in Experiment 1. There were two potential sources for such a prediction.
First, with one third of the items in Experiment 1 containing control structures in the
temporal adjunct, participants may have learned to predict such structures within
the experiment. Or, it may have been that the items themselves, along with their
associated pictures, may have simply had a high probability of containing control
structures. Evidence for this second possibility was the fact that in Experiment 2,
continuations at the adjunct clause were more likely to contain a control structure
than a pronoun referring either to the subject or object. In addition, the cloze value
of control structures in the items of Experiment 2 had a significant effect on looks
in the PRO condition of Experiment 1.
In Experiment 3, since the same experimental items were used, it is possible
that participants made the same predictions about upcoming control structures.
However, the probability of control structure continuations within the experiment
was reduced by adding fillers with adjuncts beginning with subjects other than PRO
or pronouns. This reduced the within-experiment probability of control. Further-
more, even if individual item contexts still led to prediction of a control structure,
it would still have been necessary for participants to listen longer in order to verify
that prediction.
264
That prediction played a lesser role in Experiment 3 is evident in the fact that
the cloze probability of a control structure had no significant effect on looks in the
PRO condition (see in Figure 14), in contrast to what was seen in Experiment 1.
This fact provides further evidence that prediction was involved in Experiment 1,
but not (or to a lesser extent) in Experiment 3. But it is not clear whether this
is due to a reduction of the within-experiment proportion of control structures,
or to the fact that listeners would have had to listen longer to verify predictions
about individual items. In either case, when the experimental context reduced the
likelihood of control structures, interpretation of PRO was slower.
6.4.3.3 Effect of cue length?
The point of divergence in looks to the target or competitor in the PRO
condition was 170 ms later than in the pronsubj condition. This is comparable to the
difference between the duration of the cue in the two conditions: the non-finite verb
in the PRO conditions was on average 160 ms longer than the pronoun in the pronsubj
condition. Therefore, even if resolution of the different forms of anaphora took the
same amount of time after the relevant bottom-up input, the PRO condition could
still be roughly 160 ms slower than the pronsubj condition.
In order to better control for differences in cue length, the data were also
compared beginning at the offset of the critical word, as illustrated in Figure 15. In
this comparison, there is no significant difference between initial looks in the PRO
and pronoun conditions. If anything, looks to the target in the PRO condition are
numerically higher. Major differences only emerge after around 500 ms, at which
265
point looks in the PRO condition level off as participants interpret the remainder
of the sentence. The divergence points between looks to the subject and object are
also similar in the PRO and pronsubj condition, becoming significant at 115 ms and
120 ms, respectively, after the offset of the critical word. This suggests that listeners
were able to resolve the anaphora in the PRO condition as fast as feasibly possible
given the bottom-up input.
Interestingly, the pronobj also differed from the pronsubj condition, with looks
to the subject and object diverging 125 ms later than in the pronsubj condition
when plotted from the onset of the critical word, and only 45 ms earlier than the
PRO condition. From the offset of the critical word, the pronobj condition showed
significant divergence at 195 ms, roughly 75 ms slower than the PRO or pronsubj
conditions. This may have been due to a subject bias in the resolution of the
anaphora leading to fast interpretation in the pronsubj and PRO conditions. If, on
the other hand, this subject bias were specific to overt pronouns in this experiment,
then the speed of interpretation for PRO would be even more impressive, being
only slightly slower than an overt pronoun, and even faster when measured from the
offset of the critical word.
Although plotting from the offset of the critical word does provide some control
over the effect of cue length, this type of analysis admittedly may be problematic,
since the interpretation of any given word may begin before the end of the word.
This may be especially true for the verbs in the PRO condition, in which listeners
may have been able to identify the verb and begin interpreting it before the final
syllable (ing). A different approach to control for cue length would be to normalize
266
it in the actual auditory stimuli by lengthening the duration of the pronouns and
shortening the verbs. This would allow us to plot from the onset of the critical
word and still control for cue duration, but it would also not completely solve the
problem; it would remain unlikely that participants would require the same amount
of time to correctly identify the target word.
6.5 General discussion
The experiments reported here provide evidence for the rapid interpretation
of PRO in temporal adjuncts, suggesting that structural information can be used
quickly in reference resolution. Using pronoun interpretation as a baseline, we found
that when listeners may believe an upcoming control structure to be likely, their
interpretation of PRO is slightly faster than their interpretation of overt pronouns.
When the experimental context reduced the likelihood of a control structure, PRO
is interpreted more slowly, but only insofar as the cue to anaphora is longer. After
the relevant bottom-input is received, listeners are able to resolve both the implicit
anaphora in adjunct control structures and pronominal anaphora at the same rate.
This result is interesting for a number of reasons. First, there are many dif-
ferent kinds of referential expressions and types of anaphora. Each may rely on
syntactic, discourse, and conceptual sources of information in different ways. The
interpretation of PRO in adjunct control structures is heavily dependent on struc-
tural features of the sentence, while pronouns, restricted by their φ features, rely
more on discourse information. These experiments provide evidence that structural
267
information can be used in absence of other features just as efficiently as information
guiding the interpretation of overt pronouns.
One of the differences between PRO and pronouns that make the results of
this experiment surprising is that the cue that a referential dependency is needed
in the PRO sentences, namely the nonfinite verb, does not itself have reference
to an individual as its semantic function. Lexically, the verb expresses an event
concept, or the concept of a relation to an event. The signal of anaphoric reference
to an individual bearing that relation consists in the verb’s grammatical context, a
nonfinite context lacking a subject. While comprehenders interpret the verb, they
must simultaneously find one of its participants in order to establish the referential
dependency necessary to interpret PRO. This is one instance where, as Van Berkum
(2008, p. 376) put it, it is not the case that “[f]irst you recognize each of the words,
then you look up their meaning in your mental dictionary, and then, using syntax
to guide the combination, you simply combine the meanings so that you know what
I said.” Instead, recognition of the nonfinite verb in this particular context activates
both the event concept lexically expressed by the verb, and the cue to anaphoric
resolution of the subject argument.
The non-finite verb does not even unambiguously indicate that a control rela-
tion is necessary. As was seen in the example in (5), repeated in (9), at the non-finite
verb, the sentence could still have a continuation that does not require a referen-
tial dependency between the implicit subject of the adjunct and the main clause
subject. If (9a) is continued with (9b), the implicit subject receives an arbitrary
268
interpretation, no anaphora being necessary, but this only becomes clear later in the
sentence, long after the non-finite verb.
(9) a. Mickey talked to Minnie before eating pizza at the park on Sunday. . .
b. . . . was forbidden.
And of course, as was seen in Chapter 2, a non-finite verb in a temporal adjunct
could also be part of an NOC structure, which would also not require a syntactic
dependency.
Although all of the examples in our experiment did require control by the
main clause subject, it is still possible that participants in principle would wait
to interpret the implicit subject until it was clearly necessary. However, this was
not the case; participants quickly looked to the character corresponding to the main
clause subject upon hearing the non-finite verb, evidently establishing the anaphoric
control dependency at the earliest possible indication that it might be necessary.
6.5.1 Prediction in the resolution of anaphora
The role of prediction in the resolution of anaphora has been well documented.
For example, Kehler et al. (2008) provide evidence that pronoun interpretation is
incrementally influenced by predictions listeners make about what coherence rela-
tions are likely to be at play as well as what discourse entities are most likely to
be next mentioned (see also Kehler & Rohde, 2013). In the examples in (10) from
Caramazza, Grober, Garvey, and Yates (1977), for example, listeners predict that an
269
explanation for the first clause will be given, and their interpretation of the pronoun
is biased towards whatever interpretation will satisfy that prediction. In (10a), this
explanation is likely to be a description of something Mary did, and in (10b), it
is likely to be about Jane, and listeners may be quick to assign those referents to
the pronoun, despite the fact that the rest of the sentence may favor an alternative
interpretation, as does (11). Kehler et al. also demonstrate that the same biases
are at play even without the presence of words like because, suggesting even more
strongly that prediction of coherence relations is driving pronoun resolution.
(10) a. Jane hit Mary because she had stolen a tennis racket.
b. Jane angered Mary because she had stolen a tennis racket.
(11) Jane hit Mary because she reacts violently to criticism.
