University of Southern California Dissertations and Theses

Discourse-level processing of nominal possessive constructions

(USC Thesis Other)

Discourse-level processing of nominal possessive constructions

PDF

Download

Open document

Flip pages

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

DISCOURSE-LEVEL PROCESSING OF NOMINAL POSSESSIVE CONSTRUCTIONS

by

Jesse Storbeck

A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(LINGUISTICS)

December 2022

Copyright 2022 Jesse Storbeck
ii
Acknowledgments

I would not have been able to complete my graduate education or this dissertation
without the support of many kind and talented individuals. Firstly, I would like to thank my
advisor, Elsi Kaiser, whose guidance and expertise have been essential to my development as a
researcher. In addition to improving the quality of my work, Elsi has also been an invaluable
source of encouragement and optimism when I doubted my ability to reach this point. I am also
grateful for the support she provided to me in the form of research assistantships, through which
I was able to grow my methodological and analytical skillsets significantly. Additionally, I owe a
debt of gratitude to my committee members Toby Mintz and Alexis Wellwood for their
thoughtful feedback on this work. I also thank Toby especially for his help during my earlier
years as a graduate student.
Many faculty, staff, and students have supported me throughout my time at USC—too
many to name here—but I appreciate all of you. In particular, I need to thank Guillermo Ruiz for
his assistance with a multitude of administrative tasks over the years and, more importantly, his
kind presence in the department. I also would not have made it through the PhD program without
my fellow graduate students, especially my cohort and members of the Language Processing
Lab. I will cherish my memories of our jokes in the office, adventures at conferences, and
encouraging Zoom chats during the darkest times of the pandemic. Additionally, I would like to
thank Madeline Rouse for her assistance annotating the continuation data in Chapters 2 and 3 of
this dissertation.
My family has been vital to my graduate education. My parents, Tim and Shelly, have
always encouraged my intellectual curiosity and indulged my esoteric interests. My sister,
iii
Emma, has been a reliable source of inspiration, humor, and reassurance. Without their love and
support, I never could have completed this degree.
Thank you to Alexia, who has rescued my mental health on too many occasions to count
and graciously tolerated the unusual habits and schedule of a PhD student. You have seen me
through the highs and lows of the program and always helped me find the motivation to
continue. Finally, I thank Lucinda and Sandy, my feline companions, who have slept nearby
throughout almost all of the writing of this dissertation.

iv
Table of Contents

Acknowledgments ........................................................................................................................... ii
List of Tables ................................................................................................................................ vii
List of Figures ................................................................................................................................ xi
Abstract ........................................................................................................................................ xiv
Chapter 1: Introduction ....................................................................................................................1
1.1. Overview .......................................................................................................................1
1.2. The scope of possessive constructions under investigation ..........................................2
1.3. Defining discourse prominence ....................................................................................3
1.4. Approaches to prominence in possessives ....................................................................5
1.5. Existing theory’s implications for possessives .............................................................7
1.6. The structure of this dissertation .................................................................................11

Chapter 2: Testing how possession affects discourse prominence using a sentence
continuation task ..........................................................................................................14
2.1. Introduction .................................................................................................................14
2.2. Should possession decrease or increase a referent’s prominence? .............................15
2.3. Does the type of possession relation affect prominence? ...........................................17
2.4. Experiment 1 ...............................................................................................................21
2.4.1. Participants ...................................................................................................22
2.4.2. Design and materials ....................................................................................23
2.4.3. Procedure .....................................................................................................25
2.4.4. Pre-processing of the continuation data .......................................................25
2.4.5. Predictions ....................................................................................................26
2.4.6. Results ..........................................................................................................29
2.4.6.1. Mentions of the prompt subject as continuation subject ...............30
2.4.6.2. Mentions of the prompt object as continuation subject ................32
2.4.6.3. Mentions of the prompt subject anywhere in continuations .........33
2.4.6.4. Mentions of the prompt object anywhere in continuations ...........35
2.4.6.5. Pronominalization patterns ...........................................................36
2.5. Discussion ...................................................................................................................40
2.5.1. Implications for the Complexity Hypothesis and Parasitic Possession
Hypothesis ...................................................................................................40
2.5.2. Implications for the Interaction Hypothesis and Animacy Hypothesis .......43

Chapter 3: Investigating effects of representational complexity on the discourse
representations of possessives ......................................................................................47
3.1. Introduction .................................................................................................................47
3.2. Comparing possessives, indefinites, and definites ......................................................48
v
3.3. The role of representational complexity .....................................................................50
3.4. Experiment 2 ...............................................................................................................52
3.4.1. Participants ...................................................................................................54
3.4.2. Design and materials ....................................................................................55
3.4.3. Procedure .....................................................................................................57
3.4.4. Pre-processing of the continuation data .......................................................58
3.4.5. Predictions ....................................................................................................59
3.4.6. Results ..........................................................................................................61
3.4.6.1. Mentions of the prompt subject as continuation subject ...............61
3.4.6.2. Mentions of the prompt object as continuation subject ................63
3.4.6.3. Mentions of the prompt subject anywhere in continuations .........65
3.4.6.4. Mentions of the prompt object anywhere in continuations ...........66
3.4.6.5. Pronominalization patterns ...........................................................68
3.5. Discussion ...................................................................................................................72
3.5.1. Support for the Parasitic Possession Hypothesis and the Animacy
Hypothesis ...................................................................................................73
3.5.2. Representational complexity and an unexpected adjective modification
effect ............................................................................................................75
3.5.3. Underlying mechanisms of the Parasitic Possession Hypothesis ................76
3.5.4. Combined insights from Experiments 1 and 2 .............................................79

Chapter 4: Investigating real-time processing of possession relations with web-based eye-
tracking ........................................................................................................................83
4.1. Introduction .................................................................................................................83
4.2. Experiment 3 ...............................................................................................................85
4.2.1. Participants ...................................................................................................86
4.2.2. Design and materials ....................................................................................88
4.2.2.1 Auditory stimuli .............................................................................90
4.2.2.2 Visual stimuli .................................................................................91
4.2.3. Procedure .....................................................................................................95
4.2.4. Predictions ....................................................................................................99
4.2.5. Results ........................................................................................................104
4.2.5.1. Selections of the last-mentioned referent ....................................105
4.2.5.2. Selection latencies .......................................................................106
4.2.5.3. General remarks on the eye-tracking analysis ............................109
4.2.5.4. Time windows in the eye-tracking analysis ................................112
4.2.5.5. Eye-movements during the matrix-clause ..................................115
4.2.5.6. Eye-movements during pronoun interpretation: pronoun-
onset alignment ...........................................................................120
4.2.5.7. Eye-movements during pronoun interpretation: pronoun-
offset alignment ..........................................................................126
4.3. Discussion .................................................................................................................131
4.3.1 Effects of animacy ......................................................................................132
4.3.2 Effects of possession: support for the Parasitic Possession Hypothesis .....134

vi
Chapter 5: Effects of possession type on the comprehension of ambiguous verb-phrase
ellipsis ........................................................................................................................138
5.1. Introduction ...............................................................................................................138
5.2. Ambiguous possessive pronouns in verb-phrase ellipsis ..........................................139
5.3. Connecting verb-phrase ellipsis to possession at the discourse level .......................141
5.4. Experiment 4 .............................................................................................................144
5.4.1. Participants .................................................................................................146
5.4.2. Design and materials ..................................................................................146
5.4.3. Procedure ...................................................................................................149
5.4.4. Predictions ..................................................................................................149
5.4.5. Results ........................................................................................................152
5.4.6. Discussion ..................................................................................................155
5.5. Experiments 5a–d ......................................................................................................158
5.5.1. Participants .................................................................................................161
5.5.2. Design and materials ..................................................................................161
5.5.3. Procedure ...................................................................................................164
5.5.4. Predictions ..................................................................................................164
5.5.5. Results ........................................................................................................166
5.5.6. Discussion ..................................................................................................170
5.6. General Discussion ...................................................................................................171

Chapter 6: Conclusion ..................................................................................................................173
6.1. Overview and summary ............................................................................................173
6.2. What underlies the Parasitic Possession Hypothesis? ..............................................178
6.3. Reconsidering the Complexity Hypothesis ...............................................................180
6.4. Opportunities for future work ...................................................................................181

References ....................................................................................................................................184
Appendices ...................................................................................................................................195
Appendix A: Target Stimuli for Experiment 1 ................................................................195
Appendix B: Target Stimuli for Experiment 2 ................................................................200
Appendix C: Target Stimuli for Experiment 3 ................................................................209
Appendix D: Target Stimuli for Experiments 4 and 5a–d ...............................................214
Appendix E: Model Formulae for Experiment 1 .............................................................227
Appendix F: Model Formulae for Experiment 2 .............................................................231
Appendix G: Model Formulae for Experiment 3 .............................................................234
Appendix H: Model Formulae for Experiment 4 .............................................................248
Appendix I: Model Formulae for Experiments 5a–d .......................................................250
vii
List of Tables

Table 1. Example target items (2 of 24) in Experiment 1 in each of the four conditions .............23
Table 2. Summary of fixed effects for the model predicting mentions of the preceding
subject in subject position of the continuation (Experiment 1) ......................................31
Table 3. Summary of fixed effects for the model predicting mentions of the preceding
object in subject position of the continuation (Experiment 1) ........................................32
Table 4. Summary of fixed effects for the model testing the simple effect of possession
within the animate conditions (subject-position analysis, Experiment 1) ......................33
Table 5. Summary of fixed effects for the model testing the simple effect of possession
within the inanimate conditions (subject-position analysis, Experiment 1) ...................33
Table 6. Summary of fixed effects for the model predicting mentions of the preceding
subject anywhere in the continuation (Experiment 1) ....................................................35
Table 7. Summary of fixed effects for the model predicting mentions of the preceding
object anywhere in the continuation (Experiment 1) ......................................................35
Table 8. Summary of fixed effects for the model testing the simple effect of possession
within the animate conditions (entire-continuation analysis, Experiment 1) .................36
Table 9. Summary of fixed effects for the model testing the simple effect of possession
within the inanimate conditions (entire-continuation analysis, Experiment 1) ..............36
Table 10. Summary of fixed effects for the model predicting pronominalization of the
prompt subject in subject position (Experiment 1) .......................................................38
Table 11. Summary of fixed effects for the model predicting pronominalization of the
prompt object in subject position (Experiment 1) .........................................................38
viii
Table 12. Summary of fixed effects for the model predicting pronominalization of the
prompt subject anywhere in the continuation (Experiment 1) ......................................39
Table 13. Summary of fixed effects for the model predicting pronominalization of the
prompt object anywhere in the continuation (Experiment 1) ........................................39
Table 14. Example target items (2 of 24) in Experiment 2 in each of the 8 conditions ................55
Table 15. Summary of fixed effects for the model predicting mentions of the prompt subject
in subject position of the continuation (Experiment 2) .................................................62
Table 16. Summary of fixed effects for the model predicting mentions of the prompt object
in subject position of the continuation (Experiment 2) .................................................64
Table 17. Summary of fixed effects for the model predicting mentions of the prompt subject
in any position of the continuation (Experiment 2) .......................................................66
Table 18. Summary of fixed effects for the model predicting mentions of the prompt object
in any position of the continuation (Experiment 2) .......................................................67
Table 19. Summary of fixed effects for the model predicting pronominalization of the
prompt subject in subject position of the continuation (Experiment 2) ........................70
Table 20. Summary of fixed effects for the model predicting pronominalization of the
prompt subject in all positions of the continuation (Experiment 2) ..............................70
Table 21. Summary of fixed effects for the model predicting pronominalization of the
prompt object in subject position of the continuation (Experiment 2) ..........................72
Table 22. Summary of fixed effects for the model predicting pronominalization of the
prompt object in all positions of the continuation (Experiment 2) ................................72
Table 23. Example target items (2 of 24) in Experiment 3 in each of the 4 conditions. ...............88

ix
Table 24. Summary of fixed effects for the model predicting selections of the last-
mentioned referent (Experiment 3) .............................................................................106
Table 25. Summary of fixed effects for the model predicting selection latencies relative to
the onset of the pronoun (Experiment 3) .....................................................................108
Table 26. Summary of fixed effects for the models predicting looks to the subject and object
at time intervals aligned to the onset of the matrix-clause object (Experiment 3);
significant effects are shaded ......................................................................................118
Table 27. Summary of fixed effects for the models predicting looks to the subject and object
regions at time intervals aligned to the onset of the pronoun (Experiment 3);
significant effects are shaded ......................................................................................123
Table 28. Summary of fixed effects for the follow-up models predicting looks to the subject
and object in the animate conditions (aligned to the onset of the pronoun,
Experiment 3); significant effects are shaded .............................................................125
Table 29. Summary of fixed effects for the models predicting looks to the subject and object
at various time intervals aligned to the offset of the pronoun (Experiment 3);
significant effects are shaded ......................................................................................129
Table 30. Summary of fixed effects for the follow-up models predicting looks to the subject
and object in the animate conditions (aligned to the offset of the pronoun,
Experiment 3); significant effects are shaded .............................................................130
Table 31. Summaries of fixed effects in the models for Experiment 4 .......................................154
Table 32. Summary of fixed effects for the model predicting the rate of bound variable
interpretations in Experiments 5a–d ............................................................................167

x
Table 33. Summary of fixed effects for the model predicting the rate of bound variable
interpretations in Experiments 5a and 5c ....................................................................168
Table 34. Summary of fixed effects for the model predicting the rate of bound variable
interpretations in Experiments 5b and 5d ....................................................................169

xi
List of Figures

Figure 1. Example item from Experiment 1 as it appeared to participants ...................................25
Figure 2. The proportion of continuations in Experiment 1 referring back to the prompt
subject, prompt object, or something else (error bars represent ±1 standard error) ......30
Figure 3. The proportion of continuations in Experiment 1 which mention the prompt
subject and/or prompt object (error bars represent ±1 standard error) ..........................34
Figure 4. Pronominalization likelihood (conditional probability of pronoun given mention)
in Experiment 1 for the prompt subject and object in subject position and all
positions of the continuation (error bars represent ±1 standard error) ..........................37
Figure 5. Example item from Experiment 2 as it appeared to participants ...................................58
Figure 6. The proportion of continuations in Experiment 2 in which the subject refers back
to the prompt subject (error bars represent ±1 standard error) ......................................62
Figure 7. The proportion of continuations in Experiment 2 in which the subject refers to the
prompt object (error bars represent ±1 standard error) .................................................63
Figure 8. The proportion of continuations in Experiment 2 which contain a reference to the
prompt subject (error bars represent ±1 standard error) ................................................65
Figure 9. The proportion of continuations in Experiment 2 which contain a reference to the
prompt object (error bars represent ±1 standard error) .................................................67
Figure 10. Pronominalization likelihood (conditional probability of pronoun given mention)
in Experiment 2 for the prompt subject in subject position of continuations
(error bars represent ±1 standard error) .......................................................................68

xii
Figure 11. Pronominalization likelihood (conditional probability of pronoun given mention)
in Experiment 2 for the prompt subject in all positions of the continuation (error
bars represent ±1 standard error) .................................................................................69
Figure 12. Pronominalization likelihood (conditional probability of pronoun given mention)
in Experiment 2 for the prompt object in subject position of continuations (error
bars represent ±1 standard error) .................................................................................71
Figure 13. Pronominalization likelihood (conditional probability of pronoun given mention)
in Experiment 2 for the prompt object in all positions of the continuation (error
bars represent ±1 standard error) .................................................................................71
Figure 14. An example screen layout from Experiment 3 for the target sentence The soldier
heard the plumber throughout the autumn evening because he was nimpacious ........91
Figure 15. An example screen layout from Experiment 3 for the target sentence The runner
praised her nurse during the podcast interview because she was pambly ..................92
Figure 16. An example screen layout from Experiment 3 for the filler sentence The golfer
ate the bromble because the surfer took the last slice of pizza (utilizing the top-
left, bottom-left, and right regions, with the center regions empty) ............................93
Figure 17. An example screen layout from Experiment 3 for the filler sentence The teapot
filled with spreltoom was prepared for the pirate, who drank it with honey
(utilizing the same left, center, and right regions as in the targets) .............................93
Figure 18. The screen layout for the first of three practice items in Experiment 3 (with
feedback displayed after a correct click) .....................................................................97
Figure 19. The proportion of selections of the subject as the last-mentioned referent in
Experiment 3 (error bars represent ±1 standard error) ...............................................105
xiii
Figure 20. Selection latencies in Experiment 3 relative to the onset of the pronoun (error
bars represent ±1 standard error) ...............................................................................108
Figure 21. Eye movements in Experiment 3 relative to the onset of the center-fixation cross
(animate conditions vs. inanimate conditions) ...........................................................111
Figure 22. Eye movements in Experiment 3 relative to the onset of the matrix-clause object
(each facet shows looks to the three regions of interest within a condition) .............116
Figure 23. Eye movements in Experiment 3 relative to the onset of the matrix-clause object
(facets compare conditions in the subject and object regions) ..................................117
Figure 24. Eye movements in Experiment 3 relative to the onset of the pronoun (each facet
shows looks to the three regions of interest within a condition) ................................121
Figure 25. Eye movements in Experiment 3 relative to the onset of the pronoun (facets
compare conditions in the subject and object regions) ..............................................121
Figure 26. Eye movements relative to the offset of the pronoun (each facet shows looks to
the three regions of interest within a condition) ........................................................127
Figure 27. Eye movements in Experiment 3 relative to the offset of the pronoun (facets
compare conditions in the subject and object regions) ..............................................128
Figure 28. An example target item from Experiment 4 as it appeared to participants ...............149
Figure 29. The proportion of bound variable responses by possession type in Experiment 4
(error bars represent ±1 standard error) .....................................................................153
Figure 30. The relationship between each noun’s possession likelihood and the proportion
of bound variable interpretations which it produced in Experiment 4 .......................158
Figure 31. The proportion of bound variable interpretations by possession-type condition in
Experiments 5a–d (error bars represent ±1 standard error) .......................................166
xiv
Abstract

This dissertation explores how the referents involved in nominal possessive constructions
are represented at the discourse level, how possession affects discourse prominence, and how
different semantic possession relations between the possessor and possessed referent influence
discourse-level processing.
To address these research questions I formulate two sets of competing hypotheses about
how possession affects the prominence of possessed referents. The first set of hypotheses relates
to the overall effect of possession on prominence (abstracting away from variation across
different types of possession relations): the Parasitic Possession Hypothesis claims that
involvement in a possessive construction reduces the prominence of the possessed referent, while
the Complexity Hypothesis claims the opposite—that possession increases the prominence of the
possessed referent. The second set of hypotheses relates to whether different types of possession
relations modulate the overall discourse-level effect of possession: the Interaction Hypothesis
posits an exceptional boost in prominence for possessed animates, while the Animacy
Hypothesis attributes any differences in prominence between possessed animates and possessed
inanimates to the effect of animacy (rather than any influence due to the possession relation
itself).
I test these hypotheses in five experiments. Experiments 1 and 2 (Chapters 2 and 3,
respectively) use a sentence-continuation paradigm to compare the prominence of possessed
referents to two types of simple nominals: indefinites (Experiment 1) and definites (Experiment
2). Both experiments primarily analyze the likelihood that participants re-mention referents from
a prompt, but they produce somewhat different results. In Experiment 1, I do not find strong
xv
evidence for either the Parasitic Possession Hypothesis or the Complexity Hypothesis (which
relate to the overall effect of possession on prominence); however, the results do suggest an extra
boost in prominence for possessed animates, which lends support to the Interaction Hypothesis.
Experiment 2, on the other hand, finds evidence that possession reduces the prominence of
possessed referents overall and that the type of possession relation does not modulate this
reduction in prominence. Thus Experiment 2 supports the Parasitic Possession Hypothesis and
the Animacy Hypothesis.
Experiment 3 (Chapter 4) uses visual-world eye-tracking to examine gaze patterns during
ambiguous pronoun resolution and compares how participants resolve a pronoun when one of its
potential antecedents is possessed versus definite. The results show that possessed referents are
less accessible as pronominal antecedents and therefore support the Parasitic Possession
Hypothesis—that possession reduces a referent’s prominence.
Experiments 4 and 5 (Chapter 5) examine the resolution of ambiguous possessive
pronouns in a type of verb-phrase ellipsis. Here I adopt a discourse-based approach to ellipsis
interpretation which relies on the prominence of possessed referents. I find that the type of
possession relation significantly affects ambiguity resolution in these elliptical constructions and
relate the results back to the Interaction Hypothesis.
In sum, the results of this dissertation support the Parasitic Possession Hypothesis and the
Interaction Hypothesis. I argue that possession generally decreases the prominence of possessed
referents, although different types of possession relations may modulate this overall effect. In
particular, possessed animates seem to be especially prominent in discourse.
1
Chapter 1: Introduction

1.1. Overview
A fundamental principle in psycholinguistic theory is that, when people process
language, they form mental models which include representations of mentioned entities
(“discourse referents”) (e.g. Chafe, 1974; Clark & Haviland, 1977). The referents involved in a
discourse differ in prominence (alternatively “salience”, “accessibility”, or “activation”
1
), and the
prominence of a given referent inevitably changes over time as the discourse continues (e.g.
Ariel, 1988; Chafe, 1976; Clark & Sengul, 1979; Givón, 1983; van den Broek et al., 1996).
Investigating which factors modulate prominence—as well as what kinds of referential forms are
used for referents with different prominence levels—has been a major topic of psycholinguistic
research; however, most previous work on language processing at the discourse level has focused
on simple noun phrases which include a mention of a single discourse referent (e.g. the attorney,
Kate, she). Within the smaller body of work on complex noun phrases, which include mentions
of multiple referents, even less attention has been devoted to the discourse-level behavior of
nominal possessive constructions (her car, his attorney, Bill’s sister).
While nominal possessives occur frequently in language, their discourse-level properties
are understudied, and existing theories based on simple noun phrases do not extend
straightforwardly to account for them. Unlike simple noun phrases like the car, nominal
possessives like Sam’s car mention two entities within a single nominal constituent: a possessor
(Sam) and a possession (car). In addition, possessives can express such a broad range of
semantic relations between the possessor and possessed referent (e.g. ownership, part-whole, and

1
These terms are related, but not necessarily equivalent, and have been used in different ways within linguistic
literature. See section 1.3.
2
kinship relations, among others) that describing them as a unified class requires especially
inclusive and abstract language (e.g. Haspelmath, 2017; Langacker, 1993; 1995). This
dissertation aims to shed light on how the referents involved in nominal possessive constructions
are represented in discourse, how possession affects a referent’s prominence, and how different
semantic possession relations might affect discourse-level processes.

1.2. The scope of possessive constructions under investigation
While possessive structures are not confined to the nominal domain (e.g. there are also
verbal predicate constructions like Sam has a car), this dissertation investigates nominal
possessives. Specifically, I examine adnominal possessives (or “adpossessives”), which
Haspelmath (2017) defines as:
…a construction in which a possessive relationship is expressed in a single
nominal (or NP) consisting of (i) a noun expressing the possessum (the possessed
noun) and (ii) a modifying possessor nominal or a possessive person index (i.e. a
bound person form). (p. 196)
In English, these constructions are realized in two syntactic configurations, s-genitives (e.g.
Sam’s car, the doctor’s office, my cousin) and of-genitives (e.g. the floor of the bedroom, the
capital of California, the father of the bride). I focus on adpossessives of the s-genitive variety
for two main reasons: firstly, animate possessors exhibit a broad range of semantic possession
types, including interpersonal, ownership, and part-whole relations—key for investigating how
different possession relations affect discourse-level representation—yet animates are typically
less acceptable as possessors in of-genitives (e.g. Anschutz, 1997; Rosenbach, 2002). Secondly,
the work in this dissertation involves pronouns, which can be ungrammatical or degraded as
3
possessors in of-genitive constructions (e.g. the palace of the king vs. *the palace of him; the
roof of the house vs. ?the roof of it) (Rosenbach, 2002).

1.3. Defining discourse prominence
Although prominence is an intuitive concept with clear theoretical value, it is particularly
elusive to define. Additionally, it is often discussed interchangeably with related concepts such
as salience, accessibility, and activation. Here it may be helpful to introduce a definitional
property of prominence proposed by von Heusinger & Schumacher (2019) (following
Himmelmann & Primus, 2015; Jasinskaja et al., 2015), which states, “Prominent elements are
structural attractors, i.e. they serve as anchors for the larger structures they are constituents of,
and they may license more operations than their competitors” (von Heusinger & Schumacher,
2019, p. 119). In the context of discourse, the linguistic operations which prominent referents
attract include a host of dependent variables which have been noted in the literature as indicative
of prominence: for example, subjecthood, linear order, re-mention, and reference with
pronominal or zero forms (e.g. Ariel, 1990; Chafe, 1976; Clark & Begun, 1971; Givón, 1983;
Gundel et al., 1993).
The fact that prominence is associated with a variety of empirical observations can make
it a challenging concept to investigate. There is currently no consensus among researchers which
dependent variable (or set of variables) provides the most accurate view of a referent’s
prominence. Furthermore, while prominence is often discussed as a unidimensional quantity, it
may in fact be more precisely understood as a multidimensional property. Notwithstanding these
gaps in our current understanding, it seems that the best way to advance toward a more secure
and useful theory of prominence is to investigate it through a range of dependent variables. For
4
example animates can confidently be considered more prominent in general than inanimates
precisely because, whether one considers their rate of subjecthood, their tendency to be realized
earlier in utterances, or their likelihood of pronominalization, the conclusions converge toward a
theory in which animacy confers prominence (e.g. Bock et al., 1992, Branigan et al., 2008; Dahl,
2008; Fukumura & van Gompel, 2011). For this dissertation’s investigation of prominence in
possessives, I adopt an analogous strategy by considering production-based measures like rates
of re-mention, subjecthood, and pronominalization, as well as availability as a pronominal
antecedent, which is based on comprehension.
This multidimensional approach to prominence does not mean, however, that all
dependent variables provide equivalent views of prominence. For example, a referent’s
likelihood of mention has been shown to be dissociable from its likelihood of pronominalization
(e.g. Fukumura & van Gompel, 2010; Kaiser, 2010; Kehler et al., 2008; Kehler & Rohde, 2013).
It may then be possible to measure changes in prominence through different dependent variables
and obtain conflicting results. Nevertheless, some prior work has argued that different measures
of discourse prominence can be integrated with a Bayesian approach (e.g. Kehler et al., 2008;
Kehler & Rohde, 2013). In the Bayesian equation proposed in this work, the likelihood that a
pronoun refers to a particular referent is proportional to the product of that referent’s likelihood
of mention and the likelihood that a pronoun will be produced to refer to that referent:
P(referent|pronoun) ∝ P(referent)P(pronoun|referent). The lefthand side of this expression is
oriented toward comprehension—specifically pronoun resolution—since the probability is
conditioned on the presence of a pronoun; however, the righthand side includes the rates of
mention and pronominalization, which are production-based measures. While these three terms
may be dissociable, they are still related, and a detailed investigation of the discourse properties
5
of any linguistic feature benefits from an examination of all of them. With this view of
prominence in mind, this dissertation investigates the discourse-level effects of possession with a
focus on how we comprehend and refer back to possessors and possessed referents.

1.4. Approaches to prominence in possessives
In considering the prominence of the two referents involved in a possessive construction,
there are two possible approaches; one can assess their prominence relative to each other (i.e.
does the possessor outrank the possessed referent or vice versa?). For example, in Sam’s car, one
can consider whether Sam or car is more prominent. Alternatively, one can estimate the effect of
the possession relation itself on the prominence of the two referents (i.e. how does participation
in a possession relation change the discourse status of the possessor and possessed referent?). For
example, one can compare the prominence of the head noun’s (car) referent in Sam’s car versus
the car. Both of these approaches are valid and have advantages depending on the nature of the
investigation. Neither has been explored in great depth; however, the small amount of existing
work on possessives’ discourse-level behavior has focused on the relative prominence of the
possessor and possessed referent (e.g. Chae, 2003; Di Eugenio, 1998; Walker & Prince, 1996).
Prior work on this topic comes from the literature on Centering Theory (e.g. Gordon et
al., 1993; Grosz et al., 1995; Walker et al., 1998). Centering Theory is a computational approach
based primarily on corpus data which seeks to model local coherence between the utterances
comprising a discourse. In Centering Theory, each utterance yields a list of “forward-looking
centers” (or “Cf”), which is essentially a ranking of the mentioned referents by prominence. The
ranking of the Cf list is crucially important for Centering Theory, since the highest-ranking
referent in Cf becomes that utterance’s preferred center (or “Cp”); the Cp in turn predicts the
6
backward-looking center (“Cb”), i.e. the topic, of the next utterance. The relationship within an
utterance between the Cp and Cb—as well as across utterances between the current Cb and prior
Cb—defines a set of transition states which form the crux of Centering Theory. Consequently,
the ranking of the Cf list (i.e. the relative prominence of all referents mentioned in an utterance)
is paramount for any investigation within this framework. Thus it makes sense why approaches
to possessives within Centering Theory would be primarily concerned with the relative
prominence of the possessor and possessed referent.
The most in-depth treatment of possessives within Centering Theory comes from Chae
(2003), who seeks to adjudicate between two previous accounts. The first is the Complex NP
Assumption, a more general constraint proposed by Walker & Prince (1996); this hypothesis
states that the referents in complex noun phrases (of which possessives are one type) are ranked
in the Cf list based on their left-to-right order (at least in English). Therefore, for s-genitives, the
possessor would always be higher ranked in the Cf list than the possessed referent. The second
account which Chae tests comes from Di Eugenio (1998), who suggests a more nuanced
treatment of possessives. Di Eugenio proposes that animate possessed referents immediately
outrank their possessors in the Cf list, while inanimate possessed referents are ranked
immediately below their possessors; however, Di Eugenio characterizes this approach as a
“heuristic” which requires further investigation. Based on a corpus analysis, Chae finds limited
evidence in support of Di Eugenio’s proposal and instead favors the approach which always
ranks possessors higher than their corresponding possessed referents.
Consideration of the relative prominence of the possessor and possessed referent is
clearly an appropriate approach in the context of Centering Theory; however, for the purposes of
7
this dissertation in the field of experimental psycholinguistics, there are reasons to favor an
investigation of the effect of the possession relation itself on the involved referents.
Firstly, by examining the effect of the possession relation (e.g. how prominent is the
referent for car in Sam started the car versus Sam started his car), I ground this dissertation’s
contribution more effectively in the large existing literature on the discourse behavior of simple
nominals. Secondly, from an experimental perspective, comparing the prominence of the
possessor and possessed referent may be problematic, since—even in a well-controlled
experiment—it might be unavoidable that they differ in factors such as animacy, grammatical
role, thematic role, givenness, and topicality, which are known to affect prominence (e.g. Ariel,
1990; Arnold, 2001; Bock et al., 1992; Givón, 1983; Gundel et al., 1993; Kaiser, 2011;
Stevenson et al. 1994). On the other hand, this approach makes more sense for a corpus analysis,
since a large and varied set of tokens might allay concerns of confounding factors. Finally, the
primary aim of this dissertation aim of this dissertation is to assess whether being anchored to
another referent through a possession relation impacts the discourse status of the involved
referents. The most direct way to address this aim is through measuring the effect of possession
on referents’ prominence relative to non-possessed forms.

1.5. Existing theory’s implications for possessives
This dissertation primarily investigates the discourse-level effects of possessive relations
on the involved referents. What can existing theory tell us about how these effects might play
out? In fact, prior work—including the investigations within Centering Theory I mention in
Section 1.4—gives rise to competing predictions. Here I address the three primary logical
possibilities for how involvement in a possessive construction might affect a referent’s
8
prominence. I frame these predictions in terms of how possession changes the prominence of the
possessed referent (e.g. the change to the referent for car in the car vs. her car), since these make
for a tidy minimal pair; however, under the common assumption that referents compete among
each other for limited attentional resources (e.g. Arnold & Griffin, 2007; Chiriacescu, 2015;
Fukumura & van Gompel, 2011), one could extend these predictions such that an increase in
prominence for one referent involved in the possessive would suggest a reduction in prominence
for the other.
The first logical possibility for the effect of possession on a possessed referent is that
there is none, i.e. the null hypothesis. We could assume no change in the prominence of the
referent for car in the car vs. her car, and estimate its prominence based on previously identified
discourse factors. This approach seems unsatisfactory, however, since presumably the link
between the possessor and possessed referent is represented somehow on the discourse level, and
it would be surprising for it to play no role in the linked referents’ prominence. Additionally,
prior work provides a priori reasons to believe that possession might alternatively decrease or
increase the possessed referent’s prominence.
The second possible effect of possession is that it decreases the prominence of possessed
referents. This view is indirectly supported by prior work in Centering Theory which claims that
possessors in s-genitive constructions outrank their corresponding possessed referents (e.g. Chae,
2003; Walker & Prince, 1996). If we consider a simple minimal pair such as The car is in the
garage vs. Her car is in the garage, the referent for car is likely the most prominent in the first
sentence by virtue of its subjecthood; however, in the second sentence, Centering Theory would
predict that the antecedent of the possessive pronoun outranks the referent for car. Thus, one can
extrapolate that the effect of possession on the possessed referent is a reduction in prominence.
9
In subsequent chapters, I refer to this view as the Parasitic Possession Hypothesis and explore
how the structure of possessives’ discourse representations might produce this pattern.
Thirdly, there are reasons to expect conversely that possession might increase the
prominence of the possessed referent. Comparing the car to her car syntactically, the former is a
simple noun phrase and the latter a complex noun phrase. Likewise, it seems reasonable to
expect that a possessed referent is also more complex at the discourse level by virtue of its
relation to its possessor. For example, when processing her car, a comprehender presumably
must link the discourse referent for car to the antecedent of her in order to establish an
ownership relation between the two. This linking operation need not occur with the car, and if it
does occur, it is the result of pragmatic reasoning about who the owner of the car is. Possessives’
inherent complexity thus connects my work to prior literature on representational complexity and
informativity. For example, research on memory has shown that words with a greater number of
associated features are more easily retrieved (e.g. Craik & Tulving, 1975; Fisher & Craik, 1980),
and more recent psycholinguistic work has demonstrated that antecedents with greater
representational complexity are accessed more quickly in resolving filler-gap and anaphoric
dependencies (e.g. Hofmeister, 2011; Hofmeister & Vasishth, 2014; Troyer et al., 2016).
Additionally, referents which encode more information have been shown to be more likely
antecedents for pronouns (e.g. Karimi & Ferreira, 2016; Karimi et al., 2014, 2019). Generally
speaking, these kinds of results suggest that more complex representations enjoy greater
prominence. Assuming this is the case, one might also expect that if the possessive relation links
the possessed referent to an existing salient possessor, some amount of prominence might be
transferred to the possessed referent. For these reasons, I consider the alternative hypothesis that
possession increases the prominence of possessed referents, which I call the Complexity
10
Hypothesis. The predictions of this hypothesis are directly opposed to those of the Parasitic
Possession Hypothesis discussed above.
Finally, there is an additional dimension to the preceding proposals, which is that perhaps
the discourse-level effects of possession depend on the type of semantic possession relation. For
example, is the difference in the prominence of the referent for car in her car versus the car
different from the corresponding difference in the prominence in her lawyer versus the lawyer?
This possibility is in line with the claim of Di Eugenio (1998) that the relative ranking of the
referents in a possessive construction depends on the animacy of the possessed referent.
Likewise, this dissertation explores the effect of different possession relations primarily through
manipulation of the animacy of the possessed referent (while maintaining the possessor as
animate). I focus on animacy as a determinant of possession relation not only for its connection
to Di Eugenio’s proposal, but also because animacy has been extensively researched as a
discourse-level factor.
In general, animate referents are commonly understood to be more prominent than
inanimates in discourse and memory (e.g. Dahl, 2008; Fukumura & van Gompel, 2011; Gelin et
al., 2017). Animacy affects word order with animate referents more often appearing before
inanimates and in subject position (e.g. Bock et al., 1992, Branigan et al., 2008; Dahl, 2008; Dahl
& Fraurud, 1996; Prat-Sala & Branigan, 2000). Animate referents have also been argued to have
greater conceptual accessibility than inanimates, as they tend to be pronominalized more often
(e.g. Dahl & Fraurud, 1996; Fukumura & van Gompel, 2011; Yamamoto, 1999). Additionally,
animacy is already known to affect the syntactic alternation between s-genitives and of-genitives
in English (e.g. Anschutz, 1997; Rosenbach, 2002). Thus it seems to be a natural starting point to
explore the contribution of the possession relation to discourse prominence via manipulation of
11
animacy. Additionally, in the final set experiments reported in this dissertation, I explore
different possession types within the categories of animate and inanimate possessed referents.

1.6. The structure of this dissertation
Chapters 2 and 3 report two sentence-continuation experiments which compare the
discourse status of possessives to two types of simple noun phrases: indefinites (Experiment 1)
and definites (Experiment 2). In both experiments, I use a sentence completion paradigm to
measure how likely people are to mention the possessor and possessed referent after reading a
prompt like Daniel zatted {a, his} jacket (nonce verbs were used in Experiment 1). I also
consider the form of mention for these referents (i.e. how likely the participant is to use a
pronoun for a particular referent). I analyze the results with the assumption that a higher
likelihood of mention (and pronominalization, as a secondary measure) corresponds to greater
prominence. The two experiments produce divergent results. When possessives are compared to
indefinites, I observe an interaction between animacy and possession whereby possessed
animates (e.g. her butler, his nurse) are mentioned especially often as the subjects of
continuations. This result suggests an extra boost in prominence for possessed animates;
however, when possessives are compared to definites (e.g. The carpenter judged {the, his} nurse
because…), I observe a main effect of possession whereby the rate of subjecthood is lower for
possessed referents, suggesting an overall decrease in prominence. While Experiment 1 supports
a theory which differentiates the discourse-level effect of possession depending on the type of
possession relation, Experiment 2 provides evidence for the Parasitic Possession Hypothesis.
The sentence continuation paradigm of the first two experiments relies on both
comprehension of the prompt and production of a continuation, but the dependent measures—
12
rates of mention and pronominalization—come from the production side of the Bayesian
formalism proposed by Kehler et al. (2008): P(referent) and P(pronoun|referent). In Chapters 4
and 5, I shift to purely comprehension-based methods, in that my dependent measures relate to
how participants resolve pronouns. These chapters relate to the lefthand term in the Bayesian
equation, P(referent|pronoun). Chapter 4 reports a visual-world eye-tracking experiment
(Experiment 3) in which I examine gaze patterns during ambiguous pronoun resolution. I
compare how participants resolve a pronoun when one of its potential antecedents is possessed
versus definite, using sentences like The princess saw {the, her} attorney through the glass door
because she was crompety (nonce adjectives were used in Experiment 3). The results show that
possession renders the possessed referent less accessible as an antecedent for the pronoun, which
I interpret as evidence that it is less prominent in discourse compared to a simple definite.
Although Experiments 2 and 3 use quite different methods and dependent measures (representing
different sides of Kehler et al.’s Bayesian equation), their results converge in favor of the
Parasitic Possession Hypothesis.
While Chapter 5 still deals with the interpretation of pronouns, I shift to a different
theoretical and methodological domain to test the effect of possession type on the resolution of
ambiguous possessive pronouns in verb-phrase ellipsis (VPE, e.g. Bill washed his car, and John
did, too). While VPE has historically been analyzed more at the syntax-semantics level (e.g.
Fiengo & May, 1994; Sag, 1976), discourse-based approaches also exist (e.g. Asher et al., 2001;
Hardt & Romero, 2004; Kertz, 2013). While much of the literature on VPE (and generally in
theoretical syntax and semantics) has abstracted away from different possession relations in
examinations of possessive pronouns, I find that the type of possession relation (e.g. ownership
in his skateboard versus kinship in his mother) plays a major role in how participants prefer to
13
interpret ambiguous VP ellipsis. I explore the applicability of the discourse-based theory
developed in earlier chapters to these results.

14
Chapter 2: Testing how possession affects discourse prominence using a sentence
continuation task

2.1. Introduction
A widely accepted view about discourse-level representation and processing is that
referents in a discourse vary in prominence and that the prominence of referents changes over
time as the discourse continues (e.g. Ariel, 1988; Chafe, 1976; Clark & Sengul, 1979; Givón,
1983; van den Broek et al., 1996). Many factors contribute to a referent’s prominence, but two of
the most influential and dependable predictors are grammatical role and animacy; specifically,
subjects tend to be more prominent than objects (e.g. Chafe, 1976; Crawley et al., 1990), and
animates tend to be more prominent than inanimates (e.g. Bock et al., 1992; Dahl & Fraurud,
1996). However, as I discussed in Chapter 1, prior work has tended to overlook a frequent
grammatical structure with potential to inform current theories of discourse: nominal possessive
constructions (e.g. Sam’s car, Sam’s attorney). Unlike simple nominals (a car, the attorney),
nominal possessives mention an additional referent, a possessor (Sam), as well as the head noun
(car/attorney).
Most previous work investigating the discourse prominence of referents has focused on
simple nominals, which mention just a single referent. How then are the two referents mentioned
in possessives represented in discourse? Can we treat them like two independent discourse
referents and model their prominence based on previously identified discourse factors? Or does
participation in a possessive construction modify prominence in a way which is not captured by
existing discourse theories? Furthermore, do different semantic possession relations (e.g. the
ownership relation in Sam’s car versus the interpersonal relation in Sam’s attorney) affect
15
prominence in different ways? This chapter takes a first step toward addressing these questions
with a sentence continuation experiment (Experiment 1), which tests whether and how
possessives differ from simple indefinite noun phrases on the discourse level and compares
ownership and interpersonal possession relations. Before describing Experiment 1, I outline
some competing hypotheses about prominence in possessives based on prior work. I then present
the methods and results of Experiment 1 and discuss its implications for my hypotheses about
possession.

2.2. Should possession decrease or increase a referent’s prominence?
Setting aside the null hypothesis that possession does not affect a referent’s prominence, I
consider the logical alternatives that possession either increases or decreases a referent’s
prominence. For example, is the referent for car more or less prominent in Sam’s car versus a
car? For now, I present this question in terms of possession’s effect on the possessed referent,
rather than the possessor, since this minimal pair provides an easier connection to prior work on
simple nominals; however, I will also consider the effect on the possessor in the discussion of
this chapter. Accordingly, prior work provides reasons to entertain both hypotheses: that
possessed referents are more prominent than their simple nominal counterparts and vice versa.
As I mention in Chapter 1, evidence that possession might boost the possessed referent’s
prominence comes from the psycholinguistic literature on informativity and representational
complexity. It has been shown that more complex noun phrases, which encode more information,
are generally easier to retrieve from memory and more prominent in discourse (e.g. Fisher &
Craik, 1980; Hofmeister, 2011; Karimi et al., 2014; Karimi & Ferreira, 2016; Troyer et al.,
2016). While this work has not specifically addressed possession as a source of complexity,
16
nominal possessives like Sam’s car can be viewed as more representationally complex than
simple nominals because they mention an additional referent as the possessor and encode a
semantic possession relation which is not present in simple nominals. One might therefore
suspect that possessed referents would be more prominent than simple nominals in discourse; I
refer to this alternative as the Complexity Hypothesis.
On the other hand, prior work within the framework of Centering Theory (e.g. Gordon et
al., 1993; Grosz et al., 1995; Walker et al., 1998) provides some evidence that possession might
decrease the prominence of the possessed referent. Centering Theory is a model of discourse
coherence which seeks to capture patterns in transitions from one utterance to another (e.g. what
makes a particular sequence of utterances more or less coherent). Centering Theory is
particularly relevant to my research aims because a key component of its algorithm is ranking
referents according to their prominence. Given the frequency of nominal possessive
constructions and the fact that most work within Centering Theory has been corpus-based, it is
not surprising that this framework has had to address nominal possessive constructions.
Within Centering Theory, some researchers have argued for a prominence ranking in
which possessors in s-genitives are always more prominent than their corresponding possessed
referents (e.g. in Sam’s car, the referent for Sam would outrank the referent for car). This
ranking was first proposed as the Complex NP Assumption in Walker & Prince (1996), which
posits that the referents realized within complex noun phrases—of which possessives are one
type—are ranked from most to least prominent based on their left-to-right order. In subsequent
work, Chae (2003) found support for this view in a corpus analysis which specifically examined
s-genitive possessives. One can extrapolate based on this ranking that possession decreases
prominence by considering a hypothetical minimal pair of sentences such as The car is in the
17
garage and Her car is in the garage. In the first sentence, the referent for car is likely the most
prominent by virtue of its subjecthood; however, in the second sentence, the antecedent of the
possessive pronoun outranks the referent for car. Thus the net effect of possession would be to
decrease the prominence of the possessed referent relative to a non-possessed simple nominal; as
I mention in Chapter 1, I refer to this view as the Parasitic Possession Hypothesis. Its prediction
that possession decreases the prominence of the possessed referent is directly opposed to the
prediction of the Complexity Hypothesis, which predicts an increase in prominence. Notably, the
corpus-based accounts within Centering Theory do not compare precise minimal pairs of
possessed and non-possessed referents, as an experimental psycholinguistic approach would.
Thus, Experiment 1 builds on this work by directly comparing possessives and simple nominals.

