ABSTRACT 
 
Title of document:  STRUCTURED ACCESS IN SENTENCE 
COMPREHENSION 
 
Brian W. Dillon 
Doctor of Philosophy, 2011 
 
Directed by:   Professor Colin Phillips 
    Department of Linguistics 
 
 
 
Abstract: This thesis is concerned with the nature of memory access during the 
construction of long-distance dependencies in online sentence comprehension. In 
recent years, an intense focus on the computational challenges posed by long-distance 
dependencies has proven to be illuminating with respect to the characteristics of the 
architecture of the human sentence processor, suggesting a tight link between general 
memory access procedures and sentence processing routines (Lewis & Vasishth 2005; 
Lewis, Vasishth, & Van Dyke 2006; Wagers, Lau & Phillips 2009). The present 
thesis builds upon this line of research, and its primary aim is to motivate and defend 
the hypothesis that the parser accesses linguistic memory in an essentially structured 
fashion for certain long-distance dependencies. In order to make this case, I focus on 
the processing of reflexive and agreement dependencies, and ask whether or not non-
 structural information such as morphological features are used to gate memory access 
during syntactic comprehension. Evidence from eight experiments in a range of 
methodologies in English and Chinese is brought to bear on this question, providing 
arguments from interference effects and time-course effects that primarily syntactic 
information is used to access linguistic memory in the construction of certain long-
 distance dependencies. The experimental evidence for structured access is compatible 
with a variety of architectural assumptions about the parser, and I present one 
implementation of this idea in a parser based on the ACT-R memory architecture. In 
the context of such a content-addressable model of memory, the claim of structured 
access is equivalent to the claim that only syntactic cues are used to query memory. I 
argue that structured access reflects an optimal parsing strategy in the context of a 
noisy, interference-prone cognitive architecture: abstract structural cues are favored 
over lexical feature cues for certain structural dependencies in order to minimize 
memory interference in online processing. 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  
 
 
 
 
 
 
 
 
 
 
 
 
 
STRUCTURED ACCESS IN SENTENCE COMPREHENSION 
 
by 
 
Brian William Dillon 
 
 
 
 
Dissertation submitted to the Faculty of the Graduate School of the  
University of Maryland, College Park, in partial fulfillment  
of the requirements for the degree of  
Doctor of Philosophy  
2011 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Advisory Committee: 
Professor Colin Phillips, Chair  
Professor Norbert Hornstein 
Professor William Idsardi 
Professor Jeffrey Lidz 
External:  Professor Robert DeKeyser, SLA 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
? Copyright by 
Brian William Dillon 
2011 
 
 
 
 
 
 
 
 
 
 
 
 
 
ii 
 
Acknowledgments 
 
I?d like to thank first and foremost Colin Phillips for all the support he?s given me 
over the last six years. Colin has spent a significant amount of time patiently listening 
to me and my half-baked ideas week after week, helping me to sharpen those ideas 
while simultaneously teaching me how to be a responsible and engaged scientist. 
There?s no question that he has really been an all-around top-notch advisor. I?m still 
puzzled as to why he thought I was qualified to run an EEG lab way back when, but 
I?m thankful that he gave me the chance; I wouldn?t be where I am today if he hadn?t 
thought so. My time at Maryland has been an extremely frustrating and extremely 
rewarding experience that I wouldn?t trade for anything. For all his time and energy 
that he?s given me over the years, I owe a great debt to him that I really can?t sum up 
in a paragraph. So I?ll just leave it this: thank you, Colin! 
  
I?ve also been very lucky to work with Bill Idsardi during my time at Maryland. I?m 
thankful for his all-in support and encouragement in pursuing my research ideas, and 
his incredibly diverse approach to research questions in cognitive science has been an 
inspiration along the way. From the highest-level discussions of our work to the 
minute details of hierarchical clustering, it seems there was nothing that I couldn?t 
talk to Bill about, and he was always willing to lend an ear.   
 
Jeff Lidz was also a huge help through the years. His excitement for language 
research was an important source of encouragement in frustrating times. I benefitted 
from his insight on too many occasions to count, and he was never too busy to find 
the time to talk (or, if he was, he didn?t let on). 
 
Of course, many thanks are due to Norbert Hornstein for his daily afternoon cookie 
deliveries, but more importantly, I?m thankful for his many non-cookie related visits 
to 1413 H. I?ve really enjoyed and learned a lot from our discussions over the years, 
and I?m going to miss them.    
 
So many people deserve thanks for the help and friendship they?ve given me along 
the way. Thanks to Ming Xiang, who has been a good friend and colleague since the 
beginning. Ming has a special talent for keeping things in perspective and it?s been 
great to work with her over the years. Thanks to Matt Wagers, who I?ve learned a lot 
from over the years, who very patiently taught me how to run SAT, and who?s been a 
good friend to boot.  
 
Thanks also to my good friends and classmates: Pedro Alcocer, Annie Gagliardi, and 
Shannon Hoerner have helped me time and time again to relax and not take things so 
seriously, and Alex Drummond, Dave Kush, and Terje Lohndal have given me many 
impromptu syntax lessons over the years. Thanks to Ewan Dunbar for non-stop math 
fun. Thanks to Wing Yee Chow for all her help and discussion over the years; a good 
deal of the research in this thesis would not have been possible without her help.  
 
 
 
iii 
I also feel lucky to have been part of a phenomenal lab during my time at Maryland, 
and I?m going to miss everyone from the UMD CNL lab, past and present. Thanks to 
everyone, seriously.   
 
The ideas presented in this thesis have benefitted from discussions with many, many 
people. In particular I?d like to thank Rick Lewis and Shravan Vasishth, who have 
both given me a lot of support and helpful feedback on this work. The computational 
modeling in this paper would not have been possible without Rick?s guidance. 
Additionally, I am grateful to Taomei Guo, who very generously provided me with 
support for running the Chinese experiments reported here. Special thanks are also 
due to a number of amazing researchers who have helped me develop ideas or given 
me helpful guidance at several stages in this thesis: Rajesh Bhatt, Lyn Frazier, Roger 
Levy, Brian McElree, Adrian Staub, and Amy Weinberg.  
 
The research I report here was supported by a number of outstanding research 
assistants who I?ve had a lot of fun working with. Many thanks to Peiyao Chen, 
Fengqin Liu, Alan Mishler, Mike Shvartsman, Shayne Sloggett, and Angela Stanley. 
 
Last but definitely not least: thank you, Jorge, for being my best friend throughout all 
of this and for being so supportive of my choices over the years. I?m incredibly lucky 
to have you in my life, and I hope to be so lucky for a long time to come. 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
iv 
Table of Contents 
 
 
v 
 
 
vi 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
vii 
List of Tables 
 
Table 2.1: Summary of agreement conditions in Experiment 1. Critical and spillover 
regions included in the analysis are underlined. ....................................................... 50
 Table 2.2: Summary of reflexive conditions in Experiment 1. Critical and spillover 
regions included in the analysis are underlined. ....................................................... 50
 Table 2.3: Mean judgments and standard error by subjects for Experiment 1 rating 
study. Values are on a 7-point scale where 7 is perfectly acceptable, and 1 is 
completely unacceptable. .......................................................................................... 52
 Table 2.4: Table of means (in ms where applicable) for agreement conditions for first 
pass, total time, and probability of regression. Standard error by participant is 
shown in parentheses. ............................................................................................... 58
 Table 2.5: Table of means (in ms where applicable) for reflexive conditions for first 
pass, total time, and probability of regression. Standard error by participant is 
shown in parentheses. ............................................................................................... 60
 Table 2.6: Summary of fixed effects for best-fit models on agreement conditions at 
the critical agreeing verb region, including t-values (z-values for first-pass 
regression probability data). An asterisk (*) indicates significance at ? = 0.05, 
while a cross (?) indicates significance at ? = 0.10. First-pass and total time 
coefficients are in milliseconds. ................................................................................ 62
 Table 2.7: Summary of fixed effects for best-fit models on reflexive conditions at the 
critical reflexive region, including t-values (z-values for first-pass regression 
probability data). An asterisk (*) indicates significance at ? = 0.05, while a cross 
(?) indicates significance at ? = 0.10. First-pass and total time coefficients are in 
milliseconds. ............................................................................................................. 63
 Table 2.8: Summary of agreement conditions in Experiment 2. Regions included in 
the analysis are underlined. ....................................................................................... 70
 Table 2.9: Table of means (in ms where applicable) for Experiment 2, agreement 
conditions with a singular head noun, for first pass, total time, and probability of 
regression. Standard error by participant is shown in parentheses. .......................... 71
 Table 2.10: Table of means (in ms where applicable) for Experiment 2, agreement 
conditions with a plural head noun, for first pass, total time, and probability of 
regression. Standard error by participant is shown in parentheses. .......................... 72
 Table 2.11: Summary of fixed effects for best-fit models at the critical agreeing verb 
region in Experiment 2, including t-values (z-values for first-pass regression 
probability data). An asterisk (*) indicates significance at ? = 0.05; a cross (?) 
indicates significance at ? = 0.10. First-pass and total time coefficients are in 
milliseconds. ............................................................................................................. 74
 Table 2.12: Summary of reflexive conditions in Experiment 3. Regions included in 
the analysis are underlined. ....................................................................................... 81
 Table 2.13: Table of means (in ms where applicable) for Experiment 3, reflexive 
conditions with a singular head noun, for first pass, total time, and probability of 
regression. Standard error by participant is shown in parentheses. .......................... 82
 Table 2.14: Table of means (in ms where applicable) for Experiment 3, reflexive 
conditions with a plural head noun, for first pass, total time, and probability of 
regression. Standard error by participant is shown in parentheses. .......................... 84
 
 
viii 
Table 2.15: Summary of fixed effects for best-fit models at the critical reflexive 
region in Experiment 3, including t-values (z-values for first-pass regression 
data). An asterisk (*) indicates significance at ? = 0.05; a cross (?) indicates 
significance at ? = 0.10. First-pass and total time coefficients are in milliseconds. . 84
 Table 3.1: Distribution of NP feature match and match of the inaccessible NP to 
retrieval cues across experimental conditions. ........................................................ 154
 Table 4.1: Summary of conditions in experiment: Critical ziji conditions. Critical 
conditions are 1-2 and 4-5; conditions 3 and 6-9 were included for purposes of d? 
scaling (see text). .................................................................................................... 207
 Table 4.2: Difference in parameter estimates between local and long-distance 
configurations on the critical ziji and control conditions. Values greater than 0 
indicate a processing advantage for local antecedents. Standard errors by subject 
are in parentheses. ................................................................................................... 222
 Table 4.3: Summary of conditions in Experiment 2. ..................................................... 233
 Table 4.4: Table of experimental fixed effects (coefficients in ?V, with standard 
error). Experimental fixed effects only shown if the best-fit model included a 
significant interaction of experimental effect with anteriority and laterality . ? = 
p < 0.1, *  = p < 0.05, **  = p < 0.01. ..................................................................... 239
 Table 5.1: Summary of interference properties of long-distance dependencies. ........... 262
 Table 5.2: Critical conditions from Experiment 8. Region breaks are indicated by 
slashes. .................................................................................................................... 277
 Table 5.2: Mean judgments and standard error by subjects for ziji Experiment 8 
rating study. Values are on a 7-point scale where 7 is perfectly acceptable, and 1 
is completely unacceptable. .................................................................................... 279
 Table 5.4: Critical conditions from Experiment 9. Region breaks are indicated by 
slashes. .................................................................................................................... 285
 Table 5.5: Mean judgments and standard error by subjects for ziji Experiment 8 
rating study. Values are on a 7-point scale where 7 is perfectly acceptable, and 1 
is completely unacceptable. .................................................................................... 286
 Table 5.6: Summary of fixed effects for best-fit models at the critical ziji and 
spillover regions, including t-values ....................................................................... 288
 Table 5.7: Summary of access properties of long-distance dependencies. .................... 295
 Table A.1: Constituent creation times and feature makeup for agreement conditions. . 333
 Table A.2: Schedule of retrievals and cue sets. R1 = attachment of DP1 to VP1; R2 = 
attachment of DP2 to VP1; R3 = critical retrieval to attach DP1 to BE. ................ 333
 Table A.3: Constituent creation times and feature makeup for agreement conditions. . 334
 Table A.4: Schedule of retrievals and cue sets for the structured access reflexive 
model. R1 = attachment of DP1 to VP1; R2 = attachment of DP2 to VP1; R3 = 
attachment of DP1 to VP2; R4 = attachment of REFL to VP2; R5 = critical 
retrieval of REFL?s antecedent. .............................................................................. 334
 Table A.5: Schedule of retrievals and cue sets for the feature-based access reflexive 
model. R1 = attachment of DP1 to VP1; R2 = attachment of DP2 to VP1; R3 = 
attachment of DP1 to VP2; R4 = attachment of REFL to VP2; R5 = critical 
retrieval of REFL?s antecedent. .............................................................................. 334
  
 
 
ix 
List of Figures 
 
Figure 1.1: Hypothetical processes for processing a reflexive pronoun, demonstrating 
different modes of antecedent activation. ................................................................... 6
 Figure 1.2: Structured access mechanisms use a narrow syntactic set of cues to access 
the reflexive?s antecedent (left panel). Feature-based access mechanisms deploy 
a wider range of cues that to access the representation of the antecedent (right 
panel). .......................................................................................................................... 7
 Figure 2.1: Target and distractor NPs when processing a reflexive pronoun in 
English. ..................................................................................................................... 27
 Figure 2.2: Multiple match and partial match interference configurations. For 
multiple match interference, the target memory is a perfect match to the search 
cues, but the distractor overlaps in some feature content with the target. In the 
partial match situation, neither target nor distractor is a perfect match to the 
search cues. ............................................................................................................... 29
 Figure 2.3: Interference effects in Experiments 1-3 in Pearlmutter et al (1999). The 
interference effect is the difference in RTs at the critical region that is due to 
manipulating the feature content of the distractor, as shown. ................................... 35
 Figure 2.4: Interference effects in Wagers et al (2009). .................................................. 37
 Figure 2.5: Interference effects for early (first-pass) measures in Sturt (2003). .............. 42
 Figure 2.6: Interference effects for late (total time) measures in Sturt (2003). ............... 43
 Figure 2.7: Mean first-pass reading time at the critical region in Experiment 1. Error 
bars show standard error by participants. ................................................................. 65
 Figure 2.8: Mean total reading time at the critical region in Experiment 1. Error bars 
show standard error by participants. ......................................................................... 66
 Figure 2.9: Mean first-pass reading time at the critical region in Experiment 2. Error 
bars show standard error by participants. ................................................................. 77
 Figure 2.10: Mean total reading time at the critical region in Experiment 2. Error bars 
show standard error by participants. ......................................................................... 78
 Figure 2.11: Mean first-pass reading time at the critical region in Experiment 3. Error 
bars show standard error by participants. ................................................................. 85
 Figure 2.12: Mean total reading time at the critical region in Experiment 3. Error bars 
show standard error by participants. ......................................................................... 86
 Figure 2.13: Interference effects (in ms) observed for in total time measures across 
Experiments 1-3. Error bars reflect 95% CI by participants. .................................... 88
 Figure 2.14: Total reading times (ms) at critical region, combining similar conditions 
across Experiments 1-3. Error bars are standard error by participants. .................... 89
 Figure 3.1: Average activation for target (black) and distractor (red) NPs for a 
sentence that shows partial-match interference at the agreeing verb. Incorrect 
retrievals of the distractor NP are reflected in the increased activation at the 
plural verb were. ..................................................................................................... 131
 Figure 3.2: Average activation for target (black) and distractor (red) NPs for a 
sentence that shows partial-match interference at the reflexive. Incorrect 
retrievals of the distractor NP are reflected in the increased activation at the 
plural reflexive themselves. ..................................................................................... 133
  
 
 
x 
Figure 3.3: Percentage of retrieval of distractor NP, for all parameterizations 
(n=324), at critical probe position for agreement and feature-based reflexive 
models. .................................................................................................................... 141
 Figure 3.4: Difference in interference error between agreement and feature-based 
reflexive models (n=324). Error bars indicate 95% confidence intervals. ............. 142
 Figure 3.5: Predicted interference effect ([+intr]-[-intr] conditions) for agreement and 
feature-based reflexive models (n=324). Error bars indicate 95% confidence 
intervals. .................................................................................................................. 144
 Figure 3.6: Difference in predicted interference effect (agreement-reflexive 
conditions) for all models (n=324). Error bars indicate 95% confidence intervals. 145
 Figure 3.7: Percentage of incorrect retrievals for reflexive conditions for feature-
 based and structured access models (n = 324). ....................................................... 148
 Figure 3.8: Comparison of predicted interference effects (solid) for reflexive 
conditions (n = 324) and observed reflexive interference effects from 
Experiments 1 and 3 (by participants, n = 72). Error bars indicate 95% CI. ......... 149
 Figure 3.9: Relationship between interference error and interference effect on 
average retrieval latency. Blue points indicate comparisons between grammatical 
conditions, and red points are ungrammatical comparisons, for each of 324 
parameterizations. ................................................................................................... 157
 Figure 3.10: Effect of interference in multiple match and partial match comparisons. 
The inhibition in multiple match interference is driven by decreased retrieval 
latencies on the target noun, due to feature overlap. The facilitation in partial 
match interference is an increased overlap in the distribution of target and 
distractor retrieval distributions. The race aspect of the retrieval process leads to 
an overall facilitation effect, which unambiguously indicates that incorrect 
access has occurred online. ..................................................................................... 160
 Figure 4.1: Architecture of a content-addressable memory, from Gallistel & King 
(2009). Memories consist of three bits, and each bit is probed in parallel for a 
match. In the present case, all memories that contain a 0 in third position are 
returned in response to a retrieval query. ................................................................ 181
 Figure 4.2: Hypothetical SAT curves showing a) two processes that differ in 
asymptotic accuracy only (top panel) and b) two processes that differ in 
processing speed only (bottom panel) (figure from ?ztekin & McElree 2010). 
Vertical and horizontal lines indicate that point at which each curve is at 50% of 
asymptotic accuracy. ............................................................................................... 186
 Figure 4.3: Example of a structured search process for finding ziji?s antecedent in the 
sentence Lisi shuo fengbao hai-le ziji ?Lisi said the storm harmed him?. The 
hypothetical structural cues do not allow comprehenders to rule out 
consideration of the local subject fengbao ?storm?. Thus comprehenders must 
evaluate multiple subject positions in the search for the correct antecedent Lisi. 
This structured access predicts that processing time should grow with the 
number of subject positions that need to be evaluated. .......................................... 199
 Figure 4.4: Example of feature-based access for finding ziji?s antecedent in the 
sentence Lisi shuo fengbao hai-le ziji ?Lisi said the storm harmed him?. The 
mixture of structural and semantic cues allow direct access to the correct 
 
 
xi 
antecedent Lisi. Feature-based access predicts that processing time should be 
constant with the number of subject positions that need to be evaluated. .............. 200
 Figure 4.5: SAT functions for LD and local antecedent ziji conditions with fully 
saturated models (2?-2?-2?), over average data (not averaged parameters). 
Accuracy is scaled to show proportion of asymptote; vertical bars indicate time 
point at which 50% accuracy is reached. ................................................................ 215
 Figure 4.6: SAT functions for LD and local control conditions with fully saturated 
models (2?-2?-2?), over average data (not averaged parameters). Accuracy is 
scaled to show proportion of asymptote; vertical bars indicate time point at 
which 50% accuracy is reached. ............................................................................. 216
 Figure 4.7: Average asymptotic accuracy (?) across individual participant SAT 
function fits with Bayesian parameter estimation. Error bars show ?1 SE, 
corrected for between-participant variance. ............................................................ 217
 Figure 4.8: Average rate (?) across individual participant SAT function fits with 
Bayesian parameter estimation. Error bars show ?1 SE, corrected for between-
 participant variance. ................................................................................................ 218
 Figure 4.9: Average intercept (?) across individual participant SAT function fits with 
Bayesian parameter estimation. Error bars show ?1 SE, corrected for between-
 participant variance. ................................................................................................ 219
 Figure 4.8: Average speed (?+?-1) across individual participant SAT function fits 
with Bayesian parameter estimation. Error bars show ?1 SE, corrected for 
between-participant variance. ................................................................................. 220
 Figure 5.1: Structured search for ziji forces the parser to consider the local antecedent 
position before the sub-commanding antecedent. ................................................... 272
 Figure 5.2: Direct access for ta-ziji based on semantic or discourse prominence 
allows the parser to immediately access the sub-commanding antecedent. ........... 273
 Figure 5.3: Region-by-region mean log reading times for Experiment 8. Error bars 
represent ?1 standard error, by participants, corrected for between-participant 
variance. .................................................................................................................. 283
 Figure 5.4: Region-by-region mean log reading times for Experiment 9. Error bars 
represent ?1 standard error, by participants, corrected for between-participant 
variance. .................................................................................................................. 288
 Figure 5.5: Effect of embedded animate (embedded [+animate] subject ? embedded 
[-animate] subject) on spillover reading times in Experiments 8 (ziji) and 9 (ta-
 ziji). Error bars represent 95% confidence interval, by participants. ...................... 290
 Figure 5.6: Rate of occurrence of gender attraction errors across languages. Figure 
from Lorimor, Bock, Zalkind & Sheyman (2008). ................................................. 311
  
 
 
 
 
 
 
 
 
 
 
1 
Chapter 1: Introduction  
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
In online language comprehension, the information contained in a sentence 
unfolds over time. In order to successfully understand a sentence, a comprehender 
must have a mechanism for maintaining and combining the information contained in 
each of the words of the sentence. Because language comprehenders perceive 
linguistic input in a sequential, left-to-right order, the basic act of understanding a 
sentence must make use of working memory to manage the information conveyed by 
the incoming speech.  
It is tempting to view the working memory system for linguistic 
comprehension as carrying out the relatively straightforward task of combining 
adjacent words into higher-order units of meaning and syntactic structure.  However, 
this simple picture is rapidly complicated by the fact that human language is full of 
?long-distance? dependencies between words. These are relationships between two 
non-adjacent, and possibly quite distant, words in a sentence. One clear example of 
 
 
2 
such a dependency is the relationship between a pronoun (an anaphor) and its 
referent (the antecedent). Other common examples include subject-verb agreement 
and the relationship between fronted wh-words and the verbs they combine with. 
Long-distance dependencies present unique computational challenges to the parser, 
and suggest the need for sophisticated methods of information storage and retrieval 
that are flexible enough to handle the range of these dependencies that comprehenders 
are bound to encounter in every conversation. The nature of these dependencies, and 
the fact that comprehenders on average have little trouble understanding them, lead to 
important questions about the relationship between linguistic structure and working 
memory mechanisms. In this work I will attempt to address the following theoretical 
question: how do linguistic representations and working memory processes interact 
to allow the construction and interpretation of long-distance dependencies?  
In the present work, I attempt to articulate and defend the hypothesis that 
syntactic structure provides the crucial information that aids comprehenders in 
organizing and retrieving information in linguistic working memory stores. My 
primary claim is that for certain long-distance linguistic dependencies comprehenders 
employ a structured access mechanism. For these dependencies, comprehenders 
access linguistic memory by deploying uniquely structural information, selectively 
attending to these features over otherwise useful morphological and semantic 
information. 
This claim may appear unremarkable, as there is no shortage of 
psycholinguistic research that suggests that the grammar is deployed rapidly online to 
structure incoming material (Frazier 1998; Phillips, Wagers & Lau 2010). However, 
 
 
3 
in recent years this view has been challenged on several fronts. A number of 
researchers have suggested that grammatical relations are at best deployed as violable 
constraints online alongside morphological and semantic constraints (Tabor, 
Galantucci & Richardson, 2004; Lewis & Vasishth 2005; Van Dyke 2007); at worst, 
they are not at all deployed in initial parsing (Townsend & Bever 2001; Ferreira, 
Bailey & Ferraro 2002). Ferreira & Patson (2007) provide a useful summary and a 
clear articulation of the opposite position. These challenges reflect very different 
architectural commitments, ranging from claims about the subsymbolic nature of 
online linguistic computation (Tabor et al 2004), to the primacy of heuristic strategies 
in parsing (Townsend & Bever 2001; Ferreira et al 2002), or constraints on the 
representation of linguistic information in working memory (Lewis & Vasishth 2005; 
Van Dyke 2007).  
This last challenge, the difficulty of representing structured syntactic relations 
in online working memory, is the focus of the present work. There is an emerging 
consensus that the computational properties of the sentence processor?s memory 
architecture mirror those found in domain-general working memory (McElree 2000; 
Gordon, Hendrick & Johnson 2001; Gordon, Hendrick & Levine 2002; McElree, 
Foraker & Dyer 2003; Lewis & Vasishth 2005; Lewis, Vasishth & Van Dyke 2006; 
Wagers 2008). One claim that has come to be associated this view is that the parser 
forced to construct syntactically illict representations because of the constraints that 
the memory architecture places on memory access (Van Dyke 2007; Vasishth, 
Br?ssow, Drenhaus & Lewis 2008). However, this claim of structural fallibility 
depends more on specific hypotheses about the type of information used to access 
 
 
4 
memory than hypotheses about the computational properties of the memory 
architecture itself. The primary goal of this thesis is to argue that although the 
memory architecture does place interesting constraints on representation of linguistic 
information during online comprehension, comprehenders are nonetheless able to 
engage structured access mechanisms that effectively target and access specific 
syntactic positions during parsing. 
It should be clear from the outset that in arguing for structured access, I am 
making a claim about the type of information recruited to access working memory in 
parsing long-distance dependencies, rather than an endorsement of any particular 
memory architecture; ?structured access? is intended as a general term for strategies 
that privilege structural information in accessing memory, without implying any 
commitment to a particular theory of memory. Thus the claim of structured access 
stands in contrast to mechanisms that use a wider range of morphological, syntactic 
and semantic features in parallel to access working memory, as has been advocated in 
recent years by a number of authors (Van Dyke & Lewis 2006; Van Dyke 2007; 
Vasishth et al 2008; Wagers, Lau & Phillips 2009). 
To see the difference between the two sorts of account, consider the problem 
of finding an antecedent for a bound reflexive pronoun. There are a number of 
syntactic constraints on which structural positions a reflexive?s antecedent can occupy 
(Chomsky 1981), in addition to a formal requirement of feature concord between the 
reflexive pronoun and its antecedent in English. Upon encountering a reflexive 
pronoun in English, the processor?s task is to construct a legal binding chain, which 
presumably starts by accessing potential antecedents. There are a number of ways that 
 
 
5 
antecedent reactivation might occur. For example, the parser might employ the full 
range of the information in the reflexive to find the correct antecedent, using 
morphological, syntactic, and semantic cues in parallel in a feature-based access 
mechanism (e.g. Badecker & Straub 2002; Patil, Vasishth & Lewis 2011). 
Alternatively, it might engage a structured access mechanism that selectively attends 
to syntactic information in retrieving potential antecedents (Nicol 1988; Sturt 2003a; 
Xiang, Dillon & Phillips 2009). The primary difference between these views is their 
prediction about the impact of structurally inaccessible, but feature matched nouns in 
the process of resolving the reflexive?s reference. A comparison of two sorts of 
process is considered in Figures 1.1 and Figure 1.2. One plausible procedure for 
processing a reflexive involves activating a set of candidate antecedents based on  a 
mixture of morphological, semantic, and syntactic feature information, and then 
constructing the binding dependency with one of the activated antecedents. This 
amounts to a claim of temporary, but spurious ambiguity: the parser is temporarily 
considering two antecedents for the reflexive, even though only one ends up 
grammatically licensed. This temporary ambiguity that arises in the feature-based 
account is schematized in the right panel of Figure 1.2. On a structured access 
account, however, antecedent reactivation proceeds using only syntactic information, 
and so the antecedent selection process is blind to the feature content. Only 
structurally licit antecedents are considered, as in the left panel of Figure 1.2. The 
syntactic information is deployed as a hard constraint, and ambiguity does not arise at 
any level of processing.   
 
 
 
6 
The man [CP who saw John] hurt himself ?  
Structured access Feature-based access  
{the man} {the man, John} Antecedent activation 
<the man=himself> <the man=himself> Binding  
   
Figure 1.1: Hypothetical processes for processing a reflexive pronoun, demonstrating 
different modes of antecedent activation. 
 
Interestingly, for at least some long-distance dependencies, the feature-based 
approach to memory access suggested by the right panel of Figure 1.2 appears to be 
correct. For example, in subject-verb agreement dependencies, morphological 
features appear to be used in the generation of candidates for the agreement relation, 
leading to spurious illusions of grammaticality in agreement formation (Clifton, 
Frazier & Deevy 1999; Pearlmutter, Garnsey & Bock 1999; Wagers et al 2009). This 
finding is expected given recent models of linguistic working memory, and this has 
led some authors to propose that such feature-based access is the primary manner of 
memory access (Lewis et al 2006; Van Dyke 2007; Vasishth et al 2008; Patil et al 
2011).  
The generality of feature-based access in online parsing remains unclear, 
however. A number of results appear to be suggestive of structured, rather than 
feature-based access mechanisms online (Nicol 1988; Sturt 2003a; Xiang et al 2009), 
 
 
7 
opening up 
 
Figure 1.2: Structured access mechanisms use a narrow syntactic set of cues to access 
the reflexive?s antecedent (left panel). Feature-based access mechanisms deploy a 
wider range of cues that to access the representation of the antecedent (right panel). 
 
the possibility that there are distinct strategies that comprehenders employ to organize 
and access information during parsing. This thesis explores this possibility and builds 
support for a uniquely structural access mechanism in comprehension. The main 
prediction of a structured access mechanism is that whenever it is engaged, 
comprehenders should selectively retrieve information based on its syntactic position, 
rather than its feature content. There are two types of empirical evidence that I offer 
for this claim. In chapters 2 and 3, I present experimental evidence that feature-
 matching but structurally inaccessible antecedents do not impact early memory access 
for English reflexives. Chapters 4 and 5 demonstrate the converse situation: even in 
S
 NP VP
 The man RC V NP
 who S hurt himself
 NP VP
 t V NP
 saw John
 S
 NP VP
 The man RC V NP
 who S hurt himself
 NP VP
 t V NP
 saw John
 +c-command
 +NP
 +c-command
 +NP
 +masculine
 +singular
 +c-command
 +NP
 +masculine
 +singular
 
 
8 
the presence of structurally accessible feature-matching candidates, feature-
 mismatching structurally accessible antecedents are accessed in the construction of 
long-distance reflexive dependencies in Mandarin Chinese. Across all studies, 
comprehenders appear to reactivate particular syntactic positions during 
comprehension, in the face of both inaccessible (Chapters 2-3) and accessible 
(Chapters 4-5) feature-matching material. These two sources of evidence confirm the 
central prediction of a structured access theory of memory access: it is primarily 
syntactic position, rather than feature match, that guides access to linguistic memory 
for the dependencies considered. In addition to experimental evidence, I also present 
evidence from explicit computational models that further supports the conclusion that 
memory access proceeds in a structured fashion. Lastly, I will turn to a critical 
assessment of claims that appear to run counter to my central argument, showing that 
the empirical support for feature-based access is actually rather limited.  
Although it is likely that short-term working memory places interesting 
constraints on representing syntactic hierarchy during parsing, the arguments 
presented here stress that these constraints do not inhibit the parser?s ability to engage 
in structured access. Rather, the constraints on representation that may accompany a 
noisy content-addressable cognitive architecture may in fact provide the key to 
understanding the role of structured access in parsing: in Chapter 5 I argue that 
structured access reflects an optimal strategy for an interference-prone parser. In 
particular, by limiting search cues to the most predictive and minimal set of cues, 
disruptive interference can be minimized. In general, deploying uninformative and 
superfluous cues to memory access, such as morphological features in a feature-based 
 
 
9 
access account, increases the risk of memory interference with no countervailing 
benefit for memory access. If this claim is correct, then for the dependencies 
considered here, structured access is a rational strategy for the parser to pursue. In this 
way, the disruptive effect of memory interference actually provides functional 
pressure for abstraction in parsing. 
 
Models of memory and syntactic representation 
 
The role of linguistic structure in memory access has been an active area of 
research for recent work that has focused on the architecture of linguistic memory in 
sentence comprehension. There has been an increasing amount of interest in the fine 
computational details of the memory architecture of the parser, and research on this 
front has become influential in thinking about the relationship between syntactic 
competence and online patterns of processing difficulty.  
One major goal of this recent line of research is an explicit characterization of 
the computational characteristics of the parser?s working memory architecture. The 
starting point for this work was the intuition that insights from research on working 
memory processes from other domains of cognition could ported in a fairly 
straightforward way to model the memory architecture of the language processor. 
One explicit characterization of this intuition is given by Lewis & Vasishth, who state 
that the ?goal? is to explain as much detailed psycholinguistic phenomena as 
possible with independent principles of cognitive processing? (2005: pp. 377).  The 
guiding hypothesis of this research is that a small set of general computational 
 
 
10 
principles govern memory access in sentence comprehension, just as they do in any 
cognitive task that requires retention of information in a short-term memory store 
(McElree 2000; Gordon et al 2001; Gordon et al 2002; McElree, Foraker & Dyer 
2003; Lewis & Vasishth 2005; Lewis, Vasishth & Van Dyke 2006; Wagers 2008; see 
also Greene, McKoon & Ratcliff 1992 for similar ideas in the domain of reference 
resolution). This approach has had considerable empirical purchase and has been 
supported by successful computational models (Lewis & Vasishth 2005; Vasishth et 
al 2008; Wagers 2008).  
Parsing models based on this idea form a heterogeneous group of sentence 
processing theories that are collectively referred to as cue-based approaches to 
parsing. A great deal of the excitement that surrounds these frameworks is the 
promise that whatever principles govern the parser?s behavior are the same general 
principles that are thought to govern information processing across cognitive 
domains. Though memory considerations have long been used to motivate parsing 
principles (Yngve 1960; Miller & Chomsky 1963; Kimball 1973; Frazier & Fodor 
1978), this line of research actually makes the stronger claim that the memory 
systems that enable sentence processing are essentially identical to those recruited for 
more general working memory tasks, with no role for linguistically specialized 
memory mechanisms such as hold cells or stacks (Wanner & Maratsos 1978; Marcus 
1980). This highlights the excitement inherent in these approaches; processing 
principles that have long been established and debated could in principle be a simple 
reflex of ?memory limitations? in a broad sense (see also Bever 1970), even if one 
 
 
11 
maintains that the memory systems for language are separate from those seen in other 
cognitive domains (as in Caplan & Waters 1998; Lewis & Vasishth 2005).  
Although I review the main empirical arguments for cue-based approaches to 
parsing in Chapter 3, as well as their formal characterization, it is worthwhile to 
briefly summarize the main theoretical commitments of this approach to frame the 
discussion that follows. It is important to bear in mind that each of these theoretical 
commitments are independently motivated insofar as they are drawn from theories of 
short-term memory access in more ?domain-general? areas of cognition (i.e. list 
memory, McElree & Dosher 1989). Lewis, Vasishth and Van Dyke (2006) present an 
explicit and succinct characterization of the relevant computational principles 
assumed across various implementations of the cue-based parsing view. The first, and 
arguably the most crucial, is the assumption of a content-addressable memory 
architecture. In a content-addressable architecture, stored pieces of information 
(memories) are indexed and retrieved according to the content of their representation, 
rather than their location in memory (Kohonen 1980). For example, rather than 
storing a wh-filler in a special hold cell (Wanner & Maratsos 1978) for later retrieval, 
in a content-addressable architecture it simply suffices to mark the filler with [+wh] 
content. When needed, the wh-filler can be accessed in light of bearing the crucial 
[+wh] feature in its representation, and it need not be stored in any particular location 
or cell in memory.  
There are several consequences of adopting this manner of indexing and 
retrieving memories. The first is the direct access property of these architectures 
(McElree & Dosher 1989). Direct access refers to the fact that memories with the 
 
 
12 
target content are immediately accessed in memory access, without a need to first 
traverse or check memories that do not match the desired content. This means that 
memories are retrieved in constant time relative to the size of the search space, an 
important point that forms the basis for discussion in Chapter 4. Another consequence 
of adopting a content-addressable architecture is that of retrieval or encoding 
interference (Kohonen 1980). When multiple memories contain the target content (in 
full or in part), then there is a possibility that something other than the desired 
memory will be retrieved, or that the desired memory will be more difficult to 
recover. Put differently, the degree to which a memory has unique content in its 
representation is the degree to which it may be seen as having a unique memory 
location. The less unique its content-defined ?address? is, the less reliable access to 
that memory will be. These two computational properties form the basis for the two 
main empirical arguments?arguments from time course and arguments from 
interference?that have been offered for a content-addressable architecture. 
The second main theoretical commitment of a cue-based parsing approach is 
that of a limited focus of attention. The adoption of a limited focus of attention for 
concurrent processing of elements is well supported in other cognitive domains 
(McElree & Dosher 1989; Lewis et al 2006), but somewhat less is known about the 
size or character of focal attention in sentence processing, and this question is very 
much a focus of current research (Wagers & McElree 2009). In cue-based parsing 
approaches, the assumption of a limited focus of attention attributes a good deal of 
explanatory power to our characterization of memory representation and access. This 
is because, by hypothesis, a limited focus of attention for concurrently processing 
 
 
13 
elements entails that sentence processing involves a good deal of passing information 
between the active processing state and the more passive memory representation state 
(Lewis & Vasishth 2005; Wagers 2008). If sentence-processing crucially relies on a 
skillful shunting of information into and out of active and passive processing states, 
then the manner of retrieving that information and restoring it to attention takes on a 
central role in the characterization of the parsing process. The commitment to a 
limited focus of attention will not be directly addressed in this thesis, although insofar 
as it foregrounds the role for memory access and information retrieval, it is an 
important assumption that underlies the arguments presented here.  
With these two theoretical commitments in mind, a cue-based approach to 
parsing maintains that the normal process of constructing grammatical representations 
of speech input proceeds by storing the pieces of structure in a ?passive? content-
 addressable memory store, and carrying out targeted retrievals of structure in order to 
engage processes related to the retrieved structure. In the remainder of this thesis, the 
terms memory retrieval and memory access are used interchangeably to refer to the 
process by which a given piece of structure is restored from a passive storage state to 
a state that is active for processing. In cue-based models of parsing, the generation of 
structure is by hypothesis parasitic on the retrieval process (Lewis & Vasishth 2005). 
Thus in order to draw more explicit parallels with previous work, it is useful to 
occasionally refer to memory retrieval as the generation of structure. I will use the 
term cue to describe the information that is used to access memory (or, equivalently, 
the information that is used to generate linguistic structure). A cue in the context of 
linguistic processes will generally refer to an atomic feature of the representation used 
 
 
14 
for a given piece of structure, such as a distinguishing semantic or morphological 
feature of a lexical item. A formal model of the retrieval process, and the relationship 
of retrieval cues to the properties of memory retrieval will be presented in Chapter 3. 
It can be seen that the name ?cue-based parsing? is fairly transparent: the approach 
maintains that the cues used to access memory representations represent a major 
informational bottleneck in the parsing process. 
The cue-based parsing approach is compatible with a number of different 
implementations (see, e.g. Lewis & Vasishth 2005 versus Van Dyke 2007), and there 
are a number of theoretical questions that become important upon adopting this 
general framework for understanding parsing. One is the relationship between 
prospective structure-building processes (i.e. ?active parsing?) and retrospective 
memory access processes, a topic that I will not address in detail here. A second 
theoretical question is what exactly the nature of the cues used in parsing is, and what 
their relation to a speaker?s grammatical knowledge is. An in-depth treatment of this 
question is one of the primary goals of this thesis. If we take the parallel between the 
information used to access memory and the information used to generate structure 
seriously, it can be seen that this question is an alternative way of understanding 
familiar questions of information encapsulation in the parsing process: what 
information is used to build structure, and when (Frazier & Clifton 1996)? As such, 
the substantive content of the cue set should provide an important point of debate in 
the context of cue-based parsing frameworks. For these models, the claim of 
structured access may be understood as the claim that the cues used in the initial 
stages of parsing are primarily syntactic in nature. 
 
 
15 
Because the various cue-based parsing models represent the most explicit 
attempts to integrate explicit models of short-term memory and parsing processes, 
much of the discussion in the rest of the thesis will be cast in this framework. 
Likewise, because of the existence relatively explicit computational models, I also 
explore models of my results using one implementation of a cue-based parser to 
support the experimental findings presented here. As stated above, the claim of 
structured access in the context of cue-based parser amounts to the claim that 
primarily syntactic cues are used in memory retrieval, though the theoretical status of 
a relational syntactic ?cue? in this framework is unclear. Again, however, the central 
claim of structured access is independent of the particular implementation I adopt. It 
is important to note that even though I present a model of structured access in a cue-
 based framework, these mechanisms are compatible with a much wider range of 
assumptions about the computational character of the parser. More broadly 
understood, the claim is that only syntactic information drives the generation of 
linguistic structure for a certain set of linguistic dependencies.  
 
Cue-based parsing and the psycholinguistic enterprise 
 
Even if one does not subscribe to the main theoretical commitments of cue-
 based parsing, these models provide a very interesting way of understanding familiar 
questions in psycholinguistics, and new ways of asking and testing these questions. 
One important contribution of this line of work is that it has compelled researchers to 
consider computationally complete models of parsing (Lewis & Vasishth 2005, p. 
 
 
16 
377). As Lewis & Vasishth put it, computationally complete models of parsing are 
those that give a joint characterization of the processes, memories, and control 
structures used by the parser. Memories are the temporary pieces of information that 
are relevant for a given parsing task, processes refer to operations over that 
information, and control structures may be understood as the decision principles that 
determine when different processes are applied. From this point of view, it can be 
seen that much work in sentence processing may be understood as characterizing the 
control structure of the parser. For instance, research on the behavior of the parser in 
the face of ambiguous input (e.g. Kimball 1973, Frazier 1978, et seq) is by definition 
research on the control structure of the parser, being concerned with the principles 
that govern actions taken at choice points in the parse. This work can in principle be 
carried out with relatively minimal commitments to the specifics of memories and 
processes involved in syntactic processing.  
 In the context of the broader goal of developing a computationally complete 
parser, the work on memory architecture in sentence processing may be seen as the 
complement to work on the control structure of parsing. The focus of this line of 
work, and the focus of this thesis, is on characterizing the nature of the parser?s 
memories (i.e. temporary linguistic representations) and processes (i.e. computations 
over that information). This is in no sense a new concern for psycholinguists, as many 
researchers have made substantive claims about memories and process in the parser 
(Fodor, Bever & Garrett 1974; Levelt 1974; Frazier & Fodor 1978; Wanner & 
Maratsos 1978; Marcus 1980; Berwick & Weinberg 1984; Prichett 1993; Sturt, 
Pickering & Crocker 2000, among many others). Put this way, the aim of the present 
 
 
17 
thesis and its relation to prior work can be made clearer. The claim of a structured 
access mechanism makes the claim that linguistic working memory encodes detailed 
syntactic information, and that the parser can selectively attend to this information in 
performing online parsing operations. This stands in contrast to claims that the fine-
 grained information is either not encoded (Townsend & Bever 2001; Ferreira & 
Patson 2007), or that all linguistic information is deployed in parallel during parsing 
(e.g. MacDonald, Pearlmutter & Seidenberg 1994; Vasishth et al 2008). 
Just as in psycholinguistic work that has aimed to characterize the generation 
process, one interesting question for cue-based approaches to parsing concerns the 
generality of the interaction between representation and memory access. On one 
interpretation of this literature (e.g. Greene et al 1992; Van Dyke 2007), it can be 
understood as making the strong claim that a general-purpose memory access 
mechanism and a domain-specific representation are the only components of the 
human sentence processor (see also Ratcliff & McKoon 2008). On this view, the role 
of the grammar is simply to provide declarative representations that are manipulated 
by general cognitive mechanisms. That is, linguistic representation only provides 
domain-specific memories, but no domain-specific processes. This is a view that a 
number of separate research traditions appear to be converging on (for similar 
conclusions from a different point of view, see McDonald et al 1994; Pearlmutter & 
McDonald 1995; Jurafsky 1996; Levy 2008). This stands in contrast to theories that 
posit specialized mechanisms or principles that are operative only in linguistic 
processing (Kimball 1973; Wanner & Maratsos 1978; Marcus 1980; Frazier, Clifton 
& Randall 1983; Berwick & Weinberg 1984). As before, the intuition pursued here is 
 
 
18 
straightforward: if there are existing computational principles that have explanatory 
purchase across multiple domains of cognition (content-addressability, rational 
inference, etc.), then theories that make use of independently motivated mechanisms 
have an advantage over those that invoke more specialized mechanisms.  
In comparing cue-based approaches to the broader psycholinguistic enterprise, 
it can be seen that familiar psycholinguistic questions remain of central interest. 
These involve the types of information that are recruited in online processing, as well 
as the domain-specificity of the parser?s processing routines. The adoption of a 
content-addressable memory architecture does not in and of itself provide an answer 
to the question of what information is used to guide the parse. However, these 
architectures are often associated with a corollary claim that a wide range of linguistic 
features, including morphological and semantic features, are always deployed to 
access memory during parsing, a claim that is also endorsed in constraint-based 
models of parsing (e.g. MacDonald et al 1994). The structured access claim I pursue 
here rejects this as an inherent feature of the parser: if there exists structural access 
mechanisms deployed by the parser, this claim cannot be globally true.  
As for questions of domain-specificity, one might be tempted to view the 
central claim of this thesis as a claim about domain-specific processes that apply only 
to linguistic processing. This is not necessarily the case, however; it may be that 
structured access mechanisms follow from deeper principles of cognition. In Chapter 
5, I will suggest one such possibility. Abstraction in access mechanisms is one way 
that the parser might respond to functional pressure to minimize memory interference 
in a noisy architecture. If this line of argumentation is correct, then the claim of 
 
 
19 
structured access might be true even if there is no architectural commitment to 
modularity or domain-specificity. Instead, the parser?s optimal strategy may be to pay 
attention to the most distinctive, minimal set of information to ease online processing, 
which would result in structured access for the dependencies considered here.  
 
Outline of the dissertation  
  
The thesis has two main parts. Chapter 2 introduces the argument from 
interference, one of the most well-practiced arguments for cue-based parsing models 
and content-addressable memory architectures. I investigate the empirical 
interference profile of two long-distance dependencies in English: subject-verb 
agreement and antecedent-reflexive dependencies in English. An off-line judgment 
task shows that in offline measures, agreement and reflexive feature mismatches are 
similarly anomalous for English speakers. However across three eye-tracking 
experiments, a minimal comparison of English agreement and reflexive dependencies 
reveals qualitatively different profiles with respect to online interference patterns. 
Agreement dependencies are reliably susceptible to interference online, showing a 
widely observed interference profile that suggests incorrect access to structurally 
inaccessible features (Clifton et al 1999; Pearlmutter et al 1999; Wagers et al 2009). 
In clear contrast, reflexive dependencies do not show any reliable interference from 
structurally inaccessible antecedents. This indicates that the processes used to resolve 
these two grammatical dependencies are distinct, opening up the question of how it is 
that reflexive dependencies are immune to interference online. I argue that reflexives 
 
 
20 
engage structured access to recover their antecedent. In other words, the difference 
between agreement and reflexives obtains because the reflexive dependency is 
initially constructed only with reference to the structural relation between the anaphor 
and its antecedent; the morphological features are inert in the construction of the 
antecedent-anaphor chain, in line with much typological and theoretical work on the 
representation of the binding dependency (i.e. Lidz & Idsardi 1999; B?ring 2005; 
Hornstein 2007; Kratzer 2009). The crucial conclusion from these experiments is that 
the syntactic dependencies licensed by reflexives are grammatically accurate, in the 
sense that the earliest stages of dependency formation appear to only access positions 
occupied by the local subject.  
Chapter 3 takes up a discussion of the implications of the results in Chapter 2 
for models of parsing. I critically assess the wider range of evidence from interference 
in sentence processing, and clarify the predictions of the models under consideration 
using a simple mathematical model of memory access. Using this formalization to 
model the agreement and reflexive dependencies in Chapter 2, I provide simulation 
evidence that further supports the claim that structured access, rather than passive 
memory dynamics, is the source of the empirical difference between agreement and 
reflexives observed in Chapter 2. 
In addition to supporting the hypothesis of structured access in parsing, 
Chapter 3 provides an analysis of the diverse range of experimental results that have 
been labeled as ?interference effects?. It is important to critically assess the range of 
these results, as they have been very influential in constructing models of the parser?s 
memory architecture. I argue that the range of constructions where the parser is 
 
 
21 
demonstrably ?grammatically fallible? in the early stages of processing (Wagers 2008; 
Phillips et al 2009) is narrower than generally assumed. The behavioral effects that 
risk being conflated under a single notion of an ?interference effect? actually 
correspond to two distinct underlying phenomena that are usefully dissociated in 
interpreting experimental data that investigates interference patterns in sentence 
processing. Once this distinction is made, it can be shown that for many phenomena, 
conclusions about grammatical infidelity rest on assumptions that may not be 
generally tenable. This critical assessment provides important support to the main 
claim of structured access by directly addressing potential counterexamples.  
 Chapter 4 builds upon the claim for structured access that was developed in 
Chapters 2 and 3 by examining another case of anaphor-antecedent dependency 
building, the case of long-distance reflexives in Mandarin Chinese. I provide an 
alternative argument for structured access by investigating the time course of 
reactivating antecedent noun phrases in completing local and long-distance binding 
relations. Time-course evidence in the form of speed-accuracy tradeoff (SAT) 
functions will suggest a role for structured access mechanisms in the resolution of 
long-distance reflexive dependencies, further supporting the conclusion that 
morphological and semantic features are not used to gate memory access for reflexive 
dependencies. The SAT evidence is supported by electrophysiological evidence about 
the processing of Mandarin long-distance reflexives. The ERP evidence suggests that 
familiar syntactic reanalysis processes do not drive the effect we observe in SAT, but 
that instead the difficulty observed in long-distance binding configurations shows a 
functionally similar profile to that observed in processing long-distance movement 
 
 
22 
dependencies and other memory-intensive tasks in sentence processing. Taken 
together, the investigation of the processing of ziji in Chapter 4 provides an 
alternative argument in support of structured access. Chapters 2 and 3 made the case 
that feature-matching but structurally inaccessible antecedents do not impact early 
memory access for reflexives. Chapter 4 demonstrates the converse: feature-
 mismatching structurally accessible antecedents are considered in the construction of 
a long-distance reflexive dependency, even in the presence of other feature-matched, 
accessible antecedents. This satisfies a clear prediction of structured access 
mechanisms: in both English and Chinese, structural relation to the anaphor, rather 
than feature content of potential antecedents, is the main determinant of online access. 
Chapter 5 synthesizes the results presented in Chapters 3 and 4 with extant 
results in the literature, articulating and defending a revised theory of the relation 
between grammatical dependencies and memory access. Results from both English 
and Chinese point to the conclusion that reflexive dependencies access memory using 
purely structural cues. However, there remains the question of the generality of 
structured access in comprehension, and in Chapter 5 I begin to address this. To this 
end, I present data from two self-paced reading (SPR) tasks to provide a comparison 
of the processing of the syntactic anaphor ziji with the processing of the intensified 
pronominal ta-ziji. The SPR data provide another piece of converging evidence for 
the claim that ziji initially accesses only commanding antecedent positions. In 
contrast, ta-ziji appears to access all possible antecedent positions in parallel, 
suggesting that structured access is not pursued for ta-ziji. The distinction in behavior 
between ziji and ta-ziji provides confirmation of the result of the computational 
 
 
23 
modeling in Chapter 3, which shows that the apparent structural sensitivity of the 
reflexive anaphors presented in this thesis is not an artifact of linear position. More 
generally, they serve to sharpen the hypothesis of structured access in parsing: the 
contrast between ziji and ta-ziji demonstrates that it is not simply interpretive content 
that drives the use of a structured access mechanism. Not all pronominal 
dependencies access their antecedents in a structured fashion. I suggest that the parser 
deploys a structured access mechanism for all long-distance structural dependencies 
that are crucial to interpretation. I furthermore argue that structured access is a 
reflection of functional pressure to minimize memory interference: in order to guard 
against memory interference and misinterpretation for structurally constrained long-
 distance dependencies, more minimal and abstract retrieval procedures are to be 
preferred. In light of this claim, structured access can be seen as the optimal strategy 
for recovering the correct interpretation from a long-distance dependency. 
 Chapter 6 concludes. 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
24 
Chapter 2: The argument from interference: English agreement and reflexives 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 In this chapter I present evidence that syntactically inaccessible, but feature-
 matched linguistic material is not considered while processing reflexive 
dependencies. This evidence is built on a direct contrast between reflexive 
antecedent-anaphor agreement and subject-verb agreement in English, which shows 
reliable interference from feature-matched, non-subject noun phrases in 
comprehension (Pearlmutter et al 1999; Wagers et al 2009). This contrast is of 
theoretical importance for investigating the manner in which linguistic memory is 
accessed, because subject-verb agreement and reflexive-antecedent dependencies are 
superficially very similar: to a first approximation, they both require feature concord 
with the local subject. If feature-based access is the single option for building 
agreement dependencies with the subject, the two dependencies should behave 
similarly with respect to interference effects. If reflexives engage a structured access 
mechanism, however, they should show qualitatively different patterns, with 
reflexives being insensitive to the feature content of structurally inaccessible 
antecedents at the point of memory access.  
 