(Kehler et al., 2008)
The present experiments give evidence that not only conceptual sources of
information, but also predictions about the upcoming structure of a sentence may
affect the timing of anaphora resolution. When listeners may be likely to predict
an upcoming control structure, resolution of its implicit subject is faster. This is in
line with the arguments of (Kehler, 2008) that reference resolution is not a purely
reactive process, with participants always waiting for cues before retrieving potential
antecedents. Instead, listeners may actively predict upcoming structure and likely
referents before any cue for reference is received.
One question that still remains from these experiments is the extent to which
270
these kind of structural predictions influence reference resolution in real-world lan-
guage use. Individual item contexts in Experiments 1 and 2 appeared to favor pre-
diction of control structures. But was that due to properties unique to these items
or to the simple image and discourse context? Or are control structures predicted
in similar sentences more generally? Did adding additional fillers in Experiment 3
lead to more naturalistic processing of PRO, or did it reduce the predictions that
listers make in sentence processing outside the lab? Although these experiments do
not answer these questions, it is clear that listeners can use structural predictions to
aid anaphora resolution, and that when the context changes those predictions, the
use of structural information in the resolution of PRO takes no more time than the
use of morphological information in pronoun resolution once the bottom up input
is heard.
6.5.2 Additional implications
The results of these experiments have at least two other relevant implications.
First, they call into question a previous claim that the interpretation of PRO in
adjuncts involves a “most-recent filler strategy.” As mentioned in Chapter 5, based
on the results of an eyetracking-while-reading study, Betancort et al. (2006) (fol-
lowing Frazier et al., 1983; Nicol & Swinney, 1989) suggest that upon encountering
PRO, comprehenders first consider the most (linearly) local potential antecedent,
even if they must later revise that initial interpretation to establish an anaphoric
dependency between PRO and the main clause subject.
If such a strategy were active in the current experiment, we would expect ini-
271
tial looks during the critical region in the PRO condition to be to the competitor, as
it was the most recently mentioned potential antecedent. However, in Experiment 1,
looks to the target began to rise immediately after the onset of the critical word,
while looks to the competitor quickly began to decrease, suggesting that compre-
henders were at no point considering the more recently mentioned character over
the main clause subject in the PRO conditions. Even in Experiment 3, when inter-
pretation of PRO was slower, looks to the competitor were never higher than looks
to the target in the critical region of the PRO condition. It is therefore possible that
both characters were under equal consideration, which on its own would counter the
garden-path theory. But it is also possible that interpretation of the anaphora had
simply not begun at that point, and that once resolution of the anaphora began,
the subject was quickly recognized as the only potential antecedent.
Finally, previous research on the processing of PRO has left open the possibility
that it is not actually resolved. These experiments give evidence that at least in
some contexts, PRO is resolved quickly during incremental sentence processing.
6.6 Conclusion
This chapter used visual-world eyetracking to investigate the processing of the
implicit subject in adjunct control. It has shown that the implicit subject can be
interpreted at least as quickly as overt pronouns, especially when participants may
be predicting an upcoming control structure. Studying different forms of reference
can shed light on how different sources of information are implemented during sen-
272
tence processing, and this study contributes by providing evidence that structural
information can be immediately utilized in reference resolution.
273
Chapter 7: Conclusion
This dissertation has explored the syntax and processing of adjunct control.
Part I demonstrated in detail that temporal and rationale adjuncts allow both OC
and NOC, as do some other adjuncts, contrary to prior assumptions about the
complementarity of OC and NOC. But the same is not true for all adjuncts. Some
appear to be strictly OC and others strictly NOC, contrary to implicit predictions
of the TTC as instantiated in Landau (2017). I have argued that the MTC is able
to better explain the distribution of OC and NOC in some adjuncts, as long as it
is assumed that not all adjuncts are phases. However, the MTC also struggles in
capturing the full range of data. Thus, although this dissertation does not present a
full theory of the syntax of adjunct control, it does highlight some of the difficulties
associated with it for two modern theories of control.
Part II summarized some of the previous literature on the processing of adjunct
control and presented a set of experiments investigating the timing of the resolution
of PRO in subject-controlled temporal adjuncts. It was shown that PRO can be
interpreted in temporal adjuncts just as quickly as an overt pronoun can be, once the
relevant bottom-up input is received to cue the referential dependency. This timing
274
information provides important foundational knowledge for future research into the
processing biases relevant to the interpretation of adjunct control more generally.
This concluding chapter discusses implications of this research and some re-
maining questions, setting an agenda for future research.
7.1 Implications of this research
7.1.1 Adjuncts
The study of control has most frequently focused on cases where the controlled
clause is the complement of a main clause verb. Theories of the control relation and
of controller choice have been built largely on the properties of complement control,
and adjuncts have often been ignored to a significant extent. In syntactic theories
of control such as the predominant version of the MTC, it is assumed that adjunct
control will display the same properties as complement control. Even some semantic
theories of control make the same assumption. For example, Farkas (1988) argued
that controller choice is determined by a Responsibility relation (RESP)-relation in
complement control as well as in adjuncts, although her discussion of the latter is
limited to RatCs. Other semantic theories of control such as that of Sag and Pollard
(1991) do admit that the mechanism of controller choice in adjuncts may differ from
the one involved in complement control, since they are not selected by any element
of the main clause, but the discussion of adjuncts is stated as an aside, with no
detailed analysis.
This dissertation has provided evidence that adjuncts do not all behave in the
275
same way as complements, contrary to what Hornstein (1999) and Boeckx et al.
(2010) assume. While complements require OC, some adjuncts, including temporal
adjuncts and RatCs, allow both OC and NOC. It is also not the case, though,
that all adjuncts uniformly behave differently from complements, as is implicit in
Landau (2017). Several adjuncts require strict OC. Although some of these seem to
behave like quasi-arguments, it is likely that some of them (SPCs and telic clauses,
at least), are true adjuncts.
It is unclear as of yet how the processing of adjunct control relates to the
processing of complement control. There are some reasons to expect that adjunct
control would be processed more slowly, which is in fact what Betancort et al. (2006)
found for one type of adjunct. In complement control, the embedding verb selects
a control structure, and therefore provides a highly predictive cue to the upcoming
PRO. Adjuncts, on the other hand, are not selected by any element of the matrix
clause. But there may still be cues to an upcoming control structure, such as the
temporal prepositions before, after, and while that were used in the experiments
presented here. Although there was no comparison with complement control in
these experiments, it seems unlikely that PRO in complement control could have
been interpreted any quicker. But although the adjuncts in the experiments reported
here were interpreted quite quickly, the same may not be true for other adjuncts,
such as those that have no prepositional cue, as with the absolutive clause in (1),
or in settings where NOC is more likely, as in (2).
276
(1) Mickey1 flirted with Minnie outside the schoolhouse, [PRO1 having put on
his finest suit].
(2) The hamburgers1 became much more popular at the school carnival after. . .
a. PRO1 cooking just a little bit longer.
b. PRO cooking them1 just a little bit longer.
7.1.2 The OC/NOC distinction
This dissertation also has implications for what we mean by the terms “obliga-
tory” and “non-obligatory” control. Williams (1980) applied the term “OC” to any
case of control having the properties in (3). Sentences with “NOC” were argued to
have none of these properties.
(3) OC1. Lexical NP cannot appear in the position of PRO.
OC2. The antecedent precedes the controlled PRO.
OC3. The antecedent c-commands the controlled PRO.
OC4. The antecedent is thematically or grammatically uniquely determined.
OC5. There must be an antecedent.
In short, “OC” was used to define any case of control where a control dependency
was required between an obligatory PRO and a unique antecedent.
The definition of “OC” and “NOC” in the MTC is different. According to
Hornstein (1999), OC simply entails that the relation between PRO and its an-
tecedent is syntactic–specifically, that it is a movement relation. NOC entails that
277
PRO’s interpretation is not determined syntactically. A similar definition just for
adjuncts is given in Landau (2017), but with OC entailing a predicative relation
and NOC a logophoric one. Under these definitions, a sentence that can exhibit OC
would not necessarily require that its antecedent be uniquely determined or that it
c-command PRO. Nor do these definitions strictly require PRO in such sentences
to have an antecedent, if NOC is also available for the same sentence (although
according to Boeckx and Hornstein (2007), “parsers” will rarely allow both OC and
NOC).