2.3. Does the type of possession relation affect prominence?
An additional dimension to the two aforementioned hypotheses is that possession’s effect
on the prominence of possessed referents might depend on the type of semantic possession
relation. Prior work on possessives has noted that a wide range of possession relations can be
expressed in English using the same morphosyntax, such as ownership (e.g. Sam’s car), part-
whole (e.g. Sam’s face), kinship (e.g. Sam’s uncle), and non-kinship interpersonal relations (e.g.
Sam’s doctor), among others (e.g. Haspelmath, 2017; Langacker, 1993; 1995). It does not seem
unreasonable, therefore, to consider the possibility that these different possession relations may
affect prominence in different ways. Moreover, such effects are not mutually exclusive with any
potential overall effect of possession (i.e. the Complexity Hypothesis or Parasitic Possession
Hypothesis).
18
Prior work provides some arguments in favor of a theory in which the type of possession
relation modulates possession’s effect on prominence. Within Centering Theory, there is a
proposal by Di Eugenio (1998) which provides an alternative view to Walker & Prince (1996)
and Chae (2003). Instead of always ranking possessors above possessed referents, Di Eugenio
proposes that the ranking depends on the animacy of the possession. Animate possessed referents
immediately outrank their possessors, just as Walker & Prince and Chae argue; however,
inanimate possessed referents are ranked immediately below their possessors. Accordingly, in
Sam’s car, the referent for Sam would still outrank the referent for car, but in Sam’s doctor, the
referent for doctor would outrank the referent for Sam. This account dovetails with the well-
documented cross-linguistic finding in the broader discourse literature that animate referents tend
to be more prominent than inanimate ones (e.g. Bock et al., 1992; Dahl & Fraurud, 1996; Dahl,
2008). For instance, animate referents tend to appear before inanimates and in subject position
(e.g. Branigan et al., 2008; Prat-Sala & Branigan, 2000). Additionally, animacy influences the
linguistic form with which a referent is mentioned, with animates being overall more likely to be
pronominalized than inanimates (e.g. Dahl & Fraurud, 1996; Fukumura & van Gompel, 2011;
Yamamoto, 1999).
One might also hypothesize that possession’s effect on prominence depends on the type
of possession relation because certain relations are intrinsically more salient than others.
Specifically, when the possessor and possession in a possessive construction are both animate
and human, the construction explicitly denotes an interpersonal relationship. Such relationships
may be permanent (e.g. Sam’s mother, the teacher’s son) or more transitory (e.g. their neighbor,
the accountant’s boss) and may vary with respect to the formality of the relationship’s definition
(e.g. kinship relations tend to be unambiguous, but people may disagree on who counts as a
19
friend). Nevertheless, in all possessives where both the possessor and possessed referent are
human, the two individuals stand in some kind of socially salient relation to each other (i.e. an
interpersonal relationship).
Prior work provides evidence that interpersonal relationships are more cognitively salient
than other types of relations expressed in possessives. In non-linguistic research, interpersonal
relationships have been shown to be critical for human health and well-being (e.g. Cacioppo &
Hawkley, 2009; Eisenberger & Cole, 2012). For instance, social connectedness has been linked
to longevity and disease resistance (e.g. Holt-Lunstad et al., 2010; Miller et al., 2009), while
loneliness is correlated with depression and cognitive decline (e.g. Cacioppo et al., 2006; Tilvis
et al., 2004). In short, this literature makes clear that people with stronger connections to others
tend to live longer, happier lives, while socially isolated individuals tend to suffer a variety of
negative consequences. Furthermore, prior research on memory shows that people tend to form
stronger memory representations for animate entities than for inanimates (e.g. Bonin et al., 2014;
Nairne et al., 2013). This observation has led some researchers to theorize that that better
memory for animates arose from the evolutionary importance of identifying threats, mates, and
social groups, most of which entail interpersonal relationships (e.g. Nairne, 2010; VanArsdall et
al. 2013).
Based on these findings, human cognition appears to be especially attuned to
interpersonal relationships. Therefore, one might expect possessives that express interpersonal
relationships to also be privileged in our mental representations (relative to other kinds of
possession relations). If this is true, then it is not unreasonable to predict that the referent for
doctor in Sam’s doctor would be more prominent than in the simple nominal the doctor or a
doctor. Furthermore, the increase in prominence to the possessed referent when it participates in
20
an interpersonal relation could be greater than the effect of possession on other types of
possessed referents (e.g. Sam’s car).
There are many possible types of possession relations one could compare to interpersonal
relations in order to explore the theory laid out above. As a starting point, this dissertation
compares non-kinship interpersonal relations (e.g. her electrician, his nurse) to ownership
relations (e.g. her chandelier, his jacket). These two classes are well suited to the needs of
Experiment 1, since they are natural in both possessives and in simple nominals. On the other
hand, the distribution of non-possessed kinship terms or body parts can be more restricted (e.g.
The children love {their/?an} uncle, Sam hurt {her/?the} finger).
Accordingly, in addition to the Complexity Hypothesis and Parasitic Possession
Hypothesis described in Section 2.2, I consider two supplementary theories. The first is that, as
discussed above in this section, the general cognitive privilege of interpersonal relations will
provide a superadditive boost in prominence to possessed referents like electrician in her
electrician (i.e. in excess of any overall effect on prominence as predicted by the Complexity
Hypothesis or Parasitic Possession Hypothesis). I will refer to this theory as the Interaction
Hypothesis, since it relies on the presence of a statistical interaction between possession (i.e.
possessive versus simple nominal) and animacy (i.e. human role nouns, which denote
interpersonal relations when possessed, versus inanimates, which denote ownership when
possessed). Notably, this theory entails the interaction of two distinct types of discourse-relevant
features: animacy, which is a lexical property, and possession, which is a relational property
between referents. The fact that these two domains may interact at the discourse level is
potentially informative for theories about prominence in general and what features contribute to
it.
21
On the other hand, it might be that no such interaction exists, and either the Complexity
Hypothesis or the Parasitic Possession Hypothesis is sufficient on its own to describe the
discourse prominence of possessed referents. In this case, extensive prior work demonstrating
animates’ overall advantage in discourse prominence suggests that referents for animates like
electrician would be more prominent than those of inanimates like chandelier—irrespective of
whether they are possessed or not; any increase or reduction in prominence for possessed
referents would then apply to animates and inanimates equally. I refer to this possibility as the
Animacy Hypothesis, since it predicts only a statistical main effect of animacy, with no
interaction between animacy and possession.

2.4. Experiment 1
Experiment 1 tests the two sets of competing hypotheses laid out in Sections 2.2 and 2.3.
On one hand, Experiment 1 looks for evidence that possession either increases or decreases the
prominence of possessed referents, which would respectively support either the Complexity
Hypothesis or Parasitic Possession Hypothesis. These two hypotheses address the overall effect
of possession. On the other hand, Experiment 1 also assesses whether the type of semantic
relation between the possessor and the possessed referent modulates the discourse-level effect of
possession (i.e. the Interaction Hypothesis versus the Animacy Hypothesis). As a starting point,
Experiment 1 compares two types of possession relations: interpersonal and ownership. Based on
existing research on Centering Theory and social psychology, I formulated the Interaction
Hypothesis, which predicts a superadditive boost in prominence for possessed animate referents.
Alternatively, the difference between possessed animates and possessed inanimates may be the
same as between their non-possessed counterparts. This pattern would support the Animacy
22
Hypothesis, which posits that possessives exhibit animacy effects similar to those which have
canonically been observed in simple nominals.
To test these hypotheses, I use a sentence continuation task to estimate the prominence of
competing referents. This task builds on the common assumption that a referent’s prominence is
positively correlated with its likelihood of subsequent mention in the discourse (e.g. Arnold,
2001; Givón, 1983; Kaiser, 2009; Kehler et al., 2008; Kehler & Rohde, 2013). Furthermore,
existing discourse theory suggests that the most prominent referents tend to appear as
grammatical subjects (e.g. Chafe, 1976; Gordon et al., 1993; Stevenson et al., 1994). Thus we
can measure which referents are most prominent by analyzing how often they are selected as the
subjects of continuations (e.g. Ariel, 1988; Arnold, 2001; Givón, 1983; Stevenson et al., 1994). I
also analyze mentions of entities throughout the continuation sentence as a further measure of
discourse prominence, as well as looking at the linguistic form of mentions as a measure of
entities’ conceptual accessibility (e.g. Gundel et al., 1993; Kaiser, 2009).

2.4.1. Participants
Data from 40 native English speakers are presented for analysis. All participants were 18
years of age or older and were recruited from Amazon Mechanical Turk. There were 55 total
participants who completed the experiment. I excluded 4 participants for failing attention checks
meant to detect automated responses, 2 who either used copy-paste to duplicate all or portions of
the prompt sentence or their previous responses, and 6 whose responses were consistently
nonsensical, unrelated to the prompt, ungrammatical, or fragments. Finally, the last respondents
to finish the experiment on three of the experimental lists (3 of 55) were excluded in order to
balance the number of analyzable respondents on each list. All of the participants included in the
23
analysis reported being born in the United States and identified as native speakers of English. All
reported normal or corrected-to-normal vision and normal hearing.

2.4.2. Design and materials
Experiment 1 had 56 items: 24 targets and 32 fillers. All target items followed the frame:
[name] [nonce verb in simple past tense] {a, his, her} {animate, inanimate}, e.g. Jessica rulked
an electrician. (see examples in Table 1).

Indefinite
Animate
Possessed
Animate
Indefinite
Inanimate
Possessed
Inanimate
F … M
Gender Bias
Jessica rulked
an electrician.
Jessica rulked
her electrician.
Jessica rulked
a chandelier.
Jessica rulked
her chandelier.
M … F
Gender Bias
Daniel zatted
a nurse.
Daniel zatted
his nurse.
Daniel zatted
a jacket.
Daniel zatted
his jacket.
Table 1. Example target items (2 of 24) in Experiment 1 in each of the four
conditions

I manipulated (i) whether the object was possessed or indefinite and (ii) whether the direct object
was animate or inanimate. Animate objects were all human role nouns (e.g. stockbroker, florist);
inanimate objects were all concrete alienable possessions (e.g. umbrella, jacket). Both the human
role nouns and animate objects that I used were natural in possessive constructions. To minimize
the potential for pronoun ambiguity in participants’ continuations, the names (i.e. the prompt
subjects) were unambiguous with respect to their typically associated gender (e.g. Jessica,
Daniel) and matched for each item with human role noun objects which were stereotypically
biased toward the opposite gender. Norms for role nouns’ stereotypical gender came from
Misersky et al. (2014). As some role nouns can sound unnatural when possessed (e.g. ?her
firefighter), especially with limited context, roles were checked so that they were felicitous in
24
both the indefinite and possessed conditions. The order of the genders in target items with
animate objects (as well as the gender of the grammatical subject overall) was counterbalanced.
For each target item, animate objects were paired with inanimates that matched as closely
as possible in lexical frequency in the SUBTLEXus corpus (Brysbaert & New, 2009), syllable
length, and character length. Nonce verbs (e.g. tammed, chabbed, blorned) were used because
they allowed the verb to remain constant within items. Predicates which naturally take animate
objects are often unnatural with inanimate objects (and vice versa) and thus using real verbs
while keeping the verb constant across conditions is highly constraining. The nonce verbs also
limited potential effects of verbal semantics on the continuations, for instance, from implicit
causality (e.g. Hartshorne & Snedeker, 2013) or distributional biases toward animate or
inanimate objects.
Filler items were evenly divided into two categories: those beginning with evaluative
adverbs (e.g. Incredibly, some sneevs noticed Richard.) and those containing a syntactic sluice or
sprout (e.g. The band played some shows, but I don’t know where.). These filler types were
designed to be somewhat repetitive in order to distract from the potentially repetitive nature of
the targets. Nonce words appeared in 24 of the 32 fillers and varied in their syntactic position
(i.e. subject, object, or verb).
The 2 × 2 (possessive/indefinite × animate/inanimate) design resulted in four conditions.
Conditions were distributed across four experimental lists in a Latin square (i.e. within-subject
and within-item design). Thus, each participant saw six items per condition. Targets and fillers
on lists were interleaved in a pseudorandomized fashion. Four additional lists were created by
reversing the original lists, for a total of eight experimental lists.

25
2.4.3. Procedure
Before beginning the study, participants were told that they would read prompt sentences
and write one-sentence continuations. Participants were instructed to make their continuations
natural-sounding, not to copy-paste material from previous continuations or the prompts, and to
limit their responses to a single complete sentence for each item. Participants completed three
practice items (without any possessive structures) and saw samples of acceptable continuations
and unacceptable fragments. Other than the instructions against fragments and copy-paste,
participants were assured multiple times that there were no right or wrong answers. Each
experimental item was presented on a separate page (see Figure 1).

Figure 1. Example item from Experiment 1 as it appeared to participants

2.4.4. Pre-processing of the continuation data
In order to analyze participants’ continuations, they were annotated by hand with respect
to whether they included mentions of the preceding subject and/or object. Additionally, I
annotated whether the head noun of the first independent-clause subject noun phrase in each
continuation referred to the preceding subject, preceding object, or something else (or whether it
was unclear). For example, if the prompt was Jennifer pranned her surgeon, and the participant
continued this with She asked him how long her surgery would last, the continuation would be
coded as (i) mentioning the preceding subject as subject of the continuation and (ii) mentioning
26
both the preceding subject and preceding object (for the purposes of the entire-continuation
analyses). Fragment responses were noted and later omitted from the analysis (only 4 of the 960
target responses from included participants).
A word was only counted as a mention of an entity if the head of the noun phrase referred
to that entity; for example, for the prompt Daniel zatted his jacket, an instance of his jacket in the
continuation would only count as a mention of the preceding object (jacket) and not the
preceding subject (Daniel). I analyzed the continuation in this way because, in a strict semantic
sense, the noun phrase his jacket only refers to the jacket (although it mentions both the
possessor and the possessed referent); however, it should be noted that this choice cannot
influence the subject-position analysis, since only the head noun of the subject noun phrase
(which can never be the possessor) was counted for this analysis. Furthermore, when I did try
including possessors as mentions in the entire-continuation analysis, there were no consequential
changes to the results.
Additionally, all mentions of entities from the prompt sentence were coded with respect
to their linguistic form (e.g. Daniel vs. he). For the purposes of this analysis, I divide form of
mention into pronouns and all other forms. A second coder double-coded a randomly selected
subset of 10% of participants’ responses. The first and second coder agreed on 97% of double-
coded trials.

2.4.5. Predictions
The two sets of competing hypotheses introduced in Sections 2.2 and 2.3 make different
predictions about the results. Here I frame these predictions in relation to mentions of the
preceding object, since this referent is the locus of the experimental manipulation. In the results,
27
I discuss two primary dependent variables, a referent’s likelihood of mention in subject position
of the continuation and its likelihood of mention anywhere in the continuation (including subject
position, as any grammatical or thematic role). Notably, the predictions I present here are
broadly relevant for both the subject-position and entire-continuation analyses. I consider subject
position to reflect a “winner take all” measure of prominence, under the widely held assumption
that the most prominent referent in an utterance tends to be realized as the grammatical subject.
On the other hand, I expect that mentions across the entire continuation will be a correlated but
more inclusive measure that may pick up on finer differences in prominence.
As discussed in Section 2.2, the two hypotheses which pertain to the overall (i.e. main)
effect of possession are the Complexity Hypothesis and the Parasitic Possession Hypothesis. The
Complexity Hypothesis states that possessed referents are more prominent than simple noun
phrases (indefinites, in this case) due to their representational and semantic complexity.
Accordingly, it predicts that preceding objects will be more likely to be mentioned when they are
possessed compared to indefinite. Conversely, the Parasitic Possession Hypothesis states that
possessed referents are less prominent than simple nominals, perhaps because their link to the
possessor’s referent makes them less accessible or in some way dependent on the possessor.
Thus, under the Parasitic Possession Hypothesis, preceding objects would be mentioned less
often when they are possessed. Crucially, both hypotheses predict a main effect of possession in
the statistical models but are differentiated by the direction of this effect.
In addition to a possible main effect of possession, Section 2.3 considered whether
different possession relations would modulate the prominence of possessed referents. The first
possibility discussed was the Interaction Hypothesis, which theorizes that animate possessed
referents are exceptionally prominent in discourse due to their denotation of interpersonal
28
relationships. Therefore, one would expect possessed animates to receive an extra boost in
likelihood of mention (i.e. a statistical interaction of animacy and possession). However, if no
interaction is present, such a pattern would serve as evidence for the Animacy Hypothesis. This
theory states that, if animate possessed referents are more prominent than inanimate ones, it is
due to the same well-known effects which make animate simple nominals more prominent than
inanimate ones. Thus the difference in likelihood of mention between animate possessed
referents (e.g. his nurse) and inanimate ones (e.g. his jacket) would be comparable to the
difference between their indefinite counterparts (e.g. a nurse vs. a jacket).
Implicit in the Animacy Hypothesis is the prediction that animate preceding objects will
have an overall higher likelihood of mention than inanimates (i.e. a main effect of animacy). This
pattern seems especially likely to surface in the subject position of continuations, as prior work
shows that animates are canonically preferred as subjects (e.g. Bock et al., 1992, Branigan et al.,
2008; Dahl, 2008; Dahl & Fraurud 1996; Prat-Sala & Branigan, 2000). Therefore, participants
should tend to choose animate subjects for their continuations, and, when they do choose to
mention the preceding object in subject position, they should be more likely to do so when the
preceding object is animate (regardless of possession status). Notably, Experiment 1 could
potentially find that preceding objects are most prominent when they are animate and possessed,
but this pattern could arise from the combined effects predicted by the Complexity Hypothesis
and the Animacy Hypothesis. Only a superadditive effect of animacy and possession together
serves as evidence for the Interaction Hypothesis.
Finally, regarding the relative prominence of the preceding subject and object, prior work
shows that preceding subjects are privileged and—with other factors held constant—likely to be
mentioned subsequently in subject position (e.g. Arnold et al., 2000; Gordon et al., 1993;
29
Stevenson et al., 1994). In the inanimate conditions of this experiment, I therefore predict that
preceding subjects, which are always animate, should have a significant advantage in
prominence over preceding inanimate objects; however, predictions are less clear for the animate
conditions. In these conditions, the preceding subjects (i.e. the proper names) will be in
competition for subjecthood with animate objects (i.e. the role nouns). The advantage of
preceding subjects should boost the named referents’ prominence, but the role nouns are
semantically richer than the names and may compete closely with them due to their greater
informativity (e.g. Hofmeister, 2011; Karimi et al., 2014; Karimi & Ferreira, 2016; Troyer et al.,
2016).

2.4.6. Results
I analyzed mentions of the prompt subject and object in continuation-subject position and
across the entire continuation. Accordingly, there are four main analyses corresponding to the
mentioned referent and its position. Pronominalization patterns are also considered at the end of
this section. All statistical analyses for this experiment used generalized linear mixed-effects
models implemented with the R package lme4 (Bates et al., 2015; R Core Team, 2020). All the
models were fit to the binomial outcome of mentioning the relevant entity in a particular position
(outcome = 1) or not (outcome = 0). The independent variables animacy and possession were
deviation coded (animate = 0.5, inanimate = -0.5; possessive = 0.5, indefinite = -0.5). The
models reported here have the maximal random effects structure that did not result in
nonconvergence (Barr et al., 2013). When a model’s random effects structure needed to be
reduced due to nonconvergence, the inclusion of random slopes for participants was prioritized.

30
2.4.6.1 Mentions of the prompt subject as continuation subject
I first analyzed the subject position of continuations to see whether participants
mentioned the preceding subject, object, or a third party not mentioned in the prompt (19
responses with unclear subjects were also counted in the third-party category). A visualization of
these data is given in Figure 2.

Figure 2. The proportion of continuations in Experiment 1 referring back to the
prompt subject, prompt object, or something else (error bars represent ±1 standard
error)

I fit two models which predicted the probability of the continuation subject referring (with any
type of referential form) to the preceding subject or object. Accordingly in the model for subject
mentions, trials had the binomial outcome of 1 when the prompt subject was mentioned as the
subject of the continuation and 0 for all other trials (excluding fragments); likewise, in the model
0.0
0.1
0.2
0.3
0.4
0.5
0.6
Indefinite
Animate
Possessed
Animate
Indefinite
Inanimate
Possessed
Inanimate
Proportion of continuations
Continuation subject
refers to preceding...
subject
object
other
31
for object mentions, trials had an outcome of 1 when the prompt object was mentioned as the
subject of the continuation and 0 for all other trials.
The model for the preceding subject revealed a main effect of animacy, whereby the
presence of an animate object in the prompt sentence significantly reduced the chances of the
continuation subject referring to the prompt subject (i.e. the proper name) (see Table 2 for a
summary of the model’s fixed effects). This animacy effect seems to reflect increased
competition for prominence between the preceding subject and object when the object is
animate.

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) -0.569 0.210 -2.713 0.007
animacy -0.705 0.224 -3.144 0.002
possession -0.154 0.152 -1.011 0.312
animacy:possession -0.125 0.305 -0.408 0.683
Table 2. Summary of fixed effects for the model predicting mentions of the
preceding subject in subject position of the continuation (Experiment 1)

To further test the competition between the two animate entities in the conditions where
the object was animate, I fit a model which predicted whether continuations in the animate
conditions would refer to either the preceding subject or object (ignoring the possibility of a
third-party subject). This model had a significant intercept (β = 0.640, SE = 0.248, z = 2.584, p =
0.010), signifying that in the animate conditions, the preceding object was more likely than the
preceding subject to be the referent of the continuation’s subject. As I mentioned in Section
2.4.5, this pattern may be due the differences in semantic content between the proper name
subjects and role noun objects. With all other factors held equal, one would expect the preceding
subject to be the preferred subject of the continuation; however, the role nouns encode more
32
information than the proper names, and their greater informativity may give them an advantage
in prominence.

2.4.6.2 Mentions of the prompt object as continuation subject
Having looked at continuations where the participant chose to start by referring to the
preceding subject, I now turn to continuations where the participant starts by mentioning the
preceding object, i.e. the model for mentions of the preceding object as subject of the
continuation. This model is the most relevant to the predictions in Section 2.4.5 and the
associated hypotheses. Like the model for mentions of the preceding subject, the object model
also showed a significant main effect of animacy. This finding illustrates that animate objects
had a greater chance of being mentioned in subject position compared to inanimate objects (see
Table 3 for a summary of the model’s fixed effects).

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) -0.340 0.234 -1.457 0.145
animacy 1.045 0.249 4.191 <0.001
possession 0.160 0.157 1.018 0.309
animacy:possession 0.628 0.314 2.000 0.046
Table 3. Summary of fixed effects for the model predicting mentions of the
preceding object in subject position of the continuation (Experiment 1)

This pattern reflects the well-known advantage in prominence that animacy confers. The model
did not show a main effect of possession and thus provides evidence for neither the Complexity
Hypothesis nor the Parasitic Possession Hypothesis; however, there was a significant interaction
between animacy and possession, which reflects a boost in the likelihood of mention for animate
possessions. To further examine this interaction, I fit a model to test for a simple effect of
33
possession within the animate conditions; this analysis revealed a significant simple effect of
possession (see Table 4).

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) 0.175 0.218 0.805 0.421
possession 0.47 0.211 2.228 0.026
Table 4. Summary of fixed effects for the model testing the simple effect of
possession within the animate conditions (subject-position analysis, Experiment
1)

However, when the inanimate conditions were analyzed with the same sort of follow-up model,
there was no significant simple effect of possession (see Table 5).

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) -0.878 0.313 -2.807 0.005
possession -0.277 0.254 -1.093 0.274
Table 5. Summary of fixed effects for the model testing the simple effect of
possession within the inanimate conditions (subject-position analysis, Experiment
1)

These follow-up models show that the difference between the animate possessed and animate
indefinite conditions is responsible for the significant interaction in the larger model. Therefore,
the results illustrate that, when the prompt object was both animate and possessed, it had an
especially increased chance of being the subject of the continuation. This pattern supports the
Interaction Hypothesis, which predicts a superadditive boost in prominence for possessed
animates.

2.4.6.3 Mentions of the prompt subject anywhere in continuations
I turn now to the analysis of mentions in all positions of the continuation. A visualization
of mentions anywhere in the continuation (with any referential form) is presented as Figure 3.
34
The analysis proceeded in much the same way as the subject-position analysis; first I fit two
models to predict the likelihood of a continuation mentioning the subject or object in any
position.

Figure 3. The proportion of continuations in Experiment 1 which mention the
prompt subject and/or prompt object (error bars represent ±1 standard error)

The model for the preceding subject revealed no significant main effects or interaction (see Table
6 for a summary of the model’s fixed effects). It is noteworthy that this model contrasts with the
subject-position model, which showed an animacy effect. While the animate conditions showed a
lower likelihood of the prompt subject being mentioned in subject position, by the end of the
continuation, all conditions had a similar likelihood of mentioning the prompt subject.

0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Indefinite
Animate
Possessed
Animate
Indefinite
Inanimate
Possessed
Inanimate
Proportion of continuations
Continuation contains
mention of preceding...
subject
object
35
Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) 0.119 0.181 0.659 0.510
animacy -0.025 0.209 -0.121 0.903
possession -0.074 0.144 -0.510 0.610
animacy:possession 0.152 0.289 0.528 0.598
Table 6. Summary of fixed effects for the model predicting mentions of the
preceding subject anywhere in the continuation (Experiment 1)

2.4.6.4 Mentions of the prompt object anywhere in continuations
When the probability of the preceding object being mentioned anywhere in the
continuation was analyzed, the results were similar to the corresponding model for subject
position (see Table 7).

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) 0.740 0.216 3.419 <0.001
animacy 0.277 0.153 1.810 0.070
possession 0.404 0.158 2.551 0.011
animacy:possession 0.890 0.306 2.905 0.004
Table 7. Summary of fixed effects for the model predicting mentions of the
preceding object anywhere in the continuation (Experiment 1)

The main effect of animacy was marginally significant, and the main effect of possession was
significant, but their interaction was also significant. I again sought to investigate this interaction
by testing for simple effects of possession within just the animate and inanimate conditions. Like
the corresponding subject-position analysis, these follow-up models revealed that possessed
animates were significantly more likely than their indefinite counterparts to be mentioned in
continuations (see Table 8); however, no such simple effect was present for the inanimate
conditions (see Table 9).

36
Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) 0.877 0.234 3.753 <0.001
possession 0.834 0.252 3.311 <0.001
Table 8. Summary of fixed effects for the model testing the simple effect of
possession within the animate conditions (entire-continuation analysis,
Experiment 1)

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) 0.643 0.256 2.513 0.012
possession -0.045 0.216 -0.210 0.834
Table 9. Summary of fixed effects for the model testing the simple effect of
possession within the inanimate conditions (entire-continuation analysis,
Experiment 1)

This interaction again reflects a superadditive boost in the likelihood of mention for possessed
animates and therefore lends further support to the Interaction Hypothesis. While the main effect
of possession in the larger model showed a significant increase in the likelihood of mention for
possessed objects—seemingly evidence for the Complexity Hypothesis—the fact that there was
no simple effect of possession for the inanimates casts doubt on whether there is a true overall
effect of possession here.

2.4.6.5 Pronominalization patterns
As an additional analysis, pronominalization patterns for the prompt subject and object
were analyzed both in subject position and across the entire continuation. Specifically, I looked
at the conditional probability that a pronoun would be used to refer to a particular referent (either
in subject position or the entire continuation) given that the referent was mentioned (in that
particular position). For example, for the analysis of the prompt subject in subject position, I
examined the conditional probability that the prompt subject would be referenced with a pronoun
in subject position given that the prompt subject was referenced with any form in subject
37
position. To analyze these patterns, the data were subsetted to just the continuations which
mentioned the relevant referent in the relevant position and a model was fit to the binomial
outcome of using a pronoun or not. The models were fit with the same predictors and according
to the same principles as the previously discussed models which estimate the simple probability
of mention. A visualization of the pronominalization data is given in Figure 4.

Figure 4. Pronominalization likelihood (conditional probability of pronoun given
mention) in Experiment 1 for the prompt subject and object in subject position
and all positions of the continuation (error bars represent ±1 standard error)

In the two models for subject position of the continuation, the prompt subject and object
exhibit complementary patterns (see Tables 10 and 11 for summaries of the model parameters).
The models for this position showed main effects of possession, such that possession of the
prompt object decreased the conditional probability of using a pronoun to refer to the subject and
increased the conditional probability of using a pronoun to refer to the object.
Subject of continuation Anywhere in continuation
Indefinite
Animate
Possessed
Animate
Indefinite
Inanimate
Possessed
Inanimate
Indefinite
Animate
Possessed
Animate
Indefinite
Inanimate
Possessed
Inanimate
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Conditional probability of pronoun given referent
Pronoun refers
to preceding...
subject
object
38

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) 11.765 2.926 4.021 <0.001
animacy -0.166 3.409 -0.049 0.961
possession -2.532 1.221 -2.073 0.038
animacy:possession -3.533 2.201 -1.605 0.109
Table 10. Summary of fixed effects for the model predicting pronominalization of
the prompt subject in subject position (Experiment 1)

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) -0.623 0.409 -1.523 0.128
animacy -0.634 0.326 -1.943 0.052
possession 0.624 0.288 2.163 0.031
animacy:possession -0.366 0.568 -0.644 0.520
Table 11. Summary of fixed effects for the model predicting pronominalization of
the prompt object in subject position (Experiment 1)

This pattern seems to support the Complexity Hypothesis, since a higher rate of
pronominalization is typically associated with increased accessibility and prominence (e.g.
Arnold, 2001; Fukumura & van Gompel, 2010; Gundel et al., 1993; Kaiser, 2011). Thus one can
interpret this effect as showing that possession increased the prominence of the possessed
referent and made it more accessible to pronominal reference.
The significant effects of possession did not persist across the entirety of the
continuation. In the pronominalization models for the subject and object in all
continuation positions, the effect of possession was marginally significant only in the
model for the prompt object, although it was in the same direction as the significant effect
in subject position (see Tables 12 and 13 for summaries of the model parameters).

39
Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) 3.370 0.601 5.612 <0.001
animacy -0.853 0.390 -2.189 0.029
possession 0.029 0.363 0.079 0.937
animacy:possession -0.191 0.728 -0.263 0.793
Table 12. Summary of fixed effects for the model predicting pronominalization of
the prompt subject anywhere in the continuation (Experiment 1)

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) -0.108 0.331 -0.325 0.745
animacy -0.499 0.310 -1.606 0.108
possession 0.382 0.213 1.794 0.073
animacy:possession 0.262 0.425 0.616 0.538
Table 13. Summary of fixed effects for the model predicting pronominalization of
the prompt object anywhere in the continuation (Experiment 1)

Additionally, in the model for pronominalization of the prompt subject in all continuation
positions, there was a significant effect of animacy, which showed that the presence of a second
animate referent (i.e. an animate prompt object) decreased the rate of pronominalization of the
prompt subject. This result is consistent with prior work showing that animate referents compete
closely for prominence and that the presence of multiple animates reduces rates of
pronominalization, even when the referents differ in grammatical gender (e.g. Arnold & Griffin,
2007; Chiriacescu, 2015; Fukumura & van Gompel, 2011).
Finally, it should be noted in the models for both subject position and the entire
continuation, that the overall rate of pronominalization for the prompt subject and the prompt
object are dramatically different. The highly significant positive intercepts in both the models for
the prompt subject indicate that the conditional probability of pronominalization for the subject
is significantly above 0.5. On the other hand, the non-significant intercepts in the models for the
prompt object indicate a lower rate of pronominalization. This difference is especially
noteworthy given that the rates of mention for these two referents were much closer. For
40
instance, as discussed in Section 2.4.6.2, the prompt object was actually more likely to be
mentioned in subject position of continuation than the prompt subject when the prompt object
was animate. This result supports a theory of pronominalization which dissociates the probability
of mention and the conditional probability of pronominalization, as in the Bayesian formalism of
Kehler and colleagues (e.g. Kehler et al., 2008; Kehler & Rohde, 2013).

2.5. Discussion
Experiment 1 investigates how possession affects referents’ discourse prominence. To
this end, I proposed two sets of competing hypotheses. The first set (the Complexity Hypothesis
and the Parasitic Possession Hypothesis) consider whether the overall effect of possession on the
possessed referent is to increase or decrease its prominence relative to a simple nominal. In
Experiment 1, the class of simple nominals against which I compared possessives was
indefinites. The second set of hypotheses (the Interaction Hypothesis and the Animacy
Hypothesis) ask whether the type of possession relation between the possessor and possessed
referent modulates possession’s overall effect on prominence. In Experiment 1, the two types of
possession relations I tested were ownership (e.g. her chandelier, his jacket) and interpersonal
relations (e.g. her electrician, his nurse). I tested these hypotheses using a sentence-continuation
task, in which I analyzed how likely participants were to mention referents from a prompt
sentence depending on the prompt object’s possession status and animacy.

2.5.1. Implications for the Complexity Hypothesis and Parasitic Possession Hypothesis
With regard to the first set of hypotheses, the results offer some support for the
Complexity Hypothesis but are not conclusive. In the primary analyses of Experiment 1—the
41
rates of mention in subject position and across the continuation—neither the Complexity
Hypothesis nor the Parasitic Possession Hypothesis was compellingly supported. The one
instance which hinted in favor of the Complexity Hypothesis was in the analysis of mentions of
the prompt object in all positions of the continuation. Here there was a significant main effect of
possession in the direction predicted by the Complexity Hypothesis; however, the tests for
simple main effects of possession in the animate and inanimate conditions suggested that the
main effect of possession was primarily driven by the animate conditions and may not be reliable
given the significant interaction of animacy and possession.
On the other hand, in the pronominalization patters, especially in subject position, the
significant main effects of possession provide reasonable evidence for the Complexity
Hypothesis. These effects show that when the prompt object was possessed and mentioned in
subject position of the continuation, it had a higher likelihood of being realized as a pronoun
compared to when it was indefinite in the prompt. This result suggests that possession increased
the prominence of the possessed referent. Nevertheless, some care needs to be taken in
interpreting this effect, since the pronominalization patterns were not intended to be the primary
dependent variable in this experiment. Specifically, the experiment was not designed to facilitate
this analysis, as it required rather aggressive subsetting of the results which resulted in a
relatively small dataset with imbalanced conditions. If one considers the intercepts in the models
for subject-position mentions, neither the prompt subject nor prompt object appeared as the
continuation subject at a rate significantly greater than 50% (see Tables 2 and 3). Consequently,
when modeling the conditional probability of pronominalization, the number of trials considered
is diminished by at least half compared to the analysis of simple mentions. Likewise, differences
among the conditions with regard to mentions create imbalances in the conditions when
42
modeling the rate of pronominalization. Because only the pronominalization patterns provided
support for the Complexity Hypothesis, and there was no such evidence in the analysis of
mentions, further research is needed before one can securely argue that possession increases the
prominence of the possessed referent.
It is somewhat unexpected not to find clear evidence in favor of either the Complexity
Hypothesis or the Parasitic Possession Hypothesis—at least with respect to the analysis of
mentions, which was the primary dependent variable in Experiment 1. It seems reasonable to
assume that the link between possessor and possessed referent must have some baseline effect on
discourse prominence, irrespective of exact nature of the possession relation. One reason the
results may not show such an effect is the use of indefinites as the comparison class for
possessives. Prior work has shown that possessed referents can be either discourse-new or
familiar (e.g. Barker, 2000; 2011), while indefinites typically introduce new discourse referents
(e.g. Gundel et al., 1993; Heim, 1982). Indefinites can also have a non-specific semantics which
does not pick out a particular discourse referent (e.g. e.g. Baker, 1966; Heim, 1982; Karttunen,
1968; Quine, 1956). Since both givenness and specificity are known to affect discourse
prominence (e.g. Chiriacescu & von Heusinger, 2010; Kaiser, 2011; von Heusinger, 2011), it is
possible that these factors could have masked potential overall differences between possessives
and indefinites in Experiment 1. It might be that choosing a different class of simple nominal to
compare with possessives (e.g. definites, as I test in Experiment 2) could result in a more
dependable baseline effect of possession. Therefore, at this point, I can neither endorse the
Complexity Hypothesis on the basis of the pronominalization data alone nor abandon the
Parasitic Possession Hypothesis for a lack of evidence.

43
2.5.2. Implications for the Interaction Hypothesis and Animacy Hypothesis
Regarding my second set of hypotheses, which relate to how different possession types
might modulate the overall effect of possession, there is convergent evidence from both the
subject-position and entire-continuation analyses in favor of the Interaction Hypothesis. In both
analyses, I found that participants were especially likely to mention possessed animates (in
excess of the combined independent effects of animacy and possession). Therefore, the results
indicate that that possessed animates get an exceptional boost in discourse prominence, as
predicted by the Interaction Hypothesis. I take these results as evidence that possessed animates’
special status in discourse may relate to the fact that they explicitly denote interpersonal
relationships. Prior work in the diverse research domains of health, social cognition, and memory
has shown that an individual’s relationships with other humans are of critical importance (e.g.
Cacioppo & Hawkley, 2009; Eisenberger & Cole, 2012). I suggest that the exceptional discourse
behavior of possessed animates observed in the current experiment could be linked to a more
general cognitive privilege for interpersonal relationships.
Even though the results of Experiment 1 seem to go against the Animacy Hypothesis,
which essentially predicted no interaction between animacy and possession, implied in the
Animacy Hypothesis was the prediction that there would be overall effects of animacy in line
with the significant amount of prior research on the discourse-level effects of animacy. This
aspect of the Animacy Hypothesis was robustly supported in the subject-position analysis. Here I
observed that animate prompt objects were much more likely to be promoted to the subject of the
continuation compared to inanimates. This finding is in line with existing work showing that
animates tend to be preferred as subjects (e.g. Bock et al., 1992, Branigan et al., 2008; Dahl,
2008; Dahl & Fraurud 1996; Prat-Sala & Branigan, 2000).
44
Additionally, I observed some notable animacy effects related to the relative prominence
of the prompt subject and object. It is generally accepted that the most prominent discourse
referent will appear as the subject, abstracting away from other influences that might affect the
choice of subject, such as topicality or coherence relation (Arnold, 2001; Givón, 1983; Grüter et
al., 2017, Stevenson et al., 1994). Accordingly, one can understand the subject-position patterns
in the current experiment to reflect a winner-take-all competition for discourse prominence. In
this light, it seems clear that animacy of the preceding object plays a substantial role in how that
competition for the subject position plays out; when the object is inanimate, the preceding
subject is the strongest candidate for subject position. This should not be surprising, since its
prominence should get a boost both from topicality as preceding subject and from its animacy;
however, when the preceding object is animate, the pattern flips, and the preceding object
becomes the preferred continuation subject. This pattern seems to reflect competition between
animate entities, which has previously been observed in similar tasks (e.g. Arnold & Griffin,
2007); however, it goes beyond normal competition, since the animate role nouns actually
surpass the proper names in prominence, despite the proper names occupying the more
privileged structural position in the prompt sentence.
One possible explanation for why this pattern arises is that the role nouns (in object
position) are semantically richer than the names (in subject position). When processing the
names, participants have few features beyond gender that can be reliably represented (although
some names may provide hints about age and other dimensions of identity). On the other hand,
the role nouns are quite representationally complex by comparison; in addition to the gender
biases we leveraged as part of the experimental design, the role nouns induce strong expectations
about visual appearance, the sorts of events in which the individual participates, the sorts of other
45
individuals with whom the role interacts, and societal status, among other features. These
intuitions fit well with previous studies showing that representational complexity and
informativity can increase a referent’s discourse prominence (e.g. Hofmeister, 2011; Karimi et
al., 2014; Karimi & Ferreira, 2016; Troyer et al., 2016). In the current experiment, the semantic
content of the role nouns may have given participants more features to explain or comment on in
their continuations. Additionally, prior work shows that animate objects are relatively marked
and less common than inanimate objects across a number of languages (Dahl, 2008; Dahl &
Fraurud 1996). Due to the markedness of animate objects, it is possible that participants in our
study found the animate objects especially salient and worthy of promotion to subject position.
Regarding prior work on the discourse behavior of possessives, the results of Experiment
1 address a debate within the Centering Theory literature mentioned in Section 2.2. Researchers
in this field have proposed two competing accounts for the relative prominence of possessors and
possessions in English s-genitives. The first is that possessors are always more prominent than
possessions (Chae, 2003; Walker & Prince, 1996). Alternatively, Di Eugenio (1998) has
proposed that the prominence ranking depends on the animacy of the possession. According to
this account, possessors are more prominent than inanimate possessed referents, but animate
possessed referents are more prominent than their possessors. As I noted earlier, this theory maps
neatly onto the Interaction Hypothesis, since it implies a boost in prominence for possessed
animates. The results of Experiment 1 suggest that animate and inanimate possessed referents
behave differently on the discourse level, with animate possessed referents receiving more
prominence than their animacy alone would predict. Therefore, our results are more compatible
with Di Eugenio’s account (and, consequently, the Interaction Hypothesis), which differentiates
46
the discourse behavior of animate and inanimate possessed referents and confers extra
prominence on possessed animates.

47
Chapter 3: Investigating effects of representational complexity on the discourse
representations of possessives

3.1. Introduction
Experiment 1 tested two sets of competing hypotheses about possession’s effect on
discourse prominence with a sentence continuation task. The first set of hypotheses concerned
the overall effect of possession on possessed referents, with the Complexity Hypothesis
predicting an increase in prominence and the Parasitic Possession Hypothesis predicting a
decrease. The results of Experiment 1 provided some evidence in favor of the Complexity
Hypothesis with regard to pronominalization patterns, although these analyses were not intended
as the primary measure of the experiment. Specifically, in the subject-position pronominalization
patterns, it was observed that possession increased the conditional probability that a reference to
the possessed object would be realized as a pronoun. This result provided some support to the
claim that possession increases the prominence of the possessed referent, perhaps by virtue of its
increased complexity and informational content. Nevertheless, the rates of mention examined in
Experiment 1, which formed the principal domain of analysis, did not provide clear evidence in
favor of either the Complexity Hypothesis or the Parasitic Possession Hypothesis.
On the other hand, the second set of hypotheses considered whether the effect of
possession varies with the type of possession relation. The results of Experiment 1 showed a
boost in prominence for the possessed referent when the possessive expressed an interpersonal
relation but not when it expressed ownership. Thus Experiment 1 provided support for the
Interaction Hypothesis over the Animacy Hypothesis.
48
The results of Experiment 1 bring up two primary points. Firstly, this work examines the
effect of possession relative to simple nominals. Experiment 1 compared possessives to
indefinites, but definites provide another valuable and arguably more direct comparison. Thus
Experiment 2 intends to provide a more comprehensive view of the discourse behavior of
possessives by comparing them to definites. Secondly, the results of Experiment 1 included some
effects which may have been driven by differences in representational complexity. Experiment 2
contains a more direct manipulation of representational complexity in order to explore this issue
further.