 
25 
The result of this comparison is that reflexives are found to be immune from 
interference in environments where agreement reliably demonstrates interference 
effects. The agreement interference is predicted on a straightforward feature-based 
access model, but the reflexive data are not. This indicates that the argument from 
interference?one of two primary arguments for feature-based memory access 
mechanisms in a content-addressable memory architecture?does not 
straightforwardly extend to reflexive dependencies. As reflexives do not appear to 
consider structurally inaccessible antecedents, this provides evidence that structural 
information, rather than morphological or semantic feature information, gates 
memory access during construction of these antecedent-anaphor dependencies. This 
provides the first of the two main empirical arguments I provide for structured access 
in comprehension: we do not see interference as widely as expected if feature content 
always controls access to stored information.  
I first briefly outline the argument from interference that has been made in 
support of content-addressable memory architectures, and then review some of the 
experimental evidence that has been used to make this argument for subject-verb 
agreement. Note that there is a great diversity of results that have been attributed to 
memory interference, and I reserve a fuller discussion of this evidence until Chapter 
3. In addition to a brief survey of relevant subject-verb agreement findings, I present a 
brief review of existing evidence on the processing of reflexives. Although existing 
literature appears to suggest that reflexives pattern very differently from agreement 
with respect to interference effects, this conclusion is hampered by the fact that past 
studies have varied widely with respect to their experimental materials and methods.  
 
 
26 
I present three eye-tracking experiments and one offline judgment task that 
directly compare the interference profiles of agreement and reflexives. The results of 
this comparison establish that, unlike subject-verb agreement, English reflexive 
dependencies reliably resist intrusion from feature-matched, but structurally 
inaccessible antecedents. This comparison supports the hypothesis of structured 
access in comprehension, in that syntactic position, rather than morphological feature 
content, is what gates memory access in the construction of reflexive dependencies. 
This is a crucial first step in establishing that feature-based access is not a general 
feature of linguistic comprehension: it appears that in some cases, a structured access 
mechanism in engaged that privileges syntactic position over feature match when 
accessing memory. In the discussion I take up a more in-depth consideration of the 
relation of the current results to prior findings. 
 
The argument from interference  
 
One of the most well practiced arguments for content-addressable architectures in 
sentence processing is the argument from interference. Suppose that the parser needs 
to retrieve a certain constituent for some processing (the target). Given this task, the 
argument from interference has the following general form: if feature-matching, non-
 target memories disrupt processing or memory access, then the parser is employing 
an access procedure that uses this feature content to guide retrieval, rather than 
location in the memory store. Consider Figure 2.1. Given a hypothetical memory 
retrieval for a reflexive, the local subject memory is the target: on the assumption that 
 
 
27 
the string is grammatical, the desired antecedent will be found in this position. The 
embedded noun phrase John is a distractor memory: it may have the correct 
morphological feature makeup, but it is not the target of the reflexive?s antecedent 
retrieval. The target memory is structurally accessible given the retrieval: it occupies 
the correct syntactic position. For English reflexives, this means that it occupies a c-
 commanding place in the syntactic hierarchy (Reinhart 1976; Chomsky 1981). In 
contrast, the distractor NP is structurally inaccessible. No matter how well feature-
 matched the distractor NP is, it is not an acceptable antecedent by virtue of its 
structural position. 
 
Figure 2.1: Target and distractor NPs when processing a reflexive pronoun in 
English. 
 
S
 NP VP
 The man RC V NP
 who S hurt himself
 NP VP
 t V NP
 saw John
 Target
 Distractor
 
 
28 
The argument from interference follows directly from the definition of a 
content-addressable access mechanism: the desired feature content (the cue) is 
directly matched against the contents of all stored elements in short-term working 
memory. Stored memory representations resonate with the search cues when the cues 
are identical to the content in the memory image, in the same way that a tuning fork 
spontaneously begins to resonate in the presence of a pitch that matches its own 
characteristic pitch. By hypothesis, this process of resonating with the search cue 
renders a given memory available for further processing. When a memory is made 
accessible for further processing, it is said to have been retrieved. 
In a content-addressable memory system, there is no formal requirement that 
memories be entirely distinct in terms of content, despite the fact that this content is 
used as an index for later reaccess. The prediction of interference derives from this 
fact. Without a unique manner of indexing any given memory, interference effects are 
bound to occur (Kohonen 1980). This stands in contrast to a register-based 
architecture, where each memory may be assigned a unique storage location. This 
feature of content-addressable architectures means that there may arise situations 
where the search cues do not uniquely identify the target memory. There are two 
relevant situations in which this might occur, and it is useful to distinguish them. One 
such situation is where many elements in memory contain some or all of the search 
cues, leading to a situation where multiple matching candidates resonate with the 
search cues. In this multiple match situation, the memories that resonate may be 
understood to compete with each other for selection. The alternative situation is when 
the best-matching memory only contains a subset of the desired cues, a partial match 
 
 
29 
situation. These are diagrammed in Figure 2.2. Multiple match is observed in the 
grammatical sentence in the left panel, because multiple NPs bear singular and 
masculine features. Partial match is observed in the ungrammatical counterpart of this 
sentence; the memory that is retrieved will only carry a subset of the features used to 
access memory, but the potential candidates are nonetheless distinct in feature 
content. 
 
Figure 2.2: Multiple match and partial match interference configurations. For 
multiple match interference, the target memory is a perfect match to the search cues, 
but the distractor overlaps in some feature content with the target. In the partial match 
situation, neither target nor distractor is a perfect match to the search cues. 
 
Neither situation occurs in a register-based memory. In these architectures, 
unique addressing would eliminate the possibility of multiple matches being returned 
in response to a query, and failure to find information in a specified address would 
result in a retrieval failure, rather than a partially matched memory being retrieved. 
Both the multiple match and the partial match situations are a sort of retrieval 
S
 NP VP
 The man RC V NP
 who S hurt himself
 NP VP
 t V NP
 saw John
 +c-command
 +NP
 +masculine
 +singular
 +c-command
 +NP
 +masculine
 +singular
 S
 NP VP
 The woman RC V NP
 who S hurt himself
 NP VP
 t V NP
 saw John
 +c-command
 +NP
 +masculine
 +singular
 +c-command
 +NP
 +masculine
 +singular
 Multiple match interference Partial match interference
 
 
30 
interference, a prediction of content-addressable memory architectures. The 
distinction between these two types of retrieval interference is important, but I delay 
an in-depth characterization of the behavioral predictions of each until Chapter 3, 
where I argue that conflation of this distinction can led to incorrect generalizations 
about what interference effects can tell us with respect to underlying mechanism. 
The present studies focus exclusively on partial match interference, situations 
in which there is incorrect access of memories that only contain a subset of the 
necessary retrieval cues. This is because of the two cases of interference, only partial 
match interference unambiguously signals that the parser has retrieved structurally 
inaccessible material, and because the behavioral signature of partial match 
interference is well understood in the context of subject-verb agreement. In contrast, 
multiple match interference may be attributed to a number of different underlying 
mechanisms, as I detail in Chapter 3. For these reasons, multiple match interference 
thus does not provide the strongest evidence for incorrect access during parsing. 
Partial match interference, on the other hand, does not appear to be amenable to such 
alternative explanations, and so it provides the most stringent empirical test of 
incorrect access during sentence comprehension. Note that partial matching effects 
have also been referred to as illusions of grammaticality (Phillips et al 2010), but I 
avoid this term in order to not prejudge the underlying source of the effect. Although 
this is a very natural interpretation of these effects, partial-matching interference 
effects need not be linked to processes of ungrammaticality detection in any direct 
way; this is in fact true of the model to be presented in Chapter 3, as well as the ACT-
 R model of Vasishth et al (2008). 
 
 
31 
 
Partial-match interference in subject-verb agreement 
 
To date, two long-distance dependencies have been shown to be susceptible to 
partial-match interference in comprehension: negative polarity items (NPIs) and 
subject-verb agreement. The source of the NPI interference effect is a matter of 
debate (contrast claims in Vasishth et al 2008 and Xiang et al 2009), and I will put off 
discussion of these effects for the moment. Instead I focus on subject-verb agreement, 
where the phenomenon of interest is the well-known agreement attraction effect that 
has been noted by a wide range of researchers (e.g., Kimball & Aissen 1971; Kayne 
1989; Bock & Miller 1991; den Dikken 2001). Agreement attraction occurs when the 
morphological features of a noun other than the local subject appear to control verbal 
agreement. For example, Bock and Miller (1991) presented subjects with sentence 
fragments as in (2.1): 
 
(2.1)  a. The key to the cabinet... 
  b. The key to the cabinets... 
 
Bock and Miller observed that when the inaccessible noun cabinet was plural, 
there was a marked increase in the probability that participants would produce a 
plural verb form (i.e. were), compared to when cabinet did not bear plural features. 
The agreement attraction effect has been shown to obtain in a range of environments 
and across a range of different types of interfering noun (see Eberhard, Cutting & 
Bock 2005 for a review), and appears to be sensitive to the hierarchical, rather than 
 
 
32 
linear distance between the interfering noun position and the target verb (Franck, 
Vigliocco & Nicol 2002). 
 As with NPI dependencies, there remains some debate on the nature of 
subject-verb agreement interference effect, although much of the debate centers on 
the source of agreement attraction effects in production (Bock & Miller 1991; 
Vigliocco & Nicol 1998; Franck et al, 2002; Eberhard et al 2005; Staub 2010; 
Gillespie & Pearlmutter 2011). There is a sizeable body of research that has 
documented agreement attraction in production, including cross-linguistic 
documentation of similar effects in Spanish (Anton-Mendez, Nicol & Garrett 2002), 
German (Hartsuiker, Schrifers, Bock & Kikstra 2003), Dutch (Hartsuiker et al 2003), 
French (Franck et al 2002), Slovenian (Badecker & Kuminiak 2007), and Italian 
(Vigliocco & Franck 2001), among others. Importantly for current purposes, similar 
effects are also readily observed in comprehension measures (Clifton et al 1999; 
Pearlmutter et al 1999; H?ussler & Bader 2009; Wagers et al 2009).  
A number of researchers have argued that the subject-verb agreement 
attraction effects reflect interference effects in a content addressable architecture, in 
both production and comprehension (Badecker & Kuminiak 2007; Wagers 2008; 
H?ussler & Bader 2009; Wagers et al 2009). If true, an important prediction of 
feature-based access in a content-addressable architecture is borne out. In 
comprehension, agreement attraction effects clearly constitute an argument from 
interference: structurally inaccessible, but feature-matching material is seen to exert a 
disruptive influence in the computation of subject-verb agreement. Furthermore, in 
comprehension this is a clear example of partial-match interference, occurring when 
 
 
33 
the target noun (the local subject) does not match in the desired features. In these 
situations this interference arises when the distractor memory (i.e. cabinets in (2.1)) is 
retrieved as a function of its morphological features. If a structured access mechanism 
that privileges structural information over feature content during memory access is 
employed to resolve agreement online, we should not observe these effects. This is 
because from the point of view of structured access, the subject (the key) is uniquely 
identifiable if only structural cues are used to access memory. 
 Across a number of studies, clear generalizations about the comprehension 
profile of this interference effect in agreement have emerged. For example, in one 
investigation of the interference profile of English subject-verb agreement in 
comprehension, Pearlmutter, Garnsey, and Bock (1999) employed both self-paced 
reading and eye-tracking to ask about the processing of grammatical and 
ungrammatical agreement dependencies. They examined sentences of the form in 
(2.2): 
 
 (2.2) a. The key to the cabinet was rusty from years of disuse. 
  b. The key to the cabinets was rusty from years of disuse. 
  c. *The key to the cabinet were rusty from years of disuse. 
  d. *The key to the cabinets were rusty from years of disuse. 
 
In Experiment 1, they investigated these sentences using self-paced reading, and in 
Experiment 2, they employed eye-tracking methodology. In both replications, the 
findings were consistent. In the grammatical conditions, (2.2b) was read more slowly 
than (2.2a) from the interfering noun onwards in eye-tracking, and from the critical 
verb region onwards in self-paced reading. In the ungrammatical conditions, (2.2c) 
 
 
34 
was read more slowly than (2.2d), an effect that had a similar time course to the 
slowdown observed in the grammatical conditions. In eye-tracking measures, it was 
second-pass measures (i.e. re-reading times) that showed the effect. In Experiment 3, 
they only investigated the processing of the two grammatical variants of (2.2) above, 
in addition to variants with a plural head noun. A summary of the observed effects 
across the three experiments is presented in Figure 2.3. This summary plot shows the 
magnitude of the ?interference effect?, which is obtained by subtracting the mean 
reading time for conditions that had singular intervening nouns (cabinet) from those 
that had a plural intervening noun (cabinets), for both grammatical and 
ungrammatical conditions.  
 
 
 
 
 
35 
 
Figure 2.3: Interference effects in Experiments 1-3 in Pearlmutter et al (1999). The 
interference effect is the difference in RTs at the critical region that is due to 
manipulating the feature content of the distractor, as shown. 
 
It can be seen from Figure 2.3 that across methodologies and replications the 
effect of the interfering noun position on the computation of subject-verb agreement 
was stable. The effect of a feature matched noun phrase in both ungrammatical and 
grammatical conditions was facilitatory. For the ungrammatical sentences, one way of 
understanding this effect is as an illusion of ungrammaticality (Phillips et al to 
appear), wherein the feature match with the inaccessible noun phrase leads to 
spurious acceptability, and hence eased processing, on some portion of trials in the 
experiment. This is reflected in the negative interference effects shown in Figure 2.3. 
Exp 1 Exp 2 Exp 3
 Summary of interference effects in Pearlmutter et al 1999
 Interference effect (ms
 )
 -15
 0
 -10
 0
 -5
 0
 0
 5
 0
 10
 0
 15
 0
 Grammatical
 Ungrammatical
 [The key to cabinets were...] - [The key to the cabinet were]
 [The key to cabinets was...] - [The key to the cabinet was]
 
 
36 
Pearlmutter and colleagues also observed a similar processing advantage when the 
interfering noun was singular in grammatical conditions: when the interfering noun 
was singular, matching the singular verb form, faster reading times (eased processing) 
were observed. This effect is not obviously attributable to an illusion of 
grammaticality, as the two conditions are equally grammatical, but the generalization 
in both grammatical and ungrammatical cases is straightforward. When the agreeing 
verb?s features match the inaccessible noun?s features, processing is facilitated, 
though this facilitation is much greater in the case of unacceptable sentences. 
 Wagers, Lau and Phillips (2009) presented another set of studies that 
investigated interference in agreement computation in comprehension. Like 
Pearlmutter et al (1999), they consistently found partial-match interference effects for 
agreement comprehension. They also gave evidence that interference occurred even 
when the interfering noun position was not in a linearly or hierarchically intervening 
position, as in (2.3a-d). However, unlike Pearlmutter et al (1999), they found that this 
effect was limited to ungrammatical sentences, for both PP and non-intervening 
relative clause environments. This finding was replicated across 6 experiments, 
wherein Wagers et al (2009) consistently found that in grammatical sentences (2.3a-b, 
e-f), the interfering noun?s number marking did not exert any measurable influence 
on the relevant behavioral measure (reading times in self-paced reading, and accuracy 
measures in a speeded judgment task). Instead they found that across six experiments, 
and in two very different structural configurations, ungrammatical conditions (2.3c-d, 
g-h) consistently show an effect of partial-matching interference?that is, greater 
apparent acceptability and eased measures of processing. In Experiments 2-3, they 
 
 
37 
showed that this was the case for non-intervening relative clause environments (2.3a-
 d), and in experiments 4-7 they showed this in the same prepositional modifier 
environments (2.3e-h) that Pearlmutter et al (1999) investigated. Figure 2.4 provides 
an overview of their findings across all of the experiments. 
 
 (2.3) a. The musician who the reviewer praises so highly... 
  b. The musicians who the reviewer praises so highly... 
  c. *The musician who the reviewer praise so highly... 
  d. *The musicians who the reviewer praise so highly... 
  e. The key to the cell unsurprisingly was rusty? 
  f. The key to the cells unsurprisingly was rusty? 
  g. *The key to the cell unsurprisingly were rusty? 
  h. *The key to the cells unsurprisingly were rusty? 
 
 
Figure 2.4: Interference effects in Wagers et al (2009). 
 
Exp 2 Exp 3 Exp 4 Exp 5 Exp 6
 Summary of interference effects in Wagers et al 2009
 In
 te
 rfe
 re
 nc
 e 
ef
 fe
 ct
  (ms
 )
 -6
 0
 -4
 0
 -2
 0
 0
 20
 40
 60
 Grammatical
 Ungrammatical
 
 
38 
Wagers and colleagues did not find any evidence that the distractor noun 
impacted the processing of grammatical subject-verb agreement, and their 
Experiment 6 showed that even with a relatively high-powered study, the effect was 
not reliable. Instead, they suggested that the effects observed by Pearlmutter et al 
(1999) for grammatical sentences were due to an effect of plural complexity that 
arose at the previous region, which they addressed by including adverbs that 
intervened between the attracting noun and the critical verb, as well as performing 
mixed-effects regression analyses that factored out the effect of plural complexity. 
However, despite the fact that the effect in grammatical sentences was minimal and 
unreliable, there was consistency in the findings for ungrammatical sentences: 
inaccessible feature-matched nouns resulted in facilitated processing. The partial 
feature match led to a consistent interference effect. 
Wagers and colleagues argued that cue-based retrieval interference was 
responsible for attraction in the comprehension experiments. In other words, the 
partial match that occurred in ungrammatical sentences led to incorrect retrievals of 
the distractor NP, indicated by eased processing profiles when there was an 
inaccessible feature match. The most compelling piece of evidence to this end, 
according to these authors, is the grammatical asymmetry prediction of the retrieval 
account: they did not observe an illusion of ungrammaticality, which is what would 
be expected if singular subjects with embedded plural nouns caused faulty encoding 
of the target subject?s number feature (as in other models of the agreement attraction 
effect, Eberhard et al 2005). Though their account posited a mixture of forward- and 
backward-looking processes to derive this prediction, this grammatical asymmetry 
 
 
39 
prediction holds even if one assumes a purely retrospective agreement computation 
process (a point noted in Wagers 2008). The crucial difference controlling these 
predictions is the fact that in the grammatical sentences, there is multiple-match 
interference: in the grammatical interfering condition, there are multiple singular 
nouns, but the target head noun nonetheless is a perfect fit to the search cues. In 
contrast, for the ungrammatical sentences, there is a situation of partial-match 
interference: no noun fully licenses the agreement morphology on the verb, and so 
comprehension proceeds by considering either a structurally accessible feature 
mismatch, or a structurally inaccessible feature match.  
 Because of the clear and reliable behavioral patterns associated with partial-
 match interference in subject-verb agreement in English, this dependency provides an 
important benchmark in assessing the impact of interference effects in the parser 
more generally. Across production, judgment, and comprehension tasks, there is 
ample evidence that the agreement attraction effects in comprehension arise from an 
erroneous consideration of an inaccessible noun?s morphological features. Across 
studies we observe that in grammatical sentences, no reliable effects are observed, 
and in ungrammatical sentences, eased processing is observed when there is a partial 
feature match in an inaccessible position. This interference pattern shown in subject-
 verb agreement is a clear and reliable behavioral signature of grammatically fallible, 
feature-based access in comprehension. Furthermore, these predictions follow from 
the adoption of a content-addressable model of memory access (as I demonstrate in 
Chapter 3). An important prediction of these models is that wherever feature content 
is used to access memory, agreement-style interference should be observed. Thus, in 
 
 
40 
order to evaluate the role of feature-based over structured access in comprehension, it 
is necessary to determine how widely observed these sorts of partial-matching 
interference effects are.  
 
Lack of interference in reflexive dependencies 
 
 Reflexive dependencies provide a natural point of comparison with subject-verb 
agreement in English. This is because at a very superficial level, they are subject to 
similar constraints. In English, they both require a feature-matched local subject in 
most cases. If the retrieval mechanism pools all the relevant linguistic constraints for 
a dependency to access memory, then for both dependencies the same mixture of 
morphological and structural features will be used to access memory. If this is so, 
then both agreement and reflexive dependencies should show similar interference 
profiles. However, in studies that have examined reflexive processing to date, there 
has been no reliable evidence for partial-matching interference as seen in agreement 
dependencies. A number of authors have argued from a wide variety of experimental 
results that syntactic constraints provide a hard constraint on the nouns that are 
considered for participation in a reflexive dependency (Nicol 1988; Nicol & Swinney 
1989; Clifton et al 1999; Sturt 2003a,b; Xiang et al 2009). If the conclusion of these 
authors is correct, then there is a compelling case to be made for structured access in 
comprehension: reflexives access memory using primarily structural information.  
 Sturt (2003a) used eye-tracking to determine whether or not a morphological 
feature match with structurally inaccessible antecedents exerted an influence on the 
 
 
41 
construction of the binding dependency. He examined small discourses of the form in 
(2.4) (Experiment 1) and (2.5) (Experiment 2). Note that unlike the agreement studies 
reviewed above, Sturt did not manipulate the feature content on the agreeing reflexive, 
instead manipulating the gender of the two noun positions in the sentence. In addition, 
rather than using number features, he manipulated the stereotypical gender of the 
noun phrase in the accessible position, and the actual gender of the noun phrase in the 
interfering position. Thus, none of the sentences in his experiments were globally 
ungrammatical, although because gender-biased nouns cause comprehenders to 
commit to the gender of a noun phrase, a temporary ungrammaticality arises in when 
the local subject?s stereotypical gender mismatches the reflexive?s morphological 
features. Previous ERP work demonstrates that this temporary percept of 
ungrammaticality due to violation of gender stereotype is reliable (Osterhout, Bersick, 
& Laughlin 1997). I label cases where the local subject does not agree with the 
reflexive in perceived gender as incongruent conditions and mark them with #. 
 
 (2.4) Jonathan was pretty worried at the City Hospital. 
  a. He remembered that the surgeon pricked himself with a needle. 
  b. #He remembered that the surgeon picked herself with a needle. 
  Jennifer was pretty worried at the City Hospital. 
  c. She remembered that the surgeon pricked himself with a needle. 
  d. #She remembered that the surgeon pricked herself with a needle. 
 
 (2.5) Jonathan was pretty worried at the City Hospital. 
  a. The surgeon who treated Jonathan had pricked himself with a needle. 
  b. #The surgeon who treated Jonathan had pricked herself with a needle. 
  Jennifer was pretty worried at the City hospital. 
  c. The surgeon who treated Jennifer had pricked himself with a needle. 
  d. #The surgeon who treated Jennifer had pricked herself with a needle. 
 
 In order to facilitate comparison with the agreement studies presented in the 
 
 
42 
previous section, I present the key data from Sturt?s Experiments 1 and 2 in a form 
that mirrors the design of the Pearlmutter et al (1999) and Wagers et al (2009) studies. 
In Figures 2.4 and 2.5, the interference effects for congruent and incongruent 
conditions in Sturt (2003a) are shown. The interpretation of the direction of the 
effects is identical to that in the previous section. It can be seen that the reliable 
behavioral pattern that signaled interference in subject-verb agreement 
comprehension is not found in either of these studies. 
 
 
Figure 2.5: Interference effects for early (first-pass) measures in Sturt (2003). 
Exp 1 Exp 2
 Summary of first-pass interference effects in Sturt 2003
 In
 te
 rfe
 re
 nc
 e 
ef
 fe
 ct
  (ms
 )
 -6
 0
 -4
 0
 -2
 0
 0
 20
 40
 60
 Congruent
 Incongruent
 
 
43 
 
Figure 2.6: Interference effects for late (total time) measures in Sturt (2003). 
 
 
 In neither experiment did Sturt find consistent evidence for a partial-match 
interference effect as seen in the agreement studies above; in fact, there were no 
reliable interference effects at all across measures. In Experiment 1, there were no 
reliable effects in first-pass measures, however an effect was observed in re-reading 
times in Experiment 1. For congruent conditions only, a feature-matched inaccessible 
antecedent caused faster reading times, and this difference reached statistical 
significance in pairwise and omnibus ANOVA analyses. This effect is consistent with 
partial-match interference, but this effect did not replicate in Experiment 2, and in 
fact, the direction of the numerical trend was reversed.  
 Sturt concluded that the syntactic binding constraints act as a hard constraint on 
Exp 1 Exp 2
 Summary of interference effects in Sturt 2003
 In
 te
 rfe
 re
 nc
 e 
ef
 fe
 ct
  (ms
 )
 -6
 0
 -4
 0
 -2
 0
 0
 20
 40
 60
 Congruent
 Incongruent
 
 
44 
the early stages of constructing the binding dependency. To the extent that any effect 
of the inaccessible antecedent is observed in his experiments, it is limited in its scope 
and is not consistently obtained. Although the primary picture suggested by Sturt?s 
study is one of structural fidelity, it is important to note that offline measures that 
Sturt examined in a follow-up to Experiment 1 showed that participants clearly do 
make more errors in conditions that contain a feature match in the inaccessible 
antecedent position. By asking a question such as ?Who was pricked with a needle?? 
after as the sentences used in Experiment 1, and prompting participants with both 
characters in the story, Sturt found that the number of ungrammatical interpretations 
sharply rose in response to a feature-matched, inaccessible antecedent, for both 
unmatched and matched accessible antecedent sentences.   
 The finding that reflexives resist interference was subsequently replicated by 
Xiang, Dillon and Phillips (2009) using event-related potentials (ERPs). They 
employed materials that were based on Sturt (2003a), and contrasted the three 
conditions in (3.8): 
 
(2.6) a. The tough soldier that Fred treated in the military hospital 
introduced   himself to all the nurses.  
 b. #The tough soldier that Katie treated in the military hospital 
introduced herself to all the nurses. 
c. #The tough soldier that Fred treated in the military hospital 
introduced herself to all the nurses. 
 
 They contrasted these materials with a parallel set of conditions that examined 
the impact of inaccessible, feature-matched material on the processing of NPI 
dependencies, which are known to be susceptible to partial-matching interference 
 
 
45 
effects: 
 
 (2.7)      a. No restaurants that the newspapers have recommended in their 
dining         reviews have ever...  
 b. *The restaurants that no newspapers have recommended in their 
dining reviews have ever... 
c. *Most restaurants that the newspapers have recommended in their 
dining reviews have ever... 
 
 Their logic was as follows. If partial-matching interference obtains in either 
dependency, then the condition that has feature-matched, inaccessible material (either 
a gender-matched NP or an embedded negative quantifier) should show reduced 
processing difficulty on whatever ERP component indexes the difficulty associated 
with encountering an ungrammatical dependency. For both dependencies, 
grammaticality detection was reflected in the P600 component, a positive-going ERP 
component that occurs approximately 600ms post-stimulus, and has a primarily 
posterior distribution. This component is often observed in ungrammatical 
environments, though the exact relation of this response to ungrammaticality 
detection is unclear (Friederici, Pfeifer, & Hahne, 1993; Hagoort, Brown, & 
Groothusen, 1993). For NPI dependencies, the pattern in the P600 replicated earlier 
findings of partial-matching interferene in NPIs (Drenhaus, Saddy & Frisch 2005). 
Specifically, the P600 response to the interfering condition (3.9b) was smaller than 
that associated with the fully ungrammatical condition (3.9c), again showing partial-
 matching interference. In fact, the response to the NPI interfering condition was 
statistically indistinguishable from the waveform evoked by the fully grammatical 
condition (3.9a).  
 
 
46 
 An altogether different pattern was observed for reflexives. In a P600 response 
that began at approximately 450ms after the presentation of the critical reflexive, both 
gender-incongruent conditions (3.8b-c) showed an identical positive deflection 
relative to the baseline condition (3.8a). Pairwise analyses revealed no significant 
differences between the interfering condition and the entirely incongruent condition. 
In later time windows, however, there was a trend towards a greater P600 effect in 
(3.8b) relative to (3.8c).  This did not reach significance, though it is important to 
note that this is exactly the opposite pattern to that found in the NPI conditions. It 
does not reflect the eased processing that is expected in partial-match interference 
situations. This mirrors the non-significant trends in parallel conditions in Sturt 
(2003a), where the effect of an intrusive, feature-matched subject was a numerical 
slowdown in reading times.  
 There was no trend consistent with partial-matching interference for reflexives 
in Sturt (2003a) and Xiang et al (2009) is suggestive, and matches well with 
suggestions of structural fidelity for reflexives that have been observed in other 
experiments (Nicol 1988; Nicol & Swinney 1989; Clifton et al 1999). Of all of the 
reflexive studies that have been published to date, there is no convincing 
demonstration of anything resembling the clear partial-matching interference effects 
that have repeatedly been observed for subject-verb agreement dependencies. There is 
no reliable indication that reflexives use their morphological features in retrieving 
their antecedents, which appears to support the hypothesis of structured access in 
comprehension.  
  
 
 
47 
Experiment 1: Direct comparison of agreement and reflexives 
 
 The brief survey of agreement and reflexive studies given above suggests a 
contrast in the interference profiles between the two phenomena. As noted at the 
outset, whether or not this contrast is real is an important theoretical question when 
considering the hypothesis of structured access in comprehension. From the point of 
view of feature-based access mechanisms, subject-verb agreement and reflexive 
dependencies form a close minimal pair: they both require a local subject that agrees 
in morphological features. If they show differential sensitivity to feature-matching, 
but non c-commanding linguistic material, then there is a case to be made that the 
mechanisms used to access linguistic memory are qualitatively different for the two 
dependencies. In surveying previous results, I noted that partial-matching interference 
effects are reliably observed in agreement dependencies, and that this interference 
profile is the crucial behavioral signature of feature-based memory access. For 
reflexives, no result that is consistent with this interference profile has been observed 
to date, opening up the possibility that reflexive dependencies engage a different 
strategy for memory access. Since morphological features do not appear to gate 
memory access in behavioral measures, it is tempting to conclude that reflexives 
engage in structured access.  
  However, it is difficult to make a strong case for structured access with 
current studies, due to large differences in the environment where interference has 
been found for agreement and reflexives. All reflexive studies to date have 
investigated interference from nouns that are outside the local clause that contains the 
 
 
48 
reflexive. For example, in Sturt?s (2003a) Experiment 2 and the experiment reported 
in Xiang et al (2009), the inaccessible noun was embedded inside an object relative 
clause that modified the subject noun. In these environments, it is well known that the 
magnitude of the agreement attraction effect is lessened in production, perhaps 
because of the scope of planning in production (Bock & Miller 1991) or because of 
the extra phrase-structural distance between the interferer and the head noun (Franck 
et al 2002; Eberhard et al 2005). Furthermore, agreement studies in English 
necessarily look at interference from number features, whereas reflexive studies have 
investigated the effect of (stereotypical) gender features on the computation of the 
reflexive dependency. It is possible that the role of these two features differs in 
comprehension, driving the observed differences between reflexives and agreement. 
 In order to first establish the contrast in interference profiles, and confirm 
whether or not reflexives are constructed in a fundamentally structure-sensitive 
manner, Experiment 1 presents a within-subjects comparison of subject-verb 
agreement and reflexives in English using eye-tracking. For both dependencies, the 
question is whether or not inaccessible, feature-matched noun phrases exert an 
influence on the computation of the dependency. The crucial test is whether or not the 
partial-match interference profile commonly observed for agreement obtains for 
reflexives, which would unambiguously indicate a feature-based access strategy. 
Structured access predicts that the feature match with the inaccessible antecedent 
should not affect the early stages of processing the reflexive dependency. In order to 
facilitate the comparison, the syntactic position and feature content of the interfering 
noun was held constant across dependencies. The prediction is that if structured 
 
 
49 
access is used for reflexives, Experiment 1 should replicate the partial-match 
interference pattern observed in Pearlmutter et al (1999) and Wagers et al (2009) only 
for agreement conditions; no observable such interference pattern should be seen for 
reflexives.  
 
Participants 
 
 40 members of the University of Maryland community participated in 
Experiment 1 (24 females, mean age 21.9). Participants gave informed consent, and 
were either paid $10 for their participant or received course credit. The experimental 
session, including set-up and calibration, lasted approximately one hour. 
 
Stimuli 
 
 Forty-eight item sets of the form given in Tables 2.1 and 2.2 below were 
constructed. For both agreement and reflexive dependencies, the subject noun was the 
same. In all cases, the subject head noun (NP1) was modified by a subject relative 
clause that contained the intrusive noun (NP2). In order to ensure that the overt 
gender marking in singular reflexives did not provide extra cues to antecedent identity 
above and beyond number, both were chosen to have similar gender bias, based on 
the norms in Kennison & Trofe (2003). 24 of the item sets contained a pair of male-
 biased nouns, and the remaining 24 contained a pair of female-biased nouns. The verb 
inside the relative clause never overtly expressed agreement, and neither did the main 
 
 
50 
clause verb used in the reflexive conditions. For all conditions, the subject was 
followed by an adverbial that signaled the end of the relative clause. 
 
 
 
 
 
Agreement conditions for Experiment 1  
Grammatical, no interference 
1 The new executive/ who oversaw/ the middle manager/ apparently/ was dishonest / about the company?s profits. 
Grammatical, interference 
2 The new executive/ who oversaw/ the middle managers/ apparently/ was dishonest / about the company?s 
profits. 
Ungrammatical, no interference 
3 The new executive/ who oversaw/ the middle manager/ apparently/ were dishonest / about the company?s 
profits. 
Ungrammatical, interference 
4 The new executive/ who oversaw/ the middle managers/ apparently/ were dishonest / about the company?s 
profits. 
 
Table 2.1: Summary of agreement conditions in Experiment 1. Critical and spillover 
regions included in the analysis are underlined.   
 
 
Reflexive conditions for Experiment 1  
Grammatical, no interference 
1 The new executive/ who oversaw/ the middle manager/ apparently doubted/ himself on/ most major decisions. 
Grammatical, interference 
2 The new executive/ who oversaw/ the middle managers/ apparently doubted/ himself on/ most major decisions. 
Ungrammatical, no interference 
3 The new executive/ who oversaw/ the middle manager/ apparently doubted/ themselves on/ most major 
decisions. 
Ungrammatical, interference 
4 The new executive/ who oversaw/ the middle managers/ apparently doubted/ themselves on/ most major 
decisions. 
 
Table 2.2: Summary of reflexive conditions in Experiment 1. Critical and spillover 
regions included in the analysis are underlined.   
 
 For agreement conditions, the main verb was always a present tense agreeing 
form of be (was or were), followed by a predicative adjective and a four-word 
spillover region. For reflexives the main verb was always a non-agreeing, past tense 
 
 
51 
verb that was followed immediately by a direct object reflexive. When the reflexive 
was singular, it agreed in gender with the gender bias of the two nouns in the 
sentence: thus, 24 items contained himself, and the remaining 24 items contained 
herself. As in the agreement conditions, the reflexive was followed by a four-word 
spillover region.  
 The 48 target items were mixed with 152 fillers, for a total of 200 sentences. 
In addition to the 24 ungrammatical target items, there were 12 unrelated 
ungrammatical fillers (containing an illicit NPI dependency) for a grammatical-to-
 ungrammatical ratio of 4.6:1. Half of the target items and half of the fillers were 
followed by a comprehension question. Across items, comprehension questions 
addressed various parts of the sentence; this was done in order to prevent participants 
from adopting superficial reading strategies that extracted the information needed to 
answer comprehension questions without fully comprehending the sentence. The 
entire set of experimental stimuli for experiments 1-3 can be found at 
http://people.umass.edu/bwdillon. 
 
Offline judgments 
 
 One concern with investigating the effect of number mismatch on reflexive 
dependencies in English is the fact that plural reflexives for singular antecedents are 
tolerated in situations where the speaker does not wish to commit to a particular 
gender for the antecedent. Intuition suggests that sentences such as the student hurt 
themselves during lunch break are acceptable in a colloquial register, but this effect 
 
 
52 
appears to be subject to significant dialectal variation. However, for nouns that 
overtly signal the referent?s gender, this option is degraded, as in *the girl hurt 
themselves during lunch break. In order to test whether or not the number mismatch 
in the present materials was reliably rejected, an offline judgment study was 
conducted with the 48 items from Experiment 1, equally balanced for male and 
female biased nouns. These 48 target items were mixed with 100 fillers, and the 
materials were balanced so that across the experiment, half of the sentences were 
ungrammatical. The anomalies in the fillers comprised a variety of different 
grammatical errors, including unlicensed NPIs and unlicensed verbal morphology 
(e.g. *will eating). 12 participants were asked to judge the acceptability of the 
sentences they read on a 7-point scale, where 7 was completely acceptable and 1 was 
completely unacceptable. Participants were instructed to judge the sentences with 
regard to whether or not they were acceptable in colloquial speech. The results are 
presented in Table 2.3. 
 
 [+gram,-intr] [+gram,+intr] [-gram,-intr] [-gram,+intr] 
Agreement 5.36 (?0.29) 5.08 (?0.29) 3.00 (?0.41) 3.32 (?0.43) 
Reflexives 5.60 (?0.25) 5.68 (?0.22) 3.08 (?0.37) 3.24 (?0.38) 
 
Table 2.3: Mean judgments and standard error by subjects for Experiment 1 rating 
study. Values are on a 7-point scale where 7 is perfectly acceptable, and 1 is 
completely unacceptable. 
 
 A three-way repeated measures ANOVA by participants revealed a significant 
main effect of grammaticality (F(1,11) =  34.0, p < 0.001), as well as a significant 
interaction of grammaticality with interfering noun number (F(1,11) = 5.3, p < 0.05).  
Resolving this interaction revealed that there was a significant interaction of 
 
 
53 
interfering noun number and grammaticality only for agreement conditions (F(1,11) = 
10.4, p < 0.01); the interaction did not reach significance for reflexives (F(1,11) = 0.1, 
p > 0.7).  Resolving this interaction further using paired t-tests revealed a marginal 
effect of interfering number for grammatical agreement conditions (t(11) = 1.8, p < 
0.1), and no effect in ungrammatical conditions (t(11) = -1.5, p < 0.2). Thus there was 
no reliable evidence that the interfering number had an effect on offline judgments for 
either dependency. 
 Importantly, the effect of grammaticality was highly significant for reflexives 
(F(1,11) = 31.7, p < 0.001), and the size of the grammaticality effect differed across 
dependencies. In fact, there was a slight trend towards a larger penalty for 
ungrammaticality in reflexives than there was in agreement (agreement: ?? = 2.1 ? 
0.37, reflexives: ?? = 2.5 ? 0.44). 
 The results confirm that in offline judgments, participants treat the reflexive 
and agreement anomalies in a qualitatively similar fashion. Importantly for the 
present purposes, there was no indication that the plural reflexive themselves was 
accepted to any degree with the gender-biased singular nouns in the experimental 
materials. 
 
Procedure 
 
 The 48 target item sets were distributed into 8 Latin Square lists, and five 
participants were assigned to each list. Each list was randomized along with the filler 
 
 
54 
sentences subject to the constraint that no two experimental sentences were presented 
next to each other.  
 The maximum number of characters allowed on a single line on the visual 
display was 142 characters, and all sentences in the experiment fit on one line. All 
sentences were presented in a 12-point fixed-width font (Courier), and all characters 
were 9 x 16 pixels on the display. The resolution of the visual display was 1280 x 720 
pixels on an LCD screen.  
 Eye movements were recorded using an Eyelink 1000 tower-mount eye-
 tracker. Participants had binocular vision while movements were measured, but only 
the gaze of the right eye was tracked. The tower was 32 inches from the visual 
display, giving participants approximately 5 characters per degree of visual angle. 
The eye-tracker sampled eye movements at 1000Hz.  
 Prior to beginning the experiment, participants were familiarized with the 
apparatus and given four practice trials. While seated, participants? heads were 
immobilized using a chin rest and forehead restraint that was adjusted for comfort. 
Before the experiment, and whenever necessary throughout the experiment, the 
experimenter calibrated the eye-tracker with a 9-point display to ensure an accurate 
record of eye-movements across the screen. Participants began each experimental 
trial by fixating on a marker at the beginning of the sentence. Once the fixation in the 
target region was recognized by the experimental software, the trial sentence was 
displayed all at once. Participants ended the presentation of the trial sentence by 
indicating they had finished using a response pad. On trials with a question, the 
question was presented immediately after the trial sentence, and participants indicated 
 
 
55 
their response on the same response pad. Participants were allowed to take breaks at 
their discretion throughout the experiment. At a minimum, the experimenters asked 
the participants to take one short rest during the course of the experiment. After each 
break, participants were recalibrated to ensure accurate measurement of the eye 
movements.  
 
Data Analysis 
 
 Sentences in both reflexive and agreement conditions were divided into six 
regions of interest, as indicated in Tables 2.2 and 2.3. For all conditions, the complex 
subject was divided into three regions: the head noun, its determiner and adjective 
(NP1), the relative clause complementizer and the embedded verb, and the embedded 
noun along with its determiner and adjective (NP2). The remainder of each sentence 
was divided into a pre-critical region, a critical region, and a spillover region. For 
agreement conditions, the pre-critical region consisted of the main clause adverbial, 
the critical region consisted of the agreeing form of be and the predicative adjective, 
and the spill-over region consisted of the remaining four words. For reflexive 
conditions, the pre-critical region consisted of the main clause adverbial and the main 
clause verb, the critical region consisted of the reflexive and the following 
preposition, and the spill-over region was the remaining three words in the sentence. 
The larger analysis window for the agreement conditions was adopted because of a 
high rate of skipping of the inflected auxiliary. A larger window was adopted for the 
reflexive conditions to attempt to maximize the similarity of the critical reflexive 
 
 
56 
region to the critical agreement region. It should be noted, however, that similar 
patterns of results obtain when word-by-word regioning of the critical areas is used. 
Analysis for reflexive and agreement conditions proceeded separately. I present the 
data from four regions of interest in both agreement and reflexive conditions: NP1, 
NP2, the critical agreement/reflexive region, and the spillover region.  
 I report three measures for each of our regions of interest. Early measures 
reported here are the first pass reading time and the probability of regression. First 
pass reading time (FPT) is calculated by summing all fixations in a region of interest 
after participants first enter the region until the first saccade out of that region (either 
to the right or the left). The probability of regression (PR) corresponds to the 
probability that a regression is initiated from a particular region before exiting that 
region to the right. I also report a late measure, total time (TT), which is the total sum 
of all fixations in a particular region of interest, including first pass reading time and 
any time spent re-reading the region.  
 For all measures, statistical analysis was performed using mixed-effect linear 
regressions to assess the magnitude, direction, and reliability of the experimental 
factors on reading times and probability of regression. There are a number of 
advantages to this analysis. Mixed effects models have the advantage of allowing for 
simultaneous modeling of by-subject and by-item effects. Additionally, unlike 
repeated-measures ANOVAs, they generalize readily to data sets that have missing 
values (see Baayen, Davidson & Bates 2008). This is useful in the context of eye-
 tracking: for any given trial, there is a chance that participants will simply not fixate 
in a region of interest, leading to missing data. In adopting a mixed-effect models 
 
 
57 
analysis, missing values were left out, rather than adding zeros to trials where 
participants had no fixations for a particular measure.  
The experimental fixed effects in the models were the factors GRAM (whether 
or not the sentence was grammatical), INTR (whether or not the embedded NP was 
plural), and their interaction. The fixed effects were coded using simple difference 
sum coding (grammatical conditions were coded as -.5, ungrammatical as .5; no 
interference conditions as -.5, interference as .5). Thus, all reported coefficients 
reflect the magnitude of the difference between levels of a given factor in 
milliseconds. Additionally, a fixed effect for the order of a trial in the experiment was 
considered. In addition to these fixed effects, I additionally considered random 
intercepts for participants and items, as well as random slopes for the experimental 
fixed effects by subjects and by items. In all cases, the significance of non-
 experimental fixed effects and random effects was assessed, and I report in all cases 
the best-fit model (following Baayen et al 2008; Jaeger 2008). I leave the 
experimental fixed effect structure constant across all models, because these effects 
were theoretically motivated by the design of the study. For most analyses, in 
addition to the experimental fixed effects, the best model included only random 
intercepts for subjects and items, as well as a fixed effect for trial order. The raw 
mean fixation times are presented in Tables 2.4 and 2.5 below. For the probability of 
regression measure, logistic mixed effects models were used (see Jaeger 2008), as the 
dependent measure was categorical (i.e. presence or absence of a first-pass regression 
for a given region of interest). A fixed effect was considered significant if its absolute 
t-value was greater than 2, which indicates that a given effect?s 95% confidence 
 
 
58 
interval does not include 0 (Gelman & Hill 2005); thus all reported coefficients with a 
t-value whose absolute value is greater than 2 are significant at the p < 0.05 level. 
Although I do not report the significance of non-experimental fixed effects (i.e., 
order), in most instances it did have a statistically significant coefficient in the final 
model. 
  
Results: Agreement 
 
 The by-region reading times for first-pass and total time measures for the 
agreement conditions, as well as first-pass regression probabilities, are presented in 
Table 2.4. Prior to the critical region, no significant effects for any of the 
experimental factors were observed in any measure. 
 
Agreement NP1 NP2 Critical Spillover 
First Pass     
Gram, no int  582 (30) 627 (37) 373 (18) 795 (42) 
Gram, int  588 (32) 612 (34) 389 (17) 837 (48) 
Ungram, no int 583 (36) 593 (31) 445 (23) 805 (51) 
Ungram, int 594 (39) 639 (33) 439 (21) 847 (44) 
Total Time     
Gram, no int  867 (60) 955 (63) 579 (34) 1018 (46) 
Gram, int  835 (51) 959 (50) 622 (40) 1059 (52) 
Ungram, no int 891 (63) 971 (51) 811 (57) 1140 (70) 
Ungram, int 881 (55) 1002 (70) 693 (35) 1074 (51) 
Pr(Regression)     
Gram, no int  - .18 (.03)  .13 (.03) .65 (.04) 
Gram, int  - .20 (.03) .17 (.02) .67 (.04) 
Ungram, no int - .18 (.03) .25 (.04) .71 (.04) 
Ungram, int - .13 (.02) .20 (.03) .70 (.04) 
 
Table 2.4: Table of means (in ms where applicable) for agreement conditions for first 
pass, total time, and probability of regression. Standard error by participant is shown 
in parentheses.   
 
 
 
59 
 The model results at the critical agreeing verb region (was dishonest) are 
summarized in Table 2.6. In the early first-pass measures, there was a significant 
effect of grammaticality. This was due to a significant slowdown in ungrammatical 
conditions of approximately 64ms (SE = 14.7, t = 4.25). In the spillover region, there 
were no significant fixed effects in first-pass reading times. In the total time measure, 
there was a significant effect of grammaticality, due to a slowdown of approximately 
157ms (SE = 28.0, t = 5.60) for ungrammatical conditions relative to grammatical 
conditions. Additionally, there was an interaction of interference and grammaticality 
(? = -148ms, SE = 56.4, t = -2.64). To resolve this interaction a second model was fit 
to evaluate planned pairwise comparisons, directly comparing differences due to 
interference within grammatical and ungrammatical sentences. This model revealed 
that the interaction was driven by a significant difference due to interference for 
ungrammatical sentences (? = -117.2, SE = 39.6, t = -2.96); no such difference 
obtained in grammatical environments. This pattern replicates the interference 
asymmetry noted in Wagers et al (2009). In the spill-over region for agreement 
conditions, modeling revealed a significant effect of grammaticality (? = 78.7, SE = 
33.7, t = 2.33). No other fixed effects reached significance in this region. 
 In first-pass regression probabilities, modeling revealed a significantly greater 
probability of regressing in the ungrammatical conditions relative to grammatical 
conditions at the critical agreeing verb (? = .587, SE = .180, Wald z = 3.26, p < .005). 
There was only a marginal interaction of grammaticality with interference (? = -.681, 
SE = .360, Wald z = -.183, p < .06), although this is consistent with the pattern seen 
in total time measures, and in the predicted direction. However, a model containing 
 
 
60 
planned comparisons to resolve this interaction showed no significant differences due 
to interference within either grammatical or ungrammatical conditions. In the 
spillover region, there was only a marginal effect of grammaticality on the probability 
of regression (? = .273, SE = .152, Wald z = 1.79, p < .08). 
 
 
 
 
 
 
 
 
Reflexives NP1 NP2 Critical Spillover 
First Pass     
Gram, no int  583 (31) 619 (30) 299 (12) 650 (49) 
Gram, int  579 (32) 645 (35) 295 (13) 659 (35) 
Ungram, no int 570 (36) 621 (29) 351 (17) 628 (34) 
Ungram, int 567 (32) 651 (29) 342 (16) 631 (27) 
Total Time     
Gram, no int  841 (49) 951 (55) 481 (20) 851 (48) 
Gram, int  819 (47) 907 (48) 471 (24) 846 (42) 
Ungram, no int 863 (56) 1023 (59) 588 (30) 918 (62) 
Ungram, int 890 (59) 1040 (54) 580 (30) 882 (47) 
Pr(Regression)     
Gram, no int  - .19 (.03)  .14 (.03) .73 (.04) 
Gram, int  - .18 (.03) .16 (.03) .77 (.04) 
Ungram, no int - .16 (.03) .09 (.02) .77 (.03) 
Ungram, int - .15 (.03) .10 (.02) .77 (.04) 
 
Table 2.5: Table of means (in ms where applicable) for reflexive conditions for first 
pass, total time, and probability of regression. Standard error by participant is shown 
in parentheses.   
 
 
 
 
61 
Results: Reflexives 
 
 The by-region reading times and regression probabilities for the reflexives 
conditions are presented in Table 2.5. The results of models fit at the critical reflexive 
region are shown in Table 2.7. In first-pass measures, no significant effects were 
observed prior to the critical region. At the critical reflexive region, there was a 
significant effect of grammaticality, indicating a slowdown for ungrammatical 
conditions (? = 50.7, SE = 10.5, t = 4.81). This slowdown in ungrammatical 
conditions was also observed in total time measures at the critical reflexive region (? 
= 110.1, SE = 21.8, t = 5.05). In neither measure was there an interaction of 
grammaticality and interference, nor was there a consistent trend for the direction of 
this effect across measures. Although the interaction of interference and 
grammaticality did not reach significance, to maximize the chance of showing a 
difference due to interference I performed the same planned comparisons that we 
applied to the total reading times at the critical agreement region. This model failed to 
show an effect of interference for either ungrammatical or grammatical sentences. 
Surprisingly, the effect of ungrammaticality on the probability of regression for 
reflexive measures was a decreased probability of regression (? = .494, SE = .207, 
Wald z = -2.38, p < 0.05). There was again no significant effect of either an 
interfering plural, nor was there an interaction of interference and grammaticality on 
the probability of regression from the critical reflexive region. 
  
Agreement ? SE t / z 
First pass    
GRAM 62.4 14.7 4.25* 
 
 
62 
INTR 3.8 12.5 0.31 
GRAM?INTR -21.8 25.0 -0.87 
Total Time    
GRAM 157.2 28.0 5.60* 
INTR -43.0 28.1 -1.53 
GRAM?INTR -148.5 56.4 -2.64* 
Pr(Regression)    
GRAM .587 .180 3.26* 
INTR .004 .180 0.02 
GRAM?INTR -.681 .360 -1.83? 
 
Table 2.6: Summary of fixed effects for best-fit models on agreement conditions at 
the critical agreeing verb region, including t-values (z-values for first-pass regression 
probability data). An asterisk (*) indicates significance at ? = 0.05, while a cross (?) 
indicates significance at ? = 0.10. First-pass and total time coefficients are in 
milliseconds.  
 
 At the spillover region, there were no significant effects of the experimental 
manipulations in any of the measures. Likewise, there were no significant effects on 
the total reading at NP1. At the interfering NP position (NP2), however, modeling 
revealed a significant effect of grammaticality on total reading times for reflexive 
conditions. The NP2 position had significantly higher reading times (? = 100.0, SE = 
32.0, t = 3.13) when the reflexive was ungrammatical. There were no significant 
effects for interference or the interaction of interference and grammaticality at this 
region. There were no significant effects on probability of regression at this region. 
 