These more modern versions of the notions of OC and NOC are what I have
used: OC involves some sort of grammatical dependency (movement, binding, or
coargumenthood) between PRO (whatever that label represents) and its antecedent,
while in NOC, PRO’s interpretation is determined without a grammatical depen-
dency. This definition of OC does not require that there be an antecedent in every
case. In other words, OC does not preclude NOC, contrary to what (3) suggests;
some adjuncts allow but do not require OC antecedents, permitting NOC as well.
However, there is some merit in maintaining a Williams-style concept of OC
and NOC. In complements and in some adjuncts, the antecedent is both required
and uniquely determined. In other words, OC is the only option. I have labeled this
“strict OC.” Some adjuncts allow both syntactic OC and non-syntactic NOC. And
others still require strict NOC. Landau (2015) also noted that traditional notions
of OC and NOC did not adequately capture distinctions in the types of control
that are possible. In the terms of the TTC, complement control is strictly OC,
which can be instantiated as either a predicative or a grammatically-determined
278
logophoric relation. Some adjuncts require strict predicative OC and others strict
NOC, while others allow both predicative OC and logophoric NOC. Instead of a
binary OC/NOC distinction, we need at least a three-way categorization, including
strict OC, strict NOC, and optional OC/NOC.
Within the third category, even though both types of control are possible, OC
is strongly preferred, all else equal. I have argued that this is due to processing pref-
erences. For example, parsers may prefer OC because of its simpler structure, if the
TTC (Landau, 2015) is correct. Or, under the the MTC, parsers may prefer to posit
a trace when they encounter a gap, resulting in OC, rather than a null pronominal
for NOC (Boeckx & Hornstein, 2007; Boeckx et al., 2010). In addition to or in-
stead of these, OC may be more preferred to NOC to a more general preference for
bound variable readings (Reinhart, 1983, 2006), or a preference for local (Cunnings
& Sturt, 2014) or recent (Cunnings et al., 2014) antecedents. Factors favoring NOC
may include the availability of a non-local perspective-holding potential antecedent,
whether OC would result in a strange interpretation (Landau, 2017), attachment
site of the adjunct (Landau, 2013), or other factors.
In the experiments in Chapter 6, despite the fact that temporal adjuncts some-
times allow NOC, participants only appeared to consider OC interpretations, quickly
resolving PRO to the referent of the main clause subject. Of course, within the
experiments, there were no cases of intended NOC. Still, the results give some in-
dication that the preference for OC may be strong enough that NOC is not even
considered an option if the OC reading is likely.
279
7.1.3 Constraints on reference
This research has additional consequences in the study of constraints on ref-
erence more generally. In remote control of RatCs, for example, the referent of
PRO can be restricted quite strongly, even when it is not done through grammat-
ical means. In (4a), control by the object is ruled out without further supporting
context. When the object is raised to the subject position in (4b), however, it can
remotely control PRO. But this cannot be because a syntactic dependency between
Sam and PRO is possible in (4b) but not (4a), since there can be no such dependency
in either case.
(4) a. *The police arrested Sam1. The goal was PRO1 to get himself off the
streets.
b. Sam1 was arrested by the police. The goal was PRO1 to get himself off
the streets.
This provides evidence that there can be strong constraints on reference that are
non-grammatical but that are nevertheless highly sensitive to grammatical structure.
In prior literature, pragmatic constraints on reference in general have been described
as sensitive to syntactic structure only inasmuch as it effects information structure.
Examples from Kehler (2004), repeated in (5), for example, showed that the referent
of a pronoun can be constrained by topichood. Because the subject of a passive is
a strong topic, the pronouns in (5a) and (5b) are much more likely to refer to the
demonstrators than to the city council, even if resolution of the pronoun to the
280
demonstrators results in a less coherent discourse. When the sentence is in active
voice in (6), on the other hand, conceptual constraints play the lead role, and we
simply resolve the pronoun in whichever way yields the more likely story.
(5) The demonstrators were denied a permit by the city council because. . .
a. . . . they feared violence.
b. . . . they advocated violence.
(6) The city council denied the demonstrators a permit because. . .
a. . . . they feared violence.
b. . . . they advocated violence.
Although PRO in a remote RatC is something like a free pronoun, its interpre-
tation is not constrained solely on the basis of coherence and information structure.
This can be seen in examples such as (7) and (8).
(7) *An alarm woke the boy1 up early. The goal was PRO1 to wash himself before
school.
(8) The demonstrators were denied a permit by the city council. The goal was
PRO to reduce violence.
In (7), if coherence were the only issue, then it would be surprising that PRO could
not refer to the object of the first sentence, as this would result in a sensible discourse;
boys are much more likely than alarms to wash themselves before school. Similarly,
281
if information-structural constraints were as important in RatCs as they are in
examples such as (5), without additional constraints, then the strongly preferred
interpretation of (8) should be one where PRO refers to the demonstrators. This
does not seem to be the case. PRO in (8) seems much more likely to refer to the city
council, despite the information-structurally strong position of the demonstrators.
Even more strikingly, the same interpretation is the most likely in (9). Here
information structure still favors control by the demonstrators, and this could even
arise through OC, which should further increase its likelihood. Yet we still find that
PRO is much more likely to refer to the city council.
(9) The demonstrators were denied a permit by the city council in order PRO
to reduce violence.
PRO in a RatC, then, is subject to constraints on reference that go beyond
coherence and are sensitive to structure for reasons other than information structure,
and beyond the preference for OC when possible. Sentence structure can apparently
affect who can easily be seen as responsible for a fact. This in turn affects the
reference of PRO. A. Williams and Green (2017) state that objects are generally
unable to control PRO in a RatC, even when syntactic OC is completely unavailable,
as in remote control, because they are represented as non-responsible. Thus, control
may be determined through a conceptual notion that is independent of information
structure but is nevertheless sensitive to syntactic structure.
In sum, constraints on reference can be conceptual as in (6), or purely struc-
282
tural, as with reflexives or cases of OC. Or they can also be sensitive to structure,
but still not grammatical. This can arise through the effect syntactic structure has
on information structure, as in (5), or it can be due to how syntactic structure affects
the way individuals and events are represented, as is relevant for RatCs.
7.1.4 Prediction in reference resolution
Finally, the experiments in Chapter 6 provide insights into the types of pre-
dictions comprehenders may utilize during reference resolution. Previous research
has shown that pronoun resolution is influenced by predictions about coherence re-
lationships, about which discourse entities are most likely to be mentioned, and
which entities are most likely topics (Kehler & Rohde, 2013). In the experiments
reported here, PRO’s resolution may also have been influenced by participants’ pre-
dictions about upcoming structure. In Experiment 1, participants may have been
predicting upcoming control structures, which led to faster resolution of PRO than
in Experiment 3, when control structures were less likely. Crucially, the influence of
such predictions must have been independent from predictions about the next likely
referent. In Experiment 1, if participants were simply predicting that the subject
was most likely to be next mentioned, then looks to the target character and away
from the competitor should have been just as fast for pronouns coreferring with
the subject as for PRO. This was not the case. Instead, participants may have
predicted an upcoming control structure, which entailed in this experiment corefer-
ence with the main clause subject, but when the onset of a pronoun was heard, the
283
structural prediction was abandoned, along with its corresponding subject reference
prediction.
7.2 Remaining questions and future work
In this dissertation, I have not attempted to provide a conclusive theory of
the syntax of adjunct control or of its processing. Although I believe it provides
important progress, there are many open questions. Here I discuss just a few.
7.2.1 Variability in adjunct control
As demonstrated in Chapter 4, neither the TTC nor the MTC can adequately
handle the whole range of adjuncts that allow control. The TTC predicts that
all adjuncts will allow both OC and NOC, which is not the case. The MTC as
formulated here allows for some adjuncts to be strictly OC, as long as they attach
low enough and have a small enough structure, but this also fails to account for telic
clauses. Further research is therefore needed to develop a comprehensive theory of
adjunct control, situated within a broader theory of control.
Additionally, both the TTC and the MTC accounts predict that similar pat-
terns will be seen cross-linguistically. Many adjuncts should allow both OC and
NOC. What may differ, however, is the weighting of factors favoring OC interpre-
tations over NOC in processing. As just one example, Español-Echevarŕıa (2000)
argues that Modern Greek does not allow NOC, and that therefore implicit control
284
of RatCs is impossible (compare (10a) and (10b), from Español-Echevarŕıa (2000,
pp. 101–102)).
(10) a. O Yannis vithise to plio ya na voithisi tin vassilisa.
the Yannis sank the boat for NA help-act-3s the queen.
