University of Southern California Dissertations and Theses

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

(USC Thesis Other)

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

PDF

Download

Open document

Flip pages

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content SYNTACTIC AND NON-SYNTACTIC FACTORS IN REFLEXIVE PRONOUN RESOLUTION IN
MANDARIN CHINESE
By
Jun Lyu

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

May 2023

Copyright 2023 Jun Lyu
ii
Acknowledgments

Completing this PhD dissertation has been a long and challenging journey. I cannot do so without the
support, encouragement, and guidance of many wonderful people.
First and foremost, I would like to express my heartfelt gratitude to my advisor, Elsi Kaiser, whose
mentorship and guidance have been invaluable throughout my research. Her knowledge and expertise
have helped me to navigate the challenges of my academic journey with clarity and purpose.
I would also like to thank my committee members, Audrey Li, Andrew Simpson, and Toby Mintz,
for their insightful comments and suggestions that have helped me improve my dissertation. It’s been
such an honor to have so many accomplished linguists and psycholinguists on my committee!
To the USC faculty and my screening paper committee members (Andrew Simpson, Audrey Li,
Deniz Rudin, Elsi Kaiser, Zuzanna Fuchs), I thank them for their contributions to my intellectual growth
and academic development.
This journey would have been less enriching without the psycholinguistics community at USC. I am
grateful for all the comments and suggestions on my various studies provided by Jina Song, Sarah Lee,
Ian Rigby, Linh Pham, Metehan Oguz, and many others.
I must acknowledge my parents, whose unconditional love and support have been the driving force
behind my academic journey. I dedicate this dissertation to the memory of my dad, Lyu Linhai, and my
grandpa, Lyu Chao, who had great hopes for me and unfortunately could not see me on my graduation
day.
I could not have finished this PhD at USC within 4 years without the encouragement of Yan Yaning,
who has been my companion and source of inspiration throughout my PhD journey.
iii
Let me not forget to mention my dear friend Ellen Gillen-Higgins and her family for their warm
hospitality and kindness during my stay at Stony Brook. I would also like to thank John Braile and Bob
Higgins for their companionship. I treasure our memories of jogging at the Sunken Meadow State Park.
Lastly, I would like to acknowledge the challenges I have faced during my academic journey. It was
indeed a trying but fruitful journey. These challenges have made me stronger and more resilient.

iv
Table of Contents

Acknowledgments ………………………………………………………………………………………………………………… ii
List of Tables …………………………………………………………………………………………………………………………vi
List of Figures …………………………………………………………………………………………………………………….. viii
Abstract .………………………………………………………………………………………………………………………………xi
Chapter 1: Introduction ……………………………………………………………………………………………………………1
1.1 Overview and objectives ………………………………………………………………………………………….. 1
1.2 Research background ………………………………………………………………………………………………3
1.3 Antecedent retrieval and sentence processing …………………………………………………………...33
1.4 Outline of the dissertation ………………………………………………………………………………………52
Chapter 2: Locality and recency in reflexive resolution ……………………………………………………………….54
2.1 Introduction …………………………………………………………………………………………………………54
2.2 Antecedent retrieval models .………………………………………………………………………………….57
2.3 Overview of the aims ……………………………………………………………………………………………..60
2.4 Experiment 1 ………………………………………………………………………………………………………..61
2.5 Experiment 2 ………………………………………………………………………………………………………..71
2.6 General discussion .……………………………………………………………………………………………….83
Chapter 3: Discourse prominence and reflexive resolution …………………………………………………………87
3.1 Introduction …………………………………………………………………………………………………………87
3.2 Discourse prominence and Mandarin reflexives ………………………………………………………..88
3.3 Verb semantics and Mandarin reflexives …………………………………………………………………..92
3.4 Antecedent retrieval models …………………………………………………………………………………..94
3.5 Overview of the aims ……………………………………………………………………………………………..98
3.6 Experiment 3 ………………………………………………………………………………………………………..99
3.7 Experiment 4 ……………………………………………………………………………………………………….113
3.8 General discussion ………………………………………………………………………………………………124
Chapter 4: Logophoric prominence and reflexive resolution .……………………………………………………. 133
4.1 Introduction .………………………………………………………………………………………………………. 133
4.2 Logophoricity and Mandarin reflexives .…………………………………………………………………. 136
4.3 Logophoricity and antecedent retrieval .………………………………………………………………….142
4.4 Overview of the aims ..…………………………………………………………………………………………..144
4.5 Experiment 5 .………………………………………………………………………………………………………145
4.6 Experiment 6 .……………………………………………………………………………………………………...149
v
4.7 Experiment 7 .…………………………………………………………………………………………………….. 157
4.8 General discussion .………………………………………………………………………………………………173
Chapter 5: The blocking effect and reflexive resolution .…………………………………………………………….179
5.1 Introduction .……………………………………………………………………………………………………….179
5.2 Overview of the linguistic accounts of the blocking effect .………………………………………….181
5.3 Standard and structure-mediated cue-based retrieval models .……………………………………184
5.4 Overview of the aims .……………………………………………………………………………………………186
5.5 Experiment 8 .………………………………………………………………………………………………………187
5.6 Experiment 9 .…………………………………………………………………………………………………..... 200
5.7 General discussion .………………………………………………………………………………………………216
Chapter 6: General discussion and conclusions .……………………………………………………………………….229
6.1 Summary of findings .……………………………………………………………………………………………229
6.2 Ziji and ta-ziji as exempt anaphors and their differences .…………………………………………..236
6.3 Empirical coverages of different variants of memory-based retrieval models ..………………238
6.4 Directions for future research .……………………………………………………………………………….240
References .………………………………………………………………………………………………………………………...245
Appendix 1 .………………………………………………………………………………………………………………………...260
Appendix 2 .………………………………………………………………………………………………………………………..267

vi
List of Tables

Table 1.1 Claims about the empathic and logophoric properties of ziji ……………………………………………17
Table 1.2 Predictions for the blocking effect with different approaches …………………………………………22
Table 1.3 Proportions of speech, belief, perception verbs in studies reviewed by Sloggett (2017)…………35
Table 2.1 Summary of statistics in Experiment 1a (*: < 0.05)………………………………………………………….65
Table 2.2 Summary of statistics for log-transformed RTs in Experiment 1b (*: < 0.05)………………………68
Table 2.3 Summary of statistics in Experiment 2a (*: < 0.05) ………………………………………………………..75
Table 2.4 Summary of statistics for log-transformed RTs in Experiment 2b (*: < 0.05)………………………81
Table 3.1 Number and proportions of speech, belief, and perception verbs in Experiments 3 and 4…..102
Table 3.2 Summary of statistics for Experiment 3a (*: < 0.05)……………………………………………………….105
Table 3.3 Summary of statistics for log-transformed RTs in Experiment 3b (*: < 0.05)……………………..109
Table 3.4 Summary of statistics for Experiment 4a (*: < 0.05)……………………………………………………….116
Table 3.5 Summary of statistics on the comparison of antecedent choices in Experiment 3a (ziji)
and Experiment 4a (ta-ziji) (*: <0.05) …….……………………………………………………………………..…………117
Table 3.6 Summary of statistics for RTs in Experiment 4b (*: < 0.05)…………………………………………….120
Table 4.1 Summary of statistics for Experiment 5a (*: < 0.05)………………………………………………………150
Table 4.2 Summary of statistics for log-transformed RTs in Experiment 5b (*: < 0.05). Note that
when the matrix verb is ‘heard’, the next region is the embedded subject (e.g., ‘Prof. Wang’)..…………154
Table 4.3 Summary of statistics for Experiment 6a (*: < 0.05)………………………………………………………161
Table 4.4 Summary of statistics for log-transformed RTs in Experiment 6b (*: < 0.05)…………………….165
vii
Table 4.5 Summary of statistics for Experiment 7 (*: < 0.05)………………………………………………………. 172
Table 5.1 Predictions for the blocking effect with different approaches………………………………………..184
Table 5.2 Summary of statistics for Experiment 8a (*: < 0.05)……………………………………………………..192
Table 5.3 Summary of statistics for log-transformed RTs in Experiment 8b (*: < 0.05)……………………195
Table 5.4 Summary of statistics for Experiment 9a (*: < 0.05)……………………………………………………..204
Table 5.5 Summary of statistics for between-experiment comparison (*: < 0.05)…………………………...205
Table 5.6 Summary of statistics for log-transformed RTs in Experiment 9b (*: < 0.05)…………………….209
Table 5.7 Summary of statistics for log-transformed RTs in Experiment 9c (*: < 0.05)……………………..215

viii
List of Figures

Figure 2.1 Illustration of animacy mismatch effect assuming a standard cue-based retrieval process...58
Figure 2.2 Illustration of no animacy mismatch effect assuming a structure-based retrieval process….59
Figure 2.3 Distribution of mean acceptability ratings by condition across all individual participants
(Panel A) and mean acceptability ratings at the group level (Panel B) in Experiment 1a.............……......65
Figure 2.4 Mean RTs across conditions by region in Experiment 1b……………………………………………….69
Figure 2.5 Distribution of mean acceptability ratings by condition across all individual participants
(Panel A) and mean acceptability ratings at the group level (Panel B) in Experiment 2a...................…...75
Figure 2.6 Mean RTs across conditions by region in Experiment 2b ……………………………………………..78
Figure 3.1 Predictions of the structure-based retrieval model for sentences with discourse topics …….96
Figure 3.2 Predictions of the standard cue-based retrieval model for sentences with discourse topics
…………………………………………………………………………………………………………………………………………...98
Figure 3.3 Distribution of mean proportions of local readings of ziji across individual participants
(Panel A) and mean proportions of local readings (Panel B) at the group level in Experiment 3a……..104
Figure 3.4 Mean RTs (ms) across all regions in the critical sentence in Experiment 3b……………...…….110
Figure 3.5 Distribution of mean proportions of local readings of ta-ziji across individual participants
(Panel A) and mean proportions of local readings (Panel B) at the group level in Experiment 4a……..116
Figure 3.6 Mean RTs (ms) across all regions in the critical sentence in Experiment 4b……………………120
Figure 4.1 Distribution of mean proportions of non-local readings of ziji across all individual
participants (Panel A) and mean proportions of non-local readings (Panel B) at the group level
ix
in Experiment 5a…………………………………………………………………………………………………………………. 151
Figure 4.2 Mean RTs (ms) across all regions in the critical sentence in Experiment 5b. In the
perceiver conditions, there is no matrix object (e.g., ‘others’) …………………………………………………..…155
Figure 4.3 Participants’ non-local binding preference in RTs. The non-binding preference in
real-time sentence processing is quantified by subtracting the RTs in the other-directed verb
condition from the RTs in the self-directed verb condition…………………………………………...…………….157
Figure 4.4 Distribution of mean proportions of non-local readings of ta-ziji across all individual
participants (Panel A) and mean proportions of non-local readings (Panel B) at the group level
in Experiment 6a……………………………………………………………………………………...…………..………………161
Figure 4.5 Mean RTs (ms) across all regions in the critical sentence in Experiment 6b. In the perceiver
conditions, there is no matrix object (e.g., ‘others’) …………………………………………………………………..164
Figure 4.6 Distribution of mean proportions of non-local readings of ziji or ta-ziji across all
individual participants (Panel A) and mean proportions of non-local readings (Panel B) at
the group level in Experiment 7……………………………………………………………………………….………….….172
Figure 5.1 Predictions of the standard and structure-mediated cue-based retrieval models for the
early- stage blocking effect……………………………………………………………………………………………………..186
Figure 5.2 Distribution of mean proportions of local readings of ziji across all individual
participants (Panel A) and mean proportions of non-local readings (Panel B) at the group
level in Experiment 8a…………………………………………………………………………………………………………..192
Figure 5.3 Mean RTs (ms) across all regions in the critical sentence in Experiment 8b………………..….196
Figure 5.4 Distribution of mean proportions of local readings of ziji across all individual
x
participants (Panel A) and mean proportions of non-local readings (Panel B) at the group
level in Experiment 9a……………………………………………………………………………………………………..…...204
Figure 5.5 Mean RTs (ms) across all regions in the critical sentence in Experiment 9b……………………208
Figure 5.6 Mean RTs (ms) across all regions in the critical sentence in Experiment 9c……………………214
Figure 5.7 Individual variation in subject blocking effects. Shown are the by-participant subject
blocking effect in Experiment 8b at the three spillover regions. Each dot represents a participant’s
average blocking effect across all target trials. The grey lines connect the data points of the same
participant………………………………………………………………………………………………………………………….226
Figure 5.8 Individual variation in object blocking effects. Shown are the by-participant object
blocking effect in Experiment 9c at the critical and the spillover regions. Each dot represents a
participant’s average blocking effect across all target trials. The grey lines connect the data points
of the same participant…………………………………………………………………………………………………….…..227

xi
Abstract
A reflexive pronoun (e.g., himself) depends on another referent in the discourse for interpretation. But
when multiple referents exist, a comprehender needs to identity which referent the reflexive is to be
anchored to. The main goal of this dissertation is to use linguistic experiments to examine how syntactic
factors (e.g., locality, syntactic position) and non-syntactic factors (e.g., verb semantics, logophoricity,
perspectival information) impact reflexive resolution in Chinese and how these factors influence real-
time processing. There are two general goals for this dissertation.
First, using linguistic experiments, this work aims to shed light on the linguistic nature of two
reflexives in Mandarin, the morphologically simple ziji (‘self’) and the morphologically complex ta-ziji
(‘s/he-self’), in order to re-examine conventional linguistic judgment and, more importantly, to provide
new data contributing to the advances of linguistic theories.
Second, this dissertation assesses and compares two competing antecedent retrieval models that
have been proposed in the sentence processing literature, the standard cue-based retrieval model (e.g.,
Lewis & Vasishth 2005; Jäger et al. 2015; Patil et al. 2016) and the structure-based retrieval model (e.g.,
Sturt 2003; Xiang et al. 2009; Dillon et al. 2014; Kush et al. 2015). The central question is whether
syntactic information is prioritized in early-stage processing compared to non-syntactic information.
The antecedent choice judgment results show that, from a linguistic-descriptive point of view, both
ziji and ta-ziji are subject to syntactic, semantic, and discourse-level constraints. However, the extent to
which discourse topic prominence impacts the interpretation of ziji and ta-ziji differs: ziji is more
heavily influenced by the discourse topic status of the potential antecedent compared to ta-ziji. I ascribe
xii
this difference to the different linguistic properties of these two reflexive forms: ziji is perspective-
sensitive (empathic) while ta-ziji is not.
Critical for the assessment of antecedent retrieval models, the self-paced reading studies show that
when discourse-level information is considered in the experimental design, syntax does not play a
predominant role in antecedent retrieval. Therefore, the structure-based retrieval model cannot
account for the absence of a syntax-first effect and the presence of early discourse-related effects. The
experimental results are thus more consistent with the standard retrieval model. However, a limitation
has been found with the standard retrieval model as well because it cannot predict the absence of object
blocking effect, where a 1
st
-person pronoun (sentence-external perspective center) in the object position
cannot block long-distance binding by a discourse topic empathy locus (sentence-internal perspective
center). To explain this finding, I propose a conjunction constraint to restrict antecedent retrieval to
the perspective center in a syntactically prominent (i.e., subject) position. The weak blocking effect is
presumably a late-stage effect which only appears in the offline antecedent judgment experiments.

1
Chapter 1: Introduction

1.1 Overview and objectives
A reflexive pronoun such as himself depends on another referent in the discourse for interpretation. But
when multiple referents exist, a comprehender needs to identity which referent the reflexive is to be
anchored to. The referent that the reflexive anchors to is called the antecedent, and the process of
disambiguating the reflexive is called reflexive resolution. In this dissertation, I examine how syntactic
(e.g., locality, syntactic position) and non-syntactic factors (e.g., verb semantics, discourse prominence,
perspective-taking) impact reflexive resolution in Mandarin Chinese and how these factors influence
the real-time processing of two Mandarin reflexives, ziji (‘self’) and ta-ziji (‘s/he-self’). As the process of
locating the antecedent is constrained not only by the human cognitive architecture but also by the
linguistic properties of reflexives, reflexive resolution is a topic that bridges sentence processing and
linguistic theories. Therefore, under the general topic of reflexive resolution, this dissertation has two
themes, a psycholinguistic one and a linguistic one.
First, by using online and offline experiments,
1
this dissertation aims to shed light on the linguistic
properties of Mandarin reflexives, the morphologically simple ziji and the morphologically complex ta-
ziji composed of the pronoun ta and simple reflexive ziji. The goal is to re-examine assumptions about

1
In psycholinguistic terms, “online” experiments typically measure participants’ linguistic behavior (e.g., reading
times) or neurological responses (e.g., event-related potentials) as participants process sentences incrementally,
as in a self-paced reading task. In this dissertation, the term “online” is synonymous with the term “real-time” and
should not be confused with the term “internet-based.” “Offline” experiments are ones that do not tap into
participants’ incremental reading times and only reflect participants’ final linguistic judgments/decisions.
Offline tasks include acceptability judgments and antecedent selection (also called forced choice task in this
dissertation).

2
the linguistic properties of ziji and ta-ziji and, more importantly, to provide new data that can
contribute to the advances of linguistic theories. Knowledge about the linguistic properties of ziji and
ta-ziji will also be instrumental to my investigation of the incremental processing profiles of these
reflexives. As will become clear in the following sections, despite prior in-depth theorizing about the
syntactic and discourse-pragmatic constraints that govern the distribution and interpretation of ziji and
ta-ziji (e.g., Pan, 1997, 1998, 2001; Pollard & Xue, 1998, 2001; Huang & Liu, 2001; Anand, 2006; Cole et al.,
2006; Huang et al., 2009; Wang & Pan, 2014, 2015a,b; Giblin, 2016; Charnavel et al., 2017), there are
relatively few studies examining these theories from a testable experimental approach (but see e.g.,
Schumacher et al., 2011; He & Kaiser, 2016; Han, 2020; Liu, 2020). Furthermore, there are certain open
questions to be addressed. To these ends, a major part of this dissertation is devoted to the discussion
of linguistic questions based on the results of the experiments.
Second, this dissertation aims to assess and to compare two competing antecedent retrieval models
debated in the sentence processing literature, the standard cue-based retrieval model (e.g., Lewis &
Vasishth, 2005; Van Dyke & McElree, 2006; Chen et al., 2012; Jäger et al., 2015; Patil et al., 2016) and the
structure-based retrieval model (e.g., Sturt, 2003; Xiang et al., 2009; Van Dyke & McElree, 2011; Dillon et
al., 2013; Cunnings & Sturt, 2014; Parker & Phillips, 2017). The central question is whether syntactic
information (or cue) is prioritized over non-syntactic information in early-stage processing. Prior
studies probing the role of non-syntactic constraints have often made use of gender, number, and
animacy to test if semantically matching antecedents in syntactically inaccessible positions interfere
with antecedent retrieval. However, these studies, including those on Mandarin reflexives, typically do
not directly investigate discourse-pragmatic factors (but see e.g., Sturt, 2003; Kaiser et al., 2009;

3
Cunnings & Sturt, 2014; He & Kaiser, 2016; Sloggett, 2017). This omission could bias our understanding
of the complex processes of reflexive resolution. To fill this gap, this dissertation will focus on the
impact of discourse-level information on reflexive resolution.
To accomplish the linguistic and psycholinguistic goals listed above, this dissertation combines
offline and online experimental methods to enrich our understanding of the syntactic and non-
syntactic constraints that modulate reflexive resolution in Mandarin. Note that although the
antecedent retrieval models are evaluated based on reading times, offline experiments are equally
important as they can shed light on the linguistic nature of ziji/ta-ziji. Crucially, offline experiments can
help identify which factors or linguistic cues should be included in our predictions for antecedent
retrieval models. Below, I turn to an overview of the literature to better situate the linguistic and
psycholinguistic questions addressed in this dissertation.

1.2 Research background
This work aims to further our understanding of the linguistic properties of Mandarin reflexives and to
evaluate the two antecedent retrieval models. In this section, I first introduce prior linguistic research
on ziji and ta-ziji in sections 1.2.1 and 1.2.2, respectively, to identify gaps in our knowledge from a
linguistic point of view. I will not give an exhaustive overview of all the important linguistic work on
ziji/ta-ziji (see e.g., Han (2020) for a comprehensive review). Instead, I will mainly focus on the studies
that closely relate to my research questions: how (non-)syntactic constraints impact reflexive
resolution. Once the linguistic – especially discourse-pragmatic – properties of ziji and ta-ziji have been
clarified for the purpose of this work, I will move on to a concise review of the most relevant

4
psycholinguistic studies on reflexive resolution and raise my research questions from a sentence
processing point of view.

1.2.1 Long-distance binding of ziji
This section reviews prior accounts of long-distance binding of ziji and previews some important
notions and properties related to the non-local use of ziji. Arguably, one of the key linguistic properties
of ziji is that it can be exempt from Principle A of the Binding Theory (Chomsky, 1981, 1986) or the locality
constraint. According to Principle A, a reflexive must be bound in its local domain. Generalizing over
different theoretical work (Chomsky, 1986; Huang & Liu, 2001; Charnavel & Sportiche, 2016; Charnavel
& Huang, 2018), a local binding domain is typically the minimal XP (DP, TP, CP) that contains a subject
and a reflexive.
Although ziji can certainly be bound by the local antecedent, linguists noticed early on that ziji can
be free in the local domain (e.g., Huang et al., 1984; Battistella, 1989; Tang, 1989; Yu, 1992; Xu, 1993) and
can have a non-local antecedent. See (1.1) for an example where binding of ziji by the local subject ‘Prof.
Wang’ and the non-local subject ‘Xiaoming’ are both possible.

(1.1) Xiaoming 1 tingshuo Wang laoshi 2 xihuan ziji 1/2.
Xiaoming hear Prof. Wang like self
‘Xiaoming 1 heard that Prof. Wang 2 likes self 1/2.’

5
Different hypotheses have been put forth to explain why ziji can be used as an “exempt anaphor.”
2

Earlier syntactic approaches hold that long-distance (LD) ziji is only apparently non-local, but
underlyingly, it is locally bound. The main argument is that ziji moves cyclically at the Logical Form (LF)
to adjoin to a position that has a local relationship to its target antecedent (e.g., Battistella, 1989; Huang
& Tang, 1991; Cole et al., 1990, 1993; Cole & Sung, 1994; Cole & Wang, 1996). For example, in (1.1) above,
ziji is assumed to be freely generated with a [+3] person feature.
3
Alternatively, it can inherit the [+3]
person feature from the local antecedent. As indicated in (1.2), ziji moves up the tree to adjoin with the
T 2 head and percolates its [+3] feature to T 2 (because T in Chinese is arguably vacuous), after which
subject-verb agreement for TP 2 can proceed. Ziji then moves through all the head positions to the next
TP layer to feed subject-verb agreement in TP 1. In its final position, ziji is bound by the antecedent
Xiaoming that stands in a local relation to it. This is called the head-movement analysis, presented by
Cole et al. (1990). Huang and Tang (1991) advocated a XP-adjunction analysis, implemented differently
but close in spirit (see Charnavel et al., 2017 for a discussion of the respective merits of these two
analyses). In this work, I shall refer to these two similar analyses jointly as the agreement-based account.

2
Local and non-local uses of ziji do not mean that they are two lexical items. According to Pollard and Xue (1998,
2001), ziji is said to anchor to different antecedents determined by linguistic environments. In fact, Charnavel
(2020) posits that “plain” (local) anaphors and exempt (long-distance) anaphors are simply bound by different
antecedents. While the former are bound by overt antecedents, the latter are bound by covert logophoric
operators.
3
Cole and Sung (1994: fn.26) mentioned that ziji can also be generated as [+1] or [+2]. But these parses will be
ruled out due to spec-TP agreement failure at TP
1
. Note that this minor difference in how ziji acquires a person
feature (from the local DP or base-generation) does not affect the derivation of long-distance binding through
cyclic movement.

6
(1.2) Cyclic movement of ziji at LF

However, certain properties of ziji suggest that LD binding of ziji may have a discourse-pragmatic
component. One prominent property is the “blocking effect” to be reviewed in greater detail in Section
1.2.1.2. In (1.3), if the local antecedent is changed to a 1
st
-person ‘I’ or a 2
nd
-person ‘you’, the non-local
reading is now unavailable (e.g., Huang et al., 1985; Tang, 1989; Xu, 1993; Cole & Sung, 1994). I call
pronouns or any referential elements that block LD binding of ziji blockers. Notably, even if a blocker
is in an object position, the blocking effect is still assumed to hold, as in (1.4) (e.g., Xue, 1994; Pollard &
Xue, 1998; Huang & Liu, 2001).

7
(1.3) Xiaoming 1 tingshuo wo/ni 2 xihuan ziji *1/2.
Xiaoming hear I/you like self
‘Xiaoming 1 heard that I/you 2 likes self *1/2.’

(1.4) Xiaoming 1 gaosu wo/ni 2 Wang jiaoshou xihuan ziji *1/2.
Xiaoming tell I/you Prof. Wang like self
‘Xiaoming 1 told me/you 2 that Prof. Wang 2 likes self *1/2.’

According to the agreement-based syntactic account of LD binding, the blocking effect in (1.3) is
due to a crash in subject-verb agreement. Now suppose ziji is generated with a [+1] person feature or
gets its phi-feature from the local ‘I’ in TP 2. After it has moved to TP 1 layer and percolates its [+1] person
feature to T 1, subject-verb agreement crashes because the matrix subject has an incompatible [+3]
person feature. This process is illustrated in (1.5).

8
(1.5) Derivation of the blocking effect based on the agreement-based account

However, this account predicts that an object blocker should not block LD binding because an
object blocker is not in a structural (i.e., subject) position to pass its phi-feature to ziji, hence the
prediction that there should be no object blocking effect, contrary to the observation in (1.4). A more
recent account by Giblin (2016) offers some hope in providing a syntactic solution to the object blocking
effect. I postpone the discussion of this to Section 1.2.1.2. But prior to Giblin (2016), researchers have
proposed that discourse factors, including the crucial notions of perspective or point of view (e.g., Sells,
1987; Zribi-Hertz, 1989), could play a critical role in LD binding of ziji and the blocking effect (e.g., Huang

9
et al., 1984; Pollard & Xue, 1998, 2001; Pan, 1997, 2001; Huang & Liu, 2001).
4
I call this line of approach
the discourse-based account.
According to the discourse-based account, the blocking effect can be best explained if we assume
that ziji is perspective-sensitive. In other words, non-local antecedents can bind ziji when they are
perspective centers or “domains of point-of-view” (Zribit-Hertz, 1989). A perspective center can be
either an attitude holder – a person whose perspective or point of view the sentence expresses – or an
empathy locus – a person whose perspective the speaker takes (Charnavel & Zlogar, 2015; Charnavel,
2020). Specifically for Mandarin, it has been suggested that LD ziji is necessarily empathic (Pan, 2001;
Wang & Pan, 2015a,b), meaning that one must put himself in the shoes of the perspective holder. In
other words, the antecedent of ziji needs to be an empathy locus. The term empathy is defined as “the
speaker’s identification, with varying degrees, with a person who participates in the event that he
describes in a sentence” (Kuno & Kaburaki, 1977: 628). This term is not to be confused with another
term logophoricity which I will clarify shortly.
The idea behind the discourse-based approach is that an indexical pronoun ‘I’ or ‘you’ encodes the
speaker’s own perspective. Thus, LD binding by a (non-local) sentence-internal empathy locus
Xiaoming and the sentence-external empathy locus ‘I’ or ‘you’ in (1.3-1.4) induces a clash of perspectives.

4
I will not discuss the syntactic representation of perspective or point of view which is not the focus of this
dissertation. For the purposes of this dissertation, I follow the literature on Chinese reflexives (e.g., Huang & Liu,
2001; Pan, 1997, 2001; Cole et al., 2006; Huang et al., 2009; Wang & Pan, 2015a,b; Charnavel et al., 2017) and assume
that perspective-taking in reflexive resolution is a discourse phenomenon. However, it is worth mentioning that
it has been argued in the theoretical literature that (different types of) perspectives can be syntactically encoded
as the heads of such phrasal domains as Evaluative Phrase (EvalP) (e.g., Speas & Tenny, 2003; Speas, 2004),
Sentient and Axis Phrases (Nishigauchi, 2014), and Applicative Phrase (ApplP) (Pancheva & Zubizarreta, 2018),
which would make perspective-taking a syntax-discourse interface phenomenon (Speas & Tenny, 2003; Speas,
2004). While this is a valuable theoretical topic, how perspectives are encoded – in the discourse domain or in
the syntax-discourse interface – does not alter the conclusions of this dissertation.

10
As the speaker’s own perspective is ranked higher according to the Speech Act Empathy Hierarchy
(Kuno & Kaburaki, 1977; Kuno, 1987) defined in (1.6), only local binding is allowed. The merit of this
approach is that it can account for the object blocking effect observed in (1.4) in addition to the subject
blocking effect in (1.3).

(1.6) Speech Act Empathy Hierarchy (Kuno, 1987: 212)
The speaker cannot empathize with someone else more than with himself.

Although the above discussion glosses over some important details that will be made explicit below,
it shows the focus of this work, namely, that perspective-taking can influence reflexive resolution of ziji.
As I will mention below, there are some open questions surrounding this topic. In this dissertation, I
examine from an empirical point of view two linguistic questions about ziji. First, does the discourse
role (e.g., source vs. perceiver) of the perspective center impact LD binding of ziji? Second, is there a
syntactic component to the blocking effect? I turn to these two questions next.

1.2.1.1 Logophoricity, empathy, and ziji
In this section, I will clarify the distinction between the notions of logophoricity and empathy and
explain how they relate to LD binding of ziji. The term logophoricity refers to the phenomenon that
certain pronominal forms refer to individuals whose speech, thoughts, or general states of mind are
reported (Clements, 1975: 141). The term logophor was introduced by Hagège (1974) to describe a type
of pronoun in some African languages where a logophoric pronoun must uniquely refer to an attitude

11
holder. Critical for the purposes of this work, Culy (1994) proposed a logophoric hierarchy, shown in
(1.7), to encapsulate the typological observation that logophoric roles higher on the scale – speech
initiators or sources – are more likely to be the antecedents of logophoric pronouns than logophoric
roles lower on the scale – recipients of direct perception or perceivers.

(1.7) Logophoric hierarchy (Culy, 1994: 1062)
Speech > thought > knowledge > direct perception

In the linguistics literature, logophoric pronouns are often taken to express the perspective of the
attitude holder of an utterance. As Charnavel (2021: 132) noted, “logophoricity is used in linguistics to
refer to grammatical reflexes of perspectives.” However, Culy (1997) argues that perspective and
logophoricity should not be conflated. One piece of evidence Culy cited is that multiple logophors can
appear in the same sentence in several languages. As each sentence (or point-of-view domain) carries
only one perspective center (Banfield, 1982; Zribi-Hertz, 1989; cf. Kaiser, 2022), Culy suggests that
perspective is not a core component of logophoricity.
However, Culy (1997) did not make a distinction between perspective-taking and perspective-
reporting. The idea behind each sentence having one perspective center is that each sentence can only
have one referent – empathy locus – that one takes the perspective of. It is conceivable that one
sentence can express multiple perspectives (without requiring the speaker to assume multiple
perspectives), but the speaker can only take the perspective of one referent. In fact, in Japanese, zibun
(‘self’) can have a logophoric/non-empathic use (Oshima, 2006, 2007; Nishigauichi, 2014) where the

12
external perspective center boku (‘I’) – an empathy locus that one takes the perspective of – can co-
exist with the non-local antecedent whose perspective the sentence expresses. In this case, one can take
his/her own perspective (i.e., empathy locus) while expressing the perspective of the non-local
antecedent.
However, due to the common association between logophoricity and perspective, historically, the
term “logophoric reflexive” often refers to exempt anaphors that are perspective-sensitive (e.g., Sells,
1987; Reinhart & Reuland, 1993; Huang & Liu, 2001; Nishigauchi, 2014; Charnavel, 2020). But perspective
sensitivity can cover perspective-reporting and perspective-taking. Thus, it is not accurate to call all
perspective-sensitive reflexives logophoric reflexives. Logophoricity is related to the notion of
perspective-reporting while empathy is related to the notion of perspective-taking. This distinction is
relevant to the study of Mandarin ziji because Wang and Pan (2015a, 2015b) have argued, based on
Oshima’s (2004, 2006, 2007) work, that ziji is necessarily empathic and only under certain occasions
display properties of logophoric reflexives (i.e., de se reading, more below).
To see how Wang and Pan (2015a,b) reached this conclusion, we need to take a step back to look at
Japanese zibun (‘self’) first. Oshima (2007) posits that there are three different uses of zibun, local zibun,
logophoric zibun, and empathic zibun. The difference between logophoric zibun and empathic zibun is
illustrated by Oshima (2007: 23) (The examples come from Kuno, 1978: 212-213):

(1.8) a. Taro 1-wa [boku-ga zibun 1-o but-ta] koto-o mada urande-i-ru.
Taro-TOP I-NOM self-ACC hit-PAST fact-ACC still resent-ASP-PRES
‘Taro 1 still resents that I hit self 1.’

13
b. *Taro 1-wa [boku-ga zibun 1-ni kasi-ta] okane-o nakusite-simat-ta rasii.
Taro-TOP I-NOM self-DAT lend-PAST money-ACC lose-end.up-PAST it.seems
‘It seems that Taro 1 lost the money I lend to self *1.’

LD binding by Taro is grammatical in (1.8a) but ungrammatical in (1.8b). According to Oshima,
zibun is logophoric (perspective-reporting) in (1.8a) but empathic (perspective-taking) in (1.8ab). The
absence of the blocking effect in (1.8a) indicates that the speaker is not taking the perspective of Taro,
hence not violating the Speech Act Empathy Hierarchy. Essentially, (1.8a) is simply expressing the event
from the perspective of Taro, and the speaker is not required to empathize with Taro. In contrast, (1.8b)
requires the speaker to empathize with Taro, yielding two empathy loci, Taro and ‘I’, hence the blocking
effect. The criteria in (1.9) are used by Oshima (2007: 22-23) to distinguish logophoric reflexives and
empathic reflexives.

(1.9) a. Compatibility with 1
st
-person pronouns:
A logophoric reflexive can occur with a 1
st
-person pronoun, while an empathic reflexive cannot.
b. Possibility of non-subject binding/discourse binding:
Only logophoric reflexives can be bound to a non-subject or extra-sentential antecedent.
c. De se interpretation
Only logophoric reflexives induce de se interpretations.

14
Now let us come back to Mandarin. Wang and Pan (2015a,b) argue that ziji is necessarily empathic
because the blocking effect always holds regardless of the intensional or extensional environment that
ziji appears in. Based on (1.9a,b), Wang and Pan (2015a,b) found that ziji fits the criteria for empathic
reflexives because it meets the first two criteria: the blocking effect and subject-orientation (i.e., only
subjects can bind ziji). The third criterion is trickier and subject to debate. It has been argued that ziji
must have a de se reading (Pan, 1997; Huang & Liu, 2001), meaning that the antecedent must be aware
of the event described by the sentence, exemplified by (1.10) (Wang & Pan, 2014: 748):

(1.10) [Scenario 1: Zhangsan says ‘that thief stole my purse!’]
[Scenario 2: Zhangsan says ‘that thief stole that purse!’] (Zhangsan cannot see that it was his
purse.)
Zhangsan shuo pashou tou-le ziji-de qianbao.
Zhangsan say pickpocket steal-ASP self-GEN purse
‘Zhangsan says that the pickpocket stole self’s purse.’

The sentence in (1.10) has been claimed to be true only in Scenario 1 where Zhangsan is aware that
it is his own purpose that has been stolen. However, Wang and Pan (2015a) made another claim that
the same sentence can also be true if the speaker empathizes with Zhangsan. This suggests that ziji can
be a pure empathic anaphor. Therefore, the conclusion is that ziji is necessarily empathic but
sometimes shows logophoric properties (de se reading) in intensional contexts in the presence of a
speech or belief verb (with source subjects).

15
Given the above brief overview, it seems reasonable to hypothesize that ziji is both empathic and
logophoric. Indeed, Huang and Liu (2001) argue that reference to a source constitutes the core use of
logophoric expressions, with reference to other roles lower on the logophoric hierarchy constituting
the extended uses.
5
Based on this, it is expected that ziji should be sensitive to the logophoric hierarchy
as originally proposed by Culy (1997). I shall call this the empathic logophor hypothesis.
But to my knowledge, few studies have tested the empathic and logophoric properties of ziji
experimentally. The only exceptions are Liu (2020) and Xu and Runner (submitted) who have made
preliminary steps to test the logophoric properties of ziji and ta-ziji. (I postpone the discussion of ta-ziji
to Section 1.2.2.2.) Liu (2020) compares the acceptability of LD binding by a source referent to LD
binding by a topical referent, shown in (1.11):

(1.11) a. Ju Lisi shuo, zhe-jian shi shanghai-le ziji.
According.to Lisi say this-CL event hurt-ASP self
‘According to Lisi, this event hurt self.’
b. Shuodao Lisi, zhe-jian shi shanghai-le ziji.
Speaking.of Lisi this-CL event hurt-ASP self
‘Speaking of Lisi, this event hurt self.’

5
Huang and Liu’s logophoric hierarchy (Source > Self > Pivot) mirrors Culy’s hierarchy but partitions the scale
somewhat differently.

16
In (1.11a), the non-local antecedent Lisi is a source, while in (1.11b) Lisi is a topic. Using a two-
alternative “yes/no” acceptability judgment task, Liu found the average proportion of acceptability for
(11.1a) to be close to ceiling, but the average proportion of acceptability for (11.1b) is less than 50%. This
seems to suggest that a source is more likely to license LD binding than a topical referent. However,
there is an alternative explanation for the low acceptability of (11.1b). Lisi is in an object position in (11.1b)
but in a subject position in (11.1a). Given the subject-orientation property of ziji, the contrast in
acceptability is only to be expected. Furthermore, the local antecedent (i.e., ‘this event’) is inanimate in
(1.11a,b), leaving us wondering if the logophoricity effect still remains when the local antecedent is
animate.
Indeed, an antecedent forced-choice study by Xu and Runner (submitted) suggests that when the
local antecedent is animate (e.g., ‘Name 1 thinks/hears Name 2 is underestimating ziji’), the logophoric
role (believer vs. perceiver) of the matrix subject does not influence preferences for LD binding of ziji.
However, there are likewise some limitations with Xu and Runner’s study. First and foremost, it is
conceivable that for ziji to show logophoric properties (i.e., used as an exempt, perspective-sensitive,
anaphor), the matrix subject needs to be prominent in discourse or in a discourse-prominent position,
as in (1.11). Otherwise, ziji may be interpreted by default as a syntactic anaphor. Second, Xu and Runner
did not test the two ends of the logophoric scale – source vs. perceiver – and only compared believer vs.
perceiver. Due to these reasons, a null result does not disprove the hypothesis that ziji can show
logophoric properties.
Nevertheless, Liu (2020) and Xu and Runner’s (submitted) work have laid foundations for the
current work which takes further steps towards investigating the empathic logophor hypothesis using

17
online and offline experiments. For ease of reference, I list in Table 1.1 the different claims about the
empathic and logophoric properties of ziji. The study on LD binding of ziji by empathy loci will be
reported in Chapter 3, and the study on the relation between ziji and the logophoric hierarchy will be
reported in Chapter 4.

Table 1.1: Claims about the empathic and logophoric properties of ziji.
Example theoretical work Experimental work
Property of ziji Huang & Liu
(2001)
Wang & Pan
(2015a,b)
Liu (2020) Xu & Runner
(submitted)
Empathic Yes Yes NA NA
Logophoric Yes Yes Yes No

1.2.1.2 Syntactic prominence and the blocking effect
We have seen that there are two main approaches to the blocking effect in Mandarin. On the one hand,
the agreement-based syntactic approach to the blocking effect can only account for the subject
blocking effect, while the discourse-based account predicts both subject and object blocking effects.
Thus, it seems that a discourse-based approach is empirically superior to a syntactic approach. However,
there are two alternative approaches which predict both subject and object blocking effects. As we shall
see below, these four different accounts make three classes of predictions. In this section, I first
introduce two other theories of the blocking effect, the contiguous-agreement account (Giblin, 2016)
and the unified account (Cole et al., 2006; Charnavel et al., 2017). Then, I summarize the predictions of
all four theories for the blocking effect, which will be tested in Chapter 5.

18
(1.12) Contiguous-agreement account of the blocking effect

The contiguous-agreement account, in brief, is as follows. For LD binding of ziji, the C

head at the
left edge of the matrix clause is merged with an unvalued feature uparticipant. Upon merge, this feature
probes downward for a [+participant] feature in the spec-TP. If the contiguous matrix subject carries a
[+participant] feature, the C head will be valued. All the T heads (matrix and local T heads) will then
inherit this [+participant] feature from the C head. Finally, ziji is valued with [+participant] – the same
[+participant] carried by the matrix subject – by the local T head in a T-V-reflexive chain. This is
illustrated by the left tree of (1.12) with modification from Giblin (2016: 146).
However, there is a condition that the left tree in (1.12) must satisfy: there must be no intervening
DP with a [+participant] feature. If the contiguous DP/spec-TP is [-participant], derivation crashes due
to the violation of the Contiguous Agree constraint (Nevins, 2007). This is what happens with the right
tree in (1.12) where the local DP is [+participant].

19
Let us see, under the contiguous-agreement account, how the blocking effect is derived. Bejar and
Rezac (2009) proposed a classification of features for 1
st
, 2
nd
, 3
rd
person constituents, such that 1
st
and 2
nd

person pronouns are grouped by [+participant], in contrast to 3
rd
person constituents which are [-
participant]. With this, Giblin argues that the blocking effect emerges because the C head of the matrix
clause can only find a [+participant] feature – e.g., a 1
st
person pronoun – in a non-contiguous manner
(i.e., locating the 1
st
person [+participant] ‘I’ in the local domain across the 3
rd
person [-participant]
matrix subject), thus violating the Contiguous Agree constraint. The Contiguous Agree constraint will
be satisfied if the matrix subject is a 1
st
person pronoun, which captures the observation that binding of
ziji by a matrix ‘I’ is grammatical.
6
Note crucially that this account captures both the subject and the
object blocking effect. I reproduce Giblin’s (2016: 150) illustration of the object blocking effect in (1.13).

6
One might wonder about LD binding of ziji by a 3rd person nominal as in ‘Zhangsan
1
says Lisi likes ziji
1
,’ with
Zhangsan being [-participant]. Is the Contiguous Agree constraint violated? Giblin (2016: 159) argues that “the
failure of a probe to find the goal that its features specify does not lead to ungrammaticality.” In other words,
under this view, the Contiguous Agree constraint is satisfied vacuously for LD binding of ziji by Zhangsan.

20
(1.13) Contiguous-agreement account of the object blocking effect

The contiguous-agreement account can explain the object blocking effect using syntactic means.
But it does not predict any difference in the size of the subject vs. object blocking effects. This is different
from the unified account of the blocking effect which I turn to next.
The unified account (Cole et al., 2006; Charnavel et al., 2017) acknowledges that the blocking effect
has both a syntactic component and a discourse component. This is based on the following judgment
differences:

(1.14) a. Zhangsan 1 yiwei Lisi 2 hui ba Xiaoming 3 dai hui ziji 1/2/3 de jia.
Zhangsan think Lisi will BA Xiaoming take back self GEN home
‘Zhangsan 1 thought Lisi 2 would bring Xiaoming 3 to self’s 1/2/3 home.’
b. Zhangsan 1 yiwei Lisi 2 hui ba ni 3 dai hui ziji ?1/2/3 de jia.
Zhangsan think Lisi will BA you take back self GEN home

21
‘Zhangsan 1 thought you 2 would bring Xiaoming 3 to self’s ?1/2/3 home.’
c. Zhangsan 1 yiwei wo 2 hui ba ni 3 dai hui ziji *1/2/3 de jia.
Zhangsan think I will BA you take back self GEN home
‘Zhangsan 1 thought I 2 would bring Xiaoming 3 to self’s *1/2/3 home.’

In (1.14a), there are three potential legitimate antecedents, the matrix subject Zhangsan, the local
subject Lisi, and the local object of the BA construction Xiaoming. Here, the object Xiaoming can be the
antecedent because it arguably c-commands ziji (Cole & Wang, 1996; also see Charnavel et al., 2017 for
a discussion of whether subject orientation can be reduced to c-command relation). In (1.14b), LD
binding by Zhangsan is marked due to the intervening object blocker ‘you’, where there are two
empathy loci. Note that local binding by Lisi is still allowed as local binding can be implemented
syntactically (e.g., Pan, 1997; Pollard & Xue, 1998, 2001; Huang & Liu, 2001). Finally, in (1.14c), LD reading
of ziji is simply unavailable with a subject blocker ‘I’.
The contrast in acceptability for the sentences above is cited by Cole et al. (2006) and Charnavel et
al. (2017) as evidence supporting a “dual-route” analysis of the blocking effect. The weak degradation of
LD binding in (1.14b) is claimed to be due to the conflict of perspectives between the empathy locus
Zhangsan and the speaker’s perspective (‘you’ is uttered from the perspective of the speaker) (e.g.,
Huang & Liu, 2001; Cole et al., 2006; Huang et al., 2009). The blocking effect is weak because, according
to the agreement-based analysis (Cole & Sung, 1994; Huang & Tang, 1991), the blocker ‘you’ is not in a
subject position to value ziji. In other words, LD binding in (1.14b) is degraded because the sentence
only violates the Speech Act Empathy Hierarchy but not syntax. The blocking effect is strongest with

22
(1.14c) because the subject ‘I’ is a blocker in a syntactic and a discourse sense. Therefore, the unified
account argues that the blocking effect does not have a unitary source. However, one complication with
(1.14c) is that there are two blockers in the sentence which presumably strengthens the speaker’s
perspective compared to (1.14b). This will be addressed in my study in Chapter 5.
In summary, the four accounts of the blocking effect of ziji reviewed in this section can be grouped
into three different approaches, the syntactic approach, the discourse-based approach, and the unified
approach. I present their predictions for subject and object blocking effects in Table 1.2.

Table 1.2: Predictions for the blocking effect with different approaches.
Agreement-
based account
Contiguous-
agreement account
Discourse-based
account
Unified account
Subject blocking Strong Strong Strong Strong
Object blocking None Strong Strong Weak

All four accounts predict strong blocking effect with subject blockers. But they differ in their
predictions about object blocking effects. The contiguous-agreement account and discourse-based
account predict strong object blocking effects, in contrast to the null or weaker object blocking effects
predicted by the agreement-based and the unified accounts. It should be mentioned that introspective
judgments regarding the key data are not very clear. It seems that different researchers have different
tuitions about subject and object blocking effects, which suggest inter-speaker variation. Individual
variation will also be looked into in my work on blocking effects.
Although this section focuses on different analyses of and the open questions about the blocking
effect, I should note that other factors such as verb semantics can also modulate the blocking effect, as
suggested by some theoretical work (e.g., Yu, 1992; Y. Huang, 1994). Thus, in addition to testing the

23
linguistic accounts of the blocking effect, this work will also look closely at the interaction between verb
bias semantics and the blocking effect, a topic that has produced some inconsistencies in the literature
(e.g., Yu, 1992; Y. Huang, 1994; Pollard & Xue, 1998). For example, when a blocker biasing towards local
binding and an other-directed verb biasing towards non-local binding conflict, which antecedent is
preferred? The blocking effect is the topic of Chapter 5.

1.2.2 Long-distance binding of ta-ziji
The complex reflexive ta-ziji has not received as much attention as the monomorphemic ziji,
presumably due to the assumption that ta-ziji is strictly local (e.g., Battistelle, 1989; Tang, 1989;
Battistella & Xu, 1990; Cole et al., 1990; Huang & Tang, 1991). The supposed unavailability of non-local
binding is shown below:

(1.15) Xiaoming 1 tingshuo Wang jiaoshou 2 xihuan ta-ziji *1/2-de jia.
Xiaoming hear Prof. Wang like he-self-GEN home
‘Xiaoming 1 heard that Prof. Wang 2 likes he-self *1/2’s home.’

However, it is well known that the so-called “local” reflexives can be used as exempt anaphors (e.g.,
Pollard & Sag, 1992; Reinhart & Reuland, 1993) when certain discourse conditions hold. As Pollard and
Xue (2001: 329) pointed out, “there is a general consensus that logophoricity and discourse prominence
are among the factors involved in the distribution of non-syntactic antecedents of reflexives.” In fact,
crosslinguistic studies on complex reflexives suggest that this is the case for English (e.g., Jackendoff,
1987; Zribi-Hertz, 1989; Pollard & Sag, 1992; Reinhart & Reuland, 1993; Baker, 1995), French (e.g., Zribi-

24
Hertz, 1995; Charnavel & Spotiche, 2016; Charnavel, 2020), Korean (Kim & Yoon, 2009, 2020), among
other languages (e.g., Turkish, Gracanin-Yuksek et al., 2017).
Given the emerging realization that discourse-level prominence may impact the interpretation of
complex reflexives, it is important to examine whether ta-ziji in Mandarin also displays sensitivity to
discourse-level prominence. However, to my knowledge, there is little empirical work investigating the
role of discourse-level prominence in the interpretation of ta-ziji. Below, I discuss two types of
discourse-level prominence – discourse topic prominence and logophoricity – and LD binding of ta-ziji.

1.2.2.1 Discourse topic prominence and ta-ziji
While LD binding of ta-ziji is not a frequent phenomenon, it has been noticed in the theoretical
literature that ta-ziji can be bound by a non-local antecedent, provided that the local antecedent is
inanimate (Pan, 1997, 1998; Pan & Hu, 2002). Pan (1998: 774) provided an example, shown in (1.16):

(1.16) Zhangsan shuo na-ben shu hai-le ta-ziji.
Zhangsan say that-CL book hurt-ASP he-self
‘Zhangsan said that that book hurt he-self.’

Pan argues that the binding of ta-ziji is subject to a prominence hierarchy. For example, animacy
contributes to the semantic prominence of the antecedent. In (1.16), although local, the antecedent
‘book’ is inanimate and semantically less prominent than Zhangsan. According to Pan (1998) (also see
Pan & Hu, 2002), this animacy prominence constraint outranks the locality constraint which makes the

25
LD antecedent Zhangsan an optimal candidate. If the local antecedent is both animate and local, only
local binding is allowed. This is called the “animacy blocking effect” by Pan (1998).
However, animacy prominence is a type of semantic prominence. Thus, one gap in our knowledge
regarding LD binding of ta-ziji concerns the potential impact of other types of prominence, e.g., whether
the presence of a prominent discourse topic can override the locality constraint. Here, I assume that a
discourse topic referent is more prominent at the discourse level than a non-topic referent because in
a stretch of discourse, the non-topic referent is only part of a proposition or description about the topic
referent. Essentially, I will explore whether Pan’s (1998) prominence-based constraint can be extended
to the discourse level for ta-ziji. (In Chapter 3, I will also examine how discourse topicality impacts ziji.
But discourse topicality impacts ziji and ta-ziji in different ways. See details in Chapter 3.)
The question of whether discourse topic prominence impacts the interpretation of ta-ziji is not
trivial. Note that the animacy blocking effect is reminiscent of the “intervention effect” observed for
“standard American English” (Pollard & Sag, 1992): when a local human c-commander is available, the
reflexive must be locally bound. It has been noted that the LD use of himself/herself as in (1.17) is
“ungrammatical” in standard American English, as mentioned by Pollard and Sag (1992) and Pollard
and Xue (1998), in contrast to British literary English:

(1.17) Philip was supposed to be fooling…because Desiree had undoubted explained to them the
precise nature of her relationship with himself.

26
Example (1.17), from a literary work (Lodge, 170) and cited by Zribi-Hertz (1989: 708), illustrates the
LD use of himself in British literary English. However, standard American English seems to prohibit this
use, despite the fact that the real antecedent Philip is a perspective center (empathy locus) in the
“domain of point-of-view.” The source for the intervention effect could be ascribed to computational
economy (Reuland, 2001, 2011). The idea is that binding in the syntax module is less costly than binding
in discourse. As we shall see in Section 1.3, a large body of sentence processing work supports this view
even when the non-local antecedent is prominent in discourse. Now, given the parallel between the
intervention effect and the animacy blocking effect, it is reasonable to speculate that ta-ziji might
exhibit a similar intervention effect as himself in standard American English. Alternatively, discourse
topic prominence may show crosslinguistic variation in anaphora resolution. Conceivably, ta-ziji might
be more sensitive to discourse topic prominence compared to English himself/herself.
Although the relation between discourse topic prominence and ta-ziji has not been carefully
examined before, Kim and Yoon (2020)’s study on the complex Korean reflexive caki-casin can give us
some clue about what to expect. Kim and Yoon tested the acceptability of LD binding of caki-casin by
a discourse-prominent antecedent, as in (1.18) (Kim & Yoon, 2020: 1792):

(1.18) Context sentence: Last weekend, Youngmi and Jieun attended the alumni associate meeting.
There was an election for president at the meeting. Jieun whispered to Youngmi, “I nominated
you as president.” After the meeting, when Youngmi saw Hannah later, she told her about
what she heard at the meeting.