3.2. Comparing possessives, indefinites, and definites
While Experiment 1 compared possessives to indefinites, the discourse functions of these
two nominal classes align in some respects and diverge in others. Firstly, it is worth noting that
the same familiarity conditions and attentional status which license an indefinite may also license
the types of possessives examined in Experiment 1. If one considers the prompt sentence Daniel
zatted a jacket—for the sake of clarity, assigning a semantic interpretation like put on to the
nonce verb—the indefinite noun phrase would typically be understood to introduce a new
discourse referent previously unfamiliar to the comprehender (e.g. Du Bois, 1980; Heim, 1982;
Onea & Geist, 2011). Likewise, in the prompt with a possessive Daniel zatted his jacket, the
jacket can be discourse-new and unfamiliar to the comprehender as long as it is uniquely
identifiable in the context (e.g. Barker, 2000, 2011). Given the lack of context in these prompts,
there is no reason for the comprehender to imagine a situation in which the jacket is not uniquely
identifiable. Thus, both the indefinite and possessive noun phrases in Experiment 1 can both be
felicitous under the same discourse conditions.
49
On the other hand, there are situations in which the discourse properties of indefinites and
possessives diverge significantly. Notably, indefinites can yield non-specific interpretations in
certain contexts, which do not refer to any particular referent in the discourse (e.g. Baker, 1966;
Heim, 1982; Karttunen, 1968; Quine, 1956). Such an interpretation is possible if one understands
Daniel zatted a jacket to mean Daniel needed a jacket (with a continuation like, for example, It
was cold outside). In this case, the non-specificity of a jacket is clear based on the infelicity of a
continuation like #It was blue, in which the anaphor lacks a clear antecedent. In contrast, the
non-specific interpretation is not available with the possessive prompt. This difference has the
potential to influence the relative prominence of referents in the two conditions, since prior work
has shown that non-specific referents are less likely to be re-mentioned and therefore less
prominent (e.g. Chiriacescu & von Heusinger, 2010; von Heusinger, 2011). Furthermore, it has
also been noted that possessed referents can be either unfamiliar or familiar within the discourse,
unlike indefinites which are infelicitous with previously introduced referents (e.g. Barker, 2000,
2011; Gundel et al., 1993). For instance, his jacket can be familiar in a discourse like Daniel
wore a jacket and boots. His jacket kept him warm all day, but his boots were a bit tight. This
asymmetry in familiarity status could also affect the relative prominence of the possessives and
indefinites in Experiment 1 and, therefore, their rate and form of mention in continuations (e.g.
Bock & Irwin, 1980; Chafe, 1976; Gundel et al., 1993; MacWhinney & Bates, 1978).
Considering the discourse-level differences between indefinites and possessives (aside
from the presence or absence of a possession relation, which is the main factor of interest in this
dissertation) and the potential effects on prominence these differences produce, one can take an
alternative view of the results of Experiment 1. With regard to the pronominalization patterns,
which seemed to support the Complexity Hypothesis, it could be that the reduced prominence of
50
indefinites (by virtue of differences in givenness or specificity) contributed to indefinites being
pronominalized less often. Thus, an effect which seemed like a boost in prominence for
possessed referents could be alternatively understood as a reduction for indefinites. Likewise,
with regard to the lack of an overall effect of possession in the analysis of rates of mention, these
same differences in givenness and specificity could mask a reduction in the prominence for
possessed referents, which would have supported the Parasitic Possession Hypothesis.
The differences between indefinites and possessives mentioned here should not be taken
to mean that indefinites are a wholly inappropriate comparison to possessives, since the
discourse conditions which license them can overlap; however, definites are necessary as
additional comparison to achieve a comprehensive view of possessives’ discourse behavior. In
contrast to indefinites, definites do not exhibit the aforementioned specificity alternation, and are
licensed in similar contexts as possessed referents (i.e. either familiar or discourse-new and
uniquely identifiable) (e.g. Barker, 2000, 2011; Gundel et al., 1993). For these reasons,
Experiment 2 compares possessives to definite noun phrases.

3.3. The role of representational complexity
Experiment 1 produced the unexpected result that after prompt sentences with animate
direct objects (e.g. Jessica rulked her electrician), participants were more likely to select the
prompt object as the subject of their continuation sentence compared to the prompt subject,
which was also animate. Prior work suggests that the presence of multiple animate referents in
the discourse leads to competition for prominence (e.g. Arnold & Griffin, 2007; Chiriacescu,
2015; Fukumura & van Gompel, 2011). Thus, in the case of Experiment 1, I expected that the
presence of an animate object would lead to a lower rate of continuation subjecthood for the
51
prompt subject versus when the prompt object was inanimate. This pattern was revealed in the
results as predicted. However, it was unexpected that animate prompt objects not only lowered
the rate of subjecthood for prompt subjects but were even preferred over them as continuation
subjects. This pattern seems to go against the well-known advantage in prominence for preceding
subjects (e.g. Chafe, 1976; Crawley et al., 1990; Gordon et al., 1993). In the preceding chapter, I
suggested that this result could have been due to the animate objects’ (i.e. role nouns) greater
representational complexity in comparison to the subjects (i.e. names). Furthermore, the
Complexity Hypothesis also relies on the notions of representational complexity and
informativity. As I discuss in Chapters 1 and 2, this theory assumes that the semantic link
between the possessor and possessed referent increases the possessed referent’s informational
content and therefore increases its prominence. In this chapter, I further investigate effects of
representational complexity to assess the merit of these previous arguments.
As mentioned in Chapter 1, referents which encode more information generally tend to be
more salient in memory and discourse. Words with a greater number of associated features are
more easily retrieved from memory (e.g. Craik & Tulving, 1975; Fisher & Craik, 1980), and
antecedents with greater representational complexity are accessed more quickly in resolving
linguistic dependencies (e.g. Hofmeister 2011; Hofmeister & Vasishth, 2014; Troyer et al.,
2016). Additionally, more informative referents tend to be more likely pronominal antecedents
(Karimi & Ferreira, 2016; Karimi et al., 2014, 2019). Within this literature, one technique to
manipulate a referent’s associated features—and thus its representational complexity—is
adjectival modification (e.g. Hofmeister 2011; Hofmeister & Vasishth, 2014; Karimi et al.,
2019). By integrating adjectives into similar materials as used in Experiment 1, it is possible to
modify referents’ representational complexity with minimal alteration of the prompt. In this way,
52
I can compare the effect of explicit manipulation of complexity with any effect due to the
inherent complexity of possessives.

3.4. Experiment 2
Experiment 2 has two main aims: firstly to investigate further how possession affects
discourse prominence by comparing possessives to another class of simple nominals (i.e.
definites), and secondly to examine the role of representational complexity in the results of
Experiment 1 and its relevance to possessives in general. I approach these questions with a
second sentence continuation experiment. Chapters 1 and 2 discussed how possessives can be
considered more complex than simple nominals and proposed that this increased complexity
might result in greater discourse prominence for possessed referents (i.e. the Complexity
Hypothesis). Experiment 1 found some evidence for this theory, but it came from
pronominalization patterns, which were not the main analysis. With regard to the analyses of
referents’ rates of mention, there was no support for either the Complexity Hypothesis or the
competing account that possession reduces discourse prominence (i.e. the Parasitic Possession
Hypothesis).
Thus I see two main views emerging concerning the type of complexity inherent to
possessives by virtue of the semantic relation between possessor and possessed referent. Firstly,
it might be that the discourse properties of possessives’ complexity differ fundamentally from
other the types of informativity-boosting operations which have been examined in the literature,
such as modification with adjectives and relative clauses (e.g. Hofmeister 2011; Hofmeister &
Vasishth, 2014; Karimi & Ferreira, 2016; Karimi et al., 2014, 2019; Troyer et al., 2016). While
53
more explicit means of increasing a referent’s informational content may increase prominence,
the encoding of the possession relation could behave differently at the discourse level.
Alternatively, the choice of indefinites as the comparison class for possessives in
Experiment 1 could have masked a main effect of possession that would have emerged with a
different comparison class. Therefore, Experiment 2 includes both a comparison of possessives
to a different type of simple nominal (i.e. definites) and an explicit manipulation of
representational complexity (i.e. adjectival modification) in order to further investigate potential
effects related to possessives’ inherent complexity.
Finally, this work expands on previous studies of informativity and representational
complexity by considering referents’ rates of mention and subjecthood as the primary dependent
variables. This approach complements previous work, which has primarily looked at the
accessibility of referents in resolving pronouns and filler-gap dependencies (e.g. Hofmeister
2011; Hofmeister & Vasishth, 2014; Karimi & Ferreira, 2016; Troyer et al., 2016) and rates of
pronominalization (e.g. Karimi et al., 2014, 2019).
Pronoun resolution and pronominalization have been formalized as two different
conditional probabilities under the Bayesian approach of Kehler et al. (2008) and Kehler &
Rohde (2013). In this framework, pronoun resolution—a comprehension-based statistic—is
represented as the conditional probability P(referent|pronoun) (i.e. the likelihood that a pronoun
refers to a particular referent). On the other hand, pronominalization—based on production—is a
function of P(pronoun|referent) (i.e. the likelihood that a pronoun is used to refer to a particular
referent). Their approach includes another production-based term P(referent), which is simply
the probability of the referent being mentioned with any form. This term is analogous to the
primary dependent measure I consider in Experiments 1 and 2. All three of these terms relate to
54
the referent’s prominence and represent different approaches to quantifying a highly abstract
concept. Thus this work expands on previous literature on representational complexity by
measuring prominence through a related—yet underutilized—dependent variable.

3.4.1. Participants
Data from 112 native English speakers is presented for analysis. All participants were 18
years of age or older and were recruited from Amazon Mechanical Turk. They received $3.75 in
exchange for their participation in the experiment. All participants included in the analysis
reported being born in the United States, identified as native speakers of English, and had normal
or corrected-to-normal vision and normal hearing.
There were 167 total participants who completed the experiment. As in Experiment 1, I
excluded 8 participants who failed attention checks meant to detect automated responses., 4
participants who were either born outside the US or not native English speakers, and 3
participants with vision or hearing impairments. I excluded 10 participants whose responses were
mostly duplicates (i.e. copy-pasted) or syntactically ill-formed fragments, which suggested an
automated response. An additional 12 participants were excluded because at least a quarter of
their responses were fragments, non-sequiturs, ungrammaticalities indicative of a non-native
speaker, repeated responses, and/or instances of abusive or offensive language. Lastly, to balance
the number of participants per list, 18 additional participants were excluded
2
. After all the
aforementioned exclusions, 112 participants were evenly distributed across the 16 experimental
lists (7 per list). Finally, 8 trials from these included participants (less than 0.3% of the data)
were excluded because the responses were fragments.

2
This number was unusually high due to a technical error which affected the automatic assignment of participants to
experimental lists.
55

3.4.2. Design and materials
Like Experiment 1, Experiment 2 had 56 items: 24 targets and 32 fillers. The prompts for
all target items followed the frame: [role noun
i
] [verb
PAST
] {the, her
i
, his
i
} {Ø, adjective} {role
noun
j
, inanimate
j
} because., e.g. The carpenter judged his left-handed nurse because (see
examples in Table 14).

Definite Possessive
Animate
The carpenter judged
the {Ø, left-handed} nurse because
The carpenter judged
his {Ø, left-handed} nurse because
Inanimate
The carpenter judged
the {Ø, gluten-free} cake because
The carpenter judged
his {Ø, gluten-free} cake because
Animate
The midwife knew
the {Ø, bearded } priest because
The midwife knew
her {Ø, bearded } priest because
Inanimate
The midwife knew
the {Ø, patterned} towel because
The midwife knew
her {Ø, patterned} towel because
Table 14. Example target items (2 of 24) in each of the 8 conditions

Three variables related to the direct object of the prompt were manipulated: (i) whether the
object was possessed or definite, (ii) whether it was modified with an adjective or not, and (iii)
whether it was animate or inanimate. Animate prompt objects, as well as all prompt subjects,
were human role nouns (e.g. farmer, plumber, babysitter); inanimate objects were concrete
alienable possessions (e.g. basket, refrigerator, windbreaker).
As in Experiment 1, there is an increased potential for referential ambiguity in
participants’ continuations when the prompt object is animate (e.g. she might not have a clear
antecedent if both the prompt subject and object are interpreted as female). As such, the subject
role nouns and animate object role nouns were chosen so that they contrasted with respect to
their stereotypically associated gender. As in Experiment 1, I used norming data from Misersky
56
et al. (2014) for this purpose. Additionally, I included some role nouns which are lexically
specified for gender (e.g. postman, congresswoman).
In half of the targets, the subject role noun was female-biased (e.g. cheerleader,
seamstress, beautician) with a male-biased direct object (e.g. stockbroker, bodyguard, chauffeur)
in animate conditions; in the other half, the order of the genders was reversed. As in Experiment
1, animate objects for each target item were paired with inanimates that matched as closely as
possible in lexical frequency (Brysbaert & New, 2009), syllable length, and character length.
In contrast to Experiment 1, Experiment 2 used English verbs in prompts instead of
nonce verbs. I chose verbs without strong implicit causality biases—based on Hartshorne &
Snedeker (2013), with an object-bias range from 45-69%—to ensure that the verb itself did not
strongly influence participants’ mentions of the prompt subject and object. Verbs were selected
to be as natural as possible with both animate and inanimate objects.
In this experiment, the prompt object was modified by an adjective. Because descriptors
of animates and inanimate have limited overlap, the adjectives used with animate and inanimate
nouns had to differ but were carefully controlled for objectivity, non-gradeability, length, and
informativity relative to the modified noun (e.g. clean-shaven [butler], priceless [jewel], missing
[housekeeper], broken [microwave], widowed [psychic], industrial [refrigerator]). Subjective
and perspective-sensitive adjectives were not used, as they might introduce additional linguistic
processing related to gradeability or perspective taking (e.g. Lasersohn, 2005; Stephenson, 2007;
Stojanovic, 2007).
In contrast to Experiment 1, where the continuations were separate sentences from the
prompts, continuations to all target items in Experiment 2 were subordinate clauses beginning
with because. This change afforded better control of the coherence relation between the prompt
57
and continuation
3
, which could influence participants’ mentions of the prompt referents (e.g.
Kehler, 2002).
None of the 32 filler items contained possessives or pronouns. Among the fillers, 5 were
catch trials, in which people chose from a set of words which one was most closely related to a
target word (e.g. Which of the words within the brackets is related most closely to ‘walk’?
{driving, jogging, swimming, flying}). The rest were sentence continuation trials which
resembled targets in consisting of a first clause followed by a connective (because or although)
and were similar to targets in length. I also included evaluative adverbs in clause-initial positions
in some fillers (e.g. Predictably, the vegetables disgusted the children because…; The shed
collapsed although, fortunately,…). These adverbs were intended to introduce a salient repeated
element which would distract participants from the inherent repetitiveness of the targets.
The 2 × 2 × 2 (possessive/indefinite × adjective/unmodified × animate/inanimate) design
resulted in 8 conditions. Conditions were distributed across 8 experimental lists in a Latin square
(i.e. within-subject and within-item design). Thus, each participant saw 3 items per condition.
Targets and fillers on lists were interleaved in a pseudorandomized fashion. Eight additional lists
were created by reversing the original lists, for a total of 16 experimental lists. deduce

3.4.3. Procedure
The experiment was implemented using PCIbex (Zehr & Schwarz, 2018). Before
beginning the study, participants were told that they would be asked to read prompt clauses and
to write continuations to complete the sentence. Participants were instructed that their

3
While annotating the data in Experiment 1, I attempted to classify continuations by the coherence relation between
the prompt and continuation; however, it became clear after an initial inspection of the data that for most
continuations it was impossible to confidently infer the coherence relation because of the nonce verb in the prompt.
58
continuations should satisfy three criteria: (i) that the continuation should be meaningful and
appropriate in the context of the prompt, (ii) that the prompt and continuation should combine to
yield a single complete sentence, and (iii) that the continuation by itself should be able to stand
as a complete sentence (to avoid prepositional continuations, which would lack subjects, as in
because of her suspicions). Additionally, participants were told not to use copy-paste or repeat
their answers. Participants completed two practice prompts (without any possessives or
pronouns, and with different connectives than in the experiment) and saw samples of acceptable
continuations and unacceptable fragments. As in Experiment 1, aside from warnings about
fragments and copy-paste, participants were assured that there were no right or wrong answers.
Experimental items were presented one at a time on the screen (see the example in Figure 5).

Figure 5. Example item from Experiment 2 as it appeared to participants

3.4.4. Pre-processing of the continuation data
As in Experiment 1, the continuation data from Experiment 2 were annotated in order to
perform statistical analyses. Response coding proceeded in the same way described in Section
2.4.4. Again, 10% of the data were randomly selected to be double-coded by a research assistant.
The coding schema again proved highly reliable, as the secondary coding matched my own in
over 99% of responses.

59
3.4.5. Predictions
Predictions related to the possession and animacy manipulations remain largely the same
as those for Experiment 1 (see Section 2.4.5). Like Experiment 1, Experiment 2 considers the
same two sets of hypotheses: the Complexity Hypothesis and the Parasitic Possession
Hypothesis, which relate to the overall discourse-level effect of possession, and the Interaction
Hypothesis and the Animacy Hypothesis, which relate to possible differences in the discourse-
level effect of possession depending on the type of possessive relation. As in Experiment 1, I
consider both mentions in subject position and across the entire continuation as measures of
prominence, with a greater likelihood of mention indicating increased prominence. These
measures are based on prior work showing that more prominent referents are more likely to be
mentioned subsequently in the discourse (e.g. Arnold, 2001; Givón, 1983; Kaiser, 2009; Kehler
et al., 2008; Kehler & Rohde, 2013) and that referents with the highest prominence tend to be
realized as grammatical subjects (e.g. Chafe, 1976; Gordon et al., 1993; Stevenson et al., 1994).
As in Experiment 1, the Complexity Hypothesis and the Parasitic Possession Hypothesis
predict opposing main effects of possession on the likelihood of mention of the prompt object. If
the Complexity Hypothesis is on the right track, possession should increase the prompt object’s
likelihood of mention in the continuation, while the Parasitic Possession Hypothesis predicts the
opposite pattern. Likewise, for the other set of hypotheses, the Interaction Hypothesis predicts an
extra boost in mentions of possessed animates (i.e. a statistical interaction of animacy and
possession), as was observed in Experiment 1. The Animacy Hypothesis predicts no interaction,
with the difference between possessed animates and possessed inanimates reflecting the same
difference observed in their definite counterparts.
60
The differences in the predictions of Experiments 1 and 2 emerge when we consider
representational complexity. Experiment 2 contains adjective modification as an additional
independent variable. Following prior work on informativity and representational complexity, I
expect adjective modification to increase the prominence of modified prompt objects and lead to
a higher likelihood of mention for these referents (e.g. Hofmeister, 2011; Karimi et al., 2019)
4
.
Such an effect would be realized in the statistical analysis as a main effect of adjective
modification. Additionally, it may be that adjective modification interacts with one of the other
independent variables, but there are no obvious reasons to expect such an interaction. For
instance, existing work on the discourse-level effects of adjective modification does not suggest
that increases in representational complexity affect animates and inanimates differently. On the
other hand, if Experiment 2 finds evidence for the Complexity Hypothesis, with possession
boosting the prominence of the possessed referent, it could be that adjective modification has a
diminished additional effect on possessed referents. Such a pattern would essentially be a ceiling
effect, in which two prominence-boosting factors would not combine additively due to an upper
limit on how prominent the particular referent can become.
Finally, I expect that the relative prominence of the prompt subject and object will differ
in Experiment 2 compared to Experiment 1. In the results of Experiment 1, I observed that
animate prompt objects were preferred as continuation subjects over prompt subjects, which
were also animate. This result was surprising given the well-known advantage in prominence for

4
A potential concern related to adjective modification is that the presence of the adjective might encourage the
pragmatic inference that the prompt object is being contrasted with an alternative referent. For example, if
participants see The midwife knew the bearded priest because, they might infer that there is another discourse-
relevant priest without a beard. This inference might then affect participants’ continuations in a way which is not
related to representational complexity; however, prior work on redundant adjectives has shown that speakers
frequently produce adjectives which overspecify the intended referent (e.g. Rubio-Fernández, 2016; Tarenskeen et
al., 2015; Tourtouri et al., 2019). Thus the potential for this unintended pragmatic inference should not be
particularly worrisome.
61
subjects (e.g. Chafe, 1976; Crawley et al., 1990; Gordon et al., 1993; Stevenson et al., 1994). I
hypothesized that this result may have been due to the greater informativity of the role noun
objects compared to the proper name subjects. Experiment 2 addresses this imbalance in
complexity by using role nouns for both the prompt subjects and animate objects. Thus I expect
prompt subjects to be preferred as subjects of continuations in Experiment 2.

3.4.6. Results
Analysis of the results proceeded in an extremely similar fashion as in Experiment 1.
Again mentions of the prompt subject and object were analyzed in continuation-subject position
and across the entire continuation. Accordingly, there again were four main analyses
corresponding to the mentioned referent and its position. As in Experiment 1, I also consider
pronominalization patterns at the end of this section. Binomial models were fit in the same way
as in Experiment 1, but with adjective modification added to the predictors, which were again
deviation coded (adjective = 0.5, no adjective = -0.5; possessive = 0.5, definite = -0.5; animate =
0.5, inanimate = -0.5).

3.4.6.1. Mentions of the prompt subject as continuation subject

I first analyzed the subject position of continuations by fitting a model which predicted
the probability of the continuation subject referring back (with any type of referential form) to
the prompt subject. A visualization of mentions of the prompt subject as the continuation’s
subject is given below in Figure 6.

62

Figure 6. The proportion of continuations in Experiment 2 in which the subject
refers back to the prompt subject (error bars represent ±1 standard error)

The model revealed a main effect of adjectival modification, whereby modification of the prompt
object resulted in a significantly higher likelihood of mentioning the prompt subject as the
subject of the continuation (see Table 15 for a summary of the model’s fixed effects). This effect
was surprising in that the hypotheses considered in Section 3.4.5 did not make predictions related
to adjective modification’s effect on the prompt subject, which was not the locus of the adjective
manipulation.

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) -0.012 0.226 -0.051 0.959
animacy 0.073 0.177 0.412 0.680
possession 0.086 0.128 0.670 0.503
adjective 0.267 0.134 1.989 0.047
animacy:possession -0.110 0.179 -0.615 0.539
animacy:adjective -0.060 0.179 -0.335 0.738
possession:adjective 0.034 0.179 0.188 0.851
animacy:possession:adjective 0.218 0.356 0.611 0.541
Table 15. Summary of fixed effects for the model predicting mentions of the
prompt subject in subject position of the continuation (Experiment 2)
adjective no adjective
definite possessive definite possessive
0.0
0.1
0.2
0.3
0.4
0.5
0.6
Prompt object determiner
Proportion of continuations in which
continuation subject = prompt subject
Prompt object
animacy
animate
inanimate
63

3.4.6.2. Mentions of the prompt object as continuation subject
A second model was fit which predicted the probability of the continuation subject
referring to the prompt object. A visualization of mentions of the prompt object as the
continuation’s subject is given below in Figure 7.

Figure 7. The proportion of continuations in Experiment 2 in which the subject
refers to the prompt object (error bars represent ±1 standard error)

This model showed a main effect of possession, specifically that possessed prompt objects were
significantly less likely to be mentioned as continuation subjects compared to definite prompt
objects (see Table 16 for details of the model).

adjective no adjective
definite possessive definite possessive
0.0
0.1
0.2
0.3
0.4
0.5
Prompt object determiner
Proportion of continuations in which
continuation subject = prompt object
Prompt object
animacy
animate
inanimate
64
Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) -0.640 0.202 -3.174 0.002
animacy 0.344 0.215 1.602 0.109
possession -0.226 0.105 -2.154 0.031
adjective -0.106 0.090 -1.177 0.239
animacy:possession 0.143 0.183 0.779 0.436
animacy:adjective 0.240 0.181 1.329 0.184
possession:adjective -0.187 0.180 -1.037 0.300
animacy:possession:adjective 0.409 0.361 1.132 0.258
Table 16. Summary of fixed effects for the model predicting mentions of the
prompt object in subject position of the continuation (Experiment 2)

This effect suggests that possessed prompt objects were less prominent in participants’ discourse
representations and supports the Parasitic Possession Hypothesis. Regarding animacy effects, I
expected that animate prompt objects would have a higher likelihood of subjecthood than
inanimate prompt objects (i.e. a main effect of animacy). This pattern was present in the results
of Experiment 1 and generally aligns with a well-documented advantage in discourse
prominence for animate referents; however, in Experiment 2, there was only a trend in the
expected direction which did not reach significance. Additionally, unlike in the model for the
prompt subject, adjective modification had no effect on the prompt object’s likelihood of
mention as the continuation subject, suggesting that increased representational complexity did
not confer increased discourse prominence on the modified referent.
Lastly, I analyzed participants’ choice of continuation subject between the prompt subject
and object when the object was animate. In Experiment 1, it was surprising that the prompt
object was preferred over the prompt subject as subject of the continuation when both were
animate. A possible explanation for this pattern was that the role noun objects were more
representationally complex than the named subjects in Experiment 1. In Experiment 2, since both
65
the subject and object were role nouns in the animate conditions, I expected a bias toward the
prompt subject as the continuation subject. For this analysis, continuations with subjects that
referred to neither the prompt subject nor the prompt object were removed from the data, so the
choice of subject referent could be analyzed as a binomial outcome. The intercept in the resulting
model was not significant (β = 0.319, SE = 0.238, z = 1.341, p = 0.180), indicating that the
choice between the prompt subject and object as continuation subject was not different from
chance (50%). Even though the was no clear preference between the prompt subject and object
as continuation subject, this pattern is closer to the expected result than in Experiment 1.

3.4.6.3. Mentions of the prompt subject anywhere in continuations
As in Experiment 1, mentions across the entirety of the continuation were also analyzed.
First, a model was fit which predicted the probability of mentioning the prompt subject anywhere
in the continuation. A visualization of mentions of the prompt subject in all continuation
positions is given below in Figure 8.

Figure 8. The proportion of continuations in Experiment 2 which contain a
reference to the prompt subject (error bars represent ±1 standard error)
adjective no adjective
definite possessive definite possessive
0.0
0.1
0.2
0.3
0.4
0.5
0.6
Prompt object determiner
Proportion of continuations which
mention prompt subject
Prompt object
animacy
animate
inanimate
66

Like the comparable model for continuation subject position, mentions of the prompt subject
across the continuation exhibited an effect of adjectival modification, with modification of the
prompt object significantly increasing the likelihood of mentioning the prompt subject (see Table
17 for details). As in the subject-position analysis, it was unexpected that modification of the
prompt object would increase the likelihood of mention of the unmodified prompt subject.

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) 0.310 0.205 1.510 0.131
animacy 0.121 0.096 1.262 0.207
possession 0.060 0.096 0.629 0.529
adjective 0.306 0.125 2.444 0.015
animacy:possession -0.136 0.178 -0.767 0.443
animacy:adjective -0.056 0.178 -0.313 0.754
possession:adjective -0.020 0.177 -0.115 0.909
animacy:possession:adjective 0.383 0.355 1.080 0.280
Table 17. Summary of fixed effects for the model predicting mentions of the
prompt subject in any position of the continuation (Experiment 2)

3.4.6.4. Mentions of the prompt object anywhere in continuations
Mentions of the prompt object in all positions of the continuation were also analyzed, and
a visualization of these data is given below in Figure 9.

67

Figure 9. The proportion of continuations in Experiment 2 which contain a
reference to the prompt object (error bars represent ±1 standard error)

The model predicting mentions of the prompt object in any position did not reveal any
significant effects (see Table 18 for details). Unlike the comparable model for continuation
subject position, there was no main effect of possession. Additionally, there was a numerical
trend in the expected direction for modified prompt objects to be mentioned more often than
unmodified ones; however, this pattern did not reach significance.

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) 0.641 0.193 3.325 0.001
animacy -0.012 0.221 -0.056 0.955
possession -0.079 0.09 -0.874 0.382
adjective 0.237 0.147 1.615 0.106
animacy:possession 0.147 0.181 0.811 0.417
animacy:adjective 0.045 0.183 0.244 0.807
possession:adjective 0.057 0.181 0.313 0.755
animacy:possession:adjective -0.015 0.362 -0.043 0.966
Table 18. Summary of fixed effects for the model predicting mentions of the
prompt object in any position of the continuation (Experiment 2)

adjective no adjective
definite possessive definite possessive
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Prompt object determiner
Proportion of continuations which
mention prompt object
Prompt object
animacy
animate
inanimate
68
3.4.6.5. Pronominalization patterns
As in Experiment 1, rates of pronominalization for the prompt subject and object were
analyzed in both subject position and across the entire continuation. The analysis was performed
in essentially the same was as in Experiment 1 in order to model the conditional probability of
using a pronoun given mention of a particular referent. The data were subsetted to just the
continuations which mentioned the relevant referent in the relevant position, and then a model
was fit to the binomial outcome of using a pronoun or not. The models were fit with the same
predictors and according to the same principles as the previously discussed models which
estimate the simple probability of mention.
First, the conditional probability of using a pronoun given mention of the prompt subject
was analyzed. A visualization of these data is given for subject position in Figure 10 and all
positions of the continuation in Figure 11.

Figure 10. Pronominalization likelihood (conditional probability of pronoun
given mention) in Experiment 2 for the prompt subject in subject position of
continuations (error bars represent ±1 standard error)

adjective no adjective
definite possessive definite possessive
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Prompt object determiner
Conditional probability of pronoun
given prompt subject mention
as continuation subject
Prompt object
animacy
animate
inanimate
69

Figure 11. Pronominalization likelihood (conditional probability of pronoun
given mention) in Experiment 2 for the prompt subject in all positions of the
continuation (error bars represent ±1 standard error)

I fit a model for pronominalization of the prompt subject for each of these analysis
domains. The two models show similar patterns (see Tables 19 and 20 for details). In
subject position, the only significant effect is a main effect animacy, whereby the
presence of an animate prompt object made the prompt subject less likely to be
pronominalized when it was the subject of the continuation. In the model for
pronominalization of the prompt subject across the entire continuation, the main effect of
animacy is only marginally significant, and again no other effects reached significance.
These patterns seem to reflect previously discussed competition for prominence which
occurs when multiple animate referents are present in the discourse.

adjective no adjective
definite possessive definite possessive
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Prompt object determiner
Conditional probability of pronoun
given prompt subject mention
Prompt object
animacy
animate
inanimate
70
Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) 7.195 1.205 5.970 <0.001
animacy -2.527 0.648 -3.898 <0.001
possession -0.426 0.521 -0.818 0.413
adjective -0.177 0.501 -0.353 0.724
animacy:possession 0.960 1.046 0.919 0.358
animacy:adjective 0.298 1.001 0.298 0.766
possession:adjective 0.821 1.029 0.798 0.425
animacy:possession:adjective -1.521 2.050 -0.742 0.458
Table 19. Summary of fixed effects for the model predicting pronominalization of
the prompt subject in subject position of the continuation (Experiment 2)

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) 13.758 2.187 6.292 <0.001
animacy -2.000 1.090 -1.835 0.066
possession -1.190 0.986 -1.207 0.228
adjective 0.107 0.955 0.112 0.911
animacy:possession 1.334 1.906 0.700 0.484
animacy:adjective 2.547 1.961 1.299 0.194
possession:adjective 1.989 1.931 1.030 0.303
animacy:possession:adjective -0.776 3.818 -0.203 0.839
Table 20. Summary of fixed effects for the model predicting pronominalization of
the prompt subject in all positions of the continuation (Experiment 2)

The same analysis was repeated for pronominalizations of the prompt object. A
visualization of these data is given for subject position in Figure 12 and all positions of the
continuation in Figure 13.

71

Figure 12. Pronominalization likelihood (conditional probability of pronoun
given mention) in Experiment 2 for the prompt object in subject position of
continuations (error bars represent ±1 standard error)

Figure 13. Pronominalization likelihood (conditional probability of pronoun
given mention) in Experiment 2 for the prompt object in all positions of the
continuation (error bars represent ±1 standard error)

I again fit a model for each of the analysis domains, both of which contained no significant
effects. Summaries of the models’ fixed effects appear below in Tables 21 and 22.

adjective no adjective
definite possessive definite possessive
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Prompt object determiner
Conditional probability of pronoun
given prompt object mention
as continuation subject
Prompt object
animacy
animate
inanimate
adjective no adjective
definite possessive definite possessive
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Prompt object determiner
Conditional probability of pronoun
given prompt object mention
Prompt object
animacy
animate
inanimate
72
Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) 4.831 0.847 5.705 <0.001
animacy -2.412 1.577 -1.530 0.126
possession 0.286 0.373 0.766 0.443
adjective 0.147 0.351 0.418 0.676
animacy:possession -1.099 0.748 -1.468 0.142
animacy:adjective -0.735 0.696 -1.056 0.291
possession:adjective 0.572 0.732 0.781 0.435
animacy:possession:adjective 0.225 1.467 0.153 0.878
Table 21. Summary of fixed effects for the model predicting pronominalization of
the prompt object in subject position of the continuation (Experiment 2)

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) 3.626 0.338 10.716 <0.001
animacy 0.337 0.474 0.711 0.477
possession -0.070 0.218 -0.320 0.749
adjective -0.075 0.218 -0.343 0.732
animacy:possession -0.263 0.435 -0.605 0.545
animacy:adjective 0.196 0.437 0.447 0.655
possession:adjective -0.125 0.435 -0.287 0.774
animacy:possession:adjective 0.128 0.870 0.147 0.883
Table 22. Summary of fixed effects for the model predicting pronominalization of
the prompt object in all positions of the continuation (Experiment 2)

3.5. Discussion
Similar to Experiment 1, Experiment 2 examines the discourse-level effect of possession
with a sentence continuation task. Experiment 2, however, included an additional independent
variable—adjective modification—and used prompts which differed in a few ways from the ones
in Experiment 1. The chief difference between the prompts in the two studies is that Experiment
2 compared possessives to definites, while Experiment 1 used indefinites. I have discussed in
Section 3.2 why definites might be a closer match with possessives with regard to the discourse
73
conditions which license their use, although both comparisons are applicable to the overall aims
of this dissertation.
Experiment 2 considers the same two sets of competing hypotheses addressed in
Experiment 1; the Complexity Hypothesis and Parasitic Possession Hypothesis relate to the
overall effect of possession on the possessed referent, while the Interaction Hypothesis and
Animacy Hypothesis relate to potential differences in the effect of possession depending on the
type of possession relation. The results of Experiment 1 provided some evidence in favor of the
Complexity Hypothesis in the analysis of pronominalization patterns (a secondary measure in the
experiment) and no evidence for the Parasitic Possession Hypothesis. On the other hand,
Experiment 1 revealed an extra boost in prominence for possessed animates, which supported the
Interaction Hypothesis. Thus Experiment 2 aimed to find stronger evidence for an overall
discourse-level effect of possession and also to further test the Interaction Hypothesis. Lastly,
Experiment 2 investigates whether possessives inherent complexity relative to simple nominals
might have discourse consequences similar to a previously investigated type of complexity
manipulation (i.e. adjective modification). This research question would become especially
pertinent if I observed additional evidence in favor of the aforementioned Complexity
Hypothesis.

3.5.1 Support for the Parasitic Possession Hypothesis and the Animacy Hypothesis
The results of Experiment 2 were substantially different from those of Experiment 1. In
subject position of continuations, I found that possessed prompt objects were significantly less
likely to be mentioned compared to definites. This result provides support for the Parasitic
Possession Hypothesis, which posits that possessed referents are less prominent than simple
74
nominals. Additionally, the effect of possession was not modulated by animacy (i.e. there was no
significant interaction), as one might have expected based on the results of Experiment 1. Thus,
the results of Experiment 2 do not support the Interaction Hypothesis and show no effect which
could be attributed to interpersonal relationships.
Additionally, there were no main effects of animacy or adjectival modification on object
mentions, both of which were predicted to increase the object’s likelihood of mention
(notwithstanding a numerical trend in the expected direction for the animacy manipulation).
Although there was no significant main effect of animacy, unlike in Experiment 1, the lack of an
interaction between animacy and possession still supports the Animacy Hypothesis, in so far as
this hypothesis predicts no difference in the effect of possession depending on the type of
possession relation. Furthermore, when I analyzed participants’ choice of continuation subject in
animate conditions, where both the prompt subject and prompt object are animate, the results
also differed from what I found in Experiment 1. While Experiment 1 found that animate prompt
objects were preferred as continuation subjects, there was no clear preference between prompt
subjects and objects as continuation subjects in Experiment 2. While the prompt subject was
expected to be preferred as the subject of continuations (on account of subjects’ well-
documented advantage in prominence), these results still provide some support my intuition that
the greater informativity of the role nouns relative to the names in Experiment 1 made them more
prominent.
If the results in subject position support the Parasitic Possession Hypothesis and
possessed prompt objects are lower in prominence than definites, one might wonder why the
same effect did not obtain across all positions of the continuation. Here it should be noted that
prior work emphasizes the prominence of subject position in particular (e.g. Chafe, 1976; Clark
75
& Begun, 1971; Gordon et al., 1993) and, specifically in sentence continuation experiments,
mentions in subject position have been the primary locus of analysis (e.g. Arnold, 1991; Crawley
& Stevenson, 1990; Stevenson et al., 1994). When the entirety of the continuation is considered,
there may be other factors, such as the length of the continuation or the choice of verb (e.g.
transitive vs. intransitive), which alter a referent’s likelihood of mention. While I believe that
mentions across the continuation should still be considered as a supplementary measure of
discourse prominence, the lack of a possessive effect in the analysis of all continuation positions
should not be a basis to doubt the validity of the results in subject position. Therefore, I interpret
the lower likelihood of subjecthood for possessed prompt objects as evidence in favor of the
Parasitic Possession Hypothesis.

3.5.2 Representational complexity and an unexpected adjective modification effect
Notwithstanding the pronominalization patterns in Experiment 1, the analyses of
mentions in both Experiment 1 and Experiment 2 produced no support for the Complexity
Hypothesis. Comparison of these results to previous work using adjectives to increase a
referent’s representational complexity (e.g. Hofmeister, 2011; Hofmeister & Vasishth; 2014,
Karimi et al., 2019) thus seems to suggest that the semantic complexity inherent to possessives is
different from the increase in features conferred by adjectival modification. Nevertheless, the
effect of adjective modification in Experiment 2 was unexpected. Instead of increasing the
likelihood of mention of the modified referent (i.e. the prompt object), the presence of the
adjective boosted mentions of the prompt subject in both subject position and across the entire
continuation. This effect suggests that the adjective only increased the prominence of the
76
unmodified referent. What then could be driving this effect, which has not been addressed in
previous literature on the discourse-level effects of adjective modification?
Based on these results, it seems that the adjective’s effect was not limited in scope to the
modified referent but rather affected the overall coherence of the prompt and continuation. For
example, consider the prompt The ballerina discussed the surgeon because. In this extremely
limited discourse context, participants could imagine a variety of possible causes for the
described event; however, with the addition of the adjective (i.e. The ballerina discussed the
ambidextrous surgeon because), continuations of the prompt may be slightly more constrained.
In the modified condition, the fact that the surgeon is ambidextrous can serve as the cause of the
discussion event. As such, the adjective provides extra information for the participant to leverage
as they consider the causal coherence relation between their continuation and the prompt.
Furthermore, participants are likely to use this information from the adjective, since they will
pragmatically reason that that the presence of the adjective is likely due to its relevance to the
overall discourse. In this way, the adjective may serve to support the overall coherence of the
continuation and prompt. If the coherence relation is strengthened, I suggest that the prompt
subject, which should already be the most prominent referent by default, might increase in
prominence.

3.5.3 Underlying mechanisms of the Parasitic Possession Hypothesis
If the Parasitic Possession Hypothesis is on the right track, as Experiment 2 indicates,
there is still the lingering issue of why a possessed referent should be less prominent than a
simple nominal. There are several possible explanations for why this tendency might emerge.
Firstly, compared to simple nominals, possessives include an additional mention of the
77
possessor. This additional mention presumably reactivates the possessor’s discourse
representation (or introduces it if the possessor is discourse-new) and might confer extra
prominence to the possessor at the expense of the possessed referent’s prominence; however, this
explanation is not entirely satisfying, since the results do not show that possession affects the
prompt subject’s likelihood of subjecthood or mention. If possession decreases the prompt
object’s prominence simply because it privileges the possessor (i.e. the prompt subject), one
might expect to see a complementary effect of possession for the prompt subject.
An alternative explanation for why possession could decrease the prominence of a
referent concerns the semantic link between a possessor and possessed referent. As the
participants in Experiment 2 processed the prompts, they formed two discourse representations
for the mentioned entities. Additionally, in the possessive conditions, it stands to reason that the
link between the possessor and possessed referent also has some sort of representation on the
discourse level. This view is supported by the distinction which Barker (1991, 2000) makes
between “lexical” and “extrinsic” possessives. Barker proposes that in lexical possession
interpretations, the semantics of the possession relation come from a relational head noun, e.g. in
John’s birthday, birthday is a relational noun and supplies the possession relation in its
semantics. In contrast, Barker proposes that other possessive interpretations are extrinsic and
supplied by the discourse context, e.g. John’s day in the Whose day is it to do the dishes?
Saturday is usually John’s day. The prompt objects in Experiment 2 (e.g. postman, housekeeper,
skateboard, microwave) seem more similar to the latter category. Unlike purely relational nouns
like mother or birthday, these nouns are natural to interpret without being relativized to some
other referent (e.g. the skateboard vs. #the birthday). Nevertheless, there may be some lexical
semantic specification for the type of possession relation the nouns in Experiment 2 typically
78
enter into. For instance, his microwave seems to have a default interpretation as expressing an
ownership relation, but, in the right discourse context, the possessor could also be the designer of
a particular kind of microwave. On this basis, it seems justified to assume that the possession
relations expressed in Experiment 2 were somehow present in participants’ discourse-level
representations, along with the referents which participated in the possessive. If this relation is
present in participants’ discourse representations, it stands to reason that it could affect the
prominence of the involved referents.
One theoretical mechanism which could be used to explain why possession might
decrease the prominence of possessed referents is to understand the possessive relation as
creating a hierarchical dependency structure of discourse representations. Such a theory would
posit that the discourse representations for possessed referents are dependent on those of their
possessors. This dependent relationship could render the possessed referent less accessible and
prominent on the discourse level. Such a theory could operate analogously to “file cards” for
referents in the File Change Semantics of Heim (1982). Specifically, in this theoretical
framework different linguistic circumstances can lead to a file card being amended or to the
introduction of a new file card. To extend this idea to the difference in Experiment 2 between
possessives and definites, it could be that simple nominals introduce independent file cards,
whereas possessed referents are represented as additions to the file cards of their possessors. This
structure should not be understood to make the representations of possessed referents completely
inaccessible to discourse-level operations. Such a claim would be easily disproven by a discourse
like Mary drove her car. It’s a red convertible, in which the representation for possessed car
must be available to resolve the subsequent pronoun. Instead, the dependent status of possessed
referents could be associated with greater processing cost if theses representations need to be
79
accessed or modified in subsequent discourse. The consequence of such a theory would be that
possessed referents are less prominent in discourse, i.e. the critical claim of the Parasitic
Possession Hypothesis.
I have provided here some initial suggestions about possessed referents’ discourse
representations based on the file card metaphor, but a possible future step for this work would be
a more precise theoretical account of the nature of the involved discourse representations, how
they relate to each other, and how they are created and accessed. Such an endeavor might be able
to build on ideas from Discourse Representation Theory (e.g. Kamp, 1981; Kamp & Reyle,
1993), which already includes the notion of hierarchically nested discourse representations. I
leave this undertaking for future work.

3.5.4. Combined insights from Experiments 1 and 2
Given that Experiments 1 and 2 seem to support different conclusions regarding the
critical hypotheses addressed in this dissertation, some explanation is required to understand
these results as a whole. I first consider the Complexity Hypothesis and the Parasitic Possession
Hypothesis, which pertain to the overall effect of possession on possessed referents. Experiment
1 found evidence for the Complexity Hypothesis, but only within the pronominalization patterns.
On the other hand, Experiment 2 supported the Parasitic Possession Hypothesis. What then is
responsible for this discrepancy? It seems likely that this difference is attributable to the choice
of simple nominals against which possessives are compared. In Section 3.2, I reviewed how the
licensing conditions for possessives, indefinites, and definites align. Based on the differing
discourse properties of indefinites and definites, I suggest that the differences between
Experiments 1 and 2 can be explained in a unified way. Since indefinites permit a non-specific
80
interpretation and reflect a less focused attentional state, it follows that they would be overall less
prominent than definites (e.g. Chiriacescu & von Heusinger, 2010; Gundel et al., 1993; von
Heusinger, 2011). Furthermore, Experiment 2 shows that possessed referents are less prominent
than definites. Therefore, when mentions of possessed referents were compared to mentions
indefinites in Experiment 1, it is not unreasonable to think that possession’s prominence-
reducing effect was masked by an equally strong effect of indefiniteness. This same effect might
also be responsible for the seeming increase in prominence for possessives in the
pronominalization patterns. Although this result could reflect the logic of the Complexity
Hypothesis, it is also possible that it is due to a reduction in prominence for the indefinites
because of discourse-level factors unrelated to possession. Furthermore, given that definites
seem to generally be a better match for possessives with regard to the discourse conditions which
license them, it is perhaps appropriate to weigh the results of Experiment 2 more strongly when
considering the overall effect of possession on prominence.
Experiments 1 and 2 also produced conflicting results relating to whether the type of
semantic possessive relation modulates the discourse-level effect of possession. I have
previously framed this question as a choice between the Interaction Hypothesis, which
Experiment 1 supported, and the Animacy Hypothesis, which Experiment 2 supported based on
the lack of an interaction between animacy and possession. The Interaction Hypothesis posits
that animate possessed referents get an extra boost in prominence; framed another way, this
interaction could also be characterized as possessed animates being less susceptible than
inanimates to possession’s overall effect of reducing prominence. On the other hand, the
Animacy Hypothesis states that the discourse-level effect of possession is the same for all types
of possessed referents—essentially the null hypothesis relative to the Interaction Hypothesis.
81
Because Experiment 2 provides only negative evidence against the Interaction Hypothesis, it still
seems reasonable to believe that there could be differences in the effect of possession on
prominence depending on the type of possessive relation. In order to reject Experiment 1’s
support for the Interaction Hypothesis, there would need to be a compelling alternative
explanation for the seemingly robust interaction observed between animacy and possession. For
example, if it could be argued that the animate indefinite objects in Experiment 1 were more
likely to be interpreted non-specifically than the inanimate indefinites, then it would be
reasonable to claim that the cause of the interaction was a reduction in prominence specifically
for animate indefinites because of an asymmetrical effect of specificity. I find this account not
especially persuasive and instead favor cautious adoption of the Interaction Hypothesis.
Currently, the most convincing argument seems to be that the effect of possession on possessed
referents varies with the type of possessive relation and that possessed animates’ denotation of
interpersonal relations may confer extra prominence on them. However, this topic is certainly
deserving of additional research, and I will return to it in subsequent chapters (especially in
Chapter 5).
Finally, some readers may wonder whether differences in the coherence relation between
the prompt and continuation in Experiments 1 and 2 were at all responsible for the divergent sets
of results. In Experiment 2, the connective because always produced an explanation relation
between the prompt and continuation; on the other hand, no particular coherence relation was
privileged in Experiment 1, since the prompt and continuation were separate sentences.
Furthermore, it was not possible to reliably identify the coherence relations for Experiment 1
continuations because of the semantic ambiguity introduced by the nonce verbs in the prompts.
Regarding the effect this difference might have had on the results, it firstly seems unlikely that
82
the coherence relation would have affected differences between the conditions in the two
experiments. It seems reasonable to assume that even if coherence relations cannot be identified
in Experiment 1, the likelihood of one relation or another would not vary across conditions. For
example, in the absence of any constraints on what coherence relation a participant could choose,
there does not seem to be any reason to believe that a participant would be more likely to use an
explanation relation in the indefinite conditions versus the possessive conditions. Thus, as far as
differences between conditions are concerned, the coherence relation in Experiment 1 can be
treated as if it were constant (as in Experiment 2). At the same time, it is of course possible that
if one could isolate just the continuations in Experiment 1 which had an explanation coherence
relation, the results might look closer to those in Experiment 2.