‘Yannis sank the boat in order to help the queen.’
b. *To plio vithistike ya na voithisi tin vassilisa.
the boat sink-pass-3s for NA help-act-3s the queen
‘The boat was sunk in order to help the queen.’
A complete restriction against NOC would require there to be no logophoric control
under the TTC, which is unlikely, at least for complement control. Under the MTC,
there would have to be no last-resort pronominalization option in the derivation of
control structures. Or, it may be possible that NOC is actually still allowed by the
grammar, but that OC is simply more heavily favored in Greek than in English.
This would be consistent with Kotzoglou (2016), where it is argued that NOC is
possible in gerundive adjuncts, but only if PRO is understood to refer to an implicit
argument of the matrix clause.
It is certain that other languages will also show distinct characteristics with
respect to the possible forms adjunct control may take and its possible interpreta-
tions.
7.2.2 Biases contributing to OC preference
Along these lines, a major question that remains is what the specific biases
leading to the preference for OC are, both in English and cross-linguistically. I
285
have argued that the preference is due to processing biases, but clearly structural
and other factors may contribute to the probability of choosing OC or NOC during
processing. I have speculated here about some possible factors, as well as what
kinds of sentences may favor NOC, but a more detailed exploration of these biases
is needed. One way to start would be with a systematic acceptability judgment task
where sentence-structural features can be controlled and compared. Processing ex-
periments targeting the preference for local antecedents, the preference for syntactic
versus discourse relations, and other potential factors would also be useful.
Perhaps, contrary to my assumptions, the preference for OC is not strictly
due to processing, but is in fact encoded in the grammar of adjunct control. If that
is the case, then additional experiments investigating potential processing biases
would still be enlightening. If the experimental design were to control for all possi-
ble processing factors (e.g. frequency of OC and NOC in a given structure, topicality
of potential OC and NOC antecedents, etc.), and yet there is still a preference or
advantage for OC, this may serve as compelling evidence that it is not simply a pro-
cessing effect. Instead, this would suggest a grammar-internal preference, either for
OC over NOC specifically, or for binding over discourse coreference more generally.
7.2.3 Preference for upward movement
A related question relevant to the MTC account of adjunct control put forth
in Chapter 4 is the source of the preference for upward movement over sideward
movement. I speculated that this was due to a derivational penalty being assigned to
sideward movement, which in turn reduces the relative preference for OC over NOC.
286
However, why sideward movement should incur such a penalty remains unclear. It
could perhaps, simply have to do with the fact that sideward movement requires
forming a syntactic dependency between two separate workspaces, so to speak, in the
derivation. Or it could have something to do with constraints on proper movement.
This question deserves greater attention and research.
7.2.4 Interaction of different sources of information
An additional question that merits further study is the extent to which different
sources of information interact in prediction and the resolution of anaphora. For
example, Kehler et al. (2008) found that expected coherence relations can strongly
affect the interpretation of pronouns (see also Hobbs, 1979; Kehler & Rohde, 2013).
In (11), listeners/readers expect that the second clause will provide an explanation
for the first. Therefore, the pronoun she is interpreted in the way that will result in
the most sensible explanation; in (11a), she is interpreted as ‘Mary’, but in (11b),
as ‘Jane’.
(11) a. Jane hit Mary because she had stolen a tennis racket.
b. Jane angered Mary because she had stolen a tennis racket.
But what happens when the coherence relation is incongruous with the struc-
tural constraints of the sentence? How do structural and conceptual sources of
information in anaphora resolution interact? In (12), for example, the referent of
the pronoun is most likely to be determined based on the coherence relation between
287
the two clauses, with he in (12a) referring to Donald, and he in (12b) to Mickey.
But in (13), even though the same coherence relation is likely to be predicted, PRO
in both (13a) and (13b) is much more likely to refer only to Mickey, given the strong
preference for OC. How quickly can syntactic information from the sentence override
the conceptual expectations listeners have?
(12) a. Mickey got really angry at Donald outside the schoolhouse after he
stole the new red ball he got for his birthday.
b. Mickey really angered Donald outside the schoolhouse after he stole the
new red ball he got for his birthday.
(13) a. Mickey got really angry at Donald outside the schoolhouse after PRO
stealing the new red ball he got for his birthday.
b. Mickey really angered Donald outside the schoolhouse after PRO steal-
ing the new red ball he got for his birthday.
Exploring the real-time processing of sentences such as these can help inform
us about prediction in language processing and how the mind integrates different
sources of information. It may also shed additional light on the interpretive biases
that lead to the preference for OC and what factors contribute instead to NOC.
7.2.5 Processing and representation
Further work could also be done to determine whether syntactic and discourse
dependencies have different processing profiles. Implicit in the work of Mauner and
288
Koenig (1999, 2000) is the assumption that differences between syntactic and non-
syntactic dependencies will result in differences in processing. But the experiments
in McCourt et al. (2015) cast doubt on that conclusion. At least for self-paced
reading times, syntactic OC in local RatCs adjoined to active matrix clauses and
non-syntactic NOC in remote RatCs do not exhibit different processing profiles.
Still, it is possible that in other measures, differences between syntactic and
non-syntactic anaphoric dependencies will be reflected in real-time processing. Such
measures may lead to a greater understanding of the grammatical representation of
these dependencies as well as how these dependencies are encoded and accessed in
memory.
7.2.6 The acquisition of adjunct control
One final area that merits further inquiry is how and when children acquire
adjunct control. Gerard (2016) demonstrated that by the age of four, children
assume that PRO in temporal adjuncts is controlled by the matrix subject, while at
the age of three, children accept both object and subject control equally. However,
Gerard only tested sentences that adults unanimously agreed were subject control.
The acquisition of non-subject controlled temporal adjuncts and of other types of
adjunct control is a completely open question.
289
7.3 Conclusion
This dissertation represents the beginning of a research program that aims to
increase our understanding of implicit and explicit forms of reference both in En-
glish and cross-linguistically. It may ultimately illuminate how the mind integrates
syntactic and non-syntactic sources of information in anaphora resolution. In so
doing, it will contribute to our knowledge about the use of prediction in language
processing and the process by which referential expressions access their antecedents
in memory.
290
References
Ackema, P. & Schoorlemmer, M. (1995). Middles and nonmovement. Linguistic In-
quiry, 173–197. JSTOR: 4178895
Anderson, C. (2004). The structure and real-time comprehension of quantifier scope
ambiguity (Doctoral dissertation, Northwestern University).
Arnold, J. E. (1998). Reference form and discourse patterns (Doctoral dissertation,
Stanford University).
Arnold, J. E., Eisenband, J. G., Brown-Schmidt, S., & Trueswell, J. C. (2000).
The rapid use of gender information: Evidence of the time course of pronoun
resolution from eyetracking. Cognition, 76 (1), B13–B26. doi:10.1016/S0010-
0277(00)00073-1
Arnold, J. E. & Lao, S.-Y. C. (2015). Effects of psychological attention on pro-
noun comprehension. Language, Cognition and Neuroscience, 30 (7), 832–852.
doi:10.1080/23273798.2015.1017511
Artstein, R. (2005). Quantificational arguments in temporal adjunct clauses. Lin-
guistics and Philosophy, 28 (5), 541–597. doi:10.1007/s10988-005-6921-6
Badecker, W. & Straub, K. (2002). The processing role of structural constraints on
interpretation of pronouns and anaphors. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 28 (4), 748–769. doi:10.1037/0278-7393.28.
4.748
Baker, M., Johnson, K., & Roberts, I. (1989). Passive arguments raised. Linguistic
Inquiry, 20 (2), 219–251. JSTOR: 4178625
Barr, D. J. (2008). Analyzing ‘visual world’ eyetracking data using multilevel logistic
regression. Journal of Memory & Language, 59 (4), 457–474. doi:10.1016/j.jml.
2007.09.002
Bates, D., Mächler, M., Bolker, B., & Walker, S. (2015). Fitting linear mixed-effects
models using lme4. Journal of Statistical Software, 67 (1), 1–48. doi:10.18637/
jss.v067.i01
291
Belletti, A. & Rizzi, L. (1988). Psych-verbs and θ-theory. Natural Language & Lin-
guistic Theory, 6 (3), 291–352. doi:10.1007/bf00133902
Betancort, M., Carreiras, M., & Acuña-Fariña, C. (2006). Processing controlled
PROs in Spanish. Cognition, 100 (2), 217–282. doi:10.1016/j.cognition.2005.