 
 
 
 
  
 
 
63 
Reflexives ? SE t / z 
First pass    
GRAM 50.7 10.5 4.81* 
INTR -3.6 10.6 -0.34 
GRAM?INTR -6.3 21.1 -0.30 
Total Time    
GRAM 110.1 21.8 5.05* 
INTR -2.6 19.3 -0.14 
GRAM?INTR 1.25 38.6 0.03 
Pr(Regression)    
GRAM -.494 .207 -2.38* 
INTR -.055 .207 -0.27 
GRAM?INTR -.288 .415 -0.69 
 
Table 2.7: Summary of fixed effects for best-fit models on reflexive conditions at the 
critical reflexive region, including t-values (z-values for first-pass regression 
probability data). An asterisk (*) indicates significance at ? = 0.05, while a cross (?) 
indicates significance at ? = 0.10. First-pass and total time coefficients are in 
milliseconds.  
 
Direct comparison of interference effect 
 
 Results showed a reliable effect of the number of an interfering noun for 
agreement, and no corresponding effect for reflexives. In order to provide a direct 
measure of the interference effect across dependencies, however, a direct statistical 
comparison of the size of the effect across the two dependencies is called for. Within 
ungrammatical and grammatical sentences, I compared the size of the interference 
effect, a derived measure that was calculated by subtracting the reading time for [-
 intr] conditions from [+intr] conditions for total reading times in the critical region. A 
direct test of the size of the interference effect across agreement and reflexives 
revealed that there was a significantly larger interference effect for agreement in the 
ungrammatical conditions (t(39) = -2.5, p < 0.02; agreement ? = -118.6ms ? 42.8, 
 
 
64 
reflexive ? = -7.8ms ? 29.6). There was no significant difference between agreement 
and reflexives in the grammatical conditions (t(39) = 1.4, p < 0.3; agreement ? = 
43.0ms ? 31.6, reflexive ? = -10.1ms ? 28.8).  
 
Discussion of Experiment 1 
  
 A summary of the first-pass and total reading times in the critical region, for 
agreement and reflexives, is summarized in Figures 2.7 and 2.8. For reflexive regions, 
across all three measures reported, and all other regions of interest (notably the 
interfering noun position NP2), the only significant fixed effect was the effect of 
grammaticality (GRAM). Ungrammatical conditions reliably lead to longer reading 
times, although surprisingly they led to fewer backwards regressions from the 
ungrammatical reflexive. In contrast, there were no significant effects of interference 
or the interaction of grammaticality with interference at any region or measure. 
Potentially as important as the lack of reliability in the interaction across regions and 
measures is the relative instability of the direction of the interaction effect across 
measures (Gelman & Hill 2005), showing that there was no consistent trend for 
interference in the reflexive conditions. There was thus no evidence, in any region or 
measure, for partial-matching interference in the reflexive conditions. This extends 
the findings of Sturt (2003) and Xiang et al (2009) in showing the relative robustness 
of reflexive dependencies to interference in ungrammatical conditions with number as 
well as gender.  
 
 
65 
 
Figure 2.7: Mean first-pass reading time at the critical region in Experiment 1. Error 
bars show standard error by participants. 
Agreement Reflexive
 Critical region first pass reading time in Experiment 1
 R
 T
 0
 20
 0
 40
 0
 60
 0
 80
 0
 [+gram],[-intr]
 [-gram],[+intr]
 [-gram],[-intr]
 [-gram],[+intr]
 
 
66 
 
Figure 2.8: Mean total reading time at the critical region in Experiment 1. Error bars 
show standard error by participants. 
 
 The processing of reflexive conditions clearly contrasts with the processing of 
agreement conditions, which show a qualitatively different pattern. Although 
ungrammaticality reliably slowed down the processing of both agreement and 
reflexives, only agreement showed reliable interaction effects (a significant effect at 
the critical agreement region for total times, and a marginal effect on probability of 
regression). Planned comparisons confirmed that this interaction was driven by 
interference in ungrammatical conditions. Interference led to shorter reading times in 
ungrammatical sentences, replicating earlier findings (Pearlmutter et al 1999; Wagers 
et al 2009) of clear partial-matching interference effects for agreement in 
Agreement Reflexive
 Critical region total reading time in Experiment 1
 R
 T
 0
 20
 0
 40
 0
 60
 0
 80
 0
 [+gram],[-intr]
 [-gram],[+intr]
 [-gram],[-intr]
 [-gram],[+intr]
 
 
67 
comprehension. It is worth noting that across all region and measure pairings that the 
effect of interference had the same direction: for ungrammatical sentences, an 
interfering partial-match noun phrase leads to speeded processing or fewer 
regressions, relative to sentences in which there is no partial match. 
 This contrast is telling, but it is always difficult to mount an argument based 
on null effects (the lack of an interaction in the reflexive conditions and the lack of 
any pairwise differences due to interference). This concern is mitigated by the fact 
that a direct comparison between the size of the interference in agreement and 
reflexive dependencies nonetheless indicated that the observed interference effects 
were reliably larger for agreement. In addition, the lack of reliability of the direction 
of the critical interaction effect, coupled with the very small effect sizes seen in the 
reflexive conditions, suggest a lack of interference effect when compared to the 
agreement conditions. Across a wide range of measures and regions, agreement 
reliably showed partial-matching interference effects. If the size and variance of the 
predicted interference effect should be similar for both agreement and reflexives, the 
finding that our experimental manipulation showed the effect for agreement suggests 
that the experiment had sufficient power to detect the reflexive effect. This argument 
rests on an assumption of equally sized effects that might not in fact be true, and I 
return to the question of the size of the predicted effect in Chapter 3.  
 In Figure 2.7 it can be seen that in first-pass measures, agreement and 
reflexives appear to show an identical profile. This may be due to an early processing 
stage in which they are in fact processed in the same manner. However, it may also 
reflect more ?low-level? effects. For both agreement and reflexives, the length of the 
 
 
68 
critical region (as well as the identity of the critical words) covaried with the factor of 
grammaticality. This makes it difficult to know if pattern in Figure 2.7 is due to 
effects such as word length, or if they index an early, identical ungrammaticality 
detection stage for agreement and reflexives. Experiments 2 and 3 aimed to shed 
some light on this early grammaticality effects in the first pass measures reported 
here. Experiments 2 and 3 employ designs with a full crossing of subject NP number 
and the number of the agreeing verb or reflexive, in order to further tease apart the 
processes involved in the earliest stages of processing reflected in the eye-movements 
here. If low-level factors drive the identical patterns seen in Figure 2.7, then this 
pattern should be insensitive to the head noun number. If instead the first-pass reading 
times in Figure 2.7 reflect an early grammaticality detection stage, then the pattern 
should be reversed for plural subjects. That is, lower reading times should occur for 
ungrammatical verb forms, rather than longer words. 
An important additional goal of Experiments 2 and 2 is to provide additional 
evidence for the qualitatively different processing profiles exhibited by agreement 
and reflexives. To build an even stronger case for distinct interference profiles, 
Experiments 2 and 3 attempt to replicate the interference profiles observed in the 
agreement and reflexive conditions, respectively. 
 
Experiment 2: Agreement revisited 
 
 The results of Experiment 1 suggested a qualitatively different use of 
morphological features in comprehending reflexive and agreement dependencies. For 
 
 
69 
agreement features, a feature-matching but inaccessible NP gave rise to clear partial-
 matching inference effects across regions and measures, suggesting that the 
morphological features were used to access the subject in memory. No such 
interference was observed for reflexives. However, both agreement and reflexives 
patterned alike in early first pass measures; both dependencies showed a main effect 
of ungrammaticality. To determine whether or not this reflected an early stage of 
grammaticality detection free of interference, or was instead due to length differences 
in the critical word across the factor of grammaticality (was versus were), Experiment 
2 expanded the agreement manipulation of Experiment 1 by adding the factor of head 
noun number. If the early effect of grammaticality is due to differences in word 
length, then increased reading times should be seen for all instances of were as 
opposed to was, regardless of head noun number. On the other hand, if the early 
effect genuinely reflects detection of grammaticality, then increased reading times 
should be observed whenever the verb does not match in features with the head noun. 
An additional goal of Experiment 2 was to replicate and strengthen the finding of 
partial-matching interference for agreement that was observed in Experiment 1. 
 
Participants 
 
 32 members of the University of Maryland community participated in 
Experiment 2 (24 females, mean age 21.9). Participants gave informed consent, and 
were either paid $10 or received course credit for their participation. The 
experimental session, including set-up and calibration, lasted approximately one hour. 
 
 
70 
 
Materials 
 
 The materials and fillers were identical to those in Experiment 1, with the 
exception that the four reflexive conditions were removed. Instead, an extra factor of 
head noun number was introduced into the agreement conditions, leading to a 2?2?2 
factorial design that crossed head noun number, interfering noun number, and verbal 
number. The full set of experiment conditions in Experiment 2 is shown in Table 2.8. 
 
Agreement conditions for Experiment 2  
Singular head, singular interferer, grammatical 
1 The new executive/ who oversaw/ the middle manager/ apparently/ was dishonest / about the company?s profits. 
Singular head, plural interferer, grammatical 
2 The new executive/ who oversaw/ the middle managers/ apparently/ was dishonest / about the company?s 
profits. 
Singular head, singular interferer, ungrammatical 
3 The new executive/ who oversaw/ the middle manager/ apparently/ were dishonest / about the company?s 
profits. 
Singular head, plural interferer, ungrammatical  
4 The new executive/ who oversaw/ the middle managers/ apparently/ were dishonest / about the company?s 
profits. 
Plural head, singular interferer, ungrammatical 
5 The new executives/ who oversaw/ the middle manager/ apparently/ was dishonest / about the company?s 
profits. 
Plural head, plural interferer, ungrammatical 
6 The new executives/ who oversaw/ the middle managers/ apparently/ was dishonest / about the company?s 
profits. 
Ungrammatical, no interference 
7 The new executive/ who oversaw/ the middle manager/ apparently/ were dishonest / about the company?s 
profits. 
Ungrammatical, interference 
8 The new executive/ who oversaw/ the middle managers/ apparently/ were dishonest / about the company?s 
profits. 
 
Table 2.8: Summary of agreement conditions in Experiment 2. Regions included in 
the analysis are underlined.   
 
Procedure 
  
 The experimental set-up and procedure was the same as in Experiment 1. 
 
 
 
71 
Data Analysis  
 
 Sentences in Experiment 2 were divided into identical regions as those in 
Experiment 1. As in Experiment 1, I report analyses on first-pass, total time, and 
probability of regression measures.  
 
Singular Head NP1 NP2 Critical Spillover 
First Pass     
Gram, sing  638 (26) 708 (29) 414 (15) 923 (42) 
Gram, pl  595 (21) 722 (29) 445 (17) 976 (48) 
Ungram, sing 615 (23) 674 (37) 530 (21) 975 (45) 
Ungram, pl 669 (28) 697 (31) 497 (22) 950 (45) 
Total Time     
Gram, sing  898 (60) 1064 (76) 569 (27) 1170 (65) 
Gram, pl  867 (56) 1129 (88) 656 (52) 1296 (109) 
Ungram, sing 950 (53) 1085 (94) 875 (52) 1359 (107) 
Ungram, pl 980 (59) 1139 (68) 784 (46) 1294 (80) 
Pr(Regression)     
Gram, sing  - .16 (.03)  .16 (.03) .54 (.05) 
Gram, pl  - .19 (.03) .14 (.03) .55 (.05) 
Ungram, sing - .22 (.03) .26 (.04) .64 (.04) 
Ungram, pl - .19 (.03) .18 (.02) .58 (.05) 
 
Table 2.9: Table of means (in ms where applicable) for Experiment 2, agreement 
conditions with a singular head noun, for first pass, total time, and probability of 
regression. Standard error by participant is shown in parentheses.   
 
 As in Experiment 1, all statistical analysis was performed using mixed-effect 
linear regressions. In addition to grammaticality (GRAM; whether or not the sentence 
was grammatical), there were fixed effects for plurality (PLUR; whether or not the 
embedded NP was plural), and head number (HEAD) that indicated whether or not the 
head noun was plural or singular. PLUR is identical to the fixed effect for interference 
in Experiment 1. In the context of the full factorial design, calling this factor 
interference risks being misleading. Whether or not a particular number value 
 
 
72 
?interferes? depends on the number of the verb, and so presenting this factor simply as 
embedded noun number is more straightforward. The fixed effects were coded 
identically to Experiment 1, using simple difference coding (e.g., HEAD was coded as 
-0.5 for singular heads, 0.5 for plural heads). The model-fitting procedure and 
reporting of the results was identical to Experiment 1. Significant interactions with 
head number were resolved by assessing the interaction of grammaticality and 
plurality within plural and singular head noun conditions. Any further interactions 
were subjected to the same planned pairwise comparisons that were used in 
Experiment 1 to test the effect of interference within the levels of the factor of 
grammaticality.  
 
 
Plural Head NP1 NP2 Critical Spillover 
First Pass     
Gram, sing  638 (29) 718 (35) 507 (24) 941 (57) 
Gram, pl  691 (27) 701 (24) 456 (18) 988 (42) 
Ungram, sing 665 (28) 692 (34) 473 (22) 936 (52) 
Ungram, pl 659 (29) 721 (29) 473 (16) 919 (46) 
Total Time     
Gram, sing  981 (62) 1099 (67) 784 (44) 1256 (86) 
Gram, pl  1012 (62) 1100 (68) 670 (37) 1277 (79) 
Ungram, sing 989 (61) 1057 (76) 742 (47) 1247 (70) 
Ungram, pl 997 (66) 1146 (69) 752 (36) 1287 (77) 
Pr(Regression)     
Gram, sing  - .17 (.03)  .20 (.03) .61 (.04) 
Gram, pl  - .18 (.03) .15 (.03) .56 (.05) 
Ungram, sing - .17 (.03) .16 (.02) .60 (.05) 
Ungram, pl - .21 (.04) .27 (.04) .60 (.05) 
Table 2.10: Table of means (in ms where applicable) for Experiment 2, agreement 
conditions with a plural head noun, for first pass, total time, and probability of 
regression. Standard error by participant is shown in parentheses.   
 
Results 
 
 
 
73 
 In the pre-critical NP1 region, models of both first-pass and total reading 
times suggested a slowdown for plural nouns relative to singular nouns (first pass: ? = 
32.6, SE = 14.7, t = 2.21; total time: ? = 65.2, SE = 26.6, t = 2.46). Unexpectedly, the 
model for the first-pass times also included a significant term for a three-way 
interaction of HEAD, PLUR, and GRAM (? = -161.1, SE = 59.0, t = -2.73), which was 
unexpected due to the fact that participants had not yet seen either the embedded noun 
or the matrix verb, and these were far outside of plausible parafoveal viewing regions. 
Resolving this interaction showed a significant interaction of PLUR and GRAM for 
singular head nouns (? = 97.2, SE = 41.7, t = 2.33), though resolving this interaction 
showed no significant differences in planned pairwise comparisons, suggesting that 
the apparent interaction was spurious and not driven by any true differences between 
conditions. Additionally, in the same pre-critical NP1 region, no such interactions 
were observed in total reading times.  
At the pre-critical NP2 region, no significant experimental effects were 
observed in first-pass, total time, or probability of regression measures.  
 The model fits to first-pass, total time and regression probability at the critical 
agreeing verb region are summarized in Table 2.11.  Across all measures, a 
significant effect of grammaticality is observed, as are the interactions of GRAM, 
PLUR, and HEAD, and of HEAD and GRAM, though this later interaction did not reach 
significance in probability of regression.  
 Resolving the interaction of HEAD?GRAM for singular and plural head nouns 
showed that in both first-pass and total time reading measures, there was a significant 
effect of GRAM in singular head noun conditions (first pass: ? = 83.2, SE = 16.2, t = -
 
 
74 
5.15; total time: ? = 209.9, SE = 28.2, t = -7.43); no significant effects were observed 
in plural head noun conditions. Replicating the findings of Experiment 1, a significant 
interaction of GRAM?PLUR was observed in singular head noun conditions for total 
reading times (? = -179.7, SE = 56.5, t = -3.19), but this just barely reached 
significance in the first-pass reading times (? = -64.4, SE = 32.3, t = -2.00).  
  
 ? SE t / z 
First pass    
HEAD 4.6 11.4 0.41 
PLUR -13.5 11.4 -1.18 
GRAM 37.2 11.4 3.26* 
HEAD?PLUR -22.5 22.9 -0.98 
HEAD?GRAM -92.1 22.9 -4.03* 
GRAM?PLUR -7.5 22.8 -0.33 
HEAD?GRAM?PLUR 113.8 45.7 2.49* 
Total Time    
HEAD 10.3 20.0 0.52 
PLUR -30.5 20.0 -1.53 
GRAM 116.2 20.0 5.82* 
HEAD?PLUR -36.9 40.0 -0.93 
HEAD?GRAM -187.5 39.9 -4.69* 
GRAM?PLUR -34.2 39.9 -0.86 
HEAD?GRAM?PLUR 291.0 79.8 3.65* 
Pr(Regression)    
HEAD 0.073 0.140 0.53 
PLUR -0.114 0.140 -0.82 
GRAM 0.401 0.140 2.87* 
HEAD?PLUR 0.525 0.280 1.88? 
HEAD?GRAM -0.206 0.280 -0.74 
GRAM?PLUR 0.387 0.280 1.38 
HEAD?GRAM?PLUR 1.424 0.559 2.55* 
 
Table 2.11: Summary of fixed effects for best-fit models at the critical agreeing verb 
region in Experiment 2, including t-values (z-values for first-pass regression 
probability data). An asterisk (*) indicates significance at ? = 0.05; a cross (?) 
indicates significance at ? = 0.10. First-pass and total time coefficients are in 
milliseconds.  
 
 
75 
 
 Further resolving this interaction for first-pass times showed no differences 
due to interference for either ungrammatical or grammatical conditions with singular 
head nouns, though surprisingly there was a significant difference due to interference 
for grammatical sentences in the plural head noun conditions (? = -49.4, SE = 22.8, t 
= -2.16). In total reading times, faster reading times were observed when the 
embedded noun was plural in ungrammatical environments with singular head nouns 
(? = -101.9, SE = 39.9, t = -2.56), replicating the illusion of grammaticality observed 
in Experiment 1. In addition, the interference effect in grammatical, plural head noun 
conditions was observed (? = -104.6, SE = 39.9, t = -2.62).  
 A somewhat different pattern was suggested in the probability of regression 
measure. Resolving the interaction suggested that, as with the other measures, the 
difficulty due to ungrammatical conditions was observed only for singular head noun 
conditions (? = -0.504, SE = 0.200, Wald z = 2.52, p < 0.05). No interaction of 
GRAM?PLUR was observed in singular head noun conditions, though it did reach 
significance in plural head noun conditions (? = 1.099, SE = 0.391, Wald z = 2.81, p 
< 0.01). Resolving this interaction, showed an interference effect in ungrammatical 
conditions both for plural head nouns (? = -0.698, SE = 0.266, Wald z = -2.62, p < 
0.01) and singular head nouns (? = -0.540, SE = 0.287, Wald z = -2.06, p < 0.05). 
 In the spillover region, no significant experimental fixed effects were 
observed in first-pass or total reading times. In the probability of regression measure, 
only a marginal effect of grammaticality was observed (? = 0.218, SE = 0.121, Wald 
z = 1.81, p < 0.08).  
 
 
76 
 
Discussion 
 
 The central finding from Experiment 2 is the replication of the interference 
effect in agreement for the singular head noun conditions. The first-pass and total 
reading times at the critical verb are shown in Figure 2.9 and 2.10. As in Experiment 
1, ungrammatical sentences were read more quickly in the critical verb region when 
the interfering noun position had plural features that matched the verb. Somewhat 
surprisingly, there was only limited evidence for processing difficulty related to 
ungrammaticality in the plural head noun conditions. This contrasts with a finding by 
Pearlmutter et al 1999, who found that participants were sensitive to 
ungrammaticality in reading time measures with a plural head noun. The apparent 
lack of grammaticality effect for plural head nouns was supported by an interaction of 
head number and grammaticality that was found across measures and subsequent 
pairwise comparisons.  
 
 
77 
 
Figure 2.9: Mean first-pass reading time at the critical region in Experiment 2. Error 
bars show standard error by participants. 
Singular head Plural head
 Critical region first pass reading time in Experiment 2
 R
 T
 0
 20
 0
 40
 0
 60
 0
 80
 0
 10
 00
 12
 00
 [+gram],[-plur]
 [+gram],[+plur]
 [-gram],[-plur]
 [-gram],[+plur]
 
 
78 
 
Figure 2.10: Mean total reading time at the critical region in Experiment 2. Error bars 
show standard error by participants. 
 
 
 In addition to replicating the crucial illusion of grammaticality effect, an 
important goal of Experiment 2 was to investigate the nature of the first-pass 
grammaticality effect observed in Experiment 1?s agreement conditions (see Figure 
2.7 above). In Experiment 1, both agreement and reflexives patterned alike in first-
 pass measures, with neither showing any interference effect. This led to the 
possibility that there was an early stage of processing in which agreement and 
reflexives patterned alike. However, this conclusion is not supported by the results of 
Experiment 2. Since the pattern of difficulty observed in first-pass measures in 
Experiment 2 were not identical for all instances of was vs. were, this effect is not 
Singular head Plural head
 Critical region total reading time in Experiment 2
 R
 T
 0
 20
 0
 40
 0
 60
 0
 80
 0
 10
 00
 12
 00
 [+gram],[-plur]
 [+gram],[+plur]
 [-gram],[-plur]
 [-gram],[+plur]
 
 
79 
fully driven by visual or lexical properties of the agreeing verb. This finding in itself 
is consistent with there being an early stage of processing that is identical for 
agreement and reflexives. However, the first-pass pattern for agreement in 
Experiment 1 did not replicate in Experiment 2. Instead, even in first-pass measures 
there was a marginal interaction of NP2 plurality and grammaticality that suggested 
partial-matching interference. Thus, unlike Experiment 1, there does not appear to be 
a qualitative difference in the pattern observed in early and late measures, making it 
difficult to conclude that the pattern shown in Figure 2.7 suggests a distinct, 
interference-free stage of processing for agreement. 
Although the attraction effect is usually limited to singular head noun 
environments in production (Bock and Miller 1991) and comprehension (Wagers et al 
2009), in Experiment 2 it was evident for plural head nouns conditions in first-pass 
regression probability. That is, there were fewer regressions launched when the 
inaccessible NP position matched the number features of the ungrammatical singular 
verb was. Somewhat surprisingly, there was also a significant slowdown in 
grammatical plural head noun environments: when the interfering NP2 position was 
singular, reading times in the critical region were slower than when it was plural.  
Lastly, there were also significant slowdowns at the head NP1 position when 
it was plural relative to when it was singular. This effect may be understood as a 
plural complexity effect (Wagers et al 2009), where the increased reading times 
correspond to extra processing engaged by plural nouns relative to singular nouns. 
Experiment 2 thus replicates the crucial pattern of partial-matching 
interference in the present materials, and shows that this interference is evident from 
 
 
80 
the earliest stages of processing. The lack of any interference in early measures in 
Experiment 1 was not due to an early, interference-free stage of processing. Instead, it 
is more likely to reflect a lack of power in early measures.  
  
 Experiment 3: Reflexives revisited 
 
 Together, Experiments 1 and 2 demonstrate that the computation of subject-
 verb agreement is reliably susceptible to partial-matching interference: across 
experiments, the effect of inaccessible, but feature-matching noun phrases was 
evident. In Experiment 1 no corresponding interference effect was observed for 
reflexives. Experiment 3?s main goal was to replicate this result for reflexives, which 
would build a stronger case for structured access in the case of reflexives. As in 
Experiment 2, extra conditions that examine the effect of having a plural head noun 
were included, making a fully crossed 2?2?2 factorial design parallel to the design 
for agreement in Experiment 2. 
 
Participants 
 
 32 members of the University of Maryland community participated in 
Experiment 2 (24 females, mean age 21.9). Participants gave informed consent, and 
were either paid $10 for their participant or received course credit. The experimental 
session, including set-up and calibration, lasted approximately one hour. 
 
 
 
81 
Materials 
 
 The materials and fillers were identical to those used in Experiment 1, with 
the exception that the four reflexive conditions were removed. Instead, an extra factor 
of head noun number was introduced, leading to a 2?2?2 factorial design that crossed 
head noun number, interfering noun number, and verbal number. The full set of 
experiment conditions in Experiment 2 is found in Table 2.10. 
 
Reflexive conditions for Experiment 3  
Singular head, singular interferer, grammatical 
1 The new executive/ who oversaw/ the middle manager/ apparently doubted/ himself on/ most major decisions. 
Singular head, plural interferer, grammatical 
2 The new executive/ who oversaw/ the middle managers/ apparently doubted/ himself on/ most major decisions. 
Singular head, singular interferer, ungrammatical 
3 The new executive/ who oversaw/ the middle manager/ apparently doubted/ themselves on/ most major 
decisions. 
Singular head, plural interferer, ungrammatical  
4 The new executive/ who oversaw/ the middle managers/ apparently doubted/ themselves on/ most major 
decisions. 
Plural head, singular interferer, ungrammatical 
5 The new executives/ who oversaw/ the middle manager/ apparently doubted/ himself on/ most major decisions. 
Plural head, plural interferer, ungrammatical 
6 The new executives/ who oversaw/ the middle managers/ apparently doubted/ himself on/ most major decisions. 
Ungrammatical, no interference 
7 The new executive/ who oversaw/ the middle manager/ apparently doubted/ themselves on/ most major 
decisions. 
Ungrammatical, interference 
8 The new executive/ who oversaw/ the middle managers/ apparently doubted/ themselves on/ most major 
decisions. 
 
Table 2.12: Summary of reflexive conditions in Experiment 3. Regions included in 
the analysis are underlined.   
Data Analysis  
 
 Data analysis was identical to Experiment 2. 
 
Results 
 
 
 
82 
 In first-pass and total-time measures, modeling revealed a significant effect of 
head noun number at NP1, reflecting a slow-down for plurals (first-pass: ? = 84.3, SE 
= 16.6, t = 5.01; total time: ? = 108.0, SE = 26.1, t = 4.12). Likewise, there was a 
significant effect of noun number in first pass measures at NP2, also reflecting a 
plural complexity effect (first-pass: ? = 40.3, SE = 15.7, t = 2.57; total time: ? = 79.1, 
SE = 35.7, t = 2.22). At NP2, first pass regression probabilities revealed a significant 
three-way interaction of HEAD, PLUR, and GRAM (? = -1.23, SE = 0.538, t = -2.29), but 
resolving this interaction led to no significant differences. Grammaticality effects 
were also observed in total time measures at both NP1 and NP2 regions (NP1: ? = 
87.8, SE = 26.1, t = 3.36; NP2: ? = 88.8, SE = 25.3, t = 3.50).  
 
Singular Head NP1 NP2 Critical Spillover 
First Pass     
Gram, sing  586 (28) 611 (25) 292 (11) 676 (44) 
Gram, pl  585 (29) 678 (32) 302 (13) 708 (42) 
Ungram, sing 604 (30) 656 (31) 327 (12) 664 (44) 
Ungram, pl 604 (28) 670 (32) 345 (16) 696 (38) 
Total Time     
Gram, sing  857 (63) 957 (50) 452 (24) 882 (61) 
Gram, pl  872 (74) 1029 (56) 444 (20) 886 (56) 
Ungram, sing 905 (57) 1050 (59) 574 (41) 952 (71) 
Ungram, pl 956 (65) 1151 (64) 605 (32) 968 (54) 
Pr(Regression)     
Gram, sing  - .24 (.03)  .13 (.03) .57 (.04) 
Gram, pl  - .19 (.03) .09 (.02) .47 (.05) 
Ungram, sing - .19 (.02) .15 (.03) .64 (.05) 
Ungram, pl - .21 (.03) .12 (.02) .67 (.04) 
 
Table 2.13: Table of means (in ms where applicable) for Experiment 3, reflexive 
conditions with a singular head noun, for first pass, total time, and probability of 
regression. Standard error by participant is shown in parentheses.     
 
 A summary of the fixed effects at the critical reflexive region is presented in 
Table 2.12. Both first-pass and total times revealed a similar pattern. Modeling of 
 
 
83 
both measures revealed a main effect of grammaticality (first pass: ? = 17.5, SE = 8.4, 
t = 2.07; total time: ? = 75.5, SE = 21.3, t = 3.55), as well as an interaction of head 
number and grammaticality (first pass: ? = -44.4, SE = 15.1, t = -2.92; total time: ? = 
-146.1, SE = 30.1, t = -4.85). Resolving this interaction revealed identical patterns 
across first-pass and total times. Within singular head noun conditions, there was a 
significant effect of GRAM (first pass: ? = 39.6, SE = 11.3, t = 3.49; total time: ? = 
148.5, SE = 26.1, t = 5.70), but no significant effects of GRAM were obtained when 
the head noun was plural. There were no significant effects on the probability of 
regression at the critical region. Importantly were no significant effects of the 
interfering noun?s number, nor were there interactions of interfernoun number with 
any of the other fixed effects. In addition, a significant effect of grammaticality was 
observed in the spillover region (? = 57.7, SE = 22.7, t = 2.55), with no interaction of 
grammaticality and embedded noun plurality. 
 
 
Plural Head NP1 NP2 Critical Spillover 
First Pass     
Gram, sing  681 (46) 651 (27) 311 (11) 742 (57) 
Gram, pl  688 (38) 655 (35) 306 (11) 686 (42) 
Ungram, sing 683 (39) 621 (24) 297 (10) 684 (52) 
Ungram, pl 665 (35) 695 (25) 309 (15) 707 (46) 
Total Time     
Gram, sing  933 (62) 991 (46) 486 (21) 930 (86) 
Gram, pl  967 (65) 1051 (66) 508 (26) 909 (79) 
Ungram, sing 1046 (70) 1053 (56) 513 (26) 971 (70) 
Ungram, pl 1077 (78) 1125 (57) 494 (27) 943 (77) 
Pr(Regression)     
Gram, sing  - .17 (.03)  .13 (.02) .52 (.04) 
Gram, pl  - .22 (.03) .15 (.02) .58 (.05) 
Ungram, sing - .24 (.03) .15 (.03) .59 (.05) 
Ungram, pl - .18 (.04) .18 (.04) .61 (.05) 
 
 
 
84 
Table 2.14: Table of means (in ms where applicable) for Experiment 3, reflexive 
conditions with a plural head noun, for first pass, total time, and probability of 
regression. Standard error by participant is shown in parentheses.   
 
 
 
 
 ? SE t / z 
First pass    
HEAD -11.7 7.6 -1.55 
PLUR 9.6 7.6 1.28 
GRAM 17.5 8.4 2.08* 
HEAD?PLUR -9.0 15.1 -0.60 
HEAD?GRAM -44.1 15.1 -2.92* 
GRAM?PLUR 12.1 15.1 0.80 
HEAD?GRAM?PLUR 5.9 30.2 0.19 
Total Time    
HEAD -20.8 15.0 -1.38 
PLUR 8.0 17.3 0.46 
GRAM 75.5 21.3 3.55* 
HEAD?PLUR -2.9 30.1 -0.10 
HEAD?GRAM -146.1 30.1 -4.85* 
GRAM?PLUR -2.9 37.1 -0.01 
HEAD?GRAM?PLUR -90.7 60.1 -1.5 
Pr(Regression)    
HEAD 0.242 0.158 1.54 
PLUR -0.025 0.158 -0.16 
GRAM 0.256 0.158 1.62 
HEAD?PLUR 0.452 0.316 1.43 
HEAD?GRAM -0.188 0.316 -0.59 
GRAM?PLUR 0.080 0.316 0.25 
HEAD?GRAM?PLUR -0.114 0.631 -0.18 
 
Table 2.15: Summary of fixed effects for best-fit models at the critical reflexive 
region in Experiment 3, including t-values (z-values for first-pass regression data). An 
asterisk (*) indicates significance at ? = 0.05; a cross (?) indicates significance at ? = 
0.10. First-pass and total time coefficients are in milliseconds.  
  
 
 
85 
 Discussion 
 
 The most important finding from Experiment 3 is the replication of the 
processing profile for reflexives that was observed in Experiment 1. This can be seen 
in Figures 2.11 and 2.12, which show mean first-pass and total reading times, 
respectively. In no measure or region was there a measurable effect of interfering 
noun number, nor did it interact with any of the other factors, thus providing 
additional support for the claim that reflexives are initially processed in a way that is 
blind to the feature content of an structurally inaccessible NP. 
 
Figure 2.11: Mean first-pass reading time at the critical region in Experiment 3. Error 
bars show standard error by participants. 
 
Singular head Plural head
 Critical region first pass reading time in Experiment 3
 R
 T
 0
 20
 0
 40
 0
 60
 0
 80
 0
 10
 00
 12
 00
 [+gram],[-plur]
 [+gram],[+plur]
 [-gram],[-plur]
 [-gram],[+plur]
 
 
86 
 
Figure 2.12: Mean total reading time at the critical region in Experiment 3. Error bars 
show standard error by participants. 
 
 As in Experiment 2, at the critical reflexive region comprehenders appeared to 
be less sensitive to the feature match between antecedent and reflexive when the head 
noun was plural. This is surprising in light of the fact that there was no comparable 
asymmetry in the ungrammaticality response in any other region that showed an 
effect of grammaticality. These data could suggest that the process of recognizing 
feature mismatch is sensitive to the markedness of the features involved, but this is 
entirely speculative at this point. The empirical case for this is unclear, as the 
asymmetric grammaticality detection pattern was only seen in one region. In regions 
other than the critical region, a main effect of grammaticality was observed for both 
singular and plural head nouns alike. It is an interesting possibility that deserves 
Singular head Plural head
 Critical region total reading time in Experiment 3
 R
 T
 0
 20
 0
 40
 0
 60
 0
 80
 0
 10
 00
 12
 00
 [+gram],[-plur]
 [+gram],[+plur]
 [-gram],[-plur]
 [-gram],[+plur]
 
 
87 
further study, but this result be treated with caution. This result needs to be replicated 
and extended before any conclusions about its import for theories of sentence 
processing can be drawn.  
 
Overview of Experiments 1-3 
 
 Across Experiments 1-3, a clear picture of the interference profiles for 
subject-verb agreement and subject-reflexive binding emerges. Focusing on the four 
conditions that were repeated across all experiments, i.e., the singular head noun 
conditions, it can be seen that agreement reliably shows the predicted partial-
 matching interference effects, and that reflexives consistently fail to show sensitivity 
to the number of the interfering NP. Figure 2.13 presents a cross-experiment 
summary of the observed interference effects in total time measures at the critical 
region, demonstrating the clear and reliable contrast between agreement and 
reflexives. Figure 2.14 presents a cross-experiment analysis, combining the 
conditions across experiments to highlight the processing profile evident in total 
reading times at the critical region for agreement and reflexive dependencies. The 
predictions of a structured access account for reflexives are clearly met: there is no 
observable interference for reflexives, whereas it is repeatedly observed for 
agreement. If a feature-based access strategy was pursued for reflexive binding, then 
interference should have been observed as it was in agreement. This extends prior 
findings on reflexives (Nicol 1988; Clifton et al 1999; Sturt 2003a; Xiang et al 2009), 
providing further empirical evidence in favor of structured access in comprehension. 
 
 
88 
 
 
Figure 2.13: Interference effects (in ms) observed for in total time measures across 
Experiments 1-3. Error bars reflect 95% CI by participants. 
 
 Additionally, in both Experiments 2 and 3 no interference was observed when 
the head noun was plural. The processing difficulty associated with anomaly 
detection for agreement and reflexive appeared to be smaller when the head noun is 
plural, though this was not a general finding across measures and regions. This has 
not been observed previously (see, e.g., Pearlmutter et al 1999; Wagers et al 2009), 
and further study is needed to investigate the source and reliability of these effects. 
This finding may be orthogonal to the main goal of the studies, however. It has been 
noted that in production and comprehension, agreement interference is not found in 
Agr: Exp 1 Agr: Exp 2 Ref: Exp 1 Ref: Exp 3
 Summary of interference effects in Experiments 1-3
 In
 te
 rfe
 re
 nc
 e 
ef
 fe
 ct
  (ms
 )
 -2
 00
 -1
 00
 0
 10
 0
 20
 0
 Grammatical
 Ungrammatical
 
 
89 
plural head noun environments, as singular nouns do not generally cause interference 
(Bock & Miller 1991; Pearlmutter at al 1999; Eberhard et al 2005; Wagers et al 
2009). For this reason, the agreement-reflexive contrast observed in the processing of 
singular head noun conditions provides the crucial contrast.  
 
Figure 2.14: Total reading times (ms) at critical region, combining similar conditions 
across Experiments 1-3. Error bars are standard error by participants. 
 
 
 The central conclusion licensed by these studies is that there is a reliable 
difference between the interference profile observed for agreement and reflexives. 
The overall patterns of interference are replicable for agreement and reflexives, and 
the result of the direct comparison in Experiment showed that agreement reliably 
shows more interference than reflexives. This is predicted if structured access is 
Agreement Reflexives
 Cross-experiment total reading times of critical region
 In
 te
 rfe
 re
 nc
 e 
ef
 fe
 ct
  (ms
 )
 0
 20
 0
 40
 0
 60
 0
 80
 0
 10
 00
 
 
90 
employed in resolving the reflexive dependency, as no interference from the 
structurally inaccessible NP should occur in this situation.  
One possible concern with this finding is that it rests on null-effects logic: for 
reflexive dependencies, we did not find any trace of interference, therefore there is 
none. The direct comparison between the interference effects in Experiment 1 shows 
that this concern is not well-founded, as there is a reliable difference between 
agreement and reflexives. A counter-argument to this is that reflexives and agreement 
do differ in the size of the interference effect, but that this does not reflect a 
qualitative difference between the processing of agreement and reflexives. Instead, 
reflexives may simply have a smaller interference effect in these materials, one that 
we simply failed to observe given the power of the current studies. Chapter 3 takes up 
this question in further detail, using an explicit computational model to derive 
predictions about the size of the interference effects. To preview the results, modeling 
results support the claim that the empirical contrast between agreement and reflexives 
presented here does indeed reflect a qualitative difference in the strategies used to 
access linguistic short-term memory in the two types of dependency.  
  
Discussion 
 
The finding that reflexives are processed in a manner consistent with 
structured access mechanisms is compatible with a range of studies that have 
previously been presented. Both Sturt (2003a) and Xiang et al (2009) suggested that 
initial processing of reflexives uses exclusively structural information to index and 
 
 
91 
retrieve its antecedent. To the extent that any influence of interfering nouns was 
observed, it was unreliable, appeared in later measures, and showed the opposite of 
the predicted direction for partial-matching interference effects. That is, rather than 
leading to faster processing, a feature-matched but inaccessible antecedent had the 
effect of making processing more difficult. The contention that reflexives are 
processed in a structured manner thus receives wide empirical support from a range 
of studies. 
 In line with the conclusion of structured access for reflexives, there are a 
number of other studies that have argued that the initial stages of binding an argument 
reflexive are not impacted by structurally inaccessible antecedents. One of the earliest 
and most widely cited set of results is the cross-modal lexical priming evidence 
offered by Nicol (1988; also discussed in Nicol & Swinney 1989). In her experiments, 
she presented sentences as in (3.10) to participants auditorily: 
 
(3.10) The boxer told the skier that the doctor for the team would blame 
himself / him * for the recent injury. 
 
At the time point in the sentence indicated by the asterisk, participants were presented 
with a visual word about which they were asked to register a lexical decision. The 
reaction time for the lexical decision task was measured. Nicol reasoned that if an 
antecedent was reactivated (retrieved) at the point of processing the anaphor, then 
semantic associates of that antecedent should show priming relative to a neutral 
baseline. Thus, in the example given above, if doctor is retrieved at a given probe 
 
 
92 
point, then its semantic associate nurse should show faster lexical decision than an 
unrelated noun (e.g., paper). Nicol?s findings were straightforward. When the 
sentence contained a reflexive (e.g., himself), only semantic associates of the local, 
structurally accessible noun showed priming in the lexical decision task: the size of 
the priming effect for associates of doctor was approximately 104ms, compared to -
 1ms and 11ms to boxer and skier, respectively. Interestingly, when the anaphor was 
the pronoun him instead of the reflexive, a complementary pattern of priming effects 
was seen. That is, priming to associates of boxer and skier was observed (43ms and 
58ms, respectively).  No such effect was seen for associates of the local noun doctor 
(-21ms) in the pronoun condition. Nicol suggested that this was because the familiar 
structural binding conditions of Chomsky (1981) were applied immediately online as 
hard constraints on antecedent consideration. Thus only antecedents that complied 
with Principles A and B were considered at the point of the pronominal element.    
However, some authors have argued that the initial processing of reflexives is 
in fact prone to feature-based interference. One prominent counterexample to the 
generalization of structural fidelity that I have presented is presented by Badecker and 
Straub (2002). These authors criticized the results presented in Nicol (1988), arguing 
that if all antecedents were considered in a parallel, interactive fashion along the lines 
suggested in Gernsbacher (1989) or MacDonald and MacWhinney (1990), the cross-
 modal lexical priming task would be a poor test of access to structurally inappropriate 
antecedents. Badecker and Straub argued that if antecedents are selected according to 
both inhibitory and excitatory activation modulation (as in a parallel, neural-network 
based system), then a lexical decision task that probes only for heightened activation 
 
 
93 
should not be able to determine whether incorrect antecedents are actively inhibited 
as well (Badecker & Straub 2002, p 750). If an antecedent receives some amount of 
activation because of a feature match with a reflexive, but that activation is negated 
by inhibition due to an inaccessible structural position, they argued, then the cross-
 modal lexical priming paradigm would show no difference from baseline. This should 
lead to the incorrect conclusion that this position had not been accessed during 
processing the reflexive, whereas in fact it had been accessed, receiving equal and 
opposite excitation and inhibition. Badecker and Straub thus asked if interference 
from inaccessible antecedents would be observed in a more sensitive measure, such 
as self-paced reading. In the five experiments they presented, they carried out four 
comparisons that involved reflexives or reciprocals, similar to the following 
conditions (drawn from their Experiments 3-6): 
 
 (2.8) a. Jane thought that Bill owed himself a second chance. 
  b. John thought that Bill owed himself a second chance. 
  
(2.9) a. The attorney thought that the judges were telling each other 
which       defendants had appeared as witnesses before. 
 b. The attorneys thought that the judges were telling each other 
which       defendants had appeared as witnesses before. 
 
(2.10) a. Jane thought that Beth?s brother owed himself a second chance. 
 b. Jane thought that Bill?s brother owed himself a second chance. 
 
(2.11) a. It appeared to Jane that Bill owed himself a second chance. 
 b. It appeared to John that Bill owed himself a second chance. 
 
 In Experiments 3 and 4, Badecker and Straub found evidence for what they 
called the multiple-match effect. That is, in (2.8b) and (2.9b), a slowdown was 
observed in the self-paced reading record relative to (2.8a) and (2.9a), respectively. In 
 
 
94 
Experiment 3, this was observed two words downstream from the critical reflexive, 
and in Experiment 4, they observed this effect if they pooled reading times at the four 
regions following the critical reciprocal. They did not observe any difference between 
the conditions shown in (2.10-2.11) (Experiments 5 and 6). Badecker and Straub 
concluded that for reflexives, the binding conditions are not used as an absolute filter 
on candidate generation for reflexive dependencies. They argued that the extra 
reading times observed at the reflexive reflected an extra process of candidate 
suppression when there was an extra, feature-matching (but structurally inaccessible) 
antecedent present in the string. They suggested that although it might seem that this 
is an effect of structural position?only the inaccessible subjects in (2.8) and (2.9) 
caused interference, whereas the structurally less prominent antecedents in (2.10) and 
(2.11) did not?their preferred interpretation for this finding was that the subject 
nouns impacted the resolution process through the discourse prominence normally 
associated with the subject position (Badecker & Straub, 2002, p 764).  
 It is unclear, however that these results provide a counterexample to the 
claims of structured access that I have argued for in presenting Experiments 1-3. 
There are two problems with these results. First, the extent to which interference is 
observed for reflexives in Badecker and Straub?s studies is quite limited, only 
appearing in one out of four experiments. A clear demonstration of the effect is only 
observed in Experiment 3 (2.8a-b), causing a reading time slowdown when there is 
feature-matched but inaccessible antecedents in (2.8b). However, exactly the opposite 
pattern of results was obtained by Sturt (2003a) with structurally identical materials 
(Sturt?s Experiment 1). Sturt found that sentences like (2.8b) were actually read faster 
 
 
95 
than (2.8a) in later reading time measures, an effect that he did not replicate in his 
Experiment 2. Thus, apparent exceptions to the generalization of structural fidelity for 
reflexives do not appear to be reliable across experiments; the bulk of evidence 
suggests that there is no interference from inaccessible antecedents during the initial 
processing of reflexive dependencies. 
 It is also worth noting that in a similar set of studies using self-paced reading, 
Clifton, Frazier and Deevy (1999) failed to find evidence for multiple match effects 
with related materials. In their Experiment 2, they found evidence for a general effect 
of having complex antecedents (i.e. the son of the fireman versus the fireman) on 
reading times after the verb, regardless of whether or not there was a reflexive or a 
referential DP in that position. In their Experiment 3, they used materials very similar 
to Badecker and Straub?s Experiment 3, including the reflexive comparison in (2.8). 
They did not find a significant multiple-match effect, though at the critical reflexive 
region there was a non-significant trend in the direction of a multiple-match effect.  
 A second problem with the Badecker and Straub?s effect concerns the 
direction of the effect, which is important when interpreting the underlying access 
mechanisms. In particular, the agreement interference results that have been provided 
here and elsewhere demonstrate that the clear behavioral signature of incorrectly 
considering an inaccessible antecedent is processing facilitation due to a feature-
 matched antecedent. As I demonstrate in Chapter 3, this is a direct prediction of a 
rational model of memory access. The effect observed by Badecker and Straub 
instead shows processing inhibition when there are multiple nouns that overlap in 
feature content, and so is better understood as multiple-match interference. The 
 
 
96 
opposite direction does not clearly signal that the incorrect antecedent was 
considered, and current models of memory offer many possible interpretations of 
such an effect. Importantly, models of this effect do not entail that the antecedent in 
the incorrect position was incorrectly retrieved during reflexive resolution. One 
possibility is that memory interference or ?forgetting? occurs when memories overlap 
feature content (Oberauer & Kliegl 2006), possibly due to a process of feature-
 overwriting (Nairne 1988, 1990). If this kind of interference deteriorates the quality 
of the memory representations, then the reading time slowdown in (2.8b) might be 
understood as impeded access to a degraded or unreliable memory trace. This could 
occur without any consideration of the inaccessible noun phrase at the point of 
retrieval. The crucial evidence for consideration of spurious, feature-matched 
antecedents is a partial-matching interference effect as seen in interfering agreement 
configurations, and to date this has not been reliably observed in the processing of 
reflexives. 
 
Reflexive interpretation 
  
Although online comprehension evidence is nearly unanimous that reflexives 
are processed in a structured manner, there are interesting divergences with results 
from tasks that tap interpretation, as well as production tasks.   
In a follow up to his Experiment 1, Sturt (2003a) presented a task where 
participants were asked to determine who was the agent of the reflexive action in the 
conditions presented in his Experiment 1 (repeated here as 2.12). After reading small 
 
 
97 
passages such as these, participants were asked about the patient of the action 
described in the second sentence, and given a choice between the two participants in 
the sentence (i.e. Jonathan or the surgeon).   
 
 (2.12) Jonathan was pretty worried at City Hospital. 
   a. He remembered that the surgeon pricked himself with a needle. 
   b. #He remembered that the surgeon picked herself with a needle. 
   Jennifer was pretty worried at City Hospital. 
   c. She remembered that the surgeon pricked himself with a needle. 
   d. #She remembered that the surgeon pricked herself with a needle. 
 
 Sturt found that the percentage of incorrect interpretations of the reflexive 
increased when the inaccessible antecedent (which corresponded to the topic of the 
mini-discourse) matched the features of the reflexive. Thus, in (2.12a) participants 
responded that Jonathan pricked himself with the needle in 17% of trials, and in 
(2.12d) they said that Jennifer pricked herself with a needle 31% of the time. When 
there was no feature-match between the stereotypical gender of the inaccessible NP 
and the reflexive, as in (2.12b-c), incorrect interpretations occurred at much lower 
rates (6% and 9% of trials, respectively). In a follow-up article (Sturt 2003b), Sturt 
suggested that that although the binding principles are applied faithfully in a manner 
predicted by structured access models of memory access, this initial grammatically 
accurate parse does not bleed later application of discourse-level considerations. He 
considers the possibility that the ungrammatical offline interpretations that he found 
for the sentences in Experiment 1 reflect reanalysis in favor of the discourse-central 
antecedent, as might be expected if some discourse-level factors were brought to bear 
on the dependency construction process at a later stage (in line, possibly, with 
 
 
98 
theoretical proposals by Reinhart & Reuland 1993, Pollard & Sag 1992). This 
contention was supported by the fact that this effect was not observed in Experiment 
2, where the interferer no longer held any prominent discourse status.  
 A two-stage account makes testable predictions about the time-course of the 
influence of structurally inaccessible antecedents on processing reflexive 
dependencies. In particular, one should be able to probe reflexive comprehension at 
early and late time points and observe that the influence of inaccessible antecedents 
occurs later than the influence of structurally accessible antecedents. One way to 
accomplish this would be to employ the speed-accuracy tradeoff task (SAT) to 
examine participants? accuracy on an interpretation task at various time lags from the 
presentation of the reflexive. If two processes, structural and discourse-level, are 
involved in producing the interpretation judgment, then the resulting SAT function 
should display non-monotonic growth. Such a non-monotonic function occurs 
whenever processing involves a mixture of processes with differing asymptotic 
accuracies (Ratcliff 1980; McElree & Dosher 1989). If such a two-stage process is 
correct, accuracy should show a decrease once discourse-level considerations 
influence the parsing process, indicating a late influence of inaccessible antecedents 
on comprehension.  
 In addition, if it is truly due to discourse-level considerations, as Sturt (2003b) 
suggests, then manipulating the discourse function of the interfering NP should 
modulate the availability of this effect. Reflexives are known to be sensitive to point-
 of-view effects (Kuno, 1987; Pollard & Sag 1992; for experimental evidence, see 
Runner, Sussman & Tanenhaus 2003; Kaiser 2006). For example, Kratzer (2009) 
 
 
99 
notes that the point of view of an embedded environment modulates the availability 
of bound readings of indexical pronouns in embedded environments. Borrowing this 
insight, one might test whether or not it was simply the point-of-view of the 
embedding environment that led participants to entertain incorrect interpretations in 
Sturt?s Experiment 1b. Thus, think in (2.13a) asserts that the embedded clause takes 
the point-of-view of the thinker; the subject of think is the logophoric center of the 
sentence (a pivot or holder of the viewpoint according to Sells (1987)). No such 
point-of-view shift is required by be unaware in (2.13b), and there is no sense in 
which the subject of be unaware must be the pivot or perspective-holder of the 
embedded clause. If a late influence of the logophoric center drives these interpretive 
results then (2.13a) should show interpretive errors, but (2.13b) should not show 
errors: 
 
 (2.13) a. Jennifer thinks that the surgeon pricked herself with a needle. 
    b. Jennifer is unaware that the surgeon pricked herself with a needle. 
 
 These possibilities merit further study, but for present purposes I note that a 
mixture model of reflexive comprehension of the sort described here would be 
consistent with both the offline judgment results and the claim of structured access in 
comprehension. 
  