83
Chapter 4: Investigating real-time processing of possession relations with web-based eye-
tracking

4.1. Introduction
The two sentence continuation experiments reported in Chapters 2 and 3 produced rather
different results. The results of Experiment 1 (Chapter 2) reflect an extra boost in prominence for
possessed animates. Additionally, there was some evidence in the analysis of pronominalization
patterns that possession resulted in an overall increase in the prominence of the possessed
referent; however, this result was evaluated with some caution, since the primary analyses (based
on likelihood of mention) contained no such effect. On the other hand, Experiment 2 (Chapter 3)
shows no interaction between animacy and possession. Instead the results suggest that possession
reduces the prominence of both animate and inanimate referents to a similar extent. In the
previous chapter, I compared these two sets of results and argued for why differences in the
design of the two experiments might explain the divergent findings. Based on these arguments, I
consider the results of Experiment 2 more representative of the overall discourse-level effect of
possession. Therefore, between the two previously discussed hypotheses regarding possession’s
overall effect on referents, the results so far favor the Parasitic Possession Hypothesis, which
posits that possessed referents have a lower baseline prominence compared to simple nominals.
Additionally, Experiments 1 and 2 supported different conclusions related to whether the type of
possession relation modulates the discourse-level effect of possession (i.e. the Interaction
Hypothesis versus the Animacy Hypothesis). For reasons discussed in Section 4.2.4, this chapter
will focus primarily on the overall effect of possession, although I return to differences between
possession relations in Chapter 5.
84
Given that Experiment 1 found no support for the Parasitic Possession Hypothesis,
further investigation is needed before it can be confidently adopted. Therefore, in this chapter, I
switch methodologies from sentence continuation to visual-world eye-tracking, which provides a
view into the real-time discourse-level processing of possessives and a potentially finer-grained
measure of prominence. In the sentence continuation method used in the first two experiments,
participants comprehend the prompt, build a mental model of the discourse, and use these
discourse representations to produce a continuation. In analyzing sentence continuation results,
researchers essentially infer an end state of comprehension processing based on a static view of
subsequent production. While this method can certainly produce valid, theoretically informative
results, it does not provide any direct insight into how possession affects real-time language
processing. It is well known that mental computations in language processing occur
incrementally (e.g. Altmann & Steedman, 1988; Bock & Levelt, 1994; Bock et al., 2004; Tyler
& Marslen-Wilson, 1977), and likewise prominence is understood to be a dynamic property of
referents which changes over time as the discourse continues (e.g. Ariel, 1988; van den Broek et
al., 1996; von Heusinger & Schumacher, 2019). In order to understand how discourse-level
processing of possessives plays out in real time—and to compare these findings to previously
discussed offline measures—Experiment 3 uses visual-world eye-tracking.
Visual-world eye-tracking operates under the linking hypothesis that people tend to look
at the visual correspondent of the linguistic referent which they are currently processing (with
some delay due to the motor constraints of the eyes) (e.g. Cooper, 1974; Tanenhaus et al., 1995).
While this methodology has been used to investigate a variety of linguistic domains, its
application to pronoun resolution is especially relevant to the research aims of this dissertation
(e.g. Arnold et al., 2000; Järvikivi et al., 2005; Kaiser & Trueswell, 2008; Kaiser, 2011). Because
85
pronoun resolution is typically understood to be mediated by discourse representations (e.g.
Heim, 1982; McCawley, 1979; Prince, 1981), it is an appropriate domain to test hypotheses
related to the discourse-level processing of possessives. Nevertheless, this dissertation focuses on
possession’s effect on prominence, and relating pronoun interpretation to prominence brings up
some potential difficulties which were not present in previous chapters. While examining a
referent’s likelihood of mention is a simple and convenient measure of prominence, pronoun
interpretation is known to depend on the interaction of multiple constraints (e.g. Badecker &
Straub, 2002; Kaiser & Trueswell, 2008; Kaiser, 2011). Prominence itself (and closely associated
concepts, like salience, activation, and accessibility) has often been listed among these
constraints, with more prominent referents being more likely antecedents for pronouns (e.g.
Ariel, 1990; Arnold, 2001; Gundel et al., 1993). Additionally, an assortment of linguistic factors
which themselves are associated to varying degrees with prominence have also been identified as
constraints on pronoun interpretation: among others, subjecthood (e.g. Arnold et al., 2000;
Givón, 1983; Grosz et al., 1995; Kaiser & Trueswell, 2008), implicit causality (e.g. Cozijn et al.,
2011; Hartshorne & Snedeker, 2013), and coherence (e.g. Hobbs, 1979; Kehler, 2002).
Nevertheless, if all other factors influencing pronoun interpretation can be held constant, it
should be possible to compare referents’ discourse prominence through accessibility to
pronominal reference. Indeed this strategy has been used by other researchers interested in how
other linguistic factors affect discourse status (e.g. Karimi & Ferreira, 2016).

4.2. Experiment 3
Experiment 3 is a visual-world eye-tracking study which examines the effect of
possession on the prominence of referents in a discourse. Differences in prominence are
86
estimated primarily through the proportion of looks to visual depictions of referents during
pronoun interpretation. In accordance with previous work, a higher proportion of looks to a
referent is taken to indicate greater accessibility or prominence (e.g. Arnold et al., 2000; Järvikivi
et al., 2005; Kaiser 2011). Experiment 3 uses quite similar linguistic stimuli to Experiments 1
and 2 but an entirely different methodology. Rather than writing continuations, in Experiment 3,
participants listened to sentences and clicked on the image corresponding the person or thing
they thought was mentioned last in the sentence. In target items, the last referent mentioned was
the antecedent of a pronoun, so the task was essentially to resolve the pronoun.
Thus Experiment 3 seeks to further test the hypotheses discussed in previous chapters,
specifically the Parasitic Possession Hypothesis and the Complexity Hypothesis. While
Experiment 1 produced weak support for the Complexity Hypothesis, Experiment 2 clearly
supported the Parasitic Possession Hypothesis. If the findings of Experiment 3 were to match
those of Experiment 2, the convergent evidence from multiple methodologies would strongly
support this theory. Finally, Experiment 3 is among the first psycholinguistic studies to use
webcam-based eye-tracking (e.g. Degen et al., 2021; Slim & Hartsuiker, 2021; Vos et al., 2022).
Thus the results of Experiment 3 serve not only to address the research aims of this dissertation
but also to validate and advance this emerging methodology.

4.2.1. Participants
Data from 56 native English speakers is presented for analysis. All participants were 18
years of age or older and were recruited on Prolific. All participants included in the analysis
reported being born in the United States, identified as native speakers of English, and had normal
or corrected-to-normal vision and normal hearing.
87
There were 154 total participants who completed the experiment. Data from 21
participants were excluded because they were either not native English speakers or were born
outside the US. Additionally, I excluded 2 participants with a vision or hearing impairment and 7
participants who responded that they were not wearing headphones (which participants were
instructed multiple times to use, since it was crucial to ensure that they could hear the auditory
stimuli).
The experiment itself contained some unambiguous trials, which were also used as
exclusion criteria. Among the target items, half of the trials which a participant saw contained
only one animate referent, which was unambiguously the antecedent of the pronoun and
therefore the last-mentioned referent (e.g. The princess saw her television through the glass door
because she was crompety; see Section 4.2.2). Participants who could not select the animate
referent in these trials either were not attending to or did not understand the task. On this basis,
45 participants who responded correctly to fewer than 9 of 12 unambiguous target items were
excluded. Additionally, the last-mentioned referent in filler items was unambiguous, except for
those trials which included a homophone competitor among the images (see Section 4.2.2.2). A
further 12 participants were excluded because they responded correctly to fewer than 13 of 17
unambiguous filler items.
The final exclusion criterion was based on eye-tracker calibration checks which took
place prior to each target item (see Section 4.2.3 for more on calibration procedures); 11
additional participants were excluded because their mean calibration score before target items
was lower than 40%. After all exclusions, the remaining 56 participants were evenly distributed
across the 4 experimental lists (14 participants per list). All participant exclusions were
performed before any data analysis was conducted.
88

4.2.2. Design and materials
Experiment 3 had 48 items: 24 targets and 24 fillers. As is typical in visual-world eye-
tracking, all items included an array of images and an auditory stimulus. For each target item,
participants heard a sentence which followed the frame: [role noun
i
] [verb
PAST
] {the, her
i
, his
i
}
{role noun
j
, inanimate
j
} [prepositional phrase] because {she, he} was [nonce adjective], e.g. The
princess saw the attorney through the glass door because she was crompety (see examples in
Table 23).

Definite Possessive
Animate
The princess saw the attorney
through the glass door
because she was crompety
The princess saw her attorney
through the glass door
because she was crompety
Inanimate
The princess saw the television
through the glass door
because she was crompety
The princess saw her television
through the glass door
because she was crompety
Animate
The soldier heard the plumber
throughout the autumn evening
because he was nimpacious
The soldier heard his plumber
throughout the autumn evening
because he was nimpacious
Inanimate
The soldier heard the dryer
throughout the autumn evening
because he was nimpacious
The soldier heard his dryer
throughout the autumn evening
because he was nimpacious
Table 23. Example target items (2 of 24) in Experiment 3 in each of the 4
conditions

Two variables related to the direct object of the matrix clause were manipulated: (i) whether the
object was possessed or definite and (ii) whether it was animate or inanimate. As in Experiment
2, I chose to use definites as the comparison class for possessives. There were two main reasons
behind this choice. Firstly, as discussed in Section 3.2, definites are a closer match to possessives
with regard to the conditions that license their use. Secondly, I judged that in a visual-world
89
paradigm, it might be odd to use an indefinite when the participant’s visual familiarity with the
referent would sufficiently license a definite.
As in Experiment 2, animate objects, as well as all subjects, were human role nouns (e.g.
farmer, boxer, violinist); inanimate objects were concrete alienable possessions (e.g. toothbrush,
handkerchief, mailbox). Animate and inanimate objects within the same item were again
matched as closely as possible in length and lexical frequency in the SUBTLEXus corpus
(Brysbaert & New, 2009). Because of the animacy manipulation, the pronoun after because is
ambiguous in the animate conditions and unambiguous in the inanimate conditions. Half of the
items contained she as the critical pronoun, and half contained he, with pronoun gender
counterbalanced across experimental lists.
Matrix-clause verbs were again chosen based on norming by Hartshorne & Snedeker
(2013) to be maximally equi-biased with respect to implicit causality and comparably natural in
both animate and inanimate conditions. The prepositional phrases were strongly biased toward
VP attachment in order not to increase the informational content of the direct object and
consequently increase its potential prominence (e.g. Hofmeister, 2011; Karimi & Ferreira, 2016;
Karimi et al., 2019). Similar to the prompts in Experiment 2, the subordinate clauses of target
sentences (which contained the pronouns) all began with because in order to control the
coherence relation between the two clauses. Nonce adjectives were used to avoid introducing
semantic content after the pronoun which might disambiguate its referent in the animate
conditions (for the use of nonce words in pronoun-interpretation tasks, see also Hartshorne &
Snedeker, 2013).
The filler sentences were more varied in their structure than the targets. One third of the
fillers contained a because clause similar to the targets (e.g. The flight attendant threw the
90
barstool in the garbage because she got a new couch), one third contained a relative clause (e.g.
The bowler met the chess master who always ate peanut butter and jelly sandwiches), and one
third consisted of two clauses coordinated with and then (e.g. The realtor sold the house at the
end of the street, and then the neighbor was curious). Half of the fillers contained nonce words
(e.g. The cellist dreezed some dog biscuits, which she bought for the German shepherd), which
varied in their position within the sentence and grammatical category. Lastly, 15 of the filler
sentences contained unambiguous pronouns in various syntactic positions (e.g. The waitress
picked up gluten-free bread from the baker, who made it with special flour), which were
balanced with respect to whether they referred to the initial clause’s subject or object.
The 2 × 2 (possessive/definite × animate/inanimate) design resulted in 4 conditions,
which were distributed across 4 experimental lists in a Latin square (i.e. within-subject and
within-item design). Thus, each participant saw 6 items per condition. Items were presented to
participants in random order.

4.2.2.1. Auditory stimuli
Auditory stimuli were recorded, assembled, and annotated using Praat (version 6.1.38)
(Boersma, 2002). During recording, a head-mounted microphone was used (Razer BlackShark
V2 X), and multiple repetitions of each target and filler item were elicited in a randomized order.
Target sentences’ matrix clauses were isolated and spliced together with subordinate because
clauses which were held constant within items. Accordingly, the final audio files for a given
target item shared the same because clause across conditions. Across target items, the because
clauses in the final audio files were taken from different conditions in a counterbalanced manner.
The audio files for targets were then annotated with Praat TextGrids to allow for later analysis
91
based on several time alignments. For each file, I marked (i) the onset of speech, (ii) the onset of
the matrix-clause object noun, (iii) the offset of the matrix-clause object noun, (iv) the onset of
the critical pronoun, (v) the offset of the critical pronoun, and (vi) the offset of speech.

4.2.2.2. Visual stimuli
All images in Experiment 3 were clip art (sourced from Clipart.com) on a white
background. Images were edited with Paint S (version 6.0.3) as needed. Every target item had
three images, corresponding to the matrix-clause subject, object, and complement to the
preposition (the “look-away”), as in Figures 14 and 15. The images for the animate and
inanimate objects of each item were matched as closely as possible in aspect ratio, so that image
size would not be confounded with animacy.

Figure 14. An example screen layout from Experiment 3 for the target sentence
The soldier heard the plumber throughout the autumn evening because he was
nimpacious

92

Figure 15. An example screen layout from Experiment 3 for the target sentence
The runner praised her nurse during the podcast interview because she was
pambly

The fillers had three to five images, most of which corresponded to a referent in the
sentence, as in Figures 16 and 17; however, in 9 of the 24 fillers, one of the images was a
semantic associate of the last-mentioned referent (e.g. a rhinoceros as a competitor with
elephant, an armchair as a competitor with couch). Additionally, 7 of the 24 fillers contained
multiple images corresponding to homophonous interpretations of the last-mentioned referent
(e.g. ruler accompanied by pictures of both a monarch and a measuring device, bat accompanied
by pictures of both a flying mammal and an item of baseball equipment).

93

Figure 16. An example screen layout from Experiment 3 for the filler sentence
The golfer ate the bromble because the surfer took the last slice of pizza (utilizing
the top-left, bottom-left, and right regions, with the center regions empty)

Figure 17. An example screen layout from Experiment 3 for the filler sentence
The teapot filled with spreltoom was prepared for the pirate, who drank it with
honey (utilizing the same left, center, and right regions as in the targets)
94

The layout of the screen and eye-tracking regions differed for targets and fillers. For
targets, the screen was divided into three regions: left, center, and right. These regions were used
subsequently as the interest areas in the eye-tracking analysis. The left and right regions each
occupied 35% of the width of the participant’s screen and 100% of its height. The dimensions of
these and all other visual elements are specified as percentages of the screen area, since
participants used a variety of monitor sizes and aspect ratios. The left and right regions were
aligned flush with the vertical edges of the screen. The center region was 20% of the width of the
participant’s screen and 100% of its height and was horizontally centered. This layout left two
small margins of 5% of the screen’s width between the center and side regions.
For targets, the three images for that trial were placed into the three regions, with the
look-away always in the center and the subject and object randomly assigned to the left and
right. The images were sized and positioned such that they were contained inside a box of 80%
of the region’s width and height. For example, because the left and right regions each occupied
35% of the screen’s width and 100% of its height, the images for a target item’s subject and
object were contained in a box of 28% of the screen’s width and 80% of its height. This box was
centered within its enclosing region, and the image was vertically or horizontally centered within
its enclosing box depending on which dimension limited the size of the image; however, for
almost all screen ratios, the enclosing box’s height would be significantly greater than its width,
so the image would be scaled to fill the box’s width and centered with respect to its height. This
layout guarantees that there was a margin of whitespace around the subject and object images of
at least 3.5% of the screen’s width in the horizontal dimension and at least 10% of the screen’s
height in the vertical dimension. The reason for this margin around the images was so that gaze
95
estimates which fell just outside of the image dimensions due to the inherent noisiness of
webcam-based eye-tracking were still treated as looks to the relevant region.
The screen layout was more varied for filler items because the number of images was
different across items. In addition to the three regions from the targets, six additional regions
were added by halving the original three into top and bottom portions. Thus there were nine
potential regions into which filler items could be placed; however, if, for example, the top-left
region was occupied by one image, the larger left region could not be used for another of the
item’s images. Furthermore, not all fillers utilized every possible region, so that participants
would not get accustomed to a particular layout over the course of the experiment.
The filler images were sized and positioned within the screen regions according to the
same principles as the target images. They were arranged within the potential screen locations in
a pseudo-randomized and counterbalanced fashioned. Because the last-mentioned entity in the
target trials would always appear on the left or right, in half of the 24 fillers, the last-mentioned
entity was in one of the center regions. Thus participants could not adopt a strategy of ignoring
the center images. The position of the first-mentioned entity was also counterbalanced in the
fillers, although this referent was less important to the task.

4.2.3. Procedure
Like Experiment 2, Experiment 3 was conducted over the internet using PCIbex (Zehr &
Schwarz, 2018). For eye-tracking experiments, PCIbex utilizes the JavaScript library
WebGazer.js (Papoutsaki et al., 2016) to provide gaze estimates using participants’ webcams.
The experiment was run using PCIbex version 2.0 with minor modifications to the EyeTracker
element (edits to the feedback participants saw when a calibration attempt was unsuccessful and
96
the addition of a global variable to count the total number of calibration attempts participants
used during the experiment).
Prior to the experiment, participants went through an initial calibration phase, during
which the WebGazer gaze-estimation model was trained for individual participants. During this
phase, participants fixated on dots in the center and around the edges of the screen. Their
calibration was checked at the end of the procedure with a final central dot, and they were given
a calibration score out of 100% (the percentage of webcam samples while the final dot was on
screen during which the WebGazer model estimated that they were looking at the dot).
Participants had five attempts to calibrate to at least 60%. This calibration parameter and the
others reported here were chosen based on piloting and earlier eye-tracking experiments I built
for the USC Language Processing Lab. Prior to beginning the initial calibration, participants
were given tips on how to maximize their score (e.g. to maintain the distance between their face
and the webcam, to move only their eyes when looking around the screen, not to block any
portion of their face from the webcam). If a calibration attempt was unsuccessful, participants
were given additional feedback to try to improve their score (e.g. to check the lighting of their
environment, to stare at the dots the entire time they were on the screen, to adjust their glasses so
that they did not reflect light back into the webcam). Participants who used up all five calibration
attempts but could not score at least 60% still advanced to the experimental task.
After the initial calibration phase, participants were instructed on the experimental task.
They were told that in each item, they would see pictures of people and things and hear a
sentence about them. They were told to listen to the entire sentence and then click on whichever
person or thing was mentioned last in the sentence. They were told they would have ten seconds
to click the last-mentioned person or thing once the sentence had ended. Participants were also
97
warned that many of the sentences would contain made-up words, but they would still be able to
understand enough of the meaning of the sentence to click on the last-mentioned person or thing.
Finally, participants were informed that their calibration would be checked before every item
with the same central dot as during the initial calibration and that, after these checks, they might
in some cases be asked to repeat the calibration procedure.
Participants then completed three unambiguous practice items, which provided them with
feedback after each of their answers (e.g. After The woman told the man a blorpy story about the
Statue of Liberty: That's right! The Statue of Liberty was the last thing mentioned or Not quite!
You were supposed to click the Statue of Liberty, since it was the last thing mentioned). The
practice items followed the same format and procedure as the experimental items, except that
both the sentences and the feedback were presented as text (see Figure 18 below).

Figure 18. The screen layout for the first of three practice items in Experiment 3
(with feedback displayed after a correct click)

98
Two of the three practice items contained pronouns as the last-mentioned referent, so that
participants would not misinterpret the task as clicking on the last referring expression (e.g. The
meebo told the girl to pick up the sponge and wipe the car with it). However, neither the task
instructions nor feedback ever referenced pronouns explicitly; the participants were simply told
to click on whichever person or thing was mentioned last. The task was explained this way to
minimize the potential for metalinguistic reasoning or unnatural response strategies by
participants. After the practice items, participants were informed that they were about to begin
the experiment and reminded that they would hear (not read, as in the practice) the sentences in
each item.
Before each experimental trial, participants’ calibration was checked with a central dot
and their score was recorded. Before fillers, participants proceeded to the item after the
calibration check regardless of their score. The calibration check procedure was only present
before fillers to avoid alerting participants to the different types of trials; however, no gaze
estimates were recorded during the fillers. Before targets, participants were, in some cases, asked
to repeat the calibration procedure based on their score on the calibration check. Firstly,
participants who had not scored at least 40% on their final calibration attempt during the initial
calibration phase were never asked to repeat the calibration procedure during the experiment.
Without this feature, participants whose mean pre-trial calibration score would very likely be too
low for inclusion in the analysis (see Section 4.2.1) would needlessly be asked to repeat the
calibration procedure many times. I anticipated that this situation would frustrate participants and
lead to excessive time needed to complete the experiment. For similar reasons, participants who
had already recalibrated more than eight times during preceding experimental trials (i.e. not
including attempts during the initial calibration phase) were also not asked to repeat the
99
calibration procedure on subsequent trials. Participants who did not meet either of the
aforementioned criteria and did not score at least 40% on the pre-trial calibration check were
given two attempts to recalibrate with a score of at least 40%. If they failed to score at least 40%,
they proceeded to the trial regardless. Because the pre-trial calibration checks looked the same
for all trials and participants, participants were naïve to the control flow described above.
Each trial began with the participant clicking a central button to indicate that they were
ready, after which a fixation cross was displayed. At this point in target trials, eye-gaze estimates
started to be recorded. After approximately
5
500 ms, the fixation cross was removed, and the
image array was presented. The participant was allowed to freely inspect the image array for
2000 ms before the audio started playing. After the audio ended, clicks were enabled on the
various regions of the image array. The trial ended either when the participant clicked on one of
the image regions or after 10,000 ms had elapsed after the end of the audio.

4.2.4. Predictions
In Experiment 3, the focus of the analysis is the gaze patterns during pronoun
interpretation. The design of the experiment was tailored to this domain and, therefore, the most
important predictions relate to these data. Nevertheless, the methodology of Experiment 3
produces a rich dataset, which also includes eye-gaze data at other intervals, selections of the
last-mentioned referent, and selection latencies. Some of these analyses are more exploratory
than others and have less clear predictions, but I nevertheless discuss their potential import here.

5
Durations of all trial components can be considered approximate due to variation inherent to remote experiments.
In almost all cases, the programmed (i.e. intended) duration of a particular component was only a few milliseconds
faster than its logged duration in our results (due to the time needed for PCIbex to execute commands).
100
Regarding the hypotheses tested in Experiment 3, I noted earlier that this experiment
primarily considers the Parasitic Possession Hypothesis and the Complexity Hypothesis (i.e. the
two theories relating to the overall effect of possession on the prominence of referents). The
reason for this is that the critical pronoun of the target sentences is only ambiguous in the
animate conditions, when both the matrix-clause subject and object are candidate antecedents. In
the inanimate conditions, when the pronoun can refer only to the matrix-clause subject, I do not
expect the possession manipulation to produce any differences in accessibility to pronominal
reference, since the matrix-clause object is not a grammatical antecedent. Therefore, Experiment
3 primarily investigates potential effects of possession within the animate conditions. Because
the experiment is not designed to compare effects of possession with two different types of
possession relations (as in Experiments 1 and 2), the results have limited relevance to the
Interaction Hypothesis and Animacy Hypothesis (i.e. whether different possession types
modulate the discourse-level effect of possession).
On this basis, one might wonder how I can draw firm conclusions about the overall effect
of possession on the basis of Experiment 3, given that only one type of possession relation (i.e.
interpersonal relations) is being considered. This is a valid concern, which can be addressed by
interpreting the potential results in the context of Experiments 1 and 2. If the results of
Experiment 3 demonstrate a reduction in prominence for possessed animates relative to their
definite counterparts, such a result would be consistent with the results of Experiment 2, which
also compared possessives to definites and found an overall reduction in prominence for
possessed referents. Additionally, given that Experiment 1 suggested that possession provides a
superadditive boost in prominence for animates, if Experiment 3 shows a reduction in
prominence for possessed animates, this result likely entails a reduction in prominence for
101
possessed inanimates as well. To reframe this logic, I have no reason to believe that there is an
alternative to the Interaction Hypothesis and Animacy Hypothesis in which possessed inanimates
would receive a superadditive boost in prominence.
With those concerns addressed, I turn to the specific predictions for Experiment 3 and
begin with the most relevant analysis, the eye-gaze patterns during pronoun interpretation. As
discussed above, I do not expect possession to have an effect within the inanimate conditions,
when the pronoun is unambiguous. These conditions illustrate whether participants understand
the task and serve as a baseline for the ambiguous animate conditions. Within the animate
conditions, ambiguous pronouns are expected to drive fewer looks to the object in the possessed
versus definite conditions. Specifically, for a target sentence like The ballerina discussed {the,
her} congresswoman throughout the seafood dinner because she was brillish, participants are
predicted to look less to the congresswoman when the object is preceded by her compared to the.
Such a result would indicate that the possessed object is less accessible and therefore less
prominent in discourse than its definite counterpart. This result would be consistent with
Experiment 2 and support the Parasitic Possession Hypothesis. On the other hand, if the pronoun
drives more looks to the object when it is possessed, this result would provide evidence in favor
of the Complexity Hypothesis, which posits that the additional informational content of
possessives contributes to an increase in discourse prominence relative to simple nominals.
However, this result would constitute only somewhat convincing support for the Complexity
Hypothesis, since an increase in prominence for possessed animates could be attributed to the
Interaction Hypothesis (i.e. a superadditive boost in prominence for possessed animates) with no
overall increase in prominence for possessed referents.
102
Given that Experiment 3 is concerned with referent’s accessibility to subsequent
pronominal reference, it seems reasonable to also consider participants’ selections of the last-
mentioned referent as an additional offline measure. Again, I do not expect differences between
the two inanimate conditions. In these conditions, choices of the matrix-clause subject—the only
grammatical antecedent when the object is inanimate—should be at ceiling levels. On the other
hand, in the animate conditions, the subject is still expected to be the preferred antecedent, given
the overall higher rate of pronominalization for subjects (e.g. Arnold et al., 2000; Givón, 1983;
Grosz et al., 1995; Prince, 1992); however, one would expect a reduction to some extent in the
rate of subject choices, simply due to the availability of the object as a grammatical antecedent.
This pattern should produce a main effect of animacy in the statistical model.
This animacy effect though does not inform my hypotheses about the discourse-level
effect of possession. The question of primary importance in the selection data is whether
possessed animates will be chosen less (or more) often relative to definite animates as the
antecedents of the ambiguous pronoun. If they are chosen less often, this result would indicate a
decrease in accessibility and prominence for possessed referents and support the Parasitic
Possession Hypothesis. As in the eye-gaze analysis, the opposite pattern would provide less
strong support for the Complexity Hypothesis. However, it might be that the offline selection
data is not sensitive enough to detect the changes in prominence produced by possession.
Along with the selections themselves, Experiment 3 also permits examination of the
latencies of participants’ selections. These data are by their nature more fine-grained than the
selections and thus might provide some additional insights. However, this analysis is somewhat
exploratory, as it is not typically considered in similar experiments, and there are external factors
that might influence the speed of participants’ selections (e.g. screen size and the use of a mouse
103
versus a trackpad). For this analysis, slower selection latencies in the possessed animate
conditions might indicate decreased accessibility and prominence for possessed referents;
however, this pattern would also be consistent with increased competition between the subject
and possessed object and instead support the opposite conclusion that possession makes the
object more accessible. Notwithstanding these potential issues, I do expect that selection
latencies will be significantly slower when the pronoun is ambiguous versus unambiguous (i.e. a
main effect of animacy). This result would be consistent with prior work showing that
ambiguous pronouns require more cognitive resources to resolve than unambiguous ones (e.g.
Stewart et al., 2007).
Finally, I also examine eye-gaze patterns while participants are processing the matrix-
clause object. Unlike eye movements triggered by the pronoun, which represent retrieval of
discourse referents, gaze patterns during the matrix clause represent encoding of these
representations. Thus these patterns may be more difficult to relate to the concept of prominence.
One could argue that fewer looks to possessed referents at this stage would indicate a decrease in
prominence. On the other hand, the presence of the possessive pronoun, which refers back to the
subject, might be enough to draw participants’ gaze off the object regardless of prominence.
Conversely, prior work on representational complexity has shown that more complex referents
require more time and cognitive resources to encode, despite their faster retrieval times (e.g.
Hofmeister, 2011; Hofmeister & Vasishth, 2014). Therefore, it could be that the inherent
complexity of possessives drives more looks to possessed referents during encoding. Thus the
analysis of gaze patterns during the matrix clause is more exploratory than the pronoun analysis,
although it can also serve to confirm the validity of the webcam-based methodology by showing
that participants do actually look at the object when it is mentioned.
104

4.2.5. Results
Prior to any analysis, I examined whether participants selected grammatically possible
antecedents for the critical pronouns in target trials. As discussed above, a substantial number of
participants were excluded because they did not reliably select the matrix-clause subject as the
last-mentioned referent when the matrix-clause object was inanimate; however, participants who
met the criteria for inclusion in the analysis consistently chose grammatical antecedents for the
critical pronoun. Out of 1344 target trials from included participants, there were 29
(approximately 2% of the data) where the participant did not select a grammatical pronoun
antecedent (24 selections of the look-away, 4 of the object in inanimate conditions, and 1 trial
with no selection). These trials were excluded from the statistical analyses (with one exception in
the selection analysis, which is noted below). Additionally, one trial was affected by an audio
buffering error which rendered its timing unreliable, and therefore it too was excluded. As in
Experiments 1 and 2, statistical analyses for Experiment 3 used mixed-effects models with the
maximal random effects structure that did not result in nonconvergence (Barr et al., 2013). The
independent variables animacy and possession were again deviation-coded (possessive = 0.5,
definite = -0.5; animate = 0.5, inanimate = -0.5).
In the following sections I present the results of the various analyses in Experiment 3. I
begin with the selection data and the selection latencies, before turning to the eye-tracking data. I
first review some basic observations about these data before considering looking behavior during
the matrix clause. I conclude with the most important and relevant data, which are the eye-gaze
patterns triggered by the pronoun.

105
4.2.5.1. Selections of the last-mentioned referent
Experiment 3 was designed primarily with eye-tracking analysis in mind. Nevertheless,
participants’ mouse-click selections of the last-mentioned referent were also analyzed. For this
analysis, a generalized linear mixed-effects model was fit to the binomial outcome of selecting
either the subject or object as the last-mentioned referent (subject = 1, object = 0). Trials where
the look-away was selected or no selection was made were excluded, as described above;
however, trials in the inanimate conditions where the object was selected (despite not being a
grammatical antecedent) were included in this particular model, so that all conditions potentially
included both of the binomial outcomes. The selection data are visualized below as Figure 19.

Figure 19. The proportion of selections of the subject as the last-mentioned
referent in Experiment 3 (error bars represent ±1 standard error)

The model revealed the predicted main effect of animacy (see Table 24 for a summary of the
model’s fixed effects); participants were more likely to select the matrix-clause subject as the
106
last-mentioned referent (i.e. the antecedent of the pronoun) when the object was inanimate versus
animate.

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) 4.700 1.801 2.609 0.009
animacy -7.557 3.572 -2.116 0.034
possession 0.739 0.621 1.190 0.234
animacy:possession -0.994 1.243 -0.800 0.424
Table 24. Summary of fixed effects for the model predicting selections of the
last-mentioned referent (Experiment 3)

Subject selections were at ceiling levels in the inanimate conditions, when the subject was the
only grammatical antecedent for the pronoun. When the object was animate, there were
significantly more object selections. An intercept-only model fit to a subset of just the animate
conditions showed that the subject was still the preferred antecedent for the pronoun even when
the object was also a grammatical possibility (i.e. the subject was selected significantly more
often than chance at 50%; β = 0.957, SE = 0.319, z = 3.000, p = 0.003). This result is consistent
with the widely accepted view that, with all other factors being equal, the subject is typically the
preferred pronoun antecedent (e.g. Arnold et al., 2000; Givón, 1983; Grosz et al., 1995; Prince,
1992).
There was no significant main effect of possession or interaction of possession and
animacy in the main model for the selection data, and consequently the selection data do not bear
on the Parasitic Possession Hypothesis or Complexity Hypothesis.

4.2.5.2 Selection latencies
As discussed in Section 4.2.5, participant’s selection latencies were also analyzed for
exploratory purposes. Multiple different time alignments are possible for this analysis (e.g.
107
relative to the pronoun or the offset of the sentence). While the selection latency relative to the
pronoun is not a true reaction time measure (since clicks were only permitted after the audio file
had finished playing), this alignment is potentially a better complement to the eye-tracking
analysis, which examines gaze patterns relative to the pronoun. Furthermore, because the interval
between the pronoun and the end of the audio file was constant within each item, any differences
between conditions that might be detected by analyzing the true reaction times still ought to be
revealed in this analysis.
Specifically, selection latencies were analyzed relative to both the pronoun’s onset and its
offset. While it is standard practice in psycholinguistic experiments to align analyses to the
onsets of critical words, I discovered a systematic difference in the length of the initial consonant
in some pronouns and decided to address this difference with an additional analysis aligned to
the pronoun’s offset. I describe this difference in more detail and justify the use of the alternative
alignment in Section 4.2.5.4, since it is more consequential for the eye-gaze analysis. For the
selection latencies, the two time alignments yielded model parameters which were qualitatively
the same. Therefore, only the selection latencies relative to the pronoun’s onset are reported here.
A visualization of these data is given below as Figure 20.

108

Figure 20. Selection latencies in Experiment 3 relative to the onset of the pronoun
(error bars represent ±1 standard error)

The model revealed a main effect of animacy (see Table 25 for details). In accordance with my
predictions, in the inanimate conditions, where the subject was the only grammatical antecedent
for the pronoun, participants were significantly faster to select the last-mentioned referent;
however, in the animate conditions, the presence of multiple grammatical antecedents resulted
longer selection latencies.

Fixed effect Estimate Std. Error df t value Pr(>|t|)
(Intercept) 2822.902 91.171 59.219 30.963 <0.001
animacy 1033.518 95.890 54.827 10.778 <0.001
possession 2.922 61.202 53.678 0.048 0.962
animacy:possession -29.637 114.054 1126.434 -0.260 0.795
Table 25. Summary of fixed effects for the model predicting selection latencies
relative to the onset of the pronoun

109
This result should be unsurprising, since ambiguous pronouns demand greater cognitive
resources (e.g. Stewart et al., 2007). As in the model for the selections themselves, there was no
significant main effect of possession or interaction of possession and animacy, and thus this
analysis does not provide evidence for either the Parasitic Possession Hypothesis or Complexity
Hypothesis.

4.2.5.3 General remarks on the eye-tracking analysis
Remote webcam-based eye-tracking is a novel methodology that has only recently come
into use in psycholinguistics. Webcam-based eye-tracking allows subjects to participate in
experiments without requiring an in-person visit to the lab but also presents difficulties not found
in laboratory eye-tracking, since spatial and temporal resolution are less fine-grained than with
specialized eye-tracking equipment (e.g. Slim & Hartsuiker, 2021; Vos et al., 2022).
Furthermore, data quality may vary greatly among participants, even when data from participants
with poor calibrations is discarded. For instance, while typical lab-based eye-trackers can record
gaze data at sample rates of 500 Hz and above (e.g. Carter & Luke, 2020; Vos et al., 2022; Xu et
al., 2015), the average sample rate for participants included in this analysis ranged from
approximately 9 Hz to 35 Hz, with a median of approximately 21 Hz. Thus, statistical techniques
developed for the richer and more precise data of lab-based eye-tracking may not be appropriate
for analysis of the eye-movement data in this experiment. Bearing in mind the novelty of web-
based eye-tracking and the limitations of these data, I present here a coarse-grained and
relatively simple analysis. My goal in this approach is to provide a basic starting point on which
subsequent, more complex analyses might be based when the field has decided upon accepted
statistical techniques for this type of data.
110
However, before performing any type of statistical analysis on the data, I first wanted to
qualitatively assess some of the most basic and predictable gaze patterns to assure myself (and
readers) that this novel methodology can produce interpretable and meaningful results. It should
be noted for added assurance that other researchers have previously validated webcam-based
methods by replicating results produced with laboratory eye-trackers (e.g. Degen et al., 2021;
Slim & Hartsuiker, 2021; Vos et al., 2022); however, as far as I am aware, Experiment 3 is the
first pronoun interpretation study to use webcam-based eye-tracking.
As an initial assessment of the eye-tracking data, I visually inspected the gaze patterns
beginning at the onset of the center-fixation cross. If the methodology works as intended, I
expected to see a high proportion of looks to the center while the fixation cross was on the
screen, increases in looks to the subject and object during the visual inspection period, and
distinct peaks in looks for each referent corresponding to its mention in the matrix clause. In
order to visualize the data, off-screen gaze estimates were trimmed and proportions of looks to
each of the regions of interest were calculated within 200 ms time bins. While I did not intend to
explore any differences between conditions in this qualitative assessment, the gaze patterns for
the animate and inanimate conditions were visually different enough to merit separate treatment.
Thus the proportions of looks below in Figure 21 are aggregated by time bin, region, and
animacy.

111

Figure 21. Eye movements in Experiment 3 relative to the onset of the center-
fixation cross (animate conditions vs. inanimate conditions)

For the animate conditions, Figure 21 shows the expected pattern of results. When the
fixation cross is the only thing on the screen (i.e. the first 500 ms of the trial), the proportion of
looks to the center is high, and there are few looks to the empty subject and object regions. Then,
after the subject, object, and look-away images are revealed, the proportions of looks to the
subject and object begin to rise, and there is a period of competition between these images while
the participants explore the visual array. Finally, after 2000 ms of visual inspection (2500 ms
after the fixation cross onset), distinct peaks in looks arise for the subject, object, and look-away
according to the order of their mention in the matrix clause.
On the other hand, gaze patterns in the inanimate conditions show a somewhat different
(but no less reasonable) pattern. After the images appear on screen, instead of approximately
even competition between the subject and object, looks to the animate subject noticeably
outweigh looks to the inanimate object during the visual inspection stage of the trial. This pattern
112
is not surprising given previous work showing that animates capture more visual attention than
inanimates even in the absence of linguistic context (e.g. Downing et al., 2004; New et al.,
2007). Nevertheless, once the audio begins around 2500 ms, one can again see peaks in looks to
the three referents in the expected order. Based on the sensible gaze patterns in both the animate
and inanimate conditions, I was confident in the validity of the eye-tracking data and proceeded
with the intended statistical analysis of the gaze patterns.

4.2.5.4 Time windows in the eye-tracking analysis
I began the analysis of the eye-tracking data by aligning the gaze estimates to three time
points of interest: (i) the onset of the matrix-clause object, (ii) the onset of the pronoun, and (iii)
the offset of the pronoun (e.g. The soldier heard the [i]plumber throughout the autumn evening
because [ii]he[iii] was nimpacious). The final time alignment was not a part of my original plans
for the analysis and may be surprising in the context of standard practice in psycholinguistic
experiments. Since it has been well established that comprehenders do not wait for the end of a
word to begin processing it (e.g. Allopenna et al., 1998; Tanenhaus et al., 1995), temporal
analyses are typically aligned to the onsets of critical words; however, there is a specific feature
present in this experiment’s stimuli which makes alignment to the pronoun’s offset a potentially
more appropriate (or at least equally valid) choice. As described in Section 4.2.2, all pronouns in
target items were preceded by because, with she in one half of the items and he in the other. I
observed while annotating the audio files that in the items with the feminine pronoun, the final
/z/ in because always underwent complete assimilation to the initial /ʃ/ of she, resulting in the
sequence [bikəʃːi]. Because of this coarticulation, the initial consonant of she was effectively
lengthened, and I could not identify where the onset of the pronoun would have occurred had the
113
assimilation not taken place. Therefore, I marked the onset of the pronoun for these items at the
onset of the lengthened [ʃː]. As a result, the durations of the feminine pronouns are on average
nearly 100 ms longer than those of the masculine pronouns (two-sample t-test, t = 19.547, p <
0.001; mean she duration = 194.1 ms, sd = 12.6 ms; mean he duration = 95.1 ms, sd = 12.2 ms).
On this basis, one might expect any effects related to pronoun interpretation to begin slightly
later relative to the marked pronoun onset for items containing she. Aligning analyses to the
offset of the pronoun may mitigate this variability and is unlikely to introduce a new confound,
since the coarticulatory effects between the final /i/ of the pronoun and was should not differ
across items. Because each condition within an item shared the same because clause, neither the
onset nor the offset alignment introduces any confound that would influence differences between
the conditions.
Gaze estimates were then binned into 1000 ms windows relative to the three time
alignments. Windows of 1000 ms were chosen so that there would be enough samples in a
window to mitigate data sparsity issues but still capture some aspect of the eye movements’
temporal dynamics. For all the time alignments, I included a window from -1000 ms to 0 ms
relative to the particular time point, as a means of checking gaze patterns prior to that event. For
the matrix-clause object alignment, I also examined windows from 0 ms to 1000 ms and 1000 ms
to 2000 ms after the onset. Since the median interval between the onset of the matrix-clause
object and the onset of the pronoun was 2071 ms (with a minimum of 1607 ms and a maximum
of 2562 ms), additional windows were not considered in this analysis; for all the items, a window
of 2000 ms to 3000 ms after the onset of the matrix-clause object would include a substantial
portion of time after the onset of the pronoun, which would likely alter any eye-movement
patterns attributable to the processing of the object. On the other hand, the window from 1000 ms
114
to 2000 ms only includes a small amount of time after the onset of the pronoun for about 42% of
items (40 of 96). If one considers that eye movements typically take about 200 ms to program
and execute (e.g. Hallett, 1986), the pronoun starts less than 1800 ms after the onset of the
matrix-clause object in only about 17% of the items (16 of 96). Given the large amount of data in
the 1000 ms to 2000 ms window which cannot be influenced by the pronoun and the
comparatively small amount which might be somewhat affected by it, I judged that it was
worthwhile to analyze this window.
For the pronoun alignments (i.e. time relative to both the onset and offset of the pronoun),
I considered 1000 ms windows until 4000 ms after the relevant event. As demonstrated by the
selection latency data (see Figure 20 above), participants took significantly longer to select an
antecedent for the pronoun in the ambiguous animate conditions. Since these conditions were the
primary focus for the eye-tracking analysis, the time windows for analysis were chosen based on
the distribution of selection latencies for these conditions in particular. About 46% of trials in the
animate conditions had a final eye-tracking sample recorded at least 3000 ms after the onset of
the pronoun (302 of 661); however, by 4000 ms after the onset of the pronoun, only about 23%
of animate trials had not yet ended. On this basis, I decided to use 3000 ms to 4000 ms after the
pronoun’s onset as the last analysis window. The percentages of trials still ongoing at 3000 ms
and 4000 ms were slightly lower for the alignment to the pronoun’s offset (42% and 22%,
respectively), but I nevertheless decided to analyze this window for both alignments.
After deciding on the time windows for the analysis, I trimmed off-screen gaze estimates
and calculated proportions of looks to each screen region, aggregating by participant, trial, and
time window. Thus each participant contributes one proportion of looks for each combination of
trial, screen region, and time window within a particular time alignment. I then fit mixed-effects
115
models for each screen region and time window which predicted the proportion of looks to that
region with animacy and possession as the independent variables.
For the eye-tracking visualizations, I employed a slightly different method to provide a
more nuanced and dynamic view of the gaze patterns. I again trimmed off-screen views but
calculated proportions of looks aggregated by animacy, possession, and smaller 200 ms time bins
(again relative to each of the three time alignments of interest).