04.001
Bhatt, R. (1999). Covert modality in non-finite contexts (Doctoral dissertation, Uni-
versity of Pennsylvania).
Bhatt, R. & Izvorski, R. (1998). Genericity, implicit arguments and control. In Pro-
ceedings of the 7th student conference in linguistics. MITWPL.
Blanchard, H. (1987). Pronoun processing during fixations: Effects on the time
course of information utilization. Bulletin of the Psychonomic Society, 25,
171–174.
Boeckx, C. & Hornstein, N. (2004). Movement under control. Linguistic Inquiry,
35 (3), 431–452. doi:10.1162/0024389041402625
Boeckx, C. & Hornstein, N. (2007). On (non-)obligatory control. In W. D. Davies
& S. Dubinsky (Eds.), New horizons in the analysis of control and raising
(pp. 251–262). Springer. doi:10.1007/978-1-4020-6176-9 11
Boeckx, C., Hornstein, N., & Nunes, J. (2010). Control as movement. Cambridge
University Press. doi:10.1017/cbo9780511761997
Boersma, P. (2001). Praat, a system for doing phonetics by computer. Glot Inter-
national, 5 (9/10), 341–345.
Boersma, P. & Weenink, D. (2017). Praat: Doing phonetics by computer (Ver-
sion 6.0.29). Retrieved from http://www.praat.org
Boland, J. E., Tanenhaus, M. K., & Garnsey, S. M. (1990). Evidence for the immedi-
ate use of verb control information in sentence processing. Journal of Memory
& Language, 29 (4), 413–432. doi:10.1016/0749-596X(90)90064-7
Borer, H. (1989). Anaphoric AGR. In O. A. Jaeggli & K. J. Safir (Eds.), The null
subject parameter (pp. 69–109). Dordrecht: Springer Netherlands. doi:10.1007/
978-94-009-2540-3 3
Bošković, Ž. (1996). Selection and the categorial status of infinitival complements.
Natural Language & Linguistic Theory, 14 (2), 269–304. doi:10.1007/bf00133685
292
Caramazza, A., Grober, E., Garvey, C., & Yates, J. (1977). Comprehension of
anaphoric pronouns. Journal of Verbal Learning and Verbal Behavior, 16 (5),
601–609. doi:10.1016/s0022-5371(77)80022-4
Charnavel, I. & Sportiche, D. (2016). Anaphor binding: What French inanimate
anaphors show. Linguistic Inquiry, 47 (1), 35–87. doi:10.1162/LING a 00204
Chierchia, G. (1989). Structured meanings, thematic roles and control. In G. Chier-
chia, B. H. Partee, & R. Turner (Eds.), Properties, types and meanings (Vol. 2).
Dordrecht: Kluwer. doi:10.1007/978-94-009-2723-0 4
Chomsky, N. (1965). Aspects of the theory of syntax. Cambridge.
Chomsky, N. (1972). Some empirical issues in the theory of transformational gram-
mar. In P. S. Peters (Ed.), Goals of linguistic theory (pp. 63–130). Englewood
Cliffs: Prentice Hall Inc.
Chomsky, N. (1980). On binding. Linguistic Inquiry, 11 (1), 1–46. JSTOR: 4178149
Chomsky, N. (1981). Lectures on government and binding. Foris: Dordrecht. doi:10.
1515/9783110884166
Chomsky, N. (1986). Barriers. Linguistic Inquiry Monographs. Cambridge, Mass:
MIT Press.
Chomsky, N. (1991). Some notes on economy of derivation and representation. In R.
Freidin (Ed.), Principles and parameters in comparative grammar (pp. 417–
454). Cambridge, MA: MIT Press. [Also published in Chomsky, N. (1995).
The Minimalist Program (pp. 129–166). Cambridge, MA: MIT Press.] doi:10.
7551/mitpress/9780262527347.003.0002
Chomsky, N. (1995). The Minimalist Program. Cambridge, MA: MIT Press. doi:10.
7551/mitpress/9780262527347.001.0001
Chomsky, N. (2001). Derivation by phase. In M. Kenstowicz (Ed.), Ken Hale: A life
in language (pp. 1–52). Cambridge, MA: MIT Press.
Cinque, G. (1999). Adverbs and functional heads: A cross-linguistic perspective. New
York: Oxford University Press.
Clark, R. L. (1990). Thematic theory in syntax and interpretation. London: Rout-
ledge. doi:10.4324/9781315519890
Collins, C. (2005). A smuggling approach to the passive in English. Syntax, 8 (2),
81–120. doi:10.1111/j.1467-9612.2005.00076.x
293
Comrie, B. (1984). Subject and object control: Syntax, semantics, pragmatics. In
Proceedings of the tenth annual meeting of the Berkeley Linguistics Society
(pp. 450–464). doi:10.3765/bls.v10i0.1957
Comrie, B. (1989). Language universals and linguistic typology: Syntax and mor-
phology (2nd ed.). Chicago: University of Chicago press.
Crawley, R. A., Stevenson, R. J., & Kleinman, D. (1990). The use of heuristic strate-
gies in the interpretation of pronouns. Journal of Psycholinguistic Research,
19 (4), 245–264. doi:10.1007/BF01077259
Culicover, P. W. & Jackendoff, R. (2001). Control is not movement. Linguistic In-
quiry, 32 (3), 493–512. doi:10.1162/002438901750372531
Cunnings, I., Patterson, C., & Felser, C. (2014). Variable binding and coreference in
sentence comprehension: Evidence from eye movements. Journal of Memory
& Language, 71 (1), 39–56. doi:10.1016/j.jml.2013.10.001
Cunnings, I. & Sturt, P. (2014). Coargumenthood and the processing of reflexives.
Journal of Memory & Language, 75, 117–139. doi:10.1016/j.jml.2014.05.006
Curtis, C. E., Cole, M. W., Rao, V. Y., & D’Esposito, M. (2004). Canceling planned
action: An fMRI study of countermanding saccades. Cerebral Cortex, 15 (9),
1281–1289. doi:10.1093/cercor/bhi011
Davidson, D. (1967). Truth and meaning. Synthese, 17 (3), 304. doi:10.1007/bf00485035
De Vos, M. (2012). Towards a definition of SUBJECT in binding domains and
subject oriented anaphors. Stellenbosch Papers in Linguistics, 37. doi:10.5774/
37-0-17
DeArmond, R. C. & Hedberg, N. (1998). On complements and adjuncts. In Pro-
ceedings of the Canadian Linguistics Society meeting, cahiers linguistiques
d’Ottawa (pp. 95–106). Ottawa: University of Ottawa.
Demestre, J., Meltzer, S., Garcia-Albea, J. E., & Vigil, A. (1999). Identifying the
null subject: Evidence from event-related brain potentials. Journal of Psy-
cholinguistic Research, 28 (3), 293–312. doi:10.1023/A:1023258215604
Dink, J. & Ferguson, B. (2015). eyetrackingR: An R library for eye-tracking data
analysis. Retrieved from http://eyetrackingr.com
Dowty, D. (1989). On the semantic content of the notion of ‘thematic role’. In G.
Chierchia, B. H. Partee, & R. Turner (Eds.), Properties, types, and meanings
(Vol. 1, pp. 69–130). Dordrecht: Kluwer. doi:10.1007/978-94-009-2723-0 3
294
Drummond, A. (2011). Binding phenomena within a reductionist theory of gram-
matical dependencies (Doctoral dissertation, University of Maryland). HDL:
1903/12191
Ehrlich, K. & Rayner, K. (1983). Pronoun assignment and semantic integration
during reading: Eye movements and immediacy of processing. Journal of Ver-
bal Learning and Verbal Behavior, 22 (1), 75–87. doi:10.1016/S0022-5371(83)
80007-3
Elbourne, P. (2013). Definite descriptions. OUP Oxford. doi:10.1093/acprof:oso/
9780199660193.001.0001
Epstein, S. D. (1990). Differentiation and reduction in syntactic theory: A case
study. Natural Language & Linguistic Theory, 8 (3), 313–323. doi:10 .1007/
BF00135616
Ernst, T. (2009). Speaker-oriented adverbs. Natural Language & Linguistic Theory,
27 (3), 497–544. doi:10.1007/s11049-009-9069-1
Español-Echevarŕıa, M. (2000). The interaction of obligatory and nonobligatory con-
trol in rationale clauses. In R. Billerey & B. D. Lillehaugen (Eds.), WCCFL
19 proceedings (pp. 97–110).