27
Critical sentence: Youngmi-nun 1 Jieuni-ka caki-casin 1-ul hoycang-hwupo-lo
Youngmi-TOP Jieun-NOM self-ACC president-candidate-as
chwuchenhayssta-ko Hannah-eykey malhayss-ta.
nominated-COMP Hannah-DAT said
‘Youngmi 1 told Hannah that Jieun nominated self 1 as president.’

In the context, it is clear that Youngmi had been nominated as the president. Thus, in the critical
sentence, caki-casin refers to the non-local antecedent Youngmi. In addition, Youngmi is a discourse
topic in the sentence immediately preceding the critical sentence and marked with a topic marker -nun.
Kim and Yoon (2020) discovered that 43% of the time, Korean native speakers accepted LD binding
of caki-casin. It is interesting to note that caki-casin, similar to ta-ziji, is typically considered to be a local
reflexive (e.g., Yoon, 1989; Cole et al., 1990). Yet, LD binding was judged to be borderline acceptable
when the matrix subject is prominent at the discourse level. Based on the findings from caki-casin, we
can speculate that ta-ziji, too, might show sensitivity to discourse topicality. The study on discourse
topic prominence and LD binding of ta-ziji is the topic of Chapter 3.

1.2.2.2 Logophoricity and ta-ziji
In this section, I discuss LD binding of ta-ziji and the notion of logophoricity mentioned earlier. As a
reminder, recall that in this work, a reflexive is taken to be logophoric if the reflexive is sensitive to the
logophoric hierarchy (Culy, 1997). Here, I consider the two ends of the logophoric hierarchy, source and
perceiver.

28
A search in the literature suggests that there is little discussion about logophoricity and the complex
reflexive ta-ziji, except for two recent experimental studies (Liu, 2020; Xu & Runner, submitted) which
will be reviewed below. But prior to reviewing these two studies and discussing what the findings
implicate for the current work, I would like to situate this topic in a larger context/literature by
considering the logophoric properties of complex reflexives from a cross-linguistic perspective first.
Prior empirical studies have shown that LD reflexives – including complex reflexives – show
sensitivity to logophoricity as in English himself/herself (Kaiser et al., 2009; Sloggett, 2017) and Korean
caki-casin (Kim & Yoon, 2009). In these studies, a source matrix subject is preferred as the binder
compared to a perceiver subject. For example, Kaiser et al. (2009: Exp.1,2) found that in sentences like
(1.19), although English native speakers strongly preferred local subjects – selecting Peter as the referent
– sources (indicated by told) were selected more often than perceivers (indicated by heard from). This
trending source advantage has been replicated across several tasks. Perhaps more revealingly, Kaiser et
al.’s visual world eye-tracking study shows that the source preference emerges quite early, during the
200-600ms time window.

(1.19) Peter {told/heard from} Andrew about the picture of himself on the wall.

In addition, Kaiser et al. (2009: Exp.3) investigated the proportions of looks to the LD matrix subject
when there is a local c-commanding distractor, shown in (1.20). Although the subject Greg is the
syntactic binder in the local DP domain, indicated by participants’ strong preference for it as the
referent in an offline judgment task, in real-time processing, participants looked more at the LD subject

29
Peter when Peter is a source relative to when Peter is a perceiver. Overall, Kaiser et al. (2009) provided
crucial experimental evidence using both offline and online data that English reflexives in picture NP
constructions can be logophoric. Perhaps equally revealing is the finding that a similar logophoricity
effect also exists with the Dutch reflexive zichzelf (Kaiser & Runner, 2008).

(1.20) Peter {told/heard from} Andrew about Greg’s picture of himself on the wall.

More recently, Sloggett (2017) reached a similar – albeit not entirely identical – conclusion for non-
picture NP reflexives by investigating English native speakers’ reading times for sentences like (1.21). In
contrast to Kaiser et al.’s (2009) finding that the logophoricity effect can appear across animate local
antecedents, Sloggett found that with argument-position reflexives, the source advantage is only
present when the local antecedent is an impossible antecedent for the reflexive, in particular, when the
local antecedent mismatches with the reflexive in two phi-features (number, gender) or when the local
antecedent is inanimate.

(1.21) The {librarian fem/janitor mas} {said/heard} that the schoolboys misrepresented herself…

In (1.21), the stereotypical gender of the matrix antecedent is manipulated to either match or
mismatch with the reflexive herself. The logic is that if participants attempted LD binding, the gender
mismatch between janitor and herself would lead to surprisal, hence reading slowdowns. Indeed,
reading times from eye-tracking data suggest that English native speakers showed early sensitivity

30
(indicated by first pass reading times) to the gender manipulation of the matrix subject. But this only
holds when the local antecedent is a plural masculine DP like schoolboys. When the local antecedent is
semantically congruent schoolgirl, no logophoricity effect was detected.
The above overview of logophoricity and English reflexives indicates that under certain conditions
(e.g., the reflexive is a picture NP reflexive or the local antecedent is an unlikely antecedent), complex
reflexives are more likely to be non-locally bound by source matrix subjects. In fact, another study by
Kim and Yoon (2009) shows that Korean caki-casin is sensitive to the source role of the non-local
topically marked antecedent. Interestingly, Korean native speakers’ acceptability judgement ratings
track a descending numerical trend in line with Huang and Liu’s (2001) predictions for Mandarin ziji:
LD binding by source is more acceptable than LD binding by self which is in turn more acceptable than
LD binding by pivot. Since caki-casin is morphologically complex and has a preferred local
interpretation like ta-ziji, we can speculate that ta-ziji, too, might display the logophoricity effect (i.e.,
source advantage).
However, empirical studies on ta-ziji have not reached any consensus. As mentioned in section
1.2.1.1, Liu (2020) and Xu and Runner (submitted) have investigated the logophoricity property of ta-ziji
(in addition to ziji). Liu’s (2020) two-alternative acceptability judgment study shows that LD binding of
ta-ziji is more acceptable with a source matrix subject than a topical object. Note that, like Sloggett’s
(2017) experimental design, the local antecedent in Liu’s (2020) test stimuli is also a highly unlikely
antecedent. See (1.22) for an example:

31
(1.22) a. Ju Lisi shuo, zhe-jian shi shanghai-le ta-ziji.
According.to Lisi say this-CL event hurt-ASP he-self
‘According to Lisi, this event hurt he-self.’
b. Shuodao Lisi, zhe-jian shi shanghai-le ta-ziji.
Speaking.of Lisi this-CL event hurt- ASP he-self
‘Speaking of Lisi, this event hurt he-self.’

As suggested before, the source advantage observed for (1.22a) is confounded with the grammatical
status of the LD binder: ta-ziji is a subject in the source condition but an object in the topic condition.
Perhaps, more interestingly, Liu observed that ziji shows a stronger logophoricity effect than ta-ziji. This
will be tested in Chapter 4.
The study by Xu and Runner (submitted), however, presents a different picture. Using stand-alone
sentences like ‘Name 1 {thinks/hears} that Name 2 is underestimating ta-ziji’, Xu and Runner did not find
any logophoricity effect under investigation. But note that, in their work, the logophoricity effect is not
defined as the source advantage as they compare ta-ziji’s sensitivity to belief holders (using ‘think’) vs.
perceivers (using ‘hear’). Indeed, the selection of the logophoric roles may be one of the reasons they
failed to observe a logophoricity effect. But another reason could well be that by default ta-ziji is used
anaphorically as a syntactic reflexive.
At this point, it is important to recall that in prior studies (Kaiser et al., 2009; Kim & Yoon, 2009;
Sloggett, 2017; Liu, 2020), the logophoricity effect seems to be conditional on some other factors. In
Kaiser et al. (2009), the reflexive is a picture NP reflexive, which has been argued to be an exempt

32
environment, not strictly subject to syntactic Binding Theory principles. In Kim and Yoon (2009), the
matrix subject is a discourse (and structural) topic. Finally, in Sloggett’s (2017) and Liu’s (2020) studies,
LD binding is forced upon the complex reflexives as the local DP is semantically unsuitable to be the
antecedent. Therefore, we can speculate that the logophoric interpretation of ta-ziji (arguably also for
ziji) needs some licensing factors/contexts. In my work, discourse topic prominence is a potential
licensing factor.
Overall, this section has presented evidence that complex reflexives, including ta-ziji, often show
preferences for source subjects. But the evidence is not strong due to some confounding factors. The
contribution of this work is two-fold. First, I examine the logophoricity effect using both offline and
online experimental methods. Second, I compare the strength of the logophoricity effect between ziji
and ta-ziji.

1.2.3 LD binding of ziji and ta-ziji: A summary
So far, I have introduced the key properties of ziji and ta-ziji. We have seen that ziji is necessarily
empathic, requiring the speaker/comprehender to take the perspective of its LD binder. The empathic
nature of ziji explains the blocking effect. For ta-ziji, we know that LD binding of ta-ziji is possible under
certain conditions. However, there are several open questions. For ziji, we do not know yet whether it
shows the logophoricity effect with an attitude holder (or source) as its non-local antecedent. Moreover,
the relation between the blocking effect and syntactic prominence is yet to be determined. For ta-ziji,
we do not yet know whether LD binding across an animate intervener is possible at all and whether we
can extend Pan’s (1998) prominence hierarchy to the discourse level. Furthermore, there is inconsistent

33
evidence regarding its logophoric properties. Answering these questions can further our understanding
of these reflexives and their differences.

1.3 Antecedent retrieval and sentence processing
Section 1.2 provided information about the comprehension of ziji and ta-ziji from a “static” linguistic
point of view, including what linguistic factors could ultimately impact the interpretation of these
reflexives. In sentence processing work on reflexive resolution, researchers are more concerned with
the time course with which the various linguistic factors guide antecedent retrieval. In psycholinguistic
research, it is not unusual to find a mismatch between people’s final offline interpretations and
incremental, potentially transient processing patterns: certain linguistic factors or cues may play a role
in offline judgment tasks but not necessarily in online parsing (e.g., Cunnings & Sturt, 2014). This could
mean that certain (e.g., non-syntactic) factors are accessed at the later comprehension stages while
other (e.g., syntactic) factors are accessed much earlier.
In this section, I will discuss two antecedent retrieval models that hold different views on whether
syntactic constraints or cues are prioritized at the early processing stages, the standard cue-based
retrieval model and the structure-based retrieval model. Relevant sentence processing work will be
reviewed. I will also highlight in this introductory section some gaps in prior sentence processing work
on ziji and ta-ziji and discuss how my work on ziji and ta-ziji can help shed light on these two processing
models.

1.3.1 Standard cue-based retrieval model

34
The standard cue-based retrieval model was proposed by Lewis and Vasishth (2005). This model aims
to explain moment-by-moment antecedent retrieval in real-time parsing. Clearly, from a real-time
retrieval point of view, memory is involved. In this model, memory is composed of two components,
procedural memory and declarative memory. Procedural memory is responsible for e.g., building and
attaching syntactic structures. Declarative memory, on the other hand, contains non-procedural
information, including lexical (e.g., semantic, subcategorization) information in long-term memory,
recently processed lexical items (in feature matrices) and intermediate syntactic structures in working
memory.
Because the workspace for procedural memory is limited, items not in current focus will be shipped
to declarative memory. When a new lexical item is attached to the syntactic structure, the parser needs
to check the current linguistic structure in the declarative memory and activate the relevant node for
attachment of the new lexical item. Critically, in antecedent retrieval, when an anaphor has been
encountered, a dependency needs to be established between it and an antecedent. This antecedent-
anaphor dependency is established through the association or overlap of retrieval cues encoded on the
anaphor and the potential antecedent since the earliest moments of processing.
For example, suppose the English reflexive herself probes for a local, c-commanding, singular,
feminine DP as its antecedent. The parser searches in declarative memory for a DP that carries those
features [+local, +c-cmd, +sing, +fem].
7
If a DP matches all the retrieval cues, it will be successfully
retrieved as the antecedent. The above example illustrates a case where only one “perfect” antecedent

7
I will set aside the question of how structural relations (e.g., c-command, locality) are to be encoded as cues.
See Kush (2013) for discussion. Syntactic cues are frequently included in the predictions and discussion of cue-
based retrieval models in the sentence processing literature.

35
exists. What if a sentence doesn’t have a fully matching DP? What if there are multiple antecedents
preceding the reflexive? To understand how the parser works under these situations, two important
assumptions of the standard cue-based retrieval model need to be clarified.
First, the probability of retrieving an antecedent is related to the number of matching features
(leaving aside whether retrieval success/speed has a linear or non-linear relation with the number of
matching cues; see Parker (2019) for discussion). More featural overlap means greater retrieval success
and faster retrieval times. To use (1.21) as an example, the local antecedent schoolboys mismatches with
the reflexive herself by two features, number and gender. Thus, anaphoric dependency between them
is unlikely to be established (Parker & Philips, 2017; Sloggett, 2017). If there are more than one candidate
antecedents, antecedents with more feature overlaps will be more likely to be retrieved as the
antecedent. Suppose a reflexive probes for a c-commanding, subject perspective center, then local non-
perspective center is less likely to be retrieved.
Second, the standard cue-based retrieval model assumes that all cues – semantic, discoursal,
syntactic – by default have similar weights. Again, for a reflexive herself, if the local DP is John while the
non-local DP is Mary, there is some probability that the parser will, at the early comprehension stage,
retrieve John – a syntactically legitimate but gender-incongruent referent – as the antecedent. How
about the syntactically illegitimate but gender-congruent referent Mary? The standard model predicts
that it, too, will be retrieved sometimes. In other words, a candidate with a [+local, -fem] feature matrix
and a candidate with a [-local, +fem] feature matrix are equally probable as antecedents.

36
In sum, the standard model does not prioritize accessibility of certain linguistic cues over others.
However, one could imagine a retrieval mechanism where syntactic cues are prioritized or given more
weight. I turn to this proposal below.

1.3.2 Structure-based retrieval model
The structure-based retrieval model is a variant of the standard cue-based retrieval model. It assumes a
similar retrieval process (i.e., antecedents are retrieved by probing for a set of retrieval cues). The
difference is that, in this model, syntactic cues are either prioritized or weighted more heavily than
semantic cues such as gender, number, animacy and other no-syntactic cues like discourse topicality
(e.g., Nicole & Swinney, 1989; Sturt, 2003; Van Dyke & McElree, 2011; Dillon et al., 2013; Cunnings & Sturt,
2014, 2018; Parker & Phillips, 2017).
I will note that even within the structure-based retrieval model, there are two views. One view posits
that only syntactic cues are accessible at the early processing stages. The other view holds that syntactic
cues play a predominant role but do not entirely block the accessibility of the syntactically inaccessible
antecedents. In this dissertation, I will conflate these two views as my goal is to test whether syntactic
cues necessarily have higher weights over discourse-level cues.
That said, the common assumptions about the structure-based retrieval model are that syntactic
constraints are applied early, and that the parser initially mainly attends to antecedents in syntactically
accessible (local, c-commanding, subject) positions. In other words, only when syntactic constraints
are satisfied will non-syntactic constraints be considered. Note that the term syntactic accessibility in
the psycholinguistic literature is often synonymous with syntactic legitimacy. It can also be used in a

37
broader sense to refer to antecedents which are syntactically more prominent (e.g., local, c-
commanding). For example, Cunnings and Sturt (2014) refer to non-local subjects in sentences with
picture NP reflexives as “syntactically inaccessible” even though they acknowledge that picture NP
reflexives are exempt anaphors (e.g., Pollard & Sag, 1992; Reinhart & Reuland, 1993) licensed by non-
local perspective centers. In a strict sense, the local antecedents should not then be considered as
“syntactically accessible” as the syntactic constraint of Principle A may not necessarily be involved.
Anyway, suffice it to say that under a structure-based retrieval framework, syntactic cues (e.g., c-
command, locality, subjecthood) are more prioritized compared to non-syntactic cues.
Another point about the structure-based retrieval model is that the early application of syntactic
constraints does not mean that non-syntactic cues are ignored at all processing stages. In fact, there is
some evidence that antecedents in syntactically inaccessible positions will be considered at some later
processing stages (e.g., Sturt, 2003; Cunnings & Sturt, 2014; Jäger et al., 2020). As we shall see below in
my overview of the seminal work of Sturt (2003), it was found that discourse-prominent non-local
subjects do impact reflexive resolution at the later processing stages. The view that syntactic constraints
can be defeated at the later processing stages is known as the “defeasible filter hypothesis.” However,
regardless of whether syntactic constraints are defeasible at the late comprehension stages, the core
hypothesis of the structure-based retrieval model is that only syntactic cues are utilized or predominant
at the earliest moments of reflexive resolution.
So far, most studies supporting the structure-based model inspect the role of semantic cues such as
gender, number, and animacy. Few studies investigate the role of discourse topic prominence (except
e.g., Sturt (2003), Cunnings & Sturt (2014)) and logophoric prominence (except e.g., Kaiser et al. (2009),

38
Sloggett (2017)). Neglecting discourse-level factors may create a bias in our understanding of the
memory mechanism that subserves real-time comprehension of reflexives. Thus, the focus of this work
will be on the (early) role of discourse-level prominence in antecedent retrieval.

1.3.3 A critical review of the sentence processing literature
Over the past 20 years, a large body of processing work has investigated the role of the syntactic
constraints in online comprehension of reflexives, mostly in English. Overall, the picture is decidedly
mixed, with most studies supporting the structure-based retrieval models (e.g., Nicole & Swinney, 1989;
Sturt, 2003; Gao et al., 2005; Xiang et al., 2009; Van Dyke & McElree, 2011; Dillon et al., 2013, 2016; Qian
& Wu, 2016; Cunnings & Sturt, 2014, 2018; Parker & Phillips, 2017; Wang, 2017; Jäger et al., 2020; also see
the meta-analysis in Jäger et al., 2017). A few studies have found early interference effect from non-local
and/or c-commanding antecedents (e.g., Runner et al., 2006; Kaiser et al., 2009; Thompson & Choy,
2009; Chen et al., 2012; Jäger et al., 2015; Patil et al., 2016; Sloggett, 2017; Chang et al., 2020). Still others
have mixed findings which are tied to e.g., individual variation in working memory (Cunnings & Felser,
2013) and reading speed (Yadav et al., 2022) and activation decay of antecedents (King et al., 2012).
This mixed picture presumably stems from differences in experimental designs, the nature of
specific linguistic properties in certain languages, variation in the test population (e.g., Cunnings &
Felser, 2013), type II error (see discussion in Patil et al., 2016), and statistical power (see discussion in
Jäger et al., 2020). This section is divided into two parts. In Section 1.3.3.1, I focus on studies on English
reflexives where I aim to identify the limitations or gaps in these studies. Then, in Section 1.3.3.2, I review

39
reflexive studies on ziji and ta-ziji to see whether and how those issues identified for the English
reflexives have been addressed or not been addressed in work on Chinese reflexives.

1.3.3.1 What do studies on English reflexives tell us about antecedent retrieval?
Studies on the real-time application of the locality constraint in fact predate the proposal of cue-based
retrieval models. In an early cross-modal priming study, Nicol (1988) presented participants with aural
stimuli like (1.13) with three potential antecedents boxer, skier, and doctor. After the offset of the
reflexive, participants saw a word semantically related or unrelated to any one of the three potential
antecedents. The difference in response times to semantically related and unrelated words is called the
priming effect which indicates the activation level of the antecedent. It was found that the priming
effect is largest when the word is related to the local referent doctor, suggesting local antecedents are
most activated arguably because participants tend to take the local referent as the antecedent.

(1.23) The boxer told the skier that the doctor for the team would blame himself for the injury.

Since the pioneering work of Nicol (1988), a large body of work has reached similar conclusions: the
locality constraint highly restricts antecedent retrieval at the earliest moments of reflexive resolution.
For example, in a seminal eye-tracking study on the role of the locality constraint in reflexive processing,
Sturt (2003) found that English native speakers only considered local antecedents during the early
processing stages in sentences like (1.24):

40
(1.24) a. Jonathan was pretty worried at the city hospital. He remembered that the surgeon had
pricked himself with a used syringe needle.
b. Jennifer was pretty worried at the city hospital. She remembered that the surgeon had
pricked himself with a used syringe needle.

In both sentences above, the local referent surgeon can be the antecedent for the gender-congruent
reflexive. It was hypothesized that if the discourse topic Jonathan is immediately accessible to the
language parser, then (1.24a) should lead to significant reading slowdowns at the reflexive region due
to competition between surgeon and Jonathan. However, eye-tracking measures indicating early-stage
processing – first fixation duration and first pass reading time – showed that the reading times of the
reflexive are not significantly different between (1.24a) and (1.24b). This indicates that the discourse
prominent but syntactically non-local antecedent Jonathan was not considered by English native
speakers during early stages of comprehension. Interestingly, a late eye-tracking measure – second pass
reading time – showed that reading times of the reflexive in (1.24a) were slower compared to (1.24b),
suggesting that at later processing stages discourse prominence plays a role. These results have been
cited frequently in support of the structure-based retrieval model: syntactic constraints filter out
structurally inaccessible antecedents at an early stage.
Importantly, Cunnings and Sturt (2014) replicated Sturt’s (2003) findings with picture NP reflexives.
Picture NP reflexives are reflexives in the following sentence:

(1.25) John thought that Mary’s picture of himself would be on sale.

41

In (1.25), the sentence is consistent with following scenario: Mary took a picture of John which John
thought would be on sale. Picture NP reflexives have long been known to be exempt from Principle A
(e.g., Pollard & Sag, 1992; Reinhart & Reuland, 1993). However, Cunnings and Sturt (2014) found that in
sentences like (1.26) and (1.27), British English speakers did not show signs of interference from the non-
local antecedent Jonathan or Jennifer:

(1.26) {Jonathan/Jennifer} was walking through the military barracks. {He/She} heard that the
soldier had a picture of himself in the middle of the mess hall.

(1.27) {Jonathan/Jennifer} was walking through the military barracks. {He/She} heard about the
soldier’s picture of himself in the middle of the mess hall.

Furthermore, recall that Zribi-Hertz (1989) (also see Pollard & Sag, 1998, 2001) showed that literary
British English allows empathic readings of himself/herself when the non-local antecedent is an
empathy locus. In the case of (1.26-1.27), the discourse topic Jonathan/Jennifer can be the empathy locus
as one tends to empathize with the discourse topic. Even so, British English speakers on average did not
consider non-local antecedents. This finding suggests that whenever possible, English native speakers
prefer to treat reflexives as syntactic anaphors.
However, there are some limitations to this study. First, Cunnings and Sturt (2014) did not find
locality effects during early processing stages with sentences like (1.26) and only found a marginal

42
locality effect in first pass reading times with sentences like (1.27). This does not fit well with the
hypothesis that the locality constraint applies at early comprehension stages. Second, the experiments
could be underpowered (see a similar point made by Jäger et al., 2020 regarding another study by Dillon
et al. (2013)). Low power does not allow one to make strong conclusions due to concern with Type II
error (false negative: statistics shows no effect when there is indeed one) (see e.g., Wilcox, 2017).
The above studies compared the accessibility of semantic gender and the locality constraint.
Another line of work examines the accessibility of morphological number. For example, Dillon et al.
(2013) investigated whether non-local antecedents embedded inside a relative clause interfere with
reflexive resolution in (1.28):

(1.28) The new executive who oversaw the middle manager(s) apparently doubted themselves…

In (1.28), the number features of the local antecedent new executive and the reflexive themselves
mismatch. However, the non-local antecedent middle manager(s) matches the number of the reflexive
when it is plural. The structure-based retrieval model predicts that the number cue of the syntactically
inaccessible antecedent should not facilitate processing as middle manager(s) is filtered out in the early
comprehension stages. Dillon et al. (2013) found evidence consistent with this prediction and
discovered that only in the later processing stages – indicated by total reading times of the reflexive –

43
does number match between middle managers and the reflexive facilitate processing. On this point, a
highly powered replication study by Jäger et al. (2020) showed similar results.
8

The studies discussed so far focus on gender, number, and discourse topic prominence, which
constitute the bulk of research on real-time resolution of English reflexives (also see Badecker & Straub,
2002; Felser et al., 2009; King et al., 2012; Cunnings & Felser, 2013; Parker & Phillips, 2017). A less
discussed topic is how logophoricity – the logophoric role of the antecedent – impacts reflexive
resolution. Two real-time processing studies (Kaiser et al., 2009; Sloggett, 2017) suggest that
logophoricity seems to modulate reflexive resolution.
Using visual world eye-tracking where participants listened to audio stimuli while looking at the
pictures on the computer screen, Kaiser et al. (2009) found that in sentences like (1.29-1.30) English
native speakers showed early sensitivity to the source vs. perceiver role of the matrix subject:

(1.29) Peter {told/heard from} Andrew about the picture of himself on the wall.

(1.30) Peter {told/heard from} Andrew about Greg’s picture of himself on the wall.

In (1.29), the matrix subject is the local antecedent, yet participants showed higher proportions of
looking at Peter in the source (i.e., told) condition compared to the perceiver (i.e., heard from) condition.

8
There is one crucial difference between Jäger et al.’s and Dillon et al.’s findings. In Jäger et al., when the local
antecedent matches the number of the reflexive, the number-matching syntactically inaccessible “distractor”
causes reading slowdowns at the early processing stages due to “cue overload”, i.e., competition between two
number-matching antecedents. This is not consistent with the predictions of the structure-based retrieval model
which holds that semantic cues of syntactically inaccessible antecedents are not visible to the parser.

44
In (1.30), the matrix subject is the non-local antecedent, participants similarly showed a source
advantage. Together, these findings suggest that logophoricity modulates reflexive resolution, at least
with picture NP reflexives.
Unlike Kaiser et al.’s (2009) focus on picture NP reflexives, Sloggett (2017) investigated
logophoricity and (object) argument reflexive resolution. An example is shown in (1.31).

(1.31) The librarian {said/heard} that the schoolboys misrepresented herself at the meeting.

Setting aside the details of the experimental design, Sloggett’s finding is that there seems to be a
trending source advantage, although the results are not robust. What this could mean is that when local
antecedent schoolboys is an impossible binder (due to the mismatch of two phi-features, gender and
number), English reflexives are “co-opted” as exempt anaphors which unsurprisingly show logophoric
properties. But when the local antecedent is a singular referent (e.g., schoolboy), the matrix subject does
not interfere with the processing of herself. In the linguistics literature, this is called an “intervention”
effect. Sloggett further noticed that previous studies made use of different types of verbs, reproduced
below. He suggests that the logophoric role of the antecedent may explain the absence of “intrusion”
effect from non-local antecedents.

45
Table 1.3: Proportions of speech, belief, perception verbs in studies reviewed by Sloggett (2017).
Matrix verb Intrusion
effect Publication Exp Speech Belief Perceiver
Sturt (2003) 1 30% 54% 16% no
Cunnings & Sturt (2014) 1 ⁠— 88% 12% no
2 ⁠— 86% 14% no
Parker & Phillips (2017) 1 86% 14% ⁠— yes
2 72% 28% ⁠— yes

To summarize, we can see that the locality constraint plays a very important role in real-time
processing of English reflexives. Furthermore, it seems to be the case that, in English at least,
comprehenders tend to interpret reflexives as syntactic anaphors by default. Although discourse-level
factors – discourse topic prominence and logophoricity – can modulate the processing patterns of
himself/herself, they have not been shown to play a dominant role. The general picture that emerges
from the review above is that there is a high probability that the syntactic cue of structural locality has
higher weight than non-syntactic cues – gender, number, discourse topic prominence, logophoricity –
at least in English, although it is debatable whether non-syntactic information is indeed inaccessible at
the early processing stages.

1.3.4 What can ziji and ta-ziji say about antecedent retrieval?
The literature reviewed in the previous section suggests that syntactic locality plays a major role in the
real-time comprehension of English reflexives and that there is mixed evidence regarding the early
accessibility of non-syntactic information. In this section, I review relevant studies on Chinese
reflexives to situate the studies of this dissertation.
Echoing the pioneering work of Nicol (1988), Gao et al. (2005) conducted a similar cross-modal
priming study on ziji where it was found that Chinese speakers tend to consider the local antecedent.

46
This suggests that ziji, like English reflexives, is by default parsed as a syntactic reflexive constrained by
Principle A. However, Gao et al.’s study did not inform us of whether Chinese speakers consider
semantically congruent, non-local subjects at all, as their study only probed which referent is more
activated in participants’ mental representation. Likewise, an event-related potential (ERP) study by Li
and Zhou (2010) only examined what ERP effect is triggered in long-distance binding of ziji. Studies like
these do not contribute directly to the debate concerning different types of antecedent retrieval models
(nor do they seek to do so). In fact, both the standard and the structure-based retrieval models predict
local binding preferences when local and non-local referents are plausible antecedents. This is because
the local antecedent satisfies [+local] and [+animate] while the non-local antecedent only satisfies
[+animate], which leads to higher probability of local retrieval due to complete cue overlap.
One of the core questions this dissertation is concerned with is whether non-syntactic cues in
syntactically inaccessible (e.g., non-local, non-c-commanding) positions modulate reflexive resolution.
Several studies support this hypothesis. In a self-paced reading study, Chen et al. (2012) manipulated
the animacy of the syntactically inaccessible distractor to test if semantic animacy guides antecedent
retrieval despite its syntactic inaccessibility. An example is shown below in (1.32):

(1.32) Fanduipai lingxiu biaoshi [zhe-ge shengming [zai {kangyi/kangyizhe} shikong
Opposition leader say this-CL announcement at protest/protester out.of.control
de-shihou] gaojie-le ziji de dangyuan.
DE-time warn-ASP self DE party.member
‘The opposition leader said that this announcement, when the {protest/protesters} went out

47
of control, warned self’s party members.’

At the critical region ziji and the spillover region de (also at the pre-critical verb gaojie ‘warn’ as
discussed by Jäger et al., 2015), reading times are slower when the syntactically inaccessible distractor
is animate (e.g., ‘protesters’) compared to when it is inanimate (e.g., ‘protest’). This suggests that
animacy is one of the cues that guide antecedent retrieval even though it is non-locally embedded in a
temporal adjunct clause.
In (1.32), ziji and its animate antecedent are in a non-local dependency. However, when ziji has a
local animate antecedent, Chen et al. (2012) did not find an interference effect from the syntactically
inaccessible distractor, which was explored further in a reading eye-tracking study by Jäger et al. (2015).
See (1.33) for an example:

(1.33) {Yundongyuan/Pihuating} zai {lingdui/meiti} shijia juda yali de qingkuang xia
Athelete/kayak at team.leader/media exert great pressure DE situation under
chaoyue-le ziji yigong san ci.
outperform-ASP self in.total three times
‘The {athelete/kayak}, when the {team leader/media} exerted great pressure, outperformed
self three times in total.’

Unlike Chen et al. (2012) who only manipulated the animacy of the distractor, Jäger et al. (2015)
controlled for the animacy of both the distractor (e.g., ‘team leader’ vs. ‘media’) and the local antecedent

48
(e.g., ‘athlete’ vs. ‘kayak’). On the one hand, Jäger et al. replicated Chen et al.’s finding that when the
local antecedent matches the animacy of ziji, the animate distractor (e.g., ‘team leader’) does not
interfere with processing. On the other, when the local antecedent (e.g., ‘kayak’) mismatches ziji in
animacy, animate distractors (e.g., ‘team leader’) caused reading slowdowns. However, the standard
retrieval model actually predicts reading speedups – not slowdowns – for the “non-local match”
condition relative to the “double mismatch” condition, because in the former the distractor partially
matches the retrieval cue [+animate] while in the latter the distractor does not match any retrieval cue.
To explain this, Jäger et al. (2015) invoked the concept of “cue confusability.” They argue that [+animate]
and [+c-command] are easily confusable in the interpretation of ziji because these two features often
co-occur. The consequence is similarity-based interference. It is worthy noticing that Dillon et al. (2016)
too found trending interference effects from non-local antecedents during the processing of ziji.
Thus, the studies on Mandarin ziji seem to present a very different picture compared to studies on
English reflexives. Most studies on ziji have found significant or trending interference from syntactically
non-local NPs while studies on himself/herself show divergent results. One could speculate that the
linguistic properties of ziji are the reason behind the divergent findings in English and Chinese as ziji
can appear in non-local environments more freely. In fact, a self-paced reading study by Lu (2011) shows
that Taiwan Mandarin speakers preferred non-local binding of ziji in subject positions.
9
Maybe ta-ziji
will be more similar to himself/herself in terms of strict observance to the locality constraint? It turns
out there is evidence that ta-ziji is also subject to the influence of non-syntactic information.

9
It is also possible that reflexives in subject position are likely to yield an emphatic reading bounded by matrix
subjects.

49
In a self-paced reading study, Qian and Wu (2016) manipulated the gender congruency between ta-
ziji and (non-)local antecedents. See (1.34) for an example sentence.

(1.34) {Lijun male/Limei female} shuo Huwei female zaojiu tuijian ta male/ta female-ziji …
Lijun/Limei say Huwei already recommend he/her-self
‘{Lijun/Limei} said that Huwei had already recommended {he-self/her-self}…

At the critical region ta-ziji, participants showed reading slowdowns when the gender of the
reflexive mismatches with the gender of the local antecedent (i.e., Huwei ~ ‘he-self’), suggesting that
Chinese speakers prefer local binding for ta-ziji. But at the fourth spillover region, when the local
antecedent matches the gender of the reflexive, an animate matrix subject (i.e., non-local antecedent)
caused reading slowdowns, an effect which also trended numerically at the critical region. This
numerical trend fits the prediction of the standard cue-based retrieval model although it is unclear
whether this “double-match” reading slowdown is an early effect.
In a follow-up reading eye-tracking study which included the stimuli used by Qian and Wu (2016),
Chang et al. (2020) found additional evidence that syntactically inaccessible antecedents impact
antecedent retrieval. They discovered that when the local antecedent mismatches the gender of the
reflexive (e.g., Huwei ~ ‘he-self’), gender-congruent matrix subjects facilitate real-time comprehension
in first fixation and first pass durations. This also fits the canonical predictions of the standard cue-
based retrieval model: partial cue match (matrix subject being [+animate, -local]) facilitates processing
compared to no cue match (matrix subject being [-animate, -local]). Interestingly, the gender effect

50
emerged even before the reflexive: at the verb region (e.g., ‘recommend’), an interference effect was
already present.
10

In fact, an earlier self-paced reading study by Dillon (2016) found further evidence supporting the
standard cue-based retrieval model: when both local and non-local antecedents match the animacy of
ta-ziji, participants spent more time reading the spillover region compared to the local-match condition.
In other words, the animacy cue of the local and non-local antecedents are both utilized to guide
reflexive resolution. However, this finding should be treated with cause because this “double match”
reading slowdown is only a numerical trend.
Overall, there is some evidence that the processing of ziji and ta-ziji is not strictly guided only by
the locality constraint, although local binding seems to be a default for both reflexives. However, there
are several open areas that remain to be investigated.
First, the locality bias found in prior studies on ziji and ta-ziji can be alternatively explained as a
linear recency bias, instead of a structural or syntactic locality effect. By using prenominal relative
clause structures, I aim to tease apart the respective contribution of syntactic locality and linear recency
to online reflexive resolution. This study which will be reported in Chapter 2 is the bedrock for
subsequent studies examining the interaction of syntactic and non-syntactic factors.
Second, studies on ziji and ta-ziji tend to focus on semantic cues but give insufficient attention to
the role of discourse-level prominence. Recall that studies on English reflexives (Sturt, 2003; Cunnings
& Sturt, 2014) have either found no substantive evidence of immediate accessibility of discourse-level

10
Based on the observation that relevant effects can occur before the onset of the reflexive (ziji or ta-ziji), it should
not come as a surprise that in chapters 3 and 4 of this work, similar early effects at the verb have also been
observed.

51
information or have shown that discourse-level information does not play a dominant role (Kaiser et
al., 2009; Sloggett, 2017). If so, to what extent does this processing profile characterize Chinese reflexives?
Are ziji and ta-ziji similarly influenced by discourse-level information? These questions bear on the
validity of the standard and structure-based retrieval models. Two types of discourse prominence –
discourse topic prominence and logophoric prominence – are examined in Chapter 3 and 4 of this
dissertation.
Finally, it is unclear whether and how discourse-level prominence and syntactic prominence
interact. Does the syntactic position (e.g., subject vs. object) of a discourse-prominent antecedent (e.g.,
perspective center) impact real-time processing? To answer this question, I investigate in Chapter 5
how subject and object blocking effects modulate online reflexive resolution.

1.3.5 Reflexive resolution and antecedent retrieval models: A summary
The literature review above shows that reflexive resolution is clearly influenced by syntactic
information such as syntactic locality. In most cases, it seems that syntactic locality plays a
predominant role in reflexive resolution, for both English and Mandarin reflexives. However, research
on ziji and ta-ziji suggests that semantic cues such as gender and animacy impact early-stage processing
while research on English reflexives presents a more mixed picture. Therefore, the different findings
between studies on English reflexives and those on Chinese reflexives highlight the importance of
crosslinguistic investigation. Generalizations based only on English reflexive studies limit our
understanding of the cognitive architecture underlying antecedent retrieval. For this reason, this

52
dissertation aims to extend the exploration on the interaction of syntactic and non-syntactic
information to the discourse level, which remains underdiscussed in the psycholinguistic literature.

1.4 Outline of the dissertation
This dissertation is organized into the following chapters:
Chapter 2 tests whether the notions of syntactic locality and linear recency can be disentangled.
This is an important step as the two sentence processing models make predictions regarding whether
syntactic information or cue is accessed to the exclusion of non-syntactic cues at the earliest moments
of reference resolution. Showing that the “locality bias” is indeed syntactic in nature is a prerequisite to
testing the two processing models. As the theme of Chapter 2 covers both ziji and ta-ziji, studies in this
chapter also provide us with an opportunity to compare the locality bias strengths of these two reflexive
forms and replicate findings from studies with a similar goal (Lu, 2011; Dillon et al., 2016; Wang, 2017a,b).
Chapters 3 examines the impact of discourse topic prominence on reflexive resolution in Mandarin.
The linguistic goals are to test the effects of verb semantics and discourse topic prominence on ziji and,
more importantly, on ta-ziji which is often taken to be a syntactic anaphor. Building on Pan’s (1998)
proposal of a prominence hierarchy, Chapter 3 explores whether this hierarchy can be extended to the
discourse level. The psycholinguistic goal is to examine the time course of discourse topicality on real-
time reflexive resolution.
Chapter 4 looks at logophoric prominence and reflexive resolution. The linguistic goal is to assess
whether ziji and ta-ziji show logophoric properties and whether ziji and ta-ziji differ regarding their
sensitivity to source vs. perceiver roles. The real-time processing studies in this chapter can offer

53
insights on whether logophoric prominence and discourse topic prominence are accessed at similarly
early stages.
Chapter 5 examines the blocking effect associated with ziji to assess the four accounts of the
blocking effect discussed in Section 1.2.1.2. Furthermore, we will see if perspective centers at different
syntactic positions impact real-time reflexive resolution, which not only provides with a “dynamic” view
of the blocking effect but also allows us to look into the interaction of perspective-taking and syntactic
prominence.
Finally, I will summarize the findings of the studies in this work and discuss their theoretical
implications in Chapter 6. Suggestions for future research will also be offered.

54
Chapter 2: Locality and recency in reflexive resolution

2.1 Introduction
One of the key arguments made in prior work on reflexive resolution is that by default Principle A or
the locality constraint plays an important if not dominant role in antecedent retrieval (e.g., Nicol &
Swinney, 1989; Sturt, 2003; Gao et al., 2005; Cunnings & Sturt, 2014; Dillon et al., 2013, 2016). Local search
is typically viewed as a syntactic process whereby the parser searches for a c-commanding antecedent
in the local domain. However, whether “local” search is really a syntax-driven process is open to debate.
Clarifying this question is meaningful not only from a theoretical linguistic point of view but also pivotal
to the subsequent chapters where I look closely at the time course of the accessibility of syntactic
information – including locality – and non-syntactic information.
Prior studies on English reflexives do not provide a clear answer to this question of whether local
search really reflects a syntax-driven process or simply a recency effect, although researchers have
noticed the recency vs. locality distinction. For example, Sturt (2003: Exp.2) and other researchers
(Cunnings & Felser, 2013; Dillon et al., 2013) embedded the non-local antecedent inside a postnominal
relative clause, shown in (2.1):

(2.1) The surgeon who treated Jonathan pricked himself with a used syringe needle.

The referent Jonathan inside the relative clause is linearly closer to the reflexive but is not in the
same local domain. As the linearly closer Jonathan did not interfere with reflexive resolution, some

55
researchers concluded that what drives real-time reflexive processing at the early stage is syntactic
locality only.
11
However, it should be noted that the non-local antecedent Jonathan, in addition to being
non-local, does not c-command the reflexive and is in an object position. Thus, to isolate the impact of
linear recency, we need to compare local and non-local antecedents that both c-command the reflexive
and have the same grammatical role.
Mandarin provides an ideal testing ground for disentangling linear recency and syntactic locality
with its prenominal relative clause structures. In a typical “two-antecedent” design, both local and non-
local subject antecedents c-command and linearly precede the reflexive (e.g., Gao et al., 2005; Qian &
Wu, 2016; Wang, 2017a,b; Xu & Runner, 2019), illustrated by the schema and an example in (2.2a). When
the linearly closer and syntactically local NP 2 is extracted out of the local domain and becomes the head
noun of a relative clause, NP2 is only a syntactically local antecedent through mediation of its trace/gap
(Aoun & Li, 2003), shown by the new schema and example in (2.2b). Linearly, it is no longer a recent
antecedent.

(2.2) a. Schema: NP 1 … [ TP NP 2 … (ta-)ziji]
Zhangsan shuo laoshi zunzhong (ta-)ziji.
Zhangsan say teacher respect (he-)self
Zhangsan said that the teacher respects (he-)self.

11
This conclusion may be premature as Omaki et al. (2019) showed that even non-c-commanding linearly recent
referents (James/Brianna inside the relative clause) in sentences like “the mechanic that James/Brianna hired
predicted how annoyed with himself the insurance agent was” can be retrieved as antecedents.

56
b. Schema: NP 1 … [ RC t 2 … (ta-)ziji] NP 2
Zhangsan shuo [ RC t zunzhong (ta-)ziji de laoshi] chuming-le.
Zhangsan say respect (he-)self DE teacher become.famous-ASP
‘Zhangsan said that the teacher who respects (he-)self became famous.

Now, by manipulating the semantic features of the recent, non-local antecedent NP 1 and the non-
recent, local antecedent NP 2 in (2.3a,b), we can examine the respective influence of linear recency and
syntactic locality. Suppose we want to test whether native speakers treat a referent in a certain position
in the sentence as the antecedent, we can make it semantically congruent or incongruent (e.g., animacy)
with the reflexive. If a referent is treated as the antecedent, then semantic incongruency relative to
semantic congruency should lead to reading slowdowns, when an antecedent-reflexive dependency
can be established.
To be more specific, let us see an example in (2.3). If linear recency impacts reflexive resolution (i.e.,
Chinese speakers attempt to bind the reflexive to NP 1), then in an incremental reading task, we expect
reading time slowdowns at and/or after the reflexive when NP 1 (‘radio station’) mismatches in semantic
features with the reflexive compared to when NP 1 (‘journalist’) matches the animacy of the reflexive; if
syntactic locality guides reflexive resolution, we should see reading slowdowns at and/or after NP 2 due
to semantic feature mismatch between it and the reflexive. In this work, reading slowdowns due to
animacy mismatch (‘radio station’ ~ (ta-)ziji) – often known as animacy mismatch effect – is used as one
of the diagnostics for detecting binding, following previous work (e.g., Dillon et al., 2016).

57
(2.3) a. Schema: NP 1 mismatch … [ RC t 2 … (ta-)ziji] NP 2 match
Diantai 1 mismatch biaoshi gongbu-le (ta-)ziji-de caifang mudi de jizhe 2 match
Radio.station say disclose-ASP (he-)self-DE interview purpose DE journalist
yinqi-le 9 dajia-de buman.
provoke-ASP everyone- DE anger
‘The radio station said that the journalist who disclosed (he-)self’s interview purpose had
angered everyone.’

b. Schema: NP 1 match … [ RC t 2 … (ta-)ziji] NP 2 mismatch
Jizhe 1 match biaoshi gongbu-le (ta-)ziji-de caifang mudi de diantai 2 mismatch
Journalist say disclose-ASP (he-)self-DE interview purpose DE radio.station
yinqi-le dajia-de buman.
provoke-ASP everyone- DE anger
‘The journalist said that the radio station which disclosed (he-)self’s interview purpose had
angered everyone.’

2.2 Antecedent retrieval models
In addition to probing whether syntactic locality can be disentangled from linear recency, the
experimental design in (2.3a,b) allows us to evaluate the two memory-based retrieval models reviewed
in Chapter 1, the standard cue-based retrieval model and the structure-based retrieval model. The

58
relevant retrieval cues are semantic animacy and syntactic locality (if we follow prior work by assuming
that structural relations can be modeled as retrieval cues).
First consider the standard cue-based retrieval model. This model assumes that all retrieval cues,
syntactic or not, can be immediately used for antecedent retrieval (e.g., Lewis & Vasishth, 2005; Van
Dyke & McElree, 2006; Chen et al., 2012; Jäger et al., 2015; Patil et al., 2016). On this view, processing
difficult and retrieval speed depend on cue overlap between the target antecedents and the reflexive.
More cue overlap facilitates retrieval of an antecedent compared to less cue overlap, illustrated in
Figure 2.1. In our case, NP 1 is animate in (2.3b) but inanimate in (2.3a). As ziji and ta-ziji probe for
animate antecedents, reading times (RTs) of the reflexive in (2.3b) should be faster compared to (2.3a).
This reading time difference is the animacy mismatch effect as mentioned above.

Figure 2.1: Illustration of animacy mismatch effect assuming a standard cue-based retrieval process.

Next consider the structure-based retrieval model which posits that syntactic cues are prioritized
in early-stage comprehension. Only when syntactic constraints are satisfied is non-syntactic
information accessed, under this view. Here, it may be useful to conceptualize the syntactic constraint
as a “gate” as discussed by Van Dyke and McElree (2011). When the gate is closed, semantic cues are
inaccessible at the earliest moments of processing; when the gate is open, semantic cues are accessible.
In (2.3a,b), NP 1 is in the non-local domain. Thus, the structure-based retrieval model predicts that

59
animacy of the syntactically inaccessible NP 1 should not impact reflexive resolution. In other words,
there should be no animacy mismatch effect. This process is illustrated in Figure 2.2:

Figure 2.2: Illustration of no animacy mismatch effect assuming a structure-based retrieval process.

Furthermore, by the time the gap or trace inside the relative clause is encountered, the parser
should be anticipating the upcoming head noun NP 2. Only the animacy of NP 2 should modulate the
reading times such that the semantically inanimate NP 2 in (2.3b) should lead to reading slowdowns
compared to the animate NP 2 in (2.3a). Note that NP 2 comes after the reflexive. Typically, retrieval
models aim to explain retrieval of prior antecedents before the reflexive. But since real-time linguistic
prediction has been widely attested and is an integral part of sentence processing (Kuperberg & Jaeger,
2016), it is conceivable that a structure-based retrieval mechanism can make predictive use of the
relative clause structure to anticipate the head noun to resolve the ambiguity.
Yet, there is a third possibility. If we follow prior work (e.g., Van Dyke & McElree, 2011; Parker &
Phillips, 2017) by assuming that retrieval cues can have different weights, the processing patterns of the
monomorphemic ziji and the bimorphemic ta-ziji may differ. Indeed, the idea that the same constraint
can carry different weights with different referential elements has been proposed in the literature. For
example, Kaiser and Tueswell (2008) showed that in Finnish the referential properties of pronouns and
demonstratives are impacted by linear order and syntactic role differently. While the pronoun hän is

60
more sensitive to the syntactic role (i.e., subject), the anaphoric demonstrative tämä is more sensitive
to linear order (see also Kaiser et al. 2009 on English reflexives and pronouns in picture NPs). Likewise,
ziji and ta-ziji may show different sensitivities to the locality constraint. In fact, prior psycholinguistic
studies lend support to this hypothesis although it is debated which reflexive is more constrained by
locality. Lu (2011) argued that ta-ziji is more locally constrained than ziji, while other researchers (Dillon
et al., 2016; Wang, 2017b) presented opposite evidence suggesting ta-ziji is less locally constrained
during real-time processing.
For this third scenario, I will not make any precise predictions about how different sensitivities to
the locality constraint could impact real-time processing. I simply point out that if real-time reflexive
resolution is guided by cue-based retrieval, we might find different animacy effects with ziji and ta-ziji.
If reflexive resolution is strongly constrained by syntactic locality, we might see a weaker or even an
absent animacy effect between NP 1 and whichever reflexive that is more sensitive to the locality
constraint.

2.3 Overview of the aims
The primary aim of this study is to examine whether syntactic locality is reducible to linear recency. To
this end, two sets of acceptability judgment and self-paced reading experiments were conducted.
Experiment 1 aims to replicate the recency/locality bias of the previous studies which use a ‘two-
antecedent’ design. Experiment 2 makes novel contributions by separating syntactically local from
linearly recent antecedents.

61
A secondary aim is to take another look at the debate regarding the strength of the locality
preference displayed by ziji and ta-ziji. As prior comparative studies largely conflated syntactic locality
and linear recency (Lu, 2011; Wang, 2017a,b) or put linearly recent antecedents in a non-c-commanding
position (Dillon et al., 2016), stronger “locality bias” associated with ziji or ta-ziji may be confounded by
recency or c-command effects. These limitations are avoided in Experiment 2.

2.4 Experiment 1
Experiment 1 implements a ‘two-antecedent’ design where the local NP is also the linearly closer NP.
Experiment 1a is an acceptability judgment task tapping into the final interpretation of the reflexive.
Experiment 1b is a self-paced reading experiment which reflects incremental comprehension processes.
Together, these two types of experiment tasks can give us a more comprehensive view of Chinese
speakers’ interpretation of reflexives.

2.4.1 Materials
Experiments 1a and 1b share the same stimuli. The factors Reflexive (ziji/ta-ziji) and Distance
(recent/distant) are crossed in a factorial design. Animacy match between the animate antecedent (e.g.,
‘journalist’) and the reflexive (i.e., ziji/ta-ziji) is used to create recent (local) or distant (non-local)
dependencies. Distance refers to the linear distance between the animate (plausible) antecedent and
the reflexive. An example set of the 24 target items is shown in (2.4). The subscripts indicate the text-
presentation regions in the self-paced reading task in Experiment 1b. An additional 24 filler sentences
were included. The target items were distributed into 4 lists and pseudo-randomized with the fillers.

62

(2.4) a. ziji/recent
Diantai 1 biaoshi 2 na-ming 3 jizhe 4 yijing 5 gongbu-le 6 ziji-de 7 caifang 8 mudi. 9
Radio.station say that-CL journalist already disclose-ASP self-DE interview purpose
‘The radio station said that the journalist had already disclosed the purpose of self’s interview.’
b. ziji/distant
Jizhe 1 biaoshi 2 na-jia 3 diantai 4 yijing 5 gongbu-le 6 ziji-de 7 caifang 8 mudi. 9
Journalist say that-CL radio.station already disclose-ASP self-DE interview purpose
‘The journalist said that the radio station had already disclosed the purpose of self’s interview.’
c. ta-ziji/recent
Diantai 1 biaoshi 2 na-ming 3 jizhe 4 yijing 5 gongbu-le 6 ta-ziji-de 7 caifang 8 mudi. 9
Radio.station say that-CL journalist already disclose-ASP he-self-DE interview purpose
‘The radio station said that the journalist had already disclosed the purpose of he-self’s interview.’
d. ziji/distant
Jizhe 1 biaoshi 2 na-jia 3 diantai 4 yijing 5 gongbu-le 6 he-ziji-de 7 caifang 8 mudi. 9
Journalist say that-CL radio.station already disclose-ASP self-DE interview purpose
‘The journalist said that the radio station had already disclosed the purpose of he-self’s interview.’

2.4.2 Experiment 1a: Acceptability judgment study
2.4.2.1 Participants

63
Twenty-eight Chinese speakers in mainland China participated in Experiment 1a. Participation was
remote over the internet.