Attraction in reflexive production 
 
 The primary empirical claim of this chapter is that agreement and reflexive 
 
 
100 
dependencies are constructed in fundamentally distinct manners in comprehension, 
implicating structured access in comprehension. Whereas comprehension evidence is 
near unanimous on this point, this claim is surprising from the point of view of a 
number of studies that have investigated these dependencies from the point of view of 
production. The results in this domain suggest a tighter link between the processes 
involved in agreement and reflexive feature expression than that implied by the 
comprehension evidence. Most importantly, in addition to interpretive errors, there is 
some evidence from production tasks that reflexives are equally (or more) susceptible 
to attraction effects than agreement. This evidence is important, and requires careful 
consideration in light of the claims advanced by the experiments presented here.   
 Experimental evidence to this effect is presented by Bock, Nicol, and Cutting 
(1999). These authors aimed to further explore the phenomenon of agreement 
attraction in production (e.g. Bock & Miller, 1991), and in particular, they asked 
whether or not the feature transmission processes involved in subject-verb agreement 
and pronominal agreement reflected identical underlying mechanisms. Their answer 
was both ?yes? and ?no?, because of two major findings that suggested similarities and 
differences in the way agreement and pronominal elements (reflexives and tag 
pronouns in their study) were licensed.  
 The experimental paradigm they used was the same employed by Bock and 
Miller (1991) to induce agreement attraction errors in the lab. In this task, participants 
listen over headphones to a preamble to a sentence that contains a target subject noun 
phrase. The participants are then instructed to repeat this preamble and complete the 
sentence in an appropriate fashion. Responses are scored with respect to whether or 
 
 
101 
not the correct agreement morphology is produced. In their experiments on pronouns 
and agreement, Bock and colleagues manipulated two features of the subject noun 
phrase that they cued participants with. All preambles were of the form in (2.14) 
below, where the subject NP is modified by a PP that contains an interfering NP 
position. The first factor they manipulated was the number value of the head noun, 
which was singular, plural, or collective (i.e. notionally plural but grammatically 
singular). They crossed this with the number of the interfering noun position, which 
was either singular or plural. In addition to this within-subjects manipulation, they 
manipulated the type of dependency that participants were asked to produce between 
subjects. For subjects that were in the agreement conditions, the task was identical to 
Bock & Miller (1991). Participants in the reflexive or tag pronoun conditions were 
asked to produce a continuation with an appropriate pronominal form. In order to 
ensure that participants would produce the target pronominals, for these conditions 
participants received explicit instructions that described what a reflexive or tag 
pronoun was. In addition, the verb that participants were to produce was given in the 
preamble. In other words, the task for subjects in reflexive or tag pronoun lists was 
simply to produce the either a plural or singular pronominal form after the preamble 
was played. The verbs were all transitive for the reflexive lists, and intransitive for the 
tag pronoun lists. 
 
 (2.14) The gang leader / gang leaders / gang with the dangerous rival / 
rivals     (armed / vanished) ... 
 
 The first major finding that Bock and colleagues report is that anaphors are 
 
 
102 
significantly more sensitive to the notional number of the subject noun than verbal 
agreement. Thus, for collective head nouns such as choir, participants opted for the 
plural form in verb agreement in approximately 35% of trials, but for reflexives and 
tag pronouns in the same environment, the plural form was chosen 75% of the time. 
The authors conclude that this is a major point of variation in the source of agreement 
features for reflexives and agreement. 
 However, with respect to the attraction effect (the production analog of the 
interference effects in comprehension), the authors found similar effects for both 
anaphoric elements and verbal agreement. The magnitude of the effect sizes (i.e. the 
number of inappropriate plurals that were produced following preambles like the 
gang leader with the dangerous rivals?) was similar across tag pronouns, reflexives, 
and agreement. For agreement dependencies, a singular head noun with a singular 
interfering noun elicited plural agreement in 2% of trials, but when the interfering 
noun was plural, plural forms were produced in 10% of trials. In the non-interfering, 
singular-singular conditions, reflexives and tag pronouns were pronounced in a plural 
form in 4% and 2% of trials, respectively. But as with agreement, the proportion of 
trials with plural forms increased in response to an interfering plural noun; the 
observed rate of plural production was 18% and 17% of trials for reflexives and tag 
pronouns respectively. Furthermore, the attraction effect apparently obtained in the 
same manner across all three dependencies. Namely, as has been noted for subject-
 verb agreement (Eberhard et al 2005), it seemed to be primarily a morphological 
effect, occurring above and beyond any effect of notional number engendered by the 
head noun?s number value. The authors did note that the apparent rate of attraction 
 
 
103 
for pronominal forms used in the study was approximately twice was much as for 
agreement. They suggest that this difference, if reliable, may reflect a ?two-source? 
account of attraction for anaphora, whereby incorrect coreference and agreement 
attraction both contribute to the incorrect number marking on the anaphors. Even so, 
the authors hypothesized that the source of agreement attraction is at its core identical 
for both agreement and anaphoric dependencies, and reflects primarily 
morphosyntactic agreement relations. Any residual differences that are observed 
(namely, the sensitivity to the notional number of the head noun) are accounted for as 
extra processes that are engaged in the case of anaphora. 
 Bock, Eberhard, and Cutting (2004), as well as Bock, Butterfield, Cutler, 
Cutting, Eberhard & Humphreys (2006) replicated these the two basic findings on 
pronominal agreement?increased sensitivity to notional number and canonical 
agreement attraction?with respect to tag pronouns and reflexives, respectively. As in 
Bock et al (1999), pronouns were significantly more sensitive to the notional number 
of the antecedent than was verb agreement. These studies again found similar rates of 
attraction for both pronouns and verb agreement. An important further contribution of 
these studies was the finding that pronominal agreement attraction is primarily 
determined by the morphological number, rather than notional number on the 
interfering noun. That is, in environments with a collective noun in the interfering 
position (i.e. the gang leader with the dangerous group), no increase in attraction was 
observed for anaphoric forms. This finding is surprising in light of the marked 
increase in sensitivity to the notional number of their antecedents that is observed for 
anaphoric forms in general. This was viewed as strong support for the authors? 
 
 
104 
contention that agreement attraction, in both pronominal and agreement forms, 
reflects a primarily morphosyntactic phenomenon. Even for dependencies that are 
demonstrably more dependent on notional number, attraction remains unaffected by 
the notional number of the interfering noun.  
 These results were summarized in Eberhard, Cutting, and Bock (2005), and 
formalized using a simple linear model that the authors constructed to predict rates of 
plural form production in the context of their experiments. The model states that the 
probability of producing a plural on any given agreeing form is a function of a linear 
combination of the morphological and notional number on all of the nouns contained 
in the head noun, a value that they refer to as the SAP value (Singular and Plural): 
 
 (2.15) = + ! ( !! ) 
  
 S(r) refers to a linear predictor that, when transformed using an inverse logistic 
function, predicts the probability that a plural form will be used for a dependent 
element in any given context. S(n) refers to a contribution from notional number, a 
continuously valued term that varies from unambiguously plural to unambiguously 
singular. The second term refers to the contribution from morphological number 
marking S(m) on each of the j nominal elements contained in the subject noun phrase, 
each of which has its own weight wj that is inversely correlated to the depth of 
embedding in the noun phrase.  
 The authors interpret the SAP predictor as feature transmission processes that 
are present on all trials, providing a real-valued number for the subject noun phrase. 
 
 
105 
The impressive fits of their mathematical model itself do not directly confirm this 
claim, however. The model itself is entirely compatible with alternative views of the 
agreement attraction process, such as the interference-based view adopted here and 
elsewhere (Badecker & Kuminiak 2007; Wagers et al 2009). The successful model 
fits simply reflect the fact that notional and morphological number of the subject 
NP?s constituent elements predict, in a linear fashion, the observed amount of plural 
marking on dependent elements. This view is more or less shared by a variety of 
different accounts, though the authors do note that their model differs crucially from 
that offered by Vigliocco and colleagues (Vigliocco & Frank 1998, 2001; Vigliocco 
& Hartsuiker 2002), who argued for a greater and more direct influence of notional 
number on the agreement resolution process.  
 The mathematical formalization provided by Eberhard et al (2005) clarifies 
where agreement and pronouns part ways in production. Recall that in discussing the 
agreement-pronoun contrast, Bock and colleagues (1999) argued that agreement and 
pronoun attraction reflect the same underlying process, and that to the extent that 
differences between agreement and pronouns are observed, they are due to extra 
processes that are engaged for pronouns. In this model, the exact difference between 
reflexives and agreement is clear. The crucial difference, the authors argue, is the fact 
that pronouns are lexical items in their own right, with their own set of features that 
influence the calculation of the SAP values. Thus, the calculation of the SAP linear 
predictor for the pronominal elements is modified as follows: 
 
 (2.16) !"# = 2 + ! ( !! )+  ( !"#) 
 
 
106 
 
 The formulation of the linear predictor for pronouns S(rpro) shown above 
represents a rearrangement of the terms present in the formula for the SAP value for 
pronouns given in Eberhard et al (2005, pp. 545), intended to highlight the unity in 
the process that the authors envision between agreement processes in reflexives and 
in pronouns. Pronouns reflect the SAP value of their antecedent in a manner that is 
fundamentally identical to subject-verb agreement with i) a hard constraint that there 
is a twice as large an influence of notional number for pronouns, compared to 
agreement and ii) an extra, weighted boost from the morphological features of the 
pronominal (S(mpro)). In concrete terms, the authors assume that this additional effect 
reflects the fact that in production, the notional number of a referent (at the message 
level) feeds the selection of a lexical item for the pronoun, which leads to a greater 
role for notional number in the marking process. Taken together, this SAP value 
affects later reconciliation processes that may alter the morphological features of the 
pronominal that was selected. It is this later step of reconciliation in which agreement 
attraction occurs for pronominal elements, as in agreement. Thus in the final 
formulation of the model, attraction or interference effects have an identical source 
for both agreement and pronominal elements, but the flow of information in 
production for pronominals differs in a way that increases their sensitivity to the 
notional number of the correct antecedent only. 
 Eberhard and colleagues refer to the differences between agreement and 
reflexives by referring to them with different terms: since verb agreement arises in the 
morphological marking stage, under the influence of syntax, it is structural control (p. 
 
 
107 
22). On the other hand, since the lexicon is assumed to bear more of the burden in 
determining the observed number marking in anaphora, they call the task of selecting 
lexical items to reflect similar message-level meanings in coreferential elements 
lexical concord. Though these reflect constraints on the expression of number 
marking at distinct levels of production, the authors retain the use of the superordinate 
term agreement to refer to both types of processes.  
 Importantly, even in production, it cannot be the case that agreement on verbs 
and pronominal elements are controlled by the same SAP value. Although reflexives 
and agreement appear to pattern similarly with respect to attraction in production, the 
differential sensitivity of anaphora to notional number suggests an additional process 
or source of number marking specific to pronominal elements. Bock and colleagues 
have suggested that this is due to an effect of lexical retrieval of the pronoun, as well 
as the effect of message level meaning on this selection. However, the fact still 
remains that comprehension and production show apparently very different 
interference profiles for reflexives. In comprehension, the sensitivity to attraction 
varies between the two, and in production, the sensitivity to notional number varies 
between the two. An account of the differences between agreement and reflexives 
should be able to account for these two basic differences. 
   
The difference between agreement and reflexives 
 
In comprehension, reflexive and agreement diverge sharply with respect to 
interference effects, and in production, notional number plays a greater role in the 
 
 
108 
determination of reflexive number. Nonetheless the empirical results do point to a 
limited role for shared processes across the two phenomena: both susceptible to 
superficial agreement attraction in production. A fully specified processing model of 
agreement and reflexive dependencies must account for the fundamental differences 
in the two dependencies (the reliance on notional number), as well as the apparent 
similarities (attraction effects in production).  
One hypothesis for the difference has been proposed by Bock and colleagues 
(Bock et al 1999; Eberhard et al 2005), who maintain that two features distinguish the 
processing of reflexives from the processing of agreement: first, reflexives are lexical 
items in their own right and second, that their role as referring elements causes their 
antecedent?s semantic representation to play a greater role in the dependency 
construction process. This accounts for the greater impact of notional number on 
agreement, but not the attraction effect for agreement and reflexives in 
comprehension. On this model, the attraction effects for both dependencies stem from 
a single, post-syntactic reconciliation process in the Eberhard et al (2005) model. 
However, an alternative hypothesis is that the difference between reflexives and 
agreement is that there are two dissociable processes engaged by reflexive elements: a 
process of binding chain construction and a process of feature concord. The 
separation of feature concord and the core dependency formation, combined with 
differences in information flow in production and comprehension, may be sufficient 
to understand the production-comprehension asymmetry. 
 The major claim advanced by Eberhard et al (2005) is that the process 
difference between agreement and anaphora is that for anaphoric elements, the 
 
 
109 
message-level representation of the antecedent directly guides selection of the 
pronominal from the lexicon and its notional meaning, which in turn affects the 
calculation of the SAP value. Although intuitively appealing, there are a number of 
concerns with this account of the differences. Most importantly, it is unlikely that the 
presence or absence of a lexical item to be selected is the crucial factor that divides 
pronouns and agreement. An agreeing form?s sensitivity to notional number and 
grammatical number does not reflect whether or not an item has lexical features of its 
own to contribute. Corbett notes this in his discussion of the agreement hierarchy, an 
implicational hierarchy that describes the relative likelihood of semantic (notional) 
versus syntactic (grammatical) agreement in any given language (Corbett, 2006). 
Thus although in British English nouns that routinely control plural agreement due to 
their notional number (The committee are going to decide) do not command plural 
agreement on demonstrative determiners (This committee; cf. *These committee, 
Corbett 2006). If one makes the plausible assumption that determiners are selected 
from the lexicon, then these facts are unexpected. That is, the Eberhard et al (2005) 
model would predict that subject-verb agreement should show less reliance on 
notional number, whereas there should be greater reliance on notional number for 
determiners; this is exactly the opposite of what is observed.  
In addition, it is unclear that bound pronouns reflect pre-structural lexical 
retrieval more than subject-verb agreement. First, lexical retrieval is very likely 
involved for suppletive verb forms such as was and were, and many of the results in 
comprehension involved exactly these verbs. Second, it is unclear that reflexive 
pronouns are in fact retrieved from the lexicon prior to construction of syntactic 
 
 
110 
structure. The idea that syntactic control determines a bound pronominal?s 
morphological shape is at the heart of many approaches to anaphora (Lees & Klima 
1963; Lidz & Idsardi 1999; Kratzer 2009), and from the point of view of production, 
this assumption seems natural. The model advocated by Bock and colleagues suggests 
that the lexical boost for reflexives comes from a production stage where the anaphor 
has been retrieved from the lexicon, but structure has not been created. However, the 
selection of the correct referring expression depends on the structural relation it bears 
to its antecedent, and so it would seem that lexically retrieving the pronoun prior to 
structure generation in production is an inappropriate sequence of operations. On this 
model, there is no guarantee that the retrieved pronoun would be the correct form 
given its eventual structural relation to its antecedent. If instead the correct 
pronominal form is retrieved after the creation of structure, producers could more 
easily select the appropriate anaphor. This is essentially the model advocated by Bock 
and colleagues for subject-verb agreement in English, and it is not obvious that 
anaphor selection is any different. 
As noted above, the lexical differences approach captures the greater effect of 
notional number on reflexives in production. However, in Eberhard and colleagues? 
model (2005) the same post-structural reconciliation process is invoked in both 
agreement and reflexives to account for the attraction findings. Given that the 
inherent lexical differences model addresses the more direct influence of notional 
number of pronominal agreement, It is not clear that the different comprehension and 
production with respect to interference can be made to follow from this model. 
However, if the relevant difference between agreement and reflexives is that the 
 
 
111 
binding dependency between a reflexive and its antecedents and the feature concord 
between the two forms must be modeled as distinct processes, then the production-
 comprehension asymmetry may be a function of differences in the flow of 
information for these two modes of language use.   
A common feature of contemporary theoretical accounts of reflexive pronouns 
is a separation between the process of building the binding dependency and the 
process of ensuring feature concord between antecedent and anaphor are distinct 
(Pollard & Sag 1992; Reinhart & Reuland 1993; Lidz & Idsardi 1999; B?ring 2005; 
Hornstein 2007; Kratzer 2009). This is a natural theoretical move, as the semantic 
representation of reflexives may in fact be devoid of feature content. The intuition is 
that the core dependency to be accounted for by a proper theory of reflexive anaphors 
is the mapping from a surface form such as (2.19a) to an interpretation such as 
(2.19b) (Ross 1967; Sag 1976; Hankamer & Sag 1976; Sag & Hankamer 1984; 
Fiengo & May 1994). 
 
(2.19)  a. Only John hurt himself. 
  b. ?Only John? ?x . x likes x 
 
 The representation of the dependency as a featureless bound variable ensures 
the proper interpretation (that the two arguments of the predicate obligatorily covary 
in reference), and accounts for the semantically inert nature of the morphological 
features. Theoretical accounts that separate the process of feature concord from the 
process of anaphoric interpretation are well positioned to account for the difference 
between agreement and reflexives in both production and comprehension.  
 
 
112 
 Once feature concord and binding relations are distinguished, the production-
 comprehension difference may simply reflect differences in the flow of information 
with respect to morphological features: in comprehension, the reflexive?s features 
need to be checked, and in production, those features need to be assigned. The 
attraction in production reported by Bock et al (1999) may reflect an error in a 
feature-assignment stage, which is entirely absent in comprehension. The post-
 syntactic reconciliation process in the Eberhard et al (2005) model suggests exactly 
this possibility. Furthermore, if the construction of the binding dependency precedes 
the selection of the morphological features in the feature-assignment stage, then the 
increased sensitivity to notional number may reflect the semantic value of the binding 
dependency. In contrast, there is no comparable feature-assignment stage that occurs 
in comprehension; instead, once the binding dependency is constructed in 
comprehension, the features on the incoming reflexive need to be verified. Thus it is 
possible that in both production and comprehension, the binding process occurs prior 
to the feature concord process. The different nature of the feature concord process in 
comprehension and production could in fact reduce to a difference in the fidelity of 
feature assignment versus feature checking. 
 The role of the reflexive?s morphological features in comprehension on this 
account is unclear. Although they are apparently not used in memory access, it is 
clear that the antecedent-anaphor feature mismatch is evident to comprehenders early 
(modulo the results on processing plural subjects in Experiment 3). Comprehenders 
evidently assess the feature concord rapidly online, but the time course of this feature 
check relative to the construction of the binding dependency is unclear. Likewise, the 
 
 
113 
main explanatory aim of the lexical differences model in Eberhard et al (2005) was 
the increased sensitivity to notional number for pronouns. This does not automatically 
follow from the separation of binding and feature concord processes. It is possible 
that the process of creating a referential dependency between the anaphor and its 
antecedent prior to the feature assignment stage in production drives the increased 
sensitivity to notional number (perhaps by temporarily reactivating the referent?s 
interpretive content), but there is no direct evidence for this conjecture.  
Determining which account is a better model of the difference between 
reflexives and agreement is an important goal for future research. While I have 
provided a sketch of how attraction might differ for reflexives in comprehension and 
production, the source of the increased sensitivity to notional number is not obvious. 
Although important questions about the production-comprehension link remain, the 
current data demonstrate that reflexives engage in structured access in comprehension. 
Even in comprehension, however, important questions remain unanswered. In 
particular, the scope of structured access in comprehension remains unknown. The 
conditions under which structured access is deployed are taken up in detail in Chapter 
5. 
 
Conclusion 
 
 The primary goal of this chapter was to determine whether or not subject-verb 
agreement and reflexive dependencies showed identical interference profiles in 
English. It was shown that in a tightly controlled comparison, there are systematic 
 
 
114 
differences in the role that morphological features play in the agreement and reflexive 
dependencies.  
 For agreement, the current data replicated and extended the partial-matching 
interference effects previously observed in comprehension (Pearlmutter et al 1999; 
Wagers et al 2009). This finding supported the predictions of a content-addressable 
architecture for sentence comprehension, showing that the morphological feature 
content of noun phrases in a parse is used as a cue to gate memory access in 
constructing the agreement dependency online.  
 For reflexives, however, an altogether different profile was observed. There 
was no evidence in any experiment that the morphological features of the reflexive 
were used in memory access. This conclusion rests on the repeated failure to find the 
partial-matching interference effect that was observed with agreement dependencies, 
which provides clear evidence of feature-based memory access. Instead, the available 
evidence suggested that reflexives are uniquely sensitive to the feature match with the 
structurally accessible antecedent, implicating a memory access procedure that 
leverages structural position, rather than feature match, in building the binding 
dependency. 
I suggested that the process of structured access in comprehension tracks the 
construction of a binding dependency, rather than the fact that reflexives have 
interpretive or lexical content (as in the Eberhard et al 2005 model). On this view, 
structural binding is the process that gives rise to the profile of structured access in 
comprehension.   
 
 
115 
Before a full evaluation of this hypothesis, its predictions, and its fit with the 
current data, it is important to provide a firmer basis for the claim that reflexives 
access their antecedents in a fundamentally structured fashion in comprehension. One 
remaining empirical concern with the data presented in Experiments 1-3 is the 
possibility that the qualitatively interference profiles for reflexives and agreement 
actually reflect underlying quantitative differences in what is actually the same 
memory access procedure across the two dependencies. If this objection were true, it 
would undermine the case for structured access from the current data. Chapter 3 will 
focus on formalizing a model of memory access in order to critically evaluate this 
possibility, and clarify the relation of the present findings to previous research on 
interference in sentence processing.
  
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
116 
Chapter 3: Revisiting the interference argument: optimal information retrieval 
 
 
 
 
 
 
 
 
 
 In Chapter 2, I argued that the interference profile observed for reflexives 
implicated a structured access mechanism: when comprehenders retrieve the 
reflexive?s antecedent from memory, it is primarily syntactic information that 
controls memory access. This argument is to some degree independent of specific 
architectural commitments: it is built on the observation that the behavioral signature 
that characterizes feature-based access for subject-verb agreement is consistently 
absent for reflexives. Instead, reflexives show sensitivity only to the feature content 
of structurally accessible antecedents. This suggests that structural position, rather 
than feature match, gates memory access for reflexives, a conclusion that is 
compatible with a wide range of architectural assumptions. The primary goal of this 
chapter is to provide an alternative computational argument in favor of structured 
access. By providing a simple but explicit model of structured memory access in 
reflexive dependencies, I show that structured access models capture the data better 
than feature-based models.  
 
 
117 
 Although the claim of structured access is in principle independent of 
particular architectural commitments, in the present chapter I present an explicit 
model of structured memory access that builds on existing models of memory 
architecture in sentence comprehension (Lewis et al 2006). Structured access can be 
straightforwardly implemented in a content-addressable architecture, if the retrieval 
cues are sufficiently rich to selectively access the local subject position. I demonstrate 
this with a simple structured access model based on the ACT-R architecture 
(Anderson 1989; Anderson & Milson 1989; Lewis & Vasishth 2005; Lewis et al 
2006). Providing an explicit model serves two goals. First, it provides a framework 
for understanding the relation of current results to previous work on interference 
effects in comprehension. This is important as it clarifies what conclusions are 
justified from previous work on interference effects, and makes concrete predictions 
about what data are necessary to falsify the claim of structured access that I have 
made. In addition, the explicit behavioral predictions derived from the model address 
some residual empirical concerns with Experiments 1-3. The results of this modeling 
solidify the case for structured access in comprehension by showing that the reflexive 
non-interference effect cannot be reduced to superficial differences between 
agreement and reflexives, such as linear or hierarchical position relative to the 
antecedent.  
 I start by presenting an abstract characterization of the information retrieval 
problem in sentence comprehension that draws on the seminal work by John 
Anderson (Anderson & Milson, 1989). Casting the problem as one of rational 
memory access allows the application of very general principles of optimal behavior 
 
 
118 
and rational analysis in reasoning about memory access. This abstract 
characterization is at the heart of the cognitive architecture ACT-R (Anderson 1990), 
which characterizes cognition, in part, as an optimal statistical adaption to the 
structure of the environment (Anderson 1990; Anderson & Lebiere 1998). I recruit 
this architecture as an explicit model of memory access in sentence comprehension 
(following Lewis & Vasishth 2005). With the principles of rational memory access 
and an explicit model of memory access in comprehension, I present a model of the 
agreement and reflexive data in Experiments 1-3. Using the model, I directly compare 
the predictions of feature-based and structured access models of reflexive dependency 
formation. Results show that the structured access model predicts the observed data 
significantly better than feature-based access models, providing an alternative 
argument in favor of structured access in comprehension.  
Lastly, I contrast findings that have been argued to reflect interference effects 
with the predictions of the memory model assumed here and elsewhere (Lewis & 
Vasishth 2005, Lewis et al 2006). Patterns of inhibition and patterns of facilitation 
have both been attributed to interference effects (compare claims in Van Dyke & 
McElree 2006 with those in Vasishth et al 2008), with both taken to be indicative of 
online access to incorrect items in memory. However, formal analysis and simulation 
evidence show that inhibition and facilitation have distinct mechanistic sources. This 
analysis suggests that only patterns of facilitation, as seen in cases of partial-matching 
interference, provide evidence in favor of incorrect access online. This finding 
suggests that structurally incorrect access may occur less frequently in comprehension 
 
 
119 
than is generally assumed, and makes a clear prediction about the data that is needed 
to falsify claims of structured access for reflexives. 
 
Rational memory access 
 
Memory access procedures play an important role in parsing long-distance 
dependencies. For this reason, it is important to give careful consideration to how 
memory access might proceed in the ideal case. Note that this is likely true even if 
there is a substantial role for prospective structure-building processes (Lau, 2009). 
The architectural assumption that ensures a substantial role for memory retrieval is a 
limited focus of attention (McElree 2000; McElree, Foraker, & Dyer, 2003; Lewis & 
Vasishth 2005; Lewis et al 2006, Wagers 2008). In adopting a limited focus of 
attention, we make the claim that not all information in memory can be maintained in 
an active, accessible state in parallel; when information is not in an active state, it 
needs to be retrieved or reactivated in order to participate in further processing. This 
assumption allows the parsing process to be cast, in part, as a series of retrievals that 
shunt information into and out of an active processing state that is referred to as focal 
attention (McElree et al 2003; Lewis & Vasishth 2005; Lewis et al 2006; Wagers 
2008).  
On this view, memory retrieval is an important informational bottleneck in 
sentence processing, and so a formal consideration of the properties of that access 
procedure provides potentially useful insight into the operation of the parser. In the 
context of models such as the ACT-R model of sentence processing (Lewis & 
Vasishth 2005), it is arguably the memory access mechanism that controls the most 
 
 
120 
interesting and novel predictions of the architecture (see also Van Dyke 2007). For 
this reason, in the discussion that follows I abstract away from model details that are 
arguably orthogonal to the issues at hand, such as the role of lexical access or the 
possibility of specialized buffers for dealing with particular parsing processes (a 
number of these issues are discussed in Lewis & Vasishth 2005).  
An important milestone in understanding the properties of memory access is 
Anderson and Milson?s (1989) observation that the behavior that characterizes 
memory access might be understood as rational, in the sense of performing optimally 
given the constraints on the system (see also Anderson 1990). Anderson and Milson 
cast the central problem of memory modeling as one of information retrieval, which 
has well-understood methods of optimization. This abstract approach allowed them to 
formalize how the memory access should behave in the best case, without making any 
commitments to the actual mechanisms that implement rational memory access. To 
borrow terminology familiar from linguistic theorizing, this might be understood as a 
?competence?-level characterization of the access procedure, abstracting away from 
the algorithms that implement memory access in performance in order to gain insight 
into how the system behaves. This is an important abstraction that makes conclusions 
based on this reasoning very general; this analysis is compatible with any number of 
mechanistic implementations that might achieve the same extensional results.    
At this abstract level we might first consider one statement of optimal 
behavior in the task of information retrieval under formal decision theory. If the task 
is to retrieve the target memory given some evidence (e.g. a set of retrieval cues), 
then the optimal decision rule is simple. Under these conditions, the Bayes optimal 
 
 
121 
decision?the behavior that will produce the least error in the limit?is to choose the 
memory m that has the highest probability given some evidence E: 
 
(3.1)  argmax!  
 
 This reflects a posterior belief over the probability of a given memory being 
the target, given some retrieval cues. Intuitively, this statement says that for any given 
retrieval cues, the optimal behavior for a memory system is to retrieve the memory 
best fits those cues. Put differently, it is not a good idea to retrieve memories that are 
a poor fit to your retrieval cues. This formalization is useful, as we can decompose 
this statement of optimal behavior to further gain insight into the system. This 
posterior belief about the best memory given some cues can be broken down using 
Bayes? rule into likelihood function of the evidence (i.e., the probability that a 
particular set of retrieval cues would be used if m was the target memory), and a prior 
belief about the likelihood of a memory m. Note that the denominator, the probability 
of the evidence, p(E), is constant over m and so not relevant for finding the most 
likely memory: 
 
(3.2)  ?  
 
 The assumption that memory access should be rational provides the starting 
point for Anderson?s theorizing about memory: if memory access is rational, then 
behavioral indices of memory retrieval should be monotonically related to p(m|E) 
 
 
122 
(Anderson & Milson 1989; Anderson 1990). Anderson and Milson (1989) derive an 
explicit linking between this high-level characterization of Bayesian optimal behavior 
to an information retrieval characterization of memory access, and they show that this 
simple characterization of the problem predicts the existence of frequency effects, 
recency effects, fan effects, and a number of other observed working memory 
phenomena. Bayesian optimal performance?fastest access or recall latencies, highest 
task performance?is achieved by always retrieving the memory with the highest 
posterior probability, though this abstract analysis of the system?s behavior is silent 
with respect to the actual mechanisms that implement this behavior. On the 
assumption that behavioral measures are directly related to p(m|E), these widely 
documented memory access effects may be understood as reflecting optimal 
adaptation to the statistical structure of the environment. In other words, people 
optimize their information retrieval routines to do the best they can given cognitive 
constraints, and the variety of empirical phenomena observed in working memory 
tasks reflect this simple fact. An important feature of this characterization is that it is 
independent of implementation, as it only makes the assumption that the goal of 
memory access is to get the ?right? memory for some search query. As such, the 
above characterization is compatible with a wide range of assumptions about the fine 
structure of memory.  
 For example, consider the basic fan effect in list memory (Anderson 1974). 
The fan effect refers to the phenomenon whereby the more facts are studied in 
relation to a concept, the slower participants are to retrieve any one fact about that 
concept. If one adopts a rational model of memory access, there is an easy 
 
 
123 
explanation for this effect. Informally, this is because the more memories a given cue 
is associated with, the less diagnostic that cue is: the more red objects are represented 
in memory, the less helpful a query such as ?retrieve the red object? is. The less 
diagnostic a cue is, the less likely it is to be used given a desire to access some 
memory, lowering p(E|m) above. This translates into slower reaction times for items 
that have greater fan, given the crucial linking assumption that time to access a 
memory is inversely related to the posterior probability of that memory given some 
retrieval cues.  
 This abstract characterization of the memory access problem is extremely 
general, and it provides a powerful tool for reasoning about how access should 
proceed at a high level. Using Marr?s terminology, the description of memory access 
as Bayesian optimal is an example of a computational level characterization of a 
cognitive problem (Marr, 1982). That is, it provides a general statement of the goals 
of how cognition should proceed, but it does not in itself specify the algorithms or 
processing routines that guarantee optimal function in memory access. This is an 
important point, as the comparisons that the analyst performs in determining what 
behavior would be ?rational? given a set of constraints and a goal appear to have an 
algorithmic character: we calculate the probability of a set of memories for some set 
of retrieval cues, and then select the most probable memory. However it is important 
to bear in mind that this comparison is an analytic tool rather than a claim about 
actual access algorithms. There are, however, explicit algorithmic-level descriptions 
of optimal memory access that have been developed.  
 
 
 
124 
Implementing rational retrieval: the model 
 
 The computational characterization of the memory access problem described 
above forms the basis of the ACT-R cognitive architecture (Anderson & Lebiere 
1998). The memory retrieval process in ACT-R reflects one particular 
implementation of a rational model of memory access (Anderson 1990; Anderson and 
Schooler 1991; Lewis & Vasishth 2005). The essential idea of this model that each 
memory image contained in the memory store (a chunk) is associated with a 
hypothesized activation level A that indexes its availability. This activation evaluated 
at the point of retrieval, and is intended to track the log probability that a given chunk 
is going to be needed. It is a function of a given chunk?s retrieval history and the 
current context (i.e. the current set of retrieval cues). In other words, activation is a 
function of a posterior probability (activation prior to retrieval) and a likelihood (fit 
with retrieval cues): thus the activation A for a given memory m is proportional to 
log(p(m|E)), the log of the posterior probability of the memory given the search cues  
With this understanding, the link between the abstract, formal statement of the 
optimal model of memory access and the ACT-R implementation is easier to see. The 
final activation of a chunk in ACT-R is calculated by summing terms that correspond 
to the prior (p(m)) and the likelihood (p(E|m)) in the statement of optimal memory 
access, as shown in (3.3). The history of retrievals reflects the prior probability, as 
this indexes the past usefulness of a given memory. Two terms are related to the 
likelihood of the evidence given a memory: an associative boost for retrieval cues 
match the information contained in the memory, and a penalty for retrieval cues that 
 
 
125 
are not matched by information in the chunk. Together, these implement the 
likelihood p(E|m). 
 
 (3.3)  ! = ! + ! !"! + !"! +  
 
 The term that corresponds to the prior, B, reflects the history of prior 
retrievals. It is taken by summing the time elapsed since the mth retrieval of chunk i, 
over all n retrievals of that chunk. The time elapsed since a prior retrieval is weighted 
by a decay factor d (3.4). This implements the intuition that the more times a memory 
has been accessed, the more likely it is to be accessed again in the future. 
  
 (3.4)  ! = !!!!  
 
 In addition to base activation, the activation level for a given memory at the 
point of a retrieval operation is a function of the degree to which it conforms to the 
retrieval cues. Intuitively, this is because the better the retrieval cues match any given 
memory, the more likely that memory is to be the target. There are two terms that 
accomplish this, and they correspond to the usefulness of the cues used (more 
distinctive is better) and the mismatch between features in the cues and features in the 
target (all other things being equal, it is better to avoid using search cues that 
mismatch the target?s features). The first determines the strength of association of the 
retrieval cues to the information contained in chunk i. This is done by multiplying 
each of the J retrieval cue?s weights (Wj) by the strength of association for that cue on 
 
 
126 
chunk i (Sji). Sji is reduced by the fan of a cue. As noted above, fan is related to p(E|m) 
(Anderson & Reder 1999). In other words, the more chunks in memory contain that 
cue, the greater the fan, and the lower the value of Sji. Intuitively, this is because a cue 
that has a high fan does a poor job of discriminating the target memory chunk from 
distractors, and thus should be down-weighted in trying to access memory in the most 
effective manner possible. It is calculated in the following manner: 
 
  (3.5)  !" = ? ln ( !) 
 
 Where S is a free parameter of the model (the maximum strength of a cue). 
Fanj is simply the number of items in memory that contain a given cue. In addition to 
the associative strength between the retrieval cues and the information contained in a 
memory chunk, there is also a term that penalizes the activation of a given chunk if 
any of the retrieval cues are not matched by information on the chunk. P is a 
negatively-valued free parameter in the model, known as the mismatch penalty. Mki is 
a logical value that is 1 for each of the retrieval cues that is not matched by 
information in the chunk, and 0 otherwise. Thus the total mismatch penalty is P times 
the number of cues that are not matched by the information in chunk i. 
 In addition, there is assumed to be a certain amount of stochastic noise in the 
system, , which is distributed according to a logarithmic distribution with a mean of 
0 and a variance that is a function of the ?noise? parameter s: 
  
  (3.6)  ! =
 !!
 !
 ! 
 
 
127 
 
 At the point of a retrieval, ACT-R implements rational memory access in the 
following fashion. The activation value for every memory chunk is calculated with 
respect to the search cues being used, as well as each memory chunk?s baseline 
activation, which tracks each chunk?s previous retrieval history. The memory chunk 
that has the greatest activation (equivalent to the greatest posterior probability of 
being the target memory) is the memory that is retrieved. This follows from assuming 
an optimal decision metric, and can be understood as a ?winner-take-all? situation. In 
this way, memory retrieval operates in a manner similar to that endorsed by race 
models of syntactic comprehension, which posit that in syntactic disambiguation, the 
fastest available syntactic analysis is the one that is adopted (McRoy & Hirst 1990; 
Traxler, Pickering & Clifton 1998; Van Gompel, Traxler & Pickering 2001). This is 
because the most active chunk has the shortest retrieval latency, which is calculated 
by exponentiating the activation valu, and multiplying by a latency scaling factor F: 
 
  (3.7)  ! = !!! 
  
 From (3.7) it can be seen that greater activations translate to faster access. The 
direct relation between activation (probability of retrieval in context) and retrieval 
latency ensures that the fastest memory to be retrieved is guaranteed to be the most 
likely on this model. The race aspect of the ACT-R model is important in making 
sense of the model predictions, and I return to the relationship between race models 
 
 
128 
and retrieval latencies in ACT-R below (the link is also noted in Lewis & Vasishth 
2005 p 399). 
 By adopting an explicit model of memory retrieval, it is possible to compare 
quantitative model predictions to the results from Experiments 1-3. By linking model 
predictions to experimental results, firmer conclusions about the underlying system 
may be drawn. Before presenting the results of this modeling, however, it is important 
to reiterate the crucial linking assumption between retrieval latencies and online 
reaction time measures made here and in other applications of this model to sentence 
processing (Lewis & Vasishth 2005; Vasishth et al 2008). The crucial linking 
hypothesis is that the behavioral measures that are thought to index retrieval 
operations in experimental settings are inversely and monotonically related to p(m|E). 
In other words, faster retrieval should result in faster reading times, and slower 
retrieval should result in slower reading times. Although this is a useful starting 
assumption, it need not be the case: there are any number of processes that might 
obscure the relationship between retrieval time and observed reading time, especially 
in the context of sentence processing. When predicted retrieval times are used to 
model behavioral measures such as reaction time, this implicitly assumes that no 
interpretive processes, reanalysis procedures, or error signals are indexed in the 
behavioral measure. This seems unlikely, as retrieving a memory in sentence 
comprehension is presumably carried out in the service of enabling further 
processing. It is almost certain that these processes are also reflected in reading time 
measures. However, for purposes of the current discussion, I assume with others 
(Lewis & Vasishth 2005; Lewis et al 2006; Vasishth et al 2008) that whatever extra 
 
 
129 
processes are engaged as the result of a retrieval, they do not disrupt the monotonic 
relation between memory retrieval times and online reading time measures.  
  
Modeling reflexive and agreement dependencies 
 
The relation between model predictions and experimental findings 
 
The ACT-R model described above can be used to provide quantitative 
predictions about the hypothesized retrieval processes that occur during the 
processing of long-distance dependencies. The explicit predictions derived from the 
model above have the potential to be informative in understanding the model results 
presented in Chapter 2. One possible concern with the results presented in Chapter 2 
is the assumption that the interference effect sizes for agreement and reflexives 
should be equal. This assumption is implicitly made in contrasting the presence 
versus absence of interference effects in the within-subjects comparison of agreement 
and reflexives in Experiment 1. The logic of this comparison is the following: we 
observed interference in agreement, so the experiment has the power to detect the 
interference effect. Therefore, the failure to observe it for reflexives means it is not 
there. This logic is reasonable if the interference effects for both dependencies are of 
comparable magnitude. 
However, it is not clear that the assumption of equal-sized effects is valid, 
which risks undermining this logic and weakening the conclusion. In particular, there 
are unavoidable differences between agreement and reflexives with respect to their 
 
 
130 
linear position in the string, and the parsing processes that precede the critical subject 
retrieval process for agreement and reflexives. For example, consider Figures 3.1 and 
3.2, which are derived from an ACT-R model of the retrievals involved in parsing 
agreement or reflexive dependencies, respectively. These show an average trace of 
the activation of the target and distractor NPs across the parse, leading up to the 
critical retrieval. One crucial difference between agreement and reflexives is evident: 
for reflexives, but not agreement, the target memory (i.e. the local subject) is 
reactivated just prior to the critical retrieval, giving the local subject a large activation 
advantage over the interfering NP at the point of the critical reflexive retrieval. This 
fact leads to a plausible alternative explanation of the experimental results in Chapter 
2: the lack of interference effects for reflexives may simply reflect differences in 
passive memory dynamics that are unrelated to the construction of the antecedent-
 anaphor dependency (the baseline differences hypothesis). If there is a baseline bias 
in favor of the local noun immediately after processing the matrix verb, this may 
selectively eliminate interference effects for reflexive dependencies. If the baseline 
differences hypothesis is true, then the results presented in Chapter 2 may not in fact 
provide evidence for structured access. 
 
 
131 
 
Figure 3.1: Average activation for target (black) and distractor (red) NPs for a 
sentence that shows partial-match interference at the agreeing verb. Incorrect 
retrievals of the distractor NP are reflected in the increased activation at the plural 
verb were. 
 
It is difficult to assess the impact of that this baseline bias should exert on any 
reflexive interference effects without an explicit model, and for this reason the 
quantitative fits obtained by modeling provide useful insight on the experimental 
results. There are two key questions that the models may be used to answer. The first 
is whether or not the predicted size of the interference effect for reflexives is 
demonstrably smaller than that for agreement. If this is true, then a second, more 
critical question is raised. Given the observed interference effects in Chapter 2, is 
more likely that they reflect a feature-based model or a structured access model of 
reflexives? This second question is critical for the main claim of this thesis. If a 
0 500 1000 1500 2000 2500 3000
 0.
 0
 0.
 5
 1.
 0
 1.
 5
 2.
 0
 2.
 5
 3.
 0
 3.
 5
 Time
 Act
 ivatio
 n
 Head NP
 Distractor NP
 *The new executive who t oversaw the middle managers definitely were ....
 
 
132 
feature-based model of reflexive antecedent access can adequately capture the 
experimental data, then the claim for structured access based on the data in Chapter 2 
is weakened. If, on the other hand, structured access reflexive models better fit the 
observed data, then the models provide an additional argument for structured access 
in comprehension. 
Each of the two questions was investigated in a separate computational 
experiment. Experiment 4 models the size of the predicted interference effect for 
reflexives if they retrieved their antecedent using a full set of morphological and 
structural cues. This experiment directly compares a feature-based access model of 
agreement to a feature-based access model of reflexives, to assess whether or not 
passive memory dynamics do in fact minimize any interference effect that might 
occur for reflexives. To preview the findings, the modeling results in Experiment 4 
suggest that reflexives should in fact show a reduced interference effect relative to 
agreement, even if they retrieve the subject in the exact same fashion. This finding 
makes it important to determine the answer to the second question: do structured 
access or feature-based access models better capture the experimental data in Chapter 
2? Experiment 5 directly compares the predictions of a feature-based model of 
reflexives to a structured access model of reflexives to attempt to answer this 
question.  
 
 
133 
 
Figure 3.2: Average activation for target (black) and distractor (red) NPs for a 
sentence that shows partial-match interference at the reflexive. Incorrect retrievals of 
the distractor NP are reflected in the increased activation at the plural reflexive 
themselves. 
 
Modeling feature-based and structured access 
 
 The crucial contrast that is drawn here is between feature-based and structured 
models of memory access. Feature-based models use a mixture of structural and 
feature-based cues to access memory, as is maintained by interference models of the 
agreement attraction effect (Badecker & Kuminiak 2007; Wagers 2008; Wagers et al 
2009), as well as models of NPI interference effects (Vasishth et al 2008). In contrast, 
a structured access model means that morphological feature information is not used to 
access memory; instead, only structural cues are used in retrieval. In the context of 
0 500 1000 1500 2000 2500 3000
 0.
 0
 0.
 5
 1.
 0
 1.
 5
 2.
 0
 2.
 5
 Time
 Act
 ivatio
 n
 Head NP
 Distractor NP
 *The new executive who t oversaw the middle managers apparently doubted themselves
 
 
134 
the ACT-R model, this means that the search cues are selectively restricted to only 
structural information. This has the effect of rendering access effectively blind to the 
morphological feature content of the interfering position, and so there should be no 
effect of the morphological feature content of the distractor noun on the retrieval 
process.  
Admittedly, the notion of ?structural cue? in the context of a content-
 addressable architecture is unclear, and there are a number of important issues that 
arise when considering the plausibility of representing structured linguistic 
representations on individual memories (as noted by Vasishth et al 2008; Wagers 
2008). The primary difficulty is that a relational notion like c-command is defined 
over pairs of constituents, which makes it qualitatively different from the lexical 
content of constituents (such as morphological features). Inherent content can 
straightforward be used as a retrieval cue in a content-addressable architecture, but 
relational notions like [+c-command] cannot be encoded on any single chunk, making 
it difficult to use it as a retrieval cue. However, the issue is somewhat orthogonal to 
the present concerns. Whether or not the relational notions that characterize linguistic 
structure can be adequately encoded in retrieval cues is a question that speaks to the 
feasibility of content-addressable models of sentence processing, rather than the 
possibility of structured access. For the current purposes I follow Vasishth et al 
(2008) in hypothesizing the existence of a cue that is functionally equivalent to [+c-
 commanding], even tough significant extra architectural commitments may be 
required to implement such a cue. Even if [+c-commanding] does not prove to be a 
plausible retrieval cue due to its inherently relational nature, the fact remains that in 
 
 
135 
order for a content-addressable parser of the sort presented here to function at a basic 
level of accuracy, there must be some cue that discriminates structurally accessible 
content from inaccessible content (a similar approach was taken by Lewis & Vasishth 
2005). Although I adopt the term [+c-commanding], this is merely a notational 
convenience. No matter what label it is assigned, the role of this structural cue in the 
present model is formally equivalent.  
 
The model 
 
In Experiment 4, all 8 conditions from Experiment 1 were modeled, and in 
Experiment 5, only the reflexive conditions were modeled. The structure of the 
crucial experimental sentences is repeated in (3.8). In Experiments 1-3, the structure 
of the subject noun was held constant across dependencies. The subject noun was 
always a singular, gender-biased noun that was modified by a subject relative clause. 
As before, the main clause subject noun is the target memory due to its structurally 
accessible position (i.e. it is the c-commanding local subject), regardless of its feature 
content. The distractor noun was the object NP that was embedded inside the relative 
clause; since it does not c-command the verb or reflexive, the distractor was always 
structurally inaccessible. The probe refers to the memory retrieval of theoretical 
interest. For agreement sentences, the probe retrieval was engaged by the agreeing 
verb form, and reflected the need to find and attach the subject of the VP. For 
reflexive dependencies, the probe retrieval was initiated by the reflexive element, and 
reflected the processes involved in retrieving the antecedent. 
 
 
 
136 
(3.8) a. [The new executive]target who oversaw [the middle 
manager]distractor apparently [was]probe dishonest about the 
company?s profits. 
b. [The new executive]target who oversaw [the middle 
manager]distractor apparently doubted [himself]probe on most 
major decisions. 
 
 It is important to consider the role of the free parameters in a model such as 
ACT-R, as there are a number of model parameters that the modeler is free to 
manipulate: these include the variance on the noise distribution (s), the magnitude of 
the penalty for mismatching features (M), the amount of maximum activation for each 
chunk (G), the decay rate (d), the base associative strength for a cue (S), and the 
scaling parameter (F). There are in general default settings for many parameters 
(Anderson & Lebiere 1998; Lewis & Vasishth 2005), and model fits in previous work 
have adopted a single parameter setting to assess the quantitative fit provided by the 
model. I adopt a different approach here. The focus of the modeling reported here is 
not to determine whether or not the model can accurately fit a given set of 
experimental results, as has been the focus in previous applications of ACT-R to 
sentence processing results (Lewis & Vasishth 2005; Vasishth et al 2008). Instead, I 
am comparing the fit of two separate model structures to the data, and asking which 
model is better suited to capturing the observed data. Thus rather than focusing on the 
fit of a single set of parameters, I consider the effects of a range of parameter settings 
that span values assumed in previous work. In all reported results, 324 different 
model combinations are reported. The most important advantage of this approach is 
that it allows an assessment of the robustness of a predicted effect: if an effect 
survives under various parameterizations of the model, it can be regarded as a reliable 
 
 
137 
prediction of the model. Effects that are present only under certain parameter settings 
require a greater number of assumptions to motivate; they require independent 
motivation for adopting that parameter setting, beyond noting that it predicts a given 
set of data. The sole exception to this approach is the scaling parameter F, which was 
chosen to ensure that predictions of the model were on an appropriate scale (in all 
simulations, F = 2.0).  
 The model of the crucial memory retrieval engaged by the probe provides two 
dependent measures of interest. The first measure is the proportion of retrievals that 
result in recalling the target memory position. Recall that the memory that is retrieved 
is the one that has the greatest activation at the point of retrieval, given the search 
cues; in other words, retrieval selects the most probable memory given a retrieval 
query. The percentage of accurate retrievals thus refers to the percentage of trials in 
which the target memory was the most probable memory at the point of retrieval. The 
second measure provided by the model is a measure of the interference effect, which 
is a measure that is derived from predicated retrieval latencies that the model 
generates. As in Experiment 1, the interference effect refers to the effect on predicted 
retrieval times that occurs when the distractor NP contains interfering feature content. 
It is calculated by subtracting the predicted retrieval time for [+intr] conditions from 
[-intr] conditions, and is a measure of the predicted impact of interfering material on 
retrieval times.  
The Lewis and Vasishth (2005) ACT-R sentence processing model is a fully 
specified parser, including an implementation of a left-corner parser and a lexical 
access module. In order to focus on the predictions of the memory access component 
 
 
138 
of the parser, the model presented here abstracts away from these and simply 
stipulates the schedule of retrievals required to parse the sentences in (3.8). Thus the 
model presented here is a minimal implementation of the crucial retrieval process. 
The schedule of non-critical retrievals, which controls the prior history component of 
the activation process, is given in Appendix A; this schedule is derived from 
empirical estimates of processing times from Experiments 1-3, as well as a 
consideration of the hypothesized parsing processes that engage both the target and 
distractor NPs.  
 
Experiment 4: comparison of interference effect for agreement and reflexives 
 
 Experiment 4 asked whether or not there is a substantial difference in the 
predicted size of the interference effect for reflexives relative to agreement, on the 
assumption that both types of items retrieve their antecedents in exactly the same 
manner. Thus, the model for reflexives and agreement accesses the local subject in 
the exact same manner, using a feature-based search that uses a combination of 
morphological and structural cues. As described above, the two dependencies differ 
on a number of dimensions (such as position relative to the verb and linear distance 
from the subject). Thus it is worthwhile to assess the nature of any baseline 
differences that exist. Both agreement and reflexive retrievals are modeled with the 
same feature-based access model: the subject is retrieved using the cues local subject 
and number; in addition, singular reflexives probe memory with the cue gender. Due 
to syncretism in the preterite verb form in English, the retrieval of the subject at the 
 
 
139 
verb that directly precedes the reflexive (see Figure 3.2) does not include number; 
thus there is a targeted re-activation of the local subject at this point. 
 As described above, a range of model parameterizations were considered, 
totaling 324 distinct combinations in all. For each combination, 5000 Monte Carlo 
simulations were run. Each simulation included the full series of hypothesized 
retrievals, including random noise, and provided a prediction for i) the retrieval 
latency of the most probable memory and ii) the identity of the most probable 
memory (target or distractor). Within each parameterization, the average retrieval 
latency of the most probable memory in each individual trial was calculated, 
providing an estimate of the average retrieval time that occurs for that retrieval. The 
average retrieval time does not distinguish between target and distractor retrievals; 
the predicted retrieval latency does not contain error signals generated by retrieving 
the distractor NP. In addition to the average retrieval latency, within each model 
parameterization the proportion of correct retrievals was calculated by observing on 
how many trials the target NP was correctly retrieved. All eight conditions from 
Experiment 1 were modeled, and they are repeated in a simplified form here as (3.9) 
and (3.10) for convenience: 
 
(3.9) a. The executive who oversaw the manager apparently was 
dishonest about the company?s profits.  [+gram,-
 intr] 
b. The executive who oversaw the managers apparently was 
dishonest about the company?s profits.  [+gram, 
+intr] 
c. The executive who oversaw the manager apparently were 
dishonest about the company?s profits.  [-gram,-
 intr] 
 
 
140 
d. The executive who oversaw the managers apparently were 
dishonest about the company?s profits.  [-gram, 
+intr] 
 
(3.10) a. The executive who oversaw the manager apparently doubted 
himself on most major decisions.   [+gram,-
 intr] 
b. The executive who oversaw the managers apparently 
doubted himself on most major decisions.    
[+gram, +intr] 
c. The executive who oversaw the manager apparently doubted 
themselves on most major decisions.    [-gram,-
 intr] 
d. The executive who oversaw the managers apparently 
doubted themselves on most major decisions.  
  [-gram, +intr] 
 
 The predictions for each condition were determined in independent runs. The 
interference effect was determined by subtracting the retrieval latency for [-intr] 
conditions from [+intr] separately for grammatical and ungrammatical sentences.  
 The rate of incorrect retrievals for each condition is plotted in Figure 3.3. It 
can be seen that there is a sizeable decrease in the number of incorrect retrievals in 
interfering environments for reflexives compared to agreement. To determine whether 
or not this held within each model parameterization, the difference in error rate as a 
result of interference was calculated for both grammatical and ungrammatical 
sentences alike. These values were then directly compared across reflexive and 
agreement models, for each parameter setting. The distribution of these differences is 
shown in Figure 3.4, which indicates that for almost every single model setting, the 
percentage of incorrect retrievals due to interference was lower for reflexives than for 
agreement dependencies. This indicates that the baseline activation differences that 
 
 
141 
are evident in comparing Figures 3.1 and 3.2 have a clear impact on the amount of 
interference that is expected to occur for reflexives.   
 