Evans, G. (1980). Pronouns. Linguistic Inquiry, 11 (2), 337–362. JSTOR: 4178164
Fabricius-Hansen, C. & Sæbø, K. J. (2011). Behabitive reports. In E. Brendel, J.
Meibauer, & M. Steinbach (Eds.), Understanding quotation (pp. 85–106). De
Gruyter Mouton. doi:10.1515/9783110240085.85
Faraci, R. A. (1974). Aspects of the grammar of infinitives and for-phrases (Doctoral
dissertation, MIT). HDL: 1721.1/13017
Farkas, D. F. (1988). On obligatory control. Linguistics and Philosophy, 11 (1), 27–
58. doi:10.1007/bf00635756
Felser, C. (2004). Wh-copying, phases, and successive cyclicity. Lingua, 114 (5), 543–
574. doi:10.1016/s0024-3841(03)00054-8
Ferreira, F., Bailey, K. G., & Ferraro, V. (2002). Good-enough representations in
language comprehension. Current Directions in Psychological Science, 11 (1),
11–15. doi:10.1111/1467-8721.00158
Fillmore, C. J. (1971). Types of lexical information. In D. D. Steinberg & L. A.
Jakobovits (Eds.), Semantics: An interdisciplinary reader in philosophy, lin-
guistics and psychology (pp. 370–392). Cambridge: Cambridge University Press.
295
Fox, D. (2000). Economy and semantic interpretation. Cambridge, MA: MIT Press.
Frazier, L., Clifton, C., & Randall, J. (1983). Filling gaps: Decision principles and
structure in sentence comprehension. Cognition, 13, 187–222. doi:10 . 1016/
0010-0277(83)90022-7
Gerard, J. (2016). The acquisition of adjunct control: Grammar and processing (Doc-
toral dissertation, University of Maryland). doi:10.13016/M2621K
Geurts, B. (2011). Accessibility and anaphora. In K. von Heusinger, C. Maienborn, &
P. Portner (Eds.), Semantics: An international handbook of natural language
meaning (Vol. 2, pp. 1988–2011). Berlin: Mouton de Gruyter. doi:10.1515/
9783110255072.1988
Grano, T. & Lasnik, H. (in press). How to neutralize a finite clause boundary: Phase
theory and the grammar of bound pronouns. Linguistic Inquiry.
Grimshaw, J. (1994). Minimal projection and clause structure. In B. Lust, M. Suñer,
& J. Whitman (Eds.), Syntactic theory and first language acquisition: Cross-
linguistic perspectives (Vol. 1: Heads, projections and learnability, pp. 75–83).
Hillsdale, NJ: Lawrence Erlbaum Associates.
Grosz, B. J., Weinstein, S., & Joshi, A. K. (1995). Centering: A framework for
modeling the local coherence of discourse. Computational Linguistics, 21 (2),
203–225.
Haegeman, L. (2010). The internal syntax of adverbial clauses. Lingua, 120 (3), 628–
648. doi:10.1016/j.lingua.2008.07.007
Hallman, P. (2015). Syntactic neutralization in double object constructions. Lin-
guistic Inquiry, 46 (3), 389–424. doi:10.1162/ling a 00187
Heim, I. (1983). File change semantics and the familiarity theory of definiteness. In
R. Bäuerle, C. Schwarze, & A. von Stechow (Eds.), Meaning, use, and interpre-
tation of language. Berlin: Walter de Gruyter. doi:10.1515/9783110852820.164
Heim, I. (1998). Anaphora and semantic interpretation: A reinterpretation of Rein-
hart’s approach. In U. Sauerland & O. Percus (Eds.), MITWPL (Vol. 25,
pp. 205–246).
Heim, I. (2009). Forks in the road to Rule I. In A. Schardl, M. Walkow, & M.
Abdurrahman (Eds.), Proceedings of NELS 38 (Vol. 1, pp. 339–358). Amherst,
MA: GLSA. HDL: 1721.1/71811
Heim, I. & Kratzer, A. (1998). Semantics in generative grammar. Oxford: Blackwell.
296
Heycock, C. & Kroch, A. (1999). Pseudocleft connectedness: Implications for the LF
interface level. Linguistic Inquiry, 30 (3), 365–397. doi:10.1162/002438999554110
Higgins, F. R. (1973). The pseudo-cleft construction in English (Doctoral disserta-
tion, MIT). HDL: 1721.1/12988
Hirschbühler, P. (1988). The middle and pseudo-middle in French. In D. Birdson
& J.-P. Montreuil (Eds.), Advances in Romance linguistics (pp. 97–111). Dor-
drecht: Forsi.
Hobbs, J. R. (1979). Coherence and coreference. Cognitive Science, 3 (1), 67–90.
doi:10.1207/s15516709cog0301 4
Hornstein, N. (1999). Movement and control. Linguistic Inquiry, 30 (1), 69–96. doi:10.
1162/002438999553968
Hornstein, N. (2001). Move! A minimalist theory of construal. Blackwell.
Hornstein, N. (2003). On control. In R. Hendrick (Ed.), Minimalist syntax (pp. 6–
81). Oxford, UK: Blackwell. doi:10.1002/9780470758342.ch1
Hornstein, N. (2008). A theory of syntax: Minimal operations and universal gram-
mar. Cambridge University Press. doi:10.1017/cbo9780511575129
Hornstein, N. & Kiguchi, H. (2003). PRO Gate and movement. Pennsylvania Work-
ing Papers in Linguistics, 8 (1), 101–114.
Huang, C.-T. J. (1982). Logical relations in Chinese and the theory of grammar
(Doctoral dissertation, MIT). HDL: 1721.1/15215
Huang, C.-T. J. & Liu, C.-S. L. (2001). Logophoricity, attitudes, and ziji at the
interface. In Long-distance reflexives (Vol. 33, pp. 141–195). Syntax and Se-
mantics.
Huettner, A. K. (1989). Adjunct infinitives in English (Doctoral dissertation, Uni-
versity of Massachusetts, Amherst).
Jackendoff, R. S. (1972). Semantic interpretation in generative grammar. MIT Press.
Jaeggli, O. A. (1986). Passive. Linguistic Inquiry, 17 (4), 587–622. JSTOR: 4178510
Jones, C. (1991). Purpose clauses: Syntax, thematics, and semantics of English
purpose constructions. Studies in linguistics and philosophy. Dordecht, The
Netherlands: Kluwer Academic.
297
Jones, C. F. (1985). Syntax and thematics of infinitival adjuncts (Doctoral disserta-
tion, University of Massachusetts, Amherst).
Kamide, Y., Altmann, G. T., & Haywood, S. L. (2003). The time-course of predic-
tion in incremental sentence processing: Evidence from anticipatory eye move-
ments. Journal of Memory & Language, 49 (1), 133–156. doi:10.1016/s0749-
596x(03)00023-8
Karimi, H. & Ferreira, F. (2016). Good-enough linguistic representations and online
cognitive equilibrium in language processing. The Quarterly Journal of Exper-
imental Psychology, 69 (5), 1013–1040. doi:10.1080/17470218.2015.1053951
Karttunen, L. (1976). Discourse referents. In J. D. McCawley (Ed.), Syntax and se-
mantics: Notes from the linguistic underground (Vol. 7). New York: Academic
Press.
Katada, F. (1991). The LF representation of anaphors. Linguistic Inquiry, 22 (2),
287–313. JSTOR: 4178722
Kawasaki, N. (1993). Control and arbitrary interpretation in English (Doctoral dis-
sertation, University of Massachusetts, Amherst).
Kehler, A. (2004). Discourse topics, sentence topics, and coherence. Theoretical Lin-
guistics, 30, 227–240. doi:10.1515/thli.2004.30.2-3.227
Kehler, A. (2008). Rethinking the smash approach to pronoun interpretation. In J.
Gundel & N. Hedberg (Eds.), Reference: Interdisciplinary perspectives (pp. 95–
122). Oxford University Press. doi:10.1093/acprof:oso/9780195331639.003.