2.4.2.2 Procedure
The acceptability judgment Experiment 1a was hosted on the internet-based survey platform Qualtrics
(Qualtrics, Provo, UT). On each trial, the participants read a sentence and rated its acceptability on a 1-
7 Likert scale. “1” means the sentence is ‘totally unacceptable’ (wanquan bu tongshun in Chinese); “7”
means the sentence is ‘perfectly acceptable’ (wanquan tongshun in Chinese).

2.4.2.3 Predictions
Following prior work which shows recency/locality biases for ziji and ta-ziji (e.g., Dillon et al., 2016;
Wang, 2017a), a main effect of Distance is expected: sentences with recent animate antecedents should
be more acceptable than those with distant animate antecedents. I follow prior work (e.g., Dillon et al.,
2014, 2016; Wang, 2017a,b) by assuming that ziji and ta-ziji typically do not select inanimate referents
(but see Huang & Charnavel, 2018 for exceptions with ziji which do not apply here). Thus, in this
experiment, the locality/recency bias is measured by the animacy mismatch effect. In other words,
animacy mismatch between ziji/ta-ziji and the local/recent antecedent is expected to yield lower
acceptability compared to animacy match between these elements. We might also see an interaction
of Reflexive and Distance if the recency/locality biases associated with ziji and ta-ziji are different: the
animacy mismatch effect shown by ziji may be different compared to ta-ziji.

64
2.4.2.4 Data analysis
Mixed effects ordinal regression was used to analyze acceptability ratings, implemented by the clmm
function of the ordinal package (Christensen, 2019). The fixed effects are Reflexive and Distance. The
random effects are slopes and intercepts grouped by participants and items. Sum codings were fitted
for the factors Reflexive (ziji = 0.5, ta-ziji = -0.5) and Distance (Recent = 0.5, Distant = -0.5). Satterthwaite
approximation was used to transform t-values to p-values, implemented by the lmerTest package
(Kuznetsova et al. 2015). Statistical models were fitted parsimoniously (Bates et al., 2015): simpler
models are preferred over more complex models if the models do not differ significantly. For model
comparison, an alpha level of 0.2 was adopted following Matuschek et al.’s (2017) suggestion.

2.4.2.5 Results
Figure 2.3 displays the acceptability ratings for the four conditions. See Table 2.1 for a summary of
statistics. In Figure 2.3, the left panel is a violin plot which shows the distribution of ratings for all
sentences; the right panel is a dot plot which displays the mean acceptability of the sentences in the
four conditions.
The mixed effects ordinal regression reveals a main effect of Reflexive (p = 0.01) and a main effect of
Distance (p < 0.001). These results suggest two things. First, assuming that higher acceptability is
correlated less processing cost, then sentences containing ziji are easier to process than sentences
containing ta-ziji, presumably due to the higher frequency of ziji. Often, higher frequency leads to easier
access (e.g., Bybee & Hopper, 2001). Indeed, the Center for Chinese Linguistics Corpus (Zhan, 2003)
shows that ziji is over 17 times more frequent than ta-ziji. Second, Chinese speakers prefer short-

65
distance binding over long-distance binding overall, consistent with prior findings (e.g., Jäger et al., 2015;
Dillon et al., 2016; Chang et al., 2020). The Reflexive x Distance interaction is not significant.

Figure 2.3: Distribution of mean acceptability ratings by condition across all individual participants
(Panel A) and mean acceptability ratings at the group level (Panel B) in Experiment 1a.

Table 2.1: Summary of statistics in Experiment 1a (*: < 0.05).
β SE t p
Reflexive 0.44 0.18 2.50 0.01*
Distance 0.64 0.18 3.59 0.0003*
Reflexive x Distance -0.11 0.35 -0.32 0.75

2.4.2.6 Discussion of Experiment 1a
Using acceptability judgment, Experiment 1a has replicated the locality/recency bias of ziji and ta-ziji
identified by previous offline experiments such as acceptability judgments (Dillon et al., 2016; Wang,
2017a,b) and forced choice judgments (Jäger et al., 2015; Chang et al., 2020).
Importantly, Experiment 1a did not find any difference in the locality/recency bias between ziji and
ta-ziji (i.e. no significant interaction). This contrasts with Dillon et al.’s (2016) acceptability judgment
2
4
6
Ziji Ta−ziji
Distribution of ratings
A
4.0
4.5
5.0
5.5
6.0
Ziji Ta−ziji
Mean acceptability rating
Distance
Recent
Distant
B

66
experiment which shows that only ziji, but not ta-ziji, has a locality/recency bias. The different findings
could be related to the fact that the non-local antecedent in Dillon et al.’s study is embedded inside a
relative clause and thus does not c-command the reflexive. It is possible that ziji and ta-ziji show
different sensitivities to the c-command relation as well (which I will not address in this study on
syntactic locality).
Next, I turn to Experiment 1b to investigate the real-time interpretation of ziji and ta-ziji using a self-
paced reading task.

2.4.3 Experiment 1b: Self-paced reading
2.4.3.1 Participants
Forty-two Chinese native speakers from mainland China participated over the internet.

2.4.3.2 Procedure
Experiment 1b was hosted on the internet-based experiment platform Ibex Farm (Drummond, 2013).
Participants read sentences presented region by region, indicated by the subscripts in (2.4). At the end
of each sentence, participants answered a two-alternative comprehension question that does not probe
the interpretation of the reflexive. The accuracies of participants’ answer choices were used as
indicators of attention.

2.4.3.3 Data analysis

67
Before data analysis, it was decided that participants with comprehension accuracies below 75% would
be removed due to concern with lapse of attention during the experiment. All participants passed the
screening with a mean accuracy of 95%. RTs shorter than 100ms and longer than 3000ms were removed
as outliers. RTs above 2.5 SDs of the mean were also removed. These two steps affected 3.23% of the
original data.
The coding scheme is the same as in Experiment 1a. Statistical analyses were conducted over both
raw and log-transformed RTs using mixed effects linear regression, implemented by the lme4 package.
Only statistics on log-transformed RTs are reported below but I will report both when raw and log-
transformed analyses show different results (see similar practice in Parker, 2019). This practice is
adopted for all real-time studies in this dissertation. If a statistical model fails to converge or yields a
singular fit in RT analysis, different controllers will be tried with the all_fit function from the afex
package (Singmann et al., 2022) before simplifying the model in a parsimonious manner outlined in
Experiment 1a.

2.4.4.4 Results
Figure 2.4 shows the raw reading times across conditions by region. See Table 2.2. for statistics. The
critical region is the reflexive region. Prior to the critical region, no main effect or the interaction
reaches significance at any region.

68

69

Figure 2.4: Mean RTs across conditions by region in Experiment 1b.

At the critical region, there is a main effect of Reflexive (p < 0.001), presumably because the
trisyllabic ta-ziji takes longer to read than the disyllabic ziji (Wang, 2017a; Dillon et al., 2016). The main
effect of Distance and the Reflexive x Distance interaction are not significant. At the first spillover region
(e.g., ‘interview’), there is only a main effect of Reflexive (p < 0.001), again presumably due to greater
processing effort related to ta-ziji. Finally, at the sentence-final region, the statistical model revealed a
main effect of Distance (p < 0.001): distant binding is more difficult compared to recent binding. No
other effects are significant.

2.4.4.5 Discussion of Experiment 1b
Experiment 1b replicated the frequently observed locality/recency bias in the real-time processing of
ziji and ta-ziji (e.g., Gao et al., 2005; Dillon et al., 2016; Qian & Wu, 2016; Wang, 2017a,b; Chang et al.,
2020). Similar to Experiment 1a, Experiment 1b did not find any evidence suggesting a locality/recency
Reflexive
350
400
450
500
550
600
Journalist/Radio say that radio/journalist already disclosed (ta)ziji−DE interview purpose.
RT (ms)
Reflexive
Ziji
Ta−ziji
Distance
Recent
Distant

70
bias difference between ziji and ta-ziji. This may seem counterintuitive, as cross-linguistically complex
reflexives are more likely to be locally bound than monomorphemic reflexives (e.g., Pica, 1986) which
might lead us to expect stronger locality bias for ta-ziji (see Lu, 2011). I speculate that this could be
because by default ziji is a syntactic anaphor governed by Principle A, unless triggered by discourse-
pragmatic (e.g., perspective-taking) factors. Given that the target sentences in Experiment 1b are stand-
alone sentences and do not involve perspective-sensitive contexts (see Chapters 3-5 in this dissertation),
perhaps it should not come as a surprise that ziji shows similar locality/recency bias to ta-ziji.
At any rate, the opposite claim that ziji has a stronger locality/recency bias in real-time processing
(see Dillon et al., 2016) is not supported by Experiment 1b. As mentioned above in Section 2.4.2.6, this
could be related to the different c-command relations between the reflexive and the antecedent in this
study and Dillon et al.’s study. It may also be related to the syntactic position of the reflexive: in this
work, I investigated reflexives in genitive positions which are the most frequent forms (Jia, 2020) in the
Lancaster Corpus of Mandarin Chinese (McEnerey & Xiao, 2004); in the studies of Wang (2017a,b) and
Lu (2011), the reflexives are in subject positions which are susceptible to emphatic readings (e.g., Pan,
1995; Xu, 1999).
Together with Experiment 1a, Experiment 1b has replicated the locality/recency bias for ziji and ta-
ziji when these reflexive elements are in the genitive position, which is important because without this
replication, probing the locality vs. recency effects would be pointless with genitive reflexives. Thus,
Experiment 1 has set the stage for Experiment 2 where I explore whether syntactic locality and linear
recency play distinct roles in reflexive resolution in Mandarin.

71
2.5 Experiment 2
Experiment 2 disentangles linear recency from syntactic locality and inspects the respective influence
of these two factors on reflexive resolution in Mandarin. Experiment 2a is an acceptability judgment
task, followed by Experiment 2b which is a self-paced reading task.

2.5.1 Materials
In Experiment 2, the factors Reflexive (ziji/ta-ziji) and Position (Local/Recent) were crossed in a factorial
design. The factor Position refers to the position of the animate antecedent. As mentioned before, ziji
and ta-ziji typically take animate antecedents but not inanimate antecedents. Thus, by placing
(in)animate referents in syntactically local or linearly recent positions, we can examine whether
Mandarin speakers attempt to bind the reflexives to syntactically local or linearly recent antecedents.
In the local conditions, animate antecedents are post-reflexive; in the recent conditions, animate
antecedents are pre-reflexive. Ex.(2.5) displays an example set of the target stimuli. The subscripts
indicate the reading regions in the self-paced reading experiment. In (2.5), the matrix subject is the
linearly recent but structurally non-local antecedent; the head noun of the relative clause is linearly
non-recent but structurally local. A total of 24 sets of target items were distributed into 4 lists using
Latin square. Another 24 filler sentences were mixed with the target sentences and presented to the
participants pseudo-randomly.

(2.5) a. ziji/local
Diantai 1 biaoshi 2 gongbu-le 3 ziji-de 4 caifang 5 mudi 6 de 7 jizhe 8 yinqi-le 9

72
Radio.station say disclose-ASP self-GEN interview purpose Comp journalist provoke-ASP
dajia-de 10 buman. 11
everyone-GEN anger
‘The radio station said that the journalist who disclosed self’s interview purpose had angered
everyone.’
b. ziji/recent
Jizhe 1 biaoshi 2 gongbu-le 3 ziji-de 4 caifang 5 mudi 6 de 7 diantai 8 yinqi-le 9
Journalist say disclose-ASP self-GEN interview purpose Comp radio.station provoke-ASP
dajia-de 10 buman. 11
everyone-GEN anger
‘The journalist said that the radio station which disclosed self’s interview purpose had angered
everyone.’
c. ta-ziji/local
Diantai 1 biaoshi 2 gongbu-le 3 ta-ziji-de 4 caifang 5 mudi 6 de 7 jizhe 8 yinqi-le 9
Radio.station say disclose-ASP he-self-GEN interview purpose Comp journalist provoke-ASP
dajia-de 10 buman. 11
everyone-GEN anger
‘The radio station said that the journalist who disclosed he-self’s interview purpose had angered
everyone.’
d. ta-ziji/recent
Jizhe 1 biaoshi 2 gongbu-le 3 ta-ziji-de 4 caifang 5 mudi 6 de 7 diantai 8 yinqi-le 9

73
Journalist say disclose-ASP he-self-GEN interview purpose Comp radio.station provoke-ASP
dajia-de 10 buman. 11
everyone-GEN anger
‘The journalist said that the radio station which disclosed he-self’s interview purpose had angered
everyone.’

2.5.2 Experiment 2a: Acceptability judgment
2.5.2.1 Participants
Thirty-seven Chinese speakers from mainland China participated remotely in this experiment.

2.5.2.2 Procedure
The procedure for this acceptability judgment task is the same as for Experiment 1a.

2.5.2.3 Predictions
Predictions can be made for the following three hypotheses:
First, if the ‘locality bias’ is in fact linear recency, participants should prefer reflexive binding by a
linearly recent, animate antecedent than by a post-reflexive, animate antecedent. This means that we
should see higher ratings for the recent conditions like (2.5b,d) than the local conditions like (2.5a,c).
In statistical terms, we should expect a main effect of Position.
Second, if linear recency is separable from syntactic locality, participants may or may not favor
recent binding depending which factor – linear recency or syntactic locality – plays a more prominent

74
role in reflexive interpretation. So, in this scenario, we cannot make clear predictions. However, if
participants prefer the local conditions with post-reflexive animate antecedents like (2.5a,c), we will
have clear evidence that linear recency is not the (only) reason behind the ‘locality bias’ observed in
prior studies. In this case, we can expect a main effect of Position with local conditions receiving higher
acceptability.
Finally, the acceptability difference between the recent and local conditions may be different for
ziji and ta-ziji, which means a Reflexive x Position interaction.

2.5.2.4 Data analysis
Data analysis was done in the same way as for Experiment 1a, except for the coding scheme for the novel
factor Position (Local = 0.5, Recent = -0.5).

2.5.2.5 Results
Figure 2.5 displays the distribution of the acceptability ratings for all sentences (left panel) and the
mean ratings (right panel) for all conditions. See Table 2.3 for a summary for statistics. Mixed effects
ordinal models revealed a main effect of Position (p < 0.001) because participants on average preferred
local binding over recent binding. The factor Reflexive and the Reflexive x Position interaction are not
statistically significant.

75
Table 2.3: Summary of statistics in Experiment 2a (*: < 0.05).
β SE t p
Reflexive -0.03 0.15 -0.22 0.83
Position 0.62 0.15 4.06 < 0.001*
Reflexive x Position 0.02 0.30 0.08 0.94

Figure 2.5: Distribution of mean acceptability ratings by condition across all individual participants
(Panel A) and mean acceptability ratings at the group level (Panel B) in Experiment 2a.

2.5.2.6 Discussion of Experiment 2a
The results of the acceptability judgment in Experiment 2a show that Chinese speakers prefer local
binding instead of recent binding. As outlined in our predictions in Section 2.5.2.3, this suggests that
syntactic locality is a separate factor from linear recency. However, Experiment 2a only shows that
Chinese speakers do consider syntactically local antecedents in offline end-of-sentence interpretations,
but it cannot show us whether linear recency guides reflexive interpretation. Neither does it tell us
whether syntactic locality plays a role during the early stages of processing since acceptability judgment
only taps into the final comprehension stage. Therefore, I turn to Experiment 2b to examine the
respective role of linear recency and syntactic locality in real-time reflexive resolution.
2
4
6
Ziji Ta−ziji
Distribution of ratings
A
4.0
4.5
5.0
5.5
6.0
Ziji Ta−ziji
Mean acceptability rating
Distance
Local
Recent
B

76

2.5.3 Experiment 2b: Self-paced reading
2.5.3.1 Participants
Fifty-seven Chinese native speakers in mainland China participated over the internet in Experiment 2b.

2.5.3.2 Procedure
The procedure is identical to Experiment 1b.

2.5.3.3 Predictions
In Experiment 2b, with regard to linear recency, we can make the following predictions. If Chinese
speakers attempt to bind the reflexive to the linearly recent matrix subject, inanimate matrix subjects
in local match (i.e., recent mismatch) conditions like (2.5a,c) should lead to reading slowdowns – an
animacy mismatch effect – at and/or after the reflexive has been encountered relative to recent match
conditions like (2.5b,d). In other words, we should expect a main effect of Position. As noted in Figure
2.1, the standard cue-based retrieval model in fact predicts this result: less featural overlap, in this case
animacy, between the reflexive and the potential antecedent leads to slower retrieval speed.
However, it is possible that Chinese speakers do not consider syntactically inaccessible antecedents
at early processing stages even though the recent antecedents appear first. This fits the prediction of
the structure-based retrieval model: antecedents in syntactically non-local positions will not be
retrieved at first; only post-reflexive local antecedent will be retrieved (assuming the parser engages in
predictive processing). Thus, the structure-based model only predicts a main effect of Position at and/or

77
after the head noun region: animacy only impacts processing patterns for syntactically local
antecedents (head nouns).
Finally, the animacy effects associated with the processing of ziji and ta-ziji could differ at the
reflexive. The reflexive that is more constrained by syntactic locality may be less influenced by the
availability of a linearly recent antecedent at the reflexive region. The animacy effects at the head noun
region may differ as well, but it is hard to predict what the pattern is likely to be due to potential
reflexive-antecedent reanalysis.

2.5.3.4 Results
The by-region reading times for all conditions are displayed in Figure 2.6. See Table 2.4 for statistics.
There are two critical regions, the reflexive inside the relative clause and the head noun of the relative
clause. For regions prior to the first critical region, no effects have been detected.
At the reflexive region, the main effect of Reflexive is significant (p < 0.001), reflecting the fact that
ta-ziji is longer than ziji. The Reflexive x Position interaction is also significant (p < 0.05), indicating that
the animacy of the linearly recent antecedents impacts the processing of ziji and ta-ziji in different ways.
Post-hoc pairwise comparisons show that animacy mismatch between the reflexive and the recent
antecedent (e.g., ‘radio station’) leads to reading slowdowns in the processing of ziji but not in the
processing of ta-ziji.
This interaction thus suggests that, when the reflexive is encountered, participants immediately
attempted to bind ziji, but not ta-ziji, to the linearly recent antecedent. In fact, there is an opposite
numerical trend for ta-ziji with animacy match causing slowdowns, but this trend is only marginally

78
significant in raw RT analysis (β = -28.13, SE = 15.54, t = -1.81, p = 0.07) and not significant in log-
transformed RT analysis (β = -0.04, SE = 0.03, t = -1.59, p = 0.11). I will not read too much into this
‘animacy match effect’ but will highlight the different processing patterns of ziji and ta-ziji at the first
critical region.

Figure 2.6: Mean RTs across conditions by region in Experiment 2b.

At the first spillover region (e.g., ‘interview’) of the reflexive, the statistical model shows main effects
of Reflexive and Position (ps < 0.001). These results suggest that (i) ta-ziji takes longer to process
compared to ziji and (ii) animacy mismatch between the linearly recent antecedent and the reflexive
leads to processing difficulty for both ziji and ta-ziji at a somewhat later processing stage. The Reflexive
x Position interaction is not significant. The main effect of Position was also found (ps < 0.05) at the
second (e.g., ‘purpose’) and third (i.e., DE) spillover regions. No other effects are significant.
Reflexive
Head Noun
400
500
600
Journalist/Radio say disclosed (ta)ziji−DE interview purpose DE radio/journalist provoked everyone's anger.
RT (ms)
Position
Local
Recent
Reflexive
Ziji
Ta−ziji

79
At the second critical region, the head noun of the relative clause, no effects are significant.
At the first spillover region (e.g., ‘provoked’) of the head noun, the main effects of Reflexive and
Position are significant (ps < 0.05). The Reflexive main effect indicates that the processing of the head
noun is more costly in sentences with ziji than with ta-ziji, but this is due to long reading times for the
ziji/recent condition. I will return to this in the Discussion. Importantly, the main effect of Position
means that the animacy of the structurally local head noun modulates the processing of the reflexive-
head-noun dependency for both ziji and ta-ziji. This further suggests that Chinese speakers attempted
to bind the reflexive to the syntactically local antecedent. The Reflexive x Position interaction is not
significant in log-transformed analysis but significant in raw RT analysis (β = -54.22, SE = 24.87, t = -2.18,
p = 0.03). The interaction could mean that the syntactic locality effects exhibited by ziji and ta-ziji are
different, although this should be treated with caution (see more in the Discussion).
At the second spillover region (e.g., ‘everyone’s’), no effect is significant. The final region shows a
main effect of Position (p < 0.005), similar to the pattern observed at the first spillover region of the
head noun, indicating a syntactic locality effect.

2.5.3.5 Discussion of Experiment 2b
Using the animacy mismatch effect as a diagnostic tool to detect binding between the reflexive and a
potential antecedent, Experiment 2b aims to tease apart two different factors that can modulate
reflexive resolution, syntactic locality and linear recency. The reading times in Experiment 2b show that
an animacy mismatch between (ta-)ziji on the one hand and the linearly recent and syntactically local
antecedents on the other leads to processing difficulty, thus providing empirical evidence that linear

80
recency and syntactic locality both guide antecedent retrieval. These results are broadly consistent with
the predictions of the standard cue-based retrieval model (e.g., Lewis & Vasisth, 2005; Jäger et al., 2015;
Patil et al., 2016).
However, it seems that a more fine-grained approach to the standard retrieval model is necessary.
We have seen that the processing patterns of ziji and ta-ziji are different at the reflexive region and at
the post-head-noun region. At the reflexive, participants showed immediate attempt at binding ziji to
the linearly recent antecedent due to an animacy mismatch effect but showed hesitance at binding ta-
ziji due to the lack of an animacy mismatch effect and in fact a trending animacy match effect.
To make sense of this contrasting pattern, we can hypothesize that ta-ziji is more constrained by
syntactic locality. Assuming null effect at the reflexive region, one interpretation is that, at the early
stages, linearly recent antecedents are not considered in the processing of ta-ziji. Only during the later
processing stages, as in the spillover regions, is the recent antecedent considered. If we further take
seriously the marginal animacy match effect at the ta-ziji region, we can have another interpretation:
when a linearly recent, animate antecedent is available, the parser is conflicted between two parses, the
animacy-based parse favoring recent binding and the syntax-based parse favoring local binding. This
tension leads to reading slowdowns. But crucially, regardless of which view we take, the different
processing patterns at the reflexive region suggests that ziji and ta-ziji are differentially sensitive to the
locality constraint.

81

82
Similarly, at the second critical region, the head noun region, the processing of dependency relation
between the reflexive and the head noun seems different for ziji and ta-ziji. Although the animacy of
the head noun impacts real-time processing, which is evidence for syntactic locality effect, the size of
the animacy mismatch effect differs. In combination, the main effect of Reflexive and the significant
Reflexive x Position interaction (in raw RT analysis) suggest that the animacy mismatch effect is stronger
for ziji relative to ta-ziji. On the surface, this could mean that the syntactic locality effect of ziji is stronger
compared to ta-ziji. But this runs contradictory to the conclusion reached earlier in the preceding
paragraphs. Here, I offer an alternative explanation. I posit that the animacy mismatch effect at the
head noun could be related to the strength of recency-based dependency established in the initial
parsing which can interfere with the reanalysis of a new locality-based dependency.
For ziji, when the non-local, recent antecedent is animate (e.g., ‘journalist … ziji … radio station’),
the parser is strongly committed to the recent dependency. In the same condition, when an inanimate
local antecedent appears later, it presumably costs the parser a lot of cognitive resources to reassess the
local dependency between ziji and the (inanimate) head noun. In contrast, when the recent antecedent
is inanimate (e.g., ‘radio station … ziji … journalist’), the parser is less strongly committed to the recent
dependency due to animacy mismatch, which makes the subsequent appearance of an animate local
antecedent less surprising. This contrast in cognitive costs at the head noun is reflected as a large
animacy mismatch effect for ziji. The difference between ziji and ta-ziji is that the syntactically local
dependency for the more locally constrained ta-ziji is more expected due to it being more local
compared to ziji, hence the weaker animacy mismatch effect at the head noun region.

83
Overall, the overarching finding of Experiment 2b is that both linear recency and syntactic locality
guide antecedent retrieval. In addition, the results are compatible with the standard cue-based retrieval
model which predicts early accessibility of (non-syntactic) animacy information. Finally, a form-
specific approach (e.g., Kaiser & Runner, 2008; Kaiser & Trueswell, 2008; Kaiser et al., 2009) to the cue-
based retrieval model seems necessary to account for the different processing profiles of ziji and ta-ziji.

2.6 General discussion
In this section, I will first summarize the main findings of this chapter. I will next discuss the
implications for the sentence processing models and linguistic theories.

2.6.1 Summary of results
There are two main findings from this study. First and foremost, the offline acceptability judgment and
the online self-paced reading experiments show that syntactic locality plays a distinct role from linear
recency. In the acceptability judgment Experiment 2a, the participants preferred sentences with
syntactically local animate antecedents more than sentences with linearly recent animate antecedents.
This suggests that syntactic locality should not be conflated with linear recency. The distinct influence
of syntactic locality and linear recency is more clearly seen in the self-paced reading Experiment 2b
where we see animacy mismatch effects at both reflexive and the head noun regions.
Second, the experiments in this study did not show strong evidence that the locality bias strengths
of ziji and ta-ziji are different. However, in Experiment 2b, there is evidence at the early processing
stages that ta-ziji is not immediately bound to the linearly recent antecedent. This seems to suggest that

84
ta-ziji may be more sensitive to the syntactic locality constraint than ziji is. This finding makes sense
because probabilistically and theoretically ta-ziji is less likely to take non-local antecedents than ziji
(e.g., Pica, 1986; Pan, 1998). In fact, using similar target sentences, a recent study by Lyu and Kaiser (2021)
who asked participants to select between ziji and ta-ziji shows that Chinese speakers chose more ta-ziji
in local contexts and more ziji in non-local contexts, which constitutes complementary evidence to the
present study from a sentence production point of view. In short, this study has not replicated some
prior findings where ziji shows stronger tendency for local binding than ta-ziji. As mentioned previously,
this could be related to some other factors which have not been carefully controlled for, such as c-
command relation and syntactic position of the reflexive. Future work should look more into these
additional factors.

2.6.2 Implications for sentence processing and theoretical linguistics
The present study has broader implications for sentence processing models and linguistic theories as
well. Let us first consider implications to the two antecedent retrieval models. The results of this study,
Experiment 2b in particular, are consistent with the predictions of the standard cue-based retrieval
model which does not assign any privileged status to syntactic constraints. This is because the animacy
of the syntactically inaccessible antecedent outside the local domain has been found to modulate the
processing patterns of ziji right after the reflexive has been encountered. However, the structure-based
retrieval model is not entirely refuted either due to the early Reflexive x Position interaction at the
reflexive region with ta-ziji. It is possible that, for the processing of ta-ziji, the animacy of its linearly

85
recent antecedent is indeed initially inaccessible; only later is animacy used to guide antecedent
retrieval for ta-ziji.
Given the early Reflexive x Position interaction, it seems that the relevant constraints involved do
not have identical weights to the processing of these two reflexive forms. Therefore, it may be useful to
think of real-time reflexive resolution as form-specific. No two reflexive forms assign exactly the same
weights to the same set of constraints, including syntactic locality. This view is not only theoretically
plausible in studies on synonyms (e.g., Lindsay & Aronoff, 2013; Aronoff, 2019) but has received
empirical support in prior experimental studies (Kaiser & Trueswell, 2008; Kaiser et al., 2009). The
findings of Experiment 2b can be viewed as supporting evidence for a form-specific approach to
antecedent retrieval.
Now let us consider the theoretical linguistic implication. In this chapter, I focus on the fact that
despite the possibility for long-distance binding, a syntactic locality effect is nonetheless observed
during the processing of ziji. The common assumption is that by default ziji is a syntactic reflexive.
However, this assumption leaves aside the following question: when is ziji used as a syntactic anaphor
and when is it used as an exempt anaphor? In this study, only two items have perceiver verbs while the
rest (22 items) have source verbs (e.g., ‘say’, ‘state’, ‘disclose’, ‘confess’, ‘claim’) and belief verbs (‘think’)
which make the matrix subject attitude holders. According to research on exempt anaphors, attitude
holders and empathy loci license exempt reading of reflexives (e.g., Charnavel & Zlogar, 2015; Charnavel,
2020, 2021). Thus, we could expect that the availability of an attitude holder can cue the participants to
have an exempt reading of ziji.

86
But this is contrary to our finding in Experiment 2 where Chinese speakers considered both matrix
subject – the attitude holder – and the head noun of the relative clause – the local antecedent. How do
we account for this? I argue that the recency-based binding is best conceptualized as a processing or
the “third factor” (Chomsky, 2005) phenomenon due to how memory works in language processing.
The syntactic locality effect, on the other hand, shows Chinese speakers treat ziji as a syntactic reflexive
because otherwise, syntactic locality should not play any significant role. If so, the finding of this study
is that despite the presence of an attitude holder, ziji is still used as a syntactic reflexive. In fact, this
finding is consistent with Pan and Wang’s view (e.g., Pan, 2001; Wang & Pan, 2014, 2105a,b), namely, ziji
is primarily empathic. Thus, we may expect ziji to be bound by empathy loci and not so easily by attitude
holders. In other words, attitude holders on their own are not sufficient to trigger exempt readings of
ziji. In Chapter 3, I will examine whether the presence of an explicit empathy locus triggers the exempt
reading of ziji.

87
Chapter 3: Discourse topic prominence and reflexive resolution

3.1 Introduction
The results of Experiment 2 in Chapter 2 show that the constraint of syntactic locality is separable from
linear recency. With that in mind, one of the main goals of this chapter is to assess whether syntactic
locality is prioritized relative to discourse-level constraints in the real-time parsing of ziji and ta-ziji.
The discourse-level factor this chapter focuses on is discourse topic prominence. However, note that
the factor of discourse topic prominence may have different consequences for the real-time resolution
of the ziji and ta-ziji. As mentioned in Chapter 1 (and reviewed below), discourse topicality may impact
the resolution of ziji through perspective-taking and the resolution of ta-ziji through a non-perspective-
related prominence constraint. Thus, we can evaluate the standard cue-based retrieval model (e.g.,
Lewis & Vasishth, 2005; Van Dyke & McElree, 2006; Chen et al., 2012; Patil et al., 2016) and the structure-
based retrieval models (e.g., Van Dyke, 2007; Van Dyke & McElree, 2011; Dillon et al., 2013; Parker &
Philips, 2017) more thoroughly by examining the processing patterns of two reflexive forms with
somewhat different discourse properties.
A secondary goal of this chapter is to examine the interaction of verb semantics and the locality
constraint, especially for the complex reflexive ta-ziji. Two types of verbs are tested in this study, self-
directed verbs and other-directed verbs. Self-directed verbs (or ‘introverted’ verbs) describe actions that
“one generally performs upon one’s self” and other-directed verbs (or ‘extroverted verbs’) denote
actions that one “usually performs towards others” (Haiman, 1983: 803; also see e.g. König & Siemund,
2000; Gast, 2006). In the literature on Chinese reflexives, it has been observed that the semantics of the

88
verb does not seem to contribute substantially to the interpretation of ta-ziji although it does for ziji
(Lu, 2011). However, in Lu’s (2011) work, the verb bias manipulation is context-dependent (see more
below). This study aims to complement prior work by investigating the role of verb directedness (which
inherently biases (non-)local antecedents) in reflexive interpretation. Furthermore, as explained below,
verb directedness is additionally used as a diagnostic to assess participants’ binding preferences in
incremental comprehension. Reading times allow us to examine linguistic questions which we cannot
easily answer with antecedent judgment data alone and can simultaneously shed light on the sentence
processing models.
In the following sections, I will first discuss from a theoretical linguistic angle the difference
between ziji and ta-ziji in terms of their discourse-pragmatic properties. Next, I review how verb
directedness could impact the interpretation of these reflexives. Finally, I will focus on how discourse
topic prominence can be utilized to assess the two antecedent retrieval models, for which two sets of
experiments will be reported.
3.2 Discourse topic prominence and Mandarin reflexives
The two reflexives in Mandarin, ziji and ta-ziji, show different discourse-pragmatic properties. For
instance, using Sell’s (1987) terminology, Huang and Liu (2001) showed that the long-distance use of ziji
is allowed when the antecedent is a source (intentional agent of communication), self (one whose
mental state or attitude the proposition describes), or pivot (one with respect to whose space-time
location the proposition is evaluated). (Also see Charnavel, 2020: 694 for a revised taxonomy of
discourse roles that can act as long-distance binders.) See (3.1) for an example, adapted from Huang and
Liu (2001: 156):

89

(3.1) a. Yuehan i shuo Bier j xihuan ziji i/j.
John say Bill like self
‘John i said that Bill j liked self i/j.’
b. Bier j hui bangzhu ziji i/j rang Yuehan i hen gaoxing.
Bill would help self make John very happy
‘That Bill j would help self i/j made John i very happy.’
c. Dang Bier j zai jiaoji de deng ziji i/j de shihou Yuehan i hai mei chu men.
when Bill at anxious DE wait self DE time John still not exit door
‘When Bill j was anxiously waiting for self i/j, John i had not left the house yet.’

Huang and Liu argue that source, self, and pivot roles can be perspective centers (or empathy loci)
which ziji anchors to. In other words, if the speaker puts him/herself in the shoes of the non-local
antecedent or assumes the perspective of the empathy locus, long-distance binding is licensed. In
comparison, ta-ziji does not seem to show this property because replacing ziji with ta-ziji lowers the
acceptability of the sentence significantly, shown in (3.2):

(3.2) a. Yuehan i shuo Bier j xihuan ta-ziji ??i/j.
John say Bill like self
‘John i said that Bill j liked he-self ??i/j.’

90
b. Bier j hui bangzhu ta-ziji *i/j rang Yuehan i hen gaoxing.
Bill would help he-self make John very happy
’That Bill j would help he-self *i/j made John i very happy.’
c. Dang Bier j zai jiaoji de deng ta-ziji *i/j de shihou Yuehan i hai mei chu men.
when Bill at anxious DE wait he-self DE time John still not exit door
‘When Bill j was anxiously waiting for he-sellf *i/j, John i had not left the house yet.’

The contrast in acceptability between (3.1) and (3.2) highlights a key difference between ziji and ta-
ziji: ziji can be empathic while ta-ziji cannot, at least in these stand-alone sentences. This difference
could have different consequences for the interpretation of ziji and ta-ziji in a context where the non-
local referent is the discourse topic. Below, I first describe how discourse topic prominence can
influence the interpretation of ziji before turning to ta-ziji.
According to the Topic Empathy Hierarchy (Kuno, 1987), given an event involving two participants
A and B such that A is a discourse topic, it is easier to take the perspective of A than B. Although Kuno’s
work is not based on Chinese ziji, it makes clear predictions for ziji. Suppose we have two characters in
a discourse as in (3.3), Xiaoming and Professor Wang. The first sentence introduces Xiaoming in the
subject position and describes an event about him. He can be easily construed as the discourse topic.
The Topic Empathy Hierarchy predicts that readers should prefer the perspective of Xiaoming. Thus,
perspective-sensitive ziji should tend to be interpreted as referring to the perspective center Xiaoming
(rather than Professor Wang). In this work, building on Kuno’s proposal of the link between topics and

91
empathy/perspective, I assume that any potential topic prominence effects displayed by ziji are related
to its perspective-taking property.

(3.3) Xiaoming shi banji li de youxiu xuesheng. Ke shang, ta tingshuo Wang jiaoshou
Xiaoming be class in DE excellent student lecture on he hear Prof. Wang
zhaodao-le ziji de xueshu lunwen.
find-ASP self DE academic paper
‘Xiaoming is an excellent student in the class. During the lecture, he heard that Prof. Wang
had found self’s academic paper.’

What about ta-ziji which does not allow an empathic reading in (3.2)? Will its interpretation be
influenced by the discourse topic status of the matrix subject? Previous studies have not investigated
this question in Chinese. However, Pan (1998) proposed that, as a syntactic anaphor, ta-ziji is
susceptible to a set of prominence constraints which not only include structural prominence but also
semantics-based animacy prominence. As mentioned in Chapter 1, in sentences like ‘that book hurt ta-
ziji’, non-local binding is allowed when the local referent is inanimate and unfit to be the binder.
However, in light of a large amount of work showing that grammatical role and animacy are only two
of the many factors – including discourse-level factors – that influence the prominence of referents (e.g.,
Ariel, 1990; Foraker & McElree, 2007; Kaiser, 2011; Kim & Yun, 2009, 2020), it seems reasonable to
hypothesize that ta-ziji may be sensitive to topic prominence as well. Thus, in this study, I will test
whether this prominence constraint for ta-ziji can be extended to the discourse level.

92

3.3 Verb semantics and Mandarin reflexives
So far, I have sidestepped verb semantics as a potential factor in the interpretation of reflexives. Ex.(3.3)
uses a neutral verb zhaodao (‘find’) in the second sentence. But if this verb is changed to a self-directed
verb such as fabiao (‘publish’), the reading of the reflexive becomes more ambiguous. Now, it is more
plausible that Prof. Wang has published his own paper as opposed to Xiaoming’s paper. Thus, when ziji
or ta-ziji occurs with a self-directed verb like fabiao, verb semantics, discourse topic prominence, and –
especially for ta-ziji – syntactic locality do not point to the same referent. Faced with this dilemma, how
do Chinese speakers interpret ziji and ta-ziji? In other words, how are different kinds of linguistic
information used and integrated in reflexive resolution?
Prior work by Lu (2011) on Taiwan Mandarin Chinese suggests that verb semantics does not play a
major role in the interpretation of ta-ziji. Lu tested sentences with triple-clause structures where ziji or
ta-ziji occupies the subject position of the most embedded third clause. As shown in (3.4), the verb after
(ta-)ziji biases either towards the “local” antecedent (‘receive’ biasing ‘public’) or towards the non-local
antecedent (‘deliver’ biasing ‘mailman’). It is worth emphasizing that the verb semantics manipulated
by Lu (2011) depends crucially on the context of the sentence: the mailman typically delivers the mail,
and the public receives the mail. If the agent and the recipient of the delivery action change to some
other referents (e.g., ‘doctor’ and ‘nurse’), the semantic bias could disappear, in contrast to verb
directedness where the action is inherently interpreted as directed towards the agent of the action
regardless of its identity.

93
(3.4) Youchai i xuancheng minzhong j zong huaiyi (ta-)ziji i/j meiyou anshi shoudao/toudi
mailman claim public always suspect (he-)self not on-time receive/deliver
youjian.
mail
‘The mailman i claimed that the public j always suspects that (he)-self i/j does not receive public/
deliver mailman the mail on time.’

Given that reflexives in Mandarin tend to be interpreted by default as syntactic reflexives heavily
constrained by Principle A (e.g., Dillon et al., 2014, 2016), we can have the following default prediction
for the interpretation of (ta-)ziji in (3.4). As the locality constraint pushes (ta-)ziji to look for a
structurally closer antecedent but a non-local-biasing verb such as ‘deliver’ pushes for the matrix
subject ‘mailman’, one can expect ambiguity in the interpretation of (ta-)ziji. Interestingly, Lu (2011)
found divergent interpretation patterns for ziji and ta-ziji: when the verb favors the matrix subject,
participants prefer the matrix subject as the antecedent of ziji but still prefer the closer referent as the
antecedent of ta-ziji. Lu’s account of this asymmetry boils down to different degrees of sensitivity to the
locality constraint (ziji is less sensitive to locality than ta-ziji, under this view). Lu’s finding thus suggests
that the locality constraint plays a more prominent role for ta-ziji compared to verb bias. In this study,
I will examine to what extent this is also true for the interaction of verb directedness and the locality
constraint.
In addition to examining the contribution of verb directedness to the global interpretation of
(ta-)ziji, I also use verb directedness as a diagnostic tool to probe Chinese speakers’ real-time

94
interpretative bias in sentences with and without discourse topics as they read sentences incrementally.
Here is how this diagnostic works. If participants prefer local binding, other-directed verbs favoring the
non-local referent as the antecedent should lead to reading slowdowns when the reflexive is
encountered; if participants prefer non-local binding, self-directed verbs biasing the local referent
should lead to slowdowns instead. Thus, by observing whether the self-directed or the other-directed
verbs lead to processing difficulty, we can determine participants’ interpretive bias of the reflexive in
certain contexts. This verb manipulation is similar in spirit with Lu’s design for her self-paced reading
study.

3.4 Antecedent retrieval models
The primary goal of this chapter is to examine the role of discourse topic prominence in the processing
of two Mandarin reflexives ziji and ta-ziji in order to test the standard and structure-based antecedent
retrieval models. The two retrieval cues are syntactic locality and discourse topic prominence. Note that
discourse-level cues, like structural cues, are not encoded lexically or morphologically on the
antecedent, although they are discussed in prior work (e.g., Foraker & McElree, 2007; Cunnings et al.,
2014; Jäger et al., 2015). Thus, it is an open question how syntactic and discoursal cues are encoded in
working memory compared to semantic and morphological cues. However, this should not prevent us
from incorporating these cues in the modeling of antecedent retrieval. In this section, I discuss how
discourse topic prominence can impact the processing of ziji and ta-ziji.
The earliest investigation of discourse topic prominence and real-time reflexive resolution can be
traced to Sturt (2003), as introduced in Chapter 1 (see ex.(1.24)). In Sturt’s reading eye-tracking study, it

95
was found that the discourse topicality status of the matrix subject does not intrude on the early-stage
parsing – as indicated by first pass reading times – of English reflexives.
12
Only during the later
processing stage – as indicated by second pass duration – did (British) English speakers consider the
non-local discourse topic. But even this later-stage effect does not seem to be robust: in a later study,
Cunnings and Sturt (2014) did not find any sign of topic prominence effect in the analysis of late eye-
tracking measures. These findings are noteworthy, especially because the (British) English reflexive can
be bound by an empathy locus in the ‘domain of point-of-view’ (Zribi-Hertz, 1989; Pollard & Xue, 2001).
Thus, prior work on English reflexives suggests the possibility that discourse-level information may not
be immediately accessed and that only syntactic cues (e.g., syntactic locality) are accessible at the early
processing stage, which favors the structure-based retrieval model.
If this parsing mechanism applies universally to reflexive resolution in all languages, we might
expect a similar situation in the processing of ziji and ta-ziji. This means that the discourse-level cues of
the matrix subject in (3.3) should be invisible to a parser only sensitive to structural relations at the
earliest moments of reference resolution. Now, let us consider what discourse-level cues are at play. As
suggested in Section 3.2, discourse topic prominence impacts the interpretation of ziji and ta-ziji in
different ways. For ziji, long-distance binding is mediated through perspective-taking; for ta-ziji, the
interpretation could be influenced by a general prominence constraint. For ease of exposition, I will
notate the discoursal retrieval cue for ziji as [+POV] where POV stands for “point of view” and the

12
Although Sturt (2003) did not explicitly discuss the results in relation to the discourse topicality effect, this
interpretation is supported by the data. In Sturt’s study (also see Cunnings and Sturt, 2014), the non-local subject
is always a discourse topic. The delayed accessibility of the matrix subject seems to show that discourse topic
prominence does not play an early role in real-time reflexive resolution in (British) English.

96
discoursal cue for ta-ziji as [+Top] where “Top” stands for discourse topic. Given these notations, the
antecedent retrieval process for the structure-based retrieval model can be visualized in Figure 3.1.

Figure 3.1: Predictions of the structure-based retrieval model for sentences with discourse topics.

The discourse retrieval cues [+POV] and [+Top] are shaded to indicate that these cues are not used
for retrieval of the antecedent at the early stage. Under a structure-based retrieval mechanism, syntactic
constraints such as locality block the accessibility of the non-local retrieval cues. Thus, non-local
antecedents are supposedly not considered at all. Local binding is preferred.
In contrast, the standard cue-based retrieval model allows for early accessibility of discourse-level
cues. In fact, previous studies on English reflexives showing null discourse topic effect do not disprove
the standard retrieval model, for at least two reasons. First, reflexives in natural languages can vary in
their sensitivities to discourse topic prominence. It is conceivable that the perspective-related reading
of English reflexives is only available in some highly constrained discourse context. Suppose that there
is a scale of discourse topic sensitivity. English reflexives could sit on the lower end of this scale, but this
does not mean that reflexives in Chinese are also located on the lower end. In other words, the null
discourse topic effect in the real-time processing English reflexives could be due to specific linguistic

97
properties of English reflexives. This does not necessarily mean that reflexives in other languages,
regardless of their linguistic properties, show similar insensitivity to discourse-level constraints at the
early processing stage. Second, the previous studies on English reflexives do not falsify the standard
retrieval model because Sturt (2003) and Cunnings and Sturt (2014)’s studies may suffer from a lack of
statistical power, as discussed by Patil et al. (2016).
According to the standard cue-based retrieval model, discourse-level cues will be immediately
recruited for reflexive resolution. For the processing of ziji, I follow Pollard and Xue (2001) in assuming
that an empathic reading can be triggered when language specific discourse/pragmatic constraints (e.g.,
logophoricity, discourse prominence) are satisfied. This means that the parser will search in the
memory workspace for an antecedent with a [+POV] retrieval cue. The local subject referent will not
be favored as it is not a perspective center.
13
Thus, non-local binding preference is predicted. For the
processing of ta-ziji, I similarly assume that discourse topicality triggers the search for a discourse
prominent antecedent, in line with an extended version of Pan’s (1998) proposal. However, depending
on the strength of the prominence constraint, there could be two outcomes: (i) if the Prominence
Hierarchy can be extended to the discourse level – discourse topic prominence outranks syntactic
locality – we should expect non-local binding preference for ta-ziji; (ii) if discourse topic prominence
only impacts but does not determine the interpretation of ta-ziji – discourse topic prominence does not
outrank syntactic locality – we should see no clear binding preferences as probabilities of local binding

13
This does not mean that local subjects will not be considered at all. Since it is in a subject position with a [+Subj] cue, it
may be retrieved sometimes. But overall non-local POV centers should be preferred.

98
([+Local, -Top]) and non-local binding ([-Local, +Top]) should be close. These predictions are
visualized in Figure 3.2:

Figure 3.2: Predictions of the standard cue-based retrieval model for sentences with discourse topics.

3.5 Overview of the aims
I report two sets of experiments below to further our understanding of the syntactic, semantic, and
discourse constraints that are involved in reflexive resolution in Mandarin. Critically, by examining
Chinese speakers’ real-time processing patterns of ziji and ta-ziji, this study sheds light on the two
antecedent retrieval models. There are two main aims for this chapter:
First, I aim to answer basic linguistic questions by using offline judgment experiments to lay the
foundation for the online reading experiments. In addition to showing that ziji is sensitive to discourse-
topicality and verb semantics, I ask from a descriptive perspective (i) whether the prominence
constraint for ta-ziji can be extended to the discourse level and (ii) whether syntactic locality is a
stronger constraint than verb semantics for ta-ziji. The answer to the first question has direct
consequences for the subsequent online experiment on ta-ziji: the online experiment would not be
informative if discourse topicality turns out to not have any influence in even the global offline

99
interpretation of ta-ziji. The second question is an attempt to expand Lu’s (2011) work by testing the
interplay of verb directedness and syntactic locality.
Second, I aim to test from a psycholinguistic perspective how syntactic and discourse-level
information are structured, organized, and integrated in online reflexive resolution. The goal is to
answer (i) whether syntactic cues are prioritized compared to discourse-level cues, (ii) whether the
processing patterns of ziji and ta-ziji are different, and (iii) if so, how to account for the difference.
To accomplish these two aims, I use offline forced-choice judgment and online self-paced reading
tasks. Experiments 3a and 3b test the offline and online processing of ziji, respectively, while
Experiments 4a and 4b test the offline and online processing of ta-ziji, respectively.

3.6 Experiment 3: Processing of ziji in discourse
Experiment 3 is comprised of two studies. Experiment 3a is a forced choice judgment task where
participants are asked to select which referent they think the reflexive refers to. This offline judgment
probes participants’ global interpretation of the reflexive and tests whether the linguistic variables have
any effect at all. Experiment 3b is an online self-paced reading experiment where participants read
sentences incrementally (region by region). This experiment can show the impact of syntactic and
discourse factors at different processing stages.

3.6.1 Materials and design
Experiment 3a and 3b share the same set of stimuli. Two factors, context (neutral/biased) and verb
directedness (self-/other-directed), were crossed in a factorial design. Each target item is composed of

100
two sentences, a context sentence and a critical sentence. See (3.5) for a target example. The subscripts
indicate the presentation region of the critical sentence in the self-paced reading experiment.

(3.5) a. Neutral context/Self-directed verb
Jintian shi shang wenxue ke de rizi. Ke shang, 1 Xiaoming 2 tingshuo 3 Wang jiaoshou 4
Today be have literature class DE day lecture on Xiaoming hear Wang just
ganggang 5 fabiao-le 6 ziji 7 de 8 xueshu 9 lunwen. 10
just publish-asp self DE academic paper
‘Today is the day for the literature class. During the lecture, Xiaoming heard that Professor
Wang had just published self’s academic paper.’
b. Neutral context/Other-directed verb
Jintian shi shang wenxue ke de rizi. Ke shang, 1 Xiaoming 2 tingshuo 3 Wang jiaoshou 4
Today be have literature class DE day lecture on Xiaoming hear Wang professor
ganggang 5 pigai-le 6 ziji 7 de 8 xueshu 9 lunwen. 10
just grade-asp self DE academic paper
‘Today is the day for the literature class. During the lecture, Xiaoming heard that Professor
Wang had just graded self’s academic paper.’
c. Biased context/Self-directed verb
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 tingshuo 3 Wang jiaoshou 4
Xiaoming be class in DE excellent student lecture on he hear Wang professor
ganggang 5 fabiao-le 6 ziji 7 de 8 xueshu 9 lunwen. 10

101
just publish-asp self DE academic paper
‘Xiaoming is an excellent student in the class. During the lecture, he heard that Professor
Wang had just published self’s academic paper.’
d. Biased context/Other-directed verb
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 tingshuo 3 Wang jiaoshou 4
Xiaoming be class in DE excellent student lecture on he hear Wang professor
ganggang 5 pigai-le 6 ziji 7 de 8 xueshu 9 lunwen. 10
just publish-asp self DE academic paper
‘Xiaoming is an excellent student in the class. During the lecture, he heard that Professor
Wang had just graded self’s academic paper.’

The term ‘neutral context’ refers to a context that lacks a discourse topic character. The term ‘biased
context’ refers to a context with a discourse topic character (e.g., Xiaoming) which is the (non-local)
matrix subject of the critical sentence. In biased contexts, the matrix subject is shown in the pronominal
form ta (‘s/he’) to avoid the Repeated Name Penalty (e.g., Gordon et al., 1993; Garrod et al., 1994; Almor,
1999). The critical sentence contains a self-directed verb (e.g., ‘publish’) or an other-directed verb (e.g.,
‘grade’) in the embedded clause. The verbs were selected based on native speaker judgment, but the
verbs’ directedness was confirmed in Experiment 3a, reported below.
Half of the local and non-local characters were male, and half were female. Ten different verbs
(‘hear’, ‘notice’, ‘say’, ‘learn’, etc.) were used as the matrix verb in the critical sentence. Table 3.1 lists the

102
numbers and proportions (in parentheses) of speech verbs, belief verbs, and perceiver verbs in
Experiments 3 and 4.

Table 3.1: Number and proportions of speech, belief, and perception verbs in Experiments 3 and 4.
Speech Belief Perception
Number (percentage) 7 (35%) 2 (10%) 11 (55%)

Twenty sets of target items were created and distributed into 4 lists following a Latin square design.
Thus, each participant saw a particular item once, and saw 5 targets in each of the 4 conditions over the
course of the experiment. In the self-paced reading experiment, the critical region is the reflexive ziji,
preceded by the biased verb and followed by three spillover regions: DE, a modifier, and an NP. Twenty
filler sentences were interleaved with target sentences such that each target was followed or preceded
by a filler. The fillers all contained non-reflexive pronouns.

3.6.2 Experiment 3a: Forced choice judgment
3.6.2.1 Participants
Forty-five adult Chinese native speakers born and raised in mainland China participated over the
internet. All had normal or corrected-to-normal vision.

3.6.2.2 Procedure
The experiment was run on Qualtrics (Qualtrics, Provo, UT). On target trials, the questions probed the
referent of ziji. For example, the question for (3.5) is ‘whose paper has been published/graded?’. On each

103
trial, the context sentence and the critical sentence were displayed separately. After each context
sentence, participants clicked to see the critical sentence. This was done to match the presentation set-
up of the self-paced reading Experiment 3b. After reading the critical sentence, participants answered
the comprehension question (e.g., ‘who wrote an academic paper?’) displayed on the same screen by
selecting one of two answer choices. The order of antecedent choices (e.g. ‘Xiaoming’/ ‘Professor Wang’)
was counterbalanced.