4.2.5.5 Eye-movements during the matrix-clause
While the experiment was designed with the primary intention of analyzing participants’
looking behavior during pronoun resolution, an exploratory analysis of the gaze patterns during
the matrix clause was also performed. Specifically, I wanted to examine where participants
looked while processing the object, which was the locus of the experimental manipulations. I
predicted there would be animacy effects during this time period, both because the animate
object images would attract more looks than the inanimates even before the onset of the sentence
(see Figure 21 and associated discussion in Section 4.2.5.3) and because animates are known to
be more prominent in discourse than inanimates and therefore might attract more looks during
linguistic processing (e.g. Arnold & Griffin, 2007; Bock et al., 1992, Branigan et al., 2008). In
addition to these potential animacy effects, I was also interested in whether there would be any
effect of possession (or interaction of animacy and possession) during processing of the matrix-
clause object, which might implicate differences in how possessives are encoded in discourse.
Visualizations of the gaze patterns around the onset of the matrix-clause object are given below
as Figures 22 and 23. These two figures represent the same data, but Figure 22 shows how looks
116
to the three regions of interest unfolded within each condition, while Figure 23 shows how looks
to the subject and object varied across conditions.

Figure 22. Eye movements in Experiment 3 relative to the onset of the matrix-
clause object (each facet shows looks to the three regions of interest within a
condition)

117

Figure 23. Eye movements in Experiment 3 relative to the onset of the matrix-
clause object (facets compare conditions in the subject and object regions)

Statistical analyses were performed as described in Section 4.2.5.4 (i.e. by calculating
proportions of looks to each screen region, aggregating by participant, trial, and the three 1000
ms time windows). Accordingly, three models were fit for both the subject region and the object
region (i.e. a total of six models). Each model had animacy and possession as independent
variables and the proportion of looks as the dependent variable. Table 26 below summarizes the
results of these models (with significant effects shaded).

118
Region Fixed effect -1–0 s 0–1 s 1–2 s
Subject
(Intercept)
β = 0.327
SE = 0.014
df = 45.030
t = 22.731
p < 0.001
β = 0.207
SE = 0.014
df = 53.968
t = 14.382
p < 0.001
β = 0.113
SE = 0.009
df = 45.425
t = 12.672
p < 0.001
animacy
β = -0.068
SE = 0.018
df = 21.680
t = -3.695
p = 0.001
β = 0.033
SE = 0.013
df = 1121.354
t = 2.598
p = 0.010
β = -0.009
SE = 0.010
df = 110.093
t = -0.864
p = 0.390
possession
β = 0.012
SE = 0.016
df = 55.296
t = 0.745
p = 0.459
β = 0.007
SE = 0.014
df = 53.859
t = 0.543
p = 0.590
β = 0.010
SE = 0.010
df = 90.784
t = 0.988
p = 0.326
animacy:possession
β = 0.001
SE = 0.030
df = 1104.839
t = 0.023
p = 0.981
β = 0.011
SE = 0.026
df = 1122.150
t = 0.432
p = 0.666
β = 0.038
SE = 0.019
df = 1125.934
t = 1.973
p = 0.049
Object
(Intercept)
β = 0.196
SE = 0.009
df = 35.679
t = 21.206
p < 0.001
β = 0.275
SE = 0.014
df = 51.788
t = 19.729
p < 0.001
β = 0.240
SE = 0.016
df = 56.203
t = 15.444
p < 0.001
animacy
β = 0.109
SE = 0.013
df = 1133.350
t = 8.640
p < 0.001
β = 0.034
SE = 0.018
df = 22.079
t = 1.929
p = 0.067
β = 0.060
SE = 0.015
df = 21.515
t = 4.071
p < 0.001
possession
β = -0.024
SE = 0.013
df = 55.237
t = -1.795
p = 0.078
β = -0.010
SE = 0.013
df = 1159.900
t = -0.749
p = 0.454
β = 0.005
SE = 0.013
df = 855.516
t = 0.388
p = 0.698
animacy:possession
β = 0.042
SE = 0.025
df = 1133.873
t = 1.683
p = 0.093
β = 0.035
SE = 0.026
df = 1159.036
t = 1.352
p = 0.177
β = -0.038
SE = 0.026
df = 1161.882
t = -1.459
p = 0.145

Table 26. Summary of fixed effects for the models predicting looks to the subject
and object at time intervals aligned to the onset of the matrix-clause object
(Experiment 3); significant effects are shaded

119
As predicted, the models revealed strong effects of animacy. These effects began even before the
onset of the object, as illustrated by the significant effects of animacy on both subject and object
looks in the 1000 ms prior to the object’s onset. The opposing signs of these effects indicates that
animate objects both attracted looks themselves and drew looks away from the subjects—the
expected pattern given that the regions are mutually exclusive. These results confirm my earlier
qualitative assessment that inanimate objects received fewer looks than animate objects even
before participants heard the referent mentioned. Interestingly, in the model for subject looks in
the 1000 ms immediately following the object’s onset, the direction of the animacy effect
switched (remaining significant) from the previous time window. This result indicates that the
subject received more looks during this time period when the object was animate. This effect is
curious, since one would assume that increased competition due to the presence of multiple
animate referents would reduce the proportion of looks to the subject. I return to this effect in
Section 4.3.1. Animacy’s effect on object looks during this same time window was only
marginally significant, although in the final time window, animate objects again received
significantly more looks. There was no corresponding effect of animacy on subject looks in the
final time window.
Finally, there was one significant interaction of animacy and possession for subject looks
in the final time window. To explore this interaction further, I fit two follow-up models on
animate and inanimate subsets of the data. In the animate subset, possession was only marginally
significant (β = 0.029, SE = 0.016, df = 51.618, t = 1.806, p = 0.077), and in the inanimate
subset, there was no effect of possession (β = -0.011, SE = 0.014, df = 51.357, t = -0.751, p =
0.456).
120
In sum, the analysis of the gaze patterns during the matrix clause provides little evidence
in favor of any theory about the discourse-level effect of possession; however, the fact that these
data exhibit reasonable temporal dynamics and animacy effects contributes to the impression that
the methodology is sound.

4.2.5.6 Eye-movements during pronoun interpretation: pronoun-onset alignment
Eye-gaze patterns during pronoun interpretation were designed to be the most important
and informative domain of analysis. As described in Section 4.2.5.4, I analyzed these data with
two temporal alignments: the onset and offset of the critical pronoun (e.g. she in The runner
praised her nurse during the podcast interview because she was pambly). Visualizations of the
gaze patterns aligned to the onset of the pronoun are presented below as Figures 24 and 25.
Figure 24 shows looks to the three regions within each condition, while Figure 25 shows looks to
the subject and object across conditions.

121

Figure 24. Eye movements in Experiment 3 relative to the onset of the pronoun
(each facet shows looks to the three regions of interest within a condition)

Figure 25. Eye movements in Experiment 3 relative to the onset of the pronoun
(facets compare conditions in the subject and object regions)

122
For each of the temporal alignments, I created 1000 ms windows beginning 1000 ms
prior to the event until 4000 ms after it. In the same way as for the matrix-clause object
analysis, proportions of looks were calculated for the subject and object for each
combination of participant, trial, and analysis window. These proportions of looks then
served as dependent variables in the statistical models. Animacy and possession were
again the independent variables. The results of the models based on the pronoun onset
alignment are summarized in Table 27 (with significant effects shaded).

123
Region Fixed effect -1–0 s 0–1 s 1–2 s 2–3 s 3–4 s
Subject
(Intercept)
β = 0.115
SE = 0.010
df = 47.773
t = 11.666
p < 0.001
β = 0.282
SE = 0.016
df = 51.733
t = 17.296
p < 0.001
β = 0.474
SE = 0.021
df = 53.993
t = 22.704
p < 0.001
β = 0.534
SE = 0.021
df = 42.730
t = 25.935
p < 0.001
β = 0.487
SE = 0.034
df = 37.582
t = 14.544
p < 0.001
animacy
β = -0.015
SE = 0.010
df = 1125.150
t = -1.485
p = 0.138
β = -0.045
SE = 0.015
df = 1172.210
t = -3.111
p = 0.002
β = -0.189
SE = 0.018
df = 1169.365
t = -10.711
p < 0.001
β = -0.251
SE = 0.028
df = 44.809
t = -9.026
p < 0.001
β = -0.289
SE = 0.053
df = 30.941
t = -5.431
p < 0.001
possession
β = 0.019
SE = 0.010
df = 53.274
t = 1.834
p = 0.072
β = 0.017
SE = 0.015
df = 235.155
t = 1.122
p = 0.263
β = 0.008
SE = 0.018
df = 1006.643
t = 0.453
p = 0.651
β = -0.015
SE = 0.024
df = 773.883
t = -0.614
p = 0.539
β = -0.027
SE = 0.045
df = 265.474
t = -0.612
p = 0.541
animacy:
possession
β = 0.026
SE = 0.020
df = 1126.098
t = 1.323
p = 0.186
β = 0.023
SE = 0.029
df = 1173.387
t = 0.774
p = 0.439
β = 0.078
SE = 0.035
df = 1170.659
t = 2.216
p = 0.027
β = 0.165
SE = 0.048
df = 777.418
t = 3.421
p < 0.001
β = 0.289
SE = 0.090
df = 265.495
t = 3.226
p = 0.001
Object
(Intercept)
β = 0.227
SE = 0.018
df = 52.444
t = 12.955
p < 0.001
β = 0.132
SE = 0.010
df = 42.220
t = 13.580
p < 0.001
β = 0.126
SE = 0.010
df = 55.120
t = 12.765
p < 0.001
β = 0.157
SE = 0.012
df = 47.593
t = 12.591
p < 0.001
β = 0.210
SE = 0.027
df = 39.600
t = 7.871
p < 0.001
animacy
β = 0.067
SE = 0.013
df = 19.930
t = 5.136
p < 0.001
β = 0.060
SE = 0.013
df = 21.460
t = 4.605
p < 0.001
β = 0.151
SE = 0.011
df = 1201.217
t = 13.343
p < 0.001
β = 0.217
SE = 0.019
df = 838.846
t = 11.643
p < 0.001
β = 0.272
SE = 0.041
df = 368.968
t = 6.694
p < 0.001
possession
β = 0.002
SE = 0.013
df = 55.944
t = 0.177
p = 0.860
β = -0.015
SE = 0.011
df = 1162.913
t = -1.352
p = 0.177
β = -0.005
SE = 0.011
df = 1201.533
t = -0.402
p = 0.688
β = -0.007
SE = 0.019
df = 819.849
t = -0.353
p = 0.724
β = -0.011
SE = 0.040
df = 357.879
t = -0.282
p = 0.778
animacy:
possession
β = -0.047
SE = 0.026
df = 1116.794
t = -1.822
p = 0.069
β = -0.019
SE = 0.022
df = 1161.738
t = -0.856
p = 0.392
β = -0.034
SE = 0.023
df = 1200.821
t = -1.491
p = 0.136
β = -0.049
SE = 0.037
df = 816.931
t = -1.315
p = 0.189
β = -0.137
SE = 0.080
df = 353.772
t = -1.713
p = 0.088

Table 27. Summary of fixed effects for the models predicting looks to the subject
and object regions at time intervals aligned to the onset of the pronoun
(Experiment 3); significant effects are shaded

As predicted, the models showed strong effects of animacy. Beginning at the onset of the
pronoun, all subsequent time windows showed a significant decrease in looks to the
subject when the object was animate, which suggests that the object competed for
124
prominence with the subject when it was a grammatical antecedent for the pronoun (i.e.
when the object was also animate). Additionally, there was a complementary set of
animacy effects for subject and object looks which indicate that animate objects received
more looks inanimate ones. For the object, animacy also significantly increased the
proportion of looks during the 1000 ms window prior to the onset of the pronoun. This
pattern is not surprising, however, given that animate objects were already receiving
more looks than inanimates during the final analysis window after the matrix-clause
object (see Section 4.2.5.5).
There were no significant main effects of possession, but the models did reveal
interactions between animacy and possession. For subject looks, there were significant
interactions in the 1000–2000 ms, 2000–3000 ms, and 3000–4000 ms windows.
Interestingly, there were no corresponding significant interactions in the models for
object looks, and only the -1000–0 ms and 3000–4000 ms windows were marginally
significant.
To further investigate these interactions, follow-up models were fit to test for
simple effects of possession within the animate conditions. Regardless of whether there
was an interaction in the larger model, I fit follow-up models for looks to both the subject
and the object for every analysis window, which I justify for multiple reasons. Firstly,
comparison of the definites and possessives within the animate conditions was planned
during the experiment’s design. Because the pronouns were only ambiguous in the
animate conditions, my predictions of the gaze patterns during pronoun resolution were
specifically oriented toward this subset. Secondly, as demonstrated in the selection
latency data (see Section 4.2.5.2 and Figure 20), pronoun resolution was significantly
125
faster in the unambiguous inanimate conditions, and thus gaze data for the inanimate
conditions was sparser in the later analysis windows. This asymmetry could have
contributed to either the presence or absence of interactions in the larger models, and thus
fitting follow-up models to test for simple main effects in the animate conditions was
necessary to fully address the predictions of the experiment. A summary of the results of
these models is given below as Table 28.

Region Fixed effect -1–0 s 0–1 s 1–2 s 2–3 s 3–4 s
Subject
(Intercept)
β = 0.108
SE = 0.010
df = 33.391
t = 10.892
p < 0.001
β = 0.260
SE = 0.016
df = 37.729
t = 15.884
p < 0.001
β = 0.380
SE = 0.025
df = 46.390
t = 15.339
p < 0.001
β = 0.408
SE = 0.025
df = 37.346
t = 16.229
p < 0.001
β = 0.342
SE = 0.029
df = 27.738
t = 11.887
p < 0.001
possession
β = 0.033
SE = 0.016
df = 51.321
t = 2.035
p = 0.047
β = 0.027
SE = 0.020
df = 552.257
t = 1.314
p = 0.189
β = 0.045
SE = 0.027
df = 53.535
t = 1.664
p = 0.102
β = 0.068
SE = 0.033
df = 50.914
t = 2.024
p = 0.048
β = 0.120
SE = 0.042
df = 163.710
t = 2.869
p = 0.005
Object
(Intercept)
β = 0.260
SE = 0.018
df = 39.232
t = 14.415
p < 0.001
β = 0.162
SE = 0.013
df = 30.551
t = 12.974
p < 0.001
β = 0.201
SE = 0.018
df = 47.535
t = 11.208
p < 0.001
β = 0.266
SE = 0.021
df = 36.049
t = 12.611
p < 0.001
β = 0.344
SE = 0.029
df = 31.272
t = 11.836
p < 0.001
possession
β = -0.020
SE = 0.021
df = 50.784
t = -0.962
p = 0.341
β = -0.024
SE = 0.017
df = 560.760
t = -1.373
p = 0.170
β = -0.021
SE = 0.021
df = 36.328
t = -0.980
p = 0.334
β = -0.032
SE = 0.032
df = 49.670
t = -1.005
p = 0.320
β = -0.080
SE = 0.041
df = 257.912
t = -1.941
p = 0.053

Table 28. Summary of fixed effects for the follow-up models predicting looks to
the subject and object in the animate conditions (aligned to the onset of the
pronoun, Experiment 3); significant effects are shaded

For looks to the subject, the follow-up models revealed significant effects of possession
within the animate conditions in multiple time windows. When an animate object was
possessed, the subject received significantly more looks in the 1000 ms preceding the
pronoun, as well as during the final two time windows from 2000 to 4000 ms after the
pronoun’s onset. There were no significant effects of possession on object looks in the
126
animate conditions, although the final time window contained a marginal effect of
possession in the expected direction, whereby possession decreased the proportion of
looks to animate objects during pronoun resolution.
While the significant effect prior to the pronoun’s onset may represent lingering
processing of the matrix clause, the two significant time windows from 2000 to 4000 ms
after the pronoun’s onset support the Parasitic Possession Hypothesis. While I did not
find a reduction in looks to possessed objects, because prominence is always relative to
the referents in the discourse, a boost in prominence for the subject when the object is
possessed is also expected under the Parasitic Possession Hypothesis. By this logic, the
reduced prominence of the possessed referent conferred increased prominence on the
subject, making it more available as an antecedent and driving more looks to it during
pronoun resolution

4.2.5.7 Eye-movements during pronoun interpretation: pronoun-offset alignment
As discussed in Section 4.2.5.4, a systematic difference in the durations of the
masculine and feminine pronouns caused me to consider aligning the gaze patterns to the
pronoun’s offset. This analysis may be more reliable than the onset alignment, since
looking patterns triggered by the masculine and feminine pronouns may be better aligned.
Additionally, because the clause containing the pronoun was the same within items, the
pronoun-offset alignment is unlikely to introduce spurious differences between
conditions. It is worthy of note that shifting the temporal alignment of the gaze patterns
may change the time at which one condition diverges from another; however, this study’s
127
predictions relate to differences between conditions during pronoun resolution in general
and are less concerned with the fine-grained timecourse of how those differences arise.
With these arguments in mind, I repeated essentially the same analysis as above for the
pronoun offset alignment. Visualizations of the gaze patterns are given below as Figures 26 and
27, which again present the data first by condition and then across conditions for the subject and
object regions.

Figure 26. Eye movements in Experiment 3 relative to the offset of the pronoun
(each facet shows looks to the three regions of interest within a condition)

128

Figure 27. Eye movements in Experiment 3 relative to the offset of the pronoun
(facets compare conditions in the subject and object regions)

I performed the same statistical analyses for these data as for the pronoun onset alignment in
Section 4.2.5.6. A summary of the models’ fixed effects is given below as Table 29.

129
Region Fixed effect -1–0 s 0–1 s 1–2 s 2–3 s 3–4 s
Subject
(Intercept)
β = 0.125
SE = 0.010
df = 48.439
t = 12.078
p < 0.001
β = 0.314
SE = 0.017
df = 52.078
t = 18.079
p < 0.001
β = 0.499
SE = 0.022
df = 55.552
t = 23.134
p < 0.001
β = 0.531
SE = 0.021
df = 55.259
t = 25.225
p < 0.001
β = 0.470
SE = 0.038
df = 37.518
t = 12.517
p < 0.001
animacy
β = -0.017
SE = 0.013
df = 23.167
t = -1.299
p = 0.207
β = -0.056
SE = 0.015
df = 1129.799
t = -3.608
p < 0.001
β = -0.212
SE = 0.018
df = 1168.049
t = -11.930
p < 0.001
β = -0.271
SE = 0.026
df = 743.854
t = -10.306
p < 0.001
β = -0.274
SE = 0.059
df = 31.628
t = -4.637
p < 0.001
possession
β = 0.017
SE = 0.010
df = 1104.999
t = 1.635
p = 0.102
β = 0.015
SE = 0.016
df = 54.604
t = 0.937
p = 0.353
β = 0.002
SE = 0.018
df = 1163.851
t = 0.132
p = 0.895
β = -0.004
SE = 0.027
df = 25.348
t = -0.158
p = 0.875
β = -0.038
SE = 0.049
df = 227.502
t = -0.770
p = 0.442
animacy:
possession
β = 0.034
SE = 0.021
df = 1105.059
t = 1.650
p = 0.099
β = 0.028
SE = 0.031
df = 1130.589
t = 0.893
p = 0.372
β = 0.084
SE = 0.036
df = 1169.253
t = 2.357
p = 0.019
β = 0.162
SE = 0.052
df = 726.709
t = 3.100
p = 0.002
β = 0.300
SE = 0.098
df = 229.035
t = 3.061
p = 0.002
Object
(Intercept)
β = 0.197
SE = 0.016
df = 54.362
t = 12.462
p < 0.001
β = 0.131
SE = 0.009
df = 40.275
t = 14.066
p < 0.001
β = 0.125
SE = 0.010
df = 54.962
t = 11.953
p < 0.001
β = 0.165
SE = 0.013
df = 68.463
t = 12.831
p < 0.001
β = 0.218
SE = 0.031
df = 43.921
t = 6.967
p < 0.001
animacy
β = 0.058
SE = 0.012
df = 1113.871
t = 4.687
p < 0.001
β = 0.078
SE = 0.011
df = 1184.141
t = 6.975
p < 0.001
β = 0.160
SE = 0.012
df = 1200.156
t = 13.782
p < 0.001
β = 0.238
SE = 0.028
df = 59.469
t = 8.604
p < 0.001
β = 0.262
SE = 0.047
df = 332.766
t = 5.637
p < 0.001
possession
β = -0.001
SE = 0.016
df = 20.555
t = -0.032
p = 0.975
β = -0.012
SE = 0.011
df = 1184.199
t = -1.102
p = 0.270
β = -0.005
SE = 0.012
df = 1200.348
t = -0.433
p = 0.665
β = 0.002
SE = 0.020
df = 728.139
t = 0.126
p = 0.899
β = -0.010
SE = 0.046
df = 320.182
t = -0.214
p = 0.831
animacy:
possession
β = -0.060
SE = 0.025
df = 1114.029
t = -2.426
p = 0.015
β = -0.020
SE = 0.022
df = 1184.516
t = -0.916
p = 0.360
β = -0.021
SE = 0.023
df = 1199.665
t = -0.926
p = 0.355
β = -0.074
SE = 0.039
df = 728.495
t = -1.879
p = 0.061
β = -0.164
SE = 0.091
df = 320.321
t = -1.804
p = 0.072

Table 29. Summary of fixed effects for the models predicting looks to the subject
and object at various time intervals aligned to the offset of the pronoun
(Experiment 3); significant effects are shaded

Reassuringly, all the fixed effects which were significant for the pronoun-onset alignment were
also significant in the same time windows for the pronoun-offset alignment. Additionally, there
was one additional significant interaction of animacy and possession on looks to the object
130
during the 1000 ms prior to the pronoun’s offset. While only the final time window contained a
marginal interaction for object looks with the pronoun-onset alignment, both the 2000–3000 ms
and 3000–4000 ms windows were marginally significant with the pronoun-offset alignment.
I again fit follow-up models to test for differences between the definite and possessed
conditions when the object was animate. The fixed effects in these models are summarized
below in Table 30.

Region Fixed effect -1–0 s 0–1 s 1–2 s 2–3 s 3–4 s
Subject
(Intercept)
β = 0.117
SE = 0.010
df = 34.096
t = 11.218
p < 0.001
β = 0.287
SE = 0.017
df = 39.571
t = 16.570
p < 0.001
β = 0.393
SE = 0.026
df = 46.835
t = 15.052
p < 0.001
β = 0.396
SE = 0.025
df = 35.872
t = 15.669
p < 0.001
β = 0.331
SE = 0.032
df = 26.408
t = 10.513
p < 0.001
possession
β = 0.034
SE = 0.016
df = 51.741
t = 2.179
p = 0.034
β = 0.027
SE = 0.022
df = 52.540
t = 1.250
p = 0.217
β = 0.042
SE = 0.027
df = 53.377
t = 1.565
p = 0.123
β = 0.076
SE = 0.034
df = 50.868
t = 2.205
p = 0.032
β = 0.112
SE = 0.043
df = 184.969
t = 2.590
p = 0.010
Object
(Intercept)
β = 0.226
SE = 0.016
df = 38.090
t = 14.106
p < 0.001
β = 0.170
SE = 0.013
df = 31.737
t = 13.490
p < 0.001
β = 0.204
SE = 0.019
df = 47.547
t = 10.751
p < 0.001
β = 0.283
SE = 0.022
df = 34.229
t = 12.920
p < 0.001
β = 0.350
SE = 0.031
df = 30.378
t = 11.314
p < 0.001
possession
β = -0.030
SE = 0.022
df = 19.927
t = -1.327
p = 0.200
β = -0.022
SE = 0.018
df = 561.176
t = -1.212
p = 0.226
β = -0.015
SE = 0.021
df = 56.721
t = -0.708
p = 0.482
β = -0.035
SE = 0.029
df = 431.508
t = -1.187
p = 0.236
β = -0.091
SE = 0.044
df = 236.437
t = -2.067
p = 0.040

Table 30. Summary of fixed effects for the follow-up models predicting looks to
the subject and object in the animate conditions (aligned to the offset of the
pronoun, Experiment 3); significant effects are shaded

For subject looks, the results of these models were qualitatively the same as the corresponding
pronoun onset analysis. There were significant effects of possession during the 1000 ms
preceding the offset of the pronoun and during the final two time windows from 2000 ms to 4000
ms. Notably, the pronoun-offset alignment yielded a significant effect of possession on object
looks during the final time window, whereas this effect was only marginal with the onset
131
alignment. The directionality of the effects in these follow-up models was again consistent with
the Parasitic Possession Hypothesis; possession of the object seems to decrease its prominence
and increase the prominence of the subject. This effect made the object less available as an
antecedent and drove more looks to the subject during resolution of the ambiguous pronoun.

4.3. Discussion
Experiment 3 investigated how possession affects the discourse representation of animate
and inanimate referents and how differences in discourse representation might affect referents’
accessibility as pronominal antecedents. The central research aim in this study was to further
investigate the Parasitic Possession Hypothesis (i.e., that possession reduces a referent’s
discourse prominence), which was supported by the results of Experiment 2. Additionally, I
sought to assess a competing theory, the Complexity Hypothesis, for which there was more
limited support in the results of Experiment 1. In opposition to the Parasitic Possession
Hypothesis, the Complexity Hypothesis claims that possessed referents are more prominent in
discourse than simple nominals.
While multiple dependent measures were analyzed in Experiment 3 (i.e. selections of the
last-mentioned referent, selection latencies, and eye-gaze patterns), eye-gaze patterns—
particularly during pronoun resolution—were the most important to my research aims.
Participants’ looking behavior during pronoun resolution (e.g. after hearing he in The soldier
heard his plumber throughout the autumn evening because he was nimpacious) showed that
when possessed objects were potential antecedents for ambiguous pronouns, they received fewer
looks than their definite counterparts. Conversely, subjects which were the possessors of those
objects received more looks during ambiguous pronoun resolution than non-possessor subjects.
132
This pattern provides evidence that the possessed objects were less accessible as pronominal
antecedents and therefore less prominent in discourse. Consequently, these results support the
Parasitic Possession Hypothesis.
The other dependent measures in Experiment 3 neither supported nor challenged my
theoretical proposals related to possession; however, they displayed some sensible animacy
effects which inform the other analyses and provide evidence of the task’s validity. The gaze
patterns following the matrix-clause object were particularly important in this regard, since they
illustrate (independently of the primary pronoun analysis) that webcam-based eye-tracking, a
rapidly emerging methodology, can produce reasonable data.

4.3.1 Effects of animacy
Experiment 3 was not primarily intended to investigate effects of animacy, and the
animacy effects found in the data are largely orthogonal to hypotheses about the discourse-level
effects of possession; however, given the appearance of animacy effects in Experiments 1 and 2
and the robustness of animacy effects in discourse processing in general, it is not surprising that
animacy was a significant influence on the data patterns in every domain of analysis for
Experiment 3.
In the selection data, inanimate objects were almost never chosen as the last-mentioned
referent, i.e. the antecedent of the pronoun she or he (for which an inanimate would be an
ungrammatical antecedent), whereas animate objects were often selected over animate subjects
(although subjects were still the preferred choice). This pattern leads to two conclusions: firstly,
participants understood the task (whether implicitly or explicitly) as pronoun resolution and
successfully rejected the ungrammatical antecedent in inanimate conditions; secondly, at least in
133
offline selections, animate objects did compete with subjects as viable antecedents for the
ambiguous pronoun. Similarly, the longer selection latencies in the animate conditions speak to
increased competition among referents when the pronoun was ambiguous and confirm previous
findings that ambiguous pronouns require more cognitive resources to resolve than unambiguous
ones (e.g. Stewart et al., 2007).
In the eye-tracking data, animacy exerted an influence on the gaze patterns during
processing of both the matrix-clause object and the pronoun (e.g. nurse and she respectively in
The runner praised her nurse during the podcast interview because she was pambly). At the
matrix-clause object, the general pattern of results was that animate objects received more looks
and drew eyes off the subject more often than inanimate objects did. This finding is consistent
with previous work showing that the presence of multiple animate entities increases competition
for discourse prominence (e.g. Ariel, 1990; Arnold & Griffin, 2007; Chiriacescu, 2015;
Fukumura & van Gompel, 2011). Likewise, at the pronoun, the predicted animacy effects were
observed, as the presence of an ungrammatical antecedent in the inanimate conditions (i.e. the
inanimate object) resulted in significantly more looks to the subject.
Nevertheless, there was one notable instance where these canonical animacy patterns
were reversed. In the gaze patterns from 0 to 1000 ms after the onset of the matrix-clause object,
there were significantly more looks to the subject when the object was animate versus inanimate.
This pattern seems strange at first, since one would expect increased competition for prominence
to lower the proportion of looks to the subject. An explanation for this reversal can be found by
examining the overall shape of the gaze curves for the matrix-clause object (see Figures 22 and
23). In the 1000 ms preceding the onset of the object, one can observe that looks to the subject
reach a local maximum for both the animate and inanimate conditions; however, the peak in
134
looks to the subject occurs earlier when the object is inanimate. This pattern makes sense in light
of the visual search procedure which participants undertake upon processing the sentence
subject; for example, when the subject is princess and the object is television, it should be easy
for participants to locate the princess by using the basic visual feature of personhood.
Furthermore, animates’ advantage in capturing visual attention (e.g. Downing et al., 2004; New
et al., 2007) and the linguistic bias toward animate subjects (e.g. Clark & Begun, 1971; Dahl,
2008; Itagaki & Prideaux, 1985) are additional factors which make it easier for participants to
recognize the intended referent after they hear princess. On the other hand, when the subject is
princess and the object is attorney, participants must utilize more fine-grained visual features in
order to find the image which was mentioned. Thus it seems reasonable that ambiguity would
persist longer, and the visual search would require more cognitive resources to carry out. Given
that the peak in looks to the sentence subject is delayed in animate conditions, it follows that the
shift in visual attention to the object would also begin later. Conversely, because the peak in
looks to the subject occurs earlier in the inanimate conditions, participants can move their eyes to
the object sooner after it is mentioned. This lag is then the source of the reversed animacy effect
in the single time window from 0 to 1000 ms after the object onset, whereby the presence of an
animate object results in more looks to the subject.

4.3.2 Effects of possession: support for the Parasitic Possession Hypothesis
Experiment 3 was designed principally to detect effects of possession during pronoun
resolution. In general, analysis of the gaze patterns at the pronoun supported the Parasitic
Possession Hypothesis; in the animate conditions—where the pronoun was ambiguous—
possession reduced looks to the object (i.e. the possessed referent) and boosted looks to the
135
subject (i.e. the possessor). The gaze patterns are consistent with the conclusion that possession
decreases prominence and therefore makes the possessed referent less available as an antecedent.
These effects occurred fairly late during pronoun resolution, with the earliest significant time
window in any of the applicable models beginning 2000 ms after the onset of the pronoun. While
these effects surfaced slowly, it should be clear based on their timing that they are attributable to
pronoun interpretation, rather than some earlier linguistic processing. Additionally, the fact that
these effects were only significant in the later time windows might be due to the inherent
noisiness of webcam-based eye-tracking in combination with the limited statistical power of this
dataset. With additional power or less noisy data, these effects might have appeared earlier
during pronoun resolution. Interestingly, this decrease in prominence shown by the eye-gaze
patterns did not translate into a lower proportion of selections of possessed animates as the
antecedent of the ambiguous pronoun; however, it seems likely that this offline selection
measure is not sensitive enough to detect the modulation in discourse prominence produced by
possession.
In discussing the results of Experiment 2, I considered the possibility that possession
might decrease the prominence of the possessed referent simply because possessives inherently
contain an additional mention of the possessor. It would follow logically that whatever boost in
prominence the possessor receives from this extra mention would come at the expense of the
prominence of the possessed referent. In Experiment 2, I discounted this explanation because the
effect of possession observed in the sentence continuation data pertained specifically to the
possessed object, not the possessor subject. In Experiment 3, the evidence for this alternative
theory is perhaps stronger. While the eye-gaze patterns in Experiment 3 demonstrated both
reductions in prominence for possessed referents and complementary increases in prominence
136
for possessor subjects, the effects observed on subject looks were detected in earlier time
windows, were numerically stronger, and were less sensitive to the choice of temporal alignment
(i.e. onset versus offset of the pronoun). Thus it seems more difficult in this case to reject the
argument that there is nothing special about possession per se, but rather the additional mention
of the subject in the form of the possessive pronoun reactivates the subject’s discourse
representation and subsequently makes it more accessible to the pronoun.
Nevertheless, the gaze patterns during the matrix clause provide a counterargument for
why the effects of possession in Experiment 3 are not reducible merely to an additional mention
of the subject in the possessed conditions. If the presence of the possessive pronoun in the
possessed conditions were responsible for increasing the prominence of the subject via
reactivation of its representation, we might expect to see more looks to the subject following the
possessive pronoun. This effect would likely surface as a main effect of possession, since
presumably the pronoun would reactivate the possessor regardless of whether the possession was
animate or inanimate.
Because the possessive pronoun immediately preceded the matrix-clause object, we can
check this prediction to some extent against the analysis of looks relative to the object onset.
Referring to the gaze data both preceding and following the object’s onset, there is no main
effect of possession. The only effect of possession in this set of models is an interaction of
animacy and possession in the final time window from 1000 to 2000 ms after the object’s onset;
however, this time window falls after the peak in looks to the object corresponding to its mention
in the sentence. Given that peaks in looks tend to follow the order of mention in visual-world
eye-tracking (e.g. Cooper 1974; Tanenhaus et al., 1995), it seems unlikely that the possessive
pronoun, which preceded the object, was itself driving more looks to the subject when looks to
137
the object were already subsiding. Furthermore, it is unclear why reactivation of the subject by
the possessive pronoun would produce an interaction of animacy and possession, which in this
case incorporated only a marginal simple main effect of possession in the animate conditions and
no corresponding effect in the inanimates. Thus it seems that the reduction in prominence for
possessed referents is not attributable simply to an additional mention of the possessor in the
form of the possessive pronoun.
Finally, while gaze patterns during pronoun interpretation support the Parasitic
Possession Hypothesis, both the time alignments display an interesting pattern in which the
effects show up more strongly for subject looks compared to object looks. By the basic
assumption that prominence is allocated relative to the referents present in the discourse, the
gaze patterns ought to show complementary effects. One possible explanation for this asymmetry
might be that because the subject is by default more prominent than the object, modulations in
the subject’s prominence may be easier to detect. Thus, while the true reduction in prominence
for possessed referents may be complementary to an increase in prominence for other referents,
this experiment may have been less sensitive to changes in looks to the object. Perhaps with
greater statistical power or more accurate eye-tracking methods (i.e. laboratory eye-tracking), a
more symmetrical set of effects would have surfaced for the subject and object looks.
In summary, the results of Experiment 3 support the Parasitic Possession Hypothesis,
which claims that possession reduces the prominence of possessed referents.

138
Chapter 5: Effects of possession type on the comprehension of ambiguous verb-phrase
ellipsis

5.1. Introduction
Up until this point, I have considered two sets of competing hypotheses related to the
discourse-level consequences of possession. The first set deals with the overall effect of
possession on referents, while the second asks whether the type of semantic relation between the
possessor and the possessed referent modulates possession’s overall effect. Specifically,
regarding this second set of hypotheses, I have examined the Interaction Hypothesis, which
predicts an extra boost in prominence for possessed animate referents, potentially because of the
social salience of interpersonal relationships, which possessed animates denote (see Section 2.3
for more discussion of this topic). Alternatively, I also formulated the Animacy Hypothesis,
which states that possession’s effect on the prominence of referents does not vary with the type
of semantic possession relation. Instead, the Animacy Hypothesis predicts that the difference in
prominence between possessed animates and possessed inanimates will be equivalent to the
difference between animate and inanimate simple nominals. Essentially the Animacy Hypothesis
posits that any difference observed between possessed animates and inanimates is reducible to
animates’ well-known advantage in prominence (e.g. Arnold & Griffin, 2007; Bock et al., 1992;
Branigan et al., 2008; Dahl & Fraurud, 1996) and does not represent an effect of the possession
relation itself.
So far, the results I have presented related to this second set of hypotheses have been
mixed. In Experiment 1, I found an extra boost in prominence for possessed animates, which
supported the Interaction Hypothesis. On the other hand, Experiment 2 found no such
139
interaction. This same sort of interaction was also absent in the results of Experiment 3, although
the design of this experiment made it less applicable to the Interaction Hypothesis than the prior
two. All three of these previous experiments used methodologies which have frequently been
associated with the investigation of discourse-level effects (i.e. sentence continuation and visual-
world eye-tracking). In this chapter, I return to the question of whether possession type
modulates discourse prominence, but I consider this issue through the lens of a seemingly
orthogonal topic: the interpretation of ambiguous possessive pronouns in verb-phrase ellipsis
(VPE). In this chapter, I show how patterns of ambiguity resolution in this construction can shed
light on my hypotheses about the discourse-level effects of possession. Additionally, I expect
that insights from this chapter, in combination with my other findings, can inform future inquiry
into VPE and anaphora in general—to the extent that researchers in these fields integrate
possessives into their theories.

5.2. Ambiguous possessive pronouns in verb-phrase ellipsis
The type of VPE I examine in this chapter is given in Example (1a). Such sentences
contain possessive pronouns which at some level of representation are copied to the ellipsis site
and give rise to an ambiguity regarding the interpretation of the elided clause. This ambiguity is
the result of two different mechanisms for pronoun interpretation: discourse-level coreference
and semantic-level binding (e.g. Heim, 1982; Reuland, 2001).
(1) a. Bill washed his car, and John did, too.
b. Bill
i
washed his
i
car, and John
j
did [
VP
washed his
j
car], too.
c. Bill
i
washed his
i
car, and John
j
did [
VP
washed his
i
car], too.
d. Bill
i
washed his
k
car, and John
j
did [
VP
washed his
k
car], too.
140
Example (1b) is an instance of variable binding, in which the possessive pronoun his is bound by
Bill in the first clause but John in the second (also called “sloppy identity”). Examples (1c) and
(1d), on the other hand, are instances of coreference, in which the possessive pronoun his has the
same referent (Bill or a discourse-relevant third party) in the antecedent verb phrase as it does in
the elided constituent (also called “strict identity”). The interpretation in (1d) necessitates the
presence of a salient third party in the discourse, and, without such a referent available, this
interpretation is unlikely.
Prior research has demonstrated that multiple factors affect the resolution of such
ambiguities. Firstly, many researchers have argued that the bound variable interpretation of this
construction is computationally easier than coreference (e.g. Foley et al., 2003; Frazier &
Clifton, 2000; Koorneef et al., 2011, Reuland, 2001, Vasić et al., 2006). These claims depend on
theories of differing computational demands in coreference and binding. One such theory, the
Primitives of Binding framework (Reuland, 2001) posits an economy hierarchy of pronominal
reference assignment, where operations taking place at the level of semantics (i.e. variable
binding) are less costly than operations at the level of discourse (i.e. coreference). This outlook is
supported by evidence from studies of VPE interpretation in children (e.g. Foley et al., 2003;
Guo et al., 1996), healthy adults (e.g. Frazier & Clifton, 2000; Koorneef et al., 2011; Reuland,
2001), and patients with aphasia (e.g. Grodzinsky et al., 1993; Vasić et al., 2006).
To a lesser extent, the lexical semantics of verbs and possessed nouns in ambiguous VPE
have also been shown to influence ambiguity resolution. For instance, Ong & Brasoveanu (2014)
found that a verb’s implicit causality modulated the strength of bound variable preference.
Additionally, Foley et al. (2003) found in work with children that sentences where the object
noun was inalienably possessed by the subject (e.g. Big Bird scratches his arm, and Ernie does,
141
too) resulted in more bound variable interpretations than comparable sentences with alienable
possessions (e.g. Big Bird moves his apple, and Ernie does, too). They found a related effect for
self-directed verbs (e.g. scratch), which were more biased toward bound variable interpretation
than other-directed verbs (e.g. move).
There may also be reason to believe that animacy could affect interpretational preferences
in ambiguous VPE. To the best of my knowledge, there is no prior experimental work on the
effects of animacy in guiding this type of ellipsis resolution; however, work by Dahl & Fraurud
(1996) suggests that animacy can influence the choice between bound variable and coreferential
interpretations of non-elided sentences like Examples (2a) and (2b).
(2) a. John sent his paycheck to his mother, and Bill sent it to his wife.
b. John sent his daughter to his mother, and Bill sent her to his wife.
They claim that (2a), in which the possessed noun paycheck is inanimate, is biased toward bound
variable interpretation (i.e. that Bill sent his own paycheck to his wife); whereas (2b), in which
the possessed noun daughter is animate and human, is biased toward coreferential interpretation
(i.e. that Bill sent John’s daughter to his wife). They speculate that these contrasting intuitions
might “[have] to do with differences in individuation between animates and inanimates” (Dahl &
Fraurud, 1996, p. 57). Nevertheless, the judgments they present are not supported by
experimental evidence, nor do they quantify the extent to which animacy influences the relative
bound variable biases of Examples (2a) and (2b).

5.3. Connecting verb-phrase ellipsis to possession at the discourse level
Before discussing the specifics of this chapter’s experiments, it is necessary to lay out
which of the hypotheses discussed in earlier chapters can be addressed in an examination of
142
possessive pronouns in VPE. My first set of hypotheses (i.e. the Complexity Hypothesis and the
Parasitic Possession Hypothesis), which relate to the overall effect of possession on prominence,
necessarily rely on a comparison of possessives to some other class of nominal. Specifically,
Experiment 1 compared possessives to indefinites, while Experiments 2 and 3 compared
possessives to definites. In the case of the ambiguous ellipsis discussed in Section 5.2 of this
chapter, no such comparison is possible. To illustrate this point, if one changes the possessive
pronoun in Example (1a) to a definite article (Bill washed the car, and John did, too), the result
is a fully unambiguous ellipsis with no option of variable binding. Thus, this chapter cannot
confirm or contradict my previous findings in favor of the Parasitic Possession Hypothesis
(specifically in Chapters 2 and 3).
On the other hand, this chapter does have the potential to inform debate regarding my
second set of hypotheses (i.e. the Interaction Hypothesis and the Animacy Hypothesis), which
consider whether the type of possession relation affects prominence. Particularly in Experiments
1 and 2, consideration of these hypotheses relied on the presence of absence of a statistical
interaction between animacy and possession. As I mention above, there is no way of
manipulating possession in an investigation of this type of VPE; however it is still possible to
design an experiment which can inform these hypotheses by introducing additional types of
possession relations beyond just the ownership and interpersonal relations tested in the previous
experiments. In fact, it is only possible to test some types of possession relations in an
experiment which does not compare possessives to simple nominals. For instance, with part-
whole (e.g. her hands) and kinship (e.g. her father) relations, the distribution of corresponding
non-possessed nominals can be significantly restricted (e.g. Mary washed her/#the hands, Mary
loves her/#the father). The experiments in this chapter contain an animacy manipulation similar
143
to the previous experiments but also test multiple types of possession relations which arise from
animate and inanimate possessed referents (with animate possessors). By looking for differences
within the animate and inanimate possession relations, I can address the fundamental question
asked by the second set of hypotheses: does possession type influence a referent’s prominence?
Finally, it is worthy of some explanation how I can connect ellipsis interpretation with
the primary theoretical dimension considered in this dissertation, which is discourse prominence.
The topic of VPE has historically been analyzed more at the syntax-semantics level (e.g. Fiengo
& May, 1994; Sag, 1976), and thus it might be surprising to connect its interpretation with
prominence. Nevertheless, discourse-based approaches to ellipsis interpretation do exist (e.g.
Asher et al., 2001; Hardt & Romero, 2004; Kertz, 2013), and it is commonly assumed that
coreferential interpretations like Examples (1c) and (1d) require access to discourse-level
representations (e.g. Heim, 1982; Reinhart, 2000; Reuland, 2001). On this basis, the linking
hypothesis which I adopt in this chapter is that the more prominent the possessed referent is in
sentences like Example (1a) (i.e. car in Bill washed his car, and John did, too), the more likely
that referent is to be used in the computation of a coreferential interpretation like Example (1c)
(i.e. that John washed Bill’s car). On the other hand, if the possessed referent is not particularly
prominent, it will be more difficult to use this referent for coreference, and thus the
computationally simpler bound-variable interpretation will be preferred (i.e. that John washed his
own car). This logic aligns with the central assumption of the sentence continuation paradigm
used in Experiments 1 and 2: that more prominent referents have a higher likelihood of mention
in subsequent discourse (e.g. Arnold, 2001; Givón, 1983; Kaiser, 2009; Kehler et al., 2008;
Kehler & Rohde, 2013). Furthermore, it fits with the theory that discourse prominence generally
supports computationally costly grammatical operations (e.g. von Heusinger & Schumacher,
144
2019), of which coreference can be considered one type. Finally, if the interpretational
preferences reported in Dahl & Fraurud (1996) are correct, and animate possessed referents do
lead to more coreferential interpretations in non-elided sentences like Example (2b), it would
make sense for animates’ general advantage in discourse prominence over inanimates to be
driving this pattern.