0005
Kehler, A., Kertz, L., Rohde, H., & Elman, J. L. (2008). Coherence and coreference
revisited. Journal of Semantics, 25 (1), 1–44. doi:10.1093/jos/ffm018
Kehler, A. & Rohde, H. (2013). A probabilistic reconciliation of coherence-driven and
centering-driven theories of pronoun interpretation. Theoretical Linguistics,
39 (1-2), 1–37. doi:10.1515/tl-2013-0001
Kotzoglou, G. (2016). Control in Greek gerunds: Implicit arguments and other fac-
tors. In Selected papers of the 21st international symposium on theoretical and
applied linguistics (pp. 166–183).
Kratzer, A. (1996). Severing the external argument from the verb. In J. Rooryck &
L. Zaring (Eds.), Phrase structure and the lexicon (pp. 109–137). Dordrecht:
Kluwer. doi:10.1007/978-94-015-8617-7 5
298
Kuznetsova, A., Brockhoff, P., & Christensen, R. (2014). LmerTest: Tests in linear
mixed effects models. R package version 2.0-20. Retrieved from http://CRAN.
R-project.org/package=lmerTest
Landau, I. (2000). Elements of control: Structure and meaning in infinitival construc-
tions. Studies in natural language and linguistic theory. Dordrecht, Boston:
Kluwer Academic. doi:10.1007/978-94-011-3943-4
Landau, I. (2001). Control and extraposition: The case of super-equi. Natural Lan-
guage & Linguistic Theory, 19 (1), 109–152. doi:10.1023/A:1006485514817
Landau, I. (2004). The scale of finiteness and the calculus of control. Natural Lan-
guage & Linguistic Theory, 22, 811–877. doi:10.1007/s11049-004-4265-5
Landau, I. (2013). Control in generative grammar: A research companion. Cam-
bridge University Press. doi:10.1017/cbo9781139061858
Landau, I. (2015). A two-tiered theory of control. Linguistic Inquiry Monographs.
Cambridge, MA: MIT Press. doi:10.7551/mitpress/9780262028851.001.0001
Landau, I. (2017). Adjunct control depends on voice. In C. Halpert, H. Kotek, &
C. van Urk (Eds.), A pesky set: Papers for David Pesetsky. Cambridge, MA:
MITWPL.
Landau, I. (2018, April 10). From predication to control and back again: Lessons
and puzzles. Paper presented at GLOW 41, Hungarian Academy of Science.
Larson, R. K. (1987). ‘‘Missing prepositions” and the analysis of English free relative
clauses. Linguistic Inquiry, 18 (2), 239–266. JSTOR: 4178537
Larson, R. K. (1988). On the double object construction. Linguistic Inquiry, 335–
391. JSTOR: 25164901
Lasnik, H. (1988). Subjects and the θ-criterion. Natural Language & Linguistic The-
ory, 6 (1), 1–17. doi:10.1007/BF01791588
Lasnik, H. (1991). On the necessity of binding conditions. In R. Freidin (Ed.), Prin-
ciples and parameters in comparative grammar (pp. 7–28). MIT Press.
Lasnik, H. (2001). When can you save a structure by destroying it? In M. Kim &
U. Strauss (Eds.), Proceedings of NELS 31 (Vol. 2, pp. 301–320).
Lasnik, H. & Saito, M. (1984). On the nature of proper government. Linguistic
Inquiry, 15 (2), 235–289. JSTOR: 4178382
299
Law II, F., Mahr, T., Schneeberg, A., & Edwards, J. (2016). Vocabulary size and
auditory word recognition in preschool children. Applied Psycholinguistics,
38 (01), 89–125. doi:10.1017/s0142716416000126
Lebeaux, D. (1984). Anaphoric binding and the definition of PRO. In C. Jones & P.
Sells (Eds.), Proceedings of NELS 14 (pp. 253–274). UMASS. Amherst, MA:
GLSA Publications.
Levin, B. & Rappaport Hovav, M. (1995). Unaccusativity: At the syntax-lexical
semantics interface. Linguistic Inquiry Monographs. Cambridge, Mass: MIT
Press.
Lewis, R. L. & Vasishth, S. (2005). An activation-based model of sentence processing
as skilled memory retrieval. Cognitive Science, 29 (3), 375–419. doi:10.1207/
s15516709cog0000 25
Lobeck, A. (1995). Ellipsis: Functional heads, licensing, and identification. Oxford
University Press.
Lyngfelt, B. (2000). OT semantics and control. Rutgers optimality archive, 411.
Lyngfelt, B. (2001). C-command vs. logophoric control: Interpretive OT applied to
Swedish adverbial infinitives. In Proceedings of the 18th Scandinavian Confer-
ence of Linguistics (Vol. 2, pp. 53–68).
Lyngfelt, B. (2002). Kontroll i svenskan. Den optimala tolkningen av infinitivers
tankesubjekt. [‘Control in Swedish: The optimal interpretation of implicit sub-
jects in infinitive clauses.’] (Doctoral dissertation, Acta Universitatis Gothobur-
gensis).
Lyngfelt, B. (2009a). Control phenomena. In Handbook of pragmatics (Vol. 13, pp. 1–
22). John Benjamins. doi:10.1075/hop.13.con15
Lyngfelt, B. (2009b). Towards a comprehensive construction grammar account of
control: A case study of Swedish infinitives. Constructions and Frames, 1 (2),
153–189. doi:10.1075/cf.1.2.01lyn
Manzini, M. R. (1983). On control and control theory. Linguistic Inquiry, 14 (3),
421–446. JSTOR: 4178338
Manzini, M. R. (1986). On control and binding theory. In S. Berman, J.-W. Choe,
& J. McDonough (Eds.), Proceedings of NELS 16 (pp. 322–337). Amherst,
MA: GLSA Publications.
300
Maris, E. & Oostenveld, R. (2007). Nonparametric statistical testing of EEG- and
MEG-data. Journal of Neuroscience Methods, 164 (1), 177–190. doi:10.1016/
j.jneumeth.2007.03.024
Matin, E., Shao, K. C., & Boff, K. R. (1993). Saccadic overhead: Information-
processing time with and without saccades. Perception & Psychophysics, 53 (4),
372–380. doi:10.3758/bf03206780
Mauner, G. & Koenig, J.-P. (1999). Lexical encoding of event participant informa-
tion. Brain and Language, 68, 178–184. doi:10.1006/brln.1999.2096
Mauner, G. & Koenig, J.-P. (2000). Linguistic vs. conceptual sources of implicit
agents in sentence comprehension. Journal of Memory & Language, 43 (1),
110–134. doi:10.1006/jmla.1999.2703
Mauner, G., Tanenhaus, M. K., & Carlson, G. N. (1995). Implicit arguments in
sentence processing. Journal of Memory & Language, 34, 357–382. doi:10 .
1006/jmla.1995.1016
McCourt, M., Green, J. J., Lau, E., & Williams, A. (2015). Processing implicit
control: Evidence from reading times. Frontiers in Psychology, 6, 1629. doi:10.
3389/fpsyg.2015.01629
McFadden, T. & Sundaresan, S. (2016). Failure to control is not a failure: It’s pro.
In C. Hammerly & B. Prickett (Eds.), Proceedings of NELS 46 (Vol. 3, pp. 1–
10). Amherst, MA: GLSA.
Merchant, J. (2001). The syntax of silence: Sluicing, islands, and the theory of ellip-
sis. Oxford studies in theoretical linguistics. Oxford; New York: Oxford Uni-
versity Press.
Mirman, D., Dixon, J. A., & Magnuson, J. S. (2008). Statistical and computational
models of the visual world paradigm: Growth curves and individual differences.
Journal of Memory & Language, 59 (4), 475–494. doi:10.1016/j.jml.2007.11.006
Mohanan, K. P. (1983). Functional and anaphoric control. Linguistic Inquiry, 14 (4),
641–674. JSTOR: 4178353
Nicol, J. & Swinney, D. (1989). The role of structure in coreference assignment
during sentence comprehension. Journal of Psycholinguistic Research, 18 (1),
5–19. doi:10.1007/BF01069043
Nishigauchi, T. (1984). Control and the thematic domain. Language, 60 (2), 215–
250. doi:10.2307/413640
301
Nunes, J. (1995). The copy theory of movement and linearization of chains in the
minimalist program (Doctoral dissertation, University of Maryland).
Nunes, J. (2001). Sideward movement. Linguistic Inquiry, 32 (2), 303–344. doi:10.
1162/00243890152001780
Nunes, J. (2014). Adjunct control and edge features. In P. Kosta, S. L. Franks, T.