3.6.2.3 Predictions
If the verb directedness manipulation is effective, participants should prefer local antecedents with self-
directed verbs (e.g., ‘publish’) and non-local antecedents with other-directed verbs (e.g., ‘grade’).
Furthermore, if discourse topic prominence influences the interpretation of ziji, a main effect of context
is expected: when the non-local antecedent is the discourse topic in a biased context, non-local choice
preference should be stronger relative to the neutral context.

3.6.2.4 Data analysis
An analysis of the participants’ answers to comprehension questions asked on filler trials suggests that
they all paid attention to the experiment with comprehension accuracy over 75%. Statistical analyses
were conducted in R (R Core Team, 2018). For the two factors context and verb bias, two contrasts were
fitted (biased context: 0.5, neutral context: -0.5; self-directed verb: 0.5, other-directed verb: -0.5).
Participants’ antecedent choices were analyzed using mixed-effect logistic models with the glmer
function, implemented by the R package lme4. Statistical models were first fit with random intercepts

104
and random slopes. When the model failed to converge, I simplified the model following the procedures
in Bates et al. (2015). A simpler model was preferred to a more complex model if model comparison did
not show significant differences between the two. An alpha value of 0.2 (rather than 0.05) is chosen for
model comparisons following the suggestion of Matuschek et al. (2017).

3.6.2.5 Results
Participants’ preferences for local referents as antecedents of ziji are displayed in Figure 3.3. The left
panel A shows the distribution of local choices, and the right panel B shows the mean proportions. See
Table 1 for statistical analysis.

Figure 3.3: Distribution of mean proportions of local readings of ziji across individual participants
(Panel A) and mean proportions of local readings (Panel B) at the group level in Experiment 3a.

As shown in Figure 1, when the verb is self-directed, local referents are preferred as antecedents (63%
in the biased context and 82% in the neutral context); when the verb is other-directed, non-local
referents are preferred (85% in the biased context and 77% in the neutral context). Indeed, the mixed-
effect logistic model reveals a main effect of verb bias (p < 0.001). Contexts modulate participants’
0%
25%
50%
75%
100%
Self−directed Other−directed
Local coreference
A
15 %
63 %
23 %
82 %
0%
20%
40%
60%
80%
Self−directed Other−directed
Local coreference
Context
Biased
Neutral
B

105
choices as well, as critical sentences preceded by biased contexts made participants less likely to select
local referents as antecedents (p < 0.001). The context x verb bias interaction is not significant.

Table 3.2: Summary of statistics for Experiment 3a (*: < 0.05).
β SE t-value p-value
Context -0.92 0.18 -5.01 < 0.001*
Verb 3.01 0.21 14.67 < 0.001*
Context x verb 0.35 0.36 0.97 0.33

Overall, the results of Experiment 3a show that the verb directedness manipulation is effective: in
neutral contexts, the proportions of local choices under self-directed verbs (82%) and the proportions
of non-local choices under other-directed verbs (77%) are similar. In other words, both verbs show a
preference for the referents favored by the verbs’ semantic directionality bias. In addition, discourse
topicality influences participants’ antecedent choices as the discourse topical status of the non-local
subject lowers the proportions of local choices. Nevertheless, global interpretation of ziji is primarily
determined by verb directedness. Next, I turn to Experiment 3b to examine the real-time incremental
processing of ziji.

3.6.3 Experiment 3b: Self-paced reading
3.6.3.1 Participants
Seventy-nine adult Chinese native speakers with normal or corrected-to-normal vision participated
over the internet. The experiment was run on Ibex Farm (Drummond, 2013). No participant had
participated in the offline Experiment 3a.

106
3.6.3.2 Procedure
The context sentence and the critical sentence were displayed separately. Participants first saw the
context sentence in its entirety. They then pressed the spacebar to see the critical sentence on a new
screen presented region by region, as indicated by the subscripts in (3.5). Each key press revealed the
next region and at the same time masked the previous region with a dash. After participants finished
reading the critical sentence, they pressed the spacebar again to see a comprehension question with
two answer choices. Twelve questions out of the 20 questions for the target items probed the
antecedent of ziji, and the answer choices consisted of the two candidate antecedents.
14
The remaining
8 questions asked about the context sentence with answer choices consisting of a factually correct and
a factually incorrect choice. Participants’ responses to these 8 questions and the 20 questions in filler
trials were used to calculate comprehension accuracy. The order of answer choices was randomized.

3.6.3.3 Predictions
As already discussed in Section 3.6.1 above, in neutral contexts where no character is privileged as the
discourse topic, both the standard and structure-based retrieval models predict local binding since no
discourse-level cues are involved. Since verb directedness is used as a diagnostic to test binding
preferences, we should expect other-directed verbs to lead to reading slowdowns at the reflexive region
in neutral context conditions.

14
These questions were included to ensure that participants paid attention. Prior studies on reference resolution
show that participants might engage in shallow parsing if comprehension questions do not involve interpretation
of pronouns (e.g., Stewart et al., 2007; Creemers & Meyer, 2022).

107
In biased contexts, however, the two processing models make different predictions. According to
the Topic Empathy Hierarchy (Kuno, 1987), participants should empathize with or take the perspective
of the discourse topic. The standard cue-based retrieval model thus predicts that the language parsers
will make quick use of the [+POV] cue (in addition to [+Subj]) for antecedent retrieval. As the matrix
subject constitutes a full match of retrieval cues, a preference for this non-local antecedent is expected.
This means that self-directed verbs should cause reading slowdowns. In contrast, the structure-based
retrieval model predicts a local binding preference because, at early processing stages, discourse-level
cues are invisible to the parser. Only during the later processing stages can discourse-level cues be
utilized.
In addition, a global main effect of context is also expected due to reasons unrelated to the aims of
this study. As mentioned, participants should prefer the perspective of the discourse topic after reading
the context sentence. Thus, in reading the ensuing critical sentence with the discourse topic as the
subject, we expect participants to maintain that perspective (e.g., the perspective of Xiaoming). But in
neutral contexts, participants first take their own perspective when reading the context sentence. Upon
encountering the matrix subject of the critical sentence, they then shift their perspective to that of a
new character, which has been shown to carry a cognitive cost (e.g., MacWhinney & Pleh, 1988;
MacWhinney, 2001). This means that critical sentences in biased contexts should take less time to read
than in neutral contexts.

3.6.3.4 Data analysis

108
Before data analysis, it was decided that any participant with mean accuracy below 75% on the
comprehension questions would be removed, which resulted in the removal of 6 participants. The
remaining 73 participants had a mean accuracy of 94%. Reading times (RTs) shorter than 100ms or
longer than 3000ms were trimmed. RTs above 2.5 standard deviations of the mean by region and
condition were also removed. This affected 2.71% of the original data. Statistical analyses were
conducted over both log-transformed and raw RTs using mixed-effect linear models implemented by
the lme4 package in R. Below, I report statistical results for the log-transformed data when the log-
transformed and raw RT analyses are consistent (see e.g., Parker, 2019 for a similar practice). In this
study, the two types of analyses mostly yielded consistent results. When the results of the log-
transformed and raw RT analyses diverge, I report both.

3.6.3.5 Results
The mean RTs for each region in the target sentence are in Figure 3.4. The critical region is Region 7,
ziji. A visual inspection suggests two noteworthy patterns: First, words in the biased context are
processed faster compared to the neutral context. Second, the verb directedness effects already appear
at the verb, prior to the reflexive ziji. At the reflexive itself, RTs are faster in biased contexts than in
neutral contexts, but there appears to be no effect of verb directedness.
These observations are confirmed by statistical analysis (Table 3.3). First, the main effect of context
is significant across all regions (regions 2 through 10; ps < 0.005), showing that critical sentences
preceded by biased contexts were read faster overall. This is not relevant to the main purpose of this
study but is expected under a perspective-shift account.

109

110

Figure 3.4: Mean RTs (ms) across all regions in the critical sentence in Experiment 3b.

At the verb region, in addition to a context main effect (p < 0.001), a significant context x verb bias
interaction (p < 0.001) was identified. This suggests that verb directedness has different effects in
neutral and biased contexts. Pairwise comparisons revealed that in neutral contexts, other-directed
verbs lead to RT slowdowns relative to self-directed verbs (β = -0.08, SE = 0.02, t = -3.08, p = 0.002). As
will be explained below, given similar pre-critical effect in prior studies on Chinese reflexives (Chen et
al., 2012; Jäger et al., 2015; Chang et al., 2020), I take this early effect at the verb region as indication of
predictive processing; participants anticipated the upcoming reflexive and thus showed verb
directedness effect related to reflexive processing before the arrival of the reflexive.
15
At any rate, this
reading slowdown related to other-directedness – indicating a locality bias – aligns well with the
prediction of both the standard and structure-based retrieval models. But crucially, in the biased

15
Session analysis on the first and second half of the trials was conducted. The context x verb bias interaction
consistently occurred at the verb region.
verb
ziji
400
500
600
700
800
In lecture, Xiaoming/he heard Prof.Wang just published/graded ziji DE academic paper.
RT (ms)
Context
Biased
Neutral
Verb
Self−directed
Other−directed

111
context, self-directed verbs caused longer RTs than other-directed verbs (β = 0.0, SE = 0.02, t = 2.13, p =
0.03), suggesting non-local binding preference. This fits the prediction of the standard cue-based
retrieval model.
At the critical region ziji and the three spillover regions, only main effects of context in the log-
transformed RT analysis were found (ps < 0.001). However, analysis of raw RTs for the second spillover
region (‘management’) reveals a main effect of verb bias (β = -21.27, SE = 8.08, t = -2.63, p < 0.01) and a
marginal context x verb bias interaction (β = -28.25, SE = 16.08, t = -1.76, p = 0.08). Pairwise comparisons
indicate that while there is no verb bias effect in biased contexts (-6.11, SE = 7.56, t = -0.81, p = 0.42),
other-directed verbs lead to slowdowns in neutral contexts (β = -36.69, SE = 14.04, t = -2.61, p < 0.01),
pointing to a locality bias. No other effects at the spillover regions are significant.

3.6.4 Discussion of Experiments 3a and 3b
Experiments 3a and 3b investigate how the locality constraint, verb semantics and discourse topic
prominence guide the offline and online interpretation of ziji. The antecedent judgement data show
that the interpretation of ziji is strongly influenced by verb directedness and that discourse topicality
also increases the proportions of non-local choices. These results indicate that the self- and other-
directed verbs used in this study are effective in eliciting local and non-local readings, although the
interpretation patterns are modulated by discourse topic prominence.
To assess how ziji is interpreted in real-time comprehension, participants’ incremental reading
times in the self-paced reading experiment were analyzed for each region. The standard cue-based
retrieval model predicts that participants should immediately locate the discourse topic – notated with

112
[+POV] due to its tendency to be empathized with (Kuno, 1987) – as the antecedent at the earliest
possible moments. The structure-based retrieval model predicts that only syntactic cues such as locality
information will be accessed by the parser and thus only local antecedents should be considered at the
early stages. The reading time data are in line with the standard cue-based retrieval model as a
significant context x verb bias interaction occurred as early as the pre-critical verb region; crucially, self-
directed verbs lead to reading slowdowns in biased contexts, suggesting non-local binding preferences.
The POV cue is thus immediately accessible to the parser and guides antecedent retrieval for ziji.
The finding that the context x verb bias interaction already emerged at the pre-critical verb region
may seem surprising. However, it is in line with the findings of Chen et al. (2012), as discussed by Jäger
et al. (2015). In Chen et al.’s self-paced reading study, an animate antecedent in a structurally
inaccessible position (i.e., non-local, non-c-commanding) interfere with the search for a more distant
grammatically accessible antecedent. However, the interference effect was observed even before the
onset of the reflexive ziji. The same early effect at the verb region was also found in an eye-tracking
study using a similar design to Chen et al. by Jäger et al. (2015: Exp.1). Jäger et al. (2015) argued that the
pre-critical animacy effect is either a preview effect – which does not apply to self-paced reading where
preview is impossible – or due to “the verb’s preference for an animate subject” which does not reflect
the antecedent retrieval process. Jäger et al. believed that the second explanation is a more likely reason
for the pre-critical effect. However, this explanation cannot account for the pre-critical verb bias effect
in Experiment 1b of this study because all the antecedents in the experiments are animate in this study.
Instead, I suggest that this effect is due to predictive processing which is an integral part of sentence
processing (Kuperberg & Jager, 2016). One reason for this predictive reading behavior may be related

113
to the predictability of the target sentences which always follow or precede the filler sentences. In fact,
as will be reported in Chapter 5 of this dissertation, this verb bias effect can appear at either the verb
region or at the reflexive, which seems to depend on the time available to the participants for
prediction-making. Additionally, the presence of an adverbial before the biased verb (see ex.(3.5)) may
give participants more time to engage in predictive processing. In any case, the verb bias effect in
Experiment 3b occurred earlier than predicted presumably due to predictive processing, but this does
not detract from the main findings of this study, namely that discourse-level cues [+POV] can be
immediately accessed in the early stages of processing.

3.7 Experiment 4: Processing of ta-ziji in discourse
So far, I have focused on the interpretation and the processing of ziji. In this section, I turn to ta-ziji and
report a similar set of experiments. Experiment 4a is a forced choice judgment experiment, and
Experiment 4b is a self-paced reading experiment. The offline experiment explores the interpretive
patterns of ta-ziji in contexts where verb semantics and discourse topicality interact and conflict with
the locality constraint. Moreover, this section will look into the potentially different sensitivities to
discourse topic prominence given that ta-ziji and ziji differ with regard to empathy readings
(perspective-taking), as shown in (3.2). The online experiment examines the processing patterns of ta-
ziji to see whether this reflexive form too is sensitive to discourse topicality at the early comprehension
stages.

3.7.1 Materials and design

114
The materials used in Experiment 4 are identical to Experiment 3, except that ziji was now replaced
by ta-ziji.

3.7.2 Experiment 4a: Forced-choice judgment
3.7.2.1 Participants
Forty-six adult Chinese native speakers participated via Qualtrics (Qualtrics, Provo, UT). All had normal
or corrected-to-normal vision. None had participated in Experiment 3.

3.7.2.2 Procedure
The procedure is the same as in Experiment 3a.

3.7.2.3 Predictions
If the interpretation of ta-ziji is only governed by the locality constraint, verb directedness and the
discourse topical status of the matrix subject are not expected to have any effect. Alternatively, it is
conceivable that ta-ziji, too, is susceptible to the verb semantics information and discourse topicality.
However, ta-ziji could yield different results than ziji for at least two reasons.
First, as regards the verb semantics effect, Lu (2011) found that ta-ziji shows a locality bias in an
offline judgment task even when the verb biases towards the matrix subject. As mentioned earlier in
Section 3.3, the effectiveness of the verb bias is heavily dependent on the sentence context. It remains
to be seen whether the same results will be obtained in this study involving verb directedness which is
typically context-independent. Second, as shown by (3.2), ta-ziji shows no clear sign of empathy

115
sensitivity and is assumed to be anaphoric (e.g. Pan, 1998). If the effects of topic prominence found in
Experiments 3a and 3b are tied to perspective-taking, we may find no discourse topicality effect (i.e.,
main effect of context) for ta-ziji. However, if Pan’s (1998) Prominence Hierarchy can be extended to
discourse topic prominence, a context main effect may also be observed. In other words, we may still
observe more non-local choices for the supposedly non-empathic ta-ziji when the non-local subject is
topically prominent in the discourse.

3.7.2.4 Data analysis
Data analysis was identical to Experiment 3a.

3.7.2.5 Results
As can be seen in Figure 3.5, participants’ antecedent choices are modulated by verb bias as well as
context. See Table 3.4 for statistics. The main effect of verb bias is significant (p < 0.001): participants
clearly preferred non-local binding when the verb is other-directed (Biased context: 81%; Neutral
context: 75%) and local binding when the verb is self-directed (Biased context: 77%; Neutral context:
82%). This contrasts with the antecedent choice results in Lu (2011) regarding ta-ziji where local
antecedents are preferred. I find – in line with Experiment 3a on ziji – that verb bias is a stronger cue in
offline judgment than syntactic locality in the interpretation of ta-ziji. I come back to this in the
Discussion. The main effect of context is also significant (p < 0.05), in the same direction as what was
found with ziji in Section 3.6.2.5. This novel finding indicates that the previous claim that ta-ziji is

116
sensitive to antecedent prominence (Pan, 1998; Pan & Hu, 2002) applies to the discourse level. If the
non-local antecedent is the discourse topic and thus more prominent, LD binding is more accessible.

Figure 3.5: Distribution of mean proportions of local readings of ta-ziji across individual participants
(Panel A) and mean proportions of local readings (Panel B) at the group level in Experiment 4a.

Table 3.4: Summary of statistics for Experiment 4a (*: < 0.05).
β SE t-value p-value
Context -0.42 0.18 -2.38 0.017*
Verb -3.13 0.20 -15.34 < 0.001*
Context x verb -0.26 0.35 0.46 0.46

Comparing Experiment 3a (ziji) and Experiment 4a (ta-ziji)
A visual comparison of Experiments 3a and 4a (Figures 3.3 vs. 3.5) seems to show that although ziji and
ta-ziji are both sensitive to the context manipulation, ziji exhibits a stronger topicality effect: in biased
contexts, ziji seems to elicit higher proportions of non-local antecedent choices compared to ta-ziji. To
assess the differences between ziji and ta-ziji statistically, I included reflexive type as a between-
participant predictor (contrast coding: ta-ziji = 0.5, ziji = -0.5). If ta-ziji and ziji differ in terms of the
sensitivities to the context manipulation, we should see a context x reflexive type interaction.
0%
25%
50%
75%
100%
Self−directed Other−directed
Local coreference
A
19 %
77 %
25 %
82 %
0%
20%
40%
60%
80%
Self−directed Other−directed
Local coreference
Context
Biased
Neutral
B

117
As summarized in Table 3.5, the statistical model indicates that the main effect of reflexive type is
significant (p < 0.05), suggesting that overall participants preferred non-local antecedents more with
ziji than with ta-ziji. Crucially, the context x reflexive type interaction is also significant (p < 0.05),
meaning that the interpretations of ziji and ta-ziji are impacted differently by discourse topicality.
Pairwise comparisons show that in biased contexts ziji tends to refer to non-local antecedents more
than ta-ziji does (β = 0.59, SE = 0.18, t = 3.20, p = 0.001) but they do not differ in neutral contexts (β =
0.06, SE = 0.18, t = 0.32, p = 0.75). This means that although both forms show sensitivity to discourse
topicality, the interpretation of ta-ziji is less sensitive to discourse-level prominence than ziji. I turn to
Experiment 4b next to examine whether this difference has consequences for online processing.

Table 3.5: Summary of statistics on the comparison of antecedent choices in Experiment 3a (ziji) and
Experiment 4a (ta-ziji) (*: < 0.05).

β SE t-value p-value
Context -0.68 0.13 -5.26 < 0.001*
Verb -3.08 0.14 -21.37 < 0.001*
Reflexive 0.31 0.14 2.20 0.027*
Context x Verb 0.01 0.26 0.05 0.96
Context x Reflexive 0.51 0.25 2.01 0.045*
Verb x Reflexive -0.09 0.25 -0.37 0.71
Context x Verb x Reflexive -0.59 0.50 -1.17 0.24

3.7.3 Experiment 4b: Self-paced reading
3.7.3.1 Participants
Eighty-three adult Chinese native speakers participated via the internet. The study was implemented
using Ibex Farm (Drummond, 2013). None had participated in any other studies reported in this chapter.

118
3.7.3.2 Procedure
The procedure is the same as in Experiment 3b.

3.7.3.3 Predictions
In neutral contexts without a discourse prominent topic, local antecedents should be preferred. This
means that other-directed verbs should lead to reading slowdowns. This is predicted by both the
standard cue-based and the structure-based retrieval model. However, since Experiment 4a shows that
discourse topic prominence is one of the linguistic factors that impact the interpretation of ta-ziji –
hence a discourse-level cue that guides antecedent retrieval – the standard retrieval model predicts that
discourse topicality is expected to impact the processing of ta-ziji, a context x verb bias interaction in
statistical terms. The structure-based retrieval model only predicts a locality bias (i.e., a verb bias main
effect) at the early processing stage, although it does not preclude a late-stage interaction.

3.7.3.4 Data analysis
Data analysis is the same as in Experiment 3b.

3.7.3.5 Results
The mean reading times for each region by condition are shown in Figure 3.6. See Table 3.6 for statistics.
The critical region is Region 7, ta-ziji. Before the critical region, the main effect of verb bias reaches
significance at the matrix subject (e.g. ‘Xiaoming/he’) and the embedded subject (‘Professor Wang’),
respectively (ps < 0.05). These effects are probably spurious because the biased verbs have not yet been

119
encountered. No other region shows any main effect of verb bias before the critical region, including
the verb.
16
Below, I report statistical results which relate to the predictions in Section 3.7.3.4.

Figure 3.6: Mean RTs (ms) across all regions in the critical sentence in Experiment 3a.

There are three main findings of interest. First, the main effects of context are found for all regions
except Region 1 (ps < 0.001), which is expected under a perspective shift account but not crucial to the
purpose of this study.

16
One possibility for the variation in the onset of the verb bias effect in Experiments 3b and 4b may be
related to the frequencies of ziji and ta-ziji. A reflexive form with lower frequency may be more difficult to
predict and retrieve from mental lexicon. Indeed, a search in the Center for Chinese Linguistics Corpus (Zhan et
al., 2003) shows that ziji is 17 times more frequent than ta-ziji. Future work should look more into this.
verb
ta−ziji
400
600
800
1000
In lecture, Xiaoming/he heard Prof.Wang just verbed ta−ziji DE academic paper.
RT (ms)
Context
Biased
Neutral
Verb
Self−directed
Other−directed

120

121
Second, at ta-ziji, a main effect of verb bias was observed: other-directed verb conditions are read
more slowly than self-directed verb conditions (p < 0.001) but crucially this is qualified by a context x
verb bias interaction that reaches marginal significance in the log-transformed RT analysis (p = 0.08)
and significance in the raw RT analysis (β = -49.33, SE = 18.55, t = -2.66, p < 0.01). Pairwise comparisons
in log-transformed RT analysis (and in raw RT analysis) show that self-directed and other-directed verbs
differ in the neutral context (longer RTs for other-directed verbs: β = -0.09, SE = 0.02, t = -3.99, p < 0.001),
but not in the biased context (β = -0.04, SE = 0.02, t = -1.61, p = 0.11). The finding that other-directed verbs
elicit longer RTs than self-directed verbs in the neutral context points towards a locality bias, consistent
with prior studies which only considered stand-alone sentences (e.g., Dillon et al., 2016; Qian & Wu,
2016; Wang, 2017a,b) and with the processing patterns of ziji in neutral contexts in Experiment 3b.
However, in contrast to Experiment 3b, in the biased contexts with discourse topics, no differences
between self- and other-directed verbs were found, suggesting local and non-local binding are equally
probable to Chinese speakers in real-time processing.
Finally, at the spillover regions, the context x verb bias interaction is significant at region 9
(‘management’, p = 0.005). The interaction is also significant at region 10 (‘ability’) but only in raw RT
analysis (β = -122.99, SE = 52.79, t = -2.33, p < 0.05). But pairwise comparisons in log-transformed RT
analysis suggest that for both regions, other-directed verbs caused reading slowdowns in the neutral
context only (Region 9: β = -0.04, SE = 0.02, t= -2.22, p < 0.05; Region 10: β = -0.12, SE = 0.04, t = -3.19, p =
0.001), aligning with the pattern at the reflexive region.

3.7.4 Discussion of Experiments 4a and 4b

122
Experiment 4 explores whether discourse topicality and verb bias impact the offline interpretation of
ta-ziji and how discourse topicality modulates the incremental processing patterns of ta-ziji.
The results from the offline judgment Experiment 4a show that verb bias semantics is a
predominant factor in determining the reading of ta-ziji. This means that Chinese speakers prioritize
verb bias semantics over the locality constraint. On the surface, this seems to conflict with the findings
of Lu (2011) which showed that the interpretation of ta-ziji is mainly guided by locality. However, it
should be borne in mind that Lu’s manipulation of verb bias is context-dependent. For example, the
verbs used by Lu (2011) are not inherently biased towards (non-)local antecedents as anyone can ‘deliver’
or ‘receive’ their own or others’ packages. Only by considering the identity of the agent (e.g., ‘mailman’)
or context is a specific reading (local or non-local) more felicitous. In contrast, the verbs used in this
study are inherently biased as it is semantically anomalous to ‘publish’ someone else’s paper or to ‘grade’
someone’s own paper, regardless of the identity of the agent. Thus, one contribution of the forced
choice study is showing that different types of verb information play different roles in offline reflexive
resolution.
Perhaps more importantly, the forced choice study presents evidence that ta-ziji is sensitive to the
discourse topicality status of the antecedent, thereby suggesting that the semantics-based prominence
constraint proposed by Pan (1998) could be extended to the discourse level. However, a comparative
analysis of the participants’ antecedent choices in Experiment 3a (ziji) and 4a (ta-ziji) reveals a crucial
difference: the discourse topical status of the non-local antecedent is more likely to lead to non-local
choices in the case of ziji compared to ta-ziji, arguably related to the different linguistic properties of
ziji and ta-ziji. As shown in (3.2), ta-ziji is not sensitive to perspective centers or empathy loci. Therefore,

123
the different discourse prominence effects could boil down to the different discourse mechanisms
underlying the long-distance binding of empathic ziji and non-empathic ta-ziji (Pan, 1998). Topicality-
related perspective-taking could be a stronger discourse-level cue that prompts participants to consider
non-local antecedents in the case of ziji than non-perspective-related topic prominence in the case of
ta-ziji. This speculation, though firmly grounded in theoretical work, should be investigated in future
research.
Given the crucial finding that discourse topicality is indeed one of the linguistic factors that
influence the interpretation of ta-ziji, how fast is discourse topic prominence – notated with [+Top] –
accessed by the language parser to guide antecedent retrieval? The context x verb bias interaction
suggests that the [+Top] cue is available at the earliest possible moment when the reflexive is
encountered. In biased contexts, the absence of any verb bias effect indicates that Chinese speakers
consider local and non-local binding to be equally probable, in sharp contrast with the clear locality
bias in the neutral context where no antecedent is discourse-prominent. These results are consistent
with the predictions of the standard cue-based retrieval model but do not fit well with the predictions
of the structure-based retrieval model.
One more insight we can gain from the reading times in the self-paced reading task is that Chinese
speakers did not show any binding preference in the processing of ta-ziji, which bears on the linguistic
question of whether discourse prominence is weighted more heavily than syntactic locality. If discourse
topic prominence outranks the locality constraint, we should see a non-local binding preference. The
absence of such a preference indicates that discourse topic prominence only impacts but does not
outrank syntactic locality. Together with the reading time patterns of ziji, this again highlights the

124
linguistic differences of ziji and ta-ziji: ziji can be readily bound by empathy loci due to its empathic use
while ta-ziji is an anaphoric reflexive insensitive to perspective-taking.

3.8 General discussion
This section summarizes the findings of this chapter and discusses the implications for linguistic
theories and sentence processing models.

3.8.1 Summary of results
The goal of the studies in this chapter is two-fold. First, the linguistic goal is to explore how verb
semantics and discourse topic prominence impact the reading of the complex reflexive ta-ziji. Answers
to this question can not only resolve some of the open questions in the linguistic field but can also lay
the foundation to the subsequent sentence processing study. Second, the psycholinguistic goal is to
examine the time-course of the application of the discourse-level cues related to discourse topicality –
[+POV] and [+Top]– during incremental sentence comprehension. Examining whether, when, and
how discourse topicality impacts reflexive resolution in Mandarin can shed light on the cue-based and
structure-based retrieval models.
In two sets of forced-choice and self-paced reading experiments, I manipulated the factors verb
directedness and discourse context. Self-directed verbs semantically bias local binding while other-
directed verbs bias non-local binding. The influences of the verb directedness and context are directly
measured in the offline forced-choice experiments. In the online experiments, verb directedness is used
as a diagnostic to probe participants’ binding preferences. In neutral contexts, syntactic locality is

125
expected to clash with the semantically driven expectation for non-local binding when the verb is
other-directed, because by default ziji and ta-ziji are used as syntactic reflexives (e.g., Gao et al., 2005;
Jäger et al., 2015; Dillon et al., 2014, 2016; Qian & Wu, 2016; Chang et al., 2020). If the discourse topic
status of the non-local antecedent is also used by the parser to locate the antecedent, we expect the
processing patterns to change in biased contexts. The standard cue-based retrieval model predicts
different processing patterns of ziji and ta-ziji in neutral vs. biased context, i.e., a context x verb bias
interaction, while the structure-based retrieval model predicts no difference in processing patterns in
the two contexts, i.e., a main effect of verb bias.
Experiment 3 investigated the interpretation of ziji. In the forced-choice Experiment 3a, Chinese
speakers overwhelmingly based antecedent choices on verb semantics: when the verb is self-directed,
they preferred local binding; when the verb is other-directed, they preferred non-local binding.
Moreover, Chinese speakers are more likely to consider discourse-topic referents in biased contexts,
relative to neutral contexts where no referent is topically prominent. This finding fits with prior claims
that one tends to take the perspective of a discourse topic (Kuno, 1987) and that ziji is perspective-
sensitive (e.g. Pan, 1997, 2001; Huang & Liu, 2001; Huang et al., 2009; Wang & Pan, 2015; Charnavel et al.,
2017). In the self-paced reading Experiment 3b, the reading time patterns are different in neutral and
biased contexts. Using verb directedness as a diagnostic, local binding preference was detected in
neutral contexts, while non-local binding preference was detected in biased contexts. (As discussed
earlier, the verb bias effects at the pre-reflexive region are presumably due to predictive processing.)
Thus, the context x verb bias interaction supports the standard cue-based retrieval model.

126
Experiment 4 explored similar questions in the processing of the complex reflexive ta-ziji. The
forced-choice Experiment 4a showed that, like ziji, ta-ziji is similarly subject to the influence of verb
directedness. Perhaps more importantly, the Chinese speakers’ preference for non-local binding varies
as a function of the discourse context: they are more likely to choose a non-local referent as the
antecedent when that referent is the discourse topic. This indicates that the interpretation of ta-ziji is
modulated not just by syntactic locality but also by other aspects of linguistic information. Intriguingly,
ta-ziji differs from ziji in showing a reduced sensitivity to discourse topic prominence, potentially
related to the different pragmatic properties of non-empathic ta-ziji and empathic ziji. Further evidence
showing influence of discourse topicality on the resolution of ta-ziji comes from the self-paced reading
Experiment 4b where in biased contexts Chinese participants considered local and non-local referents
to be equally probable antecedents. The linguistic implication is that discourse topic prominence does
not out-rank but simply influences the processing of ta-ziji. The psycholinguistic implication is that the
context x verb bias interaction again underscores the early accessibility of discourse topic prominence.
Taken together, Experiments 3 and 4 suggest that Chinese speakers recruit multiple sources of
information – syntactic and non-syntactic – at the early stages of reference resolution. However, the
antecedent judgment data suggest that, over time, discourse-level and syntactic constraints decay, as
participants largely based their antecedent choices on verb semantics. Another important finding is
that discourse topicality impacts the processing of ziji and ta-ziji differently, thus reflecting different
linguistic properties. Below, I first discuss the broader implications of the findings in this chapter to
linguistic theories on long-distance anaphora, before turning to a (potential) conceptual unification of
standard cue-based and structure-based retrieval models.

127

3.8.2 Long-distance binding of ta-ziji
A key contribution of this chapter is providing substantive evidence that the interpretation of ta-ziji is
not solely guided by Principle A. Verb semantics and discourse topic prominence guide reflexive
resolution as well. Of these two factors, verb semantics plays a dominant role with discourse topicality
playing a weaker role. As discussed in Section 3.7.3.7, the strong verb bias effect only trivially conflicts
with the findings of Lu (2011) because Lu’s study focuses on the pragmatic (plausibility) aspect of the
verb meaning. In this study, I focus on the semantic directedness of the verb action. The divergent
findings from Lu (2011) and this study should thus be viewed as distinct contributions to shedding light
on the linguistic nature of ta-ziji.
The novel finding that discourse topicality modulates Chinese speakers’ interpretation of ta-ziji
represents another step in the direction of uncovering the linguistic properties of this reflexive. In the
offline antecedent choice judgment task, participants are more likely to choose the topically prominent
referent as the antecedent; in the online self-paced reading task, the absence of any binding preference
in biased contexts suggests that topically prominent, non-local referents are equally considered as
plausible antecedents just as local antecedents are. Thus, we have converging evidence from offline and
online experiments that the so-called “local anaphors” (e.g., Oshima, 2007) are not necessarily local:
discourse topicality seems equally important as syntactic locality in guiding antecedent retrieval for ta-
ziji.
The discourse topicality effect associated with ta-ziji has linguistic implications in the following two
senses. First, it suggests that the Prominence Constraint originally proposed by Pan and colleagues (Pan,

128
1998; Hu & Pan, 2002) should include topic prominence, in addition to grammatical prominence and
animacy prominence. (As we shall see in Chapter 4, logophoric prominence may need to be included
as well.) However, unlike the animacy constraint which outranks the locality constraint (e.g., ‘Xiaoming 1
said that book ruined ta-ziji 1’), the discourse topic prominence does not outrank syntactic locality. It
simply influences the resolution of ta-ziji, just as much as syntactic locality does.
Second, the discourse topicality effect highlights the crosslinguistic difference in long-distance
anaphora, especially given the linguistic behavior of exempt anaphors in languages like English. In
English, local binding is preferred even when an animate referent is available in the local domain (see
e.g. Sturt, 2003; Cunnings & Sturt, 2014), which is known as the “intervention effect” (Pollard & Sag, 1992;
also see experimental work by Sturt, 2003 and Cunnings & Sturt, 2014). Only when a suitable referent
in the local domain is unavailable is the exempt reading licensed. Indeed, prior work on the English
reflexives suggest that syntactic reflexives subject to Principle A and exempt reflexives (which are
licensed extra-syntactically) are in complimentary distribution (Pollard & Sag, 1992; Reinhart &
Reuland, 1991, 1993).
For example, the exempt long-distance reading of himself in “John said that the newspaper
published a picture of himself” is allowed because the reflexive does not have an animate clausemate
antecedent. As a result, himself can only be licensed by pragmatic means. But if himself has a viable
clausemate antecedent, the claim is that it must be locally bound. This is summarized by Reuland (2001,
2011)’s Rule L: when a viable local antecedent exists, an exempt reading is impossible as this kind of
extra-syntactic reading is computationally more costly.

129
However, given the finding from Experiment 4b that non-local binding of ta-ziji in biased contexts
is equally considered as local binding, a categorical Rule L seems too strong. In fact, prior studies on
English reflexives (Runner et al., 2006; Kaiser et al., 2009) and on the Korean complex reflexive caki-
casin (Kim & Yoon, 2020) also show that exempt readings are possible when the local antecedent is
animate. As will be reported in Chapter 4, logophoric properties of non-local antecedents – source vs.
perceiver status (see e.g. Culy, 1994; Kaiser et al., 2009; Sloggett, 2017) – impact the online and offline
interpretation of ta-ziji as well, thus strengthening the argument that exempt reflexives and syntactic
reflexives are not necessarily in complementary distribution.
As a whole, the findings relating to ta-ziji suggest that, for the linguistic theory to properly reflect
the property of ta-ziji, we need to extend Pan and Hu’s (2002) original argument about the prominence
constraint such that semantic (e.g. animacy), syntactic (e.g. grammatical prominence), and discourse-
level (e.g. topic prominence) prominence can all impact the interpretation of ta-ziji. However, instead
of viewing the prominence constraint as categorically outranking the locality constraint, I submit that
prominence at different linguistics levels could have different effects or weights. In the case of syntactic
prominence and discourse topic prominence, I argue that the latter competes with but does not
outrank the former.

3.8.3 Towards a conceptual unification of two processing models
In this section, I discuss the broader implications of this study to the standard cue-based and the
structure-based retrieval models, and how these two models can be conceptually unified. The reading
times in the self-paced reading tasks fit with the predictions of the standard cue-based retrieval model.

130
However, other studies present evidence that syntactic cues are accessed first before non-structural
cues (e.g., animacy, gender, number), thereby supporting the structure-based retrieval model (e.g., Sturt,
2003; Dillon et al., 2013; Van Dyke & McElree, 2011; Cunnings & Sturt, 2014; Chang et al., 2020). Below, I
consider how studies demonstrating priority of syntactic cues can be reconciled with studies (including
this study) showing immediate accessibility of non-syntactic cues.
A quick survey of the literature on pronoun resolution shows that the studies supporting the
structure-based retrieval model are often related to the Binding Theory and c-command (e.g. Sturt, 2003;
Kazanina et al., 2007; Felser & Cunnings, 2012; Kush et al., 2015). Conceivably, these syntactic
constraints are “hard”, as opposed to “soft” constraints. The hard vs. soft constraint distinction has been
used to characterize human sentence processing (e.g., Hendriks & Spenader, 2006; Dillon et al., 2013;
Cunnings et al., 2014). A hard constraint – including (but not limited to) syntactic locality in English –
is either inviolable (e.g., Van Dyke & McElree, 2011; Dillon et al., 2013) or violable but heavily weighted
(e.g., Parker & Phillips. 2017). A weak constraint is a constraint that is violable and incurs weak penalties.
In the context of cue-based memory retrieval, a hard constraint would mean that the relevant structural
or non-structural cue instantiating this constraint is a heavily weighted cue which mainly guides
antecedent retrieval; a weak constraint would render the relevant cue less relied upon in antecedent
retrieval.
This distinction of hard and soft constraints can helpfully reconcile the different results from prior
work testing different types of syntactic constraints. For example, Kush et al. (2017) found that in a
strong crossover configuration as in (3.6a) which involves a potential violation of Principle C,
structurally inaccessible distractors (i.e., maintenance man) have no effect on pronoun resolution (as

131
would be consistent with the predictions of the structure-based retrieval model). The inaccessibility of
the distractor can be accounted for if we assume that Principle C is a hard constraint or a strong cue,
strong enough to block the accessibility of the distractor.

(3.6) a. Jane asked which maintenance man/lunch lady i it appeared that he spoke with t i …
b. Jane asked which maintenance man/lunch lady i it appeared that his boss spoke with t i …

Now consider a weak crossover configuration such as (3.6b). Here, Principle C is irrelevant and only
the weak crossover constraint is at play. Unlike the findings with sentences like (3.6a), the distractor
now interferes with pronoun processing (contrary to the predictions of the structure-based retrieval
model). This contrast could be captured if we understand the weak crossover constraint to be a weak
constraint, violable in face of semantic congruence between the binder and the pronominal variable.
In sum, the different processing patterns in the same study testing two different types of syntactic
constraints can be captured if Principle C is a hard constraint, and the weak crossover constraint is a
soft one. In other words, structural cues that trigger Principle C is a heavily weighted cue while the those
that trigger a weak crossover violation is defeasible at the early processing stages. At any rate, a
categorical structure-based retrieval model such that syntactic cues are necessarily accessed first to the
exclusion of the non-syntactic cues at the early processing stages seems untenable. (In Chapter 5, I shall
propose a weaker version of the structure-based retrieval model.)
With the distinction of hard and soft constraints in mind, I turn to the discussion of how the findings
with ziji and ta-ziji can be understood in the context of the two processing models while

132
accommodating prior study results for English reflexives. To this end, I hypothesize that the set of hard
and soft constraints vary cross-linguistically. As mentioned, English reflexives are heavily constrained
by Principle A (Sturt, 2003; Cunnings & Sturt, 2014) despite the possibility of empathic readings of
himself/herself (e.g., Zribi-Hertz, 1989; Pollard & Sag, 1992; Pollard & Xue, 1998). In contrast, Mandarin
reflexives ziji and ta-ziji are clearly influenced by discourse topicality in real-time processing. This could
mean that syntactic locality may be a softer constraint (conversely, discourse topic prominence may be
a harder constraint) in Mandarin than in English, which can be accounted for under a cue-based
memory retrieval framework. In fact, prior computational modeling approach to reflexive resolution
(e.g., Jäger et al., 2015; Parker & Phillips, 2017) has incorporated the concept of hard vs. soft constraints
into the cue-based retrieval models by adding a “prominence principle” to the model or by assigning
more weight to the structural constraints. Allowing variation in the weights of the relevant constraints
offers a solution to unifying the two views on antecedent retrieval.
However, a conceptual unification of the structure-based and cue-based memory retrieval is not
without challenges. While the differentiation of hard vs. soft constraints can be a promising way to
reconcile the different study results regarding the time-course of the accessibility of (non-)syntactic
information, there is still no clear consensus on whether hard constraints (e.g., Binding Principles) are
indeed prioritized in early processing stages (Runner et al., 2006; Kaiser et al., 2009; Dillon, 2014; Patil
et al., 2016; Drummer & Felser, 2018; Jäger et al., 2020). Therefore, more empirical work is needed to
clarify the correlation between the hard/soft constraints and the time-course of their accessibility in
real-time processing.

133
Chapter 4: Logophoric prominence and reflexive resolution

4.1 Introduction
The previous chapter investigated the role of discourse topic prominence in reflexive resolution in
Mandarin. This chapter examines whether another type of discourse prominence – logophoric
prominence – impacts Chinese speakers’ offline and online interpretation of ziji and ta-ziji. As there is
no uniform definition for the terms logophor and logophoricity (Charnavel, 2021), before I present my
research goals, I first clarify my use of these terms.
The term logophoric pronouns is first used by Hagège (1974) to refer to certain types of pronouns in
West African languages. These pronouns typically refer to the author of the indirect discourse “whose
speech, thoughts, or general state of mind are reported” (Clements, 1975: 141), also known as the attitude
holder. For example, as documented by Clements (1975: 142), the logophoric pronoun yè in Ewe can only
refer to the source of speech Kofi, as shown in (4.1) (also see Pearson, 2015). This property of referring to
the attitude holder (author of speech and thoughts) is called logophoricity.

(4.1) Kofi be yè-dzo.
Kofi say LOG-leave
‘Kofi 1 said that he 1/*2 left.’

The referents of logophoric pronouns – logophoric centers – and perspective centers are
conceptually distinct notions. However, these two notions are sometimes used synonymously (e.g.,

134
Sells, 1987; Huang & Liu, 2001; Charnavel et al., 2017). Crucially, as the notion of perspective (or point-
of-view) has often been used to account for the pragmatic properties of long-distance bound exempt
reflexives (e.g., Sells, 1987; Pollard & Sag, 1992; Reinhart & Reuland, 1993), perspective-sensitive
reflexives are sometimes called logophoric reflexives in the linguistic literature.
However, Culy (1997) argues that the notions of perspective-taking and logophoricity should be
kept apart as multiple logophoric pronouns with their respective referents can appear in the same
sentence (Kervran & Prost, 1969). If logophoric centers are perspective centers, then each sentence can
only have one logophoric center due to the constraint that there can be only one perspective center per
sentence (Banfield, 1982; Iida, 1992). Aware of this debate, I will not make any commitment to the
relation between logophoricity and perspective-taking in this work. Instead, following prior practice
(e.g., Sloggett, 2017; Xu & Runner, submitted), I will refer to Culy’s (1994) Logophoric Hierarchy in my
investigation of logophoricity, shown in Chapter 1 as (1.7) and repeated here as (4.2). Logophoric
reflexives are taken to be sensitive to the Logophoric Hierarchy (see Wang & Pan, 2015a; Sloggett, 2017
for similar views): a referent higher on the hierarchy (e.g., source of indirect speech) is more likely to
bind a logophoric reflexive than a referent lower on the hierarchy (e.g., perceiver of indirect speech).

(4.2) Logophoric Hierarchy
Speech > thought > knowledge > direct perception

In addition, throughout this dissertation, I follow Oshima (2006, 2007) by restricting the term
“logophoric reflexive” to a specific use of exempt reflexives whose referents are attitude holders (i.e.,

135
volitional agents whose speech, thought or state of mind is reported) (also see e.g., Charnavel, 2020 for
a similar use). Specifically, a logophoric reflexive can express or report the perspective of its attitude
holder but does not necessarily require the speaker to take the perspective of (or empathize with) the
attitude holder.
Crucially, logophoric reflexives bound by attitude holders are distinguished from empathic
reflexives which require the speaker to empathize with the empathy locus of the sentence (e.g., Oshima,
2006, 2007; Charnavel & Zlogar, 2015; Charnavel, 2020),
17
namely, the referent that one takes the
perspective of. Empathic reflexives are sensitive to empathy constraints, such as the Topic Empathy
Hierarchy and the Speech Act Empathy Hierarchy (Kuno & Kaburaki, 1977; Kuno, 1987). In this work, I
adopt this distinction and take logophoric reflexives to be reflexives sensitive to the Logophoric
Hierarchy and empathic reflexives to be sensitive to the empathy constraints.
Unlike Chapter 3 which examines the Topic Empathy Hierarchy, this chapter focuses on
logophoricity and aims to explore whether Mandarin reflexives ziji and ta-ziji show sensitivity to the
Logophoric Hierarchy using linguistic experiments. The experimental results will bear on our
understanding of these reflexive forms and contribute to the debate on logophoricity and long-distance
reflexives in Mandarin (e.g., Huang & Liu, 2001; Wang & Pan, 2012, 2015a,b; Liu, 2020; Xu & Runner,
submitted). In addition, I aim to examine from a sentence processing perspective the time-course of

17
Note that logophoric and empathic reflexives represent the two uses of the exempt reflexive. No claim is made
about whether logophoric and empathic uses of ziji are two different referential elements or homophones (see
Pollard & Xue, 2001 for discussion). Following Oshima’s (2007) practice for Japanese zibun, I use the terms
“logophoric ziji” and “empathic ziji” as shorthands for two different uses of ziji.

136
the accessibility of logophoric roles to examine the two memory retrieval models – the standard cue-
based and the structure-based retrieval models – in sentence processing.
This chapter is organized as follows. I first review the linguistic literature on logophoricity and
Mandarin reflexives, including more recent empirical evidence. Next, I discuss prior psycholinguistic
work on logophoricity and real-time reflexive resolution. Then, I report five experiments to address the
linguistic and psycholinguistic questions of this chapter. Finally, I discuss broad implications of the
findings for linguistic theories of anaphora and sentence processing models.

4.2 Logophoricity and Mandarin reflexives
In his pioneering work, Sells (1987) proposed a trichotomy of logophoric roles that license exempt
reflexives, source (volitional agent of communication), self (one whose mental state/attitude the
content of the proposition describes), and pivot (one with respect to whose space-time location the
content of the proposition is evaluated). Sources are most likely to be the antecedent of exempt
reflexives, followed by self and pivot roles. Huang and Liu (2001) adopted this trichotomy of discourse
roles to explain different degrees of acceptability of long-distance uses of the reflexive ziji, shown in
Chapter 3 as (3.1) and repeated here as (4.3):

(4.3) a. Yuehan i shuo Bier j xihuan ziji i/j.
John say Bill like self
‘John i said that Bill j liked self i/j.’

137
b. Bier j hui bangzhu ziji i/j rang Yuehan i hen gaoxing.
Bill would help self make John very happy
‘That Bill j would help self i/j made John i very happy.’
c. Dang Bier j zai jiaoji de deng ziji i/j de shihou Yuehan i hai mei chu men.
when Bill at anxious DE wait self DE time John still not exit door

The judgment is that long-distance binding of ziji by a source is most acceptable, followed in
descending order of acceptability by self and pivot roles. Based on this, Huang and Liu (2001) argue that
ziji has a “logophoric” use. However, given Oshima’s (2007) distinction between the empathic and
logophoric use of exempt reflexives,
18
an open question is worth asking: could this judgment reflect
different acceptabilities of two different uses of ziji, an empathic ziji bound by the empathy locus and a
logophoric ziji bound by the attitude holder?
In fact, in Charnavel’s (2020: 694) reclassification of antecedent roles for exempt reflexives, only
two roles are argued to be relevant, attitude holder and empathy locus. According to this new
perspective, in (4.3a) where the non-local antecedent ‘John’ is an attitude holder, ziji is logophoric; in
(4.3b,c) where ‘John’ is an empathic locus, ziji is empathic. Thus, the different degrees of acceptability
in (4.3) may reflect different acceptabilities of logophoric ziji and empathic ziji. Indeed, the Preference
Hierarchy of Logophoric Centers (Charnavel, 2020: 693) posits that it is easier for exempt reflexives to
refer to attitude holders than empathy loci. Given this, it seems that different degrees of acceptability

18
See Charnavel (2020: 694) for a reclassification of the three logophoric roles in Sells (1987) into two roles,
attitude holder and empathy locus.

138
of long-distance ziji in (4.3) do not provide incontrovertible evidence that ziji has logophoric properties
in line with the Logophoric Hierarchy.
In theoretical discussion, it is unclear whether ziji is empathic or logophoric or both. Wang and Pan
(2012, 2015a,b) argue that long-distance ziji is more appropriately classified as an empathic reflexive
than a logophoric reflexive. Citing Oshima’s (2007) criteria for distinguishing empathic and logophoric
uses of exempt reflexives, shown in (4.4), they found that ziji exhibits properties characteristic of
empathic reflexives: (i) ziji shows the blocking effect (e.g., Huang et al., 1984; Cole et al., 1990; Pan, 1997,
2001) where long-distance binding across a 1
st
-person pronoun – an external empathy locus – is not
allowed due to the Speech Act Empathy Hierarchy (Kuno, 1987); (ii) ziji shows subject orientation (e.g.,
Cole & Sung, 1994; Huang & Liu, 2001; Cole et al., 2006); (iii) ziji can have non-de-se readings (Wang &
Pan, 2014) despite a de se reading preference (e.g., Pan, 1997; Anand, 2006; Wang & Pan, 2012).

(4.4) Properties of logophoric and empathic uses of reflexives (Oshima, 2007: 22-23)
a. Compatibility with 1
st
-person pronouns
A logophoric reflexive can occur with a 1
st
-person pronoun, while an empathic reflexive cannot.
b. Possibility of non-subject binding/discourse binding
Only logophoric reflexives can be bound to a non-subject or extra-sentential antecedent.
c. De se interpretation
Only logophoric reflexives induce de se interpretations.

139
Based on the linguistic behaviors of ziji, Wang and Pan (2015a,b) argue that long-distance ziji is
essentially empathic and should not be labeled “logophoric,” because it does not exhibit the typical
linguistic properties of logophoric pronouns (also see Pan, 2001: 290-292 for a similar argument).
However, the availability of de se reading (i.e., the attitude holder must be aware that the event
described by the sentence involves himself/herself, see ex.(1.10) in Chapter 1) prompts Wang and Pan
(2015) to conclude that ziji can have a logophoric use when its antecedent is an attitude holder. This
opens the possibility that ziji can be logophoric or empathic, depending on the role of its antecedent.
However, the fact that ziji can have a de se reading is not a direct test of its logophoric property. For this
study, I propose a more direct test of ziji’s logophoric property by testing ziji’s sensitivity to the
Logophoric Hierarchy in (4.2): ziji should show a stronger bias to refer to a non-local source antecedent
than a perceiver antecedent.
The above discussion centers around ziji. What about ta-ziji? It turns out there is little discussion
regarding logophoricity and ta-ziji in the theoretical literature. This omission could presumably be due
to the common assumption that the bimorphemic ta-ziji is a syntactic anaphor strictly constrained by
Principle A (e.g., Huang, 1983; Battistella & Xu, 1990; Huang & Tang, 1991; Li, 1993; Cole & Sung, 1994).
However, given recent experimental work on complex reflexives in languages like Dutch, English, and
Korean (e.g., Kaiser & Runner, 2008; Kaiser et al., 2009; Sloggett, 2017; Kim & Yoon, 2020; Kaiser, 2022)
showing correlation between their interpretations and the discourse roles of antecedents, it is worth
exploring whether ta-ziji, too, is logophoricity-sensitive.

140
In fact, recent experimental work has made some effort in testing the logophoric properties of both
ziji and the less discussed ta-ziji. Using a two-alternative acceptability judgement task, Liu (2020) tested
native Chinese speakers’ responses to sentences like (4.5a,b).

(4.5) a. Ju Lisi shuo, zhe-jian shi shanghai-le (ta-)ziji.
According.to Lisi say this-CL event hurt-Asp (he-)self
‘According to Lisi, this event hurt (he-)self.’
b. Shuodao Lisi, zhe-jian shi shanghai-le (ta-)ziji.
Speaking.of Lisi this-CL event hurt-Asp (he-)self
‘Speaking of Lisi, this event hurt (he-)self.’