5.4. Experiment 4
I have multiple aims in considering ambiguous possessive pronouns in VPE. Firstly, as
described in Section 5.3, I intend to use VPE as an alternative domain in which to assess
hypotheses about whether possession type modulates the overall discourse-level effect of
possession. While Experiments 2 and 3 did not find effects attributable to possession type itself,
Experiment 1 did (supporting the Interaction Hypothesis), and thus this topic deserves additional
investigation. Experiment 4 provides a potentially greater opportunity to observe how possession
type influences discourse-level processing by testing a broader range of possession relations than
I considered in Experiments 1, 2, and 3.
Secondly, the role that possession type plays in resolving the ambiguity in sentences like
Example (1a) is theoretically interesting in its own right and understudied. Foley et al. (2003)
and Guo et al. (1996), a comparable study on children acquiring Mandarin, provide a helpful
starting point, but do not fully address the role of possession type, since the possessions in their
studies were primarily body parts and owned objects, which had to be conceptually simple and
easily imageable due to the task constraints of working with children. Furthermore, these
acquisition studies did not manipulate animacy by testing animate possessions, which, given the
judgments in Dahl & Fraurud (1996), seems like a factor that could matter for ellipsis resolution.
145
While Foley et al. and Guo et al. implicate alienability as a driver of ambiguity resolution, it is an
empirical question whether this pattern will hold in a study of adult ellipsis comprehension that
integrates a broader range of possession relations.
Lastly, given the prevalence of possessive pronouns in test cases within the semantic
binding literature, it is surprising how little attention has been paid to the contribution of the
possession relation. Researchers in this subfield have built theories around judgements of
sentences containing ambiguous possessive pronouns, but often qualities of the possessed
referent are not well controlled (e.g. Fiengo & May, 1994; Kitagawa, 1991; Oku, 1998; Sag,
1976). Many examples test possessives with kinship terms (e.g. his mother, her son), but these
vary with other animate relational nouns like friend, as well as possessions like car, among
others. In many cases, researchers in this subfield are less interested in interpretational
preferences in these types sentences than in which interpretations are simply possible for the
grammar to compute. Nevertheless, some theories have been built around degraded grammatical
acceptability: for instance, the supposedly marginal status of the bound variable interpretation of
Mary washed her car, and John did, too due to gender mismatch between the overt and elided
possessive pronoun (e.g. Kitagawa, 1991; Oku, 1998; see Section 5.5.2 for more on this topic).
This research has not examined the potential influence of the semantics of car on such
judgments. Accordingly, by emphasizing and quantifying the influence that the semantic
possession type has on the interpretation of possessive pronouns, this research is potentially
informative for future work on binding theory and ellipsis which incorporates possessives.
To accomplish these aims, Experiment 4 uses a forced-choice ellipsis-interpretation task
to test whether and how different possession relations modulate participants’ preferences for
bound-variable versus coreferential interpretations. I test four types of possessed referents which
146
give rise to different possession relations: alienably owned inanimates (e.g. bicycle, computer,
shoes), inalienable part-whole nouns (e.g. nose, eyes, reputation), inalienable kinship terms (e.g.
aunt, cousin, father), and non-kinship animate relational nouns (e.g. boss, classmate, opponent).
The alienability of the animate relational nouns is perhaps less clear than with inanimates, but
they provide an important comparison to the kinship nouns, since they are equally animate, more
alienable, and perhaps less socially salient.

5.4.1. Participants
Data from 48 native English speakers is presented for analysis. All participants were 18
years of age or older and were recruited from Amazon Mechanical Turk. All participants self-
identified as native English speakers and reported normal hearing and normal (or corrected-to-
normal) vision. Of the 54 participants who completed the experiment, 5 were excluded from the
analysis because they correctly answered fewer than 7 out of 8 catch trials, as well as one
participant who reported being born outside of the United States.

5.4.2. Design and materials
Experiment 4 had 24 target and 40 filler items. All target items followed the frame:
[name 1] [nonce verb in simple past tense] his/her [critical noun], and [name 2] did, too, as in
Example (3).
(3) Inalienable: Helen chabbed her nose, and Amanda did, too.
Ownership: Helen chabbed her jacket, and Amanda did, too.
[Animate] Relational: Helen chabbed her boss, and Amanda did, too.
Kinship: Helen chabbed her son, and Amanda did, too.
147
The two names in each target item always matched with respect to their typically associated
gender, with half of the targets female-associated and the other half male-associated. The critical
object noun was manipulated in order to test four possession types: part-whole, ownership,
[animate] relational, and kinship, as shown in Example (3). Part-whole nouns were all inanimate
and inalienable and consisted of body parts, (e.g. arm, nose) and abstract properties (e.g.
feelings, reputation). Ownership nouns were also inanimate and denoted concrete entities
typically understood to be alienable possessions (e.g. bicycle, book, jacket). The other two
conditions tested animate nouns: relational nouns expressing a non-family relation (e.g. boss,
friend, opponent) and kinship nouns expressing a family relation (e.g. aunt, brother, cousin).
Below each target sentence, there were two answer choices, as in Example (4). One
represented the bound-variable interpretation of the elided pronoun (4b), and the other
represented the coreferential interpretation (4c). Participants were instructed to select the answer
choice that they thought was most compatible with their interpretation of the sentence. The order
of the bound variable and coreferential choices was counterbalanced across conditions.
(4) a. Helen chabbed her jacket, and Amanda did, too.
b. Amanda chabbed her own jacket. [variable binding]
c. Amanda chabbed Helen's jacket. [coreference]
As in Experiment 1, nonce verbs (e.g. dreezed, jepped, swudged) were used in target items to
minimize potential effects of verb semantics. This was especially necessary in light of research
demonstrating that the verb’s lexical semantics affect the availability of bound variable versus
coreferential interpretations (e.g. Foley et al., 2003; Ong & Brasoveanu, 2014).
148
I used a within-subjects and within-items design, with the four possession-type conditions
distributed among experimental lists in a Latin square. With 24 targets, each participant saw 6
6

items per condition. Lists were pseudorandomized so that no target sentences occurred
consecutively and the number of filler items between targets was variable. Four additional
experimental lists were created by reversing the order of items in the lists generated by the Latin
square, for a total of eight lists.
Eight of the filler items were catch trials, where only one of the answer choices was
possible, as in Example (5).
(5) Monica lives in a big city, and Whitney lives in the countryside.
[Choice 1] Monica lives in a rural location.
[Choice 2] Monica lives in an urban location.
The other fillers contained ambiguities unrelated to ellipsis and pronoun resolution: prepositional
phrase attachment ambiguities (e.g. Amber and Courtney meached the woman from far away:
ambiguous between VP or NP attachment), scope/distributivity ambiguities (e.g. Cameron and
Justin talked to a stranger at the rowdy bar: ambiguous relative position of existential
quantification and conjunction), and lexical ambiguities (e.g. Steven and Samuel sprelled the
injured fisherman's cast: ambiguous between multiple contextually-supported meanings of cast).
Like target items, fillers always referred to two people with gender-matched names. Some fillers
contained nonce verbs or nouns.

6
An error resulted in the omission of one critical noun, personality, from the list of part-whole nouns in Experiment
4. Consequently, a quarter of participants saw a second trial containing the critical noun feelings. The error was
corrected in the analysis by omitting the second trial containing feelings. Prior to running Experiments 5a-d (which
used similar items), personality was properly integrated to the part-whole condition.
149
5.4.3. Procedure
Experiment 4 used a simple two-alternative forced-choice task to probe participants’
interpretational preference for ambiguous VPE constructions. Participants were instructed to
select the answer choice that they thought was most compatible with their interpretation of the
sentence. Before beginning the experiment, participants saw five example items to familiarize
them with the task. Each experimental item appeared on a separate screen and required
participants to confirm their selection by clicking an arrow to continue (see example item in
Figure 28).

Figure 28. An example target item from Experiment 4 as it appeared to participants

5.4.4. Predictions
In Section 5.3, I discussed the linking hypothesis I use to relate the likelihood of
coreferential interpretations of VPE to the discourse prominence of the possessed referent in the
antecedent clause. I adopt the view that a more prominent possessed referent makes a
coreferential interpretation more likely, since that referent becomes more available for the
computation of coreference, which is understood to be more costly than variable binding. For
example, when the comprehender processes the antecedent clause in Helen chabbed her jacket,
and Amanda did, too, if the referent corresponding to Helen’s jacket is more prominent, then it is
150
more likely that the comprehender will use that referent again when interpreting the ellipsis. The
resulting interpretation will be that Amanda chabbed Helen’s jacket (i.e. coreference). This
linking hypothesis aligns with the judgments in Dahl & Fraurud (1996) (see Examples 2a and
2b), which suggest that animate possessed referents increase the likelihood of coreference in
non-VPE constructions. The results of Experiments 1, 2, and 3 also support the idea that
possessed referents are comparable to simple nominals in that animates tend to be more
prominent than inanimates. Thus it seems reasonable to expect that greater prominence of the
possessed referent leads to a greater likelihood of coreference. With respect to the specific
conditions tested in Experiment 4, if this hypothesis is on the right track, I predict that the part-
whole and ownership nouns will exhibit a stronger bias toward variable binding (i.e. fewer
coreferential interpretations), since they are inanimate. On the other hand, the animate relational
and kinship nouns should produce more coreferential interpretations than the inanimate
conditions, owing to their greater prominence.
One of the primary goals of Experiment 4 is to examine whether the type of possession
relation modulates the overall discourse-level effect of possession. In Section 5.3, I explained
how this experiment can test my hypotheses on this topic (i.e. the Interaction Hypothesis and the
Animacy Hypothesis) despite not being able to compare possessives and simple nominals.
According to the Animacy Hypothesis, different possession relations do not contribute to the
prominence of possessed referents, and the difference in prominence between possessed
animates and inanimates is comparable to the difference between corresponding simple
nominals. Therefore, with regard to the possession relations tested in Experiment 4, the Animacy
Hypothesis predicts no difference between the part-whole and ownership conditions (both
151
inanimate), as well as no difference between the animate relational and kinship conditions (both
animate).
As an alternative to the Animacy Hypothesis, I have also introduced the Interaction
Hypothesis, which posits an extra boost in prominence for possessed animates. Identifying
evidence for this hypothesis is somewhat more difficult in Experiment 4, since I cannot rely on
the presence or absence of an interaction of animacy and possession (as in Experiments 1 and 2).
However, in discussing the results of Experiment 1, I suggested that a potential reason why
possessed animates are especially prominent might be that they denote socially salient
interpersonal relationships. If this reasoning is on the right track, one might expect gradience in
prominence depending on how socially meaningful the particular relationship is. In Experiment
4, I test two types of animate possessed referents, kinship and non-kinship relational nouns.
Assuming that kinship relations are typically more salient than non-kinship relations, one might
then presume that kinship referents are likewise more prominent. Thus, in Experiment 4, a
greater likelihood of coreference in the kinship condition than in the animate relational condition
would support the logic of the Interaction Hypothesis.
Additionally, Experiment 4 tests a theory from prior work (e.g. Foley et al., 2003; Guo et
al., 1996) that alienability of the possessed referent affects interpretational preferences in this
type of VPE. According to this theory, which I will call the Alienability Hypothesis, ambiguous
possessive pronouns in sentences like Example (1a) are more likely to receive bound variable
interpretations if the possessed referent is an inalienable possession compared to an alienable one
(i.e. his nose should yield more bound variable interpretations than his bicycle). This pattern was
observed in children acquiring VPE; however, this work only tested inalienable possessives via
part-whole relations. Experiment 4 additionally tests kinship nouns, which also represent
152
inalienable possession. Thus, with regard to the conditions tested in Experiment 4, the
Alienability Hypothesis predicts that part-whole nouns will result in more variable binding than
ownership nouns, since the part-whole relation is inalienable. Likewise, since the non-kinship
animate relational nouns are more alienable than the kinship nouns, kinship nouns should result
in more bound-variable interpretations. It is noteworthy that the predicted difference between the
animate conditions under the Alienability Hypothesis is opposite from the pattern predicted by
the Interaction Hypothesis. While the Alienability Hypothesis predicts that the kinship nouns
will result in more variable binding due to the inalienable possession relation, the Interaction
Hypothesis predicts that they will be more biased to coreference due to the greater social salience
of the possession relation.

5.4.5. Results
Statistical analyses for Experiment 4 used generalized linear mixed-effects models, which
were fit in a similar manner as in Experiments 1 and 2. Unlike those experiments, though, in
Experiment 4, I compared each condition to the three others by fitting four models, each taking a
different possession type condition as its reference level. The possession type predictor was
treatment coded, and outcomes were coded with 0 for coreference and 1 for variable binding.
The results of Experiment 4 demonstrate that possession type influences the strength of
bound variable bias (see Figure 29).

153

Figure 29. The proportion of bound variable responses by possession type in Experiment
4 (error bars represent ±1 standard error)

Descriptively, the inanimate conditions (i.e. part-whole and ownership relations) had the highest
rates of bound variable interpretation. The animate conditions (relational and kinship nouns)
received fewer bound variable interpretations, with the kinship nouns being the least likely of the
four conditions to yield a bound variable interpretation. These patterns are confirmed by the
statistical analyses. A table of the fixed effects in the models for Experiment 4 appears below as
Table 31.

154
Comparison
Reference
Level
Intercept Part-Whole Ownership Relational Kinship
Part-Whole
β = 1.885
SE = 0.355
z = 5.310
p < 0.001

β = -0.310
SE = 0.244
z = -1.273
p = 0.203
β = -2.056
SE = 0.242
z = -8.508
p < 0.001
β = -3.112
SE = 0.262
z = -11.885
p < 0.001
Ownership
β = 1.575
SE = 0.347
z = 4.534
p < 0.001
β = 0.310
SE = 0.244
z = 1.273
p = 0.203

β = -1.746
SE = 0.229
z = -7.631
p < 0.001
β = -2.802
SE = 0.249
z = -11.234
p < 0.001
Relational
β = -0.171
SE = 0.335
z = -0.509
p = 0.611
β = 2.056
SE = 0.242
z = 8.508
p < 0.001
β = 1.746
SE = 0.229
z = 7.631
p < 0.001

β = -1.057
SE = 0.219
z = -4.831
p < 0.001
Kinship
β = -1.2273
SE = 0.343
z = -3.582
p < 0.001
β = 3.112
SE = 0.262
z = 11.885
p < 0.001
β = 2.802
SE = 0.249
z = 11.234
p < 0.001
β = 1.057
SE = 0.219
z = 4.831
p < 0.001

Table 31. Summaries of fixed effects in the models for Experiment 4

The models revealed that the part-whole and ownership conditions elicited equivalently high
rates of bound variable interpretation (77.5% and 73.6% respectively). This finding contradicts
the Alienability Hypothesis, which predicted more bound variable responses in the part-whole
condition due to the inalienability of those relations.
Crucially, the two inanimate conditions triggered more bound-variable responses than
either of the animate conditions (47.6% for relational and 31.3% for kinship nouns), reflecting an
overarching effect of animacy. This pattern aligns with the judgments presented in Dahl &
Fraurud (1996), as well as the animacy effects demonstrated in this dissertation and in the wider
literature on discourse prominence.
Finally, there was a significant difference between the two animate conditions, with
relational nouns eliciting more bound-variable responses than kinship nouns. The directionality
155
of this difference opposes the Alienability Hypothesis, which predicted that the kinship nouns
would be more biased toward variable binding than the more alienable relational nouns. Instead,
this pattern is consistent with the Interaction Hypothesis, which theorized that the kinship terms
might be more prominent due to their social salience and thus more available for coreferential
interpretations.

5.4.6. Discussion
Experiment 4 investigated how different types of possessed referents affect the
interpretation of ambiguous VPE. The experiment aimed to test theories developed in previous
literature on VPE, as well as hypotheses set forth in this dissertation about the effect of
possession type on prominence. The results show that the possessed referent’s animacy is a
strong predictor of interpretational preference in VPE and provide evidence that more socially
salient interpersonal relationships (i.e. kinship) lead to greater discourse prominence for the
possessed referent.
Because the two conditions with inanimate possessed referents (part-whole and
ownership) received higher rates of bound variable interpretation than the two conditions with
animate possessed referents (relational and kinship), the judgments of Dahl & Fraurud (1996) for
non-VPE constructions are corroborated in VPE as well. Under the linking hypothesis discussed
in Section 5.3, this pattern is expected given animates’ well-known advantage in prominence
over inanimates. Additionally, these results align with the robust animacy effects demonstrated
in earlier experiments in this dissertation.
One of the primary aims of Experiment 4 was to assess whether different possession
relations modulate the overall effect of possession on a referent’s prominence. In Experiment 1, I
156
found evidence that possessed animates receive an extra boost in prominence (supporting the
Interaction Hypothesis), although this effect was not present in Experiment 2. Experiment 4
introduced alienability as a potential new dimension along which the prominence of possessed
referents might vary. Building primarily on studies of children acquiring VPE, I formulated the
Alienability Hypothesis: that inalienably possessed referents would trigger more bound-variable
responses than alienably possessed ones; therefore, according to my linking hypothesis, one
could consider alienably possessed referents more prominent.
However, I did not find evidence for the Alienability Hypothesis. There was no
difference between the inalienable part-whole condition and the alienable ownership condition.
Furthermore, the difference I found within the animate conditions directly contradicted the
Alienability Hypothesis, since the more alienable non-kinship relational nouns triggered more
bound-variable responses than the kinship nouns. Instead this pattern lends additional credence
to the idea discussed Chapter 2 that the social salience of interpersonal relations makes animate
possessed referents more prominent in discourse. In this case, the kinship nouns may have been
particularly prominent because they carry more social importance than the non-kinship relational
nouns. Under my linking hypothesis, this difference in prominence would lead to exactly the
pattern I observe in Experiment 4, with the kinship condition triggering the most coreferential
interpretations.
The results of Experiment 4 support my hypothesis linking the discourse prominence of
the possessed referent and the rate of bound-variable interpretation. Nevertheless, there could be
an alternative explanation for the observed patterns. Perhaps preference between bound-variable
and coreferential interpretations is determined simply based on real-world plausibility (e.g. in
Bill washed his car, and John did, too, the likelihood that John has his own car). This alternative
157
theory would state that the more likely John is to have his own car (assuming the comprehender
has no additional knowledge of John), the more likely a comprehender is to interpret the ellipsis
with variable binding; however, if John is unlikely to have his own car, coreference would be
preferred. To assess the validity of this alternative account, I conducted a simple norming
experiment.
In this follow-up experiment, 28 participants (who had not taken part in Experiment 4)
were asked to rate “how likely an average person is to have at least one” of 65 nouns, including
all those used in Experiment 4. If participants’ intuitions about nouns’ likelihood of possession
were driving the results in Experiment 4 in the manner described above, one would expect that
likelihood of possession would correlate positively with the proportion of bound variable
interpretations a given noun produced in Experiment 4. To test this alternative hypothesis, I
converted participants’ raw likelihood of possession ratings into z-scores and used Pearson’s
product moment correlation to assess the relationship between the mean z-scored rating for each
noun and the proportion of bound-variable interpretations which that noun produced in
Experiment 4.
Ratings of possession likelihood did not correlate significantly with rates of bound
variable interpretation (r = 0.13, p = 0.38; see Figure 30).

158

Figure 30. The relationship between each noun’s possession likelihood and the
proportion of bound variable interpretations which it produced in Experiment 4

The results of this analysis demonstrate that people do possess high-level assumptions that
certain nouns are more likely to be possessed by a person about whom they have no prior
knowledge; however, there is no evidence that these assumptions influence the rate of bound-
variable interpretation for a sentence containing that possessed noun. These results argue against
an account which prioritizes real-world plausibility as a predictor of bound variable bias. Instead,
the results indirectly support my prominence-based account.

5.5. Experiments 5a–d
The linking hypothesis discussed in Section 5.3 takes the discourse prominence of the
possessed referent to be the primary determinant of the likelihood the ellipsis will receive a
coreferential interpretation. This intuition is based on previous work which posits that computing
159
coreference requires access to discourse-level representations (e.g. Heim, 1982; Reinhart, 2000;
Reuland, 2001). A logical prediction of this approach is that if one eliminates the opportunity to
build a discourse representation for the possessed referent, bound variable interpretation should
be the only available option. Using quantified anaphora presents such an opportunity to block the
creation of a discourse representation. According to Reinhart (2000), “referential NPs form a
discourse entity that [a] pronoun can directly be covalued with, while in the case of quantified
NPs, covaluation is only possible between a pronoun and a variable.” Reinhart notes that this
property of quantified anaphora gives rise to differences in the possible interpretations of
covalued possessive pronouns in sentences like Example (6):
(6) a. Lucie thinks that only she respects her husband.
b. Every woman thinks that only she respects her husband.
Specifically, in (6a), her can either be covalued with the discourse entity Lucie or bound by the
same operator which binds she. Nevertheless, both mechanisms yield the same interpretation (i.e.
that Lucie thinks that other people do not respect her husband). However, in (6b), her cannot be
covalued with the quantified NP every woman, and instead must be bound by the same operator
which binds she. Thus the grammatical interpretation of (6b) is that every woman thinks that she
is the only person who respects her own husband. In this interpretation she and her are
referentially equivalent. On the other hand, in the ungrammatical covaluation configuration, she
would be bound, but her would be covalued with every woman. The resulting interpretation
would be that every woman thinks that she is the only person who respects every woman’s
husband; however (6b) cannot be understood to have this meaning.
160
If one applies Reinhart’s logic to the type of ellipsis examined in Experiment 4, one can
make predictions about the interpretations which should be available when quantified NPs are
placed in strategic positions of the sentence.
(7) Bill washed his car, and John did, too.
(8) Every man washed his car, and John did, too.
(9) Bill washed his car, and every man did, too.
In Example (7), both the bound variable (i.e. that John washed his own car) and coreferential (i.e.
that John washed Bill’s car) interpretations are possible, irrespective of any previously discussed
interpretational biases; however, in (8), every man does not introduce a discourse referent, and
therefore both the overt possessive pronoun and the elided one must be interpreted via binding.
This restriction rules out an interpretation in which the elided possessive pronoun is coreferential
with every man (i.e. that John washed every man’s car). Furthermore, because every man does
not introduce a discourse referent, there should likewise be no discourse referent corresponding
to car in every man’s car. Thus, the only available interpretation of (8), with the quantified NP as
the subject of the antecedent clause, requires the elided possessive pronoun to be bound by John
(i.e. that John washed his own car). The situation is different when the quantified NP is the
subject of the elided clause, as in (9). In this case, because Bill introduces a discourse referent,
both the bound-variable (i.e. that every man washed his own car) and the coreferential
interpretations (i.e. that every man washed Bill’s car) are possible.
Based on the theoretical observations discussed above, one can predict that if the target
items from Experiment 4 were modified so that the subject of the antecedent clause were a
quantified NP as in (8), the only interpretational option should be variable binding; however, if
the subject of the elided clause were a quantified NP, both bound-variable and coreferential
161
interpretations should be available. Furthermore, assuming that both interpretations are possible,
and that the choice between them is still driven by the prominence of the possessed referent, I
should observe a similar pattern of results as in Experiment 4. Because (7) and (9) have identical
antecedent clauses, discourse-level processing of the possessed referent should proceed in the
same manner for both, and thus the sentences should exhibit a similar bias toward variable
binding or coreference. I test these predictions in a series of four experiments (Experiments 5a–
d), which use modified materials based on those in Experiment 4.

5.5.1. Participants
The analysis of Experiments 5a and 5b each present data from 32 native English speakers
recruited from Amazon Mechanical Turk; Experiments 5c and 5d each present data from 24
native English speakers also recruited from Amazon Mechanical Turk. All participants were 18
years of age or older and did not participate in any related experiments. All participants self-
identified as native English speakers and reported normal hearing and normal (or corrected-to-
normal) vision. As in Experiment 4, I excluded from analysis any participant who reported being
born outside of the United States or scored below 7/8 correct on catch trials.

5.5.2. Design and materials
The experimental design and filler items in Experiments 5a-d were all the same as in
Experiment 4. The only modifications to the materials from Experiment 4 were in the target
sentences and their interpretational choices. In Experiment 5a, antecedent-clause subjects were
switched to every man or every woman (to match the typically associated gender of the original
subject). The coreferential interpretation was also updated to reflect the theoretically
162
ungrammatical coreferential interpretation of Example (8). Compare Examples (10a-c) from
Experiment 4 with Examples (11a-c) from Experiment 5a:
(10) a. Helen chabbed her jacket, and Amanda did, too.
b. Amanda chabbed her own jacket. [variable binding]
c. Amanda chabbed Helen's jacket. [coreference]
(11) a. Every woman chabbed her jacket, and Amanda did, too.
b. Amanda chabbed her own jacket. [variable binding]
c. Amanda chabbed every woman’s jacket. [coreference]
Experiment 5b made the corresponding substitution in subject position of the elided clause:
(12) a. Helen chabbed her jacket, and every woman did, too.
b. Every woman chabbed her own jacket. [variable binding]
c. Every woman chabbed Helen's jacket. [coreference]
I also created two additional sets of items, for Experiments 5c and 5d, which switched the
genders of the quantified NPs so they would contrast with the typically associated genders of the
names:
(13) a. Every man chabbed his jacket, and Amanda did, too.
b. Amanda chabbed her own jacket. [variable binding]
c. Amanda chabbed every man’s jacket. [coreference]
(14) a. Helen chabbed her jacket, and every man did, too.
b. Every man chabbed his own jacket. [variable binding]
c. Every man chabbed Helen's jacket. [coreference]
There were two main reasons for introducing gender mismatch between the names and
quantified NPs in Experiments 5c and 5d. Firstly, I thought participants might find sentences like
163
(11a) pragmatically strange, since the first clause Every woman chabbed her jacket entails that
Amanda chabbed her own jacket, so long as Amanda is a woman. Therefore, in reasoning about
why the elided clause and Amanda did, too. was present, participants might infer that Amanda is
not a woman—even though the answer choice (11b) refers to Amanda with her. Additionally, I
wanted to make sure that the bound-variable answer choice was not being selected more often as
an artifact of the entailment relationship between it and the antecedent clause. Similar pragmatic
issues could exist to a lesser degree in Experiment 5b, although, in Helen chabbed her jacket,
and every woman did, too., the second clause is still informative given that Helen is a woman.
Secondly, I wanted to explore an observation discussed at length in Oku (1998, pp. 99-
112) (see also Fiengo & May, 1994; Kitagawa, 1991; Sag, 1976) that sentences similar to
Examples (13a) and (14a), which involve gender mismatch between the possessive pronoun in
the antecedent clause and its elided counterpart under bound-variable interpretation, may have
degraded acceptability if they are interpreted with variable binding. Furthermore, these previous
theoretical accounts observe that the order of the mismatched genders can interact with the
gender-mismatch effect.
(15) a. Bill washed his car, and Susan did, too.
b. Bill
i
washed his
i
car, and Susan
j
did [washed his
j
her
j
car], too. [variable binding]
c. Bill
i
washed his
i
car, and Susan
j
did [washed his
i
car], too. [coreference]
(16) a. Mary washed her car, and John did, too.
b. ?Mary
i
washed her
i
car, and John
j
did [washed her
j
his
j
car], too. [variable binding]
c. Mary
i
washed her
i
car, and John
j
did [washed her
i
car], too. [coreference]
The essential generalization reported in previous work is that, with some degree of variation,
most native English speakers can interpret sentence (15a) with the bound variable interpretation
164
in (15b); however, many speakers find that the bound variable interpretation (16b) of sentence
(16a) is degraded. These previous accounts of ellipsis resolution have proposed additional
mechanisms in order to deal with this observation; nevertheless, these judgments have never
been submitted to rigorous experimental verification, nor did previous work acknowledge the
impact of the possessed noun’s lexical semantic properties on the acceptability of bound variable
interpretations.

5.5.3. Procedure
The task and instructions in Experiments 5a–d were the same as in Experiment 4.

5.5.4. Predictions
I have proposed that the effects observed in Experiment 4 are due to differences in
discourse prominence among the different types of possessed referents. On this basis, I predict
that when the creation of a discourse representation for the possessed referent is blocked by the
presence of a quantified NP in subject position of the antecedent clause, binding will be the only
possible mechanism for interpreting the ellipsis. Therefore, rates of bound variable interpretation
for target items in Experiment 5a should be uniformly at ceiling, regardless of possession type.
On the other hand, in Experiment 5b, where the quantified NP is the subject of the elided clause,
discourse level-processing of the possessed referent should be similar as in Experiment 4.
Therefore, the same type of animacy effect is predicted as in Experiment 4, with animate
conditions resulting in significantly more coreferential interpretations than inanimate conditions.
For Experiments 5c and 5d, predictions are similar. Just as in Experiment 5a, where the
quantified NP in the antecedent clause blocks coreference, the same high rate of bound-variable
165
interpretations is predicted across all conditions. It is unclear though whether the processing
constraint presented by gender mismatch will push participants to choose a theoretically
ungrammatical coreferential interpretation. At least in some previous work on this type of gender
mismatch, considerable interspeaker variability was reported in how unacceptable bound
variable interpretations became when the overt and elided possessive pronouns did not match in
gender (Kitagawa, 1991). Given that this effect potentially varies across speakers, and that the
lack of discourse referents for quantified NPs seems to be a comparatively clear-cut grammatical
restriction, I expect that gender mismatch will not lead to a substantial increase in the rate of
coreferential interpretation when the quantified NP is in the antecedent clause.
However, in Experiment 5d, both coreferential and bound variable interpretations should
be available. In this case, one might expect to see fewer bound-variable interpretations relative to
Experiment 5b (where gender always matches), since variable binding necessitates a mismatch in
gender between the overt possessive pronoun in the antecedent clause and the elided one within
the ellipsis site (see Examples 15b and 16b). Since this situation applies to bound-variable
readings regardless of possession type, we should see a decrease in the rate of bound-variable
interpretations in all possession conditions relative to Experiment 5b. Additionally, if judgments
that Example (15b) is more acceptable than Example (16b) are reliable, there may be an
interaction between gender mismatch and overt pronoun gender. If this is the case, one would
expect that there would fewer bound-variable interpretations in Experiment 5d when the overt
possessive pronoun is feminine.

166
5.5.5. Results
The results of Experiments 5a–d are shown below in Figure 31. In sum, Experiments 5a
and 5c, where the quantified NP is the subject of the antecedent clause, exhibited a strong bound-
variable bias in all conditions and no animacy effect, in accordance with my predictions.
However, in Experiments 5b and 5d, where the quantified NP is the subject of the elided clause,
inanimate conditions received more bound-variable interpretations than animates, again
confirming my predictions. I also found a drop in bound variable responses due to gender
mismatch with elided-clause quantified NPs, but not with antecedent-clause quantified NPs. I
found no interaction between the gender mismatch effect and the gender of the overt pronoun, as
was previously reported in theoretical literature.

Figure 31. The proportion of bound variable interpretations by possession-type condition
in Experiments 5a–d (error bars represent ±1 standard error)
167

Statistical analyses for this experiment again used generalized linear mixed-effects
models. Unlike in Experiment 4, four predictors were deviation coded: quantified-NP position
(antecedent clause = 0.5, elided clause = -0.5), animacy of the possessed referent (animate = 0.5,
inanimate = -0.5), gender mismatch (mismatch = 0.5, match = -0.5), and overt pronoun gender.
(feminine = 0.5, masculine = -0.5). To begin, a model was fit which encompassed the data from
all four sub-experiments with quantified-NP position, animacy, gender mismatch, and overt
pronoun gender as predictors. A summary of the fixed effects of this model is given below as
Table 32.

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) 5.080 0.766 6.631 <0.001
QuNP position 8.383 1.522 5.509 <0.001
animacy -0.868 0.305 -2.849 0.004
gender match -1.394 0.823 -1.694 0.090
overt pronoun -0.275 0.323 -0.851 0.395
QuNP position:animacy 1.885 0.585 3.221 0.001
QuNP position:gender match 1.794 1.647 1.089 0.276
animacy:gender match 0.966 0.577 1.675 0.094
QuNP position:overt pronoun -0.101 0.563 -0.179 0.858
animacy:overt pronoun 0.119 0.592 0.201 0.841
gender match:overt pronoun 0.419 0.560 0.748 0.454
QuNP position:animacy:gender match 1.621 1.151 1.408 0.159
QuNP position:animacy:overt pronoun 1.167 1.132 1.031 0.303
QuNP position:gender match:overt pronoun 1.369 1.121 1.221 0.222
animacy:gender match:overt pronoun -1.081 1.120 -0.966 0.334
QuNP position:animacy:gender match:overt pronoun -2.260 2.240 -1.009 0.313
Table 32. Summary of fixed effects for the model predicting the rate of bound
variable interpretations in Experiments 5a–d

As expected, this model showed a significant main effect of quantified-NP position, with
quantified NPs in the antecedent clause resulting in more bound-variable interpretations.
Additionally, quantified-NP position interacted significantly with animacy, which conformed to
168
my prediction that animacy effects similar to the one in Experiment 4 would occur with elided-
clause quantified NPs, but not with antecedent-clause quantified NPs. Interestingly, this model
also produced a significant main effect of animacy. This effect was not part of my predictions,
since I expected animacy to only play a role when the quantified NP was in the elided clause.
Given the significant interaction which supported my theory that quantified-NP position would
greatly affect the data patterns, I decided to simplify the analysis by examining Experiments 5a
and 5c (antecedent-clause quantified NP) separately from Experiments 5b and 5d (elided-clause
quantified NP). I therefore fit an additional model for each pair of sub-experiments with the same
predictors as in the larger model, except for quantified-NP position.
In the model for the first pair of sub-experiments, where the quantified-NP appeared as
the subject of the antecedent clause, I observed no significant effects of any predictors. A
summary of the fixed effects in this model is given below as Table 33.

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) 9.082 1.533 5.923 <0.001
animacy -0.034 0.591 -0.058 0.954
gender match 0.376 3.062 0.123 0.902
overt pronoun -0.295 0.596 -0.495 0.620
animacy:gender match 1.788 1.121 1.595 0.111
animacy:overt pronoun 0.430 1.142 0.377 0.706
gender match:overt pronoun 1.101 1.086 1.013 0.311
animacy:gender match:overt pronoun -2.189 2.171 -1.008 0.313
Table 33. Summary of fixed effects for the model predicting the rate of bound
variable interpretations in Experiments 5a and 5c

Despite the main effect of animacy in the larger model, there was no significant effect of
animacy when the quantified NP appeared in the antecedent clause. This result supports my
claim that binding is the comprehender’s only choice when no discourse referent exists for the
169
possessed referent. Gender mismatch likewise has no opportunity to exert an effect either, since
coreference seems to be ruled out as a grammatical rule in these cases.
The results are much different for the two sub-experiments where the quantified-NP
appeared as the subject of the elided clause. A summary of the fixed effects in the model for
Experiments 5b and 5d is given below as Table 34.

Fixed effect Estimate Std. Error z value Pr(>|z|)
(Intercept) 0.824 0.276 2.986 0.003
animacy -1.629 0.233 -7.001 <0.001
gender match -2.152 0.529 -4.067 <0.001
overt pronoun -0.244 0.220 -1.109 0.267
animacy:gender match -0.318 0.471 -0.675 0.500
animacy:overt pronoun -0.482 0.306 -1.578 0.114
gender match:overt pronoun -0.257 0.305 -0.843 0.399
animacy:gender match:overt pronoun 0.038 0.610 0.063 0.950
Table 34. Summary of fixed effects for the model predicting the rate of bound
variable interpretations in Experiments 5b and 5d

Here the expected main effect of animacy was observed, with animate possessions resulting in
significantly fewer bound-variable interpretations. There was also a significant main effect of
gender mismatch, which demonstrates that participants gave fewer bound variable responses
when the overt pronoun mismatched in gender with the elided one. This result partially agrees
with judgments from the theoretical literature; however, I found no interaction between this
gender mismatch effect and overt pronoun gender, contradicting Oku (1998). Additionally, the
effect of gender mismatch does not interact with animacy, further supporting the idea that
morphological non-identity between the antecedent and elided clauses generally interferes with
computing bound variable interpretations.

170
5.5.6. Discussion
Overall, the results of Experiments 5a-d support my hypothesis that the discourse
prominence of the possessed referent affects the availability of coreferential interpretations for
this type of VPE. The data conform to my prediction that the lack of an available discourse
referent for the possessed noun precludes coreferential interpretation. When coreference is
permitted, as is the case in Experiments 5b and 5d, there was a similar effect of animacy as in
Experiment 4. Additionally, a decrease in bound-variable interpretations was observed due to
gender mismatch when the quantified NP was in the elided-clause (Experiment 5d), which partly
confirms some judgments presented in previous theoretical literature. I attribute the gender-
mismatch penalty to an integration conflict that arises when retrieving a memory representation
of the antecedent VP which conflicts with the VP representation activated by variable binding.
However, in contrast to the judgments of Oku (1998), I did not find an interaction between the
gender mismatch effect and overt pronoun gender. This result has some bearing on theoretical
debate concerning how much semantic content is retrieved in computing bound-variable
interpretations. Specifically, there have been conflicting claims about whether the retrieved
content includes anything beyond the logical form of the antecedent clause, which presumably
would not include gender features (e.g. Avrutin, 1999; Frazier & Clifton, 2000). The overall
lower rate of bound-variable interpretations in Experiment 5d versus Experiment 5b suggests that
gender features are retrieved during bound variable-interpretations, since there would be no basis
for the gender-mismatch penalty without the presence of these features.

171
5.6. General Discussion
This chapter aimed to further investigate whether the type of possession relation affects
the prominence of possessed referents. Experiments 2 and 3 in this dissertation demonstrated that
possession generally reduces the prominence of possessed referents but found no evidence that
this effect varies with different possession relations. On the other hand, Experiment 1 found a
superadditive boost in prominence for possessed animates. This finding led to my proposal of the
Interaction Hypothesis, which attributes the added prominence of possessed animates to the non-
linguistic importance of interpersonal relationships, which these possessives denote. Given the
mixed results in preceding chapters, in this chapter, I approached my research aims through a
different theoretical and methodological domain: the interpretation of ambiguous possessive
pronouns in verb-phrase ellipsis (VPE; e.g. Bill washed his car, and John did, too). To link this
approach to the main hypotheses examined in this dissertation, I proposed a discourse-based
approach to the resolution of this type of VPE, where the prominence of the referent possessed
by the pronoun is the primary factor in whether the ellipsis is interpreted using variable binding
or coreference.
The results of Experiments 4 and 5a-d support the usefulness of my discourse-based
approach to comprehension of this type of VPE. In general, the rates of bound-variable versus
coreferential interpretations matched the expected differences in prominence among the tested
types of possessed referents. Furthermore, the difference between the animate relational and
kinship conditions in Experiment 4 supported the Interaction Hypothesis by showing that kinship
relations, which are likely more socially important than non-kinship relations, produced the
highest rates of coreferential interpretation. Therefore, according to the linking hypothesis,
possessed kinship referents may get the greatest boost in prominence.
172
Additionally, this chapter tested whether the alienability of the possessed referent in this
type of VPE affects interpretational preferences. The importance of alienability in this domain
was originally proposed in acquisition-focused work which tested a more limited range of
possession relations (Foley et al., 2003; Guo et al., 1996). The results of this chapter do not
support alienability as a determinant of interpretational preference in these constructions and
therefore do not implicate alienability as a factor which modulates the overall discourse-level
effect of possession. Nevertheless, the role of alienability may be a topic worthy of additional
investigation in future work that uses more established methodologies for measuring discourse
prominence (e.g. sentence continuation, visual-world eye-tracking). Finally, this chapter
demonstrates conclusively that the type of possession relation can influence interpretation of
ambiguous possessive pronouns; this rather general point has not often been acknowledged in
more theoretically-oriented work on binding and coreference, and thus this work may be of
interest to researchers in those fields.