Figure 3.3: Percentage of retrieval of distractor NP, for all parameterizations 
(n=324), at critical probe position for agreement and feature-based reflexive models. 
0.
 00
 0.
 05
 0.
 10
 0.
 15
 0.
 20
 0.
 25
 0.
 30
 0.
 35
 Predicted % of incorrect retrievals
 %
  r
 et
 rie
 va
 l o
 f i
 nt
 er
 fe
 rin
 g 
no
 un
 Agreement Feature-based reflexives
 +Gram,-Intr
 +Gram,+Intr
 -Gram,-Intr
 -Gram,+Intr
 
 
142 
 
Figure 3.4: Difference in interference error between agreement and feature-based 
reflexive models (n=324). Error bars indicate 95% confidence intervals. 
 
 
The predicted interference effects for agreement and reflexives are 
summarized ub Figure 3.5. The values in Figure 3.5 reflect the average retrieval time 
for [+intr] conditions minus the retrieval time for [-intr] conditions, for grammatical 
and ungrammatical sentences alike. Note that the average retrieval time for any given 
condition is a mixture of retrieval latencies for incorrect and correct retrievals (as in 
Vasishth et al 2008). There are two main findings from these results. First, the 
direction of the model?s predicted interference effect for grammatical conditions is 
not stable across parameterizations; the model?s predicted interference effect for 
grammatical sentences is instead centered closely around zero. This is an important 
observation that I return to below. The second clear finding is that the predicted size 
-0
 .1
 5
 -0
 .1
 0
 -0
 .0
 5
 0.
 00
 0.
 05
 0.
 10
 0.
 15
 Reflexive/agreement error difference
 D
 iff
 er
 en
 ce
  in
  e
 rr
 or
 Grammatical Ungrammatical
 
 
143 
of the interference effect is smaller for reflexives than it is for agreement. In 
ungrammatical sentences, the mean interference effect for agreement is 
approximately -137ms, whereas for reflexives, it is less than half that (-52ms). Thus 
reflexives are expected to show approximately half as much interference as 
agreement, an observation that holds independent of the magnitude of the interference 
effect. Figure 3.6 shows a direct comparison of the agreement and reflexive 
interference effect for each parameterization, calculated by subtracting the agreement 
interference effect from the reflexive interference effect. It can be seen that across 
parameterizations, there is no reliable difference between agreement and reflexives 
for grammatical conditions. However, in ungrammatical conditions, almost every 
parameterization predicts a smaller interference effect for reflexives than for 
agreement. 
 
 
144 
 
Figure 3.5: Predicted interference effect ([+intr]-[-intr] conditions) for agreement and 
feature-based reflexive models (n=324). Error bars indicate 95% confidence intervals.  
 
 Thus main finding in Experiment 4 is that the susceptibility of reflexives to 
interference is predicted to be significantly less than that for agreement simply due to 
baseline activation differences between the two dependencies. The local subject NP is 
reactivated at the verb immediately prior to the reflexive, and this provides an 
activation boost that lowers error rates and minimizes the interference effect. This 
generalization held true of almost every single parameterization tested.  
-4
 00
 -3
 00
 -2
 00
 -1
 00
 0
 10
 0
 20
 0
 Size of predicted interference effect
 In
 te
 rfe
 re
 nc
 e 
ef
 fe
 ct
  (ms
 )
 Agreement Reflexives
 Grammatical
 Ungrammatical
 
 
145 
 
Figure 3.6: Difference in predicted interference effect (agreement-reflexive 
conditions) for all models (n=324). Error bars indicate 95% confidence intervals.  
 
The second important finding from Experiment 4 is that there are no clear 
model predictions for grammatical sentences. This is an important result that I will 
return to in the general discussion below, but the intuition for this fact is 
straightforward. It stems from the fact that there are opposing interference effects at 
work in grammatical sentences: inhibition from similarity-based interference and 
facilitation from retrieval error. These two effects both exert an effect on grammatical 
sentences, when the interfering feature (an embedded [+singular] distractor NP when 
the probe searches for [+singular]) is both more similar to other elements in the parse 
(i.e. the target subject noun) and causes increased retrieval error. In ungrammatical 
-3
 00
 -2
 00
 -1
 00
 0
 10
 0
 20
 0
 30
 0
 Reflexive/agreement interference effect difference
 In
 te
 rfe
 re
 nc
 e 
ef
 fe
 ct
  d
 iff
 er
 en
 ce
  (ms
 )
 Grammatical Ungrammatical
 
 
146 
sentences, on the other hand, these two effects are dissociated: greater retrieval error 
occurs when the two nouns are more dissimilar. Understanding the source of the 
equivocal predictions for grammatical sentences is important in interpreting previous 
literature, and I return to a fuller discussion of this after Experiment 5. 
The main finding of Experiment 4?reduced interference for reflexives even if 
they are processed in the same manner as agreement?corroborates the concern about 
Experiments 1-3 raised above. There is a smaller predicted interference effect for 
reflexives compared to agreement. This raises the possibility that the observed 
difference between the interference profile for agreement and reflexives reflects not a 
qualitative difference in memory access strategy, but rather an underlying quantitative 
difference stemming from baseline differences between the two. Experiment 5 
directly compares the predictions of the two access strategies against the empirical 
data to ask which access strategy for reflexives is a better fit to the observed empirical 
data from Experiments 1-3.  
 
Experiment 5: comparison of access strategies for reflexives 
 
In order to assess whether a feature-based or a structured access account for 
reflexives better captures the results of Experiments 1-3, Experiment 5 directly 
compared the predictions of the two modes of access against the observed empirical 
data. Feature-based models reflexive dependencies retrieve the subject using a 
mixture of structural (i.e., local subject) and morphological cues (number for 
themselves, number and gender for himself/herself). Structured access models retrieve 
 
 
147 
the reflexive?s antecedent based only on structural cues. In the context of the current 
model, this effectively implements a targeted search for the local subject. It was seen 
in Experiment 4 that the predicted interference effect for feature-based reflexives was 
smaller than that for agreement, but there was still a clear prediction for a facilitatory 
interference effect. As no interference effect was observed in the experiments in 
Chapter 2, an important question is whether or not this finding is more likely to have 
been generated by a feature-based or structured access model for reflexives.   
The empirical data of interest are the interference effects for reflexives 
observed in Experiments 1 and 3, as reflected in total reading times at the critical 
region. Total times were chosen because they provided the largest interference effect 
for agreement, and so provided the best opportunity for an interference effect to be 
observed for reflexives, if there was indeed any interference. For each participant 
(n=40 from Experiment 1, n=32 from Experiment 3), the average total reading time 
in each condition was obtained, and an interference effect score was derived.  
The predicted error rates for both reflexive models are shown in Figure 3.7. It 
can be seen that as in Experiment 4, feature-based models of reflexive antecedent 
access are prone to greater error in the presence of feature-matched, but inaccessible 
constituents. This is not true of structured access models, which show similar rates of 
error across all four conditions. This is due to the fact that in structured access 
models, the feature manipulation across conditions is uncorrelated with the structural 
search cues; thus, error rate and feature manipulation do not covary. The result is that 
there is no differential erroneous retrieval across conditions.  
 
 
148 
 
Figure 3.7: Percentage of incorrect retrievals for reflexive conditions for feature-
 based and structured access models (n = 324).  
0.
 00
 0.
 05
 0.
 10
 0.
 15
 0.
 20
 0.
 25
 0.
 30
 0.
 35
 Predicted % of incorrect retrievals
 R
 ef
 le
 xi
 ve
 s:
  %
  r
 et
 rie
 va
 l o
 f d
 is
 tr
 ac
 to
 r 
N
 P
 Feature-based access Structured access
 +Gram,-Intr
 +Gram,+Intr
 -Gram,-Intr
 -Gram,+Intr
 
 
149 
 
Figure 3.8: Comparison of predicted interference effects (solid) for reflexive 
conditions (n = 324) and observed reflexive interference effects from Experiments 1 
and 3 (by participants, n = 72). Error bars indicate 95% CI. 
 
Figure 3.8 presents a comparison of the predicted interference effects and the 
observed interference effects in total reading times estimated from a joint 
consideration of Experiments 1 and 3. As seen in the predicted error rates, the lack of 
correlation between the search cues and experimental manipulation in structural 
access models leads to no consistent differences in retrieval time across conditions. 
For this reason, the predicted interference effects are tightly centered around zero 
milliseconds. Feature based models, on the other hand, predict consistent direct 
interference effects (processing facilitation) for ungrammatical conditions. Again, as 
-2
 00
 -1
 00
 0
 10
 0
 20
 0
 Predicted and observed reflexive interference effects
 In
 te
 rfe
 re
 nc
 e 
ef
 fe
 ct
  (ms
 )
 Feature-based access Structured access
 Model: Grammatical
 Model: Ungrammatical
 Empirical: Grammatical
 Empirical: Ungrammatical
 
 
150 
in Experiment 4, there is no clear prediction for the direction of the interference effect 
in grammatical sentences, due to the tradeoff between inhibitory and facilitatory 
interference effects. Instead, the average prediction is for no interference effect.  
The distribution of the empirical interference effect in Figure 3.8 appears to 
better accord with the predictions of the structured access model. In order to assess 
which model better predicts the observed data, Bayes Factor model comparison 
(Gallistel 2009) can be used to determine which of the two models, feature-based or 
structured access, better predicts the observed data. One way of measuring this is with 
likelihood of any one model given the observed empirical evidence, which can be 
derived from Bayes? rule as in (3.11): 
 
(3.11)  = ( | )
 !
  
 
In (3.11), the predictions of the model serve as the prior p(?|M), which is 
combined with the likelihood of the size of the interference effect given the data. The 
greater the degree to which the model predictions and empirical data converge, the 
greater the likelihood of the model that generated those predications. By taking the 
ratio of the likelihood of the structured model given the data over the feature-based 
access model, we derive the odds in favor of one model over another.  
Results indicate that the data is better fit by the predictions of the structured 
access models. For ungrammatical sentences, the odds are 3.2:1 in favor of structured 
access models over feature-based access models. This odds ratio can be interpreted in 
an intuitive manner, as the odds that the structured model is correct over the feature-
 
 
151 
based model. Generally, values of greater than 3:1 are regarded as ?substantial? 
evidence (Jeffreys 1961) in favor of a hypothesis. For grammatical sentences as well, 
there was some evidence in favor of structured access model, with the odds favoring 
them over feature-based access by 2.2:1. Although this value this is generally 
considered only weak or ?anecdotal? evidence in favor of a hypothesis, it is notable 
that the tight predictions of the structured access model of reflexives better fits the 
observed data for both comparisons. 
The results of Experiment 5 clearly support the hypothesis of structured 
access in comprehension. Crucially, for ungrammatical sentences where feature-based 
and structured access predictions differed, an evaluation of the observed interference 
effects showed that there is substantial evidence in favor of the structured access 
model over the feature-based access model. There was additionally weak evidence in 
favor of structured access models for grammatical sentences. Although feature-based 
and structured access models did not make clearly different predictions for the size of 
the interference effects, the predictions of structured access models were more 
narrowly centered around zero, which gave it a slightly better fit to the observed 
empirical data. This analysis gives important support to the hypothesis of structured 
access for reflexives. In particular, the modeling results help to alleviate one concern 
over the null-effects logic of the reflexive manipulations of Experiments 1 and 3. In 
matching specific model predictions to empirical data, Experiment 5 formulated the 
?null? as a specific range of predicted values derived from the structured access 
model, and provided positive evidence that the predictions of the structured access 
 
 
152 
model provide a better description of the empirical data than the alternative 
hypothesis of feature-based access. 
 
Interim conclusions  
 
The models constructed in Experiments 4 and 5 clarify the conclusions that 
are licensed from Experiments 1-3. Experiment 4 demonstrated that reflexives are 
predicted to show less interference than agreement simply due to their linear position 
after the matrix verb. This position confers an activation boost to the local subject, 
and reduces the predicted interference effects observed at the immediate post-verbal 
position. Error rates and interference effects for feature-based models of reflexives 
were consistently smaller than those predicted for agreement in similar models. This 
led to the possibility that the different interference profiles in Experiments 1-3 were 
due not to qualitatively different access strategies, but rather simply a reflection of an 
underlying quantitative difference in the size of the interference effect due to baseline 
activation differences in the local subject NP between agreement and reflexive 
dependencies at the point of retrieval. Experiment 5 showed that this alternative 
hypothesis did not adequately capture the empirical data, and provided further support 
for structured access models of reflexives.  
Together, the modeling experiments evaluate and reject an alternative 
explanation of the data presented in Chapter 2. The empirical observation that 
reflexives appear to access structurally commanding NPs preferentially is in fact 
indicative of an underlying structured access strategy. This confirms the first 
 
 
153 
prediction of a structured access model of memory access for reflexives: the memory 
access procedure does not retrieve feature-matched, but structurally inaccessible NPs. 
Instead, the main determinant of whether or not an antecedent is considered is its 
structural position, as expected if reflexives engage an access mechanism that 
effectively targets the local subject. 
  
The predictions of rational memory access models 
 
 The clearest result of the models presented here is the facilitation effect 
observed for interference in ungrammatical sentences.  In this section I examine the 
source of this effect and show that an analysis of the model?s behavior supports the 
claim that facilitatory interference is the evidence that is necessary to show that 
structurally inaccessible positions are retrieved during memory access. An additional 
interesting finding from Experiments 4 and 5 is that in grammatical sentences, the 
model provides no consistent prediction for the direction of the interference effect. 
An examination of these two results clarifies the architectural conclusions are 
licensed when interference effects are observed in online processing.  
There is a fairly straightforward characterization of the reason that the model 
predicts no interference for grammatical sentences. To see this, consider the summary 
of the critical conditions that is presented in Table 3.1. Here it can be seen that the 
experimental design manipulates two dimensions along which interference has been 
claimed to occur: similarity between the target and distractor NPs in memory (Gordon 
et al 2001; Gordon, Hendrick & Johnson 2004) and the match between the retrieval 
 
 
154 
probe and the distractor NP (Van Dyke & McElree 2006; Van Dyke 2007). Similarity 
between NPs in memory occurs when the two nouns in the sentence share the same 
value for a given feature; within the current experimental manipulation, this occurs in 
conditions where both target and distractor NPs are singular. On the other hand, the 
conditions also vary with respect to whether or not the distractor NP matches the 
features required by the agreeing verb (the probe-distractor match). These are 
different for the grammatical and ungrammatical sentences, reflecting the fact that the 
probing verb requires [+singular] in the grammatical case, and [+plural] in the 
ungrammatical case. 
 
 Target-
 Distracto
 r match 
Distractor- 
Probe  
match 
The executive+SG that oversaw the manager+SG definitely was+SG.. ? ? 
The executive+SG that oversaw the managers+PL definitely was+SG.. ? ? 
The executive+SG that oversaw the manager+SG definitely were+PL.. ? ? 
The executive+SG that oversaw the managers+PL definitely were+PL.. ? ? 
 
Table 3.1: Distribution of NP feature match and match of the inaccessible NP to 
retrieval cues across experimental conditions. 
 
Viewed in this way, the reason for the no interference prediction for the 
grammatical conditions is fairly straightforward. In short, interference due to the 
target-distractor NP match has an inhibitory effect on retrieval times, and interference 
due to probe-distractor match has a facilitatory effect. The pair of grammatical 
conditions covary on these two distinct dimensions of interference, and their opposing 
effects cancel each other out in aggregate. These two types of interference are not 
simultaneously present for any of the ungrammatical sentences: NP match and 
inaccessible match interference occur in different conditions. When deriving 
 
 
155 
predictions for interference in grammatical conditions, the relative size of the 
inhibition and facilitation effects are controlled by separate model parameters. Thus, 
the relative weighting of parameter over the other determines the model?s predictions 
for any one parameterization, but on average, the prediction is for no difference 
between these conditions. In ungrammatical sentences, on the other hand, the 
configuration that has an NP match is not the same as the configuration that causes 
increased retrieval error. Importantly, the underlying source of inhibition for 
matching features across the two NPs is distinct from the source of facilitation due to 
the inaccessible match. Of the two sources of interference, inhibition and facilitation, 
it is only facilitation that provides evidence that the incorrect position was retrieved.  
The source of the inhibition effect due to the NP match is straightforward, and 
it may be understood in a number of ways. In the context of the ACT-R model used 
here, it is one instance of the well-known fan effect from general working memory 
tasks (Anderson 1974; Anderson & Reder 1999). The fan effect refers to the fact that 
the more memories a retrieval cue is associated with, the longer retrieval takes. 
Above it was noted that this follows directly from a rational model of memory access. 
When there are multiple items with the same feature specification, that cue is overall 
less diagnostic. This in turn leads to a less robust retrieval process, causing slower 
access time. On this model, the inhibition comes from using a retrieval cue that is less 
effective, slowing retrieval times. 
However, other models of this inhibitory effect do not require that the 
inhibiting feature actually be used as a retrieval cue. These models predict similar 
effects whenever two memories share feature content (Nairne 1988, 1990; Oberauer 
 
 
156 
& Kliegl 2006). In these models, feature similarity is the primary cause of forgetting. 
One way to understand this is through feature-overwriting (Nairne 1988, 1990). 
Feature-overwriting can occur whenever two memories overlap in a given feature 
value, leading to that feature getting ?erased? from the representation. This leads to 
less robust representations for the memories involved, and slows subsequent recall 
(Nairne 1988). These models accord well with interference accounts of information 
loss in short-term memory, which are considered to capture most known facts about 
short-term forgetting (Ricker, AuBouchon & Cowan 2010).  
In addition to the large body of work that suggests that such feature-
 overwriting occurs in general working memory, there are a number of results in 
sentence processing that appear to require us to posit such a process. In particular, the 
results of a number of studies by Peter Gordon and colleagues (Gordon et al 2001, 
2004, 2006) suggest that it is NP similarity, rather than probe-to-item similarity, that 
causes inhibition in sentence processing. Gordon and colleagues observed inhibition 
when two nouns in a sentence were similar along dimensions that are unlikely to be 
used as retrieval cues, such as whether or not both NPs are proper names. Because of 
this, inhibitory effects due to NP similarity in memory do not straightforwardly 
license conclusions about the features used in retrieval, and more importantly for the 
present purposes, they do not entail that the interfering memory was actually retrieved 
and considered. The multiple-match effect arises in a more indirect fashion. 
However, facilitation due to probe-item match does license the conclusion that 
illicit material is retrieved online. The assertion that retrieval error leads to faster 
retrieval times is not as immediately intuitive as the similarity-based inhibition effects 
 
 
157 
that obtain in NP match environments. The facilitation that stems from retrieval error 
has been described as an illusion of grammaticality effect (Wagers 2008; Wagers et al 
2009; Phillips et al 2010), reflecting the intuition that when retrieval error happens in 
ungrammatical sentences, the sentence is perceived as grammatical, and the average 
reading time is lowered due to those trials in which no error signal was registered. 
However, the source of this effect is more general than this: retrieval error leads to 
faster retrieval latencies in grammatical and ungrammatical sentences alike, as shown 
in Figure 3.9.  
 
Figure 3.9: Relationship between interference error and interference effect on 
average retrieval latency. Blue points indicate comparisons between grammatical 
conditions, and red points are ungrammatical comparisons, for each of 324 
parameterizations. 
 
 
-5 0 5 10 15
 -4
 00
 -3
 00
 -2
 00
 -1
 00
 0
 10
 0
 Relationship between error and faciliation
 Predicted interference error (% error)
 Pr
 ed
 ic
 te
 d 
in
 te
 rfe
 re
 nc
 e 
ef
 fe
 ct
  (ms
 )
 Grammatical conditions
 Ungrammatical conditions
 
 
158 
The race model aspect of the retrieval process drives this relationship between 
retrieval error and facilitation. The crucial property is the optimal decision rule 
assumed by the model: the memory that has the highest activation in light of its past 
history and a set of retrieval cues is the memory that is retrieved. Given the 
relationship between activation and retrieval latency, this means that the fastest 
memory to respond to the search cues is always the winner, whether or not it is the 
?target? memory. This makes the retrieval process in ACT-R in important ways 
similar to race models of sentence processing (Traxler, Pickering & Clifton 1998; 
Van Gompel, Pickering & Traxler 2001; see also Frazier 1979), which maintain that 
the parser adopts the fastest syntactic analysis to become available in cases of 
ambiguity. If the parser incorrectly retrieves ungrammatical material in cases such as 
*the key to the cabinets were?, then it effectively creates a syntactic ambiguity at the 
point of retrieval, even if the resulting structure is not well-formed. This spurious 
ambiguity in the retrieval process leads to processing facilitation in the exact same 
way that ambiguity speeds processing in race models: there are two possible 
retrievals, and the faster one is adopted. This has a facilitatory effect because the 
minimum of two random variables is never greater than the minimum of either 
random variable on its own: min(x,y) is always less than or equal to min(x) or min(y). 
If incorrect retrievals occur, then the distractor and target NPs have comparable 
retrieval latency; in the aggregate, choosing the quicker option will skew the reaction 
distribution negatively (see Figure 3.10).  
?Incorrect retrieval? thus simply means that quicker relation between the probe 
and the distractor can be built more rapidly than between the probe and the target; this 
 
 
159 
provides the processing facilitation in the context of an illicit retrieval. On the 
assumption that our processing measures are monotonically related to the latency of 
retrieval, then due to the race aspect of the model, retrieval errors should facilitate 
reading times on average. On the further assumption that retrieval drives structure 
generation, this claim can be seen to be equivalent to previous claims of ?race? type 
phenomena in sentence comprehension (Traxler et al 1999; Van Gompel et al 2001). 
For this reason, this observation is more general than the ACT-R implementation. 
There is a deeper way of understanding this feature, however: facilitatory interference 
follows from a rational model of memory access, as I will show below.  
Figure 3.10 summarizes the multiple-match and partial-match interference 
scenarios. It can be seen that in the multiple match interference environments that 
occur in the grammatical sentences, the interfering case (the man who kicked John 
hurt himself) is slowed due to feature overlap between man and John. This results in 
an inhibitory interference effect in the observed RT distributions. Any facilitation due 
to retrieval error is obscured by the larger inhibitory effect of similarity. For this 
reason, it is not possible to conclude from a behavioral slowdown that incorrect 
structure has in fact been generated. In contrast, in ungrammatical sentences, there is 
no feature overlap in the interfering case (*the man who kicked Katie hurt herself). 
The partial match causes the distribution of the retrieval times for the target and 
distractor to overlap to a greater degree, causing greater retrieval error. There is no 
corresponding inhibition due to feature overlap. Because of the race aspect of the 
model, this leads to greater retrieval errors and an overall facilitatory interference 
effect.  
 
 
160 
 
Figure 3.10: Effect of interference in multiple match and partial match comparisons. 
The inhibition in multiple match interference is driven by decreased retrieval 
latencies on the target noun, due to feature overlap. The facilitation in partial match 
interference is an increased overlap in the distribution of target and distractor retrieval 
distributions. The race aspect of the retrieval process leads to an overall facilitation 
effect, which unambiguously indicates that incorrect access has occurred online. 
 
 As mentioned, the link between behavioral facilitation and incorrect structure 
generation is more general than the ACT-R model. The fact that retrieval error leads 
to facilitation can be shown to follow directly from the rational structure of the 
model. Since this holds of the description of the system at an abstract, computational 
level, it holds across implementations of the rational access model. This provides 
another way of understanding race models of syntactic ambiguity (Traxler et al 1998; 
Van Gompel et al 2001, 2005), which might be said to operate in the same ?rational? 
Grammatical sentences:
 [The man] who kicked [John] hurt himself
 [The man] who kicked [Katie] hurt himself
 Ungrammatical sentences:
 [The man] who kicked [John] hurt herself
 [The man] who kicked [Katie] hurt herself
 Target NP: 
Retrieval latency
 distribution
 Observed RT:
 distribution of 
min(target, distractor)
 Inhibition due to decreased retrieval time on target 
noun, due to multiple [+masc] nouns.
 Facilitation due to increased 
retrieval error, which stems from increased
 overlap of target and distractor RTs.
 Distractor NP: 
Retrieval latency
 distribution
 
 
161 
fashion as memory access here: in all cases, the parser is doing what it thinks is 
correct in the fastest manner possible.  
To see how this follows from a high level computational description of the 
retrieval process, suppose that we have two conditions, one where a distractor NP 
matches to some degree a search cue E and one where it does not match at all. Let 
these be mi+ and mi-, respectively. We may express the advantage of the matching 
interfering memory over the non-matching interfering memory as the following odds 
ratio: 
 
  (3.12) 
( !!| )
 ( !!| )
 ? !!
 ( !!)
 !! ( !!)
  
 
On the assumption that prior terms reflect only the history of usage (Anderson & 
Milson 1989), and that usage is determined by grammatical role in the sentence rather 
than feature content, then this odds ratio can further be reduced to the following: 
 
  (3.13) 
( !!| )
 ( !!| )
 ? !!
 !!
  
 
 From here, an advantage for mi+ is evident across a wide variety of 
assumptions about the likelihood term here: the probability of using some search term 
E for a memory should invariably be greater when the memory contains some of the 
 
 
162 
search cues (mi+) than when it does not (mi-). There is, of course, one exception: 
degenerate cases where no partial cue overlap is allowed, which would arise if the 
search cues had to perfectly match the target memory. In this case, as neither memory 
is a perfect match (mismatching on structural cues both), the likelihood of both 
interfering memories is zero. In this degenerate case, both likelihoods are equal, and 
the posterior should be equal across these (i.e. zero for both). Thus, in all situations 
the likelihood ratio of the two posterior probabilities for the two interfering memories 
is equal to or greater than one: 
 
  (3.14) 
( !!| )
 ( !!| )
 ? 1 
 
In the context of a given set of retrieval cues, the probability of the interfering 
memory is always equal to or greater than that of the non-interfering memory, a fact 
that is mirrored in other models (see, e.g., Van Dyke & McElree 2006). Crucially, 
however, the increased probability of the interfering memory actually has a 
facilitatory effect given the current linking assumptions between access times and 
memory probability. Recall that the retrieval latency of the winner (mwin) in a 
Bayesian optimal system always is the memory with the highest posterior probability: 
 
  (3.15) 
!"# = max
 !
 ( | ) 
 
 
 
163 
It follows from the result above that the posterior of the winner in the context of an 
interfering element should always be equal to or greater than the posterior of the 
winner in the context of a non-interfering element: 
 
(3.16) 
max
 !?{!!!,? }
 ( | )
 max
 !?{!!!,? }
 ( | )
 ? 1 
  
This last result guarantees that when errors occur, the average probability of 
the winning memory is greater than when they do not. Because this probability is 
inversely related to retrieval latency, then it is must be the case that incorrect retrieval 
of inaccessible material must generate equal or eased retrieval times relative to non-
 interfering material. 
The observation that ambiguity actually has a facilitatory effect in a rational 
model has been noted previously (Traxler et al 1998; Van Gompel et al 2001, 2005; 
Levy 2008). The result presented here shows that the incorrect retrievals in a content-
 addressable architecture should give rise to the same facilitation, for essentially the 
same reason: incorrect retrievals create a spurious ambiguity where there is none 
permitted by the grammar. Like cases of grammatically licensed ambiguity, the parser 
is considering which of two structures to build.  
Thus interference effects in sentence comprehension fall into two classes, 
each with distinct mechanistic sources. Similarity among items in a sentence inhibits 
recalls of those items, whether it is due to retrieval processes (Lewis & Vasishth 
 
 
164 
2005) or feature-overwriting processes (Nairne 1988, 1990; Gordon et al 2001, 2004; 
Oberauer & Kliegl 2006). In contrast, retrieval error leads to processing facilitation. 
This distinction is important to consider, as it clarifies what data can be used as 
evidence for incorrect retrieval online. In previous literature both inhibitory (Van 
Dyke & McElree 2006) and faciltatory effects (e.g. Vasishth et al 2008; Wagers et al 
2009) have been attributed to incorrect retrievals; however, the analysis presented 
here suggests that only facilitatory effects can be taken as evidence of incorrect 
retrieval online.  
 
Relation to previous work 
 
In light of the somewhat counterintuitive finding that incorrect access should 
always result in faster retrieval, in both ungrammatical and grammatical sentences, it 
is worth revisiting prior work on interference effects and the conclusions that have 
been drawn from this work.  
 There have been a range of effects that have been labeled as interference 
effects. For example, Gordon, Hendrick & Johnson (2001) used self-paced reading to 
examine sentences like the following, containing object relative clauses with either 
pronouns or common nouns as subjects: 
 
(3.17)  a. The banker that the barber praised climbed the mountain. 
   b. The banker that Ben praised climbed the mountain. 
 
 
 
165 
 At the point of processing the embedded verb praised and at the immediately 
following word, they observed processing difficulty reflected in longer reading times 
in (1a) relative to (1b). Gordon and colleagues interpreted this as an interference 
effect that obtains at the point of retrieving the subject, which is then subsequently 
attached to the verb. The feature overlap between the two subject NPs (i.e. they are 
both definite noun phrases) in (1a) makes it relatively difficult for comprehenders to 
recover the target NP, presumably due to the content-addressable property of memory 
access. They found a number of similar effects in Gordon et al (2002), as well as in 
Gordon, Hendrick & Johnson (2004, 2006). In these experiments, they found that the 
degree of feature overlap between the two NP positions correlated with observed 
processing difficulty. Thus, nouns that were more similar on some dimension (for 
example, both being proper names) inhibited each other at the point of retrieval. Not 
all dimensions of similarity caused such interference: definiteness and specificity did 
not (Gordon et al 2004). They interpreted this finding as an interference effect of the 
sort that is predicted by a content-addressable architecture.  
Importantly, in this case similarity-based interference causes inhibition and 
slower reading times. On the predictions of the model discussed above, this is not due 
to incorrect access but rather due to interference degrading the memory 
representations of similar items. This is the account of these effects that Gordon and 
colleagues adopt. For these experiments in particular, it is difficult to imagine that the 
match between the retrieval cues and the nouns is the source of difficulty. For 
example, it is difficult to imagine that retrieval cues that select proper names are used 
to find the subject for the verbs in the experiment: it is unlikely that praised in (3.17) 
 
 
166 
retrieves its subject with a feature like [+proper name]. Rather than being due to a 
degraded cue (a fan effect), this inhibition reflects the match of the noun phrases in 
memory. This would appear to support feature-overwriting accounts of similarity-
 based interference (Oberauer & Kliegl 2006; Logacev & Vasishth to appear), and 
makes it more difficult to retrieve memories that have been degraded in this manner. 
These findings suggest that interference at the point of retrieval can arise because of 
mutual inhibition among elements in a parse, and so inhibitory interference in the 
context of multiple match interference cannot license conclusions about incorrect 
access, or the content of the retrieval cues.  
 However, Van Dyke and McElree (2006) challenged the characterization of 
these effects as the result of a disrupted encoding. Their primary question was 
whether or not they could find interference effects that stemmed entirely from the 
match of the retrieval cues to the items in memory. To do this, they compared two 
clefted sentences as in (3.18a-b). In (3.18a), the clefted noun boat is perhaps the only 
plausible object for sailed. In (3.18b), the verb fixed is instead plausible with a wider 
range of object nouns. Van Dyke and McElree reasoned that the semantic cues at 
sailed should resonate more strongly with boat than the cues at fixed. In this way their 
manipulation addressed the content of the retrieval cues, rather than feature matches 
amongst items in memory.  
In addition, Van Dyke and McElree crossed the retrieval cue manipulation 
with a memory load manipulation. For memory load conditions, participants had to 
maintain a list of words in short-term memory while they read sentences. Importantly, 
the memory list contained plausible completions for fixed, but not sailed. After 
 
 
167 
memorizing the list and reading a sentence, a list recall task and a comprehension 
question were asked, in that order. Van Dyke and McElree asked whether or not the 
content of the remembered list would interfere with the retrieval of a clefted noun at 
the gap site. If the memory list contained semantically plausible candidates for the 
gap site, they reasoned, then interference effects should be observed at the gap site. In 
contrast, sentences where the gap site was not a plausible host for any of the nouns in 
the memorized list should be relatively easy to process at the retrieval site. The 
crucial comparison was between (3.18a) and (3.18b): 
 
(3.18)  a. List: TABLE-SINK-TRUCK  
It was the boat that the guy who lived by the sea sailed e in two 
days. 
   b. List: TABLE-SINK-TRUCK 
It was the boat that the guy who lived by the sea fixed e in two 
days. 
 
 Upon reaching the critical verb fixed in (3.18b), reading times were longer 
than at sailed in (3.18a) by approximately 38ms, and longer than reading times for the 
same sentences when there was no associated memory load. Van Dyke and McElree?s 
conclusion was that it was the fit of the retrieval cues of fixed to the content of the 
interfering list that caused the disruption, rather than the similarity in items in 
memory, as this was held constant across memory load conditions.  
 Van Dyke and McElree attributed this to the greater likelihood of recalling an 
incorrect item from the memory load list, due to its match with the retrieval cues. 
However, the result in the previous section calls this interpretation into doubt. In the 
absence of further assumptions about the relationship between retrieval error and 
processing difficulty, this is exactly the opposite pattern expected on the current 
 
 
168 
model. If the interfering memory list items are incorrectly retrieved, then facilitated 
processing should be observed. The inhibition that Van Dyke and McElree observe is 
instead consistent with a fan effect: the semantic cues to thematic integration at the 
verb are less distinctive for fixed than for sailed in the context of the memory load, 
which would lead to inhibition without incorrect retrieval of the nouns from the 
memory load list. An account of these results in terms of incorrect retrieval is 
possible: for instance, it may be that an incorrect retrieval of material that is not 
integrated into a linguistic parse leads to a sort of type mismatch that specifically 
inhibits further processing in the case of retrieval error. However, in the absence of 
these extra mechanisms, these results suggest similarity-based inhibition, rather than 
incorrect access as the source of difficulty. Unlike the results from Gordon and 
colleagues (2001, 2002, 2004, 2006), it seems that these results stem from a fan 
effect. 
 In another set of studies, Van Dyke and Lewis (2003) and Van Dyke (2006) 
looked at interference effects from additional syntactic material. They asked if a 
greater number of subject positions interfered with the ability of a verb to retrieve the 
target subject NP from memory. For example, Van Dyke and Lewis (2003) contrasted 
configurations such as the following (subject NPs are bolded to highlight the 
interfering positions): 
 
(3.19) a. The worker was surprised that the resident who t was living 
near the dangerous warehouse was complaining about the 
investigation. 
b. The worker was surprised that the resident who t said that the 
warehouse was dangerous was complaining about the investigation.  
 
 
 
169 
 In (3.19a), at the point of processing the critical verb phrase was complaining, 
Van Dyke and Lewis hypothesized that the subject noun phrase (i.e. the resident 
who?), would have to be retrieved from memory and that furthermore, this retrieval 
would be subject to interference effects. Thus, the presence of an additional subject 
position in (3.19b) (the warehouse), relative to (3.19a), should impact processing 
measures at the critical region. Their hypothesis was supported by the data: increased 
reading times were observed for (3.19b) relative to (3.19a). The locus of the 
interference effect was at the point of retrieval, rather than upon encountering the 
second interfering subject. They argued that the source of these effects was 
interference at a retrieval stage, rather than interference during encoding. Although 
the time course of the observed effect locates the source of the interference in 
retrieval, rather than encoding, these results do not distinguish between similarity-
 based inhibition due to feature-overwriting and interference due to less distinct 
retrieval cues. On the current assumptions, in neither case is the online processing 
profile consistent with incorrect access. In these studies, while multiple subject-like 
elements cause similarity-based inhibition, there is no direct evidence that incorrect 
elements are retrieved in early parsing. 
 In a later study, Van Dyke (2007) showed that semantic similarity further 
contributed to the interference effects observed here. She constructed additional 
conditions like those seen in (3.20), which replace warehouse in (3.19) with 
neighbor.: 
 
(3.20) a. The worker was surprised that the resident who t was living 
near the dangerous neighbor was complaining about the investigation. 
 
 
170 
b. The worker was surprised that the resident who t said that the 
neighbor was dangerous was complaining about the investigation.  
 
 Neighbors, but not warehouses, match the selectional requirements (e.g. 
animacy) for the retrieving verb phrase was complaining. Van Dyke argued that if 
both semantic and syntactic cues were used to retrieve the subject verb phrase, then 
both manipulations should result in interference effects. Increasing interference by 
increasing the number of subject NPs as well as increasing interference by increasing 
the number of NPs that match the critical verb?s selectional restrictions both caused 
slowdowns in an eye-tracking task, across early and late measures. Again, this sort of 
processing slowdown does not implicate incorrect access online. It is compatible with 
a range of mechanisms that predict inhibited access to items in memory as a function 
of their similarity to other items in memory. 
However, a cloze-format offline task probed participants? comprehension of 
the critical VP by asking them to decide which of the three NPs present in the 
sentence truthfully completed sentence fragments of the sort ? ___ was complaining 
about the investigation.? Both semantic and syntactic sources of interference 
decreased participants? accuracy on this offline task; they were more likely to select 
the interfering noun in all interfering contexts in a way that suggested additive effects 
of semantic and syntactic interference.  However, Wagers (2008) suggested that an 
independent source of complexity in the Van Dyke & Lewis (2003) materials 
complicates their interpretation; namely, that at the point of retrieval, syntactic 
interfering conditions had two clause boundaries where non-interfering conditions 
had only one. After controlling for this by contrasting lexical versus expletive 
 
 
171 
subjects, he replicated their offline interpretation findings, but did not replicate their 
finding of increased reading times in response to an additional syntactic subject 
position.  
One possible objection to the reasoning advanced here is that it is possible that 
incorrect retrievals selectively inhibit processing, and so results like those in Van 
Dyke?s study do in fact indicate incorrect access online. It is unclear what these 
processes might be, however. It seems unlikely that they are due to reinterpretation or 
ungrammaticality detection, as the offline measures suggest that the neither reanalysis 
nor detection of ungrammaticality occurs in these situations  (Sturt 2003a; Van Dyke 
2007; Wagers 2008; Wagers et al 2009; Phillips et al 2010). Since no selective 
reanalysis or ungrammaticality detection appears to be engaged in case of an incorrect 
retrieval, it may be that the parser does not make a distinction between retrieving a 
target or distractor memory, treating both situations alike. If this argument is correct, 
then there seems to be no basis for concluding that more difficult processing occurs 
when a non-target memory is retrieved online. 
 Interference effects have been obtained in a number of other studies. For 
example, Fedorenko, Babyonyshev and Gibson (2004) found processing slowdowns 
at the verb in Russian when there were multiple nouns that share identical case, 
although interestingly these effects only obtained when both abstract case and 
phonological realization were identical across the interfering elements. Suckow, 
Vasishth, & Lewis (2005) found that when there were multiple animate NPs in 
German sentences, inhibition was observed due to similarity-based interference. As 
with the results reviewed above, increasing either the similarity of the material to 
 
 
172 
other constituents in the sentence or to the retrieval cues causes inhibition. These 
results are all architecturally interesting, in that they clearly indicate that the memory 
architecture of the sentence allows items in memory to interfere with each other in 
some fashion. Again, however, they do not demonstrate the facilitatory interference 
profile that indicates incorrect retrieval, and so it is unclear what to infer about 
incorrect access online from results that suggest inhibited processing in the face of 
interference. All of these results are compatible with a parser that can index and 
correctly access the desired items in memory in accord with structured access, if one 
assumes that items in memory can directly interfere with each other through feature-
 overwriting or conflicting bindings (Logacev & Vasishth to appear).  
 In contrast to the many examples of inhibitory interference effects, there is a 
somewhat sparser body of work that demonstrates facilitatory interference effects. To 
date there has been no clear demonstration of facilitatory interference for reflexives, 
with studies either showing no detectable interference (the present work, along with 
Nicol 1988; Clifton et al 1999; Sturt 2003a; Xiang et al 2009) or inhibitory effects 
(Badecker & Straub 2002; Sturt 2003a). As noted in Chapter 2, the two dependencies 
that reliably exhibit facilitatory interference effects are subject-verb agreement and 
NPI processing (Clifton et al 1999; Pearlmutter et al 1999; Drenhaus et al 2005; 
Wagers 2008; Vasishth et al 2009; Xiang et al 2009). Across these studies, facilitation 
is seen in ungrammatical sentences that have inaccessible feature-matched elements. 
There is a wide range of results that have been attributed to interference effects, but 
the conclusion that the parser routinely constructs ungrammatical parses (see, e.g. 
Van Dyke 2007) is not obviously warranted. Only a narrow subset of dependencies?
 
 
173 
agreement and NPI processing?appear to support this conclusion, and they may be 
the exception rather than the rule when considering parsing more generally, a topic 
that I return to in Chapter 5. 
  
 Conclusion 
 
 The theoretical and computational analysis presented here strengthen the case 
for structured access in comprehension on a number of fronts. First, the explicit 
computational models of the experimental results on the processing of agreement and 
reflexives support the conclusion that reflexives access the subject using structured 
access mechanisms. The narrow predictions of the structured access model provide a 
superior fit to the experimental data than did a feature-based access model of 
reflexive antecedent access, making a stronger case for structured access in 
comprehension.  
 Second, an analysis of the predictions of a rational access memory showed 
that facilitatory, rather than inhibitory interference is the behavioral signature of 
incorrect access. It has long been noted that in certain situations, structural ambiguity 
causes facilitated processing (Van Gompel et al 2001, 2005; Levy 2008). Facilitatory 
interference is a qualitatively similar phenomenon: if the parser incorrectly retrieves 
and attaches structurally licit and illicit material alike, it creates a situation of 
(temporary) spurious ambiguity. This apparent ambiguity leads to facilitated 
processing, as observed in cases of true syntactic ambiguity (Van Gompel et al 2001, 
2005). However, this facilitatory interference is only observed for a limited range of 
 
 
174 
syntactic dependencies: the partial-matching interference effects observed in NPI and 
agreement dependencies are the only robust demonstrations of structurally incorrect 
access in online parsing. Instead, a large portion of work on interference effects in 
sentence processing focuses on inhibitory effects due to increased similarity, but these 
effects do not provide evidence that the parser has generated ungrammatical parses by 
retrieving inaccessible linguistic material.  
 These two findings build further support for structured access mechanisms in 
comprehension, by strengthening the conclusion of structured access reached in 
Chapter 2, and demonstrating that there are few counterexamples to structured access 
in the existing literature. The wider range of interference results are compatible with 
structured access in that they indicate similarity-based inhibition, rather than spurious 
retrieval of non-target memories, in online parsing processes.  
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
175 
 
Chapter 4: Processing long-distance reflexives in Mandarin Chinese 
 
 
 
 
 
 
 
 
 
 The preceding chapters built the case for structured access by investigating 
reflexive dependencies in English. It was shown that the interference profile for 
reflexives is qualitatively different than that for agreement. The observed pattern of 
interference for agreement suggests that morphological features are used to retrieve 
memory representations in constructing English subject-verb agreement 
dependencies. English reflexives, on the other hand, were reliably immune to such 
interference, indicating that feature-based access is not used to retrieve the local 
subject to resolve a reflexive?s reference. This conclusion was supported by 
computational evidence presented in Chapter 3, which showed that a structured 
access mechanism better predicted the observed set of experimental results than did a 
feature-based model. 
 This set of findings constitutes one type of argument for structured access: 
there was no behavioral evidence that comprehenders considered inaccessible 
 
 
176 
antecedents. In the current chapter, I provide a complementary argument for 
structured access. I demonstrate that comprehenders retrieve and consider feature-
 mismatched antecedents if they are structurally accessible, even if they are not the 
target of retrieval; in order to make this argument, I examine the Mandarin Chinese 
long-distance reflexive ziji. Long-distance reflexives are a good test case in 
investigating structured access because of a number of unique properties. Like 
English reflexives, they must be structurally bound, but unlike English himself, their 
antecedent can potentially be found in any clause that dominates ziji (Huang & Liu 
2001; Huang, Li, & Li 2009). Here an investigation of the processing of ziji reveals 
that comprehenders check the local subject position before a more distant (but licit) 
subject position, regardless of the semantic features of the two subjects. The observed 
contrast between local and long-distance binding provides a useful diagnostic of the 
role of syntactic structure in memory access. Rather than investigating the impact of 
an illicit noun phrase on reflexive resolution, the studies reported here more directly 
test of the role of syntactic structure in finding an antecedent. Evidence that 
unacceptable antecedents are considered based on their syntactic position, even in the 
presence of a better candidate antecedent, would provide a strong argument for 
structured access.   
A comparison of long-distance and local binding configurations for ziji 
provides evidence that this is indeed the case. Across the two experiments presented 
in this chapter, syntactic position appears to provide the primary means of accessing 
linguistic memory when comprehenders try to construct a ziji-antecedent dependency. 
In particular, comprehenders appear to preferentially access a feature-mismatching 
 
 
177 
local subject position before accessing a feature-matched, but more structurally 
distant NP. This is expected if uniquely syntactic information is used to access ziji?s 
antecedent and consideration of antecedents occurs serially.  
In order to make this argument, I rely on time course evidence using the 
speed-accuracy tradeoff paradigm (SAT; Experiment 6), as well as event-related 
potentials evidence (ERP; Experiment 7). In previous chapters, the argument for 
structured access was built on previous arguments about memory access from the 
presence or absence of interference effects. In contrast, the argument in this chapter is 
centered on the argument from the time course of processing. I first present a brief 
overview of the argument from time course for content-addressable memory 
architectures, which introduces the SAT technique and the logic of investigating the 
time course of decision-making in processing. I then present SAT time course 
evidence that shows that comprehenders are slower to accept grammatical long-
 distance ziji interpretations than local interpretations of the anaphor. This is expected 
on a structured access account, where comprehenders consult the local subject 
position before more distant antecedent positions when processing ziji; this set of 
findings is not compatible with an account in which semantic feature cues are directly 
used in the search for an antecedent. The ERP evidence replicates the locality bias 
with an alternative methodology, and shows that canonical syntactic reanalysis is not 
engaged when shifting from local to long-distance bindings. Instead, the difficulty is 
indexed by an ERP component that has been associated with working memory 
difficulty: the left anterior negativity (LAN; Kluender & Kutas 1993).  
 
 
 
178 
Linking structured access and structured search 
 
 Processing evidence on the time course of memory access has primarily been 
used to address questions of memory architecture, contrasting content-addressable 
versus serial-register memory systems. This architectural contrast does not directly 
address the structured access hypothesis, which instead makes a claim about the 
content of the information used to index and access memory during parsing. 
Nonetheless, a consideration of the argument from time course is useful in the present 
context because the structured access hypothesis makes a prediction for the time 
course of memory access. On the assumption that retrieval is serial (i.e. multiple 
memories cannot be retrieved and processed in parallel), then in situations where the 
cues used in retrieval underdetermine the target of retrieval (i.e. when there are 
numerous acceptable syntactic targets), then the need to consider multiple targets will 
result in slowed processing. This delayed processing is a consequence of pursuing a 
narrowly syntactic access strategy, because a more restricted access strategy has the 
possibility of creating a fleeting ambiguity that needs to be ruled out by checking 
multiple positions. Consider a sentence like John gave Mary a book about himself. If 
the arguments in the preceding chapters are correct, then himself retrieves antecedents 
using only structural information. Thus in trying to fix the reference of the anaphor, 
there is a fleeting ambiguity from the point of view of memory access: the parser 
could in principle access either of the local c-commanding NP positions as an 
antecedent. Although the formal constraint supplied by the gender features makes the 
sentence unambiguous, resolving this ambiguity may require checking both noun 
 
 
179 
positions to remove the unacceptable antecedent Mary from consideration. In the 
English studies in Chapter 2, I argued that the initial candidate set is smaller than 
predicted on a feature-based access account. In contrast, for sentences like John gave 
Mary a book about himself, the situation is reversed: structured access predicts that all 
structurally accessible (local, c-commanding) antecedents are in the initial candidate 
set  (John and Mary), whereas featured access limits this set only to those that match 
in formal gender features (John).  
 In this way, structured access suggests the possibility of structured search: in 
situations where the retrieval cues underdetermine the target of retrieval, the parser 
might need to search through the syntactically licit positions, considering each in turn. 
This requires the assumption that is implicit in the memory models considered thus 
far: only one memory is retrieved for processing at a time (perhaps due to limitations 
on the scope of focal attention), forcing a sequential retrieval process. If this model is 
correct, then as the number of positions to be considered grows, so does average 
processing speed. On the further assumption that some positions are more likely to be 
considered prior to others, than there should be processing speed differences 
associated with accessing syntactic positions that are distant from the probe site 
compared to local syntactic positions.  
 Thus, a structured search account maintains that the parser is not allowed to 
entirely skip positions that could in principle be relevant in memory access, given a 
set of retrieval cues. This claim differs in important ways from the claims supported 
by existing time course evidence, as I argue below. The bulk of time course evidence 
has been taken to show that memory retrievals in sentence comprehension have the 
 
 
180 
property of direct access: accessing a target memory occurs in roughly constant time 
with respect to the size of the search path, due to the fact that direct access allows the 
parser to ?skip? irrelevant memory positions in retrieving linguistic material. Because 
the focus has been on whether or not the parser considered irrelevant memories in 
memory access, this evidence does not directly speak to the prediction of a structured 
search account, which maintains that the parser cannot skip memory positions that 
could potentially be relevant but for some reason do not contain the target memory.  
 
The argument from time course  
 
Arguments from the time course of memory access have been a strong source 
of support for content-addressable memory architectures (McElree 2000; McElree et 
al. 2003; Foraker & McElree 2007; Martin & McElree 2008, 2009). The logic of the 
argument from time course is as follows: if a serial search process is used in the 
access procedure for querying linguistic memory, then the time required to retrieve 
information should grow as a function of the size of the search path. If, on the other 
hand, memory access is mediated by a content-addressable memory architecture, then 
the size of the search path should not affect retrieval times. This is because of the 
direct access property of content-addressable memory architectures. Direct access 
refers to the fact that when memory is queried, the target memory directly resonates 
with the retrieval cues, without the need to consult memories that do not match the 
retrieval cues. Direct access may be thought of as a parallel search for content of all 
memories in a given store. The architecture of a content-addressable architecture is 
 
 
181 
schematized in Figure 4.1 (from Gallistel & King 2009). In this representation, the 
memories consist of three bits, and during a memory retrieval, each bit is probed with 
the retrieval cues. Those memories that match are returned. All memories are 
evaluated with respect to their fit to the retrieval cues, and then a decision about 
which memory is passed to further processing is made. The architecture requires that 
all memories be checked and for this reason the time to query memory is constant 
over the size of the search path. 
 
Figure 4.1: Architecture of a content-addressable memory, from Gallistel & King 
(2009). Memories consist of three bits, and each bit is probed in parallel for a match. 
In the present case, all memories that contain a 0 in third position are returned in 
response to a retrieval query.  
 
 
 
182 
Importantly, the constant-time property of content-addressable architectures 
holds only over dimensions such as the size of the search space or ?physical? distance 
to the target memory. They do not generally have the constant-time access property 
on other dimensions, however. For example, in the content-addressable ACT-R 
model presented in Chapter 3, access time is constant across the number of memories, 
but varies as a function of how well the cues match the target memory. Many content-
 addressable systems have this property; consider another well-known implementation 
of a content-addressable memory system, the Hopfield Net (Hopfield, 1982). 
Hopfield nets are a type of neural network architecture that stores ?memories? in the 
connections between binary units in a fully-connected, recurrent architecture. Once a 
network has been trained on a series of memory images, querying the network with a 
noisy version of the memory causes the network to ?retrieve? the desired memory by 
settling into a stable state that corresponds to the learned pattern that most closely 
matches the input. Crucially, however, the time course of retrieval (i.e. the number of 
iterations needed to reach convergence) is a function of the degree to which the 
retrieval cue matches the target memory. The closer the match between the probe and 
the target memory, the fewer iterations are necessary for the network to settle into a 
stable memory state. Presumably, if this were a model of an actual memory retrieval 
process, a greater number of iterations would correspond to a greater retrieval latency 
on some behavioral measure. This provides an important qualification of the direct 
access property of content-addresable architectures: while location and the size of the 
search path do not affect retrieval times, the degree of match to a search cue and noise 
 
 
183 
in the system can (and on many models do) affect retrieval times, as in Hopfield Nets 
or the ACT-R model (Anderson & Lebiere 1998; Lewis & Vasishth 2005).  
It is also important to note that the property of constant-time access is not 
unique to content-addressable architectures. Standard random-access memories 
(RAM) used in digital computers are also constant-time architectures. Memories are 
retrieved by loading a query onto a series of bus-lines that are probed in parallel 
against the contents of all memory stores, looking for the location that matches the 
query and returning it in a time that is functionally constant across the size of the 
search path and location of the target memory. This stands in contrast to disk media 
such as CD-ROMs or hard-drives that need to be physically traversed to reach the 
desired address. As noted by Gallistel and King (2009), RAM architectures are in 
essence architecturally identical to content-addressable architectures, and can trivially 
implement a content-addressable look-up table if desired. This is because they share 
the time-course properties of content-addressable architectures and storage 
capabilities. The primary difference between content-addressable architectures and 
RAMs is what is allowed as a retrieval cue. In RAM architectures, retrieval ?cues? 
correspond to the location of a given memory, and parallel search proceeds over all 
possible memory locations. The contents of any one location are not considered at 
access, however. Thus constant-time access is not a unique to content-addresable 
architectures. Rather than provide definitive evidence in favor of a content-
 addressable system, then, the argument from time-course provides evidence that for 
eliminating serial search architectures from consideration. 
 