The two sentences above involve long-distance binding because the local antecedent ‘event’ is
inanimate. In (4.5a), Lisi is the source of information; in (4.5b), Lisi is a topical theme. Liu (2020) found
that while (4.5a) was found to be close to 100% acceptable, (4.5b) was judged acceptable less than 50%
of the time. This contrast was taken by Liu (2020) to mean that ziji has a logophoric use. However, this
conclusion seems premature, given that the two occurrences of Lisi are in different grammatical
positions in (4.5a) and (4.5b) respectively: Lisi is a subject in (4.5a) but an object in (4.5b). Since ziji
shows a strong subject orientation, the so-called logophoricity effect may be partially explained by
subject orientation. Indeed, the low acceptability of (4.5b) is in line with this alternative explanation.
Intriguingly, Liu also found a similar logophoricity effect with ta-ziji. Furthermore, the logophoricity

141
effect seems stronger for ziji than ta-ziji. But the different logophoricity effects are due to the higher
acceptability of ta-ziji in (4.5b) and not due to the higher acceptability of ta-ziji in (4.5a).
Xu and Runner (submitted) similarly investigated the logophoric properties of ziji and ta-ziji with
sentences such as (4.6). Using an antecedent judgment task, they found no significant difference in
antecedent choices between the “think” and “hear” conditions regardless of the reflexive type, which
seems to conflict with the findings of Liu (2020). But it should be noted that Xu and Runner did not test
the two endpoints of Culy’s Logophoric Hierarchy – source vs. perceiver – and only compared thinker
vs. perceiver roles. In addition, the local antecedent in Xu and Runner’s study is a plausible animate
antecedent, which could reduce the exempt readings of ziji and ta-ziji. (This is called the ‘intervention
effect’ in the literature.) What happens when long-distance bindings are considered and when source
roles are compared to perceiver roles? These open questions will be addressed in this chapter.

(4.6) Sunzhi {renwei/tingshuo} Guojun zai digu {ziji/ta-ziji}.
Sunzhi think/hear Guojun PROG underestimate self/he-self
‘Sunzhi thinks/hears that Guojun is underestimating self/he-self.’

In summary, prior theoretical and experimental studies have not reached a consensus on whether
ziji and ta-ziji are sensitive to the Logophoric Hierarchy. Additionally, even if we assume that these two
reflexives can have logophoric uses, it remains to be tested which reflexive is more sensitive to the
Logophoricity Hierarchy. This latter question is meaningful, since prior work suggests that different

142
referential forms with functional overlaps can show different properties (e.g., Kaiser & Truswell, 2008;
Patternson & Schumacher, 2021). This study attempts to answer these two key linguistic questions.

4.3 Logophoricity and antecedent retrieval
Section 4.2 discussed logophoricity largely from a theoretical-linguistic perspective. This section
reviews logophoricity and reflexive resolution from a psycholinguistic perspective. For the purposes of
this study, I am concerned with whether and when logophoric information is used during real-time
processing. Examining this question can shed further light on cue-based retrieval models.
In this section, I shall restrict myself only to the discussion of studies that directly control for the
logophoric role of the antecedent. Studies on the processing of “logophoric reflexives” (e.g., picture NP
reflexives) which do not directly relate to the Logophoric Hierarchy (see e.g., Foraker, 2003; Cunnings
& Sturt, 2014; Jalbert, 2018) are beyond the scope of this section.
In a visual world eye-tracking study, Kaiser et al. (2009) examined native English speakers’
processing of picture NP reflexives in sentences like (4.7a,b) where the matrix subject is either a source
or perceiver. Picture NP reflexives are typically considered exempt reflexives and can be free in the local
domain (e.g., Pollard & Sag, 1992; Reinhart & Reuland, 1993).

(4.7) a. Peter {told/heard from} Andrew about the picture of himself on the wall.
b. Peter {told/heard from} Andrew about Greg’s picture of himself on the wall.

143
For both types of sentences in (4.7a) and (4.7b), participants’ offline interpretation of reflexives
suggests that numerically they preferred a source antecedent (Peter told …) over a perceiver antecedent
(Peter heard from …). Revealingly, the eye-tracking data suggest that, in the possessorless sentences like
(4.7a), participants showed a source advantage (i.e., looked more often at source subjects compared to
perceiver subjects) 200ms after the onset of the reflexive. This means that picture NP reflexives are
sensitive to the logophoric role of the antecedent at an early stage. However, in possessor sentences like
(4.6b), no source advantage was observed, possibly due to the proximity between the possessor (e.g.,
Greg) and the reflexive, which could suggest that reference resolution is constrained by multiple factors.
At any rate, Kaiser et al. provided substantive evidence that logophoricity can be immediately accessed
during reflexive resolution in English.
In another study using a reading eye-tracking method, Sloggett (2017) examined sentences like
(4.8a,b). In the study, reading slowdowns due to gender mismatch between the reflexive and a
preceding non-local referent were used as a diagnostic to test whether binding is considered: if a
comprehender attempts to link the reflexive herself to librarian/janitor, gender mismatch should lead
to surprisal, hence reading slowdowns.

(4.8) a. The {librarian fem/janitor mas} {said/heard} that the schoolgirl misrepresented herself.
b. The {librarian fem/janitor mas} {said/heard} that the schoolboys misrepresented herself.

Sloggett (2017) found that when the local antecedent (e.g., schoolgirl) is a suitable antecedent for
the reflexive, the matrix subject does not interfere with reflexive resolution due to a lack of gender

144
mismatch effect. (This is the expected ‘intervention effect’ in English, see e.g., Pollard & Sag, 1992;
Pollard & Xue, 2001.) But when the local antecedent (e.g., schoolboys) cannot plausibly be the
antecedent, a gender mismatch effect at the reflexive region was observed – but only for the source
conditions – at the later processing stages (indicated by regression path durations and total reading
times). This suggests that herself is sensitive to the Logophoric Hierarchy when non-local binding is
considered due to implausible local binding.
Taken together, prior psycholinguistic studies suggest that English reflexives are sensitive to the
Logophoric Hierarchy during online resolution. However, the effect seems to only appear with
possessorless picture NP reflexives or when non-local binding must be considered due to implausible
local binding.
Turning to Mandarin, will ziji and ta-ziji show the logophoricity effect? If so, when does the
logophoricity effect appear? Investigating these questions bears on the standard cue-based retrieval
model. I have shown in Chapter 3 that discourse topic prominence can be used quickly at the early
processing stage to guide antecedent retrieval, which favors the standard cue-based retrieval model.
What about logophoric roles? The standard cue-based retrieval model predicts that by default all
sources of information can be accessed quickly by the parser. But it is conceivable that not all kinds of
discourse-level information can be accessed at the same processing stage. In this study, I will test the
role of logophoricity in reflexive resolution in Mandarin to further our understanding of the antecedent
retrieval process.

4.4 Overview of the aims

145
There are two broad goals for this chapter, a linguistic one and a psycholinguistic one.
First, using forced choice judgment tasks, this chapter aims to explore when non-local binding is
considered, whether ziji and ta-ziji show logophoric properties in line with the Logophoric Hierarchy
(Culy, 1994). Relatedly, if these two reflexives are sensitive to the logophoric roles of the non-local
antecedents, is there any difference regarding the strength of the logophoricity effect as suggested by
Liu (2020)?
Second, using self-paced reading, this chapter probes at what stage of reference resolution the
logophoric cues are used for retrieving potential antecedents, especially compared to the early
accessibility of discourse topicality observed in Chapter 3. This question speaks to the standard cue-
based retrieval model which by default admits immediate recruitment of all kinds of (discourse)
information for antecedent retrieval. More data on the time-course of the logophoricity effect in
Mandarin will further our understanding of this memory retrieval model.
Below, five experiments are reported to accomplish the goals of this chapter. Experiment 5
examines the logophoricity effect associated with ziji using forced choice judgment (Experiment 5a)
and self-paced reading (Experiment 5b) tasks while Experiment 6 investigates the processing of ta-ziji
using the same methods. Experiment 7 is a forced choice judgment experiment that directly compares
the logophoricity effects shown by ziji and ta-ziji.

4.5 Experiment 5: Logophoricity and processing of ziji
Experiment 5 is composed of a forced choice experiment (Experiment 5a) and a self-paced reading
experiment (Experiment 5b). Experiment 5a tests whether the logophoric role of the non-local

146
antecedent affects Mandarin speakers’ interpretation of ziji at all. If it has an impact on participants’
antecedent judgments, this paves the way for Experiment 5b where I test whether and when logophoric
roles are used during incremental real-time processing.

4.5.1 Materials and design
Experiment 5a and 5b used the same set of stimuli, adjusted from the stimuli used in Chapter 3. In this
study, the non-local subject is consistently a discourse topic so that comprehenders are biased to take
the perspective of the non-local subject (Topic Empathy Hierarchy, Kuno, 1987). As shown in Chapter
3, this topicality manipulation is very successful in pushing participants to consider non-local binding
for both ziji and ta-ziji. As discussed in Sections 4.2 and 4.3, the chances of finding logophoricity effects
are maximized when the exempt reading of reflexives is considered (e.g., Kaiser et al., 2009; Sloggett,
2017; Liu, 2020); when reflexives are used as syntactic anaphors, the logophoricity effect may not appear
due to the “intervention effect” (e.g., Xu & Runner, submitted). Therefore, keeping the discourse topic
status of the matrix subject may be conducive to discovering the logophoricity effect, if there is any.
While keeping the discourse topic status of the matrix subject constant, this study controls two
factors, the logophoric role (source/perceiver) of the matrix subject and the verb bias (self-
directed/other-directed) of the embedded verb. Following Kaiser et al. (2009), a source role is marked
by the verb phrase ‘tell (others)’; a perceiver role is marked by the verb ‘hear’. The verb ‘tell’ was used
instead of ‘say’ because ‘tell’ highlights the ‘source-recipient’ relation. See (4.9) for an example. The
subscripts indicate the reading regions in the self-paced reading experiment.

147
The verb bias variable serves the same purpose as in Chapter 3. It tests the contribution of verb
semantics to reflexive interpretation in the offline judgment task. It also serves as a diagnostic to test
participant’s binding preferences in online reading self-paced when the reflexive is encountered. As a
reminder, if local binding is preferred, we anticipate other-directed verbs to lead to reading slowdowns
at and/or after the reflexive (in case of predictive processing, this effect may appear at the verb region);
if non-local binding is preferred, self-directed verbs are expected to lead to reading slowdowns.
Specifically for this study, the stronger the preference for non-local binding, the bigger the verb
directedness effect.

(4.9) a. Source/Self-directed
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 gaosu 3 bieren 4 Wang jiaoshou 5
Xiaoming be class in DE excellent student lecture on he tell others Wang professor
fabiao-le 6 ziji 7 de 8 xueshu 9 lunwen. 10
publish-ASP self DE academic paper
‘Xiaoming is an excellent student in the class. During the lecture, he told others that
Professor Wang had published self’s academic paper.’
b. Source/Other-directed
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 gaosu 3 bieren 4 Wang jiaoshou 5
Xiaoming be class in DE excellent student lecture on he tell others Wang professor
pigai-le 6 ziji 7 de 8 xueshu 9 lunwen. 10
grade-ASP self DE academic paper

148
‘Xiaoming is an excellent student in the class. During the lecture, he told others that
Professor Wang had graded self’s academic paper.’
c. Perceiver/Self-directed
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 tingshuo 3 Wang jiaoshou 5
Xiaoming be class in DE excellent student lecture on he hear Wang professor
fabiao-le 6 ziji 7 de 8 xueshu 9 lunwen. 10
publish-ASP self DE academic paper
‘Xiaoming is an excellent student in the class. During the lecture, he heard that Professor
Wang had published self’s academic paper.’
d. Perceiver/Other-directed
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 tingshuo 3 Wang jiaoshou 5
Xiaoming be class in DE excellent student lecture on he hear Wang professor
pigai-le 6 ziji 7 de 8 xueshu 9 lunwen. 10
grade-ASP self DE academic paper
‘Xiaoming is an excellent student in the class. During the lecture, he heard that Professor
Wang had graded self’s academic paper.’

Twenty sets of target stimuli were distributed into 4 lists using a Latin square design and presented
to the participants alongside 20 filler sentences. The fillers are the same sentences as used in Chapter 3.
Comprehension questions to the filler sentences are either factually true or false. They are used as
attention checkers. Participants saw 5 different sentences for each condition. In the self-paced reading

149
Experiment 5b, the critical region is the reflexive, followed by three spillover regions, DE, a nominal
modifier, and a noun.

4.5.2 Experiment 5a: Forced choice judgment
4.5.2.1 Participants
Forty-seven participants from mainland China participated in the study remotely. All participants had
normal or corrected-to-normal vision.

4.5.2.2 Procedure
The experiment was run on PC Ibex (Zehr & Schwarz, 2018). For each trial, the context sentence and
the critical sentence were presented separately on different screens. After each trial, a comprehension
appears. The target trials probe participants’ interpretation of ziji by asking them to select one of the
two antecedent choices. The questions are the same as those used in the studies reported in Chapter 3
(e.g., ‘whose paper has been published?’). The order of antecedent choices was randomized.

4.5.2.3 Predictions
For the verb bias effect, consistent with the predictions in Chapter 3, self-directed verbs are expected
to bias local reading and other-directed verbs should bias non-local reading of ziji. As to the logophoric
roles, suppose ziji is sensitive to the Logophoric Hierarchy, source conditions are predicted to lead to
more non-local choices compared to perceiver conditions. In other words, we should see main effects
of verb bias and logophoric role.

150

4.5.2.4 Data analysis
Before data analysis, participants with comprehension question accuracies below 75% were removed,
which affected the data of 3 participants. Statistical analysis was run on R (R Core Team, 2020).
Contrasts were fitted for verb bias (self-directed: +0.5, other-directed: -0.5) and logophoric role (source:
+0.5, perceiver: -0.5). For the remaining 44 participants, mixed-effect logistic regression was run to
analyze antecedent choices using the glmer function (lme4 package). The fixed effects are verb bias and
logophoric role, and the random effects are intercepts and slopes grouped by participants and items.
Simpler models were preferred over more complex models if they do not differ significantly in model
comparisons (Bates et al., 2015). An alpha value of 0.2 instead of 0.05 is chosen (Matuschek et al., 2017)
for model comparisons.

4.5.2.5 Results
The participants’ non-local choices are plotted in Figure 4.1. See Table 4.1 for statistics. The left panel A
shows the distribution of antecedent choices by participant; the right panel B shows the mean
percentage of choices of all participants.

Table 4.1: Summary of statistics for Experiment 5a (*: < 0.05).
β SE t-value p-value
Role 0.20 0.20 1.00 0.31
Verb -3.31 0.22 -14.93 < 0.001*
Role x verb -0.93 0.40 -2.33 < 0.05*

151

Figure 4.1: Distribution of mean proportions of non-local readings of ziji across all individual
participants (Panel A) and mean proportions of non-local readings (Panel B) at the group level in
Experiment 5a.

As predicted, the main effect of verb bias is significant (p < 0.001). The main effect of logophoric role is
not significant. However, the statistical model reveals an interaction (p < 0.05). Pairwise comparisons
suggest that source matrix subjects lead to more non-local readings only in the self-directed verb
conditions (β = -0.69, SE = 0.25, t = -2.80, p < 0.01). The lack of the logophoricity effect in the other-
directed verb conditions may be because most of the participants’ choices are close to ceiling, shown
in panel A. In sum, this study provides evidence that ziji is sensitive to the logophoric role of its
antecedent, suggesting it can have a logophoric use.

4.5.3 Experiment 5b: Self-paced reading
4.5.3.1 Participants
Forty-nine adult Mandarin native speakers in mainland China participated over the internet. All had
normal or corrected-to-normal vision.

0%
25%
50%
75%
100%
Self−directed Other−directed
Non−local coreference
A
90 %
86 %
24 %
36 %
0%
25%
50%
75%
100%
Self−directed Other−directed
Non−local coreference
Role
Source
Perceiver
B

152
4.5.3.2 Procedure
The experiment was run on PC Ibex. Each trial was composed of a lead-in context sentence and a critical
sentence. The context sentence was displayed as a complete sentence. The critical sentence was shown
region-by-region on a separate screen, as indicated by the subscripts in (4.9). Upon completing a trial,
participants pressed the space kay to see a comprehension question. The presentation order of the two
answer choices for the comprehension question was randomized.

4.5.3.3 Predictions
To get probe participants’ binding preferences in real-time processing, I again rely on the verb bias
effect as a diagnostic (introduced in Chapter 3): longer RTs in the other-directed verb conditions at
and/or after ziji point to a local binding preference while longer RTs in the self-directed conditions
suggest a non-local preference.
Building on the results of Chapter 3 where we see that the empathy locus is quickly accessed as a
retrieval cue, I predict a main effect of verb bias at the earliest moments of reflexive resolution. This is
because Chinese speakers are expected to quickly treat the non-local discourse topic as the binder. If
logophoric information is also utilized at a similarly early stage, we should additionally see a verb bias
x logophoric role interaction at the same region such that the non-local preference is stronger in the
source conditions relative to the perceiver conditions. This is because both source role and topicality
together provide stronger cues (relative to topicality alone) favoring non-local binding. No predictions
can be made about the main effect of logophoricity in self-paced reading as the source and perceiver
sentences differ in sentence length and referential complexity (i.e., 3
rd
-person plural ‘others’ in the

153
source conditions), which makes this main effect tangential to my research purpose (unlike Experiment
5a).

4.5.3.4 Data analysis
Participants whose comprehension accuracies on the 28 factually (in)correct questions fell below the
75% threshold were removed from analysis. This resulted in the removal of 4 participants. The
remaining 45 participants have a mean accuracy of 94%. Extreme RTs below 100ms or longer than
3000ms were trimmed, as were RTs 2.5 SDs above the mean by region and by condition. These steps
affected 2.42% of the original data. For statistical analysis, log-transformed RTs and raw RTs were
entered into mixed-effect linear models (lmer function in lme4 package). Log-transformed RTs are
reported. But in case of inconsistency between log-transformed and raw RTs, both will be reported.

4.5.3.5 Results
Figure 4.2 displays participants’ mean RTs across regions by condition. The critical region is Region 7,
ziji. See Table 4.2 for statistics. Before the critical region, there is a main effect of verb bias at Region 1
(p < 0.05), the adverbial (e.g., ‘In lecture’), which I take to be a spurious effect as the biased verb has not
been encountered yet. At region 6, the biased verb, the main effect of logophoric role suggests that the
source conditions lead to more processing cost due to the dative object (e.g., ‘others’) which is not
present in the perceiver conditions. This is not crucial to the purpose of this study.

154

155

Figure 4.2: Mean RTs (ms) across all regions in the critical sentence in Experiment 5b. In the perceiver
conditions, there is no matrix object (e.g., ‘others’).

At the critical region ziji, the statistical model reveals two main effects of verb bias and logophoric
role (ps < 0.01). The predicted early verb bias main effect shows that self-directed verbs lead to reading
slowdowns due to a non-local binding preference, which replicates the findings from Chapter 3. The
main effect of logophoric role at this region can be regarded as a continuation of the extra processing
costs related to the representation and processing of the matrix object argument ‘others’ in the source
conditions (which also characterizes the biased verb region). The lack of a significant interaction
suggests that logophoric information is not accessed at this early stage, in contrast to the immediate
effect of discourse topicality.
At the three spillover regions, only the main effect of verb bias is significant (ps < 0.05), again
indicating that Chinese speakers show strong surprisal at seeing self-directed verbs which conflict with
the non-local binding preference. Critically, statistical analysis at the spillover regions shows no
400
500
600
700
800
In lecture, he heard/told others Prof.Wang verbed ziji DE academic paper.
RT (ms)
Verb
Self−directed
Other−directed
Role
Source
Perceiver

156
significant interaction. This could mean that logophoric prominence does not modulate real-time
processing. But in combination with the data from the forced choice judgment Experiment 5a, we can
speculate that the logophoric information is accessed at a very late stage of reflexive resolution (which
could happen during re-reading), at least for ziji.
An alternative explanation for the absence of a verb bias x logophoric role interaction could be the
small effect size of the logophoricity effect, given that in both source and perceiver conditions,
participants prefer non-local binding. A visual inspection of the two spillover regions (i.e., ‘DE’,
‘academic’) in Figure 4.2 show a trending logophoricity effect. Indeed, if we visualize the participants’
mean preference for non-local binding in source and perceiver conditions by subtracting the RTs in the
other-directed verb condition from the RTs in the self-directed verb condition (see Figure 4.3), we see
that the participants’ non-local binding preference overall seem stronger especially in the source
conditions at the ‘DE’ region. However, due to the lack of a significant interaction, there is no clear
evidence that logophoricity is used when ziji is first encountered in incremental processing.

157

Figure 4.3: Participants’ non-local binding preference in RTs. The non-binding preference in real-time
sentence processing is quantified by subtracting the RTs in the other-directed verb condition from the
RTs in the self-directed verb condition.

4.5.4 Discussion of Experiments 5a and 5b
Experiments 5a and 5b investigated whether the logophoric role of the empathy locus impacts the
interpretation of ziji. The offline judgment results show that Chinese speakers prefer non-local binding
more when the non-local antecedent is a source compared to a perceiver. This finding aligns with the
prediction of the Logophoric Hierarchy and thus provides direct evidence that long-distance use of ziji
shows logophoric properties. What this means is that while ziji can be easily interpreted as empathic in
the presence of a discourse topic (a potential empathy locus), it can also be interpreted logophorically
if that discourse topic is simultaneously an attitude holder.
−200
−100
0
100
200
300
Source Perceiver
Non−local binding preference in RTs (ms)
ziji
0
100
200
300
Source Perceiver
DE
−250
0
250
500
Source Perceiver
Non−local binding preference in RTs (ms)
academic
−500
0
500
1000
Source Perceiver
paper

158
The online self-paced reading experiment asks a related question: if logophoric information is used
to guide antecedent retrieval, when is this information accessed? Participants’ RTs do not provide a
clear answer to this question as there is no substantive evidence that logophoricity modulates
incremental processing. At best, there are numerical trends at the spillover regions where the
preference for non-local binding – indicated by longer RTs associated with self-directed verbs – seems
stronger in source conditions relative to the perceiver conditions. At any rate, the null logophoricity
effect does not support the standard cue-based retrieval model which assumes that by default all
retrieval cues will have an impact on antecedent retrieval.
One way to accommodate the results is to postulate that while discourse-level cues can be
immediately used to search for the most optimal candidate antecedents, not all such cues are used. If
so, this would prompt us to introduce modifications to the standard cue-based retrieval model. In fact,
the idea that not all cues are used equally or given equal weights forms the core of the structure-based
retrieval model where it is argued that structural cues are prioritized (e.g., Van Dyke, 20017; Van Dyke
& McElree, 2011; Dillon et al., 2013). Here, we face a similar situation, except that some types of discourse-
level cues (e.g., discourse topicality) are prioritized while other types (e.g., logophoricity) seem to be
weak and delayed.
However, as alluded to in Section 4.5.3.5, the lack of verb bias x logophoric role interaction is perhaps
not surprising given that non-local preferences are found for both source and perceiver conditions.
(Future work should look more into this numerical trend by including a much larger sample of
participants.) But even so, in contrast to the immediate verb bias effect at the reflexive region, the

159
numerically trending logophoricity effect only appears at the spillover regions, lagging the discourse
topicality effect.
Taken together, the two experiments reported above seem to show that while ziji is strongly
empathic, it is weakly logophoric. This fits Wang and Pan’s (2015) argument that long-distance use of
ziji is primarily empathic but can be logophoric when its antecedent is an attitude holder. In addition,
the incremental RT patterns do not align well with a strong version of the standard cue-based retrieval
model that all retrieval cues are immediately accessed in real-time processing.

4.6 Experiment 6: Logophoricity and processing of ta-ziji
This section focuses on logophoricity and the processing of ta-ziji. Experiment 6a is a forced-choice
judgment study and tests whether the logophoric role of the discourse topic affects the interpretation
of ta-ziji. Experiment 6b is a self-paced reading experiment designed to examine the time-course of the
logophoricity effect.

4.6.1 Materials and design
The materials in Experiment 6 are identical to the materials in Experiment 5, except ziji was replaced
by ta-ziji.

4.6.2 Experiment 6a: Forced-choice judgment
4.6.2.1 Participants

160
Forty-four participants in mainland China participated remotely over the internet. All had normal or
corrected-to-normal vision. None had participated in any previous experiment.

4.6.2.2 Procedure
The procedure was identical to Experiment 5a.

4.6.2.3 Predictions
A main effect of verb bias is predicted, following prior observations in Chapter 3. A main effect of
logophoric role is also expected if long-distance binding of ta-ziji is sensitive to logophoricity. However,
given the statistics in Experiment 5a where the logophoricity effect only appears with self-directed verb
conditions, we might see no main effect of logophoric role but a verb bias x logophoric role interaction.

4.6.2.4 Data analysis
Data analysis follows the same steps as described in Experiment 5a.

4.6.2.5 Results
Figure 4.4. displays participants’ mean preferences for non-local readings across conditions. The left
panel A shows each individual’s mean antecedent choices while the right panel B only displays the
mean percentage of non-local binding at the population level. The statistical analysis shows that only
the verb bias effect reaches significance (p < 0.001). The main effect of logophoric role and the
interaction are not significant, which seems to suggest that the logophoric role of the matrix subject

161
does not impact the interpretation of ta-ziji. Nevertheless, the logophoricity effect in the self-directed
verb conditions does reach significance in a pairwise comparison analysis (β = -0.57, SE = 0.25, t = -2.30,
p < 0.05). Thus, despite a lack of a significant interaction, we have some evidence that ta-ziji, too, is
sensitive to the Logophoric Hierarchy.
19
I will revisit this question in Experiment 7. Below, I turn to
incremental processing to further test whether and how logophoricity impacts the reading of ta-ziji.

Figure 4.4: Distribution of mean proportions of non-local readings of ta-ziji across all individual
participants (Panel A) and mean proportions of non-local readings (Panel B) at the group level in
Experiment 6a.

Table 4.3: Summary of statistics for Experiment 6a (*: < 0.05).
β SE t-value p-value
Role 0.28 0.18 1.56 0.12
Verb -2.89 0.20 -14.43 < 0.001*
Role x verb -0.37 0.35 -1.05 0.29

4.6.3 Experiment 6b: Self-paced reading
4.6.3.1 Participants

19
The sentence-final judgment data on antecedent choices (12 items out of 20) show a similar pattern with a
marginal main effect of logophoric role (β = -0.46, SE = 0.24, t = -1.87, p = 0.06).
0%
25%
50%
75%
100%
Self−directed Other−directed
Non−local coreferences
A
82 %
84 %
24 %
33 %
0%
25%
50%
75%
100%
Self−directed Other−directed
Non−local coreference
Role
Source
Perceiver
B

162
Forty-seven adult native Chinese speakers participated over the internet. All had normal or corrected-
to-normal vision. No participant had participated in any other study presented in this chapter.

4.6.3.2 Procedure
The procedure is identical to that described in Section 4.5.3.2.

4.6.3.3 Predictions
The predictions for Experiment 6b are based on the results of Experiment 4b reported in Chapter 3.
There, we see that in biased contexts with non-local antecedents as the discourse topics and with the
matrix verbs being mostly perceiver verbs, the processing pattern of ta-ziji indicates a neutral binding
preference, which suggests that Chinese speakers consider topically prominent non-local antecedents
and local antecedents equally.
Prior studies show that when non-local binding of complex reflexives is considered, complex
reflexives also show sensitivity to the logophoric scale (e.g., English: Sloggett, 2017; Chinese: Liu, 2020)
although the time-course of the logophoricity effect is debated (Kaiser et al., 2009; Sloggett, 2017). Based
on the findings of Chapter 3 and prior studies, I predict that in the source conditions, Chinese speakers
may show a non-local binding preference because the source role is an additional cue favoring non-
local binding. Using the verb bias diagnostic to assess participants’ real-time binding preferences, I
predict that in the perceiver conditions, we should see no verb bias effect, but in the source conditions,
we could see self-directed verbs leading to reading slowdowns. In statistical terms, a verb bias x
logophoric role interaction is predicted. Furthermore, if this interaction (i.e., logophoricity effect) occurs

163
at the critical region, this effect can be taken to be an early effect; if later at the spillover regions, a
delayed effect.

4.6.3.4 Data analysis
Data analysis is the same as in Experiment 5b. Two participants were removed from data analysis as
their comprehension accuracy rates were below 75%. Data from 45 participants were entered into
statistical models.

4.6.3.5 Results
The participants’ mean RTs by region and condition are plotted in Figure 4.5. See statistics in Table 4.4.
Before the critical region ta-ziji, Region 1 (e.g., ‘In lecture’) shows a main effect of logophoric role (p =
0.01), probably a spurious effect as all conditions are exactly the same at this region. The biased verb
region shows a main effect of logophoric role as well (p = 0.002), echoing what we have seen in
Experiment 5b. There, I argued that this is actually due to the processing costs related to the
representation of the object ‘others’ which is only present in the source conditions. No other effects are
significant before the critical region.
At the critical region ta-ziji, there is again a main effect of logophoric role (p < 0.01) which should be
regarded as a continuing reflection of the processing costs related to ‘others’ in the source conditions.
No other effect is significant at this region: self-directed and other-directed verb conditions show
similar RTs, replicating the study results of Experiment 4b in Chapter 3.

164
At the first spillover region (i.e., ‘DE’), a main effect of verb bias is found (p = 0.002), as overall self-
directed verbs lead to reading slowdowns. But critically, there is a predicted interaction (p < 0.05) as
the verb bias effect is different between source and perceiver conditions. Planned comparisons show
that self-directed verbs lead to significant reading slowdowns (β = -0.08, SE = 0.02, t = -3.75, p < 0.001)
in the source conditions while there is no significant verb bias effect in the perceiver conditions (β = -
0.02, SE = 0.02, t = -1.02, p = 0.31). Linking the verb bias effect to the online binding preference, this
suggests that the participants prefer long-distance binding when the topical matrix subject is a source,
but do not show any binding preference when the topical matrix subject is a perceiver (replicating the
findings of Experiment 4b on ta-ziji in Chapter 3). In other words, the processing of ta-ziji is sensitive to
the logophoric role of its topical matrix subject.

Figure 4.5: Mean RTs (ms) across all regions in the critical sentence in Experiment 6b. In the perceiver
conditions, there is no matrix object (e.g., ‘others’).
300
400
500
600
700
800
In lecture, he heard/told others Prof.Wang verbed ta−ziji DE academic paper.
RT (ms)
Verb
Self−directed
Other−directed
Role
Source
Perceiver

165

166
4.6.4 Discussion of Experiments 6a and 6b
Experiment 6a and 6b test whether the logophoric role of the topically prominent matrix subject
modulates the offline interpretation and the online processing patterns of ta-ziji. In the forced-choice
judgment experiment, a weak logophoricity effect has been identified. The antecedent choice pattern
here is similar to the pattern in Experiment 5a on ziji. In the self-directed verb conditions, source roles
lead to higher proportions of non-local binding compared to perceiver roles, which aligns with the
Logophoric Hierarchy.
The logophoricity effect is clearly seen in the participants’ reading times in the self-paced reading
task. At the spillover region ‘DE’, a significant verb bias x logophoric role interaction was discovered.
This interaction fits the prediction outlined in Section 4.6.3.3: when the matrix subject is a topically
prominent perceiver, local and non-local binding are equally considered; but when the topically
prominent subject is a source, the tendency for non-local binding should be stronger. Indeed, this is
what the reading times have shown. Self-directed verbs lead to reading slowdowns only in the source
conditions at ‘DE’, suggesting a non-local binding preference. Therefore, the online and offline
experiments both provide evidence that when the non-local referent is considered as a possible
antecedent (by virtue of being a discourse topic), ta-ziji can have a logophoric use.
One other important finding is with regard to the time-course of the logophoricity effect. The
logophoricity effect, operationalized as a verb bias x logophoric role interaction, only appears at the first
spillover region. This contrasts with the immediate discourse topicality effect we have observed earlier
in the processing of ta-ziji in Chapter 3. There, the discourse topicality effect appears at the reflexive
region. Thus, logophoricity seems to play a similar role in the processing of ta-ziji and ziji: while the

167
discourse topicality cue can quickly guide antecedent retrieval, the logophoricity cue seems to play a
role only in the later processing stage. This does not support a strong version of the standard cue-based
retrieval model where all cues are accessible for antecedent retrieval at the early processing stage. It
seems that some discourse-level information (e.g., discourse topicality) is prioritized compared to other
types (e.g., logophoricity).
There is one open question in this study. How does ta-ziji compare to ziji in the strength of the
logophoricity effect? Given that in the offline experiments, the verbs are strongly biased and constitute
the main factor in participants’ decision on antecedent choices, the logophoricity effect may be
underestimated. Below, I turn to Experiment 7 that uses neutral verbs instead of biased verbs.

4.7 Experiment 7: Comparison of logophoricity effects in ziji and ta-ziji
Experiment 7 is a forced choice judgment task and compares ziji and ta-ziji’s sensitivities to the
Logophoricity Hierarchy, following the footsteps of Liu (2020). This study also fits in a larger research
agenda testing form-specific differences between referential forms in their sensitivities to prominence
constraints (e.g., Kaiser & Trueswell, 2008; Schumacher et al., 2016; Patterson & Schumacher, 2021). To
reduce the strong influence of verb directedness on the interpretation of reflexives, relatively neutral
verbs are used.

4.7.1 Participants
Forty-nine Mandarin native speakers participated over the internet. All had normal or corrected-to-
normal vision. None had participated in other experiments reported in this chapter.

168

4.7.2 Materials and design
The materials used here were modified from the stimuli in Experiments 5 and 6. In those experiments,
biased verbs are used. In this experiment, neutral verbs (e.g., ‘reread’) are used, which are not clearly
biased towards either the local or the non-local referent of the target sentence. (Note that since this
experiment does not aim for complete semantic neutralness, no norming on verb neutralness was
conducted.) However, due to the change of verbs, some nouns following the reflexive were adjusted to
yield natural sentence meanings. But the adjustment was kept to a minimum. Two factors, reflexive type
(ziji/ta-ziji) and logophoric role (source/perceiver), were manipulated. See (4.10) for a target example
set.

(4.10) a. Ziji/Source
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 gaosu 3 bieren 4
Xiaoming be class in DE excellent student lecture on he tell others
Wang jiaoshou 5 chongdu-le 6 ziji 7 de 8 xueshu 9 lunwen. 10
Wang professor reread-asp self DE academic paper
‘Xiaoming is an excellent student in the class. During the lecture, he told others that
Professor Wang had reread self’s academic paper.’
b. Ziji/Perceiver
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 tingshuo 3 Wang jiaoshou 5
Xiaoming be class in DE excellent student lecture on he hear Wang professor

169
chongdu-le 6 ziji 7 de 8 xueshu 9 lunwen. 10
reread-asp self DE academic paper
‘Xiaoming is an excellent student in the class. During the lecture, he heard that Professor
Wang had reread self’s academic paper.’
c. Ta-ziji/Source
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 gaosu 3 bieren 4
Xiaoming be class in DE excellent student lecture on he tell others
Wang jiaoshou 5 chongdu-le 6 ta-ziji 7 de 8 xueshu 9 lunwen. 10
Wang professor reread-asp he-self DE academic paper
‘Xiaoming is an excellent student in the class. During the lecture, he told others that
Professor Wang had reread he-self’s academic paper.’
d. Ta-ziji/Perceiver
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 tingshuo 3 Wang jiaoshou 5
Xiaoming be class in DE excellent student lecture on he hear Wang professor
chongdu-le 6 ta-ziji 7 de 8 xueshu 9 lunwen. 10
reread-asp he-self DE academic paper
‘Xiaoming is an excellent student in the class. During the lecture, he heard that Professor
Wang had reread he-self’s academic paper.’

Twenty sets of target sentences were mixed with 20 filler sentences and presented pseudo-
randomly to the participants. The fillers were identical to those in previous experiments.

170
Comprehension questions for the target sentences all probed the antecedent of the reflexive; questions
about the filler sentences were used as attention checkers.

4.7.3 Procedure
The procedure for Experiment 7 is identical to the procedure for the offline Experiments 5a and 6a.

4.7.4 Predictions
Based on the findings of Experiments 5a and 6a, a main effect of logophoric role is anticipated. There
should be stronger preferences for non-local binding under source roles compared to perceiver roles.
In addition, we have seen in Chapter 3 that long-distance binding is more likely with ziji in discourse-
topic-biasing contexts than with ta-ziji, a main effect of reflexive type should also appear. Finally, if ziji
and ta-ziji show different logophoricity effects, a reflexive x logophoric role interaction is predicted.

4.7.5 Data analysis
Contrasts were fitted reflexive type (ziji: +0.5, ta-ziji: -0.5) and logophoric role (source: +0.5, perceiver: -
0.5). All participants had filler trial comprehension accuracies over 75%, with a mean accuracy rate of
98%. Mixed-effect logistic regression was used to analyze antecedent choices using the glmer function.
The fixed effects are reflexive type and logophoric role; the random effects are intercepts and slopes
grouped by participants and items. The same model comparison procedures as described in the
previous sections were used.

171
4.7.6 Results
Figure 4.6 displays the participants’ proportions of non-local antecedent choices. The left panel A
shows the individual distribution of antecedent choices; the right panel B shows the population-level
mean antecedent choices. There are three major findings. First, the main effect of reflexive type (p =
0.002) suggests that Mandarin speakers are more likely to choose topically prominent non-local
antecedents with ziji than with ta-ziji. Second, the main effect of logophoric role (p < 0.001) indicates
that, regardless of the reflexive form, preference for non-local binding is stronger under the source
conditions relative to the perceiver conditions. Lastly, there is no significant interaction, suggesting that
ziji and ta-ziji do not differ in sensitivities to the Logophoric Hierarchy.
It is also worth pointing out that the antecedent choice patterns in this experiment match the real-
time binding patterns in Experiments 5b and 6b. Using the verb bias effect as a diagnostic for binding
preferences, we have seen that when in the perceiver conditions, ziji shows a non-local binding
preference and ta-ziji does not show any clear preference, which matches the offline antecedent choice
patterns in Figure 4.6: ziji shows a mean 66% non-local binding probability while ta-ziji shows a mean
52% non-local binding probability. Intriguingly, echoing Experiment 6b, the mean probability of non-
local binding in Experiment 7 with neutral verbs is 64%, above chance (intercept-only model: β = -0.94,
SE = 0.38, t = -2.45, p = 0.01), suggesting a non-local preference for ta-ziji when the matrix subject is both
topically and logophorically prominent.

172

Figure 4.6: Distribution of mean proportions of non-local readings of ziji or ta-ziji across all individual
participants (Panel A) and mean proportions of non-local readings (Panel B) at the group level in
Experiment 7.

Table 4.5: Summary of statistics for Experiment 7 (*: < 0.05).
β SE t-value p-value
Reflexive -0.62 0.20 -3.03 0.002*
Role 0.60 0.16 3.73 < 0.001*
Reflexive x role -0.27 0.32 -0.85 0.39

4.7.7 Discussion of Experiment 7
Experiment 7 shows that when the matrix subject is topically prominent, there is a stronger preference
for non-local binding with ziji than with ta-ziji. This may have to do with the different linguistic
properties of ziji and ta-ziji, alluded to in Chapter; ziji can have an empathic use while ta-ziji is most
probably non-empathic. Furthermore, Experiment 7 replicates the findings of Experiments 5 and 6 in
showing that logophoric roles of the antecedent impact the readings of ziji and ta-ziji.
Critically, there is no statistical evidence that ziji is more sensitive to the source role than ta-ziji,
which does not align well with the findings of Liu (2020). As mentioned in Section 4.2, Liu’s results
regarding a stronger logophoricity effect for ziji vs. ta-ziji can be ascribed not to the lower acceptability
0%
25%
50%
75%
100%
Ziji Ta−ziji
Non−local coreferences
A
66 %
72 %
52 %
64 %
0%
25%
50%
75%
100%
Ziji Ta−ziji
Non−local coreferences
Role
Source
Perceiver
B

173
of the baseline condition. The current experiment does not suffer from this limitation as the dependent
variable is not the acceptability of the sentences but rather antecedent choices. (In addition, an
unreported acceptability judgment study shows that acceptabilities of these four conditions are very
similar).

4.8 General discussion
Below, I summarize the findings of the experiments and discuss the broad implications that speak to
theoretical linguistic questions and the standard cue-based retrieval model.

4.8.1 Summary of results
The goal of Chapter 4 is two-fold. First, it aims to make a linguistic contribution to our understanding
of the exempt uses of ziji and ta-ziji. The focus is on whether the non-local binding of these reflexives is
constrained by the logophoric roles – source vs. perceiver – of the topical matrix subject. On top of this,
this study follows prior work (e.g., Liu, 2020; Xu & Runner, submitted) by directly comparing the
logophoricity effects associated with the two reflexives. Second, this chapter builds on Chapter 3 (on
discourse topic prominence) and tests whether logophoric information is immediately utilized for
antecedent retrieval. The psycholinguistic goal is to examine whether all kinds of discourse-level cues
by default are immediately accessed in real-time parsing. Below, I first summarize the findings bearing
on the linguistic questions before summarizing psycholinguistic findings.
The offline forced choice judgment experiments 5a and 6a show that Mandarin speakers prefer
non-local binding more when the topical matrix subject is a source than a perceiver. This is in line with

174
the Logophoric Hierarchy (Culy, 1994). This preference for source subjects has been replicated in
Experiment 7 where the local verbs are neutral. Importantly, the source advantage does not differ
between ziji and ta-ziji. But since null results in hypothesis testing do not allow us to make strong
conclusions, future studies should further examine this question by using more testing methods.
However, a caveat is in order regarding the logophoricity effect. Note that in this study, the matrix
subject is a discourse topic. There is no data in this study on whether source antecedents attract more
non-local readings when they are not topically prominent in the discourse context. Indeed, Xu and
Runner (submitted) did not find any logophoricity effect for ziji and ta-ziji in stand-alone sentences.
This may be because the logophoricity effect only appears when the non-local binding is plausible,
either when the local inanimate referent cannot be the antecedent (e.g., Kim & Yoon, 2009; Sloggett,
2017; Liu, 2020) or when the non-local antecedent is prominent in the discourse. Under this view, the
logophoric roles alone may not be strong cues to non-local binding.
The online self-paced reading results of Experiments 5b and 6b paint a more complicated picture.
The lack of a verb bias x logophoric role interaction in Experiment 5b on ziji points to an absence of the
logophoricity effect. However, as mentioned before, this may be due to the small effect size. An
interaction may be difficult to reach significance when the participants prefer non-local readings in
both perceiver and source conditions. A non-mutually exclusive alternative explanation is that the
logohoricity effect discovered in Experiment 5a mainly happens during re-reading, which cannot be
examined with the self-paced reading paradigm used in this work. Future work using bi-directional self-
paced reading or reading eye-tracking should look more into this question.

175
Notably, the RTs in Experiment 6b on ta-ziji does clearly point to a verb bias x logophoric role
interaction, meaning that the real-time processing ta-ziji is sensitive to the logophoric role of the topical
subject. But critically, the interactions are observed in the spillover regions. Compared to the early
discourse topicality effect at the reflexive region in Experiment 4b in Chapter 3, this delayed
logophoricity effect suggests that the logophoric information is not immediately accessed in the parsing
of ta-ziji.
To summarize, the experiments in this chapter has provided substantive evidence that the
interpretations of ziji and ta-ziji are sensitive to the Logophoric Hierarchy. However, the logophoric role
of the antecedent does not seem to play an early role in real-time resolution. In the next two sections, I
will discuss the implications of these findings in the context of linguistic and psycholinguistic theories.

4.8.2 Two uses of exempt reflexives or one?
This study has shown that the probability of non-local binding of ziji and ta-ziji is in line with the
Logophoric Hierarchy: a source topical referent is more likely to be the antecedent than a perceiver
topical antecedent. But there is a theoretical question that needs to be addressed: is there one exempt
use or two in Mandarin? Recall that Oshima (2007) divides the exempt uses of reflexives into two types:
the empathic use and the logophoric use. These two uses show different properties in languages such
as Japanese. For example, the monomorphemic reflexive zibun in Japanese shows different properties
depending on the discourse factors (e.g., Oshima, 2007; Nishigauchi, 2014). However, given the findings
of this study and prior studies, it seems that the logophoric and empathic properties do not exclude
each other, at least in Mandarin. For ease of reference, I call this approach the fusional view. As the

176
distinction between empathic and logophoric uses only applies to monomorphemic reflexives, here I
focus on the properties of ziji. There are two pieces of evidence supporting the fusional view.
First, the logophoricity effect is found only when long-distance binding is considered by
participants. In the case of ziji, non-local binding is licensed when the non-local antecedent is an
empathy locus (e.g., Pan, 1997, 2001; Huang & Liu, 2001; Wang & Pan, 2015a,b). When non-local binding
is dispreferred in stand-alone sentences (e.g., Dillon et al., 2014, 2016; Jäger et al., 2015; Lyu & Kaiser,
2021; Lyu et al., 2022) presumably due to a default syntactic use, no logophoricity effect has been found
(Xu & Runner, submitted). What this suggests is that, in the case of ziji, long-distance binding by
empathy loci co-occurs with the logophoricity effect. A non-empathized non-local subject does not
show the logophoricity effect.
Second, regardless of the logophoric role of the matrix subject, the blocking effect always occurs. In
Oshima’s (2007) tripartite division of the uses of reflexives, logophoric reflexives are immune to the
blocking effect while the empathic reflexives are always subject to 1st-person blocking. But since the
blocking effect necessarily obtains in Mandarin, shown below in (4.11), the logophoric use à la Oshima
and the empathic use cannot be distinguished in Mandarin. Thus, I follow Wang and Pan (2015a) and
submit that long-distance ziji shows the logophoric effect but is primarily an empathic reflexive.

(4.11) Xiaoming 1 {gaosu bieren/tingshuo} wo 2 xihuan ziji *1/2.
Xiaoming tell others heard I like self
‘Xiaoming 1 told others/heard that I 2 like self *1/2.’

177
However, there is one question that remains to be explained. Why is there a tendency for de se
reading of ziji (Pan, 1997; Huang & Lin, 2001; Anand, 2006)? Recall Oshima’s test in (4.4) that only
logophoric reflexives are read de se. First, the de se reading preference is consistent with the argument
that ziji is both empathic and logophoric. Second, it remains to be examined to what extent Oshima’s
de se test accurately captures the distinction between the logophoric and empathic uses. In fact,
Nishiguauchi (2014: 173) mentioned that “in general, a non-de-se reading of zibun is not easy to come
by” even though a non-de-se reading is possible (Oshima, 2004: 186), as is with ziji (Wang & Pan, 2014).
Perhaps more importantly, drawing on linguistic data, Oshima (2004: 185) found that the logophoric
reading of zibun is not free from the empathy constraints, which makes the fusional view even more
plausible, especially in the case of ziji.
In short, together with the findings in prior theoretical and experimental work, the results of this
study are in line with a fusional approach to exempt readings of ziji which shows properties of empathic
and logophoric reflexives. As to the exempt use of ta-ziji, it seems to be logophoric (given the discussion
in Chapter 3) when the non-local subject is topically prominent and a source.

4.8.3 Revisiting the standard retrieval model
In the previous chapter, we have seen that discourse topicality plays an early role in antecedent retrieval,
thus challenging the structure-based retrieval model which posits that only structural cues are
accessible at the early processing stages. However, this chapter shows that the standard cue-based
retrieval model over-predicts as it seems that not all discourse-level cues are immediately accessible. In

178
this study, the logophoric effect occurs at the spillover regions, suggesting that the logophoric role of
the topical antecedent plays a later role compared to discourse topicality.
That being said, there is another possibility which remains to be tested in future work. Prior studies
suggest that the effect observed at the spillover region may not necessarily be restricted to the later
processing stage. For example, Qian and Wu (2016) discovered that gender semantics of non-local
antecedents impacts the processing of ta-ziji in stand-alone sentences but only 4 words after the critical
region. This prompts the authors to conclude that semantic cues play a later role. However, in a reading
eye-tracking study which uses the same experimental design and includes the stimuli used in Qian and
Wu, Chang et al. (2020) discovered that the gender information of the non-local antecedent interferes
with the processing of ta-ziji: in first fixation times, gender match between the non-local antecedent
and ta-ziji leads to reading slowdowns when the local antecedent is not a good semantic candidate.
Together, what these two studies (i.e., Qian & Wu, 2016; Chang et al., 2020) suggest is that spillover
effects may not necessarily be late-stage effects, at least not in a categorical sentence. (The discourse
topicality effect may still precede the logophoricity effect in a relative sense.) One possibility for the
delayed response may be due to the weak associative strength between logophoric roles and (non-local)
antecedent retrieval. Indeed, this characterization fits the linguistic speculation discussed earlier that
ziji is primarily empathic and weakly logophoric (considering the absolute blocking effect in Mandarin
vis-à-vis Japanese). Since antecedent retrieval operates on linguistic properties, it should come as no
surprise that logophoric roles are not strong cues to antecedent retrieval in real-time parsing. Future
work using eye-tracking methodology should look more into this possibility.

179
Chapter 5: The blocking effect and reflexive resolution

5.1 Introduction
In the previous chapters, the experimental data suggest that long-distance ziji can be empathic. In other
words, it can be bound by a non-local sentence-internal empathy locus. This property of ziji predicts
that a sentence-external empathy locus (e.g., the speaker/comprehender) should interfere with long-
distance binding. Indeed, as introduced in Chapter 1, a 1
st
-/2
nd
-person pronoun in a local domain which
encodes the perspective of the speaker blocks long-distance binding (e.g., ‘Xiaoming heard I had reread
ziji’s paper), because people by default take their own perspectives than others under the Speech Act
Empathy Hierarchy (Kuno & Kaburaki, 1977; Kuno, 1987). This phenomenon is called the blocking effect.
But recall that in Chapter 1, two linguistic questions regarding the blocking effect were not yet
satisfactorily resolved. First, different linguistic approaches make different predictions regarding the
object blocking effect – i.e., whether 1
st
or 2
nd
person pronouns realized in object position can trigger
blocking (e.g., ‘Xiaoming told me Professor Wang had reread ziji’s paper’). As for the two syntactic
approaches, the agreement-based account (e.g., Cole et al., 1990; Huang & Tang, 1991; Cole & Sung, 1994)
predicts no object blocking effect, while the contiguous-agreement account (Giblin, 2016) predicts a
strong object blocking effect. For the other two approaches that acknowledge the role of discourse, the
early discourse-based account (e.g., Huang et al., 1984; Pollard & Xue, 1998; Pan, 1997, 2001; Huang & Liu,
2001; Huang et al., 2009; Wang & Pan, 2015a) also predicts a strong object blocking effect, but the unified
account (Cole et al., 2006; Charnavel et al., 2017) which incorporates syntactic and discourse
components predicts a weak object blocking effect. In fact, there is no prior experimental study on the

180
object blocking effect. This study aims to fill this gap and examine these four different theoretical
accounts of the blocking effect.
Second, there is no consensus in the theoretical literature about whether verb semantics –
specifically, verb directedness – overrides the blocking effect. (For a discussion of verb directedness
relating to reflexive resolution, see Chapter 3.) Different researchers seem to have murky judgments
(e.g., Yu, 1992; Pollard & Xue, 1998; Huang, Y., 2000), which constitutes another empirical question that
this study aims to clarify.
In addition to contributing to theoretical linguistic debates, this chapter further explores memory-
based retrieval models from a psycholinguistic perspective. The previous two chapters have shown that
the structure-based retrieval model (e.g., Sturt, 2003; Van Dyke, 2007; Van Dyke & McElree, 2011; Dillon
et al., 2013; Parker & Philips, 2017) seems untenable due to the immediate accessibility of discourse-
level cues (Chapter 3: e.g., [+POV], [+Top]). However, the standard cue-based retrieval model (e.g.,
Lewis & Vasishth, 2005; Vasisth et al., 2008; Chen et al., 2012; Patil et al., 2016) is not completely
compatible with the data either (Chapter 4: logophoric roles); it seems too strong as not all discourse-
level cues are immediately accessible to guide antecedent retrieval. In this chapter, I examine a new
question which may further restrict the standard cue-based retrieval model: during real-time
processing, is the POV cue, which has been so far observed to play an early role in reflexive resolution,
constrained by the syntactic position in which it appears? Self-paced reading experiments will be used
to explore this question.
This chapter is structured as follows. Section 5.2 reviews the theoretical linguistic accounts of the
blocking effect. Section 5.3 proposes a more restricted version of the standard cue-based retrieval model,

181
which I call the structure-mediated cue-based retrieval mode and is to be distinguished from the
structure-based retrieval model. Section 5.4 is an overview of the aims of this chapter, followed by
Section 5.5 where I report two sets of experiments on subject and object blocking effects respectively.
Finally, I summarize the findings and propose a conjunction constraint on antecedent retrieval in
Section 5.6.