173
Chapter 6: Conclusion

6.1. Overview and summary
This dissertation aims primarily to address two questions about the discourse-level effect
of possession on the referents involved in possessive relations. Firstly, it considers how
possession affects the prominence of possessed referents in a general sense (i.e. regardless of the
type of semantic relation between the possessor and possessed referent, is there an overall
discourse-level effect of possession?). Secondly, given that possessives express a wide variety of
relations between the two involved referents, it also investigates whether different possession
relations modulate the aforementioned overall effect.
Based on the first of these two research questions, I formulated two competing
hypotheses about the overall effect of possession on the prominence of possessed referents: the
Complexity Hypothesis and the Parasitic Possession Hypothesis. The Complexity Hypothesis
builds on the assumption that possessives are more representationally complex than simple
nominals because they mention an additional referent (i.e. the possessor) and encode a
possession relation between the possessor and possessed referent. Based on prior work showing
that more representationally complex referents (i.e. referents which contain more information)
tend to be more prominent in discourse (e.g. Hofmeister 2011; Hofmeister & Vasishth, 2014;
Karimi & Ferreira, 2016; Karimi et al., 2014, 2019; Troyer et al., 2016), I hypothesized that
possessed referents might be more prominent in discourse than corresponding simple nominals
(e.g. the referent for car is more prominent in Sam’s car versus the car).
On the other hand, the Parasitic Possession Hypothesis predicts the opposite effect on the
prominence of possessed referents. This view draws inspiration from prior work within the
174
Centering Theory literature which claims that possessors outrank their corresponding possessed
referents (e.g. Chae, 2003; Walker & Prince, 1996). This assertion provides indirect evidence for
the view that possession decreases the prominence of the possessed referent. For example, in the
sentence The car is in the garage, Centering Theory (or essentially any other theory of discourse)
would predict that the referent for car is most prominent by virtue of its subjecthood. Bearing in
mind the theory that possessors outrank possessed referents, if one compares the preceding
sentence to Her car is in the garage, then the antecedent of the possessive pronoun her is most
prominent because it outranks the referent for car. Thus, the net effect of possession on the
possessed referent is a reduction in prominence.
Regarding the second main research question in this dissertation (i.e. do different
possession relations affect prominence differently?), I proposed two additional hypotheses: the
Animacy Hypothesis and the Interaction Hypothesis. The Animacy Hypothesis builds on
extensive prior work showing that animates tend to be more prominent in discourse than
inanimates (e.g. Bock et al., 1992, Branigan et al., 2008; Dahl, 2008; Dahl & Fraurud, 1996;
Fukumura & van Gompel, 2011; Yamamoto, 1999). The Animacy Hypothesis therefore suggests
that possessed animates are more prominent than possessed inanimates and—crucially—that this
difference in prominence will be the same as the one between animate and inanimate simple
nominals. Accordingly, regardless of whether the overall effect of possession is consistent with
the Complexity Hypothesis or Parasitic Possession Hypothesis, that effect would not vary
depending on the type of possession relation. Furthermore, I use animacy of the possessed
referent as a proxy for different possession relations, since animate possessors form different
relations with animates versus inanimates (e.g. her car expresses ownership, but her lawyer
175
expresses an interpersonal relationship). Thus the Animacy Hypothesis predicts that the type of
possession relation does not modulate the overall discourse-level effect of possession.
The Interaction Hypothesis opposes the Animacy Hypothesis. This theory takes
inspiration from an alternative account within Centering Theory (Di Eugenio, 1998) which
claims that [animate] possessors outrank inanimate possessed referents (e.g. the referent for Sam
is higher-ranking in Sam’s car), but the ranking reverses when the possessed referent is animate
(e.g. the referent for lawyer is higher-ranking in Sam’s lawyer). This theory implies an extra
boost in prominence for possessed animates. Furthermore, possessed animates denote
interpersonal relationships, which in have been shown to be particularly important in non-
linguistic cognition and human health (e.g. Cacioppo & Hawkley, 2009; Eisenberger & Cole,
2012; Nairne, 2010; VanArsdall et al. 2013). It might be that the general importance of
interpersonal relationships contributes to an increase in discourse prominence for possessed
animates. I therefore formulated the Interaction Hypothesis, which predicts a superadditive
increase in prominence for possessed animates—in excess of whatever level of prominence
would be predicted by the additive combination of animacy and the overall effect of possession.
This dissertation evaluated these two sets of competing hypotheses in five experiments.
The first two experiments used a sentence continuation paradigm to compare the discourse
properties of possessives to two types of simple nominals: indefinites (Chapter 2, Experiment 1)
and definites (Chapter 3, Experiment 2). Both experiments measured how often participants’
continuations mentioned referents from prompts like Jessica rulked {a, her} chandelier
(Experiment 1) and The midwife knew {the, her} priest because (Experiment 2). Experiment 1
showed that participants were especially likely to mention possessed animates from the prompt
both in subject position and the entirety of their continuations. This statistical interaction of
176
animacy and possession supports the Interaction Hypothesis and indicates that possessed
animates receive an extra boost in prominence. Additionally, Experiment 1 produced relatively
weaker support for the Complexity Hypothesis in an analysis of pronominalization patterns;
however, I argued that these results could be attributable to differences in the discourse
properties of indefinites and possessives which are unrelated to the presence or absence of a
possession relation. In Experiment 2, when possessives were compared to definites, there was no
interaction of animacy and possession, but the results did show an overall reduction in the rate of
subjecthood for possessed referents. Thus Experiment 2 supported the Parasitic Possession
Hypothesis, which predicts a decrease in prominence for possessed referents, and the Animacy
Hypothesis, which predicts that this decrease does not vary across different types of possession
relations.
Experiment 3 (Chapter 4) shifted from sentence continuation to a visual-world eye-
tracking paradigm, in which participants heard sentences like The soldier heard his plumber
throughout the autumn evening because he was nimpacious, and their gaze patterns during
pronoun resolution were analyzed. Unlike in the sentence continuation experiments, where the
primary dependent variable was a referent’s likelihood of mention—production-based
P(referent) in the Bayesian formalism of Kehler et al. (2008)—Experiment 3 considered the
comprehension-based likelihood that a referent is the antecedent of a pronoun, i.e.
P(referent|pronoun). The results showed that participants looked less at potential pronoun
antecedents when they were possessed compared to definite, suggesting that possessed referents
are less accessible and prominent in comprehenders’ mental models of discourse. Thus this result
also supports the Parasitic Possession Hypothesis.
177
Chapter 5 considered the resolution of ambiguous possessive pronouns in a type verb-
phrase ellipsis (VPE) which permits either bound-variable or coreferential interpretations (e.g. in
Bill washed his car, and John did, too, John could have washed his own car or Bill’s). In this
chapter, I adopted a linking hypothesis about the interpretation of this type of VPE in which the
prominence of the possessed referent contributes to the likelihood of a coreferential
interpretation. Assuming that the possessed referent from the antecedent clause needs to be
retrieved for coreference but not for binding (e.g. Reinhart, 2000; Reuland, 2001), I hypothesized
that a more prominent possessed referent in the antecedent clause would be easier to retrieve and
make coreferential interpretation of the ellipsis more likely.
Experiment 4 supported the logic of this linking hypothesis by showing that the animacy
of the possessed referent is an influential factor in how this type of ambiguous ellipsis is
interpreted. When the possessed referent was animate (e.g. her boss), participants were more
likely to assign a coreferential interpretation to the ellipsis than when the possessed referent was
inanimate (e.g. her jacket). Under the linking hypothesis, this result is consistent with animates’
general advantage in prominence over inanimates. Additionally, Experiment 4 revealed a
difference in the rate of coreferential interpretations with kinship and non-kinship animate
relational nouns (e.g. her son versus her boss). This result suggests a difference in the discourse
prominence of possessed referents which is dependent on the type of possession relation and,
therefore, contradicts the Animacy Hypothesis. Furthermore, the directionality of this difference
is consistent with the underlying logic of the Interaction Hypothesis, which argues that possessed
animates are especially prominent because of the cognitively salient interpersonal relations
which they denote. In the case of Experiment 4, if one assumes that kinship relations generally
have a higher degree of social importance than non-kinship interpersonal relations, it follows that
178
the more salient relationships would result in more prominent possessed referents and ultimately
lead to a greater likelihood of coreferential interpretation. Finally, Experiment 5 further
supported my discourse-based approach to this type of VPE by showing that the differences
between the possession types shown in Experiment 4 depend on the creation of a discourse
referent for the possessed noun. If one introduces quantification into the antecedent clause in a
way that prevents the creation of a discourse referent for the possessed noun (e.g. Every man
washed his car, and John did, too), then binding is the only option for interpretation and the type
of possession relation becomes irrelevant.
Viewed as a whole, the results of this dissertation support the Parasitic Possession
Hypothesis and the Interaction Hypothesis. It is especially notable that the evidence for both of
these hypotheses comes from a range of experimental methodologies which incorporate
production-based (i.e. sentence continuation) and comprehension-based (i.e. ambiguous pronoun
resolution) measures. Furthermore, this evidence addresses both sides of the Bayesian formalism
of Kehler et al. (2008) and shows that these hypotheses are useful under multiple definitions of
prominence. Thus this dissertation generally argues for a view in which possession decreases the
prominence of possessed referents; however, different types of possession relations may
modulate this overall effect. Specifically, possessed animates seem to receive an extra boost in
prominence, which is not accounted for by the canonical discourse-level effect of animacy alone.

6.2. What underlies the Parasitic Possession Hypothesis?
In discussing the Interaction Hypothesis, I have explicitly proposed a theory for why
possessed animates might be especially prominent on the discourse level. In this case, I have
suggested that the social importance of interpersonal relationships might exert an influence on
179
language processing by privileging the linguistic realization of these relationships at the
discourse level. However, with regard to the Parasitic Possession Hypothesis, there is no obvious
reason why possession should decrease the prominence of the possessed referent. What then
could be the underlying mechanism which contributes to this decrease in prominence?
In earlier chapters, I hinted that possession might decrease the prominence of the
possessed referent because it somehow makes the possessed referent dependent on the referent of
the possessor. In order to flesh out this idea, I appealed to the theory of File Change Semantics
(Heim, 1982) as an example of how such a dependency structure could be modeled. In File
Change Semantics, a sentence containing a possessive structure like A man’s car broke down.
would involve a file containing two file cards, Card 1 for the man and Card 2 for the car. The
first file card would include the propositions “is a man” and “possesses 2”. The second file card
would include the propositions “is a car”, “is possessed by 2”, and “broke down”.
In its basic theoretical formulation, File Change Semantics has no mechanism to capture
the notion of hierarchical dependence between file cards; however, this representational structure
could be altered such that the car’s (i.e. the possessed referent) file card would be embedded
within the man’s (i.e. the possessor). This sort of structure resembles the hierarchical discourse
representations in Discourse Representation Theory (e.g. Kamp, 1981; Kamp & Reyle, 1993).
Such a setup could make the possessed referent less accessible within the discourse model and
therefore less prominent. Importantly, I am not suggesting an explicit amendment to File Change
Semantics or DRT, since these modifications are not intended to maintain either framework as a
theory of compositional semantics. Rather, I invoke the theory of file cards (or analogous
discourse representation structures) as a useful tool for thinking about how discourse
representations are organized.
180
As I noted in Chapter 2, the dependent status of possessed referents would not preclude
access to them, as shown in the example discourse Mary drove her car. It’s a red convertible.
Here the possessed referent must be available to resolve the subsequent pronoun. Instead,
accessing a dependent possessed referent could simply be more computationally costly than
accessing an independent (or less deeply embedded) referent. Additionally, this model still
allows the dependent possessed referent to maintain the lexical semantic features which may
otherwise contribute to its prominence. Among these features, animacy would be one of the most
important, and thus this model can account for the numerous animacy effects which have been
noted in this dissertation. Furthermore, this model can also be adapted to capture the possibility
that different possession relations may modulate the overall discourse-level effect of possession.
For example, the strength of the dependency between the possessor and possessed referent could
vary depending on the possession relation. In order to capture the claims of the Interaction
Hypothesis, animate possessed referents might be less tightly linked to the representations of
their possessors, and thus they would be especially accessible and prominent (in excess of the
additive baseline effects of animacy and possession). By varying the strength of the dependency
between the possessor’s and possessed referent’s discourse representations, this theory could be
expanded to capture types of possession relations which have not yet been investigated with
regard to discourse prominence.

6.3. Reconsidering the Complexity Hypothesis
While Experiment 1 (Chapter 2) did produce some evidence (i.e. in the pronominalization
patterns) which appeared favorable to the Complexity Hypothesis, ultimately there was more
compelling support for the Parasitic Possession Hypothesis in Experiments 2 and 3. The
181
Complexity Hypothesis was originally formulated based on the assumption that the addition of a
possessor to the noun phrase headed by the possessed referent would increase the informativity
of the possessed referent. This logic makes sense based on the syntax of English s-genitives. For
example, the possessor in Sam’s car is syntactically contained within a larger nominal phrase
referring to the possessed referent car; however, the notion of dependent possessed referents at
the discourse level of representation has the opposite hierarchical structure. In this case, the
possessed referent is in some way embedded within the representation of the possessor. Under
this view, it could be argued that the possession relation is instead adding to the representational
complexity of the possessor, not the possessed referent. If the possessor becomes more
prominent as a result, the relative prominence of the possessed referent will decrease. In this
way, the same net effect predicted by the Parasitic Possession Hypothesis may arise. Thus the
lack of evidence in this dissertation for the Complexity Hypothesis may not represent incorrect
reasoning about the effect of representational complexity, but rather an unexpected difference
between how this type of complexity is realized at the syntax and discourse levels of
representation.

6.4. Opportunities for future work
This dissertation has brought up numerous avenues for potential future work. One such
opportunity lies in further exploration of the Interaction Hypothesis. While Experiments 1 and 4
(Chapters 2 and 5) contained support for the theory that possessed animates are especially
prominent in discourse, it was notable that Experiment 2 (Chapter 3) did not find the same sort of
interaction between animacy and possession which would have supported the Interaction
Hypothesis. The primary difference between Experiments 1 and 2 was that the former compared
182
possessives to indefinites, and the latter compared them to definites. In order to strengthen the
argument in favor of the Interaction Hypothesis, it would be ideal in future work to produce this
sort of interaction in a comparison of possessives and definites. Without this evidence, it might
still be possible to argue that the results of Experiment 1 were driven by discourse-level
differences between indefinites and possessives (e.g. givenness, specificity) which may have
varied between the animate and inanimate conditions. Despite the existence of this argument,
there is no indication that those differences between indefinites and possessives did in fact vary
with animacy. Furthermore, convergent evidence for the Interaction Hypothesis in Chapter 5
makes this potential worry seem less serious.
Additionally, given that possessives are a broad class of constructions, there are many
opportunities to widen the scope of inquiry beyond the limited number of possessive
constructions and semantic possession relations investigated in this dissertation. With regard to
other syntactic configurations, this dissertation notably did not address English’s alternative
nominal possessive configuration, the of-genitive (e.g. the owner of the car versus s-genitive the
car’s owner). As I explained in Chapter 1, this omission was necessary for the purposes of my
experiments because the of-genitive can be degraded with animate and pronominal possessors
(e.g. Sam’s car, his car, #the car of Sam, #the car of him; the house’s roof, its roof, the roof of
the house, #the roof of it). Nevertheless, it would be beneficial in future work to explore whether
the hypotheses supported in this dissertation can be applied to of-genitives. Additional syntactic
variants which are also worthy of future consideration are verbal possessives (e.g. Sam has a
car), non-pronominal possessors (as the experiments in this dissertation always used possessive
pronouns his or her to refer to the possessor), and other types of pronominal possessors (e.g.
their car, your car, my car).
183
Finally, more work is needed to test additional possession relations between the possessor
and possessed referent. In exploring the Interaction Hypothesis, this dissertation provides a
reasonable starting point for showing that the type of possession relation may affect the discourse
prominence of the involved referents. However, possessives express a wide variety of relations
(even if the domain of inquiry is limited just to English), and this dissertation was only able to
test a limited number of them. While Experiments 4 and 5 included four types of possession
relations, Langacker (1995) counts at least 18 such semantic relations (in English) in what is
likely not an exhaustive list. Future work ought to test more of these types of relations between
the possessor and possessed referent to see how they affect discourse prominence. Furthermore,
it could be fruitful to examine how dimensions shared by multiple types of possession relations
could be mapped onto discourse prominence in order to make predictions about how different
classes of relations might affect prominence. Chapter 5 takes a first step in this direction by
arguing that the difference observed between the kinship and non-kinship animate relational
nouns might be driven by the varying non-linguistic importance of these two types of
interpersonal relationships.

184
References

Allopenna, P. D., Magnuson, J. S., & Tanenhaus, M. K. (1998). Tracking the time course of
spoken word recognition using eye movements: Evidence for continuous mapping
models. Journal of Memory and Language, 38(4), 419-439. doi:10.1006/jmla.1997.2558
Altmann, G., & Steedman, M. (1988). Interaction with context during human sentence
processing. Cognition, 30(3), 191-238. doi:10.1016/0010-0277(88)90020-0
Anschutz, A. (1997). How to choose a possessive noun phrase construction in four easy
steps. Studies in Language, 21(1), 1-35. doi:10.1075/sl.21.1.02ans
Ariel, M. (1988). Referring and accessibility. Journal of Linguistics, 24(1), 65-87.
doi:10.1017/S0022226700011567
Ariel, M. (1990). Accessing noun-phrase antecedents. London: Routledge.
Arnold, J. E. (2001). The effect of thematic roles on pronoun use and frequency of reference
continuation. Discourse Processes, 31(2), 137-162. doi:10.1207/S15326950DP3102_02
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., & Griffin, Z. M. (2007). The effect of additional characters on choice of referring
expression: Everyone counts. Journal of Memory and Language, 56(4), 521-536.
doi:10.1016/j.jml.2006.09.007
Asher, N., Hardt, D., & Busquets, J. (2001). Discourse parallelism, ellipsis, and
ambiguity. Journal of Semantics, 18(1), 1-25. doi:10.1093/jos/18.1.1
Avrutin, S. (1999). Development of the syntax-discourse interface. Dordrecht: Kluwer
Badecker, W., & Straub, K. (2002). The processing role of structural constraints on the
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, C. L. (1966). Definiteness and indefiniteness in English (MA thesis, University of
Illinois).
Barker, C. (2000). Definite possessives and discourse novelty. Theoretical Linguistics, 26(3),
211. doi:10.1515/thli.2000.26.3.211
Barker, C. (2011). Possessives and relational nouns. In K. von Heusinger, C. Maienborn, & P.
Portner (Eds). Semantics: An international handbook of natural language meaning (pp.
1109-1130). Berlin: de Gruyter Mouton.
185
Barr, D. J., Levy, R., Scheepers, C., & Tily, H. J. (2013). Random effects structure for
confirmatory hypothesis testing: Keep it maximal. Journal of Memory and
Language, 68(3), 255-278. doi:10.1016/j.jml.2012.11.001
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
Bock, J. K., & Irwin, D. E. (1980). Syntactic effects of information availability in sentence
production. Journal of Verbal Learning and Verbal Behavior, 19(4), 467-484.
doi:10.1016/S0022-5371(80)90321-7
Bock, J. K., Irwin, D. E., & Davidson, D. J. (2004). Putting first things first. In J. Henderson, &
F. Ferreira (Eds.), The integration of language, vision, and action: Eye movements and
the visual world (pp. 249-278). New York: Psychology Press.
Bock, J. K., & Levelt, W. J. M. (1994). Language production: Grammatical encoding. In M. A.
Gernsbacher (Ed.), Handbook of psycholinguistics (pp. 945-984). San Diego, CA:
Academic Press.
Bock, K., Loebell, H., & Morey, R. (1992). From conceptual roles to structural relations:
Bridging the syntactic cleft. Psychological Review, 99(1), 150-171. doi:10.1037//0033-
295X.99.1.150
Boersma, P. (2002). Praat, a system for doing phonetics by computer [Computer software].
Amsterdam: University of Amsterdam.
Bonin, P., Gelin, M., & Bugaiska, A. (2014). Animates are better remembered than inanimates:
Further evidence from word and picture stimuli. Memory & Cognition, 42(3), 370-382.
doi:10.3758/s13421-013-0368-8
Branigan, H. P., Pickering, M. J., & Tanaka, M. (2008). Contributions of animacy to
grammatical function assignment and word order during production. Lingua, 118(2), 172-
189. doi:10.1016/j.lingua.2007.02.003
Brysbaert, M., & New, B. (2009). Moving beyond Kučera and Francis: A critical evaluation of
current word frequency norms and the introduction of a new and improved word
frequency measure for American English. Behavior Research Methods, 41(4), 977-990.
doi:10.3758/brm.41.4.977
Cacioppo, J. T., & Hawkley, L. C. (2009). Perceived social isolation and cognition. Trends in
Cognitive Sciences, 13(10), 447-454. doi:10.1016/j.tics.2009.06.005
Cacioppo, J. T., Hughes, M. E., Waite, L. J., Hawkley, L. C., & Thisted, R. A. (2006).
Loneliness as a specific risk factor for depressive symptoms. Psychology and
Aging, 21(1), 140-151. doi:10.1037/0882-7974.21.1.140
Carter, B. T., & Luke, S. G. (2020). Best practices in eye tracking research. International
Journal of Psychophysiology, 155, 49-62. doi:10.1016/j.ijpsycho.2020.05.010
186
Chae, S. (2003). Possessives in naturally occurring discourse: A centering approach. University
of Pennsylvania Working Papers in Linguistics, 9(1). Retrieved from
https://repository.upenn.edu/pwpl/vol9/iss1/6
Chafe, W. L. (1976). Giveness, contrastiveness, definiteness, subjects, topics, and point of view.
In C. N. Li (Ed.), Subject and topic. New York: Academic Press.
Chafe, W. L. (1974). Language and consciousness. Language, 50(1), 111-133.
doi:10.2307/412014
Chiriacescu, S. (2015). Effects of animacy and gender on the choice of referring
expressions. Bulletin of the Transilvania University of Brașov, Series IV: Philology &
Cultural Studies, 8(1), 19-28.
Clark, H. H., & Begun, J. S. (1971). The semantics of sentence subjects. Language and
Speech, 14(1), 34-46. doi:10.1177/002383097101400105
Clark, H. H., & Sengul, C. J. (1979). In search of referents for nouns and pronouns. Memory &
Cognition, 7(1), 35-41. doi:10.3758/BF03196932
Clark, H., & Haviland, S. (1977). Comprehension and the given-new contract. In R. O. Freedle
(Ed.), Discourse production and comprehension (pp. 1-40). Norwood, NJ: Ablex.
Cooper, R. M. (1974). The control of eye fixation by the meaning of spoken language: A new
methodology for the real-time investigation of speech perception, memory, and language
processing. Cognitive Psychology, 6(1), 84-107. doi:10.1016/0010-0285(74)90005-X
Cozijn, R., Commandeur, E., Vonk, W., & Noordman, L. G. M. (2011). The time course of the
use of implicit causality information in the processing of pronouns: A visual world
paradigm study. Journal of Memory and Language, 64(4), 381-403.
doi:10.1016/j.jml.2011.01.001
Craik, F. I., & Tulving, E. (1975). Depth of processing and the retention of words in episodic
memory. Journal of Experimental Psychology. General, 104(3), 268-294.
doi:10.1037/0096-3445.104.3.268
Crawley, R. A., Stevenson, R. J., & Kleinman, D. (1990). The use of heuristic strategies in the
interpretation of pronouns. Journal of Psycholinguistic Research, 19(4), 245-264.
doi:10.1007/bf01077259
Dahl, Ö. (2008). Animacy and egophoricity: Grammar, ontology and phylogeny. Lingua, 118(2),
141-150. doi:10.1016/j.lingua.2007.02.008
Dahl, Ö., & Fraurud, K. (Eds.). (1996). Animacy in grammar and discourse. In T. Fretheim, & J.
K. Gundel (Eds.), Reference and Referent Accessibility. Pragmatics & Beyond New
Series, (pp. 47-64). Amsterdam: John Benjamins.
187
Degen, J., Kursat, L., & Leigh, D. D. (2021). Seeing is believing: Testing an explicit linking
assumption for visual world eye-tracking in psycholinguistics. Proceedings of the Annual
Meeting of the Cognitive Science Society, (43). Retrieved
from https://escholarship.org/uc/item/6182t9jb
Di Eugenio, B. (1998). Centering in Italian. In M. Walker, A. Joshi & E. Prince (Eds.), Centering
theory in discourse (pp. 115-137). New York: Oxford University Press.
Downing, P. E., Bray, D., Rogers, J., & Childs, C. (2004). Bodies capture attention when nothing
is expected. Cognition, 93(1), B27-B38. doi:10.1016/j.cognition.2003.10.010
Du Bois, J. W. (1980). Beyond definiteness: The trace of identity in discourse. In W. L. Chafe
(Ed.), The pear stories: Cognitive, cultural, and linguistic aspects of narrative
production (pp. 203-274). Norwood, NJ: Ablex.
Eisenberger, N. I., & Cole, S. W. (2012). Social neuroscience and health: Neurophysiological
mechanisms linking social ties with physical health. Nature Neuroscience, 15(5), 669-
674. doi:10.1038/nn.3086
Fiengo, R., & May, R. (1994). Indices and identity. Cambridge: MIT Press.
Fisher, R. P., & Craik, F. I. (1980). The effects of elaboration on recognition memory. Memory
& Cognition, 8(5), 400-404. doi:10.3758/BF03211136
Foley, C., Núñez del Prado, Z., Barbier, I., & Lust, B. (2003). Knowledge of variable binding in
VP–Ellipsis: Language acquisition research and theory converge. Syntax, 6(1), 52-83.
doi:10.1111/1467-9612.00056
Frazier, L., & Clifton Jr, C. (2000). On bound variable interpretations: The LF-only
hypothesis. Journal of Psycholinguistic Research, 29(2), 125-140.
doi:10.1023/A:1005136826534
Fukumura, K., & van Gompel, R. P. G. (2010). Choosing anaphoric expressions: Do people take
into account likelihood of reference? Journal of Memory and Language, 62(1), 52-66.
doi:10.1016/j.jml.2009.09.001
Fukumura, K., & van Gompel, R. P. G. (2011). The effect of animacy on the choice of referring
expression. Language and Cognitive Processes, 26(10), 1472-1504.
doi:10.1080/01690965.2010.506444
Gelin, M., Bonin, P., Méot, A., & Bugaiska, A. (2018). Do animacy effects persist in memory for
context? Quarterly Journal of Experimental Psychology (2006), 71(4), 965-974.
doi:10.1080/17470218.2017.1307866
Givón, T. (1983). Topic continuity in discourse: A quantitative cross-language study.
Philadelphia: John Benjamins Publishing Company.
188
Gordon, P. C., Grosz, B. J., & Gilliom, L. A. (1993). Pronouns, names, and the centering of
attention in discourse. Cognitive science, 17(3), 311-347.
Grodzinsky, Y., Wexler, K., Chien, Y. C., Marakovitz, S., & Solomon, J. (1993). The breakdown
of binding relations. Brain and Language, 45(3), 396-422. doi:10.1006/brln.1993.1052
Grosz, B. J., Joshi, A. K., & Weinstein, S. (1995). Centering: A framework for modeling the
local coherence of discourse. Computational Linguistics, 21(2), 203-225.
Grüter, T., Rohde, H., & Schafer, A. J. (2017). Coreference and discourse coherence in L2: The
roles of grammatical aspect and referential form. Linguistic Approaches to
Bilingualism, 7(2), 199-229. doi:10.1075/lab.15011.gru
Gundel, J. K., Hedberg, N., & Zacharski, R. (1993). Cognitive status and the form of referring
expressions in discourse. Language, 69(2), 274-307. doi:10.2307/416535
Guo, F., Foley, C., Chien, Y., Chiang, C., & Lust, B. (1996). Operator-variable binding in the
initial state: A cross-linguistic study of VP ellipsis structures in Chinese and
English. Cahiers De Linguistique - Asie Orientale, 25(1), 3-34.
doi:10.3406/clao.1996.1490
Hallett, P. E. (1986). Eye movements. In K. R. Boff, L. Kaufman & J. P. Thomas
(Eds.), Handbook of perception and human performance (pp. 10-112). New York: Wiley.
Hardt, D., & Romero, M. (2004). Ellipsis and the structure of discourse. Journal of
Semantics, 21(4), 375-414. doi:10.1093/jos/21.4.375
Hartshorne, J. K., & Snedeker, J. (2013). Verb argument structure predicts implicit causality:
The advantages of finer-grained semantics. Language and Cognitive Processes, 28(10),
1474-1508. doi:10.1080/01690965.2012.689305
Haspelmath, M. (2017). Explaining alienability contrasts in adpossessive constructions:
Predictability vs. iconicity. Zeitschrift für Sprachwissenschaft, 36(2), 193-231. doi:
10.1515/zfs-2017-0009
Heim, I. R. (1982). The semantics of definite and indefinite noun phrases (Doctoral dissertation,
University of Massachusetts).
Himmelmann, N., & Primus, B. (2015). Prominence beyond prosody – a first approximation. In
A. De Dominicis (Ed.), Prominence in linguistics. Proceedings of the PS-prominenceS
International Conference (pp. 38-58) Viterbo: DISUCOM Press.
Hobbs, J. R. (1979). Coherence and coreference. Cognitive Science, 3(1), 67-90.
doi:10.1207/s15516709cog0301_4
Hofmeister, P. (2011). Representational complexity and memory retrieval in language
comprehension. Language and Cognitive Processes, 26(3), 376-405.
doi:10.1080/01690965.2010.492642
189
Hofmeister, P., & Vasishth, S. (2014). Distinctiveness and encoding effects in online sentence
comprehension. Frontiers in Psychology, 5, 1237. doi:10.3389/fpsyg.2014.01237
Holt-Lunstad, J., Smith, T. B., & Layton, J. B. (2010). Social relationships and mortality risk: A
meta-analytic review. PLoS Medicine, 7(7), e1000316.
doi:10.1371/journal.pmed.1000316
Itagaki, N., & Prideaux, G. D. (1985). Nominal properties as determinants of subject
selection. Lingua, 66(2), 135-149. doi:10.1016/S0024-3841(85)90292-X
Järvikivi, J., van Gompel, R. P. G., Hyönä, J., & Bertram, R. (2005). Ambiguous pronoun
resolution: Contrasting the first-mention and subject-preference accounts. Psychological
Science, 16(4), 260-264. doi:10.1111/j.0956-7976.2005.01525.x
Jasinskaja, K., Chiriacescu, S., Donazzan, M., von Heusinger, K., & Hinterwimmer, S. (2015).
Prominence in discourse. In A. Dominicis (Ed.), Prominence in linguistics. Proceedings
of the PS-prominenceS International Conference (pp. 134-153). Viterbo: DISUCOM
Press.
Kaiser, E. (2009). Investigating effects of structural and information-structural factors on
pronoun resolution. In M. Zimmermann, & C. Féry (Eds.), Information structure from
different perspectives (pp. 332-353). New York: Oxford University Press.
doi:10.1093/acprof:oso/9780199570959.003.0014
Kaiser, E. (2010). Investigating the consequences of focus on the production and comprehension
of referring expressions. International Review of Pragmatics, 2(2), 266-297.
doi:10.1163/187731010X528368
Kaiser, E. (2011). Focusing on pronouns: Consequences of subjecthood, pronominalisation, and
contrastive focus. Language and Cognitive Processes, 26(10), 1625-1666.
doi:10.1080/01690965.2010.523082
Kaiser, E., & Trueswell, J. C. (2008). Interpreting pronouns and demonstratives in Finnish:
Evidence for a form-specific approach to reference resolution. Language and Cognitive
Processes, 23(5), 709-748. doi:10.1080/01690960701771220
Kamp, H. (1981). A theory of truth and semantic representation. In J. Groendijk, T. Janssen, &
M. Stockhof (Eds.), Formal Methods in the Study of Language: Part 1 (pp. 177-321).
Amsterdam: Mathematical Centre.
Kamp, H., & Reyle, U. (1993). From discourse to logic. Dordrecht: Kluwer.
Karimi, H., Diaz, M., & Ferreira, F. (2019). "A cruel king" is not the same as "a king who is
cruel": Modifier position affects how words are encoded and retrieved from
memory. Journal of Experimental Psychology. Learning, Memory, and
Cognition, 45(11), 2010-2035. doi:10.1037/xlm0000694
190
Karimi, H., & Ferreira, F. (2016). Informativity renders a referent more accessible: Evidence
from eyetracking. Psychonomic Bulletin & Review, 23(2), 507-525. doi:10.3758/s13423-
015-0917-1
Karimi, H., Fukumura, K., Ferreira, F., & Pickering, M. (2014). The effect of noun phrase length
on the form of referring expressions. Memory & Cognition, 42(6), 993-1009.
doi:10.3758/s13421-014-0400-7
Karttunen, L. (1968). What do referential indices refer to?. Santa Monica, Calif: Rand.
Kehler, A. (2002). Coherence, reference, and the theory of grammar. Stanford, Calif: CSLI Publ.
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), 1-37.
doi:10.1515/tl-2013-0001
Kertz, L. (2013). Verb phrase ellipsis: The view from information structure. Language
(Baltimore), 89(3), 390-428. doi:10.1353/lan.2013.0051
Koornneef, A. W., Avrutin, S., Wijnen, F., & Reuland, E. (2011). Tracking the preference for
bound-variable dependencies in ambiguous ellipses and only-structures. Syntax and
semantics (pp. 67-100) doi:10.1163/9781780523750_004
Langacker, R. (1995). Possession and possessive constructions. In R. E. MacLaury, & J. R.
Taylor (Eds.), Language and the cognitive construal of the world (pp 51-79). Berlin: de
Gruyter Mouton.
Langacker, R. W. (1993). Reference-point constructions. Cognitive Linguistics, 4(1), 1-38.
doi:10.1515/cogl.1993.4.1.1
Lasersohn, P. (2005). Context dependence, disagreement, and predicates of personal
taste. Linguistics and Philosophy, 28(6), 643-686. doi:10.1007/s10988-005-0596-x
Macwhinney, B., & Bates, E. (1978). Sentential devices for conveying givenness and newness:
A cross-cultural developmental study. Journal of Verbal Learning and Verbal
Behavior, 17(5), 539-558.
McCawley, J. D. (1979). Presupposition and discourse structure. In C. K. Oh, & D. A. Dinneen
(Eds.), Syntax and semantics 11: Presupposition (pp. 371-388). New York: Academic
Press.
Miller, G., Chen, E., & Cole, S. W. (2009). Health psychology: Developing biologically
plausible models linking the social world and physical health. Annual Review of
Psychology, 60(1), 501-524. doi:10.1146/annurev.psych.60.110707.163551
191
Misersky, J., Gygax, P. M., Canal, P., Gabriel, U., Garnham, A., Braun, F., Chiarini, T.,
Englund, K., Hanulikova, A., Öttl, A., Valdrova, J., Von Stockhausen, L., & Sczesny, S.
(2014). Norms on the gender perception of role nouns in Czech, English, French,
German, Italian, Norwegian, and Slovak. Behavior Research Methods, 46(3), 841-871.
doi:10.3758/s13428-013-0409-z
Nairne, J. S. (2010). Adaptive memory: Evolutionary constraints on remembering. In B. H. Ross
(Ed.), Psychology of learning and motivation (pp. 1-32). San Diego, CA: Academic
Press. doi:10.1016/S0079-7421(10)53001-9
Nairne, J. S., VanArsdall, J. E., Pandeirada, J. N. S., Cogdill, M., & LeBreton, J. M. (2013).
Adaptive memory: The mnemonic value of animacy. Psychological Science, 24(10),
2099-2105. doi:10.1177/0956797613480803
New, J., Cosmides, L., & Tooby, J. (2007). Category-specific attention for animals reflects
ancestral priorities, not expertise. Proceedings of the National Academy of Sciences -
PNAS, 104(42), 16598-16603. doi:10.1073/pnas.0703913104
Oku, S. A theory of selection and reconstruction in the minimalist perspective (Doctoral
dissertation, University of Connecticut).
Onea, E., & Geist, L. (2011). Indefinite determiners and the pragmatics of referential
anchoring. International Review of Pragmatics, 3(2), 194-227.
doi:10.1163/187731011X597514
Ong, M., & Brasoveanu, A. (2014). Strict and sloppy reflexives in VP ellipsis. In C. Piñón
(Ed.), Empirical issues in syntax and semantics 10 (pp. 251–268).
Papoutsaki, A., Sangkloy, P., Laskey, J., Daskalova, N., Huang, J., & James Hays. (2016).
WebGazer: Scalable webcam eye tracking using user interactions. Proceedings of the
25th International Joint Conference on Artificial Intelligence (IJCAI) (pp. 3839-3845).
Prat-Sala, M., & Branigan, H. P. (2000). Discourse constraints on syntactic processing in
language production: A cross-linguistic study in English and Spanish. Journal of Memory
and Language, 42(2), 168-182. doi:10.1006/jmla.1999.2668
Prince, E. F. (1981). Toward a taxonomy of given-new information. In P. Cole (Ed.), Radical
pragmatics (pp. 233-255). New York: Academic Press.
Prince, E. F. (1992). The ZPG letter: Subjects, definiteness, and information-status. In S. A.
Thompson, & W. Mann (Eds.), Discourse description: Diverse linguistic analyses of a
fund-raising text (pp. 295-325). Philadelphia: John Benjamins.
Quine, W. V. (1956). Quantifiers and propositional attitudes. The Journal of Philosophy, 53(5),
177-187. doi:10.2307/2022451
R Core Team (2020). R: A language and environment for statistical computing. R Foundation for
Statistical Computing, Vienna, Austria.
192
Reinhart, T. (2000). Strategies of anaphora resolution. In H. Bennis, M. Everaert & E. J. Reuland
(Eds.), Interface strategies (pp. 295–325). Amsterdam: Royal Academy of Sciences.
Reuland, E. (2001). Primitives of binding. Linguistic Inquiry, 32(3), 439-492.
doi:10.1162/002438901750372522
Rosenbach, A. (2002). Genitive variation in English. Berlin/Boston: De Gruyter, Inc.
Rubio-Fernández, P. (2016). How redundant are redundant color adjectives? an efficiency-based
analysis of color overspecification. Frontiers in Psychology, 7, 153.
doi:10.3389/fpsyg.2016.00153
Sag, I. A. (1976). Deletion and logical form (Doctoral dissertation, Massachusetts Institute of
Technology).
Slim, M. S., & Hartsuiker, R. (2021). Moving visual world experiments online? A web-based
replication of Dijkgraaf, Hartsuiker, and Duyck (2017) using PCIbex and WebGazer.js.
doi:10.31234/osf.io/5adgf
Stevenson, R. J., Crawley, R. A., & Kleinman, D. (1994). Thematic roles, focus and the
representation of events. Language and Cognitive Processes, 9(4), 519-548.
doi:10.1080/01690969408402130
Stewart, A. J., Holler, J., & Kidd, E. (2007). Shallow processing of ambiguous pronouns:
Evidence for delay. Quarterly Journal of Experimental Psychology (2006), 60(12), 1680-
1696. doi:10.1080/17470210601160807
Stojanovic, I. (2007). Talking about taste: Disagreement, implicit arguments, and relative
truth. Linguistics and Philosophy, 30(6), 691-706. doi:10.1007/s10988-008-9030-5
Tamina Stephenson. (2007). Judge dependence, epistemic modals, and predicates of personal
taste. Linguistics and Philosophy, 30(4), 487-525. doi:10.1007/s10988-008-9023-4
Tanenhaus, M. K., Spivey-Knowlton, M. J., Eberhard, K. M., & Sedivy, J. C. (1995). Integration
of visual and linguistic information in spoken language
comprehension. Science, 268(5217), 1632-1634. doi:10.1126/science.7777863
Tarenskeen, S. L., Broersma, M., & Geurts, B. (2015). Overspecification of color, pattern, and
size: Salience, absoluteness, and consistency. Frontiers in Psychology, 6, 1703.
doi:10.3389/fpsyg.2015.01703
Tilvis, R. S., Kahonen-Vare, M. H., Jolkkonen, J., Valvanne, J., Pitkala, K. H., & Strandberg, T.
E. (2004). Predictors of cognitive decline and mortality of aged people over a 10-year
period. The Journals of Gerontology. Series A, Biological Sciences and Medical
Sciences, 59(3), M268-M274. doi:10.1093/gerona/59.3.m268
193
Tourtouri, E. N., Delogu, F., Sikos, L., & Crocker, M. W. (2019). Rational over-specification in
visually-situated comprehension and production. Journal of Cultural Cognitive
Science, 3(2), 175-202. doi:10.1007/s41809-019-00032-6
Troyer, M., Hofmeister, P., & Kutas, M. (2016). Elaboration over a discourse facilitates retrieval
in sentence processing. Frontiers in Psychology, 7, 374. doi:10.3389/fpsyg.2016.00374
Tyler, L. K., & Marslen-Wilson, W. D. (1977). The on-line effects of semantic context on
syntactic processing. Journal of Verbal Learning and Verbal Behavior, 16(6), 683-692.
doi:10.1016/S0022-5371(77)80027-3
van den Broek, P., Risden, K., Fletcher, C. R., & Thurlow, R. (1996). A “landscape” view of
reading: Fluctuating patterns of activation and the construction of a stable memory
representation. In B. K. Britton & A. C. Graesser (Eds.), Models of understanding
text (pp. 173-196). New York: Psychology Press. doi:10.4324/9781315806143-11
VanArsdall, J. E., Nairne, J. S., Pandeirada, J. N. S., & Blunt, J. R. (2012). Adaptive
memory: Animacy processing produces mnemonic advantages. Experimental
Psychology, 60(3), 172-178. doi:10.1027/1618-3169/a000186
Vasic, N., Avrutin, S., & Ruigendijk, E. (2006). Interpretation of pronouns in VP-ellipsis
constructions in Dutch Broca's and Wernicke's aphasia. Brain and Language, 96(2), 191.
doi:10.1016/j.bandl.2005.04.003
von Heusinger, K. (2011). Specificity. In K. von Heusinger, C. Maienborn & P. Portner
(Eds.), Semantics: An international handbook of natural language meaning (pp. 1025-
1057). Berlin: Mouton de Gruyter.
von Heusinger, K., & Chiriacescu, S. (2010). Discourse prominence and pe-marking in
Romanian. International Review of Pragmatics, 2(2), 298-332.
doi:10.1163/187731010X528377
von Heusinger, K., & Schumacher, P. B. (2019). Discourse prominence: Definition and
application. Journal of Pragmatics, 154, 117-127. doi:10.1016/j.pragma.2019.07.025
Vos, M., Minor, S., & Ramchand, G. C. (2022). Comparing infrared and webcam eye tracking in
the Visual World Paradigm. Glossa Psycholinguistics, 1(1). doi:10.5070/G6011131
Walker, M. A., & Prince, E. F. (1996). A bilateral approach to givenness: A hearer-status
algorithm and a centering algorithm. In T. Fretheim & J. K. Gundel (Eds.), Reference and
referent accessibility (Pragmatics & Beyond New Series 38) (pp. 291-306). Amsterdam:
John Benjamins.
Walker, M., Joshi, A., & Prince, E. (1998). Centering theory in discourse. New York: Oxford
University Press.
194
Xu, P., Ehinger, K. A., Zhang, Y., Finkelstein, A., Kulkarni, S. R., & Xiao, J. (2015).
Turkergaze: Crowdsourcing saliency with webcam based eye tracking.
arXiv:1504.06755.
Yamamoto, M. (1999). Animacy and reference: A cognitive approach to corpus linguistics.
Amsterdam: John Benjamins.
Yoshihisa Kitagawa. (1991). Copying identity. Natural Language & Linguistic Theory, 9(3),
497-536. doi:10.1007/BF00135356
Zehr, J., & Schwarz, F. (2018). PennController for Internet Based Experiments (IBEX).
doi:10.17605/OSF.IO/MD832

195
APPENDIX A
Target Stimuli for Experiment 1

Jason joiped a congresswoman.
Jason joiped his congresswoman.
Jason joiped a wheelbarrow.
Jason joiped his wheelbarrow.
Brian sibbed a waitress.
Brian sibbed his waitress.
Brian sibbed a basket.
Brian sibbed his basket.
Robert dasped a secretary.
Robert dasped his secretary.
Robert dasped a television.
Robert dasped his television.
Daniel zatted a nurse.
Daniel zatted his nurse.
Daniel zatted a jacket.
Daniel zatted his jacket.
Helen swudged a tattooist.
Helen swudged her tattooist.
Helen swudged a unicycle.
Helen swudged her unicycle.
196
Megan gerped a chiropractor.
Megan gerped her chiropractor.
Megan gerped a clarinet.
Megan gerped her clarinet.
Emily pletched a stockbroker.
Emily pletched her stockbroker.
Emily pletched a saxophone.
Emily pletched her saxophone.
Jessica rulked an electrician.
Jessica rulked her electrician.
Jessica rulked a chandelier.
Jessica rulked her chandelier.
Melissa lupted a chauffeur.
Melissa lupted her chauffeur.
Melissa lupted a backpack.
Melissa lupted her backpack.
Sarah chooped a plumber.
Sarah chooped her plumber.
Sarah chooped a scarf.
Sarah chooped her scarf.
Rebecca gweeshed a bodyguard.
Rebecca gweeshed her bodyguard.
Rebecca gweeshed an umbrella.
197
Rebecca gweeshed her umbrella.
Brad churbed a receptionist.
Brad churbed his receptionist.
Brad churbed a screwdriver.
Brad churbed his screwdriver.
Carol debbed a butler.
Carol debbed her butler.
Carol debbed a violin.
Carol debbed her violin.
Stephanie shupped a mechanic.
Stephanie shupped her mechanic.
Stephanie shupped a flashlight.
Stephanie shupped her flashlight.
Kimberly chabbed a chef.
Kimberly chabbed her chef.
Kimberly chabbed a fork.
Kimberly chabbed her fork.
Jennifer pranned a surgeon.
Jennifer pranned her surgeon.
Jennifer pranned a blanket.
Jennifer pranned her blanket.
Veronica brilted a priest.
Veronica brilted her priest.
198
Veronica brilted a towel.
Veronica brilted her towel.
James fammed a hairdresser.
James fammed his hairdresser.
James fammed a thermometer.
James fammed his thermometer.
Anthony jolfed a florist.
Anthony jolfed his florist.
Anthony jolfed a toaster.
Anthony jolfed his toaster.
Kevin blorned a stewardess.
Kevin blorned his stewardess.
Kevin blorned a dictionary.
Kevin blorned his dictionary.
Joseph tammed a babysitter.
Joseph tammed his babysitter.
Joseph tammed a microwave.
Joseph tammed his microwave.
Mark meared a housekeeper.
Mark meared his housekeeper.
Mark meared a stereo.
Mark meared his stereo.
Fred risped a nanny.
199
Fred risped his nanny.
Fred risped a broom.
Fred risped his broom.
William dreezed a psychic.
William dreezed his psychic.
William dreezed a helmet.
William dreezed his helmet.