 
184 
 The first argument for content-addressable architectures of comprehension, 
the argument from interference, has been made with a number of experimental 
techniques (behavioral reading measures, electrophysiological measures, off-line and 
on-line patterns of judgments). In contrast, the argument from time-course has 
typically been made exclusively with the speed-accuracy tradeoff (SAT) paradigm 
(also known as the response-signal paradigm; see Wickelgren 1977).  The SAT 
procedure probes participants for responses at cued intervals, producing a curve that 
describes subjects? accuracy as a function of time. As the decision processes tapped 
by the SAT task are by hypothesis parasitic on the retrieval of the information needed 
to support them, the time course of the decision is taken to be proportional to the time 
needed to access the information in memory.  
The SAT task has been applied to language processing by Brian McElree and 
colleagues to ask about the role that structural or linear distance plays in the speed or 
the accuracy of linguistic processing. As mentioned, studying the speed of processing 
typically involves SAT measures, which license inferences about the time course of 
retrieval mechanisms employed in sentence processing by providing a direct measure 
of the time course of information accrual. Traditional RT paradigms are limited in 
how informative they are about the dynamics of memory processes. This is because 
participants can trade speed and accuracy in a task (Wickelgren, 1977), opting to 
spend more time processing for greater accuracy on a task. For this reason, estimating 
a single mean reaction time in a given experimental condition (or even a single 
RT/accuracy pair) can obscure differences between the success of a process, on the 
one hand, and true processing speed, on the other. By providing information about 
 
 
185 
task accuracy as a function of time, the resulting SAT functions allow researchers to 
build a picture of the dynamics of task completion that dissociates the speed of 
completing the task (measured in the dynamics parameters, see below) and the 
accuracy of task completion (referred to as the asymptotic accuracy parameter). For 
example, consider Figure 4.2 (from ?ztekin & McElree 2010), which shows 
hypothetical SAT functions for two processes that differ in the speed of processing 
(4.2, bottom panel), as well as processes that differ only in their asymptotic accuracy 
(4.2, top panel). The circled points on each curve represent hypothetical reaction 
times that could have been sampled from these processes in a simple RT task. Note 
that the single RT/accuracy pairs do not reveal if the observed difference is due to an 
underlying difference in processing speed or processing accuracy. For this reason, 
time-course measures such as SAT are essential to making an argument about the 
time course of processing. SAT functions allow the experimenter to quantify both the 
speed and the probability of successfully completing a given process, and provide 
dissociable measurements of both. 
 
 
 
186 
 
Figure 4.2: Hypothetical SAT curves showing a) two processes that differ in 
asymptotic accuracy only (top panel) and b) two processes that differ in processing 
speed only (bottom panel) (figure from ?ztekin & McElree 2010). Vertical and 
horizontal lines indicate that point at which each curve is at 50% of asymptotic 
accuracy. 
 
 
 The relatively direct measure of processing speed that is measured with SAT 
curves can be used to more straightforwardly test predictions about the time course of 
processing. In the domain of memory access, these predictions are clear. The 
prediction for a parallel-access retrieval mechanism is that all relevant representations 
should be accessed with similar temporal dynamics, as all memories are probed in the 
same processing step. That single processing step is constant across the location of 
the target in the memory store, and so should result in constant access speeds across 
the growth part of the curve. Memory access that has any serial component, by 
contrast, opens up the possibility that some representations are contacted before 
others. These extra processing steps should be reflected in differences in temporal 
 
 
187 
dynamics of the SAT function. For example, in a fully serial search, these changes 
would be observed in response to changes in the structure of the search space (e.g., 
more intervening material or greater hierarchical distance to the retrieval target). In 
much work that has made architectural arguments from the time course of memory 
access, the primary contrast of interest is between content-addressable architectures 
and serial-register architectures, as the different commitments that each architecture 
makes to the time course of memory access are clear. 
 Some of the earliest applications of this argument focused on unstructured 
lists in memory (McElree & Dosher 1989). McElree and Dosher (1989) found that the 
time course of recognition for words in an unstructured list was constant across serial 
position, with the exception of the most recent item, which was assumed to be in 
focal attention. They concluded from this finding that retrieval of items in the list is 
mediated by a content-addressable, direct-access mechanism. In contrast, the speed of 
a search process should be modulated by factors like linear position or size of the 
search cohort (Sternberg 1969). This architectural conclusion is licensed in part 
because of the unstructured nature of the word list. That is, in this context, it seems 
that the identity or content of the probe word is the only information that is available 
to be used in accessing the memorized list at the point where participants are required 
to make a recognition judgment. Thus is linear position or size of the list do not 
impact speed, the word identity itself?its content?must be guiding access. 
However, an important difference between unstructured memory lists and sentences 
concerns the availability of structure to guide memory access. As argued above, the 
nature of the cues used to access memory control how retrieval events proceed in 
 
 
188 
sentence comprehension. The rich linguistic structure in the form of possible 
morphological, syntactic and semantic cues makes it more difficult to directly draw 
architectural conclusions from this argument in the domain of syntactic processing. 
For this reason, although the debate often contrasts content-addressable and search-
 based memory architectures, in the domain of sentence comprehension it is difficult 
to address this question without also considering the nature of the information used to 
index and access memory. 
 In order to start addressing the question of memory architecture in the domain 
of sentence processing, McElree and colleagues have investigated the time course of 
dependency completion for a number of different linguistic dependencies. These 
include filler-gap dependencies (McElree, 2000; McElree et al., 2003), subject-verb 
thematic dependencies (McElree et al., 2003), pronoun antecedent resolution (Foraker 
& McElree, 2007), and verb phrase ellipsis (Martin & McElree, 2008, 2009). By and 
large, the results from these studies point in the same direction. They have generally 
indicated that the retrieval of information during sentence processing occurs at a 
constant rate regardless of the distance from the probe point to the target, suggesting a 
general content-addressable (or parallel-access) architecture during sentence 
processing. 
 One influential example of the argument of time course was presented by 
McElree, Foraker and Dyer (2003). They constructed experimental sentences as in 
(4.1). The topicalized NP the scandal needs to be interpreted as the object of the most 
deeply embedded verb in all examples. McElree and colleagues interpolated varying 
amounts of clausal material between the fronted NP and the gap position, ranging 
 
 
189 
from no intervening material in (4.1a) to two intervening clauses in (4.1c). In doing 
so they increased both the linear and the hierarchical distance to the gap site. They 
reasoned that if the difficulty associated with longer filler-gap distances was due to an 
increase in search path in conjunction with a serial access memory architecture, then 
the difficulty associated with longer filler-gap distances should affect the speed 
(dynamics) portion of the resulting SAT functions. On the other hand, a content-
 addressable architecture should show no effects of distance on retrieval dynamics. 
Instead, only the terminal accuracy (asymptote; probability of correct retrieval) 
should be affected. 
 
(4.1) a. It was the scandal that the celebrity relished e. 
b. It was the scandal that the model believed that the celebrity 
relished e. 
c. It was the scandal that the model believed that the journalist 
reported that the celebrity resished e. 
 
 McElree and colleagues found that access time (as indexed by the SAT 
curve?s dynamics parameters) for the filler was constant across all three examples in 
(4). The distance between the filler and gap was only evident in the pattern of 
terminal accuracies: the longer the distance, the less likely participants were to 
retrieve the correct filler to interpret at the gap site. This finding was compatible with 
a content-addressable memory access mechanism, in that it displayed constant-time 
access and loss of memory fidelity with increasing temporal decay or interference. 
McElree and colleagues found similar results across a number of other constructions 
including subject-verb integration (the book ripped versus the book that the editor 
admired ripped) and double gap topicalization constructions (this is the album that 
 
 
190 
the stamps were difficult to mount e in e versus this is the album that the stamps 
which obviously angered the fussy collector were difficult to mount e in e).  
 One possible objection to these results, noted by McElree et al (2003), Martin 
& McElree (2008) and Wagers (2008) is that these results rest on data from the 
processing of dependencies that are often thought to involve predictive resolution 
strategies. The most well-known of these hypotheses is the ?active-filler? strategy for 
processing wh-movement dependencies, which posits that the parser actively tries to 
find a gap for a filler held in memory (Wanner & Maratsos 1978; Frazier 1987; 
Frazier & Flores d?Arcais 1989). If this is the case, then the interpretation of these 
data is unclear. If the filler is held in a special store or engages special parsing 
routines (as in Wanner & Maratsos 1978), then it is not clear that the filler actually is 
more ?distant? in (4.1c) than (4.1a), making it difficult to argue that the constant 
access dynamics implicate a content-addressable architecture. Furthermore, there are 
independently motivated reasons for thinking that the intermediate gap sites (the 
[spec,CP] positions along the path between wh-element and its gap site) might 
contain some representation of the moved constituent, with support for this coming 
both from theories of grammatical organization (Chomsky 1973, 1977; McCloskey 
2001) and processing considerations (Frazier & Clifton 1989; Gibson & Warren 
2004). If this line of reasoning is correct, then it is not clear that there is any 
difference in the distance between the filler and gap in the examples in (4.1). 
 To address this concern and provide support for content-addressable 
architectures, it is necessary to investigate dependencies that are not amenable to 
anticipatory or predictive processing strategies. In order to make a stronger case for 
 
 
191 
content-addressable architectures in comprehension, Martin and McElree (2008, 
2009) examined the processing of verb phrase ellipsis (VPE) constructions like those 
in (4.2).  
 
 (4.2)   a. The editor [VP admired the author?s writing], but the critics did 
not e. 
b. The editor [VP admired the author?s writing], but everyone at 
the publishing house was surprised to hear that the critics did 
not e. 
 
 Unlike the filler-gap dependencies tested in earlier SAT studies (McElree 
2000; McElree et al 2003), these VPE examples do not generally allow 
comprehenders to adopt a predictive strategy for resolving the antecedent of the 
ellipsis: upon completing a VP, it is not explicitly marked as participating in an 
upcoming ellipsis dependency. Despite the entirely retrospective nature of the VPE 
dependency, Martin and McElree again found that the antecedent VP was retrieved in 
constant time, no matter how distant it was from the ellipsis site. Thus the findings of 
these studies confirm and extend those of prior studies: extra hierarchically 
intervening material between the antecedent and the ellipsis site did not produce any 
differences in the dynamics of memory retrieval. Instead of impacting processing 
speed, distance between the ellipsis site and the antecedent VP instead led to a lower 
probability of retrieval, as reflected in lower accuracy (a lower asymptotic portion of 
the curve). 
 Similar findings have been found for filler-gap dependencies (McElree & 
Griffith 1998; McElree 2000; McElree et al 2003), VPE (Martin & McElree 2008, 
2009), and coreferential antecedent-pronoun relations (Foraker & McElree 2007), 
 
 
192 
among others. The repeated failure to find effects of the size of the search space in the 
SAT paradigm constitutes the argument from time-course against serial search 
memory architectures. For the range of dependencies considered, McElree and 
colleagues conclude that a content-addressable architecture best captures the observed 
time course evidence. This conclusion aligns nicely with the argument from 
interference, which provides a qualitatively different source of evidence that points in 
the same direction: during sentence comprehension, access to memory (the contents 
of the parse) is gated by a procedure that is constant-time (with respect to search path) 
and prone to interference, two defining features of content-addressable memory 
systems (see also Lewis et al 2006). 
 
Chinese long-distance anaphors 
  
 On the current evidence, it has been argued that long-distance dependencies in 
natural language do not engage a serial memory search (McElree et al 2003; Martin & 
McElree 2008). In all long-distance dependencies that have been considered, the 
speed of processing is constant across distance and the size of the search path. 
However there is one significant point of variation between the dependencies that 
were tested that could be crucial to understanding the pattern of results across studies. 
As mentioned, dependencies such as wh-movement and clefting that have been 
studied using the SAT paradigm are prospective dependencies. In prospective 
dependencies, the left edge of the dependency clearly signals information that will be 
relevant farther downstream in parsing. In wh-movement, for example, the fronted 
 
 
193 
wh-word signals to the comprehender that a gap will occur at a later point in the 
sentence, and thus potentially allows for preprocessing strategies that may obviate the 
need for memory retrieval (as already noted in McElree et al., 2003). Not all 
dependencies that have been studied using SAT methods have this property: in 
particular, verb-phrase ellipsis is a fully retrospective dependency that cannot be 
reliably anticipated ahead of the integration site (a point noted by Martin & McElree, 
2008). However, there are a number of independent reasons why a structured search 
might not be deployed in finding the antecedent VP for VPE. First, there are no 
structural constraints on the location of the antecedent verb phrase (Johnson 2001), 
and so there is no reason to suspect that structural information should be used to 
narrow the search space. Second, and perhaps more importantly, in the case of Martin 
& McElree?s studies (2008, 2009), the experimental design only included one 
potential, complete VP antecedent. For this reason, although VPE findings do suggest 
a content-addressable architecture that does not need to serially consider irrelevant 
material, they do not rule out a serial consideration of syntactically licit candidate 
positions.  
 For these reasons, in order to probe for the existence of structured search, it is 
crucial to test a dependency that is retrospective, and for which multiple syntactic 
positions might contain the tail of the dependency. One such dependency is the 
relation between the Mandarin long-distance reflexive ziji and its antecedent. Upon 
reaching the reflexive element ziji, comprehenders need to initiate a search for an 
antecedent. This is an entirely retrospective process, as there are no cues prior to ziji 
that signal the presence of a dependency (with the exception of contextual bias and 
 
 
194 
inherently reflexive verbs in Chinese, see Jin 2003; Li & Zhou 2010). There are a 
number of key properties of ziji?s interpretation that make structured search a 
potentially useful option. These are i) long-distance binding possibilities; ii) syntactic 
constraints on possible linguistic antecedents; and iii) blocking effects. In order to 
satisfy these constraints, ziji must consider a potentially unbounded search space, and 
within that space, search must be limited to antecedents that occupy particular 
syntactic positions relative to both ziji and to other potential antecedents. This 
situation provides one context where a structured search would provide an efficient 
way to satisfy these constraints during the construction of a dependency between ziji 
and its antecedent. 
 The search space for ziji is unbounded because, unlike the English reflexives 
himself/herself, ziji does not require that its antecedent be in the same clause. Ziji is 
an example of the cross-linguistically well-attested class of long-distance reflexives. 
Long-distance reflexives are pronominal elements that have structural requirements 
on their linguistic antecedents, but unlike English reflexives, long-distance reflexives 
may be bound outside of their minimal clause  (B?ring 2005). Local and long-
 distance binding possibilities for ziji can be seen in (1) and (2). 
 
 (4.1)  Lisi  nongshang-le  ziji  
   Lisi harm-PERF self 
   ?Lisi harmed herself? 
 
 (4.2)  Zhangsani  shuo Lisij  nongshang-le  zijii/j 
   Zhangsan says Lisi harm-PERF self 
   ?Zhangsan says that Lisi harmed him / herself? 
 
 Like many long-distance reflexives, ziji imposes a number of constraints on 
potential linguistic antecedents (B?ring 2005; Huang et al 2009). There are significant 
 
 
195 
syntactic constraints placed on antecedents: they must be subjects and they must be 
contained in the same clause or a higher clause than ziji (Huang & Liu, 2001). In 
addition to these syntactic constraints, there are a number of discourse-pragmatic 
constraints on the use of ziji. Antecedents must be animate and sentient, and must be 
prominent in the current discourse (Xue, Pollard & Sag 1994; Huang & Liu 2001). In 
the absence of an appropriate antecedent in the immediate sentential context, ziji may 
refer to the speaker, presumably as a reflex of the prominent discourse status that is 
automatically afforded to the speaker (Kuno 1987; Huang & Liu 2001). There is 
ongoing debate about the relative prominence of discourse-pragmatic and structural 
licensing conditions on ziji (Huang, Cole & Hermon, 2006; Huang et al 2009), but it 
is relatively uncontroversial that both sorts of constraints are operative, and that 
accurate resolution of the antecedent-anaphor dependency requires the comprehender 
to systematically exclude inaccessible referents from consideration.  
 In addition to limiting a potentially unbounded search space to a subset of 
syntactically or pragmatically accessible antecedents, correct resolution of the 
antecedent of ziji also requires verifying that potential antecedents stand in particular 
syntactic relations to each other. Consider the following sentences (from Huang & 
Liu 2001: p.2-3; example (5) modified from their (11), p.6): 
 
 (4.3)  Zhangsani  renwei Lisij hen zijii/j 
   Zhangsan think Lisi hate ziji 
   ?Zhangsan thinks that Lisi hates himself/him.? 
 
 (4.4)  Zhangsani  renwei nij hen ziji*i/j 
   Zhangsan think you hate ziji 
   ?Zhangsan thinks that you hate yourself/*him.? 
 
 
 
196 
 (4.5)  Nii  renwei Zhangsanj hen ziji?i/j 
   You think Zhangsan hate ziji 
   ?You think that Zhangsan hates himself/?you.? 
 
 Although the matrix subject Zhangsan is generally considered an acceptable 
antecedent for ziji as in (3), when the local subject is a second-person pronoun ni, 
then this binding possibility is no longer considered acceptable (4). These examples 
show an instance of a person blocking effect, where the presence of a first- or second-
 person antecedent blocks binding from higher subjects (Huang & Liu 2001). 
Sentences (3) and (4) illustrate the blocking effect for ziji: in (4), assessing the feature 
content of the local subject is necessary to determine whether or not the matrix 
subject is an acceptable antecedent. Similarly, it has been reported that singular 
subjects block binding from higher plural subject positions (Tang 1989; Huang & 
Tang 1991). Thus, it is not only the presence of a pronoun with first- or second-
 person features that impacts the availability of other subjects, but also its relative 
position to those subjects. 
 Experimental evidence to date suggests that there is a locality bias in 
processing ziji. For example, Li & Zhou (2010) provide ERP evidence that long-
 distance binding of ziji elicits a P300/600 component relative to local or ambiguous 
binding of ziji. I take up this finding in more detail in our discussion of Experiment 7. 
In addition to ERP results, a number of cross-modal priming studies have shown a 
bias towards reactivating the local antecedent upon reaching ziji. In one study, Gao, 
Liu & Huang (2005) showed that in sentences such as ?the teacher asked the 
journalist to respect ziji?, reaction times for probes related to the object the journalist 
were faster than reaction times for probes related to the matrix subject the teacher. 
 
 
197 
Liu (2009) replicated and extended this result by varying the SOA of the probe word 
relative to ziji, probing for lexical decisions at 0ms, 160ms and 370ms after the offset 
of ziji. In a very suggestive finding, reaction times to probes related to the local 
subject were faster at the 0ms SOA, but this pattern reversed at the 160ms probe 
latency, suggesting greater reactivation for the long-distance antecedent. For these 
same two SOAs, the opposite pattern of facilitation was observed for the pronoun ta. 
At the longest SOA, both antecedents were equally activated for both pronominals. 
This result is suggestive of a structured search, with early consideration of the local 
antecedent and relatively delayed consideration of the long-distance antecedent only 
for ziji, and not for other pronominal elements. However, the reaction times for the 
lexical decision task do not provide a direct window on the time course of activating 
antecedents for ziji. 
 Taken together, the structural constraints and blocking configuration 
constraints on antecedents provide a good context to observe a structured search, if 
structured search is indeed an option for the parser to pursue. Structured search is 
here understood as a process in which potential antecedent positions are sequentially 
considered as potential antecedents, and this follows naturally from the structured 
access hypothesis and the serial retrieval of memories assumed by cue-based parsing 
models like ACT-R. This is because if only syntactic information is used to access 
memory, then multiple positions may need to be considered before an acceptable, 
feature-matched antecedent is found. Although the need for speed in language 
comprehension is an important factor that may favor parallel-access mechanisms (cf. 
Lewis et al 2006), the parser?s ability to reliably recover the correct interpretation 
 
 
198 
may also be a factor in determining an optimal online strategy for constructing a 
linguistic dependency. By limiting consideration to specific structural positions, the 
task of finding an antecedent for ziji can arguably more straightforwardly respect 
structural and blocking constraints. This is because the structural fidelity is 
presumably enhanced if only structural cues are deployed initially, and the blocking 
constraints can be naturally implemented by terminating the search if the local subject 
contains a ?blocking? feature specification. A schematic comparison of the antecedent 
retrieval process is shown in Figures 4.3 and 4.4. If only structural cues are used to 
access memory, then when the correct feature-matched antecedent is distant, multiple 
operations are needed to recover the correct antecedent, as in Figure 4.3. If, however, 
all semantic and structural cues are used in parallel to retrieve the antecedent, then 
only a single retrieval step is needed to retrieve the antecedent. This is shown in 
Figure 4.4, where the enriched cue set allows the antecedent to be recovered in a 
single access step. 
 
 
 
199 
 
Figure 4.3: Example of a structured search process for finding ziji?s antecedent in the 
sentence Lisi shuo fengbao hai-le ziji ?Lisi said the storm harmed him?. The 
hypothetical structural cues do not allow comprehenders to rule out consideration of 
the local subject fengbao ?storm?. Thus comprehenders must evaluate multiple 
subject positions in the search for the correct antecedent Lisi. This structured access 
predicts that processing time should grow with the number of subject positions that 
need to be evaluated. 
S
 NP VP
 Lisi V S
 say NP VP
 storm V NP
 harm ziji
 +subject
 +c-command
 +subject
 +c-command
 
 
200 
 
Figure 4.4: Example of feature-based access for finding ziji?s antecedent in the 
sentence Lisi shuo fengbao hai-le ziji ?Lisi said the storm harmed him?. The mixture 
of structural and semantic cues allow direct access to the correct antecedent Lisi. 
Feature-based access predicts that processing time should be constant with the 
number of subject positions that need to be evaluated. 
 
 If even the retrospective, structurally constrained ziji-antecedent dependency 
is resolved using a feature-based, parallel-access mechanism, then a strong case can 
be made for such a feature-based access mechanism across all levels of sentence 
processing. If, on the other hand, structured search is implicated in the processing of 
ziji, then it would suggest that the parser can engage in a structured search of 
linguistic memory for at least some linguistic long-distance dependencies, due to a 
selective limiting of the search cues to only structural cues. 
 
Experiment 6: SAT Evidence 
 
S
 NP VP
 Lisi V S
 say NP VP
 storm V NP
 harm ziji
 +subject
 +c-command
 +animate
 +sentient
 +source of communication
 
 
201 
 Experiment 6 employed the multiple-response speed-accuracy tradeoff (MR-
 SAT) procedure to provide time-course evidence about whether or not ziji accesses all 
potential antecedent positions in parallel, or whether or not certain positions are 
checked before others. Figure 4.3 presented a schematized comparison between a 
serial, structured access to potential antecedents, and Figure 4.4. shows the alternative 
mechanism, direct-access using semantic features. I follow the same logic employed 
in prior SAT studies (McElree 2000; McElree et al 2003; Foraker & McElree 2007; 
Martin & McElree 2008, 2009). If all possible antecedent positions are considered in 
parallel for ziji, then local and non-local antecedents should show similar processing 
speed (identical dynamics parameters in the SAT function). If, on the other hand, 
certain positions are checked prior to other positions, then we should observe 
different processing speeds for local and long-distance bindings. I also predict that 
due to the nature of the blocking effects described above, any speed advantage should 
favor local antecedents. The availability of the long-distance binder depends on the 
nature of the local binder, and so a reasonable strategy would be to check a local 
subject before a non-local subject, in order to determine whether the non-local subject 
position should even be considered. Furthermore, the linear recency of the local 
subject may itself confer an advantage over the long-distance subject. 
  
Participants 
 
Twenty college students from Beijing Normal University participated in the 
experiment. Data from 2 participants were excluded due to atypical (sigmoidal) SAT 
 
 
202 
curves. The remaining 18 participants included 10 females, and had a mean age of 
23.5 years. Each participant completed six 1-hour sessions spaced at least a day apart, 
in addition to a 1-hour practice session for familiarization with the multiple-response 
speed-accuracy tradeoff procedure (MR-SAT; see below). All participants were 
native Mandarin Chinese speakers and had normal or corrected-to-normal vision. 
They were paid 35 RMB per hour for their participation in the experiment.  
 
Materials 
 
Experiment 6 investigated the processing of Mandarin sentences that contained a 
matrix attitude verb and an embedded transitive complement clause. Two 
experimental factors were manipulated. One was the position of an animate subject; it 
was either the subject of the matrix clause, the subject of the local clause, or not 
present. The second factor was the identity of the object, which was either ziji, an 
acceptable definite NP, or an unacceptable definite NP. Since ziji forms a long-
 distance binding dependency with its antecedent, the predictions of a structured 
search are that the position of the target antecedent should impact processing time. 
Control conditions with an acceptable definite NP (thus with no binding dependency) 
were included to ensure that any processing speed differences that are observed are 
not due to baseline differences due to the sentential context. Because a definite NP 
object does not form a binding dependency with its animate subject, no processing 
speed differences should obtain when this factor is manipulated for control 
conditions.  
 
 
203 
This design provided three critical reflexive conditions designed to investigate 
the processing of ziji (conditions 1-3 in Table 1 below). Based on the position of the 
animate subject, ziji either took a long-distance antecedent (LD antecedent condition; 
1), a local antecedent (local antecedent condition; 2), or had no antecedent in the 
sentence (no antecedent condition; 3). In these conditions, the animacy of the subjects 
was the factor that controlled which antecedent ziji was forced to take. As noted 
above, an animate NP in either the main clause or embedded subject NP position can 
function as a grammatical antecedent for ziji. In addition to the critical ziji condition, 
two control sentences were constructed with definite descriptions instead of ziji in the 
embedded object position (LD control, local control, no antecedent control; 
conditions 4-6). The well-formed control conditions replaced ziji with a full NP that 
was a plausible object of the embedded verb (e.g., the batsman in conditions 4-6 
below); whereas the unacceptable control conditions replaced ziji with a full NP that 
was an implausible object of the embedded verb (e.g., glasses in conditions 7-9 
below), resulting in a semantic anomaly. The anomaly conditions (conditions 7-9) 
and the conditions with no animate subjects (conditions 3,6 and 9) were included for 
purposes of d? scaling, as described below, and so they were not directly analyzed. 
The primary experimental contrast was the effect of subject animacy on the speed and 
accuracy of processing sentences that contained ziji, and sentences that contained 
acceptable definite NP objects. 
Therefore, each set of items consisted of three critical (conditions 1-3) and six 
control sentences (conditions 4-9). The control sentences additional helped to prevent 
participants from forming strategies based on the pattern of animate and inanimate 
 
 
204 
referents that preceded the critical object NP. Thus, because of the structure of the 
control conditions, neither the presence of ziji nor the acceptability of the 
continuation was predictable from the preamble. This was done to ensure that all 
processing measures reflect processes initiated at the object NP itself. All three 
conditions consisted of a main clause that contained a verb of reporting, and an 
embedded transitive clause. Additionally, to increase complexity and difficulty of the 
task, a temporal adverbial clause was interpolated between the embedded subject and 
the embedded verb. In all conditions, an animate NP was also used as the subject of 
the temporal adverbial phrase. However, since it occupied a structural position that 
does not c-command ziji, it was not a grammatical antecedent for ziji. In the local 
antecedent and control conditions, the main clause NP was always a ?media? noun 
(e.g. book, documentary, memo) to ensure compatibility with the meaning of the 
main clause verb. In the no antecedent condition, both structurally accessible NPs 
were inanimate, and therefore this condition did not contain a grammatical antecedent 
for ziji. None of the inanimates used in any position could be construed 
metonymically; metonymic interpretations of inanimates (i.e. the newspaper being 
used to refer to the employess of the newspaper) may be used as antecedents for ziji.  
Forty sets of the 9 sentence types (5 acceptable and 4 unacceptable) were 
generated. The 360 sentences were equally distributed in 6 presentation lists, one for 
each of the 6 sessions, to minimize the repetition of content material within a session. 
Crucially, no two instances of ziji sentences (conditions 1-3) from the same set were 
included in a single presentation list. Within a session, each participant viewed 206 
sentences, of which 60 were drawn from the current study. Since only one third of 
 
 
205 
target sentences contained ziji, the critical ziji conditions comprised around 10% of all 
sentences within and across sessions. The order of presentation within a session was 
randomized.  
 
Procedure 
 
I employed the multiple-response SAT paradigm, following Wickelgren, Corbett & 
Dosher (1980). Stimulus presentation, timing, and response collection were all carried 
out on a personal computer using the Linger software by Doug Rohde (available at 
http://tedlab.mit.edu/~dr/Linger/). Each trial began with a 500 ms fixation cross 
presented in the center of the screen. Each word appeared in the center of the screen 
for 400 ms, followed by 200 ms of blank screen. All words were presented using 
simplified Chinese characters, and the last word of each sentence was marked with a 
period (?). Immediately prior to the onset of the final word, a series of 18 auditory 
response cues (50 ms, 1000 Hz tone) was initiated. The cues occurred every 350 ms, 
and the final word of the sentence remained on the screen. Participants were trained 
to initially respond by pressing both response keys simultaneously to indicate an 
undecided response. They were then trained to give a response after each tone, and to 
switch their response to either the ?accept? or ?reject? key as soon as they could. 
Importantly, they were also trained to modify their responses if their assessment 
changed. During the 1-hour practice session, participants were told that some of the 
sentences were complex, but nevertheless were meaningful sentences. Each 
 
 
206 
participant performed six 1-hour sessions, and in each they saw one of the lists of 
materials. The order of lists was randomized across participants. 
 
 
 
# Condition Example 
1 LD antecedent ???    ?? ???? [? ?? ?? ?? ?? ???] ??? ?? 
Coach Zhang say [that report [when team not  perform well time] underestimate ziji] 
?Coach Zhang says that that report underestimated self when the team was doing poorly.? 
2 Local antecedent ??? ?? ??? [? ?? ?? ?? ?? ???] ??? ?? 
Auto-biography say [coach Zhang [when team not perform well time] underestimate ziji] 
?The auto-biography says that coach Zhang underestimated self when the team was doing 
poorly.? 
3 No antecedent *??? ?? ???? [? ?? ?? ?? ?? ???] ??? ?? 
*Auto-biography say [that report [when team not perform well time] underestimate ziji] 
* ?The auto-biography says that that report underestimated self when the team was doing 
poorly.? 
4 LD control  ??? ?? ???? [? ?? ?? ?? ?? ???] ??? ????? 
Coach Zhang say [that report [when team not  perform well time] underestimate that batsman] 
?Coach Zhang says that that report underestimated the batsman when the team was doing 
poorly.? 
5 Local control  ??? ?? ??? [? ?? ?? ?? ?? ???] ??? ????? 
Auto-biography say [coach Zhang [when team not perform well time] underestimate that 
batsman] 
?The auto-biography says that coach Zhang underestimated the batsman when the team was 
doing poorly.? 
6 No antecedent control ??? ?? ???? [? ?? ?? ?? ?? ???] ??? ????? 
Auto-biography say [that report [when team not perform well time] underestimate that 
batsman] 
?The auto-biography says that that report underestimated the batsman when the team was 
doing poorly.? 
7 LD control anomaly *??? ?? ???? [? ?? ?? ?? ?? ???] ??? ?? 
*Coach Zhang say [that report [when team not  perform well time] underestimate glasses] 
*?Coach Zhang says that that report underestimated the glasses when the team was doing 
poorly.? 
8 Local control anomaly *??? ?? ??? [? ?? ?? ?? ?? ???] ??? ?? 
*Auto-biography say [coach Zhang [when team not perform well time] underestimate glasses] 
* ?The auto-biography says that coach Zhang underestimated the glasses when the team was 
doing poorly.? 
9 No antecedent control 
anomaly 
*??? ?? ???? [? ?? ?? ?? ?? ???] ??? ?? 
*Auto-biography say [that report [when team not perform well time] underestimate glasses] 
* ?The auto-biography says that that report underestimated the glasses when the team was 
doing poorly.? 
 
 
 
 
207 
Table 4.1: Summary of conditions in experiment: Critical ziji conditions. Critical 
conditions are 1-2 and 4-5; conditions 3 and 6-9 were included for purposes of d? 
scaling (see text). 
 
Data Analysis 
 
To derive the full time-course information in SAT analysis, d? scores are calculated 
by comparing the judgments in an acceptable condition and a closely matched 
unacceptable condition at each of the response points, within participants. The 
resulting series of d? values at each time point t is fit using a shifted exponential 
function: 
 
  (4.6)  d ' = ?(1? e?? (t?? ) )  , t > ?,  
d ' = 0    , otherwise  
 
Here, d? is the standard measure of discrimination: d? = z(hits)-z(false alarms) 
(Wickens, 2001). The shifted exponential in (4.6) describes the growth of accuracy 
over time t (in ms) with three parameters: asymptote (?), rate (?), and intercept (?). 
The current experiment used common scaling in its design: in order to derive the d? 
scores for LD and local antecedent conditions, their hit rates for the critical conditions 
(Conditions 1 and 2) were scaled against the pooled false alarm rate at the 
corresponding time lag from the three ungrammatical controls (Conditions 7-9), 
following McElree & Griffith (1998). Both critical ziji and control conditions were 
scaled using this pooled scaling, to allow a more straightforward comparison of any 
potential differences in the resulting time courses that may arise. All differences 
 
 
208 
between ziji and control conditions stem from differences in the hit rates between the 
two. One reason for adopting this measure, instead of scaling against the no-
 antecedent ziji condition was the somewhat high acceptability of the no antecedent 
condition (see below). If the no antecedent condition were used for scaling purposes, 
the false alarm rate used in the d? calculations would have been high, leading to low 
d? scores (less than 1 in most cases). For such low scores, small variations in hit rate 
would case very large changes in the observed d?, leading to unreliable SAT function 
estimates. An additional benefit is that pooled scaling produced SAT curves for the 
ziji conditions and the control conditions with comparable asymptotic accuracy, 
ensuring that neither floor nor ceiling effects could be responsible for any observed 
effects. Thus, this scaling measure allows direct comparison between the critical ziji 
conditions and their non-ziji acceptable control counterparts (conditions 4 and 5), and 
we can straightforwardly estimate the dynamics of the acceptable judgment alone 
(i.e., the successful completion of the dependency). 
 In order to determine whether the SAT functions for these conditions differed 
in asymptote (?), rate (?), or intercept (?), the analysis proceeded in two steps: a 
model selection analysis and a parameter estimation analysis (Li & Smith 2009). In 
the model selection analysis, the best fit model was determined using the adjusted R2-
 statistic using a hierarchical model-testing scheme over the averaged and individual 
data, an approach pursued in prior work on SAT in sentence comprehension (McElree 
et al 2000, 2003; Foraker & McElree 2007; Martin & McElree 2008, 2009). In the 
parameter estimation analysis, only the fully specified model was considered, and any 
differences between the critical conditions on the parameters of interest were assessed 
 
 
209 
using familiar hypothesis testing measures. In all analyses the SAT function given 
above was fit to the measured d? values at each time lag. In order to obtain parameter 
estimates, it is necessary to fit non-linear regressions of the SAT function (1) against 
the observed d? score. In the present study, I employed Gibbs sampling, a Monte 
Carlo method for approximating the posterior of a distribution of two or more random 
variables. Gibbs sampling is often used as a form of Bayesian statistical inference 
(Gelman & Hill 2005). For the current analysis, Gibbs sampling was used to estimate 
the posterior distributions over the three SAT parameters, and the best-fit model for a 
given set of data was chosen by selecting the median of the resulting posterior 
distributions. Note that this method of fitting SAT curves contrasts with the method 
often employed in the psycholinguistic SAT literature (STEPIT fits, McElree & 
Griffith 1998; McElree et al. 2000, 2003; Martin & McElree). In this approach, the 
SAT function is fit with an iterative hill-climbing algorithm (Reed, 1976; similar to 
STEPIT, Chandler, 1969), and fit quality was assessed by an adjusted R2-statistic, 
which measures the variance accounted for by the fit, with a penalty for an increasing 
number of model parameters. In almost all cases the two methods provide identical 
results, and I primarily employ the Gibbs sampling method because in addition to 
providing information about the estimate of a value, it provides an estimate of the 
variance associated with that value. However, to ensure that the estimation technique 
did not change the qualitative pattern of results, I also report STEPIT fits in the 
parameter estimation section. A direct comparison of STEPIT and Bayesian fits, as 
well as code for implementing the models here, is available at 
http://people.umass.edu/bwdillon/. 
 
 
210 
 For all Monte Carlo simulations, 4000 samples were used as a burn-in period, 
followed by 15,000 iterations of the sampler. 3 parallel sampling chains were run in 
each simulation, and convergence was checked by evaluating the potential scale 
reduction factor (?r-hat?) statistic (Gelman & Hill 2005). For all reported parameter 
values, the potential scale reduction factor was effectively 1, indicating satisfactory 
convergence of the MCMC chains.  
 In addition to model-fitting on the d? data, analysis was performed on 
participant and item mean final acceptance rates (the empirical accuracy), which was 
obtained by taking the average rate of acceptance over the last response point.  
 Data from two participants were excluded due to unreliable dynamics 
estimates. The empirical d? scores from these participants appeared to be better fit by 
a sigmoidal rather than an exponential function, leading to unrealistically large and 
unreliable differences in the critical conditions in the crucial intercept and rate 
parameters when fit with the SAT function in (4.6).  
 
Empirical Accuracy Analysis 
 
 The rate of acceptance of for the critical ziji conditions was 85% for the LD 
antecedent condition, 81% for the local antecedent condition, and 48% for the no 
antecedent condition. The rate of acceptance for the corresponding acceptable control 
conditions was 91%, 87% and 91%. A mixed-effects logistic linear regression was fit 
with orthogonal fixed-effect contrasts for locality (comparing long-distance versus 
local antecedent conditions) and binding (comparing both ziji conditions with 
 
 
211 
antecedents to the no antecedent condition), as well as random slopes for subject, 
item, and session. For critical ziji conditions, there was a significant effect of binding 
(? = - 1.3, z = -17.5, p < .0001) and a marginal effect of locality (? = .28 , z = 1.94, p 
< .06). For control conditions, the observed empirical accuracy  was a significant 
effect of locality only (? = .70 , z = 3.97, p < .0001). For unacceptable controls, 
accuracy for long-distance, local, and no antecedent controls was 99%, with no 
differences between these conditions. 
 
Model Selection Analysis 
 
For the model selection analysis, individual participants? data was fit with a series of 
nested models separately for control and ziji conditions. This analysis compared long-
 distance and local animate configurations. The model-fitting analysis pitted a series of 
nested models (including shared or separate parameters for the two conditions of 
interest for each of the intercept, rate and asymptote) against each other on adjusted 
R2, following McElree et al (2003) and Li & Smith (2009). This was done by 
participants to determine if extra parameters led to a significantly better model fit, as 
measured by adjusted R2-statistic. Adjusted R2 (4.7) gives an estimate of the variance 
accounted for by the (non-linear) regression against the SAT curve, weighted by the 
number of parameters used in constructing that curve (k). In (4.7), d refers to the 
observed d? value, ?d  refers to the predicted d? value, and n refers to the number of 
data points. 
 
 
 
212 
  (4.7)  R2 =1?
 (di ? ?di )2 / (n ? k)i=1
 n?
 (di ? d )2 / (n ?1)i=1
 n?
  
 
 For the critical ziji conditions, adding separate asymptote parameters for LD 
and local antecedents led to a reliable increase in adjusted R2 across participants (?? 
= 0.03 ? 0.01; t(17) = 3.9, p < 0.01). Compared to the two-asymptote baseline, the 
addition of an extra rate parameter led to a small but reliable increase in adjusted R2 
(?? = 0.002 ? 0.0005; t(17) = 4.3, p < 0.001). Any model that included an extra 
intercept parameter led to a significantly poorer fit on adjusted R2. Thus the best 
fitting model across participants allotted separate asymptotes and separate rate 
parameters for local and long-distance antecedents for ziji (2?-2?-1?).  
 For control conditions, the addition of an extra asymptote for long-distance 
and local controls led to a marginally significant increase in adjusted R2 (?? = 0.032 
? 0.016; t(17) = 2.1, p < 0.055). Neither additional rate parameters nor additional 
intercept parameters for LD and local control environments led to a reliable increase 
in adjusted R2. Across participants, the best-fitting model for control conditions had 
shared dynamics parameters for LD and local conditions, but separate asymptotes 
(2?-1?-1?).  
  The model-fitting analysis suggests that the critical ziji conditions have 
different processing speeds, as allotting separate rate parameters ? to LD and local 
antecedent conditions led to a reliable increase in adjusted R2. However, with the 
model-fitting approach one might be concerned by the possibility of over-fitting, a 
concern magnified by the small differences in adjusted R2 among the best fit models 
 
 
213 
for any given participant?s data (Liu & Smith 2009). For this reason it is important to 
ensure that the parameter estimates across participants display a consistent ordering 
(see, e.g., McElree et al 2003) by directly analyzing the estimated SAT parameters 
across participants. 
 
Parameter Estimation Analysis 
 
 The focus of the parameter estimation analysis, following Liu & Smith 
(2009), was the magnitude of the difference between parameters for local and long-
 distance antecedent configurations. Here this was done by fitting fully saturated 
models (2?-2?-2?) for each participant and testing for consistent ordering of 
parameters across participants. In addition to presenting absolute parameter values for 
the processing of local and LD conditions, I also present the differences between the 
two values of a given parameter across these conditions. This can be understood as 
the advantage enjoyed by one condition over the over in speed (for differences in 
dynamics parameters) or accuracy (for differences in the asymptote).  
 Parameter estimates for each of the participants included in the analysis were 
estimated using the Bayesian method, although I also present fits from the hill-
 climbing algorithm employed in other SAT work (McElree 2000; McElree et al 2003; 
Martin & McElree 2008). For Monte Carlo fits, the median of the resulting posterior 
distribution for each participant was chosen as the estimate and submitted to further 
analysis. Figure 4.7 summarizes the differences in all dynamics parameters for both 
critical ziji and control conditions, for each individual participant. Values above 0 in 
 
 
214 
the figures indicate a dynamics advantage for local antecedent configurations. In 
addition to rate and intercept estimates, I report the composite processing measures 
speed (?-1 + ?), a measure that helps to guard against parameter tradeoffs in the 
dynamics estimates (Carrasco, Giordano & McElree, 2006).  In all cases, I present the 
best estimate for the size of the difference between the parameters in local and long-
 distance configurations. For this comparison, we take the long-distance configuration 
to be the ?basic? configuration; thus, all parameter differences presented below reflect 
the dynamics advantage in local antecedent configurations.  
 The average advantage for local antecedents in the ziji and control conditions 
is summarized numerically in Table 4.2. Fully saturated model fits to average data are 
presented in Figures 4.5 and 4.6, scaled to proportion of asymptote in order to 
highlight the differences in the growth portion of the SAT curve. Figures 4.7-4.10 
show the mean parameter values for accuracy, rate, intercept and speed for both ziji 
and control conditions. The results from both methods of parameter estimation are 
largely in agreement. However, the STEPIT fits appear in general to trade greater 
speed in the rate parameter for local configurations for later intercepts; no such trade-
 off occurs for Bayesian fits. However, for the compound speed measure, where these 
dynamics parameter tradeoffs are controlled, the estimates from both approaches are 
closely matched.  
 
 
215 
 
Figure 4.5: SAT functions for LD and local antecedent ziji conditions with fully 
saturated models (2?-2?-2?), over average data (not averaged parameters). Accuracy 
is scaled to show proportion of asymptote; vertical bars indicate time point at which 
50% accuracy is reached. 
 
0 1000 2000 3000 4000
 0.
 0
 0.
 2
 0.
 4
 0.
 6
 0.
 8
 1.
 0
 SAT functions for Local and Long-distance ziji dependencies
 Processing time (ms)
 Proportion of asymptotic accurac
 y
 Long-distance antecedent
 Local antecedent
 Difference = 56ms
 
 
216 
 
Figure 4.6: SAT functions for LD and local control conditions with fully saturated 
models (2?-2?-2?), over average data (not averaged parameters). Accuracy is scaled 
to show proportion of asymptote; vertical bars indicate time point at which 50% 
accuracy is reached. 
 
 For the critical ziji comparison, parameter estimates from STEPIT fits 
indicated no reliable difference in estimated asymptote (lambda) parameters between 
the two critical ziji conditions. There was a reliable difference between the ziji 
conditions for rate (beta) parameters, as well as for the compound speed measure 
(t(17) = 2.47, p < 0.05 and t(17) = 2.52, p < 0.05 , respectively). In addition, for the 
intercept (delta) parameter there was a marginal effect of condition (t(17) = -2.09, p 
< 0.06). For Bayesian fits, the pattern of findings was qualitatively similar. There 
were no reliable differences for local versus long-distance configurations in either the 
0 1000 2000 3000 4000
 0.
 0
 0.
 2
 0.
 4
 0.
 6
 0.
 8
 1.
 0
 SAT functions for Local and Long-distance ziji dependencies
 Processing time (ms)
 Proportion of asymptotic accurac
 y
 Long-distance control
 Local control
 Difference = - 6ms
 
 
217 
asymptote parameter or the intercept parameter. There were, however, reliable effects 
in both the rate parameter (t(17) = 2.21, p < 0.05) and the compound speed measure 
(t(17) = 2.28, p < 0.05).  
 
Figure 4.7: Average asymptotic accuracy (?) across individual participant SAT 
function fits with Bayesian parameter estimation. Error bars show ?1 SE, corrected 
for between-participant variance. 
 
Average asymptotic accuracy ? for ziji and control conditions
 Acc
 ur
 acy
  d
 '
 0
 1
 2
 3
 4
 Local condition
 LD condition
 Ziji Control
 
 
218 
 
Figure 4.8: Average rate (?) across individual participant SAT function fits with 
Bayesian parameter estimation. Error bars show ?1 SE, corrected for between-
 participant variance. 
  
Average rate ? for ziji and control conditions
 R
 at
 e 
(s?
 1 )
 0
 1
 2
 3
 4
 Local condition
 LD condition
 Ziji Control
 
 
219 
 
Figure 4.9: Average intercept (?) across individual participant SAT function fits with 
Bayesian parameter estimation. Error bars show ?1 SE, corrected for between-
 participant variance. 
 
Average intercept ? for ziji and control conditions
 In
 te
 rc
 ep
 t (s
 )
 0.
 0
 0.
 2
 0.
 4
 0.
 6
 0.
 8
 1.
 0
 Local condition
 LD condition
 Ziji Control
 
 
220 
 
Figure 4.8: Average speed (?+?-1) across individual participant SAT function fits 
with Bayesian parameter estimation. Error bars show ?1 SE, corrected for between-
 participant variance. 
 
 On control conditions, STEPIT fits revealed no reliable differences between 
the local and long-distance configurations for any parameter. Likewise, Bayesian 
estimates of the parameters revealed only a marginal effect in the asymptote (?) 
parameter (t(17) = -1.75, p < 0.1), and no other reliable differences between the 
control conditions in any of the dynamics parameters.  
 It is unclear whether the processing speeds of the control conditions and the 
ziji conditions are directly comparable, as the processes involved in rendering a 
judgment in the two cases are distinct. Nonetheless, planned pairwise comparisons 
were performed to directly compare parameter differences for the ziji conditions to 
Average processing speed ? + ??1 for ziji and control conditions
 Pr
 oc
 ess
 in
 g 
sp
 ee
 d 
(ms
 )
 0
 50
 0
 10
 00
 15
 00
 20
 00
 Local condition
 LD condition
 Ziji Control
 
 
221 
those of the control conditions. For STEPIT fits, the locality advantage in ziji 
conditions over control conditions was significant for both rate and speed parameters 
(?: t(17) = 2.98, p = < 0.01, ?-1 + ?: t(17) = 2.29, p < 0.05). The same pattern was 
found for the Bayesian fits, although this effect was only marginally significant for 
the compound speed measure (?: t(17) = 2.69, p < 0.05, ?-1 + ?: t(17) = 1.74, p < 0.1). 
 One possible concern about the analysis presented here is that the fully 
saturated models presented here posit separate asymptote parameters for the critical 
ziji conditions. If the true model does not contain separate asymptotes for each 
condition, a potential worry is that non-significant trends in accuracy drive the 
observed dynamics effects. However, the accuracy effect for the critical ziji 
conditions was nearly reliable in the empirical accuracy analysis, and so it is unlikely 
that modeling a given participant?s data with a single asymptote model is appropriate. 
Furthermore, since almost every participant presented data that was significantly 
better fit with a two-asymptote model, modeling the data with a single asymptote 
would have led to significant distortions of the estimates of the dynamics parameters. 
For participants who have greater asymptotic accuracy for LD antecedent conditions, 
a single-asymptote model inappropriately increases the processing speed estimates for 
long-distance configurations, and conversely for participants who show the opposite 
pattern of accuracy.  
 
Parameter Ziji Control 
 Bayesian fit STEPIT fit Bayesian fit STEPIT fit 
Asymptote 
(d?) 
-0.14 (? .15) -0.15 (? .14) -0.20 (? .11) -0.17 (? .11) 
Rate (s-1) 0.42 (? .20) 0.86 (? .35) 0.04 (? .15) -0.04 (? .13) 
Delta (s) -0.01 (? .01) -0.32 (? .15) 0.00 (? .02) 0.02 (? .02) 
Speed (s) 0.09 (? .04) 0.09 (? .04) -0.02 (? .05) 0.03 (? .04) 
 
 
222 
 
Table 4.2: Difference in parameter estimates between local and long-distance 
configurations on the critical ziji and control conditions. Values greater than 0 
indicate a processing advantage for local antecedents. Standard errors by subject are 
in parentheses. 
 
Discussion 
 
 Both analyses of the SAT data indicate that the critical ziji conditions differed 
in processing speed, and that the control conditions did not. Model-fitting analyses 
showed that for the critical ziji comparison, the best fitting model attributed separate 
rate (?) parameters and asymptote parameters (?) to local and long-distance 
antecedent conditions. This indicated that the two ziji conditions reliably differed in 
speed and accuracy across conditions. The control comparison, on the other hand, 
showed no such advantage for extra dynamics parameters; only separate asymptotic 
accuracy parameters for each control condition reliably improved model fit. This 
analysis was supplemented with an analysis of the resulting parameter estimates, in 
order to check for consistent ordering of parameters. This analysis showed that in 
both rate and speed measures of the ziji conditions, the local antecedent was 
processed reliably faster across participants than was the LD antecedent condition. No 
such difference was observed in the control conditions, although there was a marginal 
effect of asymptotic accuracy for these conditions.  
 Thus the two analyses of the SAT data show that long-distance and local ziji 
conditions differ reliably in processing speed. Furthermore, the analysis of the 
direction of the parameters indicates that when ziji?s antecedent is local, it is 
reaccessed faster than when it is contained in a dominating clause. The most direct 
 
 
223 
measure of this effect is the compound speed measure indicated in Figure 4.11. The 
speed measure combines both rate and intercept parameters, and provides a measure 
of average processing speed in seconds. It can be seen that on average, local ziji 
antecedent configurations have approximately a 90ms processing advantage over the 
long-distance configurations.   
 
Figure 4.11: By-participants summary of dynamics advantage (in both rate and speed 
parameters) for local antecedent over LD antecedent conditions for ziji and control 
conditions. Participants are ranked by size of advantage; order is not identical across 
ziji and control conditions. Error bars represent 1 standard deviation of the posterior 
distribution of that parameter?s estimate.  
 
!
 !
 ! !
 !
 !
 ! ! !
 ! ! !
 !
 !
 ! !
 !
 !
 ?2
 ?1
 0
 1
 2
 Dynamics advantage for local antecedents
 Rate 
(s
 <1 )
 ?2
 ?1
 0
 1
 2
 Dynamics advantage for local control condition
 Rate 
(s
 <1 )
 ?0.5
 0.0
 0.5
 Participants
 Speed (s
 )
 !
 !
 !
 !
 !
 !
 ! !
 ! ! ! ! ! !
 !
 !
 !
 !
 ?0.5
 0.0
 0.5
 Participants
 Speed (s
 )
 ziji conditions
 control conditions
 !
 