Radeva-Bork, & L. Schürcks (Eds.), Minimalism and beyond: Radicalizing the
interfaces (pp. 79–108). Amsterdam/Philadelphia: John Benjamins. doi:10 .
1075/lfab.11.04nun
Parker, D., Lago, S., & Phillips, C. (2015). Interference in the processing of adjunct
control. Frontiers in Psychology, 6 (1346). doi:10.3389/fpsyg.2015.01346
Parker, D. & Phillips, C. (2017). Reflexive attraction in comprehension is selective.
Journal of Memory & Language, 94, 272–290. doi:10.1016/j.jml.2017.01.002
Pearlmutter, N. J., Garnsey, S. M., & Bock, K. (1999). Agreement processes in
sentence comprehension. Journal of Memory & Language, 41 (3), 427–456.
doi:10.1006/jmla.1999.2653
Percus, O. (2000). Constraints on some other variables in syntax. Natural Language
Semantics, 8 (3), 173–229. doi:10.1023/A:1011298526791
Percus, O. & Sauerland, U. (2003). On the LFs of attitude reports. In M. Weis-
berger (Ed.), Proceedings of Sinn and Bedeutung 7 (pp. 228–242). Universität
Konstanz. Konstanz.
Pietroski, P. M. (2005). Events and semantic architecture. Oxford University Press.
doi:10.1093/acprof:oso/9780199244300.001.0001
Pires, A. (2007). The derivation of clausal gerunds. Syntax, 10 (2), 165–203. doi:10.
1111/j.1467-9612.2007.00100.x
Pollard, C. & Sag, I. A. (1992). Anaphors in English and the scope of binding theory.
Linguistic Inquiry, 23 (2), 261–303. JSTOR: 4178768
Postal, P. M. (1971). Cross-over phenomena. New York.: Holt, Rinehart and Win-
ston.
R Core Team. (2014). R: A language and environment for statistical computing.
Retrieved from http://www.R-project.org/
Reinhart, T. (1981). Pragmatics and linguistics: An analysis of sentence topics.
Philosophica, 27 (1), 53–94.
302
Reinhart, T. (1983). Coreference and bound anaphora: A restatement of the anaphora
questions. Linguistics and Philosophy, 6 (1), 47–88. doi:10.1007/bf00868090
Reinhart, T. (2006). Interface strategies. Cambridge, MA: MIT Press.
Rizzi, L. (1990). Relativized minimality. Linguistic Inquiry Monographs. Cambridge,
MA: MIT Press.
Roberts, C. (1987). Modal subordination, anaphora, and distributivity (Doctoral dis-
sertation, University of Massachusetts, Amherst).
Roelofsen, F. (2008). Anaphora resolved (Doctoral dissertation, Institute for Logic,
Language, and Computation, Universiteit van Amsterdam).
Roeper, T. (1987). Implicit arguments and the head-complement relation. Linguistic
Inquiry, 18 (2), 267–310. JSTOR: 4178538
Ross, J. R. (1967). Constraints on variables in syntax (Doctoral dissertation, MIT).
HDL: 1721.1/15166
Ross, J. R. (1969). Guess who? In R. I. Binnick, A. Davison, G. M. Green, & J. L.
Morgan (Eds.), Papers from the fifth regional meeting of the Chicago Linguistic
Society (pp. 252–286). Chicago Linguistic Society. doi:10.1093/acprof:oso/
9780199645763.003.0002
Rothstein, S. (2004). Structuring events: A study in the semantics of lexical aspect.
Blackwell Publishing Ltd. doi:10.1002/9780470759127
Sag, I. A. (1976). Deletion and logical form (Doctoral dissertation, University of
Pennsylvania).
Sag, I. A. & Pollard, C. (1991). An integrated theory of complement control. Lan-
guage, 63–113. doi:10.2307/415539
Saito, M. & Murasugi, K. (1990). N′-deletion in Japanese. In UConn working papers
in linguistics (Vol. 3, pp. 80–107).
Schlenker, P. (2003). A plea for monsters. Linguistics and Philosophy, 26 (1), 29–
120. JSTOR: 25001879
Searle, J. (1965). What is a speech act? In M. Black (Ed.), Philosophy in America
(pp. 221–239). London: Allen and Unwin.
Sells, P. (1987). Aspects of logophoricity. Linguistic Inquiry, 18 (3), 445–479. JS-
TOR: 4178550
303
Speas, M. (2006). Economy, agreement, and the representation of null arguments. In
P. Ackema, P. Brandt, M. Schoorlemmer, & F. Weerman (Eds.), Arguments
and agreement. Oxford: Oxford University Press.
Stewart, A. J., Holler, J., & Kidd, E. (2007). Shallow processing of ambiguous pro-
nouns: Evidence for delay. The Quarterly Journal of Experimental Psychology,
60 (12), 1680–1696. doi:10.1080/17470210601160807
Stewart, A. J., Pickering, M. J., & Sanford, A. J. (2000). The time course of the in-
fluence of implicit causality information: Focusing versus integration accounts.
Journal of Memory & Language, 42 (3), 423–443. doi:10.1006/jmla.1999.2691
Stromdahl, L. (2018, January 5). The control of infinitival adjuncts and ditransitive
constructions. Poster presented at the Annual Meeting of the LSA, Salt Lake
City, UT.
Sundaresan, S. (2016). Perspective is syntactic: Evidence from anaphora. Unpub-
lished manuscript, University of Leipzig.
Uriagereka, J. (1999). Multiple spell-out. In S. D. Epstein & N. Hornstein (Eds.),
Working minimalism. Current studies in linguistics. Cambridge, Mass: MIT
Press.
Van Berkum, J. J. A. (2008). Understanding sentences in context: What brain waves
can tell us. Current Directions in Psychological Science, 17 (6), 376–380. doi:10.
1111/j.1467-8721.2008.00609.x
Venables, W. N. & Ripley, B. D. (2002). Modern applied statistics with S (4th ed.).
New York: Springer. doi:10.1007/978-0-387-21706-2
Whelpton, M. (1995). The syntax and semantics of infinitives of result in English
(Doctoral dissertation, Oxford).
Whelpton, M. (2001). Elucidation of a telic infinitive. Journal of Linguistics, 37 (2).
doi:10.1017/s0022226701008842
Whelpton, M. (2002). Locality and control with infinitives of result. Natural Lan-
guage Semantics, 10 (3), 167–210. doi:10.1023/A:1022107429786
Williams, A. (2015). Arguments in syntax and semantics. Cambridge University
Press. doi:10.1017/cbo9781139042864
Williams, A. & Green, J. J. (2017). Why control of PRO in rationale clauses is not
a relation between arguments. In A. Lamont & K. Tetzloff (Eds.), Proceedings
of NELS 47 (Vol. 3, pp. 233–246). Amherst, MA: GLSA.
304
Williams, E. (1974). Rule ordering in syntax (Doctoral dissertation, MIT). HDL:
1721.1/14770
Williams, E. (1980). Predication. Linguistic Inquiry, 11 (1), 203–238. JSTOR: 4178153
Williams, E. (1985). PRO and subject of NP. Natural Language & Linguistic Theory,
3 (3), 297–315. doi:10.1007/bf00154265
Williams, E. (1992). Adjunct control. In R. K. Larson, S. Iatridou, U. Lahiri, & J.
Higginbotham (Eds.), Control and grammar (pp. 297–321). Studies in linguis-
tics and philosophy. Dordrecht: Kluwer. doi:10.1007/978-94-015-7959-9 9
Winter, B. & Wieling, M. (2016). How to analyze linguistic change using mixed
models, Growth Curve Analysis and Generalized Additive Modeling. Journal
of Language Evolution, 1 (1), 7–18. doi:10.1093/jole/lzv003
Wyngaerd, G. V. & Zwart, J.-W. (1999). Antecedent-contained deletion as deletion.
In R. van Bezooijen & R. Kager (Eds.), Linguistics in the Netherlands 1999.
Amsterdam/Philadelphia: John Benjamins.
Yoshida, M., Dickey, M. W., & Sturt, P. (2013). Predictive processing of syntactic
structure: Sluicing and ellipsis in real-time sentence processing. Language and
Cognitive Processes, 28 (3), 272–302. doi:10.1080/01690965.2011.622905
Zubizarreta, M. L. (1982). On the relationship of the lexicon to syntax (Doctoral
dissertation, MIT). HDL: 1721.1/15654
305