5.2 Overview of the linguistic accounts of the blocking effect
This section briefly reviews the four linguistic accounts which were introduced earlier in Chapter 1.
Here, I only summarize and highlights the linguistic operations that yield the subject and object
blocking effect.
The agreement-based account (e.g., Cole et al., 1990; Huang & Tang, 1991; Cole & Sung, 1994; Cole &
Wang, 1996) derives the blocking effect from phi-feature (here, person) agreement. It is postulated that
ziji cyclically adjoins to the local T-head in Logical Form (LF). The T-head inherits the base-generated
person feature (e.g., [+3 person]) from ziji. If the local subject is a 1
st
-person pronoun, agreement
between the local subject ([+1 person]) and the T head ([+3 person]) will crash, hence the blocking
effect. (An alternative account is that the local DP such as a 1
st
-person pronoun passes its phi-feature to
ziji which is unlabeled at its base position. Long-distance agreement between the 1
st
-person marked ziji
and the non-local 3
rd
-person subject cannot proceed, due to person feature mismatch.) However, this
account predicts that if an object is a 1
st
-person pronoun, the blocking effect should not occur because
only subjects agree with T heads. Objects are not in a structural position to interfere with long-distance
binding between ziji and subjects.

182
However, the discovery of the object blocking effect (Xue et al., 1994; Pollard & Xue, 1998
20
) – which
supposes that long-distance binding is unavailable at all – resuscitates an earlier discourse-based
account (Huang et al., 1984). For example, a sentence like ‘Zhangsan 1 said I 2 like ziji *1/2’ does not allow
long-distance binding because the indirect speech ‘I like ziji’ is arguably derived from a direct speech ‘I
like me’ (Huang et al., 1984). As there are two instances of 1
st
-person pronouns – ‘I’ and ‘me’ –
representing the perspectives of the speaker and the agent of the indirect speech (i.e., Zhangsan)
respectively, there is a clash of perspectives. Since the 1
st
-person perspective is prioritized over the 3
rd
-
person perspective, local binding by the external empathy locus (i.e., 1
st
-person pronoun) is preferred.
Pan and colleagues (Pan, 1997, 2001; Wang & Pan, 2015a,b) implement the blocking effect somewhat
differently in technical details (see Huang & Liu, 2001: 182-184 for discussion) but hold a similar view
regarding the crucial role of perspective-taking/empathy. One advantage of this approach is that the
object blocking effect can be explained in terms of perspective clash; regardless of the syntactic position
of the blocker, long-distance binding of ziji is not allowed when an external empathy locus is present.
In contrast to the discourse-based account, Giblin (2016) proposes a syntactic explanation, the
contiguous-agreement account, that can also capture the object blocking effect. The key component of
this account is that an unvalued [uparticipant] feature carried by the sentence-peripheral C head
probes down to the contiguous DP to be valued (see ex.(1.12) in Chapter 1). If the contiguous DP carries
a [+participant] feature – here, only 1
st
-/2
nd
-person pronouns – the C head will be valued. Once valued,
this [+participant] feature will be passed down to all the T heads on the syntactic tree. Finally, ziji can

20
Pollard and Xue (1998) only suggest that a pragmatic account could better account for the blocking effect as
this is beyond the scope of their work. This account should hinge on the “contextual prominence of the discourse
participants” (Pollard & Xue, 1998: 301).

183
be valued with [+participant] in a T-V-reflexive chain and is co-indexed with the 1
st
-/2
nd
-person pronoun
with [+participant] in the matrix subject position. This process can yield long-distance binding by a 1
st
-
/2
nd
-person pronoun. However, if the 1
st
-/2
nd
-person pronoun with the [+participant] feature is not
contiguous with the unvalued C head, agreement can only occur in a non-contiguous manner by
crossing a 3
rd
-person matrix subject which is [-participant]. This results in the violation of the
Contiguous Agree Constraint (Nevins, 2007). The blocking effect thus appears. Note that this account
holds regardless of the subject/object position of the blocker (also see ex.(1.13) in Chapter 1).
Both the discourse-based account and the contiguous-agreement account predict the strength of
the subject and object blocking effects to be similar. But Cole et al. (2006) noticed that the object
blocking effect may be weaker (see ex.(1.14) in Chapter 1). Based on this observation, the unified account
(Cole et al., 2006; Charnavel et al., 2017) proposes that both agreement-based blocking and perspective
blocking may be involved: the weaker object blocking effect is due to the violation of the discourse
empathy constraint but not the violation of syntax (as an object blocker is not in a subject position).
Note that the agreement-based blocking here refers to the cyclic agreement account, not the
contiguous-agreement account, because the latter predicts a strong blocking effect with object blockers.
However, the examples used by Cole et al. (2006) to illustrate the weaker object blocking effect are
potentially confounded with the number of blockers in their example sentence, which calls for a more
systematic investigation.
In summary, the four different accounts of the blocking effect make three types of predictions for
the object blocking effect, which is summarized in Chapter 1 but repeated here in Table 5.1. The
agreement-based account predicts no object blocking effect; the discourse-based and the contiguous-

184
agreement accounts predict strong object blocking effect; the unified account predicts a weak object
blocking effect.
Table 5.1. Predictions for the blocking effect with different approaches.
Agreement-
based account
Contiguous-
agreement account
Discourse-based
account
Unified account
Subject blocking Strong Strong Strong Strong
Object blocking None Strong Strong Weak

5.3 Standard and structure-mediated cue-based retrieval models
The standard cue-based retrieval model assumes that cues at different levels, including discourse-level
cues, can impact reflexive resolution at the early processing stage. Indeed, we have seen in Chapter 3
that empathy-related discourse topicality can immediately modulate the reading times of ziji even
though syntactically the discourse topic is outside the local domain.
So far, we have been considering the empathy loci that are realized in subject position. What
happens when an empathy locus is in a syntactically less prominent object position? For the purpose
of this chapter, let us consider the 1
st
-person object blocker which is anchored to the
speaker/comprehender of the sentence. If the accessibility of an empathy locus is mediated by its
syntactic position, we may see delayed or even no object blocking effect during real-time reflexive
resolution. (This is separate from the open question of whether object blockers induce the blocking
effect at all in offline judgments). To distinguish this hypothesis from the structure-based retrieval
model which denies the early accessibility of non-local, discourse-prominent antecedents (e.g., Sturt,
2003; Cunnings & Sturt, 2014), I call this updated model the structure-mediated cue-based retrieval
model. The idea is that non-syntactic cues can be immediately utilized by the language parser, but
antecedents encoded with these non-syntactic cues must be in a syntactically prominent position;

185
otherwise, they may not be accessed at the early processing stage. In a way, this new model can be
viewed as a combination of the standard cue-based and the structure-based retrieval models.
The blocking effect provides a good testing ground for examining the standard and structure-
mediated cue-based retrieval models because both syntactic and perspective-related cues are relevant:
the subject vs. object position of the blocker can theoretically impact the strength of the blocking effect,
at least according to the agreement-based and the unified accounts. (Of course, this is an empirical
question, which I test in the offline studies described below.) If both syntactic and perspective-related
factors indeed modulate the blocking effect, the standard and structure-mediated cue-based retrieval
models would make different predictions regarding the object blocking effect.
The standard model predicts that a sentence-external empathy locus (e.g., 1
st
-person blocker) –
notated with [+POV] short for “point-of-view” – should interfere with long-distance binding by a
sentence-internal empathy locus (i.e., 3
rd
-person discourse topic subject). When the external empathy
locus is a subject (i.e., [+SUB, +POV]), local binding should be preferred as soon as ziji is encountered,
because the subject blocker ‘I’ is closer to ziji and stronger than the internal empathy locus (Kuno,1987).
Importantly, as to the object blocking effect during online processing, the standard model predicts that
(i) the blocking effect should occur immediately regardless of the syntactic position of the blocker due
to partial [+POV] feature match and (ii) the object blocking effect should be weaker – the object
blocking is [-SUB] – but not absent in reflexive processing.
In contrast, the structure-mediated cue-based retrieval model predicts immediate blocking effects
by subject blockers only. Object blockers are not predicted to interfere with long-distance binding by
the discourse topic subject at the early processing stage for two reasons. First, ziji searches for subjects

186
(i.e., subject orientation). Second, object blockers are not in a syntactic position which triggers the
agreement-based blocking effect (e.g., Cole et al., 1990, 2006). Note that this model’s predictions only
apply to the early processing stage. Whether the object blocking emerges at the later processing stages
is beyond its predictions. Figure 5.1 illustrates the different predictions of these two processing models.

Figure 5.1. Predictions of the standard and structure-mediated cue-based retrieval models for the early-
stage blocking effect.

5.4 Overview of the aims
This study has three overarching goals, two linguistic goals and a psycholinguistic one.
First, this study aims to examine the four different accounts of the blocking effect using offline
forced choice judgment tasks. The specific question is whether object blockers weaken long-distance
binding and, if so, how strong the object blocking effect is compared to the subject blocking effect.
Second, using the same offline experiment, this study additionally tests whether verb directedness
(i.e., self- vs. other-directed verbs) can override the blocking effect, which is debated in the theoretical

187
literature (e.g., Yu, 1992; Pollard & Xue, 1998; Huang, Y., 2000). Clarification on the interaction of verb
semantics and the blocking effect will enrich our understanding of this phenomenon.
Finally, using self-paced reading tasks, this study compares the two psycholinguistic models on
antecedent retrieval, the standard and the structure-mediated cue-based retrieval models. The former
predicts early subject and object blocking effects while the latter only predicts a subject blocking effect.
This chapter presents two sets of experiments. Each set is composed of an offline forced choice
judgment experiment and one or more self-paced reading experiments. Experiment 1 investigates the
subject blocking effect; Experiment 2 investigates the object blocking effect.

5.5 Experiment 8: Subject blocking effect
Experiment 8 is composed of two tasks, an offline forced choice task (Experiment 8a) and a self-paced
reading task (Experiment 8b). The offline task explores how verb directedness interacts with the local
external empathy locus (i.e., ‘I’), the blocker, in the subject position. The online self-paced reading task
uses the verb bias effect as a diagnostic to test Mandarin speakers’ binding preferences as soon as they
encounter ziji in incremental reading, with the focus being whether and when the subject blocker
interferes with non-local binding.

5.5.1 Materials and design
The experimental stimuli in Experiment 8a and 8b are the same. The stimuli were modified from those
in Chapter 3. In this study, the context sentence introduces a character which is the topical subject –
the empathy locus – of the subsequent critical sentence. Two factors were manipulated, the blocker type

188
(blocker/non-blocker) of the local referent and verb bias (self-directed/other-directed). The subject
blocker is the external empathy locus ‘I’; the subject non-blocker is a 3
rd
-person referent (e.g., ‘Prof.
Wang’). As in previous studies, the verb bias factor tests the impact of verb meaning on antecedent
choice in the offline judgment task and is a diagnostic to test participants’ real-time interpretation of
the reflexive. When encountering ziji, if participants show reading slowdowns in sentences with other-
directed verbs, this indicates that they attempt local binding; if reading slowdowns occur in sentences
with self-directed verbs, this means participants prefer non-local binding. See (5.1) for an example set
of target sentences.

(5.1) a. Blocker/Self-directed
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 tingshuo 3 wo 4 ganggang 5
Xiaoming be class in DE excellent student lecture on he hear I just
fabiao-le 6 ziji 7 de 8 xueshu 9 lunwen. 10
publish-ASP self DE academic paper
‘Xiaoming is an excellent student in the class. During the lecture, he heard that I had just
published self’s academic paper.’
b. Blocker/Other-directed
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 tingshuo 3 wo 4 ganggang 5
Xiaoming be class in DE excellent student lecture on he hear I just
pigai-le 6 ziji 7 de 8 xueshu 9 lunwen. 10
grade-ASP self DE academic paper

189
‘Xiaoming is an excellent student in the class. During the lecture, he heard that I had just
graded self’s academic paper.’
c. Non-blocker/Self-directed
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 tingshuo 3 Wang jiaoshou 4
Xiaoming be class in DE excellent student lecture on he hear Prof. Wang
ganggang 5 fabiao-le 6 ziji 7 de 8 xueshu 9 lunwen. 10
just publish-ASP self DE academic paper
‘Xiaoming is an excellent student in the class. During the lecture, he heard that Professor
Wang had just published self’s academic paper.’
d. Non-blocker/Other-directed
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 tingshuo 3 Wang jiaoshou 4
Xiaoming be class in DE excellent student lecture on he hear Prof. Wang
ganggang 5 pigai-le 6 ziji 7 de 8 xueshu 9 lunwen. 10
just grade-ASP self DE academic paper
‘Xiaoming is an excellent student in the class. During the lecture, he heard that Professor
Wang had just graded self’s academic paper.’

Twenty sets of target sentences were interleaved with 20 filler sentences and presented to the
participants pseudo-randomly. The types of the comprehension questions on target and filler trials are
identical to those in Chapters 3 and 4. Twenty factually (in)correct filler questions on filler trials were
used to screen out participants who did not pay close attention to the offline experiment. In the self-

190
paced reading experiment, an additional 8 comprehension questions that did not target the
interpretation of the reflexive were also used as attention checkers. (Recall that on target trials in the
self-paced reading experiment, 12 questions probe the interpretation of the reflexive as a secondary
measure of participants’ interpretive preferences.) The subscripts in (5.1) indicate the word regions of
the critical sentence. The critical region is the reflexive ziji.

5.5.2 Experiment 8a: Forced choice judgment
5.5.2.1 Participants
Fifty adult Mandarin speakers from mainland China participated over the internet. All participants
had normal or corrected-to-normal vision and had not participated in other experiments reported in
this study.

5.5.2.2 Procedure
Participants participated through Qualtrics (Qualtrics, Provo, UT). The context sentence and the
subsequent critical sentence are presented on separate screens. For all the target trials in this offline
experiment, participants answered a two-alternative comprehension question after finishing the
critical sentence (e.g., ‘who wrote an academic paper?’). The purpose of the question is to probe
participants’ interpretation of ziji. The order of the antecedent choices is balanced.

5.5.2.3 Predictions

191
We predict a main effect of verb bias, given that self-directed verbs should lead to more local
interpretations compared to other-directed verbs. A main effect of blocker type is also expected: When
the local subject is an external empathy locus ‘I,’ we expect more local choices than when the local
subject is not an empathy locus (e.g., ‘Prof. Wang’). Importantly, if verb directedness overrides the
blocking effect, in the blocker/other-directed conditions, we expect to see a significant non-local choice
preference.

5.2.2.4 Data analysis
Before data analysis, participants’ mean accuracies on filler questions were analyzed. A threshold of 75%
accuracy was applied. All participants passed the screening and had 92% mean accuracies. The factors
blocker type and verb bias were sum-coded in R (blocker: +0.5, non-blocker: -0.5; self-directed verb: +0.5,
other-directed verb: -0.5). Mixed-effect logistic regression was run to analyze participants’ binary
antecedent choices. The fixed effects are the two independent variables; the random effects are the
intercepts and slopes grouped by participants and items. If a complex model does not differ
significantly from a simpler model in model comparisons, the simpler model is used (Bates et al., 2015).

5.2.2.5 Results
Figure 5.2 shows participants’ antecedent choices. The left panel shows the distribution of each
participant’s mean preference for local antecedents; the right panel shows the entire population
sample’s mean preference for local antecedents. See Table 5.2 for statistics.

192

Figure 5.2: Distribution of mean proportions of local readings of ziji across all individual participants
(Panel A) and mean proportions of non-local readings (Panel B) at the group level in Experiment 8a.

Table 5.2: Summary of statistics for Experiment 8a (*: < 0.05).
β SE t-value p-value
Blocker 0.78 0.16 4.89 < 0.001*
Verb 2.03 0.23 9.01 < 0.001*
Blocker x verb -1.67 0.32 -5.26 < 0.001*

The factors blocker type and verb bias are, as expected, significant (p’s < 0.001), which suggests that
overall subject blockers make Mandarin speakers more likely to consider local binding than non-
blockers and that self-directed verbs lead to stronger local binding preference compared to other-
directed verbs. In addition, the blocker type x verb bias interaction is significant (p < 0.001). Pairwise
comparisons show that the blocking effect only appears in the other-directed conditions (β = 1.65, SE =
0.24, t = 6.96, p < 0.001) but not in the self-directed verb conditions (β = -0.04, SE = 0.22, t = -0.21, p =
0.84). This interaction is unexpected. One possible explanation is that since most of the choices are
non-local when the embedded verbs are self-directed, there is not too much space for the blocking
effect to show up. But it should be noted that the sizable blocking effect in the other-directed verb
0%
25%
50%
75%
100%
Self−directed Other−directed
Local coreference
A
46 %
18 %
68 %
70 %
0%
25%
50%
75%
100%
Self−directed Other−directed
Local coreference
Blocker
Blocker
Non−blocker
B

193
conditions confirms the judgment in the theoretical literature that 1
st
-person pronouns interfere with
(although not in a categorical way) non-local binding.
However, as regards the crucial question of how verb directedness interacts with blocking, the
results in Figure 5.2 seem to suggest that verb directedness and blocking compete. Neither overrides
the other. On the one hand, the 1
st
-person blocker does not outrank verb semantics because between
the two blocker conditions (self-directed/blocker and other-directed/blocker conditions), verb
directedness modulates the antecedent choice patterns (β = -1.26, SE = 0.22, t = -5.61, p < 0.001). On the
other hand, as we have seen, verb directedness does not override the blocker as shown by the main
effect of blocker type. In fact, when verb meaning and the blocker pull in different directions, as in the
other-directed/blocker condition, there is no clear binding preference (46% local choice on average).
Overall, this experiment confirms the judgment that 1
st
-person pronouns can block long-distance
binding. However, the blocking effect is not absolute and interacts with verb semantics. In the next
section, I turn to how the blocking effect plays out during incremental sentence processing.

5.5.3 Experiment 8b: Self-paced reading
5.5.3.1 Participants
Forty-eight adult Mandarin native speakers from mainland China participated over the internet. No
participant took part in any other experiment reported in this study.

5.5.3.2 Procedure

194
This experiment used the same materials as in Experiment 8a and was run on Ibex Farm (Drummond,
2013). On each trial, the participants read a context sentence introducing a discourse topic (e.g.,
Xiaoming). The context sentence was displayed as a whole. Participants then read a subsequent critical
sentence about the discourse topic region by region. This sentence was followed by a two-alternative
comprehension question. The order of the answer choices was randomized.

5.5.3.3 Predictions
Using the verb bias effect as a diagnostic to test participants’ online binding preference, we expect that
at or even before the critical reflexive region (see Experiment 3b in Chapter 3), participants should
prefer non-local binding by the sentence-internal empathy locus (i.e., discourse topic) in the non-
blocker conditions. If this is the case, self-directed verbs should lead to reading slowdowns at the biased
verb or reflexive region, relative to other-directed verbs. In contrast, when the local referent is a
sentence-external empathy locus (i.e., ‘I’), both standard and structure-mediated cue-based retrieval
models predict a strong interference effect (see Figure 5.1). This is because the 1
st
-person blocker is in a
syntactically prominent position and constitutes a full match of the retrieval cues (i.e., [+POV, +SUB]).
Thus, we should anticipate other-directed verbs to cause reading slowdowns at or even before the
critical region.

5.5.3.4 Data analysis
All participants had accuracy rates above 75%. The mean accuracy rate for the entire group is 94%. RTs
outside the 100ms and 3000ms range were removed as outliers. So were RTs 2.5 SDs above the mean RT

195
by region and by condition. This affected 3.37% of the original RTs. Mixed-effect linear regressions were
run over log-transformed RTs and raw RTs. Below, I only report results on log-transformed RTs. Raw RT
results will also be reported when discrepancies exist between log-transformed and raw RT analyses.

5.5.3.5 Results
The participants’ mean RTs of the critical sentence by region and by condition are plotted in Figure 5.3.
For statistics, see Table 5.3. The critical region is Region 7, ziji. Before the critical region, Region 4 (e.g.,
‘Prof. Wang/I’) shows a main effect of blocker type (p = 0.008), which I attribute to the fact that the 3
rd
-
person referent (e.g., ‘Prof. Wang’) is syllabically longer and referentially more complex than the 1
st
-
person pronoun wo (‘I’). The same main effect persists at Region 5 (e.g., ‘just’). At Region 6, the biased
verb, there are no effects related to verb bias, i.e., no main effect of verb bias and no blocker type x verb
bias interaction. There is, however, a main effect of blocker type, which, unlike the previous two regions,
is due to longer RTs associated with the blocker conditions. This reading slowdown, which spills to the
critical region, is presumably attributable to the switch of perspective from the sentence-internal
empathy locus, the 3
rd
-person discourse topic, to the external empathy locus, ‘I’.

196

Figure 5.3: Mean RTs (ms) across all regions in the critical sentence in Experiment 8b.

At the critical region ziji, there is no effect of verb bias, no blocker type x verb bias interaction, but
there is a main effect of blocker type, as mentioned above. At the first spillover region (i.e., DE), the
blocker type x verb bias interaction is significant (p = 0.001), indicating that the verb bias effect differs
in the blocker and non-blocker conditions. When the local referent is a 1
st
-person blocker, other-
directed verbs lead to significant reading slowdowns (β = 0.08, SE = 0.02, t = 3.26, p = 0.001). This means
that participants prefer local binding when a blocker intervenes. In contrast, when the local referent is
a 3
rd
-person DP, self-directed verbs elicit reading slowdowns (marginal in log-transformed RT analysis:
β = -0.04, SE = 0.02, -1.70, p = 0.09; significant in raw RT analysis: β = -24.86, SE = 11.72, t = -2.12, p < 0.05),
meaning that participants prefer non-local binding by a discourse topic when no external empathy
locus intervenes.

verb ziji
500
750
1000
1250
In lecture, he heard Prof.Wang/I just verbed ziji DE academic paper.
RT(ms)
Blocker Type
Blocker
Non−blocker
Verb
Self−directed
Other−directed

197

198
Overall, what RTs in Experiment 8b suggest is that perspective-taking impacts real-time reflexive
resolution and that 1
st
-person empathy locus is stronger than a 3
rd
-person empathy locus. Below, I
discuss the results of the Experiments 8a and 8b.

5.5.4 Discussion of Experiments 8a and 8b
Experiment 8 examines the classic subject blocking effect (e.g., Huang & Liu, 2001) – the impact of the
1
st
-person pronoun in subject position on the interpretation of its clausemate ziji – from linguistic and
psycholinguistic perspectives. Experiment 8a uses the offline forced choice judgment task to explore
how the blocking effect interacts with verb semantics. Experiment 8b approaches the subject blocking
effect from a real-time processing angle; it tests participants’ binding preferences the moment they see
ziji and the time-course of the subject blocking effect. Below, I discuss the main findings.
Experiment 8a confirms the subject blocking effect. But critically, the subject blocking effect is not
absolute, contrary to what is sometimes assumed in prior theoretically-oriented work (e.g., Pollard &
Xue, 1998: 301; Huang & Liu, 2001: 142). Blocking turns out to be only one of the factors that determine
the binding directions of ziji. As we have seen, in offline antecedent choices, verb directedness plays as
strong a role as blocking. Indeed, when verb directedness and the blocker bias different referents,
participants did not show any clear binding preference for local or non-local antecedents at the final
stages when they need to weigh different types of information. Nevertheless, Experiment 8a confirms
that an external empathy locus interferes with long-distance binding of perspective-sensitive ziji and
provides a foundation for the online experiment, Experiment 8b.

199
Experiment 8b is concerned with how and when the subject blocker interferes with non-local
binding of ziji. As for the question of “how”, the verb bias effects suggest that Mandarin speakers prefer
non-local binding by a sentence-external empathy locus, the discourse topic, when there is no subject
blocker but local binding when a local subject blocker appears. This aligns with the predictions of the
standard cue-based retrieval model and the structure-mediated cue-based retrieval models proposed
in Section 5.3, but not with the unfocused structured-based retrieval model discussed in Chapter 3.
But Experiment 8b does not straightforwardly answer the question of “when” the subject blocking
effect occurs. The two processing models this chapter focuses on predict an immediate blocking effect
at or even before the reflexive. However, the subject blocking effect and the topicality effect only
appeared at the spillover regions. While it is unclear why the verb bias effects are delayed, this is not
surprising for self-paced reading studies which tax participants’ cognitive resources and can lead to
delayed reaction (see Badecker & Straub, 2002, for evidence that the anticipated effects can shift in
regions in self-paced reading). Here, I argue that the delayed empathy-related effects in the non-blocker
and blocker conditions should not concern us excessively.
In fact, there is reason to believe that the blocking effect may be an early effect. This is because the
discourse-topicality effect (i.e., binding by sentence-internal empathy loci) occurs at the same time as
the subject blocking effect (i.e., binding by sentence-external empathy loci). We have seen in the studies
in the previous two chapters and, as we shall see below, in Experiment 9 as well, that this discourse-
topicality effect appears quite early (at the pre-critical biased verb or the reflexive region) and is
convincingly an early effect. Thus, the subject blocking effect which occurs at the same processing stage
as the topicality effect is arguably also an early effect. However, this speculation remains to be tested in

200
future studies using other research methods that are cognitively less challenging (e.g., eye-tracking
during normal reading).

5.6 Experiment 9: Object blocking effect
Experiment 9 tests the blocking effect when the blocker is in a syntactically less prominent object
position. It includes an offline forced choice task (Experiment 9a) and two self-paced reading
experiments (Experiment 9b, 9c). The offline Experiment 9a examines whether object blockers are
effective at all in interfering with non-local binding. The results of Experiment 9a will also be compared
to the subject blocking effect in Experiment 8a. The online Experiments 9b and 9c examine whether
and when the object blocking effect appears and can shed light on the standard and structure-mediated
cue-based retrieval models.

5.6.1 Materials and design
Experiment 9 uses the same design as Experiment 8. There are two factors, blocker type (blocker/non-
blocker) and verb bias (self-directed/other-directed). In the blocker conditions, the 1
st
-person pronoun
wo is in an object position; in the non-blocker conditions, wo is replaced by referentially simple words
such as disyllabic bieren (‘others’) and dajia (‘everyone’). The matrix verb is always a source verb gaosu
(‘told’) to subcategorize for the object blocker wo, unlike Experiment 8 where most of the verbs are
perceiver verbs like tingshuo (‘hear’). The manipulation of verb directedness serves the same purpose
as in Experiment 8. Experiment 9b only differs from Experiment 9c in not having the adverbial (e.g.,
‘just’). See (5.2) for an example set.

201

(5.2) a. Blocker/Self-directed
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 gaosu 3 wo 4 Wang jiaoshou
Xiaoming be class in DE excellent student lecture on he tell me Prof. Wang
ganggang 5 fabiao-le 6 ziji 7 de 8 xueshu 9 lunwen. 10
just publish-ASP self DE academic paper
‘Xiaoming is an excellent student in the class. During the lecture, he told me that Prof. Wang
had just published self’s academic paper.’
b. Blocker/Other-directed
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 gaosu 3 wo 4 Wang jiaoshou
Xiaoming be class in DE excellent student lecture on he tell me Prof. Wang
ganggang 5 pigai-le 6 ziji 7 de 8 xueshu 9 lunwen. 10
just grade-ASP self DE academic paper
‘Xiaoming is an excellent student in the class. During the lecture, he told me that Prof. Wang
had just graded self’s academic paper.’
c. Non-blocker/Self-directed
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 gaosu 3 bieren 4
Xiaoming be class in DE excellent student lecture on he tell others
Wang jiaoshou 5 ganggang 6 fabiao-le 7 ziji 8 de 9 xueshu 10 lunwen. 11
Prof. Wang just publish-ASP self DE academic paper
‘Xiaoming is an excellent student in the class. During the lecture, he told others that Prof.

202
Wang had just published self’s academic paper.’
d. Non-blocker/Other-directed
Xiaoming shi banji li de youxiu xuesheng. Ke shang, 1 ta 2 gaosu 3 bieren 4
Xiaoming be class in DE excellent student lecture on he tell others
Wang jiaoshou 5 ganggang 6 pigai-le 6 ziji 7 de 8 xueshu 9 lunwen. 10
Prof. Wang just grade-ASP self DE academic paper
‘Xiaoming is an excellent student in the class. During the lecture, he told others that Prof.
Wang had just graded self’s academic paper.’

Twenty sets of target items plus 20 filler sentences were interleaved and presented pseudo-
randomly to the participants. The target sentences were modified based on the sentences in
Experiment 8. The filler sentences are the same as in Experiment 8. The subscripts in (5.2) indicate each
region presented in the self-paced reading experiment. Experiment 9c does not have the adverbial
region (e.g., ‘just’).

5.6.2 Experiment 9a: Forced choice judgment
5.6.2.1 Participants
Forty-eight adult native speakers of Mandarin with normal or corrected-to-normal vision participated
over the internet. None had participated in any other experiment in this study.

5.6.2.2 Procedure

203
The same procedure described in Section 5.5.2.2 was followed.

5.6.2.3 Predictions
Experiment 9a aims to test the strength of the object blocking effect and evaluates the empirical basis
of the four blocking accounts. Recall from the review Section 5.2 that these four accounts make three
types of predictions. The agreement-based account predicts no blocking effect, which means an
absence of the main effect of blocker type. In contrast, the other three accounts all predict a main effect
of blocker type. The contiguous-agreement and the discourse-based accounts predict the object
blocking effect to be similar to the subject blocking effect. The unified account predicts the object
blocking effect to be weaker compared to the subject blocker effect. To compare the subject and object
blocking effects, I will conduct a between-experiment analysis by including the blocker position as a
between-participant factor (see below). Finally, to be complete in predictions, a main effect of verb bias
is also anticipated.

5.6.2.4 Data analysis
Data analysis is the same as described in 5.5.2.4.

5.6.2.5 Results
Figure 5.4 shows the distribution of the participants’ antecedent choices (left panel A) and their mean
preference for local antecedents (right panel B). See Table 5.4 for a summary of statistics.

204

Figure 5.4: Distribution of mean proportions of local readings of ziji across all individual participants
(Panel A) and mean proportions of non-local readings (Panel B) at the group level in Experiment 9a.

Table 5.4: Summary of statistics for Experiment 9a (*: < 0.05).
β SE t-value p-value
Blocker 0.56 0.15 3.82 < 0.001*
Verb 1.69 0.15 11.16 < 0.001*
Blocker x verb -0.23 0.29 -0.81 0.42

As expected, the mixed-effect logistic model shows a main effect of verb bias (p < 0.001), as self-
directed verbs are more likely to lead to local binding preferences relative to other-directed verbs.
Importantly, the factor blocker type also reaches significance (p < 0.001), suggesting that object blockers
impact long-distance binding of ziji. This finding is incompatible with the prediction of the agreement-
based account. The blocker type x verb bias interaction is not significant.

Comparison of subject and object blocking effects
Experiment 9a alone does not allow us to tease apart the three other accounts of the blocking effect. To
this end, I compare the subject vs. object blocking effect in Experiment 8a and 9a. The factor of blocker
0%
25%
50%
75%
100%
Self−directed Other−directed
Local coreference
A
31 %
19 %
62 %
54 %
0%
25%
50%
75%
100%
Self−directed Other−directed
Local coreference
Blocker
Blocker
Non−blocker
B

205
position is included (contrast coding: object = 0.5, subject = -0.5) in a mixed-effect statistic model. The
summary of statistics is shown in Table 5.5.

Table 5.5: Summary of statistics for between-experiment comparison (*: < 0.05).

β SE t-value p-value
Blocker 0.65 0.11 6.13 < 0.001*
Verb 1.80 0.11 16.34 < 0.001*
Position -0.46 0.16 -2.81 0.005*
Blocker x verb -0.93 0.21 -4.39 < 0.001*
Blocker x position -0.19 0.21 -0.92 0.36
Verb x position -0.18 0.21 -0.87 0.38
Blocker x verb x position 1.43 0.42 3.41 < 0.001*

There are two effects worth noticing. First, the main effect of blocker position (p = 0.005) means that
the local binding preference in the object blocking Experiment 9a is weaker compared to the subject
blocking Experiment 8a. This is expected and probably due to the difference in matrix verbs mentioned
in Section 5.6.1. As we have seen in Chapter 4, source verbs (e.g., ‘told’) are more likely to elicit non-local
choices than perceiver verbs (e.g., ‘hear’). Thus, this main effect should not be surprising.
Second, the three-way interaction is significant (p < 0.001). To better understand the interaction,
two models with blocker type and blocker position as fixed effects were fitted for self-directed and other-
directed verb groups separately. When the verb is self-directed, the blocker type x blocker position
interaction is marginal (β = 0.56, SE = 0.29, t = 1.94, p = 0.05). This is because, as we have seen earlier,
the blocking effect only appears in Experiment 9a (object position) but not in Experiment 8a (subject
position) (due to ceiling effects in the self-directed verb conditions and not directly relevant to our
purpose). When the verb is other-directed, a blocker type x blocker position interaction to the opposite

206
direction appears (β = -0.99, SE = 0.32, t = -3.11, p = 0.002), because the blocking effect is stronger in the
subject blocking Experiment 8a than in the object blocking Experiment 9a. Thus, this finding fits the
predictions of the unified account, but not the other three accounts.

5.6.3 Experiment 9b: Self-paced reading

5.6.3.1 Participants
Seventy-six native speakers of Mandarin in mainland China participated over the internet. All had
normal or corrected-to-normal vision and only participated once in this study.

5.6.3.2 Procedure
The self-paced reading procedure is identical to the procedure in Section 5.5.3.2. in Experiment 8b.

5.6.3.3 Predictions
As Experiment 8a has revealed a syntactic component to the blocking effect, the standard cue-based
retrieval model predicts that (i) an external empathy locus (i.e., object blocker) will interfere with long-
distance binding by a topical subject but (ii) the object blocking effect should be weaker compared to
the subject blocking effect observed in Experiment 8b. Thus, this model predicts a blocker type x verb
bias interaction at or even before the critical region ziji.
In contrast, the structure-mediated cue-based retrieval model hypothesizes that only discourse-
level cues in syntactically licensed positions are retrieved at the early moment of reference resolution.

207
The object blocker wo is theoretically not in a structural position to syntactically block long-distance
binding. Thus, this model predicts no object blocking effect as at the early processing stage – in other
words, a main effect of blocker type and no blocker type x verb bias interaction.

5.6.3.4 Data analysis
Before RT analysis, participants with mean comprehension accuracy rates below 75% were removed,
which affected 8 participants. The remaining 68 participants achieved a mean accuracy rate of 90%. RT
trimming follows the same procedure as described in Experiment 8b, which resulted in the removal of
2.45% of the original data. Mixed-effect linear regressions were fitted for raw RT and log-transformed
RT analyses. Statistics based on log-transformed RT results are reported below. But in the case of
discrepancy, raw RTs will also be reported.

5.6.3.5 Results
Participants’ mean RTs for the critical region by condition and by region are plotted in Figure 5.5. See
Table 5.6 for statistics. For all the regions prior to the biased verb, no effects are significant.
At the pre-critical biased verb region, there is a significant blocker type x verb bias interaction (p <
0.005). As we have seen in Chapter 3 (Experiment 3b), with ziji, it is possible for the effect to emerge
prior to the critical region. Planned comparisons show that in the blocker conditions, self-directed verbs
lead to reading slowdowns (β = 0.07, SE = 0.03, t = 2.58, p = 0.01), suggesting a non-local binding
preference. This means that an object blocker does not impede long-distance binding. Surprisingly, in
the non-blocker conditions, there is no significant verb bias effect in log-transformed RT analysis (β = -

208
0.03, SE = 0.02, t = -1.28, p =0.2) but in the raw RT analysis, other-directed verbs lead to slowdowns (β =
-22.39, SE = 10.77, t = -2.08, p < 0.05) which indicates a local binding preference. I return to this
unexpected “3
rd
-person blocking effect” in the discussion section 5.6.4 where I offer an explanation and
propose a modified Experiment 9c. No significant effects have been observed at the critical region, ziji.

Figure 5.5: Mean RTs (ms) across all regions in the critical sentence in Experiment 9b.

verb
ziji
300
400
500
600
700
In lecture, he told others/me Prof.Wang just verbed ziji DE academic paper.
RT(ms)
Blocker
Blocker
Non−blocker
Verb
Self−directed
Other−directed

209

210
5.6.4 Discussion of Experiments 9a and 9b
Experiment 9 tested the blocking effect with the object blocker ‘me’, and thus – in combination with
Experiment 8, provides a direct comparison of subject and object blocking effects. The forced choice
judgment results from Experiment 9a show that native speakers of Mandarin on average have higher
preferences for local antecedents in the presence of an object blocker ‘me’, which does not favor the
agreement-based account. In addition, the comparative analysis of the subject (Experiment 8a) and
object blocking effects (Experiment 9a) indicates that the object blocking effect is weaker, which aligns
with the prediction of the unified account. This means that the blocking effect presumably involves
both syntactic and discourse mechanisms.
In addition to the key findings above, there are two minor findings which I turn to in this paragraph.
First, the average probability of local binding in Experiment 9a is lower compared to Experiment 8a.
This should not be surprising because all the matrix verbs in Experiment 9a are source verbs while most
of the verbs in Experiment 8a are perceiver verbs. As we have seen in Chapter 4, ziji shows logophoric
properties such that source subjects are more likely to lead to non-local binding – or reduce local
binding probability – than perceiver subjects. Thus, this result strengthens the findings of Chapter 4.
Second, the blocking effect is present in both self-directed and other-directed verb conditions in
Experiment 9a, unlike Experiment 8a where the blocking effect is absent with self-directed verbs. Here,
I speculate that the overall weaker local binding preference in self-directed verb conditions in
Experiment 9a leaves more room for the discourse-induced blocking effect to appear; in Experiment 8a,
some responses may have reached ceiling. At any rate, the two offline experiments consistently show
blocking effects.

211
The self-paced reading Experiment 9b examines whether and when the object blocker interferes
with long-distance binding by the sentence-internal empathy locus. According to the standard cue-
based retrieval model, the perspective-sensitive ziji probes for a discourse cue [+POV] and a syntactic
cue [+SUB], among others
21
, both of which are equally accessible to the parser. Since the object blocker
constitutes a partial match of the retrieval cues, it will induce a mild blocking effect. In contrast, the
structure-mediated cue-based retrieval model predicts that only discourse cues in syntactically
accessible positions (i.e., subject, c-command) can be retrieved during the early comprehension stages.
The very early verb bias effect indicating a non-local binding preference in the blocker condition
suggests that the object blocker is not effective at all in interfering with non-local binding. This finding
is thus consistent with the structure-mediated cue-based retrieval model.
One unexpected finding from Experiment 9b is that, with 3
rd
-person objects (e.g., ‘others’),
participants’ RTs point to a weak local binding preference which only reaches significance in the raw
RT analysis. This ‘3
rd
-person blocking effect’ is unpredicted by any antecedent retrieval model. I
speculate that this effect could be due to working memory burden related to similarity-based
interference and dependency length. The accessibility of NPs is known to depend on the interference
from other similar NPs (Gordon et al. 2001; 2004; 2006) and referential complexity (Warren & Gibson
2002). In the non-blocker conditions, there are three 3
rd
-person NPs (e.g. ‘he’, ‘others’, ‘Prof. Wang’)
whereas in the blocker conditions, there are only two. Thus, too many 3
rd
-person NPs in working

21
C-command may also need to be considered. But in this case, as pointed out by an anonymous reviewer of the
published version of this chapter (Lyu & Kaiser, 2021), it is debatable whether or not the object blocker ‘me’ c-
commands ziji. If both the matrix subject (e.g., ‘he’) and the object blocker ‘me’ c-command ziji, including [+c-
command] in the predictions or not is crucial for the purpose of this chapter.

212
memory potentially makes sentence processing challenging. This challenge may be compounded by
the long distance between the discourse topic and the reflexive (7 words compared to 6 in Experiment
8b). Therefore, in Experiment 9c, I would like to replicate Experiment 2b with a slightly simpler design
by removing the adverbial region.

5.6.5 Experiment 9c: Self-paced reading
The goal of Experiment 9c is to replicate the absence of object blocking effect. This study also aims to
replicate the non-local binding preference with non-blockers, observed in the previous two chapters
and in Experiment 8b of this chapter, using simpler sentences without adverbials.

5.6.5.1 Participants
Sixty-two Mandarin native speakers in mainland China with normal or corrected-to-normal vision
participated over the internet. None had participated in any other experiment reported in this study.

5.6.5.2 Procedure
The procedure is identical to that in Experiment 9b.

5.6.5.3 Predictions
If the removal of adverbials reduces participants’ working memory burden, a main effect of verb bias is
expected. Regardless of the person of the object DP, native speakers should prefer non-local binding of

213
ziji, which means self-directed verbs should lead to reading slowdowns in blocker and non-blocker
conditions alike at or before the critical region.

5.6.5.4 Data analysis
All participants had above-threshold comprehension accuracies. The mean comprehension accuracy
of the test population is 93%. The trimming of extreme RTs affected 2.73% of the original data. The
same contrast coding and statistical analyses described in Experiment 9b apply here.

5.6.5.5 Results
Figure 5.6 displays the participants mean RTs of the critical sentence by region and condition. See Table
5.7 for statistics. Before the critical region ziji, no effect reach significance. The critical region shows a
main effect of verb bias (p = 0.001). The main effect of blocker type and the blocker type x verb bias
interaction are not significant. No effect is significant at the first spillover region ‘DE’.
The second spillover region (e.g., ‘academic’) only shows a significant blocker type x verb bias
interaction (p < 0.05). However, pairwise comparisons suggest that the two trending verb bias effects
to opposite directions in the blocker and non-blocker conditions are not significant (blocker condition:
β = -0.04, SE = 0.03, t = -1.64, p =0.10; non-blocker condition: β = 0.03, SE = 0.02, t = 1.30, p = 0.19).

214

Figure 5.6: Mean RTs (ms) across all regions in the critical sentence in Experiment 9c.

Finally, the final region (e.g., ‘paper.’) shows a main effect of verb bias (p < 0.001) and a significant
blocker type x verb bias interaction (p < 0.01), but no main effect of blocker type. Planned comparisons
indicate that in the non-blocker conditions, self-directed verbs lead to significant reading slowdowns
at this region (β = 0.19, SE = 0.04, t = 4.99, p < 0.001) while in the blocker condition, there is no significant
verb bias effect (β = 0.04, SE = 0.04, t = 0.87, p = 0.39). I will not read too much into this interaction, but
this could be an indication that at least some participants may experience an object blocking effect at
the sentence-final stage. (I will come back to individual differences in General discussion.)

verb ziji
400
500
600
700
800
In lecture, he told others/me Prof.Wang verbed ziji DE academic paper.
RT(ms)
Blocker
Blocker
Non−blocker
Verb
Self−directed
Other−directed

215

216
5.6.5.6 Discussion of Experiment 9c
With the adverbial removed from the critical sentence, Experiment 9c shows results in line with the
previous experiments without 1
st
-person blockers. We now find that, as anticipated, Mandarin speakers
prefer non-local binding as soon as they see ziji. Following previous assumptions, I argue that the verb
bias effect – self-directed verbs leading to reading slowdowns in the non-blocker conditions – at the
critical region is an early effect, compatible with both the standard and the structure-mediated cue-
based retrieval models. In addition, the non-local preference in the non-blocker conditions seems to
suggest that ‘3
rd
-person blocking effect’ observed in Experiment 9b, with longer sentences, could be due
to working memory load unrelated to empathy blocking.
Perhaps more importantly, the early non-local preference in the blocker conditions replicates the
absence of object blocking effect during early-stage processing observed in Experiment 9b. Combing
the results from the offline and online experiments, we can reasonably speculate that the weak object
blocking effect in Experiment 9a occurs after the sentence is close to finish (as indicated by the blocker
type x verb bias interaction at the final region) or during re-reading. I will leave this as an open question
for future research. In sum, together with the other experiments reported in this chapter, Experiment
9c lends further support to the unified account and the structure-mediated cue-based retrieval model.

5.7 General discussion
The five experiments reported in this chapter provide insights into the interaction of syntactic,
semantic, and discoursal information in the interpretation of the Chinese reflexive ziji. The studies in

217
this chapter also update our knowledge regarding the memory-based retrieval models of sentence
processing. Below, I summarize and discuss the findings of this chapter.

5.7.1 Summary of results
This chapter has three main aims. First, I aim to test and compare the subject and object blocking effects
to examine the four different accounts of the blocking effect. Second, I investigate whether the blocking
effect is absolute in face of competing semantic information, i.e., verb directedness. Third, by examining
the time-course of the blocking effect in real-time parsing of ziji, I explore how discourse-level and
syntactic cues are retrieved at the earliest moments of reference resolution. Below, I summarize the
study results surrounding these three questions.
First, Experiments 8a and 9a found significant blocking effects with subject and object blockers,
respectively. Broadly construed, this means that syntactic prominence – the syntactic position of the
1
st
-person blocker – and discourse prominence – Speech Act Empathy Hierarchy (e.g., Kuno, 1987) –
both modulate the blocking effect. More specifically, the presence of the object blocking effect (e.g.,
Xue et al., 1994; Pollard & Xue, 1998; Huang & Liu, 2001; Huang et al., 2009), unpredicted by the
agreement-based account, and the fact that subject blocking is stronger, unpredicted by the contiguous-
agreement account and the discourse-based account alike, suggests that the unified account fits better
with the empirical findings.
Second, this study has produced strong evidence that verb semantics can compete with the
blocking effect. Previous theoretical work has not reached consensus on this question, with
contradicting examples (e.g., Yu, 1992; Y. Huang, 1994; Pollard & Xue, 1998; Huang & Liu, 2001). The two

218
offline experiments have clarified this question by looking at verb directedness. To use Experiment 8a
as an example, the impact of the verb directedness effect is manifested in two aspects. On the one hand,
when other-directed verbs and the 1
st
-person blocker point to different referents, the mean proportion
of local coreference is 46%, inconsistent with an absolute blocking effect. On the other hand, within
the two blocker conditions, the mean preference for local binding decreases from 68% to 46% with the
intervention of verb directedness. These results strengthen the point made earlier that the blocking
effect involves syntactic, semantic, and discourse-pragmatic components.
Finally, the RTs from the three self-paced reading studies further our understanding of cue-based
retrieval. Recall that, in Chapter 4, I argue that the standard cue-based retrieval model’s prediction is
too strong as not all discourse-level cues are retrieved immediately: logophoricity plays a later role than
discourse topicality. In this study, I show that the standard cue-based retrieval model should be
restricted even further. In the studies in Chapters 3 and 4, the [+POV] and [+SUB] cues appear in
tandem as the discourse topic (i.e., sentence-internal empathy locus) is always a subject. In Experiment
9b, however, the object blocker only matches the [+POV] cue but not the [+SUB] cue. This gives us an
opportunity to explore whether the [+POV] cue alone impacts reflexive resolution.
The results from Experiment 9b and 9c provide converging evidence that an object blocker does
not seem to interfere with long-distance binding by a discourse-topic subject, at least at the very early
moment of reflexive resolution. Only at the final reading region of Experiment 9c do we see weak
evidence of object blocking effect (i.e., the blocker type x verb bias interaction). Thus, these results are
more in line with a structure-mediated cue-based retrieval model, which acknowledges that discourse-
level cues are immediately accessible (in contrast to the structure-based retrieval model discussed in

219
Chapter 3 that predicts local binding preference) but posits that only discourse cues in syntactically
accessible positions are visible to the parser at the early stages of processing. In the section below, I will
discuss how this can be conceptually integrated into cue-based retrieval models.

5.7.2 A conjunction constraint in the cue-based retrieval model
We have seen that the RT data support a structure-mediated cue-based retrieval model as ziji is only
sensitive to [+POV] cues in syntactically accessible (subject) positions. In other words, ziji only searches
for an antecedent that matches both [+POV] and [+SUB] cues. An external empathy locus with only a
[+POV] cue (i.e., object blocker ‘me’) or a local 3
rd
-person NP with only a [+SUB] cue does not interfere
with non-local binding. Therefore, we are essentially dealing with a conjunction constraint on parsing.
The conjunction constraint proposed here incorporates both syntactic and discourse-level cues.
The notion of constraint conjunction has, in fact, long been proposed in prior linguistic work within
the Optimality Theory (OT) framework (e.g., Smolensky 1993; Moreton & Smolensky 2002; Lubowicz
2002). OT views language output as the outcome of constraint interactions and has been used
previously to account for sentence processing patterns such as attachment ambiguity (Stevenson &
Smolensky 2006). Therefore, it should not come as a surprise that a linguistically motivated conjunction
constraint can guide antecedent retrieval, just as Binding Theory does. However, it should be noted that
the definition of a conjunction constraint that we use is not exactly the same as what is defined in OT,
although they are two sides of the same coin. In OT, a conjunction constraint C 1&C 2 dictates that C 1 and
C 2 must not be violated at the same time in a certain domain (e.g., segment, syllable, word) while here,
C 1 and C 2 (i.e. [+POV] and [+SUB]) must be satisfied at the same time in the point-of-view domain.

220
Simply put, a conjunction constraint is formulated as a “markedness” constraint in OT but a
“faithfulness” constraint in antecedent retrieval. Despite the minor difference, this OT-inspired
proposal provides a possibility to integrate discourse-level cues into an antecedent retrieval process
that is also syntactically constrained.
In fact, constraint conjunction can explain the divergent findings in prior psycholinguistic work on
English reflexives. Going back to the pioneering work of Sturt (2003), the second eye-tracking
experiment (see (5.3) below for an example therein) did not show any early (or late) interference from
a syntactically inaccessible but gender-congruent referent (i.e., Johnathan or Jennifer) for the processing
of himself/herself. However, Patil et al. (2016) emphasized that the distractor (i.e., Johnathan/Jennifer) is
in an object position. Given an earlier study by Van Dyke and McElree (2011) which found that
grammatical roles are highly weighted in sentence processing, Patil et al. promoted the distractor to the
subject position, shown in (5.4). In the redesigned experiment, using reading eye-tracking, Patil and
colleagues discovered an early interference effect (first pass regression probability) from the non-c-
commanding, non-local but gender-congruent referent (i.e., Fred or Katie). What this suggests is that
only semantically congruent referents in the subject position (but not in the object position) are
retrieved. Conceptually, this means that both [+SUB] and [+SEM] must be satisfied, consistent with the
spirit of constraint conjunction.

(5.3) {Jonathan/Jennifer} was pretty worried at the City Hospital.
The surgeon who treated {Jonathan/Jennifer} had pricked {himself/herself} with a used
syringe needle.

221
(5.4) The tough soldier that {Fred/Katie} treated in the military hospital introduced
{himself/herself} to all the nurses.

Finally, I emphasize that the conjunction constraint is language- and grammar-specific. Whether
or not a cue has to combine with another cue to be visible to the parser depends on the identity or
importance of the cue and the grammar of a language. I do not claim that an antecedent is only visible
to the parser when all the cues encoded on that antecedent and the anaphoric expression are fully
matched. In fact, we know that this over-simplified view is questionable given prior studies on
antecedent retrieval. For example, studies on ziji and ta-ziji (‘s/he-self’) frequently found that a local
referent only matching the reflexive on [+locality] and [+c-command] but not on [+animate] is still
erroneously retrieved as the antecedent (e.g., Jäger et al., 2015; Dillon et al., 2016; Qian & Wu, 2016; Wang,
2017; Chang et al., 2020). This suggests that antecedent retrieval does not require full cue match. The
peculiar property of [+POV] that it must combine with [+SUB] to be visible in the point-of-view domain
is determined by the linguistic property of ziji, because the perspective-sensitive ziji must be bound by
a subject POV center, hence the necessity for the conjunction constraint. Therefore, the forms of the
constraints active in antecedent retrieval are heavily constrained by the grammar of a particular
language.

5.7.3 An alternative account
There is an alternative account that also holds the promise of accounting for the presence of the subject
blocking effect and the absence of the object blocking effect. This model is called the extended cue-

222
based retrieval model, developed by Jäger et al. (2015). Below, I discuss how this model can also
accommodate the reading time results in this study.
As we have seen, the standard cue-based retrieval model fails to predict the absence of the object
blocking effect. This is because the standard model posits that all cues have equal weights, and thus the
[+POV] cue alone should be able to impact antecedent retrieval at an early processing stage, which does
not fit the findings of Experiment 9b and 9c. However, if linguistic prominence can be integrated into
the cue-based retrieval model such that interference from the blocker varies as a function of its syntactic
position, the RT results can still be accounted for.
This is exactly what has been proposed in the extended cue-based retrieval model of Jäger et al.
(2015). In an eye-tracking study, Jäger et al. (2015) crossed the animacy of local and non-local referents
to investigate whether referents in non-local positions influence the processing of ziji. See (5.5a,b) for
examples where ziji is bound by local and non-local animate antecedents respectively (the local-
match/non-local match and local-mismatch/non-local mismatch conditions are not shown in Jäger et al.
(2015: 11) but included in their experiment). Note that without an empathy locus, ziji tends to be
interpreted as a syntactic anaphor, governed by Principle A of the Binding Theory (e.g., Gao et al., 2005;
Dillon et al., 2014, 2016; Wang, 2017a,b; Lyu et al., 2022).