 
224 
 This finding supports the structured search hypothesis. Comprehenders appear 
to access the local antecedent before accessing the long-distance subject, a finding 
that is compatible with a number of architectural implementations. If ziji activates its 
antecedent with purely structural information, like English herself/himself, then this 
finding is expected; the sort of iterated structural access mechanism that would derive 
this processing profile is presented in Figure 4.3. It can be seen that, on the 
assumption that the local antecedent is more active, then recovery of the the long-
 distance antecedent requires multiple retrieval processes.  
 This finding stands in contrast to previous SAT work, which has generally 
found that memory access in sentence processing occurs with essentially constant 
time. I argued that the syntactically constrained, fully retrospective nature of ziji 
makes it an ideal dependency for testing the possibility of structured search, if such 
an access mechanism is ever engaged. The evidence presented here provides positive 
evidence that the parser does engage in structured search. Although they stand in 
contrast to prior SAT findings, this finding about the access profile of ziji converges 
nicely with prior findings on the processing of ziji, suggesting both serial access (Gao 
et al 2005; Liu 2009) and a locality preference (Liu 2009; Li & Zhou 2010).  
 One unexpected finding was the high acceptance rate of the no antecedent ziji 
condition, which participants accepted on 48% of trials. It is known that in the 
absence of linguistic binding, ziji can refer to the speaker (Huang & Liu 2001). In the 
context of this experimental task, it is unclear how participants calibrated the point of 
view of the sentences they observed. The perceived point of view of the sentence is 
presumably the main factor that controls the recoverability of this indexical 
 
 
225 
interpretation of ziji. Since this aspect was not controlled, however, it is unclear how 
participants perceived these unbound ziji sentences. The relationship of unbound ziji 
to the linguistically bound examples presented here is an important question, but I 
delay further discussion of it until after Experiment 7.  
 The SAT evidence presented here is indicative of a structured search strategy, 
but there remain some unaddressed questions. If structured search proceeds by 
checking and rejecting the local subject position before constructing a dependency 
with an acceptable, but distant subject position, as in Figure 4.3, then a process that 
functionally approximates reanalysis is engaged in recovering a long-distance 
interpretation. One important question is the relationship between structured search 
and reanalysis processes that are engaged by this type of structured search and those 
that are engaged in more familiar cases of ambiguity, such as garden-path sentences 
(Bever 1970; Kimball 1973; Frazier & Fodor 1978). 
 A related question concerns the automaticity of this process. Does structured 
search reflect a strategy that participants adopt in the context of a particular 
experimental manipulation, or is it indicative of a general structure-building process 
that is uniformly invoked when processing long-distance reflexives? This is a general 
concern about SAT experiments; repeated exposure to the experimental materials 
over multiple hour-long sessions increases the likelihood that participants detect the 
crucial experimental manipulation and adapt their processing strategies accordingly 
while performing the task. It is possible that comprehenders adopted a strategy to 
consider for rendering quick decisions about the acceptability on ziji by performing a 
superficial check of the two subject positions as if they were words in a list. This 
 
 
226 
would suffice to perform the task, but would presumably not reflect the actual 
processes engaged by building an anaphoric dependency. Although the control 
conditions were designed to disrupt this kind of anticipatory strategy, it is possible 
that comprehenders nonetheless engaged this sort of (relatively) conscious 
strategizing to perform the judgment task. 
Experiment 7 investigated long-distance and local ziji bindings using event-
 related brain potentials (ERP) to investigate these questions. ERP is an attractive 
technique to complement the SAT evidence presented here. The SAT evidence 
provides crucial support to the contention that it is processing speed that differs 
between local and long-distance ziji binding, but the nature of the processes that slow 
computation for long-distance antecedents remains unclear. The ERP signal can be 
broken down into a number of distinct waveform components, each of which is 
thought to index different aspects of linguistic processing. Determining which ERP 
components index the locality advantage could shed light on the nature of the 
structured search process.  
 
Experiment 7: ERP Evidence 
 
 The evidence in Experiment 6 shows that the process of reaccessing a long-
 distance antecedent for ziji occurs more slowly than it does for local antecedents. The 
explanation for this effect on a structured search account is that the local subject is 
first selectively reaccessed, and then rejected due to a poor fit with ziji. The extra 
processing steps involved in rejecting the local antecedent and accessing the long-
 
 
227 
distance antecedent cause a slow-down in processing time when compared to the 
local antecedent. However, the SAT data do not indicate the exact processes that 
cause this slow-down. In Experiment 7, I investigate the processing of ziji using ERP 
in order to determine which stages of processing are impacted when a long-distance 
dependency needs to be constructed. An important feature of ERPs is that they 
provide a multidimensional view on language processing, with multiple components 
indexing arguably distinct cognitive processes that support linguistic processing. For 
this reason, in determining which ERP component indexes the locality advantage in 
the SAT materials, we may gain further insight into the nature of the processed speed 
advantage enjoyed by local antecedents relative to long-distance antecedents. 
The exact functional significance of many ERP components is a matter of 
active research and debate, and there appears to be a many-to-many mapping between 
hypothesized linguistic processes and ERP components. To a first approximation, 
however, semantic and syntactic processing are characteristically associated with 
different ERP components. For instance, words that are anomalous with respect to 
morphological or syntactic features have long been recognized to generate the P600 
response, a late posterior positivity that generally peaks around 600 ms post-stimulus 
(Friederici, Pfeifer, & Hahne, 1993; Hagoort, Brown, & Groothusen, 1993; Osterhout 
& Holcomb, 1992), as well as an earlier anterior negativity termed the (E)LAN 
(Kluender & Kutas 1993; Coulson, King, & Kutas, 1998; Friederici et al., 1993; 
Hagoort, Wassenaar, & Brown, 2003; Lau, Stroud, Plesch, & Phillips, 2006; Neville, 
Nicol, Barss, Forster, & Garrett, 1991). Although these components are sensitive to 
similar factors (morphosyntactic wellformedness), they are widely regarded as 
 
 
228 
reflecting distinct processes in the computation of a syntactic representation 
(Friederici 1995; Hahne & Friederici 1999; Hagoort 2003; Bornkessel & Schlesewsky 
2006). On the other hand, semantic anomalies in otherwise syntactically well-formed 
sentences typically elicit a central negativity around 400 ms known as the N400 
(Kutas & Hillyard, 1980; Kutas & Federmeier, 2000; Lau, Phillips, & Poeppel, 2008). 
These characterizations are far from exceptionless, however. For example, there 
appear to be instances of ??semantic?? error that engender P600 responses (Kolk, 
Chwilla, van Herten, & Oor, 2003; Kim & Osterhout, 2005; Kuperberg, 2007), as 
well as N400 effects that have been linked to processing case anomalies (Hopf, 
Bayer, Bader, & Meng, 1998) or syntactic reanalysis effects (Bornkessel, 
Schlesewsky & McElree, 2004). 
Interestingly, there is support for the idea that early and late ERP components 
seem to characterize automatic and controlled processes, respectively (e.g. Hahne & 
Friederici 1999; Pulverm?ller, Shytrov, Hasting & Carlyon 2008). For example, it has 
been noted that later components such as the P600 are readily modulated by the 
proportion of ungrammatical sentences in an experimental session (Coulson et al 
1998, Hahne & Friederici 1999). This suggests that these responses reflect relatively 
controlled processes, and that experimental participants modulate the processing 
routines that underlie this component in order to adapt to the experimental 
environment. Earlier components, such as the LAN, are apparently robust to this 
manipulation (Hahne & Friederici 1999). As it is relatively invariant across task 
manipulation, these earlier responses have been argued to reflect fast, automatic 
 
 
229 
processes involved in structure-building (Friederici 1995; Hahne & Friederici 1999; 
Pulverm?ller et al 2008). 
Here these distinctions are potentially helpful in shedding light on the SAT 
data presented in Experiment 6. One question that remained from Experiment 6 was 
whether or not the observed speed delay for long-distance antecedents was related to 
strategic processing, or if it is best characterized as reflecting early, automatic 
processes necessary for constructing an anaphoric dependency. On the assumption 
that the early structure-building components (e.g. the (E)LAN) index more automatic 
structure-building processes, then we may gain some insight into this question. If we 
observe a LAN effect at ziji when its antecedent is long-distance, then there is a case 
to be made that the processes that contribute to the SAT slowdown reflect early, 
automatic processing routines associated with building a structural anaphoric 
dependency. However, if later components such as the P600 are observed, then this 
claim is not necessarily supported. A P600 opens the possibility that the shift from 
local to long-distance antecedents involves more controlled processes, potentially 
similar to those observed in other syntactically ambiguous structures such as garden-
 path sentences. A number of authors have noted that P600 effects obtain in situations 
of grammatical ambiguity (Osterhout, Holcomb & Swinney 1994; Hopf et al., 1998; 
Friederici, Hahne & Saddy 2002; Hagoort, 2003; Kaan & Swaab, 2003). If the 
reanalysis suggested by Figure 4.3 reflects reanalysis processes that are more widely 
engaged in sentence processing, or a conscious, task-specific strategy, then one 
expects to observe a P600 in long-distance antecedent situations.  
 
 
230 
 Previous ERP work on the processing of ziji has demonstrated that long-
 distance antecedents elicit a posterior positivity (P300/600) relative to local 
antecedents, supporting the locality bias for ziji (Li & Zhou 2010). However, this 
study differed in important ways from the materials in the SAT study. In Li and 
Zhou?s study, both the local and the long-distance antecedent positions contained 
animate NPs that were featurally compatible with ziji. In their manipulation, the 
semantics of the embedded verb served to disambiguate the antecedent of ziji, as the 
verb was either inherently reflexive or anti-reflexive (Li & Zhou 2010). Thus it is 
contextual or thematic knowledge that comprehenders need to draw upon to exclude 
the local subject from consideration, but the dependency between the local subject 
and the anaphor is well-formed. In contrast, the dependency between the inanimate 
local subject and the anaphor in the SAT study is not formally well-formed, which 
drives a further search for an acceptable antecedent as in Figure 4.3. It is possible that 
the use of contextual or semantic knowledge to exclude a licit binding dependency in 
case of a true ambiguity is an entirely different process from the structured search for 
an antecedent, which presumably occurs prior to the contextual evaluation of the 
antecedent-anaphor dependency. For this reason it is difficult to straightforwardly 
conclude that the P300/P600 observed by Li and Zhou is the electrophysiological 
index of structured search. If the structured access account is correct, then only when 
the local subject is unacceptable is a more extensive consideration of potential 
antecedents required. Thus in order to more directly compare the ERP reflex of ziji?s 
locality bias with the SAT results, I present an ERP study of the processing of long-
 
 
231 
distance and local ziji in environments that were parallel to those tested in the SAT 
study.  
 
Participants 
 
Twenty-four college students from Beijing Normal University participated in 
the experiment, including 13 females (mean age 22). The 24 participants (13 females) 
had a mean age of 22, were all healthy, native speakers of Mandarin Chinese with no 
history of neurological disorder, and all were strongly right-handed based on the 
Edinburgh handedness inventory (Oldfield, 1971). All participants gave informed 
consent and were paid 50 RMB/hour for their participation, which lasted around 2? 
hours, including set-up time. 
 
Materials 
 
The experimental materials consisted of three conditions designed to 
investigate the processing of ziji (conditions 1-3 in Table 4.3 below). These three 
conditions matched the three critical ziji conditions in Experiment 6 on a number of 
important dimensions. Ziji either took a long-distance antecedent (LD antecedent 
condition; 1), a local antecedent (local antecedent condition; 2), or had no antecedent 
in the sentence (no antecedent condition; 3). As noted above, an animate NP in either 
the main clause or embedded subject NP position can function as a grammatical 
antecedent for ziji. As in Experiment 6, all three conditions consisted of a main clause 
 
 
232 
that contained a verb of reporting, and an embedded clause, of which ziji was the 
object. Likewise, the binding possibilities for ziji were manipulated by manipulating 
the animacy of the subject NPs: LD antecedent conditions contained an animate 
matrix subject, and local antecedent contained an animate embedded subject. All 
other pre-critical NPs were inanimate, and in the no antecedent condition, both 
subject NPs were inanimate, and therefore this condition did not contain a 
grammatical antecedent for ziji. In order to ensure that the critical ziji was not in a 
sentence final position, all conditions followed ziji with a conjunction (e.g. ye ?also?, 
que ?but?) and second clause continuation. Although the processing profile of 
conjoined anaphors is known to be different from that of single argument anaphors 
(see, e.g. Harris, Wexler & Holcomb 2000; Burkhardt 2005), none of the conjunctions 
in the experimental materials can be used as NP-level conjunctions, ensuring that 
participants did not interpret ziji as part of a conjoined NP.  
120 sets of the 3 sentence types were generated. The 360 sentences were 
equally distributed in 3 presentation lists and subjects were assigned to each list in a 
Latin Square fashion. Within each list, the 120 targets were interleaved with 240 
unrelated fillers of similar complexity, and the list was divided into six blocks of 40 
sentences. The order of the 6 blocks was randomized across subjects. Of the 360 
sentences in each session, half were considered to be acceptable, half were considered 
to be unacceptable.  
 
# Condition Example 
1 LD antecedent ???? [?? ????? ??]? ?? ???? 
Chef say [deep-fryer scalded ziji], so (pro) resigned. 
?The chef said the deep fryer scalded him, so he resigned.? 
 
 
233 
2 Local antecedent ???? ??[?? ????? ??]? ?? ???? 
Medical report say [chef scalded ziji], so (pro) resigned. 
?The medical report said the chef scalded himself, so he resigned.? 
3 No antecedent */????? ?? [?? ????? ??]? ?? ???? 
*/?Medical report say [deep-fryer scalded ziji], so (pro) resigned. 
*/? ?The medical report says that the deep fryer scalded self, so he/she/I 
resigned.? 
 
Table 4.3: Summary of conditions in Experiment 2. 
 
Procedure 
 
 Participants were comfortably seated in a dimly lit testing room around 
100 cm in front of a computer monitor. Sentences were presented one word at a time 
in white letters on a black background in 30 pt simplified Chinese characters. Each 
sentence was preceded by a fixation cross. Participants pressed a button to initiate 
presentation of the sentence, which began 1000 ms later. Each word appeared on the 
screen for 400 ms, followed by 200 ms of blank screen. The last word of each 
sentence was marked with a period, and 1000 ms later a question mark prompt 
appeared on the screen. Participants were instructed to read the sentences carefully 
without blinking and to indicate with a button press whether the sentence was an 
acceptable Mandarin sentence. Each experimental session was preceded by a 12-trial 
practice session that included both grammatical and ungrammatical sentences. 
Participants received feedback and were able to ask clarification questions about the 
task during the practice session. The experimental session was divided into six blocks 
of 75 sentences each. Breaks were permitted after each block as necessary.  
 
 
 
234 
EEG Recording 
 
 EEG was recorded from 32 Ag/AgCl electrodes, mounted in an electrode cap 
(Electrocap International): midline: Fz, FCz, Cz, CPz, Pz, Oz; lateral: FP1/2, F3/4, 
F7/8, FC3/4, FT7/8, C3/4, T7/8, CP3/4, TP7/8, P4/5, P7/8, O1/2.  Recordings were 
referenced online to the right mastoid, and re-referenced offline using linked 
mastoids. An additional electrode was placed on the left and right outer canthus, and 
above and below the left eye to monitor eye movements and eye blinks. EEG and 
EOG recordings were amplified and sampled at 1000 Hz using a bandpass filter of 
0.1-70 Hz. Impedances were kept below 5 k?. 
 
EEG Analysis 
 
 All analyses were conducted over single trial epochs, consisting of the 100 ms 
preceding and the 1000 ms following the critical presentation of ziji, normalized 
using a 100 ms pre-stimulus baseline. In order to exclude motion and ocular artifacts, 
normalized epochs with activity greater than ? 50 ?V were removed, as were trials 
that had a peak-to-peak voltage difference of greater than 100 ?V in the EOG (Luck 
2005). The total rejection rate with these criteria was 19%, ranging between 18%-
 20% across critical conditions. Averaged waveforms were filtered offline using a 20 
Hz low-pass filter for presentation purposes; however, all statistics were performed 
on unfiltered data. The latency intervals that were analyzed statistically were chosen 
based on previous conventions in the ERP sentence processing literature: 0-200 ms, 
 
 
235 
200-400 ms, 400-600 ms, 600-800 ms, and 800-1000 ms, Regions of interest were 
defined as follows: left anterior (FT7, F3, FC3), midline anterior (FZ, FCZ, CZ), right 
anterior (F4, FC4, FT8), left posterior (TP7, CP3, P3), midline posterior (CPZ, PZ, 
OZ), and right posterior (CP4, P4, TP8).  
 In order to assess the reliability of the effects elicited by the experimental 
manipulations, I employed linear mixed effects (LME) modeling (Pinheiro & Bates 
2000). There are a number of advantages to this approach over traditional approaches. 
One important advantage in the current context is that this approach readily 
accommodates missing data (Gelman & Hill 2005; Baayen et al 2008). Since epoch 
rejection rates tend to vary a good deal across participants in ERP studies, LME 
models are an attractive analysis option.  
  Analysis proceeded separately for each time interval, with average values for 
each epoch within that time interval estimated for the electrodes included in the 
analysis. The best-fit model for each time interval was selected by hierarchical model 
comparison, by adding terms for fixed effects and checking for a significant increase 
in model likelihood using a ?2?test. The logic and interpretation of this approach is 
qualitatively similar to the hierarchical interpretation of ANOVA results that is often 
reported in ERP papers: fixed effects (factors) and their interactions are evaluated to 
ensure that they explain a significant amount of variance, and the resulting best-fit 
model is then evaluated to determine the nature of these effects. For the present 
experiment, there were two fixed effects of electrode position, anteriority (with the 
levels posterior and anterior) and laterality (with the levels left, midline, and right), 
as well as two orthogonal experimental contrasts for binding and locality. Main 
 
 
236 
effects and interactions for electrode position and order were entered into the models 
prior to the planned orthogonal contrasts using simple difference coding. The binding 
contrast assessed the effect of having a linguistic antecedent in the sentence, and 
compared the no antecedent condition to both LD and local antecedent conditions 
(coding: .5 for the antecedent conditions, -.25 for both LD and local antecedent 
conditions).  The locality contrast assessed the effect of having a local antecedent, 
and directly compared local and LD antecedent conditions (coding: .5 for LD 
antecedents, -.5 for local antecedents), 0 for no antecedent). Random intercepts for 
trial and participants were included; random intercepts for items and random slopes 
for the experimental fixed effects did not significantly increase the likelihood for any 
model, and so were not included in the final model fits described here. This exclusion 
did not change the pattern of results found. For each time interval, we describe the 
best-fitting model and the fixed effect coefficients. All p-values for linear model 
coefficients were estimated using MCMC methods as implemented in the LanguageR 
R package (Baayen 2008), with n=10,000 samples. 
 
Results: Behavioral Data 
 
The average rate of acceptance was 81% for the long-distance binding 
condition, 73% for the local binding condition, and 51% for the no-binding condition. 
A logistic linear mixed effects model with crossed random intercepts for participants 
and items revealed significant effects of both the binding contrast (? = -.81, z = -14.0, 
 
 
237 
p < .0001) as well as the locality contrast (? = .47, z = 4.2, p < .0001). An analysis of 
reaction times did not reveal any significant differences among the conditions. 
 
Results: ERP Data 
 
 The grand average ERPs are presented in Figure 4.12. A summary of the fixed 
effects for binding and locality are presented in Table 4.4. In the 0-200ms time 
interval, there were no significant fixed effects for experimental contrasts. In the 200-
 400ms time interval, the best-fitting model included an interaction of binding with 
anteriority and laterality. Resolving this interaction revealed a broadly distributed 
negativity for unbound ziji, relative to the two bound ziji conditions. Numerically, this 
effect was largest over posterior regions. In the 400-600ms time window, analysis 
revealed a negativity for the LD antecedent relative to the local antecedent binding 
conditions, which reached significance in the mid and left anterior ROIs. A negativity 
was also observed in the 600-800ms time window, with a focus over the mid and 
right anterior ROI. No experimental fixed effects were observed between 800-
 1000ms. 
 
 
238 
 
Figure 4.12: Grand average ERPs for ziji conditions in Experiment 7, low-pass 
filtered for visual presentation. Only electrodes included in the ROI analyses are 
presented. 
 
 
 
 
 
 
 
  
Long-distance antecedent
 Local antecedent
 No antecedent
 3 +V
 -3 +V
 1000ms
   OZ
   PZ
  CPZ
   CZ
  FCZ
   FZ
  FT7
  FC3
   F3
  FC4
  FT8
   F4
   P3
  CP3
  TP7
   P4
  CP4
  TP8
 
 
239 
 0-200ms 200-400ms 400-600ms 600-800ms 800-1000ms 
Binding      
Left Ant. - -.64 ? .44 - - - 
Mid Ant. - -1.18 
? .44** 
- - - 
Right Ant. - -.91 ? .44* - - - 
Left Post. - -.79 ? .44? - - - 
Mid Post. - -1.22 
? .44** 
- - - 
Right Post. - -1.07 ? .44* - - - 
Locality      
Left Ant - - -1.02 ? .43* -.73 ? .47 - 
Mid Ant. - - -.91 ? .43* -.90 ? .47? - 
Right Ant. - - -.75 ? .43? -.99 ? .47* - 
Left Post. - - -.84 ? .43? -.72 ? .47 - 
Mid Post. - - -.18 ? .43 -.13 ? .47 - 
Right Post. - - -.31 ? .43 -.48 ? .47 - 
 
Table 4.4: Table of experimental fixed effects (coefficients in ?V, with standard 
error). Experimental fixed effects only shown if the best-fit model included a 
significant interaction of experimental effect with anteriority and laterality . ? = p < 
0.1, *  = p < 0.05, **  = p < 0.01. 
 
Discussion 
 
 The ERP results revealed two distinct components associated with the 
processing of ziji in our materials. For the binding contrast, an early negativity 
associated with the processing of unbound ziji was observed in the 200-400ms time 
window. This negativity had a primarily central-posterior distribution. This is 
consistent with an N400 effect, though visual inspection of the ERPs suggests that the 
peak of this negativity is slightly earlier than the canonical N400 effect. For the 
locality contrast, a qualitatively different component was observed, differing both in 
time course and distribution. LD antecedents caused an increased negativity over the 
 
 
240 
anterior ROIs in the 400-600ms window, with some effects also being observed in the 
subsequent 600-800ms window. The distribution of this component is consistent with 
a (L)AN effect, although the time course is somewhat later than the canonical LAN 
effect.   
 One important insight from these results is that the processes involved in 
recovering a long-distance bound interpretation and revising to an indexical 
interpretation of ziji are at least partially distinct. The exact interpretation of this 
effect depends on assumptions about the identity of the component, and that 
component?s functional significance. For example, if this component is best 
characterized as an N400, then it might be understood as indexing the difficulty in 
accessing information in the lexicon (e.g. Lau et al 2010). On this interpretation, this 
effect suggests that ziji is less predicted in the no antecedent conditions. Another 
important distinction between indexical and bound ziji is that in order to recover an 
indexical interpretation, participants must change the point-of-view of the sentences 
(Huang & Liu 2001; Anand 2006). If this is necessary when encountering ziji in the 
context of a discourse with no animate participants, then the negativity may reflect 
this process of context-shifting.  
The difficulty associated with long-distance interpretations of ziji was 
observed in a (L)AN component at approximately 400ms post-stimulus.. Anterior 
negativities of this sort have been linked to a variety of morphosyntactic violations as 
well as working memory difficulty associated with forming long-distance 
dependencies (Neville et al 1991; Friederici et al., 1993; Kluender & Kutas, 1993; 
King & Kutas, 1995; Coulson et al, 1998; Hagoort et al 2003). If structured search 
 
 
241 
proceeds as in Figure 4.3, then the observed anterior negativity is compatible with 
either interpretation of the (L)AN component. On the assumption that the LAN 
indexes a morphological violation, then it might be caused by the mismatch between 
the animacy features of the local subject and ziji when the local subject is reaccessed 
and considered. Alternatively, if the observed negativity reflects working memory 
difficulty (as argued by Kluender & Kutas 1993), then it may be indexing the fact that 
more retrieval operations are necessary to recover the long-distance antecedent (as in 
Figure 4.3), leading to a greater anterior negativity. Both of these interpretations are 
compatible with a view of structured search as an early process engaged during the 
construction of anaphoric dependencies, but further work is necessary to distinguish 
between them. 
 Interestingly, recovering a long-distance interpretation did not cause ERP 
components that are commonly observed during syntactic or semantic reanalysis. 
Reanalysis that is associated with syntactic ambiguity or garden-path type sentences 
is often reflected in a posterior P600 component (Osterhout et al 1994; Friederici et al 
2002; Hagoort, 2003; Hopf et al 2003; Kaan & Swaab, 2003), although it has been 
noted that in some cases the N400 may reflect this sort of reanalysis processes 
(Bornkessel et al 2006). Neither component was observed in the processing of long-
 distance antecedents. This supports the contention that the relative processing 
disadvantage shown by long-distance interpretations of ziji in Experiment 6 is not due 
to the same reanalysis processes that are engaged by processing garden-path type 
sentences. It is important to note that these results stand in contrast to those reported 
by Li & Zhou (2010), who did report that long-distance interpretations of ziji caused 
 
 
242 
both P300 and P600 effects. However, in their study, the local subjects were always 
animate and compatible with ziji; long-distance interpretations of ziji were forced by 
manipulating the verb semantics.  In this case, if retrieval of the local subject yielded 
a feature-matched antecedent, it is unclear what processing would be necessary to 
reanalyze to a long-distance interpretation. If structured search terminates upon 
finding a compatible antecedent, then it is possible that entirely different reanalysis 
processes were required to recover the distant antecedent in that study, causing a 
P300/600 rather than an anterior negativity. If this reasoning is correct, then this 
account generates testable predictions for future work about when the different ERP 
components will be observed in processing ziji. In particular, the difficulty associated 
with the structured search necessary to find an initial, accceptable antecedent for ziji 
should be reflected in a LAN. On the other hand, difficulty associated with selecting 
among multiple antecedents based on contextual information should be associated 
with P600 effects.  
 If the distinction between early, automatic processes and later controlled 
processes is correct (Hahne & Friederici 1999), then an interesting claim about ziji 
can be made. The finding that structured search causes a LAN effect is compatible 
with the view that the (relatively difficult) process of recovering a long-distance 
antecedent for ziji is a fast, automatic process engaged by encountering a bound 
instance of ziji. This is consistent with the contention that the SAT dynamics effects 
seen in Experiment 6 were not due to task-specific strategies that participants adopted 
over the course of the SAT experiment. On the strongest interpretation of this result, 
 
 
243 
this pattern of results suggests that the structured search for an antecedent reflects the 
earliest structure-building procedures engaged upon encountering ziji.  
 Overall, the ERP effects provide an alternative source of evidence in support 
of structured search by confirming that fully grammatical (and preferred) long-
 distance interpretations of ziji are associated with extra processing difficulty. The 
time course of the difficulty associated the LD antecedent conditions is consistent 
with structured search reflecting an early, automatic process, rather than later, 
controlled processes that are sometimes associated with syntactic reanalysis 
(Osterhout et al 1994; Hahne & Friederici 1999; Friederici et al., 2002). The (L)AN 
associated with the long-distance interpretation of ziji may reflect either the feature 
mismatch with the local subject, or the extra memory retrievals necessary to reaccess 
the distant antecedent. Interestingly, the response that indexes the shift to an indexical 
interpretation of ziji temporally preceded the difficulty associated with long-distance 
interpretations. Furthermore, since no LAN was seen in this condition, it is possible 
that the decision to pursue an indexical interpretation of ziji obviated the need for the 
parser to check the local subject position at all. This may indicate that participants 
quickly decide to pursue a shift in point-of-view rather than attempting to bind ziji to 
an antecedent within the sentence. The nature of this decision remains unclear, but I 
tentatively suggest that the lack of any animate discourse entities may have biased 
readers in favor of a first-person interpretation of ziji. Thus there is a quick use of top-
 down discourse information about which parse to pursue, but once the decision to 
pursue a linguistically bound interpretation has been engaged, the parser engages a 
structured search. As in Experiment 6, however, participants varied a great deal in the 
 
 
244 
degree to which they accepted the no antecedent ziji condition. Further study that 
more carefully controls the perceived point-of-view of the sentences is necessary to 
get a clearer picture of the processing of unbound ziji.  
 
General Discussion 
  
 The current experiments investigated the processing of antecedent-anaphor 
dependencies involving the Mandarin Chinese long-distance reflexive ziji using both 
time-course analysis and electrophysiological measures. The SAT results showed that 
in constructing a ziji-antecedent dependency, local antecedents are accessed more 
rapidly than long-distance antecedents. This locality processing advantage was 
reflected in both the rate and speed measures in SAT, and this difference in dynamics 
was absent from control conditions without ziji. These results stand in contrast to 
previous SAT studies, which suggested that processing advantages due to locality 
only impacted asymptotic accuracy (i.e., the probability of successfully computing 
the dependency; McElree 2000; McElree et al 2003; Martin & McElree 2008).  
The ERP results confirmed the locality advantage for ziji antecedents. Long-
 distance interpretations of ziji elicited a (L)AN component relative to local 
interpretations of ziji rather than a P600 copmonent, a finding that suggests that the 
locality advantage seen in the SAT experiment is not due to strategies employed in 
the context of a long SAT experiment, but rather is due to automatic structure-
 building processes required to form an antecedent-anaphor dependency.  
 
 
245 
 Both experiments support the hypothesis that comprehenders initially consider 
a feature-inappropriate, but structurally accessible antecedent when pursuing a bound 
interpretation of ziji. This is a direct prediction of structured access in sentence 
comprehension: if access is guided by structural position, rather than feature-content, 
then participants have no choice but to initially consider the local subject. On this 
model, there is no information available to the parser during the course of 
constructing the binding dependency to exclude this position from consideration. This 
informational bottleneck forces participants to engage a structured search when there 
are multiple structurally licit positions to consider for forming a dependency. If 
comprehenders only employ structural information to retrieve anaphoric antecedents, 
then they have a natural mechanism for limiting the dependency to structurally 
appropriate antecedents only. As discussed above, such a mechanism is agnostic with 
respect to certain architectural commitments. In particular, it not obviously at odds 
with other time-course results in retrospective dependencies that suggest a role for 
content-addressable memory architectures in sentence processing (e.g. Martin & 
McElree, 2008). The structured search process may reflect iterated retrievals from an 
underlyingly content-addressable architecture (as in the model in Chapter 3), 
provided that the processor has access to cues that can effectively distinguish the 
local subject from other positions in the sentence. By iterating retrievals with varying 
sets of retrieval cues, a content-addressable architecture can implement this sort of 
structured search when the need to faithfully implement structural constraints 
outweighs the need for speed in dependency formation. The relevant difference 
between the ziji dependency and the VP-ellipsis dependencies examined by Martin 
 
 
246 
and McElree (2008, 2009), on this viewpoint, is that only for ziji is the parser willing 
to suffer a slower retrieval procedure in the service of structural accuracy. 
 
Locality bias in ziji dependencies 
  
 The time course evidence presented here is compatible with theories of 
dependency construction that invoke notions of structured search. Furthermore, it 
suggests that at the point of initiating the antecedent-anaphor dependency, 
comprehenders initially access and check the local subject position. This is not 
consistent with the view that all ziji-antecedent dependents are considered in parallel, 
which would predict constant SAT dynamics for all antecedent positions. An 
interesting finding is that participants appear to prefer to access the local subject 
position upon encountering ziji. This does not necessarily follow from the formulation 
of structured search given above, which only holds that multiple syntactically licit 
positions need to be checked in an effectively serial order. If comprehenders preferred 
to access the highest subject in a sentence, and search downwards, this would be 
entirely compatible with structured search as I have presented it. This leads to an 
interesting question of why the local subject is preferred in this case.    
One possibility is that the locality bias reflects a general advantage for 
linguistic material contained within the local clause relative to material outside the 
local clause. In other words, the local subject may be available more quickly by virtue 
of its being contained within the local clause, which is still is in the process of being 
parsed when ziji is encountered. This account is consistent with studies on sentence 
 
 
247 
recall that suggest that the local clause has a privileged role in online sentence 
processing (Jarvella, 1971; Jarvella & Pisoni, 1970). This account does not predict 
any necessary preferred order of access for subject positions outside of the local 
clause. This may also be equivalent to the claim that that some elements remain 
concurrently available in the focus of attention while others are displaced and must be 
later retrieved (McElree, 2006; Jonides, Lewis, Nee, Lustig, Berman & Moore, 2008). 
If the local subject remains in the focus of attention, while the long-distance subject 
requires retrieval, then the observed distinction in dynamics would be predicted. This 
interpretation seems less likely, in light of findings from list memory experiments that 
indicate that the focus of attention is extremely limited in size and scope, 
corresponding to just one task-relevant encoding (McElree & Dosher, 1989; McElree, 
1998). If only one element occupies focal attention before ziji is processed, it is likely 
to be the verb, in anticipation of the upcoming object NP. However, the available data 
on the contents of the focus of attention is limited for connected linguistic 
representations, which have considerably richer structure than do word lists. It is 
known that full clauses are sufficient to displace information about their embedding 
environment (McElree et al., 2003; Wagers & McElree, 2009). In the context of the 
present experiment, the temporal adjunct clause that intervened between the subject 
and the verb seems likely to have pushed the local subject out of the focus of 
attention. However, it is presently unknown whether information about the subject is 
later restored and thus focally available during the initial processing of verb-phrase 
internal arguments. 
 
 
248 
Alternatively, the locality advantage may reflect a strict ordering of access, 
where subject positions are accessed in progressively dominating positions. This 
ordering could reflect a useful strategy for processing ziji in particular, due to the 
blocking constraints on ziji. Consider (4.4) and (4.5), repeated as (4.7) and (4.8) 
below: 
   
 (4.7)  Zhangsani  renwei nij hen ziji*i/j 
   Zhangsan think you hate ziji 
   ?Zhangsan thinks that you hate yourself/*him.? 
 
 (4.8)  Nii  renwei Zhangsanj hen ziji?i/j 
   You think Zhangsan hate ziji 
   ?You think that Zhangsan hates himself/?you.? 
 
These sentences display the blocking effect for ziji. The important generalization 
about these effects is that ziji cannot access subjects that dominate first- or second-
 person subjects (Huang & Liu 2001); similarly, it has been reported that ziji cannot be 
bound by singular subjects that dominate plural subjects (Tang 1989; Huang & Tang 
1991). If comprehenders consider antecedent positions from the most local to the 
most distant in a strict order, then this constraint can be easily implemented by 
terminating search upon reaching an indexical or plural subject. This intuition is 
reflected in grammatical accounts of ziji that invoke cyclic movement of ziji to 
progressively higher subject positions in the derivation of an antecedent-anaphor 
chain (Pica 1986; Cole, Hermon & Sung 1990; Cole & Sung 1994; Cole, Hermon & 
Lee 2001). This account very naturally explains blocking effects. On these accounts, 
the feature match between ziji and a subject is evaluated at each subject position, 
from the most local to the most distant.  
 
 
249 
 Both of the retrieval-based accounts presented above share a common feature 
in that they require the use of positional or structural information in retrieval, and thus 
support the existence of structured access mechanisms. Note that this requirement 
does not hold if the local subject is maintained in focal attention. In order to 
implement a search that serially samples subject positions, or that preferentially 
accesses information in the local clause, the positional information inherent in those 
two specifications needs to be available to guide memory access. It is possible, in 
principle, to empirically distinguish the two accounts. The fully serial, bottom-to-top 
search and the locality bias account reflects a make distinct predictions about the time 
course of activating long-distance subjects that are two versus three clauses distant 
from ziji, as in (4.9).  
 
 
 (4.9)  Lisii shuo Xiaomingj renwei fengbao hai-le
  zijii/j/k 
   Lisi  says Xiaoming thinks storm hate ziji 
   ?Lisi says Xiaoming thinks the storm harmed himself/her.? 
 
On either account, the inappropriate fengbao should be considered before either Lisi 
or Xiaoming. If the locality bias observed here reflects a general advantage for the 
local clause, then participants should be equally likely to consider Lisi or Xiaoming in 
the second stage of access. In SAT, this would mean constant access dynamics for 
non-local position, reflecting the fact that on some trials Lisi is the first non-local 
antecedent considered, and on the rest, Xiaoming is considered. If the locality bias 
indicates a strategy whereby comprehenders progressively consider more distant 
subject positions, then SAT dynamics should show that the intermediate subject 
 
 
250 
position is reliably accessed before the highest subject position. Future work will 
examine the role of hierarchical distance beyond the local clause boundary in an 
attempt to tease apart these competing hypotheses.  
 
Alternative accounts of the data 
 
An alternative account of the current results is that the difference in 
processing dynamics between dependencies with local and long-distance antecedents 
reflects a reanalysis from a local interpretation to the long-distance interpretation. 
This is consistent with some linguistic accounts of ziji that have suggested that ziji as 
a local anaphor is lexically distinct from ziji when its antecedent is distant, based on 
differences in meaning and pragmatics of usage in these different environments (e.g., 
Huang & Liu, 2001; Anand, 2008). Reanalysis has been noted to cause delays in SAT 
dynamics parameters (McElree, 1993; Bornkessel et al 2004), and thus if the first 
option that comprehenders attempt upon recognizing ziji is the local-antecedent 
interpretation, they should fail and require lexical reanalysis in the long-distance 
antecedent conditions. 
Note, however, that the LAN we observed in the ERP record is not associated 
with general syntactic reanalysis. Instead, the ERP component that is commonly 
associated with reanalysis processes is the P600 (Friederici et al., 2002; Hagoort, 
2003; Hopf et al., 2003; Kaan & Swaab, 2003), though it has been argued  that certain 
subcases of reanalysis (specifically, reanalysis related to case marking) engender 
N400 effects (Hopf et al 1998; Bornkessel et al 2004). On the assumption that local 
 
 
251 
and long-distance interpretations of ziji required distinct lexical items and structural 
analyses, then it can be argued that P600 effects should be observed. Thus given 
relatively well-accepted assumptions about the functional significance of common 
language-related ERP components, it can be argued that the processing slowdown 
related to long-distance interpretations of ziji does not reflect general syntactic 
reanalysis mechanisms.  
 As mentioned above, however, it is unclear whether an account that invokes 
reanalysis is an entirely distinct alternative to accounts that invoke structured search. 
In particular, the notion of structured search that we have suggested posits that in 
order to recognize that a local antecedent for ziji is inappropriate in a sentence with a 
long-distance antecedent for ziji, the local subject position must first be retrieved and 
rejected due to its unacceptability as an antecedent for ziji (due to either 
morphosyntactic constraints, or possible discourse-level interpretive constraints). It is 
not unreasonable to call this intermediate step in a structured search procedure a step 
of ?reanalysis.? The main conceptual difference between a reanalysis account and a 
structured search account is that in the latter, the local subject is initially targeted for 
access using targeted cues that single out the local subject. This need not be so on a 
reanalysis-style account of these results, which instead would posit that the local 
subject is initially retrieved because more local phrasal material is more active in the 
parse. However, this local activation advantage would need to be so great as to cause 
access of the local subject during retrieval despite the fact that the long-distance 
subject has a number of semantic features (e.g., animacy, sentience, being a source of 
communication) that are known to invite anaphoric co-reference (Kaiser, Runner, 
 
 
252 
Sussman, & Tanenhaus, 2009). In contrast, the local subject contains no semantic 
cues to support retrieval. None of the nouns used in this position supported 
metonymic interpretations that could license ziji, such as the corporation being used 
to refer to the corporation?s employees. Such an account seems unlikely in light of the 
fact that the local subject is not always reliably retrieved in other instances where it 
needs to be retrieved, such as English subject-verb agreement. In computing subject-
 verb agreement, an inaccessible feature-match causes incorrect access during online 
parsing, showing that the feature match can in fact overcome any bias for the local 
subject. The data presented here thus seem to favor structured access strategies over 
reanalysis causes by general memory dynamics and a preference for material 
contained in the local clause. 
 A related alternative account might be formulated in terms of the ACT-R 
model in Chapter 3, which posits retrieval time varies as a function of the fit to the 
search cues. It may be the case that it is simply more difficult to retrieve the more 
distant antecedent, but that the intermediate antecedent is not in fact considered in the 
course of recovering the distant antecedent. However, this alternative explanation 
requires the search cues to be limited to structural cues only. Otherwise, the semantic 
cues contributed by the distant antecedent would potentially make it a much better fit 
with the retrieval engaged by ziji. This would predict that the long-distance feature-
 matched antecedent of ziji is retrieved more quickly than the local feature-
 mismatched noun phrase. But if only structural search cues are used, it is unclear on 
what grounds comprehenders retrieve the distant subject over the local subject. It 
seems that on a structured access account with serial retrieval of memories, there is a 
 
 
253 
significant portion of trials for which the local subject with be retrieved, leading to a 
situation where structured search is necessary to recover the distant interpretation. 
 
Linguistic and discourse antecedents for ziji 
 
 One consistent finding from both Experiments 6 and 7 is that sentences that 
do not contain an explicit linguistic antecedent for ziji are not consistently rejected as 
unacceptable by speakers. This is not entirely surprising, as it is possible to interpret 
ziji as a first-person indexical that refers to the speaker of an utterance. It is not clear 
what relation the process of arriving at this egocentric interpretation bears to the 
process of finding a linguistic antecedent in memory. In Experiment 7, the no 
antecedent condition elicited an apparent N400 relative to both local and LD 
antecedent conditions. Interestingly, the no antecedent condition appeared to pattern 
in between LD and local antecedent binding conditions with regards to the anterior 
negativity observed. If the time course implied by the ERP components reflects the 
order of operations, then it suggests that the egocentric interpretation is not an 
?elsewhere? interpretation that is adopted only after an exhaustive search of the parse 
fails to return a licit antecedent. Instead, it would suggest that participants engage a 
quick decision to pursue either a bound or indexical interpretation of ziji before any 
antecedent reactivation occurs. Interestingly, the offline judgment data suggests that 
this interpretation is not as readily accepted as interpretations with explicit linguistic 
antecedents, which may suggest some difficulty in recovering this interpretation. This 
apparent difficulty may be an artifact of the tasks used in the present set of studies: it 
 
 
254 
is difficult to know what perspective participants adopt when interpreting out-of-
 context sentences in a laboratory setting. Further work that explicitly manipulates the 
perceived perspective of the sentences may be useful in determining the source of this 
difficulty in the present tasks. 
If it is the case that the egocentric interpretation is considered before an 
exhaustive search of the parse, then this raises the question of what information 
allows comprehenders to decide between linguistically binding ziji and adopting an 
egocentric interpretation. One possibility is that top-down information about the 
discourse model is applied to disambiguate early: in the case where there are no 
sentient entities in the discourse, the egocentric interpretation is the only licit option. 
In other cases of structural ambiguity, it has been argued that comprehenders use the 
number of discourse entities in a heuristic fashion to decide which parse to pursue 
(van Berkum, Brown & Hagoort 1999b). Thus, in a context with two girls, the 
ambiguous input John told the girl that?, comprehenders quickly pursue a relative 
clause interpretation of that. In a context with only one girl, however, that is 
preferentially parsed as introducing a sentential complement to the verb told. A 
similar discourse heuristic might be used to disambiguate which interpretation of ziji 
is to be pursued. If the discourse contains appropriate sentient entities that could 
potentially antecede ziji, then comprehenders deploy structured search to check that 
those discourse entities stand in an appropriate structural relation to the anaphor. A 
discourse-based heuristic for making a decision about how to parse ziji generates 
testable predictions. If the discourse contains a sentient entity that is linguistically 
introduced in a syntactically inaccessible position, then access to egocentric 
 
 
255 
interpretations of ziji should be inhibited relative to sentences that introduce no 
sentient actors to the discourse model. In the case of an inaccessible antecedent, 
comprehenders may be tempted into considering a bound interpretation for ziji that is 
not globally correct. 
At this point this suggestion remains speculative, as the present study does not 
provide evidence to support this notion. The relationship between linguistically and 
non-linguistically bound ziji in online processing routines remains an important 
unresolved question, and I leave this to future research. 
 
Conclusion 
 
 In this chapter I examined the time-course of antecedent-anaphor dependency 
construction using the Mandarin Chinese long-distance anaphor ziji. It was found that 
local antecedents are accessed more rapidly than long-distance antecedents, 
suggesting that the information necessary to complete local antecedent dependencies 
is present before the information needed for long-distance antecedent dependencies. 
This finding was supported by converging evidence from ERPs, which showed that 
long-distance bindings of ziji elicited a LAN component relative to local antecedent 
bindings.  
These findings are compatible with several implementations, but the crucial 
features that all accounts share is that the local subject is reactivated using primarily 
structural cues, without regard to the semantic fit of their content. This satisfies an 
important prediction of the hypothesis of structured access: as in English reflexives in 
 
 
256 
Chapters 2 and 3, it is structural position, rather than morphological or semantic 
compatibility with the anaphor, that guides memory access. But while the argument 
from interference showed that structured access provided a more restricted initial 
candidate set, the evidence presented here demonstrated the complementary 
prediction of structured access. For long-distance reflexives like Mandarin ziji, 
structured access actually leads to an inappropriately large initial candidate space 
relative to a feature-based access mechanism. In the SAT and ERP experiments 
presented here, I showed that this larger potential candidate space causes measurable 
delays in processing speed, suggesting that comprehenders serially check potential 
antecedent positions. 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
257 
 
Chapter 5: Structured access across dependency types 
 
 
 
 
 
 
 
 
 
 The preceding chapters have argued that it is syntactic position, rather than 
morphological or semantic feature content, that guides memory access in both 
English argument reflexives and Chinese long-distance reflexives. For both of these 
dependencies, the hypothesis of structured access appears to accurately describe the 
manner in which they are parsed. There remain a number of important questions 
about structured access in sentence comprehension, however. In particular, it does not 
appear to be the only manner in which information is retrieved from memory in 
online parsing: agreement dependencies have been repeatedly shown to use 
morphological features in a direct-access fashion (Clifton et al 1999; Pearlmutter et al 
1999; Wagers et al 2009; Chapter 2 of the present thesis). The selective deployment 
of structured access mechanisms for constructing long-distance dependencies leads to 
an important theoretical question: what is the role of structured access in 
comprehension, and how generally is such a mechanism used in online processing?   
 
 
258 
 This chapter presents a first attempt at addressing this question. There is one 
straightforward hypothesis that suggests itself based on the contrast between 
agreement and reflexives. Namely, it may simply be the case that anaphoric 
dependencies are constructed at a different level of representation than are agreement 
dependencies. For example, agreement might be computed over the syntactic parse, 
but anaphoric relations might instead be stated as interpretive rules over logical form 
representations (Jackendoff 1972; Wasow 1972; Chomsky & Lasnik 1993; Fiengo & 
May 1994). If structured access is simply a property of referential dependencies, then 
structured access should hold equally well for all anaphor-antecedent dependencies. 
In order to test this hypothesis, in this chapter I contrast the processing of ziji with the 
intensified pronoun ta-ziji that has related but distinct licensing properties from ziji 
(Pan 1998, 2000; Huang 2000; Bergeton 2007). Self-paced reading evidence confirms 
that ziji initially accesses the local subject position before considering other subject 
positions. However, in a closely matched experiment, ta-ziji shows a qualitatively 
different processing profile in identical contexts, suggesting immediate direct access 
to all licit antecedent positions in parallel.  
The contrast in behavior between ziji and ta-ziji provides important insight 
into the role of structured access in comprehension. It demonstrates that structured 
search is not deployed for all anaphoric dependencies: ziji and ta-ziji both participate 
in anaphoric dependencies, and both are presumably of similar relevance to the task 
of computing the message intended by any given utterance. Nonetheless, they display 
qualitatively different access profiles, casting doubt on the hypothesis that all 
interpreted dependencies use structured access mechanisms. The processing profile 
 
 
259 
associated with structured memory access is not a general property of interpretive 
procedures (as opposed to morphological or syntactic dependencies).  
These findings suggest that interpretive content is not a sufficient condition 
for the parser to engage structured search mechanisms. In the last part of this chapter, 
I consider the possibility that interpreted dependencies that are subject to syntactic 
licensing conditions are likely to engage structured search. If this is so, then 
structured search may be understood as an optimal strategy for satisfaction of 
grammatical constraints in online parsing processes. If the goal of parsing is to 
reconstruct the intended meaning of the input string as quickly and accurately as 
possible, then a more narrowly syntactic access mechanism may convey advantages 
in speed or accuracy when a) a dependency requires a constrained search of the parse 
space, and b) when failure to build the correct dependency would result in 
misunderstanding.  
If this line of reasoning is correct, then structured access is an adaptive 
strategy to minimize interference and maximize interpretive fidelity in a noisy 
memory environment. By narrowing retrieval cues to the minimal distinctive set 
necessary to accomplish the task at hand, structured access minimizes interference. 
The selective deployment of this strategy reflects the fact that structured access can 
be costly in terms of processing time, as seen in Chapter 4, and so is only worth 
engaging when syntactically constrained dependencies contribute to meaning. Even 
this restricted characterization of the mechanism predicts a wide application of this 
mechanism, however: general thematic integration operations, most syntactic 
attachment decisions, and bound anaphor dependencies (among others) are all 
 
 
260 
syntactic operations that have direct interpretive reflexes as well as syntactic licensing 
conditions. Across the range of these dependencies, structured access should be 
apparent.  
The ?optimal strategy? argument for abstract structural retrieval cues is similar 
to arguments in Mitchell and colleagues? Tuning Hypothesis (Mitchell, Cuetos, 
Corley & Brysbaert 1995). On both accounts, the pressure to state parsing procedures 
at an abstract structural level is motivated by functional pressures, either the need to 
derive robust structural frequency estimates (Mitchell et al 1995) or to minimize 
interference from distracting elements in memory (on the present account). 
 
A puzzle for the hypothesis of structured access in comprehension 
 
 In Chapters 2-4, it was seen that syntactic information provides the primary 
means of accessing memory for certain dependencies. This argument rests on the 
observation that English reflexives and Mandarin long-distance reflexives are able to 
initiate accurate retrieval of the local subject of their clause, apparently without 
interference from other feature-matched noun phrases in the parse tree, whether or not 
they occupy a licit antecedent position. The crucial conclusion these studies license is 
that the parser can in principle accurately target and retrieve particular syntactic 
positions, at least for the dependencies under consideration.  
 However, this conclusions poses a puzzling question in light of the results in 
Chapter 2: if it is the case that the parser can in principle accurately retrieve and 
check the local subject, why don?t comprehenders do this when checking subject-verb 
 
 
261 
agreement in English? If comprehenders can accurately index the local subject for 
retrieval during sentence processing, then agreement processing could plausibly 
proceed by accurately retrieving the local subject and then verifying that its features 
match the features of the verb. However, this does not appear to be the strategy that 
comprehenders pursue: subject-verb agreement in English reliably shows interference 
effects, which suggests that morphological features are used to access memory in a 
content-adressable fashion.  Why is it that comprehenders engage different access 
strategies for the two dependencies? 
 One intuitive possibility is that since reflexives participate in a semantically 
interpreted dependency, they initiate a retrieval of the local subject in a qualitatively 
different manner. It may simply be the case that reflexive dependencies and 
agreement dependencies are constructed at different levels of representation, as would 
be expected if reflexive dependencies are best understood as constraints on 
interpretation, rather than structure-building processes (Jackendoff 1972; Wasow 
1972; Chomsky & Lasnik 1993). If one makes a secondary assumption that the parser 
constructs somewhat shallower or less structured parses before shunting constituents 
off to more structured stores for interpretation (Frazier & Fodor 1978), then this 
prediction could be easily accommodated. Fallibility may simply reflect the memory 
access mechanisms that are deployed to retrieve information from uninterpreted, 
temporary syntactic stores. In contrast, structured access mechanisms may be the 
preferred manner of access for structure built at a more compact or global level of 
representation. A number of well-studied parsing models have an architecture in 
which material is associated in a less structured manner before being passed to a more 
 
 
262 
compact, global, or structured representation (see, e.g., the two-stage model in Frazier 
& Fodor 1978 or Townsend & Bever 2001). For any parsing model that posits 
separate levels of representation that differ on the degree to which linguistic 
information is structured in memory, then the distinction between agreement and 
reflexives might simply reduce to differences in the access strategies employed by 
different representational levels in the parse. 
 This account appears to capture a straightforward intuition behind the 
agreement and reflexive dependency contrast: reflexives mean something, and 
agreement generally does not, at least in English. To evaluate this intuition, consider 
the range of grammatical dependencies that have been shown to be susceptible to 
facilitatory interference, briefly summarized in Table 5.1. If the reasoning in Chapter 
3 is correct, then those dependencies in 5.1 that do not show facilitatory interference 
are those that are constructed with structured access mechanisms. The range of 
dependencies that have been considered remains somewhat sparse; conclusions drawn 
from such a narrow sample should be approached with caution.  
 