200
APPENDIX B
Target Stimuli for Experiment 2

The dressmaker saw her chandelier because
The dressmaker saw her brass chandelier because
The dressmaker saw her electrician because
The dressmaker saw her bald electrician because
The dressmaker saw the chandelier because
The dressmaker saw the brass chandelier because
The dressmaker saw the electrician because
The dressmaker saw the bald electrician because
The bride noted her mailbox because
The bride noted her rusted mailbox because
The bride noted her chauffeur because
The bride noted her Canadian chauffeur because
The bride noted the mailbox because
The bride noted the rusted mailbox because
The bride noted the chauffeur because
The bride noted the Canadian chauffeur because
The manicurist heard her dryer because
The manicurist heard her electric dryer because
The manicurist heard her plumber because
The manicurist heard her Southern plumber because
201
The manicurist heard the dryer because
The manicurist heard the electric dryer because
The manicurist heard the plumber because
The manicurist heard the Southern plumber because
The actress lost her jewel because
The actress lost her priceless jewel because
The actress lost her butler because
The actress lost her clean-shaven butler because
The actress lost the jewel because
The actress lost the priceless jewel because
The actress lost the butler because
The actress lost the clean-shaven butler because
The wrestler replaced his microwave because
The wrestler replaced his broken microwave because
The wrestler replaced his housekeeper because
The wrestler replaced his missing housekeeper because
The wrestler replaced the microwave because
The wrestler replaced the broken microwave because
The wrestler replaced the housekeeper because
The wrestler replaced the missing housekeeper because
The butcher praised his refrigerator because
The butcher praised his industrial refrigerator because
The butcher praised his psychic because
202
The butcher praised his widowed psychic because
The butcher praised the refrigerator because
The butcher praised the industrial refrigerator because
The butcher praised the psychic because
The butcher praised the widowed psychic because
The boxer saved his sandwich because
The boxer saved his expired sandwich because
The boxer saved his secretary because
The boxer saved his part-time secretary because
The boxer saved the sandwich because
The boxer saved the expired sandwich because
The boxer saved the secretary because
The boxer saved the part-time secretary because
The pilot forgot his handkerchief because
The pilot forgot his embroidered handkerchief because
The pilot forgot his babysitter because
The pilot forgot his Australian babysitter because
The pilot forgot the handkerchief because
The pilot forgot the embroidered handkerchief because
The pilot forgot the babysitter because
The pilot forgot the Australian babysitter because
The janitor found his cellphone because
The janitor found his stolen cellphone because
203
The janitor found his stewardess because
The janitor found his brown-eyed stewardess because
The janitor found the cellphone because
The janitor found the stolen cellphone because
The janitor found the stewardess because
The janitor found the brown-eyed stewardess because
The fisherman reviewed his screwdriver because
The fisherman reviewed his pocket-sized screwdriver because
The fisherman reviewed his hairdresser because
The fisherman reviewed his self-employed hairdresser because
The fisherman reviewed the screwdriver because
The fisherman reviewed the pocket-sized screwdriver because
The fisherman reviewed the hairdresser because
The fisherman reviewed the self-employed hairdresser because
The firefighter missed his telescope because
The firefighter missed his antique telescope because
The firefighter missed his florist because
The firefighter missed his award-winning florist because
The firefighter missed the telescope because
The firefighter missed the antique telescope because
The firefighter missed the florist because
The firefighter missed the award-winning florist because
The ballerina discussed her blanket because
204
The ballerina discussed her hand-sewn blanket because
The ballerina discussed her surgeon because
The ballerina discussed her ambidextrous surgeon because
The ballerina discussed the blanket because
The ballerina discussed the hand-sewn blanket because
The ballerina discussed the surgeon because
The ballerina discussed the ambidextrous surgeon because
The farmer evaluated his wineglass because
The farmer evaluated his plastic wineglass because
The farmer evaluated his congresswoman because
The farmer evaluated his vegetarian congresswoman because
The farmer evaluated the wineglass because
The farmer evaluated the plastic wineglass because
The farmer evaluated the congresswoman because
The farmer evaluated the vegetarian congresswoman because
The magician observed his fireplace because
The magician observed his marble fireplace because
The magician observed his nanny because
The magician observed his English nanny because
The magician observed the fireplace because
The magician observed the marble fireplace because
The magician observed the nanny because
The magician observed the English nanny because
205
The sheriff corrected his thermometer because
The sheriff corrected his digital thermometer because
The sheriff corrected his receptionist because
The sheriff corrected his temporary receptionist because
The sheriff corrected the thermometer because
The sheriff corrected the digital thermometer because
The sheriff corrected the receptionist because
The sheriff corrected the temporary receptionist because
The soldier fetched his basket because
The soldier fetched his empty basket because
The soldier fetched his waitress because
The soldier fetched his short-haired waitress because
The soldier fetched the basket because
The soldier fetched the empty basket because
The soldier fetched the waitress because
The soldier fetched the short-haired waitress because
The carpenter judged his cake because
The carpenter judged his gluten-free cake because
The carpenter judged his nurse because
The carpenter judged his left-handed nurse because
The carpenter judged the cake because
The carpenter judged the gluten-free cake because
The carpenter judged the nurse because
206
The carpenter judged the left-handed nurse because
The midwife knew her towel because
The midwife knew her patterned towel because
The midwife knew her priest because
The midwife knew her bearded priest because
The midwife knew the towel because
The midwife knew the patterned towel because
The midwife knew the priest because
The midwife knew the bearded priest because
The queen needed her windbreaker because
The queen needed her imported windbreaker because
The queen needed her groundskeeper because
The queen needed her American groundskeeper because
The queen needed the windbreaker because
The queen needed the imported windbreaker because
The queen needed the groundskeeper because
The queen needed the American groundskeeper because
The cheerleader rescued her skateboard because
The cheerleader rescued her cracked skateboard because
The cheerleader rescued her postman because
The cheerleader rescued her blonde postman because
The cheerleader rescued the skateboard because
The cheerleader rescued the cracked skateboard because
207
The cheerleader rescued the postman because
The cheerleader rescued the blonde postman because
The seamstress analyzed her bracelet because
The seamstress analyzed her silver bracelet because
The seamstress analyzed her mechanic because
The seamstress analyzed her divorced mechanic because
The seamstress analyzed the bracelet because
The seamstress analyzed the silver bracelet because
The seamstress analyzed the mechanic because
The seamstress analyzed the divorced mechanic because
The beautician discovered her stockpot because
The beautician discovered her copper stockpot because
The beautician discovered her tattooist because
The beautician discovered her certified tattooist because
The beautician discovered the stockpot because
The beautician discovered the copper stockpot because
The beautician discovered the tattooist because
The beautician discovered the certified tattooist because
The saleswoman recorded her saxophone because
The saleswoman recorded her refurbished saxophone because
The saleswoman recorded her stockbroker because
The saleswoman recorded her fraudulent stockbroker because
The saleswoman recorded the saxophone because
208
The saleswoman recorded the refurbished saxophone because
The saleswoman recorded the stockbroker because
The saleswoman recorded the fraudulent stockbroker because
The princess investigated her dresser because
The princess investigated her vintage dresser because
The princess investigated her bodyguard because
The princess investigated her blue-eyed bodyguard because
The princess investigated the dresser because
The princess investigated the vintage dresser because
The princess investigated the bodyguard because
The princess investigated the blue-eyed bodyguard because

209
APPENDIX C
Target Stimuli for Experiment 3

The artist drew the hairdresser with the permanent marker because she was fandorous
The artist drew the screwdriver with the permanent marker because she was fandorous
The artist drew her hairdresser with the permanent marker because she was fandorous
The artist drew her screwdriver with the permanent marker because she was fandorous
The ballerina discussed the congresswoman throughout the seafood dinner because she was
brillish
The ballerina discussed the wineglass throughout the seafood dinner because she was brillish
The ballerina discussed her congresswoman throughout the seafood dinner because she was
brillish
The ballerina discussed her wineglass throughout the seafood dinner because she was brillish
The singer needed the florist before the concert rehearsal because she was thibbly
The singer needed the latte before the concert rehearsal because she was thibbly
The singer needed her florist before the concert rehearsal because she was thibbly
The singer needed her latte before the concert rehearsal because she was thibbly
The backpacker rescued the professor from the frozen river because she was sterchiful
The backpacker rescued the football from the frozen river because she was sterchiful
The backpacker rescued her professor from the frozen river because she was sterchiful
The backpacker rescued her football from the frozen river because she was sterchiful
The runner praised the nurse during the podcast interview because she was pambly
The runner praised the cake during the podcast interview because she was pambly
210
The runner praised her nurse during the podcast interview because she was pambly
The runner praised her cake during the podcast interview because she was pambly
The photographer evaluated the housekeeper before the birthday party because she was ribulous
The photographer evaluated the microwave before the birthday party because she was ribulous
The photographer evaluated her housekeeper before the birthday party because she was ribulous
The photographer evaluated her microwave before the birthday party because she was ribulous
The bride reconsidered the therapist over the holiday break because she was bompiful
The bride reconsidered the refrigerator over the holiday break because she was bompiful
The bride reconsidered her therapist over the holiday break because she was bompiful
The bride reconsidered her refrigerator over the holiday break because she was bompiful
The princess saw the attorney through the glass door because she was crompety
The princess saw the television through the glass door because she was crompety
The princess saw her attorney through the glass door because she was crompety
The princess saw her television through the glass door because she was crompety
The violinist found the secretary on the park bench because she was guptuous
The violinist found the newspaper on the park bench because she was guptuous
The violinist found her secretary on the park bench because she was guptuous
The violinist found her newspaper on the park bench because she was guptuous
The nun blessed the accountant after the community barbecue because she was lupative
The nun blessed the toothbrush after the community barbecue because she was lupative
The nun blessed her accountant after the community barbecue because she was lupative
The nun blessed her toothbrush after the community barbecue because she was lupative
The boxer discovered the tattooist on the social media app because she was randacious
211
The boxer discovered the stockpot on the social media app because she was randacious
The boxer discovered her tattooist on the social media app because she was randacious
The boxer discovered her stockpot on the social media app because she was randacious
The chef touched the psychic with the greasy spatula because she was ronkulous
The chef touched the helmet with the greasy spatula because she was ronkulous
The chef touched her psychic with the greasy spatula because she was ronkulous
The chef touched her helmet with the greasy spatula because she was ronkulous
The businessman observed the dentist through the open window because he was tadgey
The businessman observed the medal through the open window because he was tadgey
The businessman observed his dentist through the open window because he was tadgey
The businessman observed his medal through the open window because he was tadgey
The firefighter saved the butler from the collapsed building because he was thruppish
The firefighter saved the jewel from the collapsed building because he was thruppish
The firefighter saved his butler from the collapsed building because he was thruppish
The firefighter saved his jewel from the collapsed building because he was thruppish
The pilot investigated the mechanic after the overnight burglary because he was mindible
The pilot investigated the fireplace after the overnight burglary because he was mindible
The pilot investigated his mechanic after the overnight burglary because he was mindible
The pilot investigated his fireplace after the overnight burglary because he was mindible
The wizard forgot the postman during the tropical vacation because he was quimpiful
The wizard forgot the raincoat during the tropical vacation because he was quimpiful
The wizard forgot his postman during the tropical vacation because he was quimpiful
The wizard forgot his raincoat during the tropical vacation because he was quimpiful
212
The scientist tested the bodyguard for the deadly virus because he was shoobly
The scientist tested the handkerchief for the deadly virus because he was shoobly
The scientist tested his bodyguard for the deadly virus because he was shoobly
The scientist tested his handkerchief for the deadly virus because he was shoobly
The athlete corrected the teacher regarding the inaccurate time because he was saporous
The athlete corrected the computer regarding the inaccurate time because he was saporous
The athlete corrected his teacher regarding the inaccurate time because he was saporous
The athlete corrected his computer regarding the inaccurate time because he was saporous
The wrestler threw the architect onto of the hardwood floor because he was whemulous
The wrestler threw the microphone onto of the hardwood floor because he was whemulous
The wrestler threw his architect onto of the hardwood floor because he was whemulous
The wrestler threw his microphone onto of the hardwood floor because he was whemulous
The soldier heard the plumber throughout the autumn evening because he was nimpacious
The soldier heard the dryer throughout the autumn evening because he was nimpacious
The soldier heard his plumber throughout the autumn evening because he was nimpacious
The soldier heard his dryer throughout the autumn evening because he was nimpacious
The guitarist missed the landscaper during the nationwide tour because he was merpative
The guitarist missed the coffeemaker during the nationwide tour because he was merpative
The guitarist missed his landscaper during the nationwide tour because he was merpative
The guitarist missed his coffeemaker during the nationwide tour because he was merpative
The janitor recorded the barber on the security camera because he was chulpish
The janitor recorded the mailbox on the security camera because he was chulpish
The janitor recorded his barber on the security camera because he was chulpish
213
The janitor recorded his mailbox on the security camera because he was chulpish
The carpenter reviewed the surgeon in the anonymous survey because he was dempety
The carpenter reviewed the steak in the anonymous survey because he was dempety
The carpenter reviewed his surgeon in the anonymous survey because he was dempety
The carpenter reviewed his steak in the anonymous survey because he was dempety
The farmer fetched the priest before the winter storm because he was cranchible
The farmer fetched the towel before the winter storm because he was cranchible
The farmer fetched his priest before the winter storm because he was cranchible
The farmer fetched his towel before the winter storm because he was cranchible

214
APPENDIX D
Target Stimuli for Experiments 4 and 5a–d
7

William dreezed his nose, and Mark did, too.
Mark dreezed his own nose.
Mark dreezed William's nose.
William dreezed his father, and Mark did, too.
Mark dreezed his own father.
Mark dreezed William's father.
William dreezed his bicycle, and Mark did, too.
Mark dreezed his own bicycle.
Mark dreezed William's bicycle.
William dreezed his boss, and Mark did, too.
Mark dreezed his own boss.
Mark dreezed William's boss.
Laura shroped her eyes, and Kathleen did, too.
Kathleen shroped her own eyes.
Kathleen shroped Laura's eyes.
Laura shroped her cousin, and Kathleen did, too.
Kathleen shroped her own cousin.
Kathleen shroped Laura's cousin.

7
Items listed in this appendix are shown as they appeared in Experiment 4. For Experiments 5a–d, either the
antecedent-clause subject or elided-clause subject was replaced with every man or every woman, as described in
Section 5.5.2. The interpretation options were likewise modified with corresponding substitutions.
215
Laura shroped her computer, and Kathleen did, too.
Kathleen shroped her own computer.
Kathleen shroped Laura's computer.
Laura shroped her classmate, and Kathleen did, too.
Kathleen shroped her own classmate.
Kathleen shroped Laura's classmate.
James swudged his foot, and Robert did, too.
Robert swudged his own foot.
Robert swudged James's foot.
James swudged his uncle, and Robert did, too.
Robert swudged his own uncle.
Robert swudged James's uncle.
James swudged his shoes, and Robert did, too.
Robert swudged his own shoes.
Robert swudged James's shoes.
James swudged his acquaintance, and Robert did, too.
Robert swudged his own acquaintance.
Robert swudged James's acquaintance.
Sharon pleached her ears, and Rebecca did, too.
Rebecca pleached her own ears.
Rebecca pleached Sharon's ears.
Sharon pleached her brother, and Rebecca did, too.
Rebecca pleached her own brother.
216
Rebecca pleached Sharon's brother.
Sharon pleached her car, and Rebecca did, too.
Rebecca pleached her own car.
Rebecca pleached Sharon's car.
Sharon pleached her opponent, and Rebecca did, too.
Rebecca pleached her own opponent.
Rebecca pleached Sharon's opponent.
Anthony sibbed his fingers, and Charles did, too.
Charles sibbed his own fingers.
Charles sibbed Anthony's fingers.
Anthony sibbed his sister, and Charles did, too.
Charles sibbed his own sister.
Charles sibbed Anthony's sister.
Anthony sibbed his sandwich, and Charles did, too.
Charles sibbed his own sandwich.
Charles sibbed Anthony's sandwich.
Anthony sibbed his coworker, and Charles did, too.
Charles sibbed his own coworker.
Charles sibbed Anthony's coworker.
Melissa gweeshed her leg, and Stephanie did, too.
Stephanie gweeshed her own leg.
Stephanie gweeshed Melissa's leg.
Melissa gweeshed her nephew, and Stephanie did, too.
217
Stephanie gweeshed her own nephew.
Stephanie gweeshed Melissa's nephew.
Melissa gweeshed her garden, and Stephanie did, too.
Stephanie gweeshed her own garden.
Stephanie gweeshed Melissa's garden.
Melissa gweeshed her supervisor, and Stephanie did, too.
Stephanie gweeshed her own supervisor.
Stephanie gweeshed Melissa's supervisor.
Daniel dobbed his feelings, and Fred did, too.
Fred dobbed his own feelings.
Fred dobbed Daniel's feelings.
Daniel dobbed his grandfather, and Fred did, too.
Fred dobbed his own grandfather.
Fred dobbed Daniel's grandfather.
Daniel dobbed his book, and Fred did, too.
Fred dobbed his own book.
Fred dobbed Daniel's book.
Daniel dobbed his enemy, and Fred did, too.
Fred dobbed his own enemy.
Fred dobbed Daniel's enemy.
Helen chabbed her hometown, and Amanda did, too.
Amanda chabbed her own hometown.
Amanda chabbed Helen's hometown.
218
Helen chabbed her son, and Amanda did, too.
Amanda chabbed her own son.
Amanda chabbed Helen's son.
Helen chabbed her jacket, and Amanda did, too.
Amanda chabbed her own jacket.
Amanda chabbed Helen's jacket.
Helen chabbed her colleague, and Amanda did, too.
Amanda chabbed her own colleague.
Amanda chabbed Helen's colleague.
Edward risped his haircut, and Brian did, too.
Brian risped his own haircut.
Brian risped Edward's haircut.
Edward risped his mother, and Brian did, too.
Brian risped his own mother.
Brian risped Edward's mother.
Edward risped his skateboard, and Brian did, too.
Brian risped his own skateboard.
Brian risped Edward's skateboard.
Edward risped his teammate, and Brian did, too.
Brian risped his own teammate.
Brian risped Edward's teammate.
Emily tiched her feelings, and Jennifer did, too.
Jennifer tiched her own feelings.
219
Jennifer tiched Emily's feelings.
Emily tiched her aunt, and Jennifer did, too.
Jennifer tiched her own aunt.
Jennifer tiched Emily's aunt.
Emily tiched her newspaper, and Jennifer did, too.
Jennifer tiched her own newspaper.
Jennifer tiched Emily's newspaper.
Emily tiched her friend, and Jennifer did, too.
Jennifer tiched her own friend.
Jennifer tiched Emily's friend.
Kevin churbited his reputation, and Timothy did, too.
Timothy churbited his own reputation.
Timothy churbited Kevin's reputation.
Kevin churbited his daughter, and Timothy did, too.
Timothy churbited his own daughter.
Timothy churbited Kevin's daughter.
Kevin churbited his backpack, and Timothy did, too.
Timothy churbited his own backpack.
Timothy churbited Kevin's backpack.
Kevin churbited his partner, and Timothy did, too.
Timothy churbited his own partner.
Timothy churbited Kevin's partner.
Sandra pranned her job, and Margaret did, too.
220
Margaret pranned her own job.
Margaret pranned Sandra's job.
Sandra pranned her grandmother, and Margaret did, too.
Margaret pranned her own grandmother.
Margaret pranned Sandra's grandmother.
Sandra pranned her guitar, and Margaret did, too.
Margaret pranned her own guitar.
Margaret pranned Sandra's guitar.
Sandra pranned her roommate, and Margaret did, too.
Margaret pranned her own roommate.
Margaret pranned Sandra's roommate.
Jason jolfed his nose, and Ronald did, too.
Ronald jolfed his own nose.
Ronald jolfed Jason's nose.
Jason jolfed his father, and Ronald did, too.
Ronald jolfed his own father.
Ronald jolfed Jason's father.
Jason jolfed his bicycle, and Ronald did, too.
Ronald jolfed his own bicycle.
Ronald jolfed Jason's bicycle.
Jason jolfed his boss, and Ronald did, too.
Ronald jolfed his own boss.
Ronald jolfed Jason's boss.
221
Carol shupped her eyes, and Jessica did, too.
Jessica shupped her own eyes.
Jessica shupped Carol's eyes.
Carol shupped her cousin, and Jessica did, too.
Jessica shupped her own cousin.
Jessica shupped Carol's cousin.
Carol shupped her computer, and Jessica did, too.
Jessica shupped her own computer.
Jessica shupped Carol's computer.
Carol shupped her classmate, and Jessica did, too.
Jessica shupped her own classmate.
Jessica shupped Carol's classmate.
Jonathan zatted his foot, and Larry did, too.
Larry zatted his own foot.
Larry zatted Jonathan's foot.
Jonathan zatted his uncle, and Larry did, too.
Larry zatted his own uncle.
Larry zatted Jonathan's uncle.
Jonathan zatted his shoes, and Larry did, too.
Larry zatted his own shoes.
Larry zatted Jonathan's shoes.
Jonathan zatted his acquaintance, and Larry did, too.
Larry zatted his own acquaintance.
222
Larry zatted Jonathan's acquaintance.
Kimberly blorned her ears, and Sarah did, too.
Sarah blorned her own ears.
Sarah blorned Kimberly's ears.
Kimberly blorned her brother, and Sarah did, too.
Sarah blorned her own brother.
Sarah blorned Kimberly's brother.
Kimberly blorned her car, and Sarah did, too.
Sarah blorned her own car.
Sarah blorned Kimberly's car.
Kimberly blorned her opponent, and Sarah did, too.
Sarah blorned her own opponent.
Sarah blorned Kimberly's opponent.
Howard dacked his fingers, and Phillip did, too.
Phillip dacked his own fingers.
Phillip dacked Howard's fingers.
Howard dacked his sister, and Phillip did, too.
Phillip dacked his own sister.
Phillip dacked Howard's sister.
Howard dacked his sandwich, and Phillip did, too.
Phillip dacked his own sandwich.
Phillip dacked Howard's sandwich.
Howard dacked his coworker, and Phillip did, too.
223
Phillip dacked his own coworker.
Phillip dacked Howard's coworker.
Veronica fammed her leg, and Shelby did, too.
Shelby fammed her own leg.
Shelby fammed Veronica's leg.
Veronica fammed her nephew, and Shelby did, too.
Shelby fammed her own nephew.
Shelby fammed Veronica's nephew.
Veronica fammed her garden, and Shelby did, too.
Shelby fammed her own garden.
Shelby fammed Veronica's garden.
Veronica fammed her supervisor, and Shelby did, too.
Shelby fammed her own supervisor.
Shelby fammed Veronica's supervisor.
Joseph gorped his feelings, and Randall did, too.
Randall gorped his own feelings.
Randall gorped Joseph's feelings.
Joseph gorped his grandfather, and Randall did, too.
Randall gorped his own grandfather.
Randall gorped Joseph's grandfather.
Joseph gorped his book, and Randall did, too.
Randall gorped his own book.
Randall gorped Joseph's book.
224
Joseph gorped his enemy, and Randall did, too.
Randall gorped his own enemy.
Randall gorped Joseph's enemy.
Victoria linded her hometown, and Sophia did, too.
Sophia linded her own hometown.
Sophia linded Victoria's hometown.
Victoria linded her son, and Sophia did, too.
Sophia linded her own son.
Sophia linded Victoria's son.
Victoria linded her jacket, and Sophia did, too.
Sophia linded her own jacket.
Sophia linded Victoria's jacket.
Victoria linded her colleague, and Sophia did, too.
Sophia linded her own colleague.
Sophia linded Victoria's colleague.
Noah chooped his haircut, and Brad did, too.
Brad chooped his own haircut.
Brad chooped Noah's haircut.
Noah chooped his mother, and Brad did, too.
Brad chooped his own mother.
Brad chooped Noah's mother.
Noah chooped his skateboard, and Brad did, too.
Brad chooped his own skateboard.
225
Brad chooped Noah's skateboard.
Noah chooped his teammate, and Brad did, too.
Brad chooped his own teammate.
Brad chooped Noah's teammate.
Zoe meared her feelings, and Megan did, too.
Megan meared her own feelings.
Megan meared Zoe's feelings.
Zoe meared her aunt, and Megan did, too.
Megan meared her own aunt.
Megan meared Zoe's aunt.
Zoe meared her newspaper, and Megan did, too.
Megan meared her own newspaper.
Megan meared Zoe's newspaper.
Zoe meared her friend, and Megan did, too.
Megan meared her own friend.
Megan meared Zoe's friend.
Russell jepped his reputation, and Seth did, too.
Seth jepped his own reputation.
Seth jepped Russell's reputation.
Russell jepped his daughter, and Seth did, too.
Seth jepped his own daughter.
Seth jepped Russell's daughter.
Russell jepped his backpack, and Seth did, too.
226
Seth jepped his own backpack.
Seth jepped Russell's backpack.
Russell jepped his partner, and Seth did, too.
Seth jepped his own partner.
Seth jepped Russell's partner.
Savannah tammed her job, and Olivia did, too.
Olivia tammed her own job.
Olivia tammed Savannah's job.
Savannah tammed her grandmother, and Olivia did, too.
Olivia tammed her own grandmother.
Olivia tammed Savannah's grandmother.
Savannah tammed her guitar, and Olivia did, too.
Olivia tammed her own guitar.
Olivia tammed Savannah's guitar.
Savannah tammed her roommate, and Olivia did, too.
Olivia tammed her own roommate.
Olivia tammed Savannah's roommate.

227
APPENDIX E
Model Formulae for Experiment 1

glmer formula for the model predicting mentions of the preceding subject in subject
position of the continuation:
subject mention in subject position ~ animacy * possession
+ (1 + animacy | subject)
+ (1 | item)

glmer formula for the model predicting whether continuations in the animate conditions
would refer to either the preceding subject or object (ignoring the possibility of a third-
party subject):
subject versus object mention in subject position ~ possession
+ (1 | subject)
+ (1 | item)

glmer formula for the model predicting mentions of the preceding object in subject position
of the continuation:
object mention in subject position ~ animacy * possession
+ (1 + animacy + possession | subject)
+ (1 | item)

228
glmer formula for the model testing the simple effect of possession within the animate
conditions:
object mention in subject position ~ possession
+ (1 + possession | subject)
+ (1 | item)

glmer formula for the model testing the simple effect of possession within the inanimate
conditions:
object mention in subject position ~ possession
+ (1 + possession | subject)
+ (1 | item)

glmer formula for the model predicting mentions of the preceding subject anywhere in the
continuation:
subject mention ~ animacy * possession
+ (1 + animacy | subject)
+ (1 | item)

glmer formula for the model predicting mentions of the preceding object anywhere in the
continuation:
object mention ~ animacy * possession
+ (1 + possession | subject)
+ (1 | item)
229

glmer formula for the model testing the simple effect of possession (in all positions of the
continuation) within the animate conditions:
object mention ~ possession
+ (1 + possession | subject)
+ (1 | item)

glmer formula for the model testing the simple effect of possession (in all positions of the
continuation) within the inanimate conditions:
object mention ~ possession
+ (1 | subject)
+ (1 | item)

glmer formula for the model predicting the conditional probability of using a pronoun
given reference to the prompt subject in subject position:
subject pronominalization in subject position ~ animacy * possession
+ (1 + animacy | subject)
+ (1 | item)

glmer formula for the model predicting the conditional probability of using a pronoun
given reference to the prompt object in subject position:
object pronominalization in subject position ~ animacy * possession
+ (1 | subject)
230
+ (1 + animacy | item)
glmer formula for the model predicting the conditional probability of using a pronoun
given reference to the prompt subject anywhere in the continuation:
subject pronominalization ~ animacy * possession
+ (1 | subject)
+ (1 | item)

glmer formula for the model predicting the conditional probability of using a pronoun
given reference to the prompt object anywhere in the continuation:
object pronominalization ~ animacy * possession
+ (1 + animacy | subject)

231
APPENDIX F
Model Formulae for Experiment 2

glmer formula for the model predicting mentions of the preceding subject in subject
position of the continuation:
subject mention in subject position ~ animacy * possession * adjective modification
+ (1 + animacy + possession + adjective modification | subject)
+ (1 + animacy + possession + adjective modification | item)

glmer formula for the model predicting whether continuations in the animate conditions
would refer to either the preceding subject or object (ignoring the possibility of a third-
party subject):
subject versus object mention in subject position ~ possession * adjective modification
+ (1 possession + adjective modification | subject)
+ (1 possession + adjective modification | item)

glmer formula for the model predicting mentions of the preceding object in subject position
of the continuation:
object mention in subject position ~ animacy * possession * adjective modification
+ (1 + animacy + possession | subject)
+ (1 + animacy + possession | item)

232
glmer formula for the model predicting mentions of the preceding subject anywhere in the
continuation:
subject mention ~ animacy * possession * adjective modification
+ (1 + animacy + possession + adjective modification | subject)
+ (1 + adjective modification | item)

glmer formula for the model predicting mentions of the preceding object anywhere in the
continuation:
object mention ~ animacy * possession * adjective modification
+ (1 + animacy | subject)
+ (1 + animacy + adjective modification | item)

glmer formula for the model predicting the conditional probability of using a pronoun
given reference to the prompt subject in subject position:
subject pronominalization in subject position ~ animacy * possession * adjective modification
+ (1 | subject)

glmer formula for the model predicting the conditional probability of using a pronoun
given reference to the prompt object in subject position:
object pronominalization in subject position ~ animacy * possession * adjective modification
+ (1 + animacy | subject)
+ (1 + animacy | item)

233
glmer formula for the model predicting the conditional probability of using a pronoun
given reference to the prompt subject anywhere in the continuation:
subject pronominalization ~ animacy * possession * adjective modification
+ (1 | subject)

glmer formula for the model predicting the conditional probability of using a pronoun
given reference to the prompt object anywhere in the continuation:
object pronominalization ~ animacy * possession * adjective modification
+ (1 + animacy | subject)
+ (1 + animacy | item)

234
APPENDIX G
Model Formulae for Experiment 3

glmer formula for the model predicting selections of the last-mentioned referent:
last-mentioned selection ~ animacy * possession
+ (1 | subject)
+ (1 + animacy | item)

glmer formula for the model predicting selections of the last-mentioned referent in the
animate conditions:
last-mentioned selection ~ 1
+ (1 | subject)
+ (1 | item)

lmer formula for the model predicting selection latencies relative to the onset of the
pronoun:
last-mentioned selection ~ animacy * possession
+ (1 + animacy + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the onset of the
matrix-clause object (-1000 ms to 0 ms):
proportion of subject looks ~ animacy * possession
235
+ (1 + possession | subject)
+ (1 + animacy | item)

lmer formula for the model predicting looks to the subject aligned to the onset of the
matrix-clause object (0 ms to 1000 ms):
proportion of subject looks ~ animacy * possession
+ (1 + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the onset of the
matrix-clause object (1000 ms to 2000 ms):
proportion of subject looks ~ animacy * possession
+ (1 + animacy + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the onset of the
matrix-clause object (1000 ms to 2000 ms; interaction follow-up, animate conditions):
proportion of subject looks ~ possession
+ (1 + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the onset of the
matrix-clause object (1000 ms to 2000 ms; interaction follow-up, inanimate conditions):
236
proportion of subject looks ~ possession
+ (1 + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the object aligned to the onset of the matrix-
clause object (-1000 ms to 0 ms):
proportion of object looks ~ animacy * possession
+ (1 + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the object aligned to the onset of the matrix-
clause object (0 ms to 1000 ms):
proportion of object looks ~ animacy * possession
+ (1 | subject)
+ (1 + animacy | item)

lmer formula for the model predicting looks to the object aligned to the onset of the matrix-
clause object (1000 ms to 2000 ms):
proportion of object looks ~ animacy * possession
+ (1 + possession | subject)
+ (1 + animacy | item)

237
lmer formula for the model predicting looks to the subject aligned to the onset of the
pronoun (-1000 ms to 0 ms):
proportion of subject looks ~ animacy * possession
+ (1 + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the onset of the
pronoun (0 ms to 1000 ms):
proportion of subject looks ~ animacy * possession
+ (1 + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the onset of the
pronoun (1000 ms to 2000 ms):
proportion of subject looks ~ animacy * possession
+ (1 + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the onset of the
pronoun (2000 ms to 3000 ms):
proportion of subject looks ~ animacy * possession
+ (1 + animacy | subject)
+ (1 | item)
238

lmer formula for the model predicting looks to the subject aligned to the onset of the
pronoun (3000 ms to 4000 ms):
proportion of subject looks ~ animacy * possession
+ (1 + animacy | subject)
+ (1 + possession | item)

lmer formula for the model predicting looks to the object aligned to the onset of the
pronoun (-1000 ms to 0 ms):
proportion of object looks ~ animacy * possession
+ (1 + possession | subject)
+ (1 + animacy | item)

lmer formula for the model predicting looks to the object aligned to the onset of the
pronoun (0 ms to 1000 ms):
proportion of object looks ~ animacy * possession
+ (1 | subject)
+ (1 + animacy | item)

lmer formula for the model predicting looks to the object aligned to the onset of the
pronoun (1000 ms to 2000 ms):
proportion of object looks ~ animacy * possession
+ (1 | subject)
239

lmer formula for the model predicting looks to the object aligned to the onset of the
pronoun (2000 ms to 3000 ms):
proportion of object looks ~ animacy * possession
+ (1 | subject)
+ (1 | item)

lmer formula for the model predicting looks to the object aligned to the onset of the
pronoun (3000 ms to 4000 ms):
proportion of object looks ~ animacy * possession
+ (1 | subject)
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the onset of the
pronoun (-1000 ms to 0 ms; planned comparison of animate conditions):
proportion of subject looks ~ possession
+ (1 + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the onset of the
pronoun (0 ms to 1000 ms; planned comparison of animate conditions):
proportion of subject looks ~ possession
+ (1 | subject)
240
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the onset of the
pronoun (1000 ms to 2000 ms; planned comparison of animate conditions):
proportion of subject looks ~ possession
+ (1 + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the onset of the
pronoun (2000 ms to 3000 ms; planned comparison of animate conditions):
proportion of subject looks ~ possession
+ (1 + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the onset of the
pronoun (3000 ms to 4000 ms; planned comparison of animate conditions):
proportion of subject looks ~ possession
+ (1 + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the object aligned to the onset of the
pronoun (-1000 ms to 0 ms; planned comparison of animate conditions):
proportion of object looks ~ possession
241
+ (1 + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the object aligned to the onset of the
pronoun (0 ms to 1000 ms; planned comparison of animate conditions):
proportion of object looks ~ possession
+ (1 | subject)
+ (1 | item)

lmer formula for the model predicting looks to the object aligned to the onset of the
pronoun (1000 ms to 2000 ms; planned comparison of animate conditions):
proportion of object looks ~ possession
+ (1 + possession | subject)
+ (1 + possession | item)

lmer formula for the model predicting looks to the object aligned to the onset of the
pronoun (2000 ms to 3000 ms; planned comparison of animate conditions):
proportion of object looks ~ possession
+ (1 + possession | subject)
+ (1 + possession | item)

lmer formula for the model predicting looks to the object aligned to the onset of the
pronoun (3000 ms to 4000 ms; planned comparison of animate conditions):
242
proportion of object looks ~ possession
+ (1 | subject)
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the offset of the
pronoun (-1000 ms to 0 ms):
proportion of subject looks ~ animacy * possession
+ (1 + animacy | subject)
+ (1 + animacy | item)

lmer formula for the model predicting looks to the subject aligned to the offset of the
pronoun (0 ms to 1000 ms):
proportion of subject looks ~ animacy * possession
+ (1 + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the offset of the
pronoun (1000 ms to 2000 ms):
proportion of subject looks ~ animacy * possession
+ (1 + possession | subject)
+ (1 | item)

243
lmer formula for the model predicting looks to the subject aligned to the offset of the
pronoun (2000 ms to 3000 ms):
proportion of subject looks ~ animacy * possession
+ (1 | subject)
+ (1 + possession | item)

lmer formula for the model predicting looks to the subject aligned to the offset of the
pronoun (3000 ms to 4000 ms):
proportion of subject looks ~ animacy * possession
+ (1 + animacy | subject)
+ (1 | item)

lmer formula for the model predicting looks to the object aligned to the offset of the
pronoun (-1000 ms to 0 ms):
proportion of object looks ~ animacy * possession
+ (1 + possession | subject)
+ (1 + possession | item)

lmer formula for the model predicting looks to the object aligned to the offset of the
pronoun (0 ms to 1000 ms):
proportion of object looks ~ animacy * possession
+ (1 | subject)
+ (1 | item)
244

lmer formula for the model predicting looks to the object aligned to the offset of the
pronoun (1000 ms to 2000 ms):
proportion of object looks ~ animacy * possession
+ (1 | subject)

lmer formula for the model predicting looks to the object aligned to the offset of the
pronoun (2000 ms to 3000 ms):
proportion of object looks ~ animacy * possession
+ (1 + animacy | subject)

lmer formula for the model predicting looks to the object aligned to the offset of the
pronoun (3000 ms to 4000 ms):
proportion of object looks ~ animacy * possession
+ (1 | subject)
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the offset of the
pronoun (-1000 ms to 0 ms; planned comparison of animate conditions):
proportion of subject looks ~ possession
+ (1 + possession | subject)
+ (1 | item)

245
lmer formula for the model predicting looks to the subject aligned to the offset of the
pronoun (0 ms to 1000 ms; planned comparison of animate conditions):
proportion of subject looks ~ possession
+ (1 + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the offset of the
pronoun (1000 ms to 2000 ms; planned comparison of animate conditions):
proportion of subject looks ~ possession
+ (1 + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the offset of the
pronoun (2000 ms to 3000 ms; planned comparison of animate conditions):
proportion of subject looks ~ possession
+ (1 + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the subject aligned to the offset of the
pronoun (3000 ms to 4000 ms; planned comparison of animate conditions):
proportion of subject looks ~ possession
+ (1 + possession | subject)
+ (1 | item)
246

lmer formula for the model predicting looks to the object aligned to the offset of the
pronoun (-1000 ms to 0 ms; planned comparison of animate conditions):
proportion of object looks ~ possession
+ (1 + possession | subject)
+ (1 + possession | item)

lmer formula for the model predicting looks to the object aligned to the offset of the
pronoun (0 ms to 1000 ms; planned comparison of animate conditions):
proportion of object looks ~ possession
+ (1 | subject)
+ (1 | item)

lmer formula for the model predicting looks to the object aligned to the offset of the
pronoun (1000 ms to 2000 ms; planned comparison of animate conditions):
proportion of object looks ~ possession
+ (1 + possession | subject)
+ (1 | item)

lmer formula for the model predicting looks to the object aligned to the offset of the
pronoun (2000 ms to 3000 ms; planned comparison of animate conditions):
proportion of object looks ~ possession
+ (1 | subject)
247
+ (1 | item)

lmer formula for the model predicting looks to the object aligned to the offset of the
pronoun (3000 ms to 4000 ms; planned comparison of animate conditions):
proportion of object looks ~ possession
+ (1 | subject)
+ (1 | item)

248
APPENDIX H
Model Formulae for Experiment 4

glmer formula for the model predicting the likelihood of a bound-variable interpretation
(reference level = part-whole):
ellipsis interpretation ~ possession type
+ (1 | subject)
+ (1 | item)

glmer formula for the model predicting the likelihood of a bound-variable interpretation
(reference level = ownership):
ellipsis interpretation ~ possession type
+ (1 | subject)
+ (1 | item)

glmer formula for the model predicting the likelihood of a bound-variable interpretation
(reference level = animate relational):
ellipsis interpretation ~ possession type
+ (1 | subject)
+ (1 | item)

glmer formula for the model predicting the likelihood of a bound-variable interpretation
(reference level = kinship):
249
ellipsis interpretation ~ possession type
+ (1 | subject)
+ (1 | item)

250
APPENDIX I
Model Formulae for Experiments 5a–d

glmer formula for the model predicting the likelihood of a bound-variable interpretation in
Experiments 5a–d:
ellipsis interpretation ~ quantified-NP position * animacy * gender mismatch * overt pronoun
gender
+ (1 + quantified NP position | subject)
+ (1 + animacy | item)

glmer formula for the model predicting the likelihood of a bound-variable interpretation in
Experiments 5a and 5c:
ellipsis interpretation ~ animacy * gender mismatch * overt pronoun gender
+ (1 + gender mismatch | subject)
+ (1 | item)

glmer formula for the model predicting the likelihood of a bound-variable interpretation in
Experiments 5b and 5d:
ellipsis interpretation ~ animacy * gender mismatch * overt pronoun gender
+ (1 + animacy + gender mismatch | subject)
+ (1 | item)

Abstract (if available)

Abstract This dissertation explores how the referents involved in nominal possessive constructions are represented at the discourse level, how possession affects discourse prominence, and how different semantic possession relations between the possessor and possessed referent influence discourse-level processing.

To address these research questions I formulate two sets of competing hypotheses about how possession affects the prominence of possessed referents. The first set of hypotheses relates to the overall effect of possession on prominence (abstracting away from variation across different types of possession relations): the Parasitic Possession Hypothesis claims that involvement in a possessive construction reduces the prominence of the possessed referent, while the Complexity Hypothesis claims the opposite—that possession increases the prominence of the possessed referent. The second set of hypotheses relates to whether different types of possession relations modulate the overall discourse-level effect of possession: the Interaction Hypothesis posits an exceptional boost in prominence for possessed animates, while the Animacy Hypothesis attributes any differences in prominence between possessed animates and possessed inanimates to the effect of animacy (rather than any influence due to the possession relation itself).

I test these hypotheses in five experiments. Experiments 1 and 2 (Chapters 2 and 3, respectively) use a sentence-continuation paradigm to compare the prominence of possessed referents to two types of simple nominals: indefinites (Experiment 1) and definites (Experiment 2). Both experiments primarily analyze the likelihood that participants re-mention referents from a prompt, but they produce somewhat different results. In Experiment 1, I do not find strong evidence for either the Parasitic Possession Hypothesis or the Complexity Hypothesis (which relate to the overall effect of possession on prominence); however, the results do suggest an extra boost in prominence for possessed animates, which lends support to the Interaction Hypothesis. Experiment 2, on the other hand, finds evidence that possession reduces the prominence of possessed referents overall and that the type of possession relation does not modulate this reduction in prominence. Thus Experiment 2 supports the Parasitic Possession Hypothesis and the Animacy Hypothesis.

Experiment 3 (Chapter 4) uses visual-world eye-tracking to examine gaze patterns during ambiguous pronoun resolution and compares how participants resolve a pronoun when one of its potential antecedents is possessed versus definite. The results show that possessed referents are less accessible as pronominal antecedents and therefore support the Parasitic Possession Hypothesis—that possession reduces a referent’s prominence.

Experiments 4 and 5 (Chapter 5) examine the resolution of ambiguous possessive pronouns in a type of verb-phrase ellipsis. Here I adopt a discourse-based approach to ellipsis interpretation which relies on the prominence of possessed referents. I find that the type of possession relation significantly affects ambiguity resolution in these elliptical constructions and relate the results back to the Interaction Hypothesis.

In sum, the results of this dissertation support the Parasitic Possession Hypothesis and the Interaction Hypothesis. I argue that possession generally decreases the prominence of possessed referents, although different types of possession relations may modulate this overall effect. In particular, possessed animates seem to be especially prominent in discourse.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Dynamics of multiple pronoun resolution

PDF

Discourse level processing and pronoun interpretation

PDF

Processing the dynamicity of events in language

PDF

Vietnamese pronouns in discourse

PDF

Syntactic and non-syntactic factors in reflexive pronoun resolution in Mandarin Chinese

PDF

Interpretation of pronominal forms across languages

PDF

When things are left unsaid: existential and anaphoric implicit objects in discourse

PDF

Exploring the effects of Korean subject marking and action verbs’ repetition frequency: how they influence the discourse and the memory representations of entities and events

PDF

Considering the effects of disfluent speech on children’s sentence processing capabilities and language development

PDF

The balance of scalar implicature

PDF

Prosody and informativity: a cross-linguistic investigation

PDF

A discourse analysis of teacher-student classroom interactions

PDF

Interaction between prosody and information structure: experimental evidence from Hindi and Bangla

PDF

Effects of language familiarity on talker discrimination from syllables

PDF

Towards a correlational law of language: three factors constraining judgement variation

PDF

What 'you' and 'I' can say about reference resolution and non-structural constraints

PDF

Schema architecture for language-vision interactions: a computational cognitive neuroscience model of language use

PDF

Comparative iIlusions at the syntax-semantics interface

PDF

The grammar of individuation, number and measurement

PDF

Asymetrical discourse in a computer-mediated environment

Asset Metadata

Creator Storbeck, Jesse (author)

Core Title Discourse-level processing of nominal possessive constructions

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Linguistics

Degree Conferral Date 2022-12

Publication Date 09/17/2022

Defense Date 08/15/2022

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag discourse,language processing,OAI-PMH Harvest,possessives,psycholinguistics

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Kaiser, Elsi (committee chair), Mintz, Toben (committee member), Wellwood, Alexis (committee member)

Creator Email jesse.storbeck@gmail.com,jstorbec@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC111996064

Unique identifier UC111996064

Legacy Identifier etd-StorbeckJe-11209

Document Type Dissertation

Format application/pdf (imt)

Rights Storbeck, Jesse

Type texts

Source 20220917-usctheses-batch-981 (batch), University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright. The original signature page accompanying the original submission of the work to the USC Libraries is retained by the USC Libraries and a copy of it may be obtained by authorized requesters contacting the repository e-mail address given.

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Repository Email cisadmin@lib.usc.edu