(5.5) a. Zhe-xie shuju biaoming zhe-ge shaonian danwu-le ziji zhengzheng san nian
This-CL data indicate this-CL boy waste-ASP self wholly three year
‘These data suggest that this boy had wasted self’s life for three whole years.’
b. Zhe-ge shaonian biaoming zhe-xie shuju danwu-le ziji zhengzheng san nian

223
This- CL boy indicate this-CL data waste-ASP self wholly three year
‘This boy suggests that these data had wasted self’s life for three whole years.’

Using the reading eye-tracking method, Jäger et al. (2015) found that when the local referent is
semantically congruent with ziji, the non-local animate (i.e., semantically congruent) distractor causes
reading slowdowns in both early and later comprehension stages. This suggests that the semantic cue
[+Animate] of the syntactically inaccessible distractor is retrieved, which leads to an inhibitory effect.
When the local referent is inanimate and the non-local distractor is animate, a similar inhibitory effect
was observed but only at the later processing stage (i.e., spillover region). These results can only be
partially explained by the standard cue-based retrieval model: while the early inhibitory effect in the
local-match/non-local match condition is predicted, the delayed inhibitory effect in the local-
mismatch/non-local match condition is unexpected.
To address this unexpected inhibitory effect, Jäger et al. (2015) added a “prominence principle” that
can scale the activation levels of the local antecedent and the non-local distractor. Through
computational simulation, they found that the new extend model can correctly predict the outcome of
the eye-tracking data. Perhaps more importantly, they further predicted that the grammatical role of
the potential antecedents (subject vs. object) should also modulate the interference effect, which is
exactly what this study on the blocking effect has found. According to the extended model, 1
st
-person
blockers in the subject position should lead to stronger interference to non-local binding of ziji than a
syntactically less prominent object blocker. Investigating these predictions with computational
modeling is an important direction for future work.

224
Insofar as this study is concerned, both the structure-mediated and extended cue-based retrieval
models can make correct predictions. I remain neutral regarding which of these two models have more
merits and will aim to disentangle these two accounts in future work.

5.7.4 Individual variation in subject and object blocking effects
In this subsection, I discuss individual variation in the blocking effects during online processing. The
motivation for this exploratory analysis is two-fold. First, this allows us to further compare the subject
and object blocking effects at the individual level. Second, and perhaps more importantly, while there
is no clear object blocking effect in self-paced reading (i.e., self-directed verbs do not lead to longer RTs
in the object blocker conditions), there is a significant blocker type x verb bias interaction in the final
region in Experiment 9c. As suggested in Section 5.6.5.5, this could be due to some degree of individual
variation. To test whether the object blocking effect is entirely absent during real-time processing for
all participants, individual object blocking effects (in addition to subject blocking effects) are examined
in this section.
22

As argued in this chapter, the two key components of the blocking effect are the empathy-related
constraint (or the Speech Act Empathy Hierarchy) and the syntactic position of the blocker which
presumably prevents cyclic movement of ziji in LF. Importantly, during real-time parsing, these two
constraints apply conjointly. A [+POV] cue or a [+SUB] cue alone has not been found to influence the

22
Note that individual variation is only focused on in this chapter and not in other chapters. The reason is that
in previous chapters, statistical analyses have produced clear evidence for the relevant effects under
investigation (e.g., locality and recency effects, discourse topicality effects, logophoricity effects); but in this
chapter, the object blocking effect in real-time processing is unclear, hence the motivation for further exploratory
analysis.

225
processing of empathic ziji. However, so far, we have been only focusing on the average binding
preferences of Mandarin speakers during real-time parsing, with little attention to individual variation
in sensitivities to these syntactic and discourse-level constraints. It is conceivable that different
individuals may assign different weights to the same set of constraints. Indeed, Yadav et al. (2022)
analyzed 13 datasets and found that syntactic (c-command, locality) and semantic (number) cues are
weighted differently among participants. Some participants assign similar weights to syntactic and
semantic cues while others assign more weight to syntactic cues.
23
Similarly, regarding the weighs of
discourse-level and syntactic cues, Mandarin speakers may also show variation. Below, I discuss
individual differences by drawing on the RT data in Experiments 8b (subject blocking effect) and 9c
(object blocking effect).
In the self-paced reading study, the blocking effect is quantified as the verb bias effect, i.e., the RTs
of a region in the other-directed verb condition minus the RTs of the same region in the self-directed
verb condition. After applying this procedure to each participant in Experiments 8b and 9c, we can
derive the blocking effect for each participant across all target trials in the experiment. Figure 5.7 shows
each participant’s average blocking effect in target trials at the three spillover regions where the
blocking effect is statistically significant in Experiment 8b. Figure 5.8 shows individual object blocking
effects in Experiment 9c at the critical and spillover regions where no significant blocking effect (on
average) has been observed. In these two graphs, each dot represents a participant’s blocking effect.
Each participant’s RT data across the regions are connected by grey lines.

23
In Yadav et al.’s (2022) analysis, c-command and locality are not disentangled and treated as one syntactic
constraint. Here, I treat them as two different cues in my description.

226

Figure 5.7: Individual variation in subject blocking effects. Shown are the by-participant subject
blocking effect in Experiment 8b at the three spillover regions. Each dot represents a participant’s
average blocking effect across all target trials. The grey lines connect the data points of the same
participant.

Here, I assume that if a participant is sensitive to the blocker, the participant should show a blocking
effect in at least one of the regions in the plots. Thus, we should see at least one dot above 0ms on the
y-axis in one of the regions. If a participant is not sensitive to the blocker at all, we anticipate the dots
in all regions for that participant to be equal to or less than 0ms.
24
As we can see, in Figure 5.7, most
participants show strong subject blocking effects. In fact, only 1 participant consistently showed below
0ms blocking effect across all three regions considered.

24
This over-simplified analysis does not take into account the time-course or the change of the blocking effect. It
also omits potential noise in the data. However, as each participant sees multiple target trials, some noise may
cancel each other out.
0
500
1000
1500
DE acdemic paper
Individual subject blocking effects (ms)
Region
DE
acdemic
paper

227

Figure 5.8: Individual variation in object blocking effects. Shown are the by-participant object blocking
effect in Experiment 9c at the critical and the spillover regions. Each dot represents a participant’s
average blocking effect across all target trials. The grey lines connect the data points of the same
participant.

In contrast, the participants’ RTs in Figure 5.8 reveal a more nuanced picture. Most of the calculated
object blocking effects hover around 0ms or below 0ms as in the case of the critical region ziji. This
strengthens the findings with regard to the role of syntactic position in the blocking effect. However,
there seems to be more variation in the object blocking effect compared to the subject blocking effect.
Among the 62 participants, 8 or 12.90% consistently show below 0ms blocking effect in all four regions;
that is, they consistently considered the sentence-internal empathy locus (i.e., discourse topic) as the
antecedent. Another 15 participants or 24.19% only showed the object blocking effect at only one of the
four regions, suggesting that object blockers only have very weak influence for these participants.
−500
0
500
1000
ziji DE acdemic paper
Individual object blocking effect (ms)
Region
ziji
DE
acdemic
paper

228
Interestingly, 7 participants or 11.29% consistently show object blocking effects in all regions. This
suggests that while some individuals (12.90%) strongly prefer the perspective of the discourse topic,
others (11.29%) strongly prefer the 1
st
-person perspective, with most people showing no strong
preferences. Perhaps this explains the different judgments regarding the strength of the object blocking
effect in the theoretical literature (e.g., Pollard & Xue, 1998; Cole et al., 2006; Charnavel et al., 2017).
Overall, the above individual analysis provides further evidence supporting the main conclusions
for this chapter. However, there are clear individual differences, especially on the object blocking effect.
Following Yadav et al.’s (2022) theory, I speculate that Mandarin speakers assign different weights to
the two discourse constraints, i.e., the Topic Empathy Hierarchy (biasing the discourse topic) and the
Speech Act Empathy Hierarchy (biasing the local antecedent). Individual variation in the object
blocking effect comes partially as a result of the different weighting schemes in native speakers’ mental
representations.

229
Chapter 6: General discussion and conclusions

6.1 Summary of findings
This dissertation aims to deepen our understanding of the syntactic and non-syntactic factors that
influence the offline and online interpretation of the two reflexives in Mandarin, ziji and ta-ziji. The
factors considered include syntactic locality, linear recency, verb semantics, discourse topicality,
logophoricity (i.e., source sensitivity), and interaction of empathy loci (i.e., the blocking effect). To this
end, 9 sets of offline and online experiments were conducted to shed light on theoretical and
psycholinguistic aspects of reflexive interpretation and processing. Below, I review these research
questions and summarize the findings. As there are two themes for this dissertation, a theoretical
linguistic theme and a psycholinguistic theme, I first summarize the questions and findings on the
linguistic front before turning to the psycholinguistic front.

6.1.1 Linguistic findings
The flow of this dissertation mirrors Oshima’s tripartite division of three different uses of reflexives in
languages, the syntactic use, the empathic use, and the logophoric use. The first fundamental question
that this dissertation investigates is concerned with the syntactic use which has been long assumed to
be subject to Principle A of the Binding Theory (e.g., Chomsky, 1986) or the locality constraint. However,
it is an open question whether the local coreference preference of ziji and ta-ziji is due to a syntactic
constraint or a (potentially memory-related) linearity constraint or even both.

230
In Chapter 2, using the prenominal relative clause construction in Mandarin, I disentangled the
influence of linearly recent (but structurally non-local) and syntactically local (but linearly non-recent)
antecedents on the interpretation of ziji and ta-ziji. The offline acceptability judgment Experiment 2a
and the online self-paced reading Experiment 2b provide empirical evidence that syntactic locality and
linear recency should not be conflated. This finding suggests that ziji and ta-ziji by default can indeed
be used as ‘plain’ syntactic reflexives governed by Principle A.
However, when used as exempt anaphors – anaphors that have antecedents outside the local
domain – these two reflexives seem to show different linguistic properties, as documented in the
literature (e.g., Pan, 1998, 2001). The common assumption is that when the local referent is animate,
only ziji, but not ta-ziji, can have a non-local antecedent. The non-local or exempt reading of ziji is
arguably related to its discourse-pragmatic properties
25
: the speaker or comprehender needs to take the
perspective of the intended antecedent (e.g., Pan, 1997, 2001; Huang & Liu, 2001; Huang et al., 2009;
Wang & Pan, 2014, 2015a,b). This illustrates the empathic use of ziji. Indeed, the empathic nature of ziji
makes it subject to empathy-related constraints, including the Topic Empathy Hierarchy and the
Speech Act Empathy Hierarchy (Kuno & Kaburaki, 1977; Kuno, 1987). In contrast, ta-ziji does not easily
take non-local antecedents. But its sensitivity to the semantic prominence constraint (Pan, 1998; Pan &
Hu, 2002) raises the question of whether discourse prominence (e.g., discourse topic prominence,
logophoric prominence) also impacts its interpretation, which is the focus of Chapters 3 and 4.

25
As mentioned in fn. 4 in Chapter 1, this dissertation does not make any conclusions regarding whether
perspective can be represented syntactically. The conclusions are solely based on the experimental findings and
only speak to the different linguistic theories and hypotheses presented in this dissertation.

231
In Chapter 3, I investigated how discourse topicality impacts the interpretation of ziji and ta-ziji.
Given the general consensus that ziji can have an empathic use, I speculate that it should be constrained
by Kuno’s Topic Empathy Hierarchy. Indeed, the discourse topicality effect has been observed in the
forced choice Experiment 3a. In addition, using verb bias effects as a diagnostic for detecting binding
preferences, Experiment 3b shows that Mandarin speakers overall prefer non-local binding of ziji when
its non-local referent is a discourse topic, which, according to the Topic Empathy Hierarchy, is
preferably an empathy locus. This anticipated result reinforces the argument made in the theoretical
literature that ziji can be used as an empathic anaphor. Unlike ziji, ta-ziji does not easily lend itself to
an empathic use as Mandarin speakers seem to show a neutral binding preference. This contrast could
reflect a key difference in the linguistic properties of these two reflexive forms (but see discussion in
Section 6.2). However, despite differences in discourse topicality effects, these two reflexives share the
common property of being highly sensitive to verb directedness information, even for ta-ziji in
Experiments 4a and 4b.
Chapter 4 looks at another type of discourse-level information, logophoricity. The logophoricity
effect is measured by comparing the non-local binding probabilities with source and perceiver
antecedents. The offline forced choice Experiments 5a and 6a found statistically significant
logophoricity effects for ziji and ta-ziji alike. Furthermore, Experiment 7 does not show any significant
difference between these two reflexives in terms of the strength of the logophoricity effect. This seems
to suggest that logophoric roles impact the interpretation of these reflexive similarly (cf. Liu, 2020).
However, it is worth mentioning one notable difference between the present work and previous studies:
the source/perceiver antecedents in this study are already topically prominent (cf. e.g., Kaiser et al.,

232
2009; Sloggett, 2017; Liu, 2020; Xu & Runner, submitted). Whether logophoric prominence alone –
independent of discourse topicality status of the antecedent – impacts the non-local binding
probabilities of ziji and ta-ziji is an important topic for future research.
Critically, we see in the self-paced reading Experiments 5b and 6b that logophoricity modulates
participants’ interpretations when abstracting away from the impact of verb directedness (which is now
used as a diagnostic). In the case of ziji, the preference for non-local antecedents is stronger, albeit
numerically, with a source (topical) antecedent. This strengthens Wang and Pan’s (2015b) view that
non-local or exempt ziji is necessarily empathic but optionally logophoric in the presence of certain
attitude holders (here, source role). In the case of ta-ziji, one novel finding is that during online reading,
Mandarin speakers showed a non-local binding preference when the non-local antecedent is both a
discourse topic and a source. This challenges the conventional view that ta-ziji cannot have an exempt
use when the local referent is animate (Pan, 1998). Together with the findings of Chapter 3, Chapter 4
shows that ziji and ta-ziji have more similarities than previously thought. Both reflexive forms are
sensitive to discourse topicality and logophoricity.
Finally, Chapter 5 links empathic ziji to the Speech Act Empathy Hierarchy and tests the different
accounts of the blocking effect experimentally by comparing the subject and object blocking effects.
The agreement-based account predicts no object blocking effect; the discourse-based account and the
contiguous-agreement account predict similar blocking effects with subject and object blockers; the
unified account predicts the object blocking effect to be weaker than the subject blocking effect. A
direct comparison of the blocking effects in Experiments 8a and 9a suggests that the unified account is

233
empirically superior. A secondary linguistic finding of Chapter 5 is that verb directedness competes but
does not override the (subject) blocking effect.

6.1.2 Psycholinguistic findings
This section summarizes findings related to psycholinguistic questions and memory-based antecedent
retrieval models. There are two important questions that the online studies in this dissertation aim to
address. First, how are different types of linguistic information or cues utilized by the language parser
during real-time reflexive resolution? Specifically, do structural constraints have priority over non-
structural constraints? Second, at what processing stages are non-structural cues accessed by the parser?
These two questions relate to the standard cue-based and structure-based retrieval models. Relating to
the topics in this dissertation, the standard cue-based retrieval model predicts semantic and discourse-
pragmatic information to impact antecedent retrieval at the earliest moments. In contrast, the
structure-based retrieval model predicts that only syntactic or structural cues are accessed at the early
processing stages, although at the later stages non-structural cues could be accessible as well.
Chapter 2 examines semantic animacy and syntactic locality as cues to antecedent retrieval. It was
found in Experiment 2b that the animacy of the linearly recent but structurally non-local referent
immediately impacts the real-time processing of ziji and ta-ziji alike. But one intriguing difference in
the processing patterns of these two reflexive forms is that, when interpreted by default as syntactic
anaphors, the weight of the structural cues seems to differ for ziji and ta-ziji. The structural cue of
locality presumably carries more weight to the processing of ta-ziji than ziji. Nevertheless, the
processing patterns of ziji and ta-ziji are more in line with the predictions of the standard cue-based

234
retrieval model as syntactic locality does not block the animacy cue from being accessed at the critical
reflexive region.
The next three chapters are devoted to exploring the accessibility of discourse-level cues in online
reflexive resolution. Chapter 3 investigates how discourse topic prominence guides the processing of
ziji and ta-ziji. For ziji, the crucial question is whether, how, and when discourse topicality impacts its
interpretation as sentences are processed incrementally. Unlike previous studies which have frequently
found a strong locality bias for ziji (e.g., Gao et al., 2005; Dillon et al., 2014, 2016; Jäger et al., 2020),
Experiment 3b found that Mandarin speakers can quickly make use of discourse topicality to locate the
non-local referent as the preferred antecedent (empathy locus). As discussed in Chapter 3, this is linked
to two interconnected phenomena: (i) people tend to empathize with the discourse topic than other
referents; (ii) ziji can have an empathic use when an empathy locus is present. What is crucial for our
purposes is that the topical empathy locus immediately induces an empathic use, thus providing clear
evidence that discourse topicality or the point-of-view (POV) cue is accessed at the earliest moments.
Similarly, discourse topicality modulates the processing of ta-ziji in Experiment 4b as well despite a
weaker verb bias effect due to its different linguistic property (i.e., ta-ziji is arguably not an empathic
reflexive, unlike ziji). Thus, the online studies in Chapter 3 fit the predictions of the standard cue-based
retrieval model better.
Chapter 4 probes the logophoric role (source vs. perceiver) of the non-local antecedent and the
time-course of its accessibility. Strikingly, although the online self-paced reading Experiments 5b and
6b show evidence that logophoric role is used by Mandarin speakers, this information is utilized slightly
later (at the spillover regions), which contrasts with the early accessibility of discourse topicality. Based

235
on the different findings in terms of the time-course of the accessibility of discourse topicality and that
of logophoric roles, I proposed a revision to the standard cue-based retrieval model such that not all
discourse-level cues play an early role, only some do.
Chapter 5 deals with the blocking effect relating to ziji and explores how the structural position of
POV cues restrain retrieval of antecedents. Experiment 8b confirms the accessibilities of external POV
cues (i.e., 1
st
-person perspective) and the Speech Act Empathy Hierarchy: the 1
st
-person POV cue in the
subject position (i.e., [+POV, +SUB]) blocks the accessibility of the 3
rd
-person POV cue (i.e., the
discourse topic). However, in Experiments 9b and 9c, when the 1
st
-person blocker is relegated to an
object position (i.e., [+POV, -SUB]), no blocking effect is observed at any early stage. Only during the
last moments of reference resolution, as in the final region in Experiment 9c and in the offline
Experiment 9a, did the participants start to show evidence of a weak object blocking effect. These
results suggest that only POV cues in syntactically prominent positions induce blocking at an early stage.
The broad implication for sentence processing system is that the cue-based retrieval model should be
restricted further in addition to the earlier revision proposed in Chapter 4 (i.e., only certain discourse-
level information is accessible early on).
To summarize, the results of a series of self-paced reading studies suggest that both the standard
cue-based and the structure-based retrieval models make predictions that seem overly categorical. On
the one hand, it is not the case that non-structural cues (i.e., semantic and discourse-level cues) are not
immediately accessible to the parser. On the other, it seems that not all discourse-pragmatic cues (e.g.,
logophoric roles) play an early role. Furthermore, in the case of the blocking effect, only discourse-level

236
cues (i.e., POV) in syntactically prominent positions (i.e., subject position) guide antecedent retrieval
at an early stage.

6.2 Ziji and ta-ziji as exempt anaphors and their differences
The key take-away of this dissertation from a linguistic point of view is that ziji has a strong tendency
to be used as an empathic reflexive when discourse contexts license this use. Additionally, ziji has the
property of logophoric pronouns. However, the logophoricity effect seems weak and is not independent
of the empathic use (as ziji is always subject to the blocking effect). Thus, ziji does not neatly fit Oshima’s
tripartite division of reflexives into three distinct uses. (In comparison, zibun in Japanese seems be
immune to the blocking effect when used as a logophoric reflexive.) In contrast, based on the average
behaviors of the Mandarin participants, data on ta-ziji shows a different pattern regarding discourse
topicality. In the various sections of this dissertation, I follow the theoretical linguistic literature by
arguing that the weaker discourse topicality effect shown by ta-ziji can be related to the working
hypothesis that it is anaphoric and not empathic. If this is the case, we could argue that there is a
qualitative difference between these two reflexive forms: ziji can be empathic but ta-ziji cannot.
However, there is at least another possibility which has not been thoroughly discussed. Ziji and ta-
ziji may not be categorically different. Both could be empathic. However, the weaker discourse
topicality effect could be, among other factors, due to different kinds of underlying patterns.
Conceivably, the neutral binding preference in the self-paced reading Experiment 4b with ta-ziji could
be a mixture of different individual preferences. For example, while some participants may treat ta-ziji
as a ‘plain’ anaphor but be sensitive to both syntactic locality and discourse topicality (an extension of

237
Pan’s (1998) prominence constraint for anaphors), others may start to treat ta-ziji as an empathic
reflexive; still others may be not sensitive to topical prominence at all and only use ta-ziji in a way that
is strictly constrained only by Principle A. Together, these different patterns could yield a processing
pattern that does not show any binding preference.
Suppose the above hypothesis has some credibility which awaits further investigation, one might
ask what underlies individual variations? One possible reason could be the influence of different
Chinese dialects that the participants speak. For example, the monomorphemic reflexive jighei (‘self’)
in Cantonese shows properties different from ziji as the former can violate the subject orientation
constraint (Chan, 2012). This suggests that reflexives with similar morphological complexities in
different dialects may show different properties. Thus, similarly, regarding ta-ziji, the complex reflexives
in different Chinese dialects could impact the processing of ta-ziji in such aspects like perspective-
sensitivity through language transfer. I leave this topic to future research as this is speculative and not
critical to the evaluation of the antecedent retrieval models which mainly predict average processing
behaviors (but see Yadav et al., 2022 on modeling individual processing behaviors).
In fact, the novel finding from the self-paced reading Experiment 6b on ta-ziji in Chapter 4 –
Mandarin speakers prefer non-local binding in incremental reading when the topical subject is a source
– is in line with the possibility that Mandarin speakers may treat ta-ziji as empathic when multiple cues
favor long-distance binding. When the non-local antecedent is a discourse topic and a source, Mandarin
speakers could start to take the perspective of the antecedent. This contrasts with Experiment 4b in
Chapter 3 where the non-local referent is still a discourse topic but not a source and where no binding
preference has been observed for ta-ziji. However, it should be pointed out that the study results of

238
Experiment 6b can be alternatively explained without resorting to the notion of perspective-sensitivity.
Under the constraint-based OT framework for the analysis of ta-ziji (e.g., Pan & Hu), the non-local
binding preference can be analyzed as the outcome of constraint interactions in the resolution of an
anaphor.
Given the limited data and the scope of this dissertation, the open questions mentioned above
cannot be adequately addressed. However, these questions could lead to new discoveries that can
ultimately push forward our understanding of the linguistic behaviors of ta-ziji. I plan to look more
closely at these questions in my future research.

6.3 Empirical coverages of different variants of memory-based retrieval models
The novel contribution of this dissertation is that it has shed light on the accessibility of discourse-level
cues during antecedent retrieval. The theoretical proposal which yields the structure-mediated cue-
based retrieval model, in fact, builds on several variants of the memory-based retrieval model originally
proposed by Lewis and Vasishth (2005). The purpose of this section is to put the structure-mediated
cue-based retrieval model in the context of memory-based retrieval models in general and compare the
current proposal with previous proposals. It should be emphasized that the retrieval model proposed
here is by no means intended as a refutation of the previous models. In fact, it only adds to an
accumulating discussion on memory-based retrieval models by sharing insights from online reflexive
resolution in Mandarin.
Before comparing the structure-mediated cue-based retrieval model with other models, let us first
consider the findings of Chapters 3-5 relating to the real-time processing of ziji and ta-ziji. The first

239
major finding is that some, but not all, discourse-level cues can be accessed at the early processing
stages. This does not fit nicely with either the standard cue-based retrieval model or the structure-based
retrieval model. The former cannot explain why a prominent logophoric role such as source cannot play
an early role, while the latter cannot explain why POV cues are accessible at the early moments of
reference resolution. This means that the standard cue-based retrieval model over-generates (in the
sense that it predicts all non-syntactic cues are accessible at early stages) and the structure-based
retrieval model under-generates (in the sense that it fails to predict the early accessibility to of some
non-syntactic cues). The second major finding is that not all POV cues are equally accessible. Only POV
cues in syntactically prominent or accessible (recall the agreement-based account of the blocking effect
(e.g., Cole et al., 1990; Huang & Tang, 1991; Cole & Sung, 1994)) positions impact reflexive resolution at
a relatively early processing stage. This suggests that the standard cue-based retrieval model should be
restricted even further. Even a restricted variant of the standard model motivated by the study results
of Chapter 4 makes wider predictions than the data suggests.
Based on these findings, I proposed in Chapter 5 a structure-mediated cue-based retrieval model.
This model shares with the standard model the proposal that syntactic cues do not categorically block
the parser’s access to non-syntactic cues (e.g., semantic cues, discourse cues). However, as elaborated
in Chapter 5, it differs from the standard model in the following crucial aspect. The structure-mediated
cue-based retrieval model posits that syntax still plays an important role in accessing discourse-level
(e.g., POV) cues. In this dissertation, Chapter 5 shows that only POV cues in the subject position are
accessible; even though at the very late comprehension stage, POV cues in the object position could

240
also lead to some weak blocking effect. To account for this, I proposed a conjunction constraint which
is incorporated into the structure-mediated cue-based retrieval model.
In addition, to provide better empirical coverage for the experimental findings in the various
chapters, the structure-mediated cue-based retrieval model also needs to accommodate the fact that
not all retrieval cues are equally accessible or weighted. Some cues may be more prominent or more
heavily weighted compared to other cues. Indeed, some computational psycholinguistic studies (e.g.,
Jäger et al., 2015; Engelmann et al., 2019) show that the standard model is only partially successful in
predicting the actual results. Thus, the structure-mediated cue-based retrieval model proposed in this
dissertation assumes that different retrieval cues may carry different weights. And perhaps more
importantly, I also hypothesize, in line with the form-specific approach to reference resolution (e.g.,
Kaiser & Runner, 2008; Kaiser & Trueswell, 2008; Kaiser et al., 2009), that even the same cue can carry
different weights to different reflexive forms (as we have seen in Chapter 2).
Overall, the online self-paced reading studies in this dissertation provide insights into memory-
based retrieval models. The results regarding Mandarin ziji and ta-ziji can be used to improve our
understanding of the online reference resolution process. A future direction for this work would be to
unify the merits of various proposals and to modify the structure-mediated cue-based retrieval model
drawing on a wider spectrum of crosslinguistic data.

6.4 Directions for future research
So far, I have summarized and discussed the contributions of this dissertation. However, there are some
open questions that can benefit from more thorough investigation. In fact, some of these questions have

241
been mentioned above, including individual variation and influence from Chinese dialects. I will not
go into depth on the topic of influence from Chinese dialects regarding reflexive resolution. But I should
mention that more recent work (e.g., Hu et al., 2022) has found minor inconsistencies in inter-speaker
judgments due to the influence of regional dialects (e.g., Beijing Mandarin Chinese and Guangzhou
Cantonese Chinese), although it seems that Mandarin speakers largely agree on most of the linguistic
judgments (e.g., Chen et al., 2020; Hu et al., 2022). To my knowledge, few studies have systematically
investigated individual variation on the interpretation of reflexives in Chinese dialects (but see Cole et
al., 2001 and Chan, 2012). Future work can investigate whether reflexives in regional Chinese dialects
show different linguistic properties than (ta-)ziji in Mandarin and how these potential differences may
impact people’s reflexive resolution in Mandarin through transfer.
Another question which is left open in this dissertation pertains to the syntactic position of the
reflexive and its influence on the interpretation and the processing of ziji and ta-ziji. As we have seen,
ziji and ta-ziji in genitive position display linguistic properties similar to what we would expect for
reflexives in argument (direct object) position. These include (i) sensitivity to the syntactic locality
constraint when no discourse context is provided (Chapter 2) or when the context does not privilege
the non-local referent in the discourse (Chapter 3) and (ii) sensitivity to discourse-level prominence
(especially for genitive ziji). However, based on the comparison with other psycholinguistic studies
which use direct object reflexives, there seems to be some minor differences. One minor difference
concerns the different strengths of the locality bias displayed by ziji and ta-ziji. As mentioned in Chapter
2 (section 2.4.4.5), while Dillon et al. (2016) have found a stronger locality bias associated with ziji
compared to ta-ziji, the current work has found no such evidence. One potential explanation could be

242
related to the different syntactic positions of the reflexives in Dillon et al.’s (2016) study and the studies
in this work. Future studies can clarify the processing profiles of reflexives in different syntactic
positions.
In fact, linguistic and psycholinguistic studies which focus on genitive reflexives are relatively
scarce (but see e.g., Chen et al., 2012; Charnavel & Sportiche, 2016; Li & Juffs, 2018; Charnavel, 2020). The
studies on genitive reflexives in Mandarin in this dissertation should thus be seen as contributing to the
investigation of reflexives in non-argument positions. That said, one potential direction for future work
could be to examine genitive reflexives in other languages such as English (his/her own), French (son/sa
propre), Japanese (e.g., zibun-no), Korean (e.g., caki-uy), Norwegian and Swedish (e.g., sin), Vietnamese
(cua minh), Finnish, and others, and how these reflexives are processed.
Crosslinguistically, genitive reflexives can show different linguistic properties. On the one hand,
genitive reflexives may have an emphatic use. For example, his/her own in English is sometimes called
an ‘intensifier’ (see e.g., Baker, 1995). As Zribi-Hertz (1995) argues, his/her own in English obligatorily
obeys the locality constraint in some contexts but is exempt from the locality constraint in other
contexts. From the angle of sentence processing, one might wonder whether the locality constraint is
prioritized during real-time reflexive resolution for emphatic reflexives such as his/her own despite the
possibility for non-local binding, similar to what prior work has found regarding exempt (picture NP)
reflexives in English (e.g., Sturt, 2003; Cunnings & Sturt, 2014).
On the other hand, there are other languages – including Chinese, French (Charnavel, 2020),
Japanese, Korean, Norwegian (Hestvik & Philip, 2001), Swedish (Kaiser, 2003), and Vietnamese (Doan,
2022) – in which genitive reflexives do not carry special ‘intensification’ meaning or emphasis. For

243
instance, son/sa propre in French is by default used as a syntactic reflexive bound by the local
antecedent and can be readily interpreted as non-emphatic (e.g., Charnavel & Sportiche, 2016;
Charnavel, 2020). However, there is some preliminary evidence that reflexives in different grammatical
positions can obey (or disobey) the locality constraint to different degrees. For example, Li and Juffs
(2018) discovered that in Japanese the bare reflexive zibun (‘self’) is more likely to have a non-local
reading in dative and accusative (direct object) positions than in nominative and genitive positions.
Similar effects related to the grammatical roles of the Korean bare reflexive caki (‘self’) have also been
observed in Li and Juffs’s study. Thus, it is not only possible that reflexives in different grammatical
positions may have different linguistic behaviors but also possible that the impact of the grammatical
positions of the reflexives may differ crosslinguistically. This is a valuable direction for future work.
So far, I have discussed how we can expand the research project to look more into the linguistic
properties of reflexives in different dialects/languages and in different grammatical positions. However,
employing more research methods to complement the findings of this work is also important. There
are two methods that I aim to use in future work.
The first one is eye-tracking during reading, which provides clearer indices of early vs. later
comprehension stages. In fact, one limitation of the self-paced reading Experiments 5b and 6b is that
the logophoricity effect at the spillover regions may not necessarily be a late-stage effect. For example,
in a self-paced reading study on the processing of ta-ziji, Qian and Wu (2016) found a very late semantic
(i.e., gender) intrusion effect, which seems to suggest that, at the early stages, only Principle A plays a
dominant role. But using a larger set of stimuli with the same design, a reading eye-tracking study by
Chang et al. (2020) found indications of early accessibility of semantic gender. The discrepancy between

244
these findings could be due to methodological differences. Thus, in future work, Mandarin speakers’
processing patterns should be examined more closely using more experimental methods such as eye-
tracking during reading.
The second method that can help to evaluate memory-based retrieval models is computational
modeling. Although not crucial to the qualitative predictions of the online experiments in this
dissertation (e.g., whether POV cues can be accessible at all at the early stages), computational
modeling can make quantitative predictions (e.g., how big the animacy/gender/verb bias effect is in
milliseconds according to the memory retrieval model). This method is invaluable for approximating
the actual memory retrieval process in sentence processing.
In conclusion, this dissertation has produced several novel contributions from a theoretical
linguistic and a psycholinguistic perspective. But more insights into reflexive resolution can be gained
by considering individual-level variation and crosslinguistic differences and by using a broader range
of research methods.

245
References

Almor, A. 1999. Noun-phrase anaphora and focus: The informational load hypothesis. Psychological
Review, 106(4), 748–765.

Anand, P. 2006. De De Se. Ph.D. dissertation. MIT.

Badecker, W. & K. Straub. (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.

Baker, C. L. 1995. Contrast, discourse prominence, and intensification, with special reference to locally
free reflexives in British English. Language, 71(1), 63–101.

Banfield, A. 1982. Unspeakable Sentences: Narration and Representation in the Language of Fiction.
Boston: Routledge and Kegan Paul.

Bates, D., R. Kliegl., S. Vasishth & H. Baayen. 2015. Parsimonious mixed models. arXiv preprint arXiv:
1506.04967.

Battistella, E. 1989. Chinese reflexivization: a movement to INFL approach. Linguistics, 27(6), 987–1012.
Béjar, S. & M. Rezac. 2009. Cyclic agree. Linguistic Inquiry, 40(1), 35–73.

Bybee, J. L. & Hopper, P. J. 2001. Introduction to frequency and the emergence of linguistic structure. In
J. L Bybee & Hopper, P. J. (eds.), Frequency and the Emergence of Linguistic Structure, pp. 1–24.
Amsterdam: Johns Benjamins.

Chang, W., Y. Duan, J. Qian, F. Wu, X. Jiang & X. Zhou. 2020. Gender interference in processing Chinese
compound reflexive: Evidence from reading eye-tracking. Language, Cognition and Neuroscience,
35(10), 1355–1370.

Charnavel, I. & D. Sportiche. 2016. Anaphor binding: What French inanimate anaphors show. Linguistic
Inquiry, 47(1), 35–87.

Charnavel, I. & Y. Huang. 2018. Inanimate ziji and Condition A in Mandarin. In W. B. Bennett et al. (Eds.),
Proceedings of the 35
th
West Coast Conference on Formal Linguistics, pp. 132–141. Somerville, MA:
Cascadilla Proceedings Project.

246
Charnavel, I. 2020. Logophoricity and locality: A view from French anaphors. Linguistic Inquiry, 51(4):
671–723.

Charnavel, I. 2021. Logophoricity, perspective, and reflexives. Annual Review of Linguistics, 7, 131–155.

Charnavel, I., & C. Zlogar. 2015. English Reflexive Logophors. In K. Ershova et al. (eds.), Proceedings of
the 51st annual meeting of the Chicago Linguistic Society, 83–98. Chicago, IL.

Charnavel, I., P. Cole, G. Hermon & C.-T. J. Huang. 2017. Long-distance anaphora: Syntax and discourse.
In M. Everaert & H. C. van Riemsdijk (eds.), The Wiley Blackwell Companion to Syntax, pp. 2321–2402.
Hoboken, NJ: John Wiley & Sons, Inc.

Chen, Z., L. Jäger & S. Vasishth. 2012. How structure-sensitive is the parser? Evidence from Mandarin
Chinese. In Britta Stolterfoht & Sam Featherston (eds.), Empirical Approaches to Linguistic Theory:
Studies in Meaning and Structure (Studies in Generative Grammar), 43–62. Berlin: Mouton de
Gruyter.

Chen, Z., Y. Xu & Z. Xie. 2020. Assessing introspective linguistic judgments quantitatively: The case of
Syntax of Chinese. Journal of East Asian Linguistics, 29, 311–336.

Chomsky, N. 1981. Lectures on Government and Binding. Dordrecht: Foris.

Chomsky, N. 1986. Knowledge of Language: Its Nature, Origin, and Use. New York: Praeger.

Clements, G. N. 1975. The logophoric pronoun in Ewe: Its role in discourse. Journal of West African
Languages, 2, 141–177.

Cole, P. & C. Wang. 1996. Antecedents and blockers of long-distance reflexives: The case of Chinese ziji.
Linguistic Inquiry, 27(3): 357–390.

Cole, P. & L.-M. Sung. 1994. Head movement and long-distance reflexives. Linguistic Inquiry, 25(3), 355–
406.

Cole, P., G. Hermon & C. L. Lee. 2001. Grammatical and discourse conditions on long distance reflexives
in two Chinese dialects. In Peter Cole, Gabriella Hermon & C.-T. James Huang (eds.), Long Distance
Reflexives, 141–195. New York: Academic Press.

Cole, P., G. HermoN & C.-T. J. Huang. 2006. Long-distance anaphors: An Asian perspective. In Everaert
et al. (Eds.), Blackwell Companion to Syntax, Vol. 3. Blackwell Publishers.

Cole, P., G. Hermon & L.-M. Sung. 1990. Principles and parameters of long-distance reflexives. Linguistic
Inquiry, 21(1), 1–22.

247

Culy, C. 1994. Aspects of logophoric marking. Linguistics, 32(6), 1055–1094.

Culy, C. 1997. Logophoric pronouns and point of view. Linguistics, 35, 845–859.

Cunnings, I. & P. Sturt. (2014). Coargumenthood and the processing of reflexives. Journal of Memory and
Language, 75, 117–139.

Cunnings, I., C. Patterson & C. Felser. 2014. Variable binding and coreference in sentence
comprehension: Evidence from eye movements. Journal of Memory and Language, 71, 39–56.

Dillon, B. 2014. Syntactic memory in the comprehension of reflexive dependencies: An overview.
Language and Linguistics Compass, 8/5, 171–187.

Dillon, B., W.-Y. Chow, M. Wagers, T. Guo, F. Liu & C. Phillips. 2014. The structure-sensitivity of memory
access: Evidence from Mandarin Chinese. Frontiers in Psychology, 5. 1025.

Dillon, Brian, W.-Y. Chow & M. Xiang. 2016. The relationship between anaphor features and antecedent
retrieval: Comparing Mandarin ziji and ta-ziji. Frontiers in Psychology, 6, 1966.

Doan, Q. N. T. 2022. Anaphoric Dependencies in Vietnamese – A Syntactic Approach. Ph.D. dissertation.
Utrecht University.

Drummer, J.-D. & C. Felser. 2018. Cataphoric pronoun resolution in native and non-native sentence
comprehension. Journal of Memory and Language, 101, 97–113.

Drummond, Alex. 2013. Ibex Farm (Version 0.3.9). http://www.spellout.net/.

Eisner, J. & N. A. Smith. 2010. Favor Short Dependencies: Parsing with Soft and Hard Constraints on
Dependency Length. In H. Bunt, P. Merlo & J. Nivre (eds.), Trends in Parsing Technology. Text,
Speech and Language Technology (vol 43). Dordrecht: Springer.

Felser, C. & I. Cunnings. 2012. Processing reflexives in a second language: The timing of structural and
discourse-level constraints. Applied Psycholinguistics, 33(3), 571–603.

Gao, L., Z. Liu & Y. Huang. 2005. Ziji shi shei? Dui yueshu yuanze de shiyan yanjiu [Who is ziji? An
experimental study about the binding principle]. Yuyan kexue [Linguistic Sciences], 4, 39–50.

Garrod, S., D. Freudenthal & E. A. Boyle. 1994. The role of different types of anaphor in the on-line
resolution of sentences in a discourse. Journal of Memory and Language, 33(1), 39–68.

248
Gast, V. 2006. The Grammar of Identity: Intensifiers and Reflexives in Germanic Languages. New York:
Routledge.

Giblin, I. 2016. Agreement Restrictions in Mandarin Long-distance Binding. Ph.D. dissertation. MIT.

Gordon, P. C., B. J. Grosz & L.A. Gillion. 1993. Pronouns, names, and the centering of attention in
discourse. Cognitive Science, 17(3), 311–347.

Gordon, P. C., R. Hendrick & M. Johnson. 2001. Memory interference during language processing.
Journal of Experimental Psychology: Learning, Memory and Cognition, 27(6): 1411–1423.

Gordon, P. C., R. Hendrick & M. Johnson. 2004. Effects of noun phrase type on sentence complexity.
Journal of Memory and Language, 51(1): 97–114.

Gordon, P. C., R. Hendrick, M. Johnson & Y. Lee. 2006. Similarity-based interference during language
comprehension: Evidence from eye tracking during reading. Journal of Experimental Psychology:
Learning, Memory and Cognition, 32(6): 1304–1321.

Gracanin-Yuksek, M., S. Lago, D. F. Safak, O. Demir & B. Kirkici. 2017. The interpretation of contextual
and syntactic information in the processing of Turkish anaphors. Journal of Psycholinguistic
Research, 46: 1397–1425.

Hagège, C. 1974. Les pronoms logophoriques. Bulletin de la Société de Linguistique de Paris, 69, 287–310.

Haiman, J. 1983. Iconic and economic motivation. Language, 59(4), 781–819.

Han, P. 2020. Investigating extra-sentential factors’ roles in interpretating Mandarin ziji: The interaction
of extra-sentential (co-)reference and intra-sentential coreference. Ph.D. dissertation. University of
Calgary.

He, X. & E. Kaiser. 2016. Processing the Chinese reflexive “ziji”: Effects of featural constraints on anaphor
resolution. Frontiers in Psychology, 7, 284.

He, Xiao. 2014. What ‘you’ and ‘I’ can say about reference resolution and non-structural constraints. PhD.
Dissertation. University of Southern California.

Hestivik, A. & W. Philip. 2001. Syntactic vs. logophoric binding: Evidence from Norwegian child
language. In Peter Cole, Gabriella Hermon & C.-T. James Huang (eds.), Long Distance Reflexives, 141–
195. New York: Academic Press.

249
Hu, H., A. Li, Y. Patterson, J. Huang & C.-J. C. Lin. 2022. Examining the replicability of grammaticality
judgments in Chinese journal articles: Dialectal influences and sources of variability. Talk at the
34th annual Conference on Human Sentence Processing. Mar 24–26, Santa Cruz, CA.

Huang, C.-T. J. & C.-S. L. Liu. 2001. Logophoricity, attitudes, and ziji at the interface. In Peter Cole,
Gabriella Hermon & Huang, C.-T. James (eds.), Long-distance reflexives, 141–195. New York:
Academic Press.

Huang, C.-T. J., & C.-C. J. Tang. 1991. The local nature of long-distance reflexive in Chinese. In J. Koster
& E. Reuland (Eds.), Long Distance Anaphora, 263–282. Cambridge: Cambridge University Press.

Huang, C.-T. J., Y. Huang, E. Teng & R. Tiedeman. 1984. Zhongwen de zhaoying ci yu zhongwen jiaoxue
[Reflexives in Chinese and the teaching of Chinese]. In Proceedings of the First World Conference on
Chinese Language, 205–215. Taipei: World Chinese Language Association.

Huang, C.-T. J., Y.-H. A. Li & Y. Li. 2009. The Syntax of Chinese. Cambridge: Cambridge University Press.

Huang, Y. 1994. The Syntax and Pragmatics of Anaphora. New York: Cambridge University Press.

Huang, Y. 2000. Anaphora: A Cross-linguistic Approach. New York: Cambridge University Press.

Iida, M. 1992. Context and Binding in Japanese. Ph.D. dissertation. Stanford University.

Jäger, L. A., D. Mertzen, J. Van Dyke & S. Vasishth. 2020. Interference patterns in subject-verb agreement
and reflexives revisited: A large-sample study. Journal of Memory and Language, 111, 104063.

Jäger, L. A., F. Engelmann & S. Vasishth. 2015. Retrieval interference in reflexive processing:
Experimental evidence from Mandarin, and computational modeling. Frontiers in Psychology, 6, 617.

Jia, G. 2020. Yinghan fanshen daici gainian jichu duibi yanjiu [A contrastive study of the conceptual basis
for reflexives in English and Chinese]. Waiyu yu waiyu jiaoxue [Foreign Languages and Their
Teaching], 2, 60–68.

Kaiser, E. & J. C. Trueswell. 2008. Interpreting pronouns and demonstratives in Finnish: Evidence from
a form-sepcific approach to reference resolution. Language and Cognitive Processes, 23(5), 709–748.

Kaiser, E. & J. Runner. 2008. Intensifiers in German and Dutch anaphor resolution. In N. Abner & J.
Bishop (Eds.), Proceedings of the 27
th
West Coast Conference on Formal Linguistics, pp.265–273.
Somerville, MA: Cascadilla Proceedings Project.

Kaiser, E. 2003. Encoding (non)locality in anaphoric relations. In Nelson, D. & S. Manninen (eds.),
Generative Approaches to Finnic and Saami Linguistics, 269–293. Stanford: CSLI Publications

250
Kaiser, E. 2022. Do perspective-sensitive anaphors and subjective adjectives exhibit perspectival
uniformity? An experimental investigation. Glossa: a journal of general linguistics, 7(1), 1–52.

Kaiser, Elsi, J. T. Runner, R. S. Sussman & M. K. Tanenhaus. 2009. Structural and semantic constraints
on the resolution of pronouns and reflexives. Cognition, 112(1), 55–80.

Kazanina, N., E. F. Lau, M. Lieberman, M. Yoshida, & C. Phillips. 2007. The effect of syntactic constraints
on the processing of backwards anaphora. Journal of Memory and Language, 56(3), 384–409.

Kervran, M. & A. Prost. 1969. Les Parlers Dogons I: Donno So. Université de Dakar Documents
Linguistiques, 16. Dakar: Universite de Dakar.

Kim, E. H. & J. Yoon. 2020. Experimental evidence supporting the overlapping distribution of core and
exempt anaphors: Re-examination of long-distance bound caki-casin in Korean. Linguistics, 58(6),
1775–1806.

Kim, J.-Y. & J. H. Yoon. 2009. Long-distance bound local anaphors in Korean – An empirical study of the
Korean anaphor caki-casin. Lingua, 119, 733–755.

King, J., C. Andrews & M. Wagers. 2012. Do reflexives always find a good antecedent for themselves?
Poster presentation at the 25th CUNY Conference on Human Sentence Processing. Mar 14–16. NYC.

König, E. & P. Siemund. 2000. Intensifiers and reflexives: A typological perspective. In Z. Frajzyngier &
T. Walker (eds.), Reflexives: Forms and Functions. 41–74. Amsterdam: Benjamins.

Kuno, S. & E. Kaburaki. 1977. Empathy and syntax. Linguistic Inquiry, 8(4), 627–672.

Kuno, S. 1987. Functional Syntax: Anaphora, Discourse, and Empathy. Chicago: The University of Chicago
Press.

Kuznetsova, A., P. B. Brockhoff & R. H. B. Christensen. 2015. Package ‘lmerTest’. R package (Version 2.0).
http://CRAN.R-project.org/package=lmerTest.

Lewis, R. L. & S. Vasishth. 2005. An activation-based model on sentence processing as skilled memory
retrieval. Cognitive Science, 29: 375–419.

Li, N. & A. Juffs. 2018. Identifying co-reference of zibun and caki: The case of reflexives in Japanese and
Korean. In Xie, Z. (ed.), Buckeye East Asian Linguistics (Vol. 3), 64–74. Ohio State University.

Li, X. & X. Zhou. 2010. Who is ziji? ERP responses to the Chinese reflexive pronoun during sentence
comprehension. Brain & Research, 1331, 96–104.

251
Li, Y. 1993. What makes long-distance reflexives possible? Journal of East Asian Linguistics, 10, 135–166.

Liu, Y. 2020. Logophoricity and Mandarin exempt reflexives. University of Pennsylvania Working Papers
in Linguistics, 26(1), Article 18.

Lu, H.-Y. 2011. The Effects of Verb Bias, Context, and Tasks on Mandarin Chinese Reflexives. Ph.D.
dissertation. University of Illinois Urbana-Champaign.

Lubowicz, A. 2002. Derived environment effects in Optimality Theory. Lingua, 112: 243–280.

Lyu, J. & E. Kaiser. 2021. Unpacking the blocking effect: Syntactic prominence and perspective-taking in
antecedent retrieval in Mandarin Chinese. Glossa: a journal of general linguistics, 6(1), 136. 1–41.

Lyu, J., W. Hong & Y. Yan. 2022. Hanyu fanshen daici jiagong zhong de juyu xiaoying he jinyin xiaoying
[Locality and recency effects in the processing of Chinese reflexives]. Shijie hanyu jiaoxue [Chinese
Teaching in the World], 36(1), 115–129.

MacWhinney, B. & C. Pleh. 1988. The processing of relative clauses in Hungarian. Cognition, 29(2), 95–
141.

MacWhinney, B. 2001. The emergence of grammar from perspective. In D. Pecher & R. A. Zwann (eds.),
Grounding Cognition: The Role of Perception and Action in Memory, Language, and Thinking, 198–223.
New York: Cambridge University Press.

Matuschek, H., R. Kliegl, S. Vasishth, H. Baayen & D. Bates. 2017. Balancing Type I error and power in
linear mixed models. Journal of Memory and Language, 94, 305–315.

Moreton, E. & P. Smolensky. 2002. Typological consequences of local constraint conjunction. In L.
Mikkelsen & C. Potts (Eds.), Proceedings of the 21st West Coast Conference on Formal Linguistics, 306–
319. Cambridge, MA: Cascadilla Press.

Nicol, J. & D. Swinney. 1989. The role of structure in coreference assignment during sentence
comprehension. Journal of Psycholinguistic Research, 18, 5–19.

Nicol, J. L. 1988. Coreference Processing during Sentence Comprehension. Ph.D. dissertation. MIT.

Nishigauchi, T. 2014. Reflexive binding: awareness and empathy from a syntactic point of view. Journal
of East Asian Linguistics, 23, 157–206.

Oshima, D. Y. 2004. Zibun revisited: empathy, logophoricity, and binding. In Proceedings of the 20
th

NWLC, Vol 23 of University of Washington Working Papers, pp.175–190. Seattle: University of
Washington.

252

Oshima, D. Y. 2006. Perspectives in Reported Discourse. Ph.D. Dissertation. Stanford University.

Oshima, D. Y. 2007. On empathic and logophoric binding. Research on Language and Computation, 5,
19–35.

Pan, H. (1998). Closeness, prominence, and binding theory. Natural Language & Linguistic Theory, 16(4),
771–815.

Pan, H. & J. Hu. 2002. Hanyu fuhe fanshen daici yu yingyu fanshen daici bijiao yanjiu [Prominence and
locality: A comparative study of Chinese compound reflexives and English reflexives]. Waiyu
jiaoxue yu yanjiu [Foreign Language Teaching and Research], 34(4), 241–247.

Pan, H. 1997. Constraints on Reflexivization in Mandarin Chinese. New York: Garland Publishing, Inc.

Pan, H. 2001. Why the blocking effect? In Jan Koster & Eric Reuland (eds.), Long Distance Anaphora,
279–316. Cambridge: Cambridge University Press.

Pancheva, R. & Zubizarreta, M. L. 2018. The person case constraint: The syntactic encoding of
perspective. Natural Language and Linguistic Theory, 36, 1291–1337.

Parker, D. & C. Philips. 2017. Reflexive attraction in comprehension is selective. Journal of Memory and
Language, 94, 272–290.

Parker, D. 2019. Cue combinatorics in memory retrieval for anaphora. Cognitive Science, 43, e12715.

Patil, U., S. Vasishth & R. L. Lewis. 2016. Retrieval interference in syntactic processing: The case of
reflexive binding in English. Frontiers in Psychology, 7, 329.

Pearson, H. 2015. The interpretation of the logophoric pronoun in Ewe. Natural Language Semantics, 23,
77–118.

Pica, P. 1986. On the nature of the reflexivization cycle. Proceedings of the 17
th
North East Linguistics
Society, Article 9.

Pollard, C. & I. A. Sag. 1992. Anaphors in English and the scope of Binding Theory. Linguistic Inquiry,
23(2): 261–303.

Pollard, C. & P. Xue. 1998. Chinese reflexive ziji: Syntactic reflexives vs. non-syntactic reflexives. Journal
of East Asian Linguistics, 7(4), 287–318.

253
Qian, J. & F. Wu. 2016. Yueshu yuanze A zai hanyu fuhe fanshen daici jiagong zhong de youxian zhiyue
xiaoying [Immediate effects of Binding Principle A on complex reflexive processing in Chinese].
Xiandai waiyu [Modern Foreign Languages], 39, 495–506.