 
Table 5.1: Summary of interference properties of long-distance dependencies.   
 
 Facilitatory 
interference 
English subject-verb agreement   ?  
Spanish subject-verb agreement ? 
English reflexives ? 
Mandarin Chinese reflexives ? 
German, English NPIs ? 
 
 
263 
Subject-verb agreement in English provides perhaps the clearest example of a 
dependency that is clearly prone to grammatical illusions due to feature-matched, but 
inaccessible material. Similar comprehension results have been found in Spanish 
(Alcocer & Phillips 2009; Lago, Alcocer & Phillips 2011). Another example of a 
dependency that is prone to grammatical illusions is NPI dependencies, both in 
English  (Xiang, Dillon & Phillips 2006, 2009), and in German (Drenhaus et al 2005; 
Vasishth et al 2008). This appears to run counter to the generalization that structured 
access holds of interpreted or semantic dependencies. However, it is not clear that this 
represents the same phenomenon as agreement interference. For instance, Xiang and 
colleagues (2009) argued that the NPI effect was not due to a feature overlap with an 
inaccessible negative element. Instead, they argued that the best account of these data 
was one that invoked the generation of spurious pragmatic inferences due to the 
interaction of embedded negation and the semantics of the restrictive relative clause. 
Additional support for this view comes from a number of subsequent studies. It has 
been shown that NPI interference is negatively correlated with an individual?s 
pragmatic reasoning skills (as measured by the Autistic Quotient; Baron-Cohen, 
Wheelwright, Skinner, Martin & Clubley 2001), which supports the claim that 
?overactive? pragmatics is the source of this effect. For individuals who have 
impaired pragmatic abilities, as indexed by large AQ scores, the NPI interference 
effect is diminished, while the agreement attraction effect is unaffected (Xiang, Grove 
& Giannakidou 2011). Additionally, self-paced reading evidence presented by Parker 
& Phillips (2011) suggests that negation embedded in complement clauses does not 
cause interference in the same manner as negation embedded inside restrictive 
 
 
264 
relative clauses, which has formed all of the evidence for NPI interference to date. If 
complement clauses change the pragmatics involved with these sentences without 
drastically changing the structure, then the overactive pragmatics account of NPIs is 
supported. 
If it is true that the NPI illusion effect is generated by an altogether separate 
mechanism, then agreement provides the only clear and reliable case of a long-
 distance dependency that is structurally fallible online. It appears that the range of 
evidence for feature-based access is rather limited. 
Although the evidence for feature-based, grammatically fallible memory 
access seems restricted, it must be acknowledged that the evidence for structured 
access is equally narrow. Just as agreement provides the only clear case of fallible, 
feature-based direct access dependency construction, reflexive dependencies in 
English and Chinese as investigated here provide the clearest examples of structured 
access in Table 5.1. Given the data that is currently available, the question over the 
role of structured access in comprehension apparently reduces to a comparison 
between verbal agreement and reflexives. Initial hypotheses about the role of 
structured access in comprehension must at a minimum account for this distinction, 
but there are many dimensions on which the two dependencies differ. As suggested 
above, perhaps the most salient difference between the two dependencies is that only 
reflexives involve a referential dependency. This intuition leads to a natural first 
hypothesis about the range of dependencies that engage structured access: if a 
dependency involves constructing an interpreted anaphoric dependency then the 
parser operates in a more structured, accurate fashion. 
 
 
265 
Experiments 8 and 9 test this view by comparing the Mandarin long-distance 
reflexive ziji to the related Mandarin anaphor ta-ziji. This comparison is interesting 
because both anaphors have very similar licensing environments, although they have 
been argued to have very different underlying structure (Pan 1998, 2000; Bergeton 
2007). Both may be used as reflexive anaphors bound by the local subject (Huang, Li 
& Li 2009).  However, unlike ziji, ta-ziji is more likely to function as a contrastive 
pronominal element in many environments, leading to the possibility that apparently 
syntactic constraints on ta-ziji are epiphenomenal (K?nig & Siemund 1999; Bergeton 
2007). If ziji and ta-ziji pattern alike with respect to the access of non-local elements, 
then the hypothesis that structured access is a general property of interpreted 
anaphoric dependencies is supported. If they pattern differently, however, then this 
would suggest that structured access is deployed for satisfaction of a narrower range 
of grammatical constraints. 
 
Revisiting Chinese anaphors 
 
Recall that Mandarin Chinese ziji is a long-distance reflexive: it is subject to a 
number of structural constraints, requiring either a linguistic antecedent in a certain 
structural configuration or defaulting to an indexical interpretation. In a sentence like 
(5.1), ziji can be bound by either the local or the long-distance antecedent. There is 
broad agreement that linguistically bound ziji is essentially a syntactic anaphor, even 
if there are additional discourse-pragmatic constraints on potential antecedents 
 
 
266 
(Huang et al 2009). In accord with this characterization, I showed in Chapter 4 that 
ziji engages a structured search through the parse space to find its antecedent. 
 
 (5.1)  Zhangsani  shuo Lisij  nongshang-le  zijii/j 
   Zhangsan says Lisi harm-PERF self 
    ?Zhangsan says that Lisi harmed him / himself? 
 
In addition to ziji, however, there is another Mandarin pronominal that may be 
used to indicate that two arguments of a verb corefer: ta-ziji. Many discussions of ta-
 ziji simply state that it is a Principle A anaphor with a distribution that more or less 
mirrors English himself (Chomsky 1981), without further discussion (see, e.g., Huang 
& Liu 2001; Huang et al 2009). This characterization appears to hold in (5.2), where 
unlike ziji, ta-ziji cannot be bound by the distant antecedent. Ta-ziji is a bimorphemic 
compound that consists of the pronoun ta affixed with the reflexive element ziji, 
meaning literally him-self. The characterization of ta-ziji as a local anaphor squares 
well with a well-known cross-linguistic generalization about the morphological 
complexity of long-distance versus local reflexives. It has been noted that 
morphological complexity correlates with locality restrictions (Pica 1986). Simplex 
anaphors more often tend to allow long-distance binding, whereas morphologically 
complex forms are generally subject to stricter locality conditions. This fact has been 
taken as support for a head-movement analysis of long-distance binding, which 
predicts the phrase/head asymmetry in locality restrictions (Cole & Sung 1994). For 
this reason, from a typological point of view the analysis of ta-ziji as a principle A 
anaphor makes a lot of sense. 
 
 
 
267 
 (5.2)  Zhangsani  shuo Lisij  nongshang-le  taziji*i/j 
   Zhangsan says Lisi harm-PERF self 
   ?Zhangsan says that Lisi harmed himself/*him? 
 
However, this characterization of ta-ziji obscures more complex licensing 
conditions. The distribution of ta-ziji and English local reflexives diverge when a 
wider range of data is considered, weakening the case for ta-ziji as a Principle A 
anaphor. For instance, long-distance bindings of ta-ziji are readily obtained simply by 
altering the animacy of the local subject in (5.2): 
 
 (5.3)  Zhangsani  shuo naben  shuj  nongshang-le  tazijii/*j 
   Zhangsan says that-CL book harm-PERF self 
   ?Zhangsan says that book harmed him? 
 
No such contrast is found for English reflexives (c.f. *John said that the book 
hurt himself). A number of authors have suggested that in light of data like (5.3), the 
characterization of ta-ziji as a Principle A anaphor cannot be maintained (Pan 1998, 
2000; Huang 2000; Bergeton 2007). The exact nature of ta-ziji remains contentious, 
but existing accounts share a common insight: ta-ziji preferentially takes ?prominent? 
antecedents, where prominence is defined as some mixture of lexical semantic 
properties (e.g. animacy), structural properties (locality, dominance) and discourse 
properties (the ability to be contrasted with a contextually relevant set of individuals). 
For example, Pan (1998, 2001) accounts for this data by positing that the relevant 
notion of prominence is determined by a graded scale of animacy, ranging from 
animate human entities to inanimates. Given this notion of prominence, Pan suggests 
that the binding domain for ta-ziji is computed relative to the highest, most prominent 
 
 
268 
NP. Bergeton (2007) suggests, instead, that ta-ziji is essentially an intensified 
pronominal identical to he himself in English (a view that was endorsed by Tang 1989 
as well). On his account, intensification is licensed when the antecedent may be 
contrasted with a set of alternatives in the discourse. Interestingly, even for 
unambiguously pronominal elements in English (intensified pronouns such as he 
himself) a contrast similar to that in (5.2) obtains: 
 
 (5.4)  a. ?John thinks Mary said that he himself went to the 
store. 
 b. John thinks Mary said that she herself went to the store. 
 
(5.4a) is anomalous, except on the marginal reading where the focus domain 
associated with the intensifier himself is the VP (i.e. John went to the store himself). 
The contrast in (5.4) supports Bergeton?s claim that the discourse licensing conditions 
on intensification can mimic syntactic locality conditions (for similar observations, 
see Baker 1995; Zribi-Hertz 1995; K?nig & Siemund 1999). This mimicry is apparent 
even for English subject pronominals.  
On either style of account, the strict locality of the binding domain for ta-ziji 
suggested by examples like (5.2) is epiphenomenal: ta-ziji may well take long-
 distance antecedents if they are sufficiently prominent, either morphosyntactically or 
in a discourse context. Similar conclusions have been reached for Norwegian sig selv 
by L?drup (2009). The distribution of sig selv and L?drup?s account for this 
distribution make sig selv appear quite similar to ta-ziji. The intensification conveyed 
by the extra selv or ziji element appears to mimic Principle A effects, although given 
 
 
269 
appropriate configurations of animate discourse entities, long-distance readings are 
readily obtained. 
If ta-ziji is an intensified pronominal whose licensing conditions are stated 
primarily in terms of animacy (Pan 1998, 2000) or discourse (Bergeton 2007) 
prominence, then unlike ziji, there is no need to limit antecedent search to structurally 
licit positions. If the apparent structural constraints are epiphenomenal, then simply 
accessing antecedents that are prominent on the relevant dimensions would achieve 
the desired result. Prominence, rather than solely structural position, is the more 
direct cue to ta-ziji?s antecedent, and so it is reasonable to think that an effective 
parsing strategy for ta-ziji will make direct reference to these cues rather than (or in 
addition to) syntactic structure.  
If this line of argumentation is correct, then ziji and ta-ziji make an interesting 
minimal pair to test the domain of application of structured access procedures. In 
many cases, the binding possibilities for these anaphors overlap, but there is reason to 
suspect that the locality restrictions on ziji and ta-ziji are of a qualitatively different 
nature. If structured access is a strategy that is applied to interpreted dependencies, 
regardless of their underlying grammatical constraints, then we expect ziji and ta-ziji 
to both show the characteristics of structured search: preferential access of local 
subject positions, and delayed access to distant or inaccessible positions. However, an 
alternative possibility is that structured search is more narrowly applied to 
dependencies that have particular structural requirements on them. If it is the nature 
of the underlying structural constraints that drives structured access in reflexive 
dependencies, then ta-ziji should not show the behavioral signature of structured 
 
 
270 
access. Instead, we expect to observe direct access to licit, prominent antecedents 
regardless of position in the structure, based solely on their feature-compatibility with 
the anaphor.  
Experiments 8 and 9 test the behavioral profile of these two anaphors to 
determine whether structured access reflects a fact about interpreted dependencies 
more generally, or if it instead is an expression of the particular linguistic constraints 
on ziji. Of critical interest is whether or not distant antecedents for ta-ziji are accessed 
in a direct-access manner, or if they pattern with ziji showing structured access and 
impeded processing when the antecedent is structurally distant. 
 
The scope of structured access: contrasting ziji & ta-ziji 
   
 In order to test the manner in which ziji and ta-ziji access their antecedents, I 
focus on the processing difficulty associated with accessing sub-commanding 
antecedents. Sub-command refers to structural configurations where non-c-
 commanding antecedents may bind an anaphor, provided that the antecedents meet 
certain structural conditions. Sub-command has been long noted in descriptions of the 
licensing conditions on ziji (Tang 1989). Consider the following example from Tang 
(1989, p  100): 
 
 (5.5)  [Zhangsani tou dongxi de] shishi bei zijii  de laoban faxian-le. 
  [Zhangsan steal thing DE] fact     by  self DE boss   discover-
 PERF 
?The fact that Zhangsan stole something was discovered by his 
boss.? 
 
 
 
271 
 In this example, Zhangsan is able to bind ziji, despite the fact that it does not 
c-command the anaphor; it is not a sister to a node that dominates the anaphor at any 
level of representation. C-command has long been noted as an important condition on 
reflexive binding and quantifier-variable relationships (Reinhart 1976). However, the 
notion of c-command is not sufficient to account for the range of structural positions 
that a bound variable?s antecedent may occupy in Mandarin. To account for the 
Mandarin data, Tang (1989) formulated the subcommand condition as in (5.6), which 
generalizes the c-command constraint to include examples such as (5.5) above. Tang 
notes that subcommand is furthermore subject to the constraint that embedded NPs 
cannot antecede ziji if they are contained within a potential binder of ziji. Thus in 
(5.5), because shishi  ?fact? is not a potential binder of ziji, Zhangsan may bind ziji in 
virtue of it being contained in a commanding subject position.  
 
 (5.6)   ? sub-commands ? iff  
   a. ? c-commands ?, or  
b. ? is an NP contained in an NP that c-commands ? or that sub-
 commands ?, and any argument containing ? is in subject position. 
 
 Because binding from embedded nouns is only possible when the local subject 
is not a potential binder for either ziji or ta-ziji, the binding possibilities for embedded 
NPs is dependent on the identity of the head noun, as seen in examples (5.7) and 
(5.8). In these sentences the binding possibilities for ziji and ta-ziji are the same. 
When the local subject is animate (?seamstress? in 5.7), binding from the embedded 
NP is not possible for either anaphor. However, for both anaphors when the local 
subject is inanimate, the embedded subject Mrs. Zhang is a licit binder.  
 
 
272 
The pair of sentences in (5.7) and (5.8) provide an interesting test case that 
could be used to test the manner in which the two anaphors ziji and ta-ziji access their 
antecedents. Note that when the local noun is the anaphor?s antecedent, as in (5.7), 
structural and semantic prominence are aligned. Because structural cues (i.e. local 
subject) and semantic/discourse prominence cues (i.e. salient animate entity) align in 
this case, both ta-ziji and ziji should behave identically. Even if ziji and ta-ziji use 
qualitatively distinct mechanisms to access their antecedent, for sentences like (5.7) 
they should display similar behavior: they both immediately access the local subject.  
 
Figure 5.1: Structured search for ziji forces the parser to consider the local antecedent 
position before the sub-commanding antecedent. 
S
 NP VP
 RC N V NP
 S de boutique harmed ziji
 NP VP
 Lisi V NP
 visit t
 +sub-command
 +subject
 +sub-command
 +subject
 
 
273 
 
Figure 5.2: Direct access for ta-ziji based on semantic or discourse prominence 
allows the parser to immediately access the sub-commanding antecedent. 
 
Interestingly, the predictions of feature-based direct access and structured 
access diverge when it comes to processing binding from a subcommand position, as 
in (5.8). If structured search is deployed, the embedded subject position cannot be 
directly accessed. Instead, the local subject boutique in (5.8) must be checked, and 
only after its incompatibility is verified can the embedded position be accessed. This 
is schematized for sentence (5.8) in Figure 5.1. Alternatively, if feature-based direct 
access is employed, then access should occur in a single processing step, as in Figure 
5.2. Thus in order to determine whether structural or feature-based access of 
antecedents is engaged for either anaphor, sentences like (5.8) provide the key piece 
of data. The structured access pattern in Figure 5.1 is expected to hold for ziji based 
on the structured search strategy findings in Chapter 4. Since only structural 
S
 NP VP
 RC N V NP
 S de boutique harmed ta-ziji
 NP VP
 Lisi V NP
 visit t
 +sub-command
 +animate
 +human
 
 
274 
information is used to initially bind ziji, then the local subject position should be first 
checked and rejected, requiring extra processing to access the embedded sub-
 commanding position.  
 
 (5.7)  [Zhang taitaii  guanggu de] n?caifengj hai-le  ta-
 ziji*i/i/ziji*i/i. 
  [Mrs. Zhang    visit     DE] seamstress harm-PERF ta-ziji / 
ziji. 
?The seamstress that Mrs. Zhang visits harmed herself.? 
 
 (5.8)  [Zhang taitaii  guanggu de] shizhuangdianj hai-le  ta-
 zijii/*i/zijii/*i 
  [Mrs. Zhang    visit     DE] boutique           harm-PERF ta-ziji / 
ziji. 
?The boutique that Mrs. Zhang visits harmed her.? 
 
The crucial question for the current purposes is whether or not ta-ziji accesses 
the embedded position in the same manner. As described above, there is good reason 
to suspect that ta-ziji is preferentially processed as an intensified or emphatic 
pronoun. If semantic prominence is the primary constraint on licit ta-ziji-antecedent 
relationships, then it is reasonable to expect that ta-ziji accesses its antecedent directly 
using prominence information. A direct-access mechanism for ta-ziji would leverage 
discourse or semantic prominence to recall its antecedent, and thus would access the 
embedded position in a single processing step (Figure 5.2). 
Determining whether ta-ziji engages a structured search as in Figure 5.1, or a 
direct-access retrieval for accessing its antecedent as in Figure 5.2, provides a first 
step in determining the range of application of a structured access mechanism in 
comprehension. If ta-ziji patterns like ziji in accessing sub-commanding antecedents 
in a structured manner, then the hypothesis that structured access is employed in 
 
 
275 
interpreted dependencies is supported. However, if ta-ziji accesses its antecedents in a 
direct-access manner, then it would suggest that structured access is deployed to 
satisfy a narrower range of grammatical constraints. 
Following arguments laid out in Chapter 2 and 3 above, in order to determine 
if feature-based or structured access to a distant or inaccessible antecedent is 
deployed, it is crucial to contrast the processing of the anaphor with a distant 
antecedent (5.9a) to a baseline where there is no antecedent linguistically represented 
(5.9b).  
 
 (5.9) a. [Zhang taitaii  guanggu de] shizhuangdianj hai-le  ta-
 zijii/*i/zijii/*i 
  [Mrs. Zhang    visit     DE] boutique           harm-PERF ta-ziji / 
ziji. 
?The boutique that Mrs. Zhang visits harmed her.? 
 
  b. [Meiti   baodao  de] shizhuangdianj hai-le  ta-
 zijii/*i/zijii/*i 
  [Media   report   DE] boutique           harm-PERF ta-ziji / 
ziji. 
?The boutique that the media reported on harmed her.? 
 
The logic of this comparison is identical to that in the experiments presented 
in Chapters 2 and 3. If direct access employed discourse or semantic features is 
deployed for ta-ziji, then (5.9a) should be easier to process than (5.9b) at the point of 
processing the anaphor?s reference. On the other hand, if structured search if used, 
they should be at least equally difficult upon reaching the anaphor, reflecting the fact 
that they both appear unacceptable when only the local subject is considered.  
 
 
276 
Experiment 8 tests the processing impact of sub-commanding binding of ziji 
using self-paced reading, and Experiment 9 tests identical environments with ta-ziji. I 
delay discussion of either experiment until after both have been presented. 
 
Experiment 8: Ziji and sub-commanding antecedents  
   
Participants 
 
41 students from the University of Maryland community participated in the 
experiment. All participants were native Mandarin Chinese speakers from mainland 
China, and all had normal or corrected-to-normal vision. They were paid $10 per hour 
for their participation in the experiment. 
 
Stimuli 
 
 The experimental materials consisted of four conditions designed to investigate 
the effect of semantically coherent, but sub-commanding antecedents on the 
processing of ziji. There were two potential antecedent positions: the local subject 
position and the embedded subject position. As in Experiments 6 and 7, binding 
possibilities were manipulated by manipulating the animacy of the two subject 
positions. The conditions are summarized in Table 5.2; the local subject factor refers 
to the animacy of the local subject, and the distant subject factor refers to the animacy 
of the embedded subject. I refer to the [+local,+distant] condition as the multiple 
 
 
277 
match condition. The [+local,-distant] and [-local,+distant] conditions were the local 
match and distant match conditions, respectively. The no match condition had no 
animate NPs ([-local,-distant]). The distant (sub-commanding) antecedent position 
was the subject of an object relative clause that modified the main clause subject, 
This antecedent position was always the first word in the sentence, due to the head-
 final order of Mandarin relative clauses. The local (main clause) subject was always 
in fourth position.  
 In order to avoid wrap-up effects at the critical region, the ba construction was 
used. This construction uses a particle ba to mark the direct object, which is then 
moved to preverbal position. So that the two arguments of the main clause were not 
linearly adjacent, a temporal adverbial was placed between the local subject and the 
ba-marked ziji. Between ziji and the final verb, a manner adverbial was placed in 
order to provide an extra spillover region.    
Condition Example 
Multiple match 
[+local,+distant] 
[??? / ?????] / ?? / ???  / ???? /  ?  / ?? / ??? / ???. 
Mrs. Zhang / often visit DE / that-CL / seamstress / last week / BA / ziji / not careful / harm-PERF. 
 ?The seamstress that Mrs. Zhang often visits carelessly hurt herself last week.? 
Local match 
[+local,-distant] 
[??/ ???] / ?? / ???  / ???? /  ?  / ?? / ??? / ???. 
Media / report on DE / that-CL / seamstress / last week / BA / ziji / not careful / harm-PERF. 
 ?The seamstress that the media reported on carelessly hurt herself last week.? 
Distant match 
[-local,+distant] 
[??? / ?????] / ?? / ???/ ???? /  ?  / ?? / ??? / ???. 
Mrs. Zhang / often visit DE / that-CL / boutique / last week / BA / ziji / not careful / harm-PERF. 
 ?The boutique that Mrs. Zhang often visits carelessly hurt her last week.?  
No match 
[-local,-distant] 
[??/ ???] / ?? / ???/ ???? /  ?  / ?? / ??? / ???. 
Media / report on DE / that-CL / boutique / last week / BA / ziji / not careful / harm-PERF. 
 ?The boutique that the media reported on carelessly hurt self last week.? 
 
Table 5.2: Critical conditions from Experiment 8. Region breaks are indicated by 
slashes.  
 
 
 
 
278 
 Of the four experimental conditions, only the no match condition required ziji to 
take an extra sentential antecedent. The multiple-match, local match, and distant 
match conditions were all grammatical with a bound ziji interpretation. 
18 sets of these four conditions conditions were produced, and distributed into 
four lists in a pseudo-Latin square fashion. They were combined with 77 fillers, 
including materials from an unrelated experiment, for a total of 95 sentences. The 
ratio of acceptable-to-unacceptable sentences varied slightly from list to list due to the 
pseudo-Latin square list, but remained between 45 and 55 % acceptable. The fillers 
included 10 sentences that contained ba followed by non-anaphoric NPs in order to 
prevent anticipation of ziji.  
 
Procedure 
 
Sentences were presented using a moving-window self-paced reading 
paradigm, using the Linger software. Each sentence was presented in black characters 
on a white screen, and no sentence was more than one line long. All sentences were 
presented using simplified Chinese characters. The sentences were segmented into 9 
regions according to native speaker intuitions about where best to insert boundaries 
(as marked in Table 5.2), resulting in regions that ranged from one character (i.e. ba) 
to at most 6 characters (i.e. yishuticaoguanjun, ?gymnastics champion?). Sentences 
initially appeared as a series of dashes that obscured the words, and by pressing the 
space bar participants were able to sequentially reveal each region for a self-
 determined amount of time. Each region was remasked after the participant finished 
 
 
279 
reading it. After each sentence, a comprehension question was presented in its 
entirety on the screen, and participants were instructed to press f for yes, and j for no. 
In the critical experimental sentences, the comprehension question queried a part of 
the sentence that did not have to do with ziji?s referential dependency. Feedback was 
given for incorrect responses. 
 
Offline judgments 
 
 In order to confirm that the embedded subjects are reliably accepted as 
antecedents for ziji (as reported in Tang 1989), the experimental materials were used 
for an offline judgment study. 22 participants were asked to judge the acceptability of 
the sentences they read on a 7-point scale, where 7 was completely acceptable and 1 
was completely unacceptable. Participants were instructed to judge the sentences with 
regard to whether or not they were acceptable in colloquial speech. The results are 
presented in Table 5.2. Data was gathered using the IbexFarm over the internet; 
students from Beijing Normal University?s psychology department were recruited as 
participants. 
 
Multiple match 
[+local,+distant] 
Local match 
[+local,-disant] 
Distant match 
[-local,+distant] 
No match 
[-local,-distant] 
4.77 (?0.29) 5.08 (?0.29) 4.15 (?0.41) 3.29 (?0.43) 
 
Table 5.2: Mean judgments and standard error by subjects for ziji Experiment 8 
rating study. Values are on a 7-point scale where 7 is perfectly acceptable, and 1 is 
completely unacceptable. 
  
 
 
280 
 A two-way repeated measures ANOVA by subjects revealed a significant 
main effect of local noun animacy (F(1,21) =  22.2, p < 0.001), as well as a 
significant interaction of local noun animacy with distant noun animacy (F(1,21) = 
10.2, p < 0.01).  In addition, there was a marginal main effect of distant noun animacy 
(F(1,21) = 3.08, p < 0.1). Resolving this interaction further using planned pairwise t-
 tests revealed a marginal difference between multiple match and local match 
conditions (t(21) = -1.92, p < 0.8), and a reliable difference between distant match 
and no match conditions (t(21) = 2.87, p < 0.01).  
 In addition, a repeated measures ANOVA on judgment times revealed that 
conditions with local antecedents were judged more quickly than conditions where 
the local antecedent was inanimate (F(1,21) = 4.50, p < 0.05).  
 These results confirm that sentences with embedded antecedents are 
considered to be more acceptable than sentences with no linguistically represented 
antecedent for ziji. Importantly, embedded antecedents for ziji were judged 
significantly better than no antecedent conditions, indicating that participants were 
willing to consider the embedded position, confirming the data in Tang (1989).   
 
Data Analysis 
 
 Reading times at the critical anaphor and the spillover region were submitted 
for statistical analysis. It is common practice in self-paced reading data to reject 
outlying data points based on a cutoff criterion (e.g., more than 2.5 standard 
deviations from the mean for a given region / condition pairing). This is necessary to 
 
 
281 
deal with outliers, which can significantly distort estimates of average reading time 
(Ratcliff 1993). Instead of adopting a rejection threshold, however, I instead log-
 transformed all reaction times prior to analysis. In addition to minimizing the impact 
of reaction time outliers, this method has the additional benefit of making the reaction 
time data close to normally distributed, satisfying an important assumption of linear 
model analysis (Gelman & Hill 2005). No data was excluded from analysis.  
 Statistical analysis was performed using mixed-effect linear regressions to 
assess the magnitude, direction, and reliability of the experimental factors on reading 
times. The experimental fixed effects in the models were the factors LOCAL (whether 
or not the sentence was grammatical), DISTANT (whether or not the embedded NP was 
plural), and their interaction. The fixed effects were coding using simple difference 
sum coding (inanimate levels were coded as -.5, animate levels as .5). Thus all 
reported coefficients reflect the magnitude of the difference between levels of a given 
factor (in log-transformed RT space). In addition to these fixed effects, I additionally 
considered random intercepts for subjects and items, as well as random slopes for the 
experimental fixed effects by subjects and by items. In all cases, the significance of 
non-experimental fixed effects and random effects was assessed, and I report in all 
cases the best-fit model (following Baayen et al 2008; Jaeger 2008). I leave the 
experimental fixed effect structure constant across all models, because these effects 
were theoretically motivated by the design of the study. For most analyses, in 
addition to the experimental fixed effects, the best model included only random 
intercepts for subjects and items, as well as all a fixed effect for trial order. Word by 
word reading times are presented in Figure 5.3; for presentational purposes, pre-
 
 
282 
critical regions are collapsed into three regions: relative clause (RC), head noun 
phrase (NP), and pre-critical prepositional phrase (PP ba).  
 
Results 
 
Mean reading times are presented in Figure 5.3. Statistical analysis of the 
critical and spillover regions is summarized in Table 5.3. At the critical ziji region, 
modeling revealed that conditions with a local inanimate subject were read 
significantly slower than those with animate local subjects (? = -0.063, SE: 0.026, 
pMCMC < 0.05). This effect was also observed at the spillover region (? = -0.178, 
SE: 0.034, pMCMC < 0.0001), in addition to a main effect of the embedded noun?s 
animacy (? = 0.094, SE: 0.034, pMCMC < 0.01). Planned comparisons revealed that 
the animate embedded nouns caused a significant slowdown when the head noun was 
inanimate (? = 0.149, SE: 0.049, pMCMC < 0.01), but not when it was animate  (? = 
0.040, SE: 0.049, pMCMC < 0.5).  
 
 ? SE t / z 
ziji    
LOCAL -0.063 0.026 -2.44* 
DISTANT 0.008 0.026 0.31 
LOCAL?DISTANT -0.011 0.052 -0.21 
    
Spillover    
LOCAL -0.178 0.034 -5.19*** 
DISTANT 0.094 0.034 2.73** 
LOCAL?DISTANT -0.110 0.069 -1.58 
 
Table 5.3: Summary of fixed effects for best-fit models at the critical ziji and 
spillover regions, including t-values  
 
 
 
283 
 
 
Figure 5.3: Region-by-region mean log reading times for Experiment 8. Error bars 
represent ?1 standard error, by participants, corrected for between-participant 
variance. 
 
Experiment 9: Ta-ziji and sub-commanding antecedents 
 
Participants 
 
70 students from the University of Maryland community participated in the 
experiment. All participants were native Mandarin Chinese speakers from mainland 
China, and all had normal or corrected-to-normal vision. They were paid $10 per hour 
for their participation in the experiment. 
6.
 0
 6.
 2
 6.
 4
 6.
 6
 6.
 8
 7.
 0
 7.
 2
 Mean log reading time, by region
 Region
 lo
 g(
 R
 T
 )
 RC NP PP ba ziji Adv V
    
No match
 Distant match
 Multiple match
 Local match
 
 
284 
 
Stimuli 
 
 The materials were largely identical to those from Experiment 8; for 
convenience they are repeated in Table 5.4. Two important changes were made to 
these materials. First, all instances of ziji were replaced with ta-ziji. Ta is a third-
 person pronoun that does not distinguish male, female, animate, and inanimate 
referents phonologically. However, gender and animacy distinctions are maintained 
in the writing system: ta (?) is used for male humans, ta (?) for female humans, and 
ta (?) for animals and inanimates. In order to prevent this extra morphological 
information from biasing selection of ta-ziji?s antecedent, the materials were 
additionally modified so that within an experimental item set, the animate nouns in 
each position were of the same gender. Thus, as in Experiment 8, the acceptability of 
the two referents as antecedents for the anaphor turned solely on their structural 
status. Half of the revised materials had male nouns, and the other half had female 
nouns.  
 
Condition Example 
Multiple match 
[+local,+distant] 
[??? / ?????] / ?? / ???  / ???? /  ?  /??? / ??? / ???. 
Mrs. Zhang / often visit DE / that-CL / seamstress / last week / BA / ta-ziji / not careful / harm-
 PERF. 
 ?The seamstress that Mrs. Zhang often visits carelessly hurt herself last week.? 
Local match 
[+local,-distant] 
[??/ ???] / ?? / ???  / ???? /  ?  / ??? / ??? / ???. 
Media / report on DE / that-CL / seamstress / last week / BA / ta-ziji / not careful / harm-PERF. 
 ?The seamstress that the media reported on carelessly hurt herself last week.? 
Distant match 
[-local,+distant] 
[??? / ?????] / ?? / ???/ ???? /  ?  / ??? / ??? / ???. 
Mrs. Zhang / often visit DE / that-CL / boutique / last week / BA / ta-ziji / not careful / harm-
 PERF. 
 ?The boutique that Mrs. Zhang often visits carelessly hurt herself last week.?  
 
 
285 
No match 
[-local,-distant] 
[??/ ???] / ?? / ???/ ???? /  ?  / ??? / ??? / ???. 
Media / report on DE / that-CL / boutique / last week / BA / ta-ziji / not careful / harm-PERF. 
 ?The boutique that the media reported on carelessly hurt herself last week.? 
 
Table 5.4: Critical conditions from Experiment 9. Region breaks are indicated by 
slashes.  
 
As in Experiment 8, 18 sets of these four conditions were produced, and 
distributed into four lists in a pseudo-Latin square fashion. They were combined with 
77 fillers for a total of 95 sentences. The ratio of acceptable-to-unacceptable 
sentences varied slightly from list to list due to the pseudo-Latin square design, but 
remained between 45 and 55 % acceptable. The fillers included 10 sentences that 
contained ba followed by non-anaphoric NPs in order to prevent anticipation of ziji.  
  
Procedure 
 
The experimental procedure was identical to Experiment 8. 
 
Offline judgments 
 
 19 additional participants were asked to judge the acceptability of the 
experimental sentences on  a 7-point scale, where 7 was completely acceptable and 1 
was completely unacceptable. Participants received the same instructions as in the 
judgment pre-test in Experiment 8. The results are presented in Table 5.5. Data 
collection and participant recruitment was identical to Experiment 8. 
 
 
 
 
286 
Multiple match 
[+local,+distant] 
Local match 
[+local,-disant] 
Distant match 
[-local,+distant] 
No match 
[-local,-distant] 
4.12 (?0.34) 4.96 (?0.27) 5.06 (?0.27) 4.03 (?0.33) 
 
Table 5.5: Mean judgments and standard error by subjects for ziji Experiment 8 
rating study. Values are on a 7-point scale where 7 is perfectly acceptable, and 1 is 
completely unacceptable. 
  
 A two-way repeated measures ANOVA by subjects revealed only a significant 
interaction of local noun animacy with distant noun animacy (F(1,18) = 32.5, p < 
0.001). Resolving this interaction further using planned pairwise t-tests revealed 
significant differences between multiple match and local match conditions (t(18) = -
 3.26, p < 0.01), and a reliable difference between distant match and no match 
conditions (t(18) = 4.88 p < 0.001).  
 A repeated-measures ANOVA on judgment times revealed no significant 
differences between the conditions.  
 The judgment results show that sentences with embedded antecedents for ta-
 ziji are considered to be as acceptable as sentences with local antecedents; both are 
more acceptable than sentences with no linguistically represented antecedent. 
Interestingly, the multiple match condition was rated as significantly worse than the 
local match condition. One interpretation of this result is that the intensified 
pronominal is less felicitous in an out-of-the-blue discourse context where there are 
two equally acceptable antecedents it might refer to. If true, this finding may be taken 
to support the view that ta-ziji is preferentially perceived as an intensified, contrastive 
pronoun. 
 
 
 
287 
Data Analysis 
 
 Data analysis was identical to Experiment 8. 
 
Results 
 
Mean reading times are presented in Figure 5.4; as in Figure 5.3, pre-critical 
regions are collapsed for purposes of presentation. Statistical analysis of the critical 
and spillover regions is summarized in Table 5.6. At the critical ziji region, there were 
was a significant effect of local noun animacy, such that inanimate local nouns led to 
a slowdown (? = -0.089, SE: 0.027, pMCMC < 0.001). At the spillover region there 
was a main effect of local noun animacy (? = -0.081, SE: 0.023, pMCMC < 0.001), as 
well as interaction of local and distant noun phrase animacy (? = 0.165, SE: 0.046, 
pMCMC < 0.001). In addition, there was a marginal effect of the embedded noun?s 
animacy (? = -0.062, SE: 0.033, pMCMC < 0.07). Planned comparisons revealed that 
the animate embedded nouns caused a significant speed-up when the head noun was 
inanimate (? = -0.103, SE: 0.032, pMCMC < 0.01), but there was a marginal 
slowdown for embedded animates when the head noun was animate  (? = 0.062, SE: 
0.033,  pMCMC < 0.07).  
 
 
 
 
 
 
288 
 
 ? SE t / z 
ta-ziji    
LOCAL -0.089 0.027 -3.32*** 
DISTANT -0.030 0.023 -1.29 
LOCAL?DISTANT 0.076 0.047 1.63 
    
Spillover    
LOCAL -0.081 0.023 -3.51*** 
DISTANT -0.020 0.023 -1.90? 
LOCAL?DISTANT 0.165 0.046 3.58*** 
 
Table 5.6: Summary of fixed effects for best-fit models at the critical ziji and 
spillover regions, including t-values  
 
 
Figure 5.4: Region-by-region mean log reading times for Experiment 9. Error bars 
represent ?1 standard error, by participants, corrected for between-participant 
variance.  
6.
 2
 6.
 4
 6.
 6
 6.
 8
 7.
 0
 7.
 2
 7.
 4
 Mean log reading time, by region
 Region
 lo
 g(
 R
 T
 )
 RC NP PP ba ta-ziji Adv V
    
No match
 Distant match
 Multiple match
 Local match
 
 
289 
 
Discussion 
 
 In Experiment 8, increased reading times were observed for both distant and 
no match conditions. At the point of the anaphor, there was a significant main effect 
of local subject animacy, and in the spillover region, there were main effects of both 
local and distant subject animacy, with animate distant subjects causing longer 
reading times in the spillover region. In Experiment 9, there was also an impact of 
local subject animacy, followed by an interaction of local and distant subject 
animacy. As in Experiment 8, inanimate local subjects caused longer reading times, 
but in contrast to the pattern of reading times observed in Experiment 8, there was a 
significant facilitation for distant animate subjects when the local subject was 
inanimate. 
These results suggest that ziji and ta-ziji contrast with respect to the difficulty 
of recovering a sub-commanding antecedent. This pattern of results suggests that the 
two anaphors recruit qualitatively different strategies for accessing their antecedents 
online, despite superficially similar locality requirements. Results suggest that ta-ziji 
is able to use semantic or discourse prominence to directly access its antecedent, 
regardless of structural position. This pattern was not observed for ziji, which again 
appeared to check the semantically inappropriate local subject position before 
considering a wider range of antecedents. 
 
 
290 
 
Figure 5.5: Effect of embedded animate (embedded [+animate] subject ? embedded 
[-animate] subject) on spillover reading times in Experiments 8 (ziji) and 9 (ta-ziji). 
Error bars represent 95% confidence interval, by participants.  
 
This difference in access strategy is apparent in the summary of the reading 
time differences between the critical comparisons for ziji and ta-ziji in Figure 5.5. 
This summary presents the effect of an embedded animate subject, which is obtained 
by subtracting the reading time for (5.10b) from (5.10a). It can be seen that the effect 
of a distant animate subject for ziji dependencies is similar regardless of whether or 
not the head noun is animate. This observation is confirmed by statistical analysis: at 
the anaphor, there was only a main effect of local subject animacy, followed by main 
effects for local and distant subject animacy in the spillover region. In the spillover 
-3
 00
 -2
 00
 -1
 00
 0
 10
 0
 20
 0
 30
 0
 Effect of embedded animate
 m
 s
 Ziji: Exp 8 Ta-ziji: Exp 9
 Local subject +animate
 Local subject -animate
 
 
291 
region, an embedded animate subject caused a slowdown in reading times that was 
indistinguishable between animate and inanimate head nouns.  However, for ta-ziji, 
there was a significant interaction of local subject and distant subject animacy, such 
that animate embedded subjects with inanimate local subjects caused facilitated 
processing.   
 
(5.10)  a. [Zhang taitaii  guanggu de] shizhuangdianj hai-le  ta-
 zijii/*i/zijii/*i 
  [Mrs. Zhang    visit     DE] boutique           harm-PERF ta-ziji / 
ziji. 
?The boutique that Mrs. Zhang visits harmed her.? 
 
  b. [Meiti   baodao  de] shizhuangdianj hai-le  ta-
 zijii/*i/zijii/*i 
  [Media   report   DE] boutique           harm-PERF ta-ziji / 
ziji. 
?The boutique that the media reported on harmed her.? 
 
The results of Experiment 8 are consistent with the structured search account 
of ziji given in Chapter 4: comprehenders are initially sensitive to the feature 
mismatch between the local subject and the anaphor, suggesting that the local subject 
is selectively reaccessed based on its structural position. In the spillover region, the 
embedded animate exerts an approximately equal slowdown in reading times 
regardless of the local subject?s animacy. This suggests that whatever processes drive 
the slowdown observed for embedded animates, they are independent of the local 
noun?s animacy. Importantly, sub-commanding antecedents did not lead to faster 
reading times relative to a no antecedent condition. This provides further evidence 
against a semantic feature-based access account for ziji-antecedent dependencies, and 
stands in contrast to the results for ta-ziji. If ziji made use of the animacy information 
 
 
292 
in retrieving its antecedent, then an embedded antecedent should have caused ziji to 
be read more quickly than the no antecedent baseline.  
Interestingly, these present results provide another example where online 
processing difficulty diverges from offline acceptability findings: ziji with embedded 
antecedents was considered more acceptable than the no antecedent condition in the 
judgment task, but appeared to require more processing in the online reading task. 
This is surprising in itself, as processing difficulty is generally highly correlated with 
acceptability.  
One possible objection to these findings is that the different processing 
profiles observed for ziji and ta-ziji may not be due to differences in access 
procedures, but rather due to differences in processing difficulty in the no match 
baseline condition. In Chapter 4 I suggested that comprehenders might be able to use 
a discourse-based heuristic to avoid the need for structured search when an indexical 
interpretation of ziji is needed. The possibility of this alternative route to resolving 
ziji?s reference may meant that the no match condition is more difficult for ta-ziji than 
it is for ziji. No such strategy is possible for ta-ziji, as the 3rd-person pronoun ta 
blocks the possibility of a first-person construal of the anaphor. The SPR data 
presented here do not rule out this possibility, but this interpretation seems unlikely in 
light of the relationship between the local animate subject conditions and the distant 
subject conditions. Participants clearly spend more time reading the distant match 
condition than the local match conditions for ziji, whereas for ta-ziji, the distant 
match condition patterns with the local match conditions. Since distant match ta-ziji 
is not associated with an appreciable increase in reading times or decrease in 
 
 
293 
acceptability over grammatical baselines, there does not appear to be any extra 
processing difficulty associated with recovering a sub-commanding antecedent for ta-
 ziji. This is consistent with a direct access mechanism for recovering the distant 
antecedent for ta-ziji as in Figure 5.2. Conversely, the processing difficulty and 
degraded acceptability for distant antecedents for ziji is consistent instead with a 
structured access account as in Figure 5.1   
Another alternative explanation for the difference between ziji and ta-ziji is 
that the information contained in the written form of ta provided an extra cue to 
antecedent identity. As mentioned, ta distinguishes male, female and inanimate 
antecedents when written, although its spoken form does not. This cannot be ruled out 
on from the present data, but it seems unlikely that written cues to gender would be 
used while linguistic constraints (on animacy) are not for ziji. Even if true, however, 
the basic argument I present here holds: structured access is not a property of 
interpreted dependencies. Gender cues for ta-ziji serve as pointers to its antecedent, 
but they do not serve the same role for English himself.  
It seems plausible that the difference in access strategies for ziji and ta-ziji 
stems from deeper underlying differences between the two. This supports the claim 
that the two anaphors are only superficially similar. In order to draw inferences about 
grammatical constraints from the online processing profiles, however, clear linking 
assumptions between grammatical constraints and access strategies need to be made. 
Thus before taking up further discussion of the two types of anaphor here, it is 
necessary to first revisit the question of structured access that I started this chapter 
with. 
 
 
294 
 
Structured access as syntactic parsing  
 
Since ta-ziji appears to use semantic information as a pointer to the memory 
representation of its antecedent, while ziji does not, we are led to reject the first 
intuitive hypothesis that structured access is a property of interpreted or referential 
dependencies. With the new results concerning the access properties of these different 
anaphors in hand, the empirical scope of structured access can now be slightly 
expanded, as in Table 5.7 (from Table 5.1 above). In addition to the dependencies that 
I have considered in this thesis, there are a number of other results that may bear on 
the development of a theory of structured access. These include English VPE (Martin 
& McElree 2008, 2009) and English cross-sentential anaphora (Foraker & McElree 
2007). Though SAT results have shown that a range of dependencies proceed in a 
feature-based direct access fashion, these two dependencies are especially interesting 
in the current context because they are best described as retrospective dependencies; 
the head of the dependency (the antecedent NP / VP) does not signal that it will be 
retrieved later on. Thus no prospective processing is likely to be engaged, and the 
processing of these dependencies may be taken to reflect retrieval processes. 
Interestingly, both cross-sentential anaphora and VPE in English appear to retrieve 
their antecedents in a feature-based direct access fashion. 
However, in addition to reflexives, there is processing evidence that suggests 
another dependency that may engage a structured access mechanism: relative clause 
(RC) attachment in Dutch (van Berkum, Brown & Hagoort 1999b; Brysbaert & 
 
 
295 
Mitchell 2000). It appears that when retrieving a noun phrase to attach a relative 
pronoun to, extra-syntactic information is not used to guide this parsing decision. 
Mitchell et al (1995) argued that this is because abstract information such as major 
syntactic category is the correct ?grain size? over which decisions about RC 
attachment height should be stated. If the RC attachment decision is modeled a 
retrieval for an NP attachment site that is engaged by the relative pronoun, then the 
observation that grammatical gender is not used in this retrieval process is similar to 
the claim of structured access for reflexives advanced here.  
 
Table 5.7: Summary of access properties of long-distance dependencies.   
 
Dutch RC attachment processing provides an interesting parallel to the present 
work. This is because the relative pronoun inflects for gender, imposing a clear 
formal constraint on its attachment site above and beyond the structural conditions on 
attachment; the head noun and the relative pronoun must agree in gender features. 
Nonetheless, it appears that gender information is not used in finding an attachment 
site for the relative pronoun, just as number was not used in determining a reflexive?s 
 Facilitatory 
Interference 
English subject-verb agreement   ?  
Spanish subject-verb agreement ? 
English reflexives ? 
Mandarin Chinese ziji ? 
German, English NPIs ? 
English VP ellipsis ? 
English cross-sentential anaphora  ? 
Mandarin Chinese ta-ziji ? 
Dutch relative clause attachment  ? 
 
 
296 
antecedent in Experiments 1 and 3. For example, Brysbaert and Michell (1996) 
showed that comprehenders do not apply gender constraints to resolve attachment 
ambiguities in situations of where the relative pronoun can be attached to one of two 
NP hosts. As in English, the Dutch sentence in (5.11a) is globally ambiguous: the 
relative pronoun die can be construed as modifying the higher or lower phrase. 
Brysbaert and Mitchell noted that speakers preferred to construe the RC as modifying 
the higher NP (de zoon).  
 
(5.11)  a. De zoon van de actrice  [die         op het balkon zat?] 
The son-COM of  the actress-COM [that-COM on the balcony sat] 
?The son of the actress who was on the balcony?? 
 
  a. Het zoontje van de actrice  [die         op het balkon zat?] 
The son-NEU of  the actress-COM [that-COM on the balcony sat] 
?The son of the actress who was on the balcony?? 
  
 Interestingly, Brysbaert and Mitchell found that the preference for attaching 
the relative pronoun to the higher noun was the same for (5.11b), even though the 
head noun zoontje mismatches in grammatical gender. Thus (5.11b) was significantly 
more difficult to process than (5.11a), but not until later disambiguating information 
was received (Brysbaert & Mitchell 1996, 2000). Comprehenders did not appear to 
use the gender information early enough in parsing to eliminate the ambiguity of 
(5.11b).  
 Van Berkum and colleagues (1999) presented an additional argument for a 
purely structural parsing strategy for attaching relative pronouns to their NP hosts. 
Using event-related potentials, they showed that a mismatch in gender features 
between the relative pronoun and its host NP did not aid comprehenders in 
 
 
297 
disambiguating a relative / complement clause ambiguity. Dat is both a relative 
pronoun for neuter gender nouns, and a complementizer that introduces an embedded 
clause. For a string such as David vertelde de actrice dat (David told the actress-COM 
that-NEU / that), the gender mismatch between the common gender actrice and the 
relative pronoun interpretation of dat should preclude pursuing a relative clause 
interpretation. However, using event-related potentials, van Berkum and colleagues 
argued that participants nonetheless attempted to parse dat as a relative pronoun in 
these cases (given a supportive discourse context), encountering difficulty later in the 
sentence when further input disambiguated the sentence.  
 This finding is in line with the structured access account advocated here. The 
parsing decision and attachment of the relative clause pronoun proceeds blind to 
gender information, which is only evaluated once the dependency has been 
constructed. This surprising finding parallels the results in Chapter 2: despite its 
usefulness, the morphological feature information in reflexive and relative clause 
pronouns is not deployed for early parsing decisions. Brysbaert and Mitchell (2000) 
describe this surprising finding in a way that parallels the results in Chapters 2 and 3: 
 
[F]or some as yet unexplained reason, grammatical gender information does 
not appear to play as rapid and efficient a role in guiding syntactic processing 
as might have been expected from the formal constraints such cues place on 
the structures of sentences.   (Brysbaert & Mitchell pp 465) 
 
In light of the findings presented here, this conclusion may be sharpened. It is 
not simply gender features that the parser selectively ignores, but number features in 
English reflexives as well. More generally, however, a consideration of the 
dependencies in Table 5.7 suggests that this general statement does not hold across all 
 
 
298 
of the syntactic dependencies that have been studied. In particular, for dependencies 
such as subject-verb agreement, these features are not ignored. Instead, they are 
directly and immediately deployed in the construction of subject-verb agreement 
dependencies. The narrow focus on syntactic information is thus not a feature of all 
syntactic dependencies, nor is it a feature of interpreted dependencies.  
 Given the evidence that has accumulated to date, a strong but simple 
hypothesis about the relation between syntactic structure and memory access online 
may be entertained. Setting aside subject-verb agreement for the moment, it may be 
that structured access is simply a feature of all processing of syntactic dependencies. 
This claim makes a direct link between syntactic constraints and the information used 
to access memory online.  
A direct link between the information needed to satisfy structural constraints 
for syntactic dependencies and the information used to access memory and generate 
structure online is arguably the null hypothesis. For long-distance dependencies that 
have are primarily structural in nature, the parser appears to use only structural 
constraints to access and index memory. Reflexive dependencies are unambiguously 
structure-dependent on a wide range of approaches. Most grammatical theories of 
anaphoric interpertation treat reflexive anaphor-antecedent dependencies as ?core? 
phenomena whose interpretation is immediately constrained by their syntactic 
context, despite a wide variety of views about the role that syntax plays in 
constraining non-reflexive pronominal interpretation (Wasow 1972; Chomsky 1981; 
Reinhart & Reuland 1993; Fiengo & May 1994; Pollard & Sag 1994; B?ring 2005). 
 
 
299 
Likewise, relative clause attachment (i.e. the relation between the relative pronoun 
and its NP host) is similarly constrained by structural factors. 
However, if a long-distance dependency does not involve building direct 
syntactic relations between two elements, there is no reason to expect that it should 
use structural access based on (5.12) above. The dependency between a pronoun and 
its antecedent, for example, is not generally modeled as a syntactic relation between 
the two; instead, the two elements are simply taken to denote the same entities in the 
discourse model (B?ring 2005). Although there are apparent structural constraints on 
where a pronoun?s antecedent could be found, it is not clear that these are constraints 
on where the antecedent cannot be located syntactically, rather than constraints on 
possible coreference (Lasnik 1976). In particular, a principle B violation cannot be 
alleviated by placing the desired antecedent in a licit structural position: compare 
*Johni hit himi to I saw Johni. *Johni hit himi. This suggests that describing 
syntactically accessible antecedent positions is not the correct way to capture the 
distribution of coreferential pronouns.  There are potentially even fewer constraints 
on where the antecedent for VP ellipsis may be found (Johnson 2001), with the only 
plausible constraints being that there be some VP in the discourse that can fill the 
ellipsis site. As with constraints on coreference, apparent structural constraints on 
where VPE?s antecedent may be found may be reduced to other factors. It seems 
unlikely that either VPE or coreference relations are dependent on a direct syntactic 
dependency built between the proform and its antecedent. In light of this, it is perhaps 
not surprising that both of these comprehenders engage feature-based direct access 
mechanisms when resolving the reference of these dependencies (Foraker & McElree 
 
 
300 
2007; Martin & McElree 2008, 2009). Structured access would simply be a poor 
strategy for these dependencies, inappropriately narrowing search to a subset of 
positions in a potentially very large discourse space.  Likewise, if ta-ziji is more 
appropriately construed as a contrastive or intensified pronoun