R Core Team. 2018. R: A language and environment for statistical computing. http://www.R-project.org/.

R., Tanya & E. J. Reuland. 1993. Reflexivity. Linguistic Inquiry, 24(4): 657–720.

Reuland, E. J. 2001. Primitives of binding. Linguistic Inquiry, 32(3): 439–492.

Reuland, E. J. 2011. Anaphora and Language Design. Cambridge, MA: MIT Press.

Runner, J. T., R. S. Sussman & M. K. Tanenhaus. (2006). Processing reflexives and pronouns in picture
noun phrases. Cognitive Science, 30, 193–241.

Schumacher, P. B., W. Bisang & L. Sun. 2011. Perspective in the processing of the chinese reflexive ziji:
ERP evidence. In I. Hendrickx et al. (Eds.), Anaphora Processing and Applications. DAARC 2011.
Lecture Notes in Computer Science (Vol. 709), pp. 119–131. Berlin: Springer.

Sells, P. 1987. Aspects of logophoricity. Linguistic Inquiry, 18(3), 445–479.

Sloggett, S. 2017. When errors aren’t: How comprehenders selectively violate Binding Theory. Ph.D
dissertation. University of Massachusetts, Amherst.

Smolensky, P. 1993. Harmony, markedness, and phonological activity. Paper presented at Rutgers
Optimality Workshop 1.

Speas, P. & C. L. Tenny. 2003. Configurational properties of point of view roles. In A. M. Di Sciullo (Ed.),
Asymmetry in Grammar: Syntax and Semantics (Vo. 1), pp. 315–344. Amsterdam: John Benjamins.

Speas, M. 2004. Evidentiality, logophoricity and syntactic representation of pragmatic features. Lingua,
114, 255–276.

Stevenson, S. & P. Smolensky. 2006. Optimality in sentence processing. In P. Smolensky & G. Legendre
(Eds.), The Harmonic Mind: From Neural Computation to Optimality-theoretic Grammar (Vol. 2), pp.
307–338. Cambridge, MA: MIT Press.

Sturt, Patrick. 2003. The time-course of the application of binding constraints in reference resolution.
Journal of Memory and Language, 48(3), 542–562.

Tang, C.-C. J. 1989. Chinese reflexives. Natural Language and Linguistic Theory, 7: 93–121.

254
Thompson, C. & J. J. Choy. Pronominal resolution and gap filling in agrammatical aphasia: Evidence
from eye movements. Journal of Psycholinguistic Research, 38(3), 255–283.

Trueswell, J. C., M. K. Tanenhaus & C. Kello. 1993. Verb-specific constraints sentence processing:
Separating effects of lexical preference from garden-paths. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 19(3), 528–553.

Van Dyke, J. & B. McElree. 2006. Retrieval interference in sentence comprehension. Journal of Memory
and Language, 55, 157–166.

Van Dyke, J. & B. McElree. 2011. Cue-dependent interference in comprehension. Journal of Memory and
Language, 65, 247–263.

Van Dyke, J. 2007. Interference effects from grammatically unavailable constituents during sentence
processing. Journal of Experimental Psychology: Learning, Memory, and Cognition, 33(2), 407–430.

Vasishth, S., S. Brüssow, R. L. Lewis & H. Drenhaus. 2008. Processing polarity: How the ungrammatical
intrudes on the grammatical. Cognitive Science, 32, 685–712.

Wang, Y. & H. Pan. 2015a. Empathy and Chinese long-distance reflexive ziji—Remarks on Giorgi (2006,
2007). Natural Language and Linguistic Theory, 33, 307–322.

Wang, Y. & H. Pan. 2015b. Hanyu changjuli ziji de yiqinglun jieshi. [An empathic account of Chinese
long-distance anaphor ziji]. Waiyu jiaoxue yu yanjiu [Foreign Language Teaching and Research],
47(6), 814–825.

Wang, Y. 2017. Yuyi yueshu haishi jufa yueshu?—Hanyu jiandan he fuhe fanshen daici de hanyu jiagong
yanjiu. [Semantic-constrained or syntactic-constrained?—Processing Chinese bare and compound
reflexives]. Xinli kexue [Journal of Psychological Science], 40, 527–533.

Wang, Y. 2017a. Yuyi yueshu haishi jufa yueshu?—Hanyu jiandan he fuhe fanshen daici de hanyu
jiagong yanjiu. [Semantic-constrained or syntactic-constrained?—Processing Chinese bare and
compound reflexives]. Xinli kexue [Journal of Psychological Science], 40. 527–533.

Wang, Y. 2017b. Hanyu jiandan he fuhe fanshendaici zhidai xiaojie de yandong yanjiu [An eye-
movement study on the online processing of Chinese bare and compound reflexives]. Shanghai
Jiaotong Daxue Xuebao [Journal of SJTU], 25, 97–106.

Warren, T. & T. Gibson. 2002. The influence of referential processing on sentence complexity. Cognition,
85: 79–112.

Wilcox, R. 2017. Modern Statistics for the Social Science and Behavioral Sciences. Boca Raton: CRC Press.

255

Xiang, M., B. Dillon & C. Phillips. 2009. Illusory licensing effects across dependency types: ERP evidence.
Brain & Language, 108(1), 40–55.

Xu, Liejiong. 1993. The long-distance binding of ziji. Journal of Chinese Linguistics, 21: 123–141.

Xu, Y., & J. Runner. Submitted. Memory retrieval in reflexives processing: Evidence against the local
search hypothesis.

Xue, Ping, Carl J. Pollard & Ivan A. Sag. 1994. A new perspective on Chinese reflexive ziji. In Raul
Aranovich, William Byrne, Susanne Preuss & Martha Senturia (eds.), Proceedings of the 13th West
Coast Conference on Formal Linguistics, 432–447. CLSI Publications.

Yadav, H., D. Paape, G. Smith, B. Dillon & S. Vasishth. 2022. Individual differences in cue weighting in
sentence comprehension: An evaluation using approximate Bayesian computation. Discoveries in
Cognitive Science, 6, 1–24.

Yu, William. 1992. Challenging Chinese reflexive data. The Linguistic Review, 9(3), 285–294.

Zhan, W., R. Guo & Y. Chen. 2003. The CCL Corpus of Chinese Texts: 700 million Chinese Characters, the
11th Century B.C. – present. Available online at the website of Center for Chinese Linguistics (CCL)
of Peking University. http://ccl.pku.edu.cn:8080/ccl_corpus/

Zribi-Hertz, A. 1989. Anaphor binding and narrative point of view: English reflexive pronouns in
sentence and discourse. Language, 65(4), 695–727.

Zribi-Hertz, A. 1995. Emphatic or reflexive? On the endophoric character of French lui-même and similar
complex pronouns. Journal of Linguistics, 31, 333–374.

256
Appendix 1

This appendix shows the structures of the mixed-effect linear and logistic models used for data analyses
in the various chapters of the dissertation. Maximal models with random intercepts and slopes grouped
by participants and items are first fitted. If a model fails to converge, simpler models will be fitted by
removing random slopes. Furthermore, if a more complex model does not differ significantly from a
simpler model in model comparisons, the simpler model is adopted.

Experiment 1a
clmm(Score ~ Reflexive*Distance + (1|Participant) + (1|Item), data)

Experiment 1b
Region 1: lmer(logRT ~ Reflexive*Distance + (1|Participant) + (1|Item), R1)
Region 2: lmer(logRT ~ Reflexive*Distance + (1|Participant) + (1|Item), R2)
Region 3: lmer(logRT ~ Reflexive*Distance + (1|Participant) + (1|Item), R3)
Region 4: lmer(logRT ~ Reflexive*Distance + (1|Participant) + (1|Item), R4)
Region 5: lmer(logRT ~ Reflexive*Distance + (1|Participant) + (1|Item), R5)
Region 6: lmer(logRT ~ Reflexive*Distance + (1|Participant) + (1|Item), R6)
Region 7: lmer(logRT ~ Reflexive*Distance + (1|Participant) + (1|Item), R7)
Region 8: lmer(logRT ~ Reflexive*Distance + (1|Participant) + (1|Item), R8)
Region 9: lmer(logRT ~ Reflexive*Distance + (1|Participant) + (1|Item), R9)

257

Experiment 2a
clmm(Score ~ Reflexive*Position + (1|Participant) + (1|Item), data)

Experiment 2b
Region 1: lmer(logRT ~ Reflexive*Position + (1|Participant) + (1|Item), R1)
Region 2: lmer(logRT ~ Reflexive*Position + (1|Participant) + (1|Item), R2)
Region 3: lmer(logRT ~ Reflexive*Position + (1|Participant) + (1|Item), R3)
Region 4: lmer(logRT ~ Reflexive*Position + (1|Participant) + (1|Item), R4)
Region 5: lmer(logRT ~ Reflexive*Position + (1|Participant) + (1|Item), R5)
Region 6: lmer(logRT ~ Reflexive*Position + (1|Participant) + (1|Item), R6)
Region 7: lmer(logRT ~ Reflexive*Position + (1|Participant) + (1|Item), R7)
Region 8: lmer(logRT ~ Reflexive*Position + (1|Participant) + (1|Item), R8)
Region 9: lmer(logRT ~ Reflexive*Position + (1|Participant) + (1|Item), R9)
Region 10: lmer(logRT ~ Reflexive*Position + (1|Participant) + (1|Item), R10)
Region 11: lmer(logRT ~ Reflexive*Position + (1|Participant) + (1|Item), R11)

Experiment 3a
glmer(Choice ~ Context*Verb + (1|Participant)+(1|Item), data, family = binomial)

Experiment 3b

258
Region 1: lmer(logRT ~ Context*Verb + (1 + Context|Participant) + (1 + Context|Item), R1)
Region 2: lmer(logRT ~ Context*Verb + (1|Participant) + (1|Item), R2)
Region 3: lmer(logRT ~ Context*Verb + (1 + Context + Verb|Participant) + (1 + Context + Verb|Item),
R3)
Region 4: lmer(logRT ~ Context*Verb + (1|Participant) + (1|Item), R4)
Region 5: lmer(logRT ~ Context*Verb + (1|Participant) + (1|Item), R5)
Region 6: lmer(logRT ~ Context*Verb + (1 + Context + Verb|Participant) + (1 + Context|Item), R6)
Region 7: lmer(logRT ~ Context*Verb + (1|Participant) + (1|Item), R7)
Region 8: lmer(logRT ~ Context*Verb + (1|Participant) + (1|Item), R8)
Region 9: lmer(logRT ~ Context*Verb + (1 + Context|Participant) + (1|Item), R9)
Region 10: lmer(logRT ~ Context*Verb + (1|Participant) + (1|Item), R10)

Experiment 4a
glmer(Choice ~ Context*Verb + (1|Participant) + (1|Item), data, family = binomial)

Experiment 4b
Region 1: lmer(logRT ~ Context*Verb + (1|Participant) + (1|Item), R1)
Region 2: lmer(logRT ~ Context*Verb + (1 + Context|Participant) + (1|Item), R2)
Region 3: lmer(logRT ~ Context*Verb + (1 + Context + Verb|Participant) + (1 + Context|Item), R3)
Region 4: lmer(logRT ~ Context*Verb + (1 + Context|Participant) + (1|Item), R4)
Region 5: lmer(logRT ~ Context*Verb + (1 + Context|Participant) + (1|Item), R5)

259
Region 6: lmer(logRT ~ Context*Verb + (1|Participant) + (1|Item), R6)
Region 7: lmer(logRT ~ Context*Verb + (1 + Context + Verb|Participant) + (1|Item), R7)
Region 8: lmer(logRT ~ Context*Verb + (1 + Context|Participant) + (1|Item), R8)
Region 9: lmer(logRT ~ Context*Verb + (1 + Context|Participant) + (1|Item), R9)
Region 10: lmer(logRT ~ Context*Verb + (1 + Context|Participant) + (1|Item), R10)

Experiment 5a
glmer(Choice ~ Role*Verb + (1|Participant) + (1|Item), data, family = binomial)

Experiment 5b
Region 1: lmer(logRT ~ Role*Verb + (1 + Role|Participant) + (1|Item), R1)
Region 2: lmer(logRT ~ Role*Verb + (1 + Role|Participant) + (1|Item), R2)
Region 3: lmer(logRT ~ Role*Verb + (1|Participant) + (1|Item), R3)
Region 4: lmer(logRT ~ Role*Verb + (1|Participant) + (1|Item), R4)
Region 5: lmer(logRT ~ Role*Verb + (1|Participant) + (1|Item), R5)
Region 6: lmer(logRT ~ Role*Verb + (1|Participant) + (1 + Verb|Item), R6)
Region 7: lmer(logRT ~ Role*Verb + (1|Participant) + (1|Item), R7)
Region 8: lmer(logRT ~ Role*Verb + (1 + Verb|Participant) + (1|Item), R8)
Region 9: lmer(logRT ~ Role*Verb + (1 + Verb|Participant) + (1|Item), R9)
Region 10: lmer(logRT ~ Role*Verb + (1 + Verb|Participant) + (1|Item), R10)

260
Experiment 6a
glmer(Choice ~ Verb*Role + (1|Participant) + (1|Item), data, family = binomial)

Experiment 6b
Region 1: lmer(logRT ~ Role*Verb + (1|Participant) + (1|Item), R1)
Region 2: lmer(logRT ~ Role*Verb + (1|Participant) + (1|Item), R2)
Region 3: lmer(logRT ~ Role*Verb + (1|Participant) + (1|Item), R3)
Region 4: lmer(logRT ~ Role*Verb + (1|Participant) + (1|Item), R4)
Region 5: lmer(logRT ~ Role*Verb + (1 + Role + Verb|Participant) + (1 + Role + Verb|Item), R5)
Region 6: lmer(logRT ~ Role*Verb + (1 + Verb|Participant) + (1 + Verb|Item), R6)
Region 7: lmer(logRT ~ Role*Verb + (1 + Role|Participant) + (1 + Verb|Item), R7)
Region 8: lmer(logRT ~ Role*Verb + (1 + Role|Participant) + (1|Item), R8)
Region 9: lmer(logRT ~ Role*Verb + (1 + Role|Participant) + (1 + Role|Item), R9)
Region 10: lmer(logRT ~ Role*Verb + (1 + Verb|Participant) + (1 + Verb|Item), R10)

Experiment 7
glmer(Choice ~ Reflexive*Role + (1 + Reflexive|Participant) + (1|Item), data, family = binomial)

Experiment 8a
glmer(Choice ~ Blocker*Verb + (1|Item) + (1 + Verb|Participant), data, family = binomial)

261
Experiment 8b
Region 1: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R1)
Region 2: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1 + Verb|Item), R2)
Region 3: lmer(logRT ~ Blocker*Verb + (1 + Verb|Participant) + (1 + Blocker|Item), R3)
Region 4: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R4)
Region 5: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R5)
Region 6: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R6)
Region 7: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R7)
Region 8: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R8)
Region 9: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R9)
Region 10: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R10)

Experiment 9a
glmer(Choice ~ Blocker*Verb + (1|Participant) + (1|Item), data, family = binomial)

Experiment 9b
Region 1: lmer(logRT ~ Blocker*Verb + (1 + Blocker + Verb|Participant) + (1 + Blocker + Verb|Item), R1)
Region 2: lmer(logRT ~ Blocker*Verb + (1 + Blocker + Verb|Participant) + (1 + Blocker + Verb|Item), R2)
Region 3: lmer(logRT ~ Blocker*Verb + (1 + Blocker + Verb|Participant) + (1 + Verb|Item), R3)
Region 4: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R4)
Region 5: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R5)

262
Region 6: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R6)
Region 7: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R7)
Region 8: lmer(logRT ~ Blocker*Verb + (1 + Verb|Participant) + (1|Item), R8)
Region 9: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R9)
Region 10: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R10)
Region 11: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R11)

Experiment 9c
Region 1: lmer(logRT ~ Blocker*Verb + (1 + Verb|Participant) + (1 + Verb|Item), R1)
Region 2: lmer(logRT ~ Blocker*Verb + (1 + Verb|Participant) + (1 + Verb|Item), R2)
Region 3: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R3)
Region 4: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R4)
Region 5: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R5)
Region 6: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R6)
Region 7: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R7)
Region 8: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R8)
Region 9: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R9)
Region 10: lmer(logRT ~ Blocker*Verb + (1|Participant) + (1|Item), R10)

263
Appendix 2

This appendix contains the Chinese target sentences used in all the experiments reported in the
dissertation.

Experiments 1a & 1b
1. {电台/记者}表⽰{那名记者/那家电台}已经公布了{⾃⼰/他⾃⼰}的采访⽬的。
2. {公司/律师}表⽰{那名律师/那家公司}不⼩⼼丢失了{⾃⼰/他⾃⼰}的个⼈资料。
3. {画廊/画家}表⽰{那名画家/那家画廊}经常出售{⾃⼰/他⾃⼰}的获奖作品。
4. {医院/医⽣}表⽰{那名医⽣/那家医院}时常宣传{⾃⼰/他⾃⼰}的医学观点。
5. {影视公司/戏剧演员}表⽰{那名喜剧演员/那家影视公司}不⼩⼼搞砸了{⾃⼰/他⾃⼰}的
⾸场表演。
6. {照相馆/摄影师}声称{那名摄影师/那家照相馆}准备出版{⾃⼰/他⾃⼰}的摄影作品。
7. {法院/官员}声称{那名官员/那家法院}已经提交了{⾃⼰/他⾃⼰}的贪污证据。
8. {餐厅/厨师}认为{那名厨师/那家餐厅}还未透露{⾃⼰/他⾃⼰}的调料配⽅。
9. {学校/教师}表⽰{那所学校/那名教师}屡次修改{⾃⼰/他⾃⼰}的教学⼤纲。

264
10. {医院/⽼⼈}觉得{那位⽼⼈/那家医院}根本不了解{⾃⼰/他⾃⼰}的⾝体状况。
11. {部队/⼠兵}表⽰{那名⼠兵/那个部队}擅⾃曝光了{⾃⼰/他⾃⼰}的保密协议。
12. {⼯⼚/⼯⼈}透露{那名⼯⼈/那家⼯⼚}擅⾃篡改{⾃⼰/他⾃⼰}的⼯作记录。
13. {公司/总裁}发现{那名总裁/那家公司}打算公布{⾃⼰/他⾃⼰}的⼯资收⼊。
14. {出版社/⽼作家}声称{那名⽼作家/那家出版社}已经公布了{⾃⼰/他⾃⼰}的早期作品。
15. {电视台/舞蹈家}透露{那名舞蹈家/那家电视台}刚刚录制完{⾃⼰/他⾃⼰}的舞蹈表演。
16. {组织/卧底}意识到{那个卧底/那个组织}不⼩⼼暴露了{⾃⼰/他⾃⼰}的真实⾝份。
17. {搬家公司/妇联主任}表⽰{那名妇联主任/那家搬家公司}正在整理⾃⼰的⾏李和⾐物。
18. {书店/教授}表⽰{那名教授/那家书店}正在整理{⾃⼰/他⾃⼰}的书稿和⽇记。
19. {报社/记者}坦⾔{那名记者/那家报社}擅⾃发表了{⾃⼰/他⾃⼰}的⼀篇⽂章。
20. {警局/侦探}表⽰{那名侦探/那个警局}不慎暴露了{⾃⼰/他⾃⼰}的真实⾝份。
21. {物业/房客}表⽰{那名房客/那家物业}未能保护{⾃⼰/他⾃⼰}的个⼈财产。
22. {唱⽚公司/流⾏歌⼿}以为{那名流⾏歌⼿/那家唱⽚公司}想要策划{⾃⼰/他⾃⼰}的个⼈
巡演。
23. {服装店/设计师}表⽰{那名设计师/那家服装店}正在展⽰{⾃⼰/他⾃⼰}的服装设计。

265
24. {医院/护⼠}表⽰{那名护⼠/那家医院}到处宣扬{⾃⼰/他⾃⼰}的劳资纠纷。

Experiments 2a & 2b
1. {电台/记者}表⽰公布了{⾃⼰/他⾃⼰}的采访⽬的的{记者/电台}引起了⼤家的不满。
2. {公司/律师}表⽰丢失了{⾃⼰/他⾃⼰}的个⼈资料的{律师/公司}已经公开道歉。
3. {画廊/画家}表⽰出售了{⾃⼰/他⾃⼰}的获奖作品的{画家/画廊}经常投资慈善事业。
4. {医院/医⽣}表⽰宣传了{⾃⼰/他⾃⼰}的医学观点的{医⽣/医院}接受了病⼈的锦旗。
5. {影视公司/戏剧演员}表⽰搞砸了{⾃⼰/他⾃⼰}的⾸场表演的{喜剧演员/影视公司}发表
了道歉声明。
6. {照相馆/摄影师}声称出版了{⾃⼰/他⾃⼰}的摄影作品的{摄影师/照相馆}拟定了销售⽅
案。
7. {法院/官员}声称提交了{⾃⼰/他⾃⼰}的贪污证据的{官员/法院}引起了媒体的关注。
8. {餐厅/厨师}认为透露了{⾃⼰/他⾃⼰}的调料配⽅的{厨师/餐厅}做事不计后果。
9. {学校/教师}表⽰修改了{⾃⼰/他⾃⼰}的教学⼤纲的{学校/教师}仍在征询意见。
10. {医院/⽼⼈}觉得不了解{⾃⼰/他⾃⼰}的⾝体状况的{⽼⼈/医院}表现得过于紧张。

266
11. {部队/⼠兵}表⽰擅⾃曝光了{⾃⼰/他⾃⼰}的保密协议的{⼠兵/部队}正在接受调查。
12. {⼯⼚/⼯⼈}透露篡改了{⾃⼰/他⾃⼰}的⼯作记录的{⼯⼈/⼯⼚}早已臭名昭著。
13. {公司/总裁}发现公布了{⾃⼰/他⾃⼰}的⼯资收⼊的{总裁/公司}频频散步谣⾔。
14. {出版社/⽼作家}声称公布过{⾃⼰/他⾃⼰}的早期作品的{⽼作家/出版社}即将举办发布
会。
15. {电视台/舞蹈家}透露录制了{⾃⼰/他⾃⼰}的舞蹈表演的{舞蹈家/电视台}具有很⾼的⼈
⽓。
16. {组织/卧底}意识到暴露了{⾃⼰/他⾃⼰}的真实⾝份的{卧底/组织}已经转移了阵地。
17. {搬家公司/妇联主任}表⽰整理了⾃⼰的⾏李和⾐物的{妇联主任/搬家公司}要求调整价
格。
18. {书店/教授}表⽰整理了{⾃⼰/他⾃⼰}的书稿和⽇记的{教授/书店}忘了签署协议。
19. {报社/记者}坦⾔发表了{⾃⼰/他⾃⼰}的⼀篇⽂章的{记者/报社}犯了⼀个错误。
20. {警局/侦探}表⽰暴露了{⾃⼰/他⾃⼰}的真实⾝份的{侦探/警局}及时配合了⾏动。
21. {物业/房客}表⽰保护了{⾃⼰/他⾃⼰}的个⼈财产的{房客/物业}已经联系了警⽅。

267
22. {唱⽚公司/流⾏歌⼿}以为策划了{⾃⼰/他⾃⼰}的个⼈巡演的{流⾏歌⼿/唱⽚公司}制定
了⾏程安排。
23. {服装店/设计师}表⽰展⽰过{⾃⼰/他⾃⼰}的服装设计的{设计师/服装店}瞄准了⾼端市
场。
24. {医院/护⼠}表⽰宣扬了{⾃⼰/他⾃⼰}的劳资纠纷的{护⼠/医院}没意识到问题的严重性。

Experiments 3a & 3b
1. {今天是上⽂学课的⽇⼦/⼩明是班级⾥的优秀学⽣}。课上，{⼩明/他}听说王教授刚刚
{发表/批改}了⾃⼰的学术论⽂。
2. {⼤厦⾥正在举⾏⼀个重要的会议/董事长主持了今天早晨的例会}。会上，{董事长/他}注
意到李主任⼀直{炫耀/称赞}⾃⼰的管理才能。
3. {区政府忙碌的⼀天终于结束了/⼩芳是区政府的⼀位业务员}。下班后，{⼩芳/她}说杨姐
向上级{隐瞒/揭露}了⾃⼰的真实⽬的。
4. {昏暗的酒楼⾥客⼈稀少/省长惴惴不安、夜不能寐}。饭桌上，{省长/他}透露书记很可能
{承认/举报}了⾃⼰的灰⾊收⼊。

268
5. {今天是⼀年⼀度的圣诞节/张霞⼗分喜爱刚买的跑车}。派对上，{张霞/她}听说李莉竟然
{抵押/看上}了⾃⼰的最新款宝时捷。
6. {昨天领导们⼀起吃了顿饭/胡镇长早已对权⼒⽃争失去了兴趣}。饭局上，{胡镇长/他}得
知政委已决定{放弃/剥夺}⾃⼰的⾏政权⼒。
7. {天底下就没有不透风的墙/⼩美是班⾥的⼊团积极分⼦}。昨天，{⼩美/她}听说李芳已经
{提交/调查}了⾃⼰的个⼈档案。
8. {集团与合作伙伴进⾏了⼀次谈判/周鹏参加了集团的⼀次商业谈判}。事后，{周鹏/他}得
知吴部长并没有{发挥/表扬}⾃⼰的谈判才能。
9. {国企的⽇⼦其实很无聊/朱洁对国企的⽣活很反感}。⽐如，{朱洁/她}注意到周阿姨经常
{卖弄/吹捧}⾃⼰的语⾔天赋。
10. {试卷批改⼯作很快就结束了/⼩华是班级⾥最聪明的学⽣}。成绩出来后，{⼩华/他}推
断李明肯定{写完/抄袭}了⾃⼰的数学试卷。
11. {新闻发布会上⼈头攒动/乐乐早已是各⼤媒体的红⼈}。⾯对媒体时，{乐乐/她}说王婷
悄悄地{贡献/窃取}了⾃⼰的股票分红。

269
12. {警局⾥的⽓氛极其凝重/愁眉苦脸的李⽼板最终报了警}。⾯对质询时，{李⽼板/他}表
⽰⽼吴早已{抛售/盗取}了⾃⼰的企业资产。
13. {六⼗年代，政治氛围格外紧张/张师傅对⼈⼀直⼼存戒备}。⼀天，{张师傅/他}发现徐
班长打算{隐藏/曝光}⾃⼰的海外关系。
14. {傍晚，电视台⼤楼⾥灯⽕通明/傍晚，⽟洁在电视台接受采访}。演播室⾥，{⽟洁/她}
透露冰冰前年{捐献/继承}了⾃⼰的多处房产。
15. {⼜到了开班会的⽇⼦/娜娜是班级⾥的卫⽣委员}。班会上，{娜娜/她}表⽰雪婷刚才{承
认/批评}了⾃⼰的鲁莽⾏为。
16. {新⼀批的积极分⼦都很优秀/⽟霞最近表现得⼗分积极}。昨天，{⽟霞/她}听说⼩雪已
经{提交/批准}了⾃⼰的⼊党申请。
17. {股东⼤会刚刚结束/杜芸在焦急地等待会议结果}。很快，{杜芸/她}得知佳佳在⼤会上
{交代/反对}了⾃⼰的收购计划。
18. {庭院⾥的邻居们都聊着天/⼀向默不作声的叶倩终于说话了}。刚才，{叶倩/她}透露丁
丽上个⽉{缴纳/克扣}了⾃⼰的房租和押⾦。

270
19. {这场政治风波终于落幕了/宋迪在政治博弈中站错了队伍}。报纸上，{宋迪/他}看见领
导发⽂{放弃/剥夺}了⾃⼰的竞选资格。
20. {公司刚刚开了⼀个会/朱军在会议上勇于发⾔}。听说，{朱军/他}觉得总监应该{撤回/
奖励}⾃⼰的营销⽅案。

Experiments 4a & 4b
1. {今天是上⽂学课的⽇⼦/⼩明是班级⾥的优秀学⽣}。课上，{⼩明/他}听说王教授刚刚
{发表/批改}了他⾃⼰的学术论⽂。
2. {⼤厦⾥正在举⾏⼀个重要的会议/董事长主持了今天早晨的例会}。会上，{董事长/他}注
意到李主任⼀直{炫耀/称赞}他⾃⼰的管理才能。
3. {区政府忙碌的⼀天终于结束了/⼩芳是区政府的⼀位业务员}。下班后，{⼩芳/她}说杨姐
向上级{隐瞒/揭露}了她⾃⼰的真实⽬的。
4. {昏暗的酒楼⾥客⼈稀少/省长惴惴不安、夜不能寐}。饭桌上，{省长/他}透露书记很可能
{承认/举报}了他⾃⼰的灰⾊收⼊。

271
5. {今天是⼀年⼀度的圣诞节/张霞⼗分喜爱刚买的跑车}。派对上，{张霞/她}听说李莉竟然
{抵押/看上}了她⾃⼰的最新款宝时捷。
6. {昨天领导们⼀起吃了顿饭/胡镇长早已对权⼒⽃争失去了兴趣}。饭局上，{胡镇长/他}得
知政委已决定{放弃/剥夺}他⾃⼰的⾏政权⼒。
7. {天底下就没有不透风的墙/⼩美是班⾥的⼊团积极分⼦}。昨天，{⼩美/她}听说李芳已经
{提交/调查}了她⾃⼰的个⼈档案。
8. {集团与合作伙伴进⾏了⼀次谈判/周鹏参加了集团的⼀次商业谈判}。事后，{周鹏/他}得
知吴部长并没有{发挥/表扬}他⾃⼰的谈判才能。
9. {国企的⽇⼦其实很无聊/朱洁对国企的⽣活很反感}。⽐如，{朱洁/她}注意到周阿姨经常
{卖弄/吹捧}她⾃⼰的语⾔天赋。
10. {试卷批改⼯作很快就结束了/⼩华是班级⾥最聪明的学⽣}。成绩出来后，{⼩华/他}推
断李明肯定{写完/抄袭}了他⾃⼰的数学试卷。
11. {新闻发布会上⼈头攒动/乐乐早已是各⼤媒体的红⼈}。⾯对媒体时，{乐乐/她}说王婷
悄悄地{贡献/窃取}了她⾃⼰的股票分红。

272
12. {警局⾥的⽓氛极其凝重/愁眉苦脸的李⽼板最终报了警}。⾯对质询时，{李⽼板/他}表
⽰⽼吴早已{抛售/盗取}了他⾃⼰的企业资产。
13. {六⼗年代，政治氛围格外紧张/张师傅对⼈⼀直⼼存戒备}。⼀天，{张师傅/他}发现徐
班长打算{隐藏/曝光}他⾃⼰的海外关系。
14. {傍晚，电视台⼤楼⾥灯⽕通明/傍晚，⽟洁在电视台接受采访}。演播室⾥，{⽟洁/她}
透露冰冰前年{捐献/继承}了她⾃⼰的多处房产。
15. {⼜到了开班会的⽇⼦/娜娜是班级⾥的卫⽣委员}。班会上，{娜娜/她}表⽰雪婷刚才{承
认/批评}了她⾃⼰的鲁莽⾏为。
16. {新⼀批的积极分⼦都很优秀/⽟霞最近表现得⼗分积极}。昨天，{⽟霞/她}听说⼩雪已
经{提交/批准}了她⾃⼰的⼊党申请。
17. {股东⼤会刚刚结束/杜芸在焦急地等待会议结果}。很快，{杜芸/她}得知佳佳在⼤会上
{交代/反对}了她⾃⼰的收购计划。
18. {庭院⾥的邻居们都聊着天/⼀向默不作声的叶倩终于说话了}。刚才，{叶倩/她}透露丁
丽上个⽉{缴纳/克扣}了她⾃⼰的房租和押⾦。

273
19. {这场政治风波终于落幕了/宋迪在政治博弈中站错了队伍}。报纸上，{宋迪/他}看见领
导发⽂{放弃/剥夺}了他⾃⼰的竞选资格。
20. {公司刚刚开了⼀个会/朱军在会议上勇于发⾔}。听说，{朱军/他}觉得总监应该{撤回/
奖励}他⾃⼰的营销⽅案。

Experiments 5a & 5b
1. ⼩明是班级⾥的优秀学⽣。课上，他{告诉别⼈/听说}王教授{发表/批改}了⾃⼰的学术论
⽂。
2.董事长主持了今天早晨的例会。会上，他{告诉别⼈/得知}李主任{炫耀/称赞}了⾃⼰的管
理才能。
3.⼩芳是区政府的⼀位业务员。下班后，她{告诉别⼈/听说}杨姐{隐瞒/揭露}了⾃⼰的真实
⽬的。
4.省长惴惴不安、夜不能寐。饭桌上，他{告诉别⼈/得知}书记{承认/举报}了⾃⼰的灰⾊收
⼊。

274
5.张霞⼗分喜爱刚买的跑车。派对上，她{告诉别⼈/听说}李莉{抵押/看上}了⾃⼰的最新款
宝时捷。
6. {胡镇长早已对权⼒⽃争失去了兴趣。饭局上，他{告诉别⼈/听说}政委{放弃/剥夺}了⾃
⼰的⾏政权⼒。
7.⼩美是班⾥的⼊团积极分⼦。昨天，她{告诉别⼈/听说}李芳{提交/调查}了⾃⼰的个⼈档
案。
8.周鹏参加了集团的⼀次商业谈判。事后，他{告诉别⼈/得知}吴部长{发挥/表扬}了⾃⼰的
谈判才能。
9.朱洁对国企的⽣活很反感。⽐如，她{告诉别⼈/听说}周阿姨{常卖弄/常吹捧}⾃⼰的语⾔
天赋。
10.⼩华是班级⾥最聪明的学⽣。成绩出来后，他{告诉别⼈/听说}李明{写完/抄袭}了⾃⼰
的数学试卷。
11.乐乐早已是各⼤媒体的红⼈。⾯对媒体时，她{告诉别⼈/得知}王婷{贡献/窃取}了⾃⼰的
股票分红。

275
12.愁眉苦脸的李⽼板最终报了警。⾯对质询时，他{告诉别⼈/得知}⽼吴{抛售/盗取}了⾃
⼰的企业资产。
13.张师傅对⼈⼀直⼼存戒备。⼀天，他{告诉别⼈/听说}徐班长{隐藏/曝光}了⾃⼰的海外
关系。
14.傍晚，⽟洁在电视台接受采访。演播室⾥，她{告诉别⼈/得知}冰冰{捐献/继承}了⾃⼰
的多处房产。
15.娜娜是班级⾥的卫⽣委员。班会上，她{告诉别⼈/听说}雪婷{承认/批评}了⾃⼰的鲁莽
⾏为。
16.⽟霞最近表现得⼗分积极。昨天，她{告诉别⼈/得知}⼩雪{提交/批准}了⾃⼰的⼊党申
请。
17.杜芸在焦急地等待会议结果。很快，她{告诉别⼈/听说}佳佳{交代/反对}了⾃⼰的收购
计划。
18.⼀向默不作声的叶倩终于说话了。刚才，她{告诉别⼈/听说}丁丽{缴纳/克扣}了⾃⼰的
房租和押⾦。

276
19.宋迪在政治博弈中站错了队伍。报纸上，他{告诉别⼈/看见}领导{放弃/剥夺}了⾃⼰的
竞选资格。
20.朱军在会议上勇于发⾔。听说，他{告诉别⼈/听说}总监{撤回/奖励}了⾃⼰的营销⽅案。

Experiments 6a & 6b
1. ⼩明是班级⾥的优秀学⽣。课上，他{告诉别⼈/听说}王教授{发表/批改}了他⾃⼰的学术
论⽂。
2.董事长主持了今天早晨的例会。会上，他{告诉别⼈/得知}李主任{炫耀/称赞}了他⾃⼰的
管理才能。
3.⼩芳是区政府的⼀位业务员。下班后，她{告诉别⼈/听说}杨姐{隐瞒/揭露}了她⾃⼰的真
实⽬的。
4.省长惴惴不安、夜不能寐。饭桌上，他{告诉别⼈/得知}书记{承认/举报}了他⾃⼰的灰⾊
收⼊。
5.张霞⼗分喜爱刚买的跑车。派对上，她{告诉别⼈/听说}李莉{抵押/看上}了她⾃⼰的最新
款宝时捷。

277
6. {胡镇长早已对权⼒⽃争失去了兴趣。饭局上，他{告诉别⼈/听说}政委{放弃/剥夺}了他
⾃⼰的⾏政权⼒。
7.⼩美是班⾥的⼊团积极分⼦。昨天，她{告诉别⼈/听说}李芳{提交/调查}了她⾃⼰的个⼈
档案。
8.周鹏参加了集团的⼀次商业谈判。事后，他{告诉别⼈/得知}吴部长{发挥/表扬}了他⾃⼰
的谈判才能。
9.朱洁对国企的⽣活很反感。⽐如，她{告诉别⼈/听说}周阿姨{常卖弄/常吹捧}她⾃⼰的语
⾔天赋。
10.⼩华是班级⾥最聪明的学⽣。成绩出来后，他{告诉别⼈/听说}李明{写完/抄袭}了他⾃
⼰的数学试卷。
11.乐乐早已是各⼤媒体的红⼈。⾯对媒体时，她{告诉别⼈/得知}王婷{贡献/窃取}了她⾃⼰
的股票分红。
12.愁眉苦脸的李⽼板最终报了警。⾯对质询时，他{告诉别⼈/得知}⽼吴{抛售/盗取}了他
⾃⼰的企业资产。

278
13.张师傅对⼈⼀直⼼存戒备。⼀天，他{告诉别⼈/听说}徐班长{隐藏/曝光}了他⾃⼰的海
外关系。
14.傍晚，⽟洁在电视台接受采访。演播室⾥，她{告诉别⼈/得知}冰冰{捐献/继承}了她⾃
⼰的多处房产。
15.娜娜是班级⾥的卫⽣委员。班会上，她{告诉别⼈/听说}雪婷{承认/批评}了她⾃⼰的鲁
莽⾏为。
16.⽟霞最近表现得⼗分积极。昨天，她{告诉别⼈/得知}⼩雪{提交/批准}了她⾃⼰的⼊党
申请。
17.杜芸在焦急地等待会议结果。很快，她{告诉别⼈/听说}佳佳{交代/反对}了她⾃⼰的收
购计划。
18.⼀向默不作声的叶倩终于说话了。刚才，她{告诉别⼈/听说}丁丽{缴纳/克扣}了她⾃⼰
的房租和押⾦。
19.宋迪在政治博弈中站错了队伍。报纸上，他{告诉别⼈/看见}领导{放弃/剥夺}了他⾃⼰
的竞选资格。

279
20.朱军在会议上勇于发⾔。听说，他{告诉别⼈/听说}总监{撤回/奖励}了他⾃⼰的营销⽅
案。

Experiment 7
1. ⼩明是班级⾥的优秀学⽣。课上，他{告诉别⼈/听说}王教授重读了{⾃⼰/他⾃⼰}的学术
论⽂。
2.董事长主持了今天早晨的例会。会上，他{告诉别⼈/得知}李主任重申了{⾃⼰/他⾃⼰}的
管理才能。
3.⼩芳是区政府的⼀位业务员。下班后，她{告诉别⼈/听说}杨姐说出了{⾃⼰/她⾃⼰}的真
实⽬的。
4. 省长刚刚出席了⼀个私⼈酒席。饭桌上，他{告诉别⼈/得知}书记修订了{⾃⼰/他⾃⼰}的
⼯作报告。
5.张霞⼗分喜爱豪华跑车。派对上，她{告诉别⼈/听说}李莉喜欢{⾃⼰/她⾃⼰}的最新款宝
时捷。

280
6. {胡镇长早已对权⼒⽃争失去了兴趣。饭局上，他{告诉别⼈/听说}政委了解{⾃⼰/他⾃⼰}
窘迫处境。
7.⼩美是班⾥的⼊团积极分⼦。昨天，她{告诉别⼈/听说}李芳丢失了{⾃⼰/她⾃⼰}的⼊团
资料。
8.周鹏参加了集团的⼀次商业谈判。事后，他{告诉别⼈/得知}吴部长发现了{⾃⼰/他⾃⼰}
的谈判才能。
9.朱洁是国企的⼀名财务主管。⼀天，她{告诉别⼈/听说}周阿姨常谈及{⾃⼰/她⾃⼰}的语
⾔天赋。
10.⼩华是班级⾥的学习委员。成绩出来后，他{告诉别⼈/听说}李明不相信 {⾃⼰/他⾃⼰}
的数学成绩。
11.乐乐早已是各⼤媒体的红⼈。前不久，她{告诉别⼈/听说}王婷公布了{⾃⼰/她⾃⼰}的持
股份额。
12.⼩李最近⼀直愁眉苦脸。上周，他{告诉别⼈/听说}⽼吴评估了{⾃⼰/他⾃⼰}的企业资
产。

281
13.张师傅在军队⾥⼯作了很多年。⼀天，他{告诉别⼈/听说}徐班长动⽤了{⾃⼰/他⾃⼰}
的海外关系。
14.⽟洁上周在电视台接受了采访。访谈中，她{告诉⼤家/听说}冰冰很重视{⾃⼰/她⾃⼰}
的投资项⽬。
15.娜娜是班级⾥的卫⽣委员。班会上，她{告诉别⼈/听说}雪婷汇报了{⾃⼰/她⾃⼰}的勇
敢⾏为。
16.⽟霞最近表现得⼗分积极。昨天，她{告诉别⼈/得知}⼩雪检查了{⾃⼰/她⾃⼰}的⼊党
申请。
17.杜芸在焦急地等待会议结果。会前，她{告诉别⼈/听说}佳佳将提及{⾃⼰/她⾃⼰}的收
购计划。
18.叶倩平时很少和别⼈交流。⼀天，她{告诉别⼈/听说}丁丽付了{⾃⼰/她⾃⼰}的房租和
押⾦。
19.宋迪是政治局的青年代表。今天，他{告诉别⼈/听说}领导考虑了{⾃⼰/他⾃⼰}的竞选
资格。

282
20.朱军在会议上勇于发⾔。会后，他{告诉别⼈/听说}总监修改了{⾃⼰/他⾃⼰}的营销⽅
案。

Experiments 8a & 8b
1.⼩明是班级⾥的优秀学⽣。课上，他听说{王教授/我}刚刚{发表/批改}了⾃⼰的学术论⽂。
2.董事长主持了今天早晨的例会。会上，他发现{李主任/我}⼀直{炫耀/称赞}⾃⼰的管理才
能。
3.⼩芳是区政府的⼀位业务员。下班后，她说{杨姐/我}向上级{隐瞒/揭露}了⾃⼰的真实⽬
的。
4.省长惴惴不安、夜不能寐。饭桌上，他透露{书记/我}很可能{承认/举报}了⾃⼰的灰⾊收
⼊。
5.张霞⼗分喜爱豪华跑车。派对上，她听说{李莉/我}竟然{抵押/看上}了⾃⼰的最新款宝时
捷。
6.胡镇长早已对权⼒⽃争失去了兴趣。饭局上，他得知{政委/我}已决定{放弃/剥夺}⾃⼰的
⾏政权⼒。

283
7.⼩美是班⾥的⼊团积极分⼦。昨天，她听说{李芳/我}已经{提交/调查}了⾃⼰的个⼈档案。
8.周鹏参加了集团的⼀次商业谈判。事后，他得知{吴部长/我}并没有{发挥/表扬}⾃⼰的谈
判才能。
9.朱洁是国企的⼀名财务主管。最近，她注意到{周阿姨/我}经常{卖弄/吹捧}⾃⼰的语⾔天
赋。
10.⼩华是班级⾥最聪明的学⽣。成绩出来后，他推断{李明/我}肯定{写完/抄袭}了⾃⼰的
数学试卷。
11.乐乐早已是各⼤媒体的红⼈。⾯对媒体时，她说{王婷/我}悄悄地{贡献/窃取}了⾃⼰的股
票分红。
12.愁眉苦脸的李⽼板最终报了警。⾯对质询时，他表⽰{⽼吴/我}早已{抛售/盗取}了⾃⼰
的企业资产。
13.张师傅对⼈⼀直⼼存戒备。⼀天，他发现{徐班长/我}打算{隐藏/曝光}⾃⼰的海外关系。
14.傍晚，⽟洁在电视台接受采访。演播室⾥，她透露{冰冰/我}前年{捐献/继承}了⾃⼰的
多处房产。

284
15.娜娜是班级⾥的卫⽣委员。班会上，她表⽰{雪婷/我}刚才{承认/批评}了⾃⼰的鲁莽⾏
为。
16.⽟霞最近表现得⼗分积极。昨天，她听说{⼩雪/我}已经{提交/批准}了⾃⼰的⼊党申请。
17.杜芸在焦急地等待会议结果。很快，她得知{佳佳/我}在⼤会上{交代/反对}了⾃⼰的收
购计划。
18.⼀向默不作声的叶倩终于说话了。刚才，她透露{丁丽/我}上个⽉{缴纳/克扣}了⾃⼰的
房租和押⾦。
19.宋迪在政治博弈中站错了队伍。报纸上，他看见{领导/我}发⽂{放弃/剥夺}了⾃⼰的竞
选资格。
20.朱军在会议上勇于发⾔。听说，他觉得{总监/我}应该{撤回/奖励}⾃⼰的营销⽅案。

Experiments 9a, 9b & 9c
1.⼩明是班级⾥的优秀学⽣。课上，他告诉{别⼈/我}王教授(刚刚){发表/批改}了⾃⼰的学
术论⽂。

285
2.董事长主持了今天早晨的例会。会上，他告诉{别⼈/我}李主任(⼀直){炫耀/称赞}⾃⼰的
管理才能。
3.⼩芳是区政府的⼀位业务员。下班后，她告诉{同事/我}杨姐(向上级){隐瞒/揭露}了⾃⼰
的真实⽬的。
4.省长惴惴不安、夜不能寐。饭桌上，他告诉{别⼈/我}书记(很可能){承认/举报}了⾃⼰的
灰⾊收⼊。
5.张霞⼗分喜爱豪华跑车。派对上，她告诉{别⼈/我}李莉(竟然){抵押/看上}了⾃⼰的最新
款宝时捷。
6.胡镇长早已对权⼒⽃争失去了兴趣。饭局上，他告诉{别⼈/我}政委(已决定){放弃/剥夺}
⾃⼰的⾏政权⼒。
7.⼩美是班⾥的⼊团积极分⼦。昨天，她告诉{别⼈/我}李芳(已经){提交/调查}了⾃⼰的个
⼈档案。
8.周鹏参加了集团的⼀次商业谈判。事后，他告诉{别⼈/我}吴部长(并没有){发挥/表扬}⾃
⼰的谈判才能。

286
9.朱洁是国企的⼀名财务主管。⼀天，她告诉{别⼈/我}周阿姨(经常){卖弄/吹捧}⾃⼰的语
⾔天赋。
10.⼩华是班级⾥最聪明的学⽣。成绩出来后，他告诉{⽼师/我}李明(肯定){写完/抄袭}了⾃
⼰的数学试卷。
11.乐乐早已是各⼤媒体的红⼈。就在刚才，她告诉{记者/我}王婷(悄悄地){贡献/窃取}了⾃
⼰的股票分红。
12.愁眉苦脸的李⽼板最终报了警。⾯对质询时，他告诉{警⽅/我}⽼吴(早已){抛售/盗取}了
⾃⼰的企业资产。
13.张师傅对⼈⼀直⼼存戒备。⼀天，他告诉{别⼈/我}徐班长(打算){隐藏/曝光}⾃⼰的海外
关系。
14.傍晚，⽟洁在电视台接受采访。演播室⾥，她告诉{观众/我}冰冰(前年){捐献/继承}了⾃
⼰的多处房产。
15.娜娜是班级⾥的卫⽣委员。班会上，她告诉{别⼈/我}雪婷(刚才){承认/批评}了⾃⼰的鲁
莽⾏为。

287
16.⽟霞最近表现得⼗分积极。昨天，她告诉{别⼈/我}⼩雪(已经){提交/批准}了⾃⼰的⼊党
申请。
17.杜芸在焦急地等待会议结果。私下⾥，她告诉{别⼈/我}佳佳(在⼤会上){交代/反对}了⾃
⼰的收购计划。
18.⼀向默不作声的叶倩终于说话了。刚才，她告诉{别⼈/我}丁丽(上个⽉){缴纳/克扣}了⾃
⼰的房租和押⾦。
19.宋迪在政治博弈中站错了队伍。不久前，他告诉{⼤家/我}领导(发⽂){放弃/剥夺}了⾃⼰
的竞选资格。
20.朱军是总部新⼊职的员⼯。聚会上，他告诉{⼤家/我}总监(应该){撤回/奖励}⾃⼰的营销
⽅案。

Abstract (if available)

Abstract A reflexive pronoun (e.g., himself) depends on another referent in the discourse for interpretation. But when multiple referents exist, a comprehender needs to identity which referent the reflexive is to be anchored to. The main goal of this dissertation is to use linguistic experiments to examine how syntactic factors (e.g., locality, syntactic position) and non-syntactic factors (e.g., verb semantics, logophoricity, perspectival information) impact reflexive resolution in Chinese and how these factors influence real-time processing. There are two general goals for this dissertation.

First, using linguistic experiments, this work aims to shed light on the linguistic nature of two reflexives in Mandarin, the morphologically simple ziji (‘self’) and the morphologically complex ta-ziji (‘s/he-self’), in order to re-examine conventional linguistic judgment and, more importantly, to provide new data contributing to the advances of linguistic theories.

Second, this dissertation assesses and compares two competing antecedent retrieval models that have been proposed in the sentence processing literature, the standard cue-based retrieval model (e.g., Lewis & Vasishth 2005; Jäger et al. 2015; Patil et al. 2016) and the structure-based retrieval model (e.g., Sturt 2003; Xiang et al. 2009; Dillon et al. 2014; Kush et al. 2015). The central question is whether syntactic information is prioritized in early-stage processing compared to non-syntactic information.

The antecedent choice judgment results show that, from a linguistic-descriptive point of view, both ziji and ta-ziji are subject to syntactic, semantic, and discourse-level constraints. However, the extent to which discourse topic prominence impacts the interpretation of ziji and ta-ziji differs: ziji is more heavily influenced by the discourse topic status of the potential antecedent compared to ta-ziji. I ascribe this difference to the different linguistic properties of these two reflexive forms: ziji is perspective-sensitive (empathic) while ta-ziji is not.

Critical for the assessment of antecedent retrieval models, the self-paced reading studies show that when discourse-level information is considered in the experimental design, syntax does not play a predominant role in antecedent retrieval. Therefore, the structure-based retrieval model cannot account for the absence of a syntax-first effect and the presence of early discourse-related effects. The experimental results are thus more consistent with the standard retrieval model. However, a limitation has been found with the standard retrieval model as well because it cannot predict the absence of object blocking effect, where a 1st-person pronoun (sentence-external perspective center) in the object position cannot block long-distance binding by a discourse topic empathy locus (sentence-internal perspective center). To explain this finding, I propose a conjunction constraint to restrict antecedent retrieval to the perspective center in a syntactically prominent (i.e., subject) position. The weak blocking effect is presumably a late-stage effect which only appears in the offline antecedent judgment experiments.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Dynamics of multiple pronoun resolution

PDF

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

PDF

Vietnamese pronouns in discourse

PDF

Discourse-level processing of nominal possessive constructions

PDF

Factors affecting relative clause processing in Mandarin

PDF

Discourse level processing and pronoun interpretation

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

Processing the dynamicity of events in language

PDF

Dominant language influence in acquisition and attrition of the Chinese reflexive ziji By Chinese-English bilinguals

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

Discourse particles in Mandarin Chinese

PDF

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

PDF

A unified syntactic account of Mandarin subject nominals

PDF

Towards the unity of movement: implications from verb movement in Cantonese

PDF

How to think before you speak: getting from abstract thoughts to sentences

PDF

Number marking and definiteness in Bangla

PDF

Perspective in Turkish complementation

PDF

Interpretation of pronominal forms across languages

PDF

Comparative iIlusions at the syntax-semantics interface

Asset Metadata

Creator Lyu, Jun (author)

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

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Linguistics

Degree Conferral Date 2023-05

Publication Date 03/13/2023

Defense Date 02/14/2023

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

Tag anaphora,Chinese,discourse,OAI-PMH Harvest,reflexive,sentence processing,syntax

Format theses (aat)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Kaiser, Elsi (committee chair), Li, Audrey (committee member), Mintz, Toben (committee member), Simpson, Andrew (committee member)

Creator Email junlyu@shisu.edu.cn,junlyu@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC112764297

Unique identifier UC112764297

Identifier etd-LyuJun-11498.pdf (filename)

Legacy Identifier etd-LyuJun-11498

Document Type Dissertation

Format theses (aat)

Rights Lyu, Jun

Internet Media Type application/pdf

Type texts

Source 20230313-usctheses-batch-1009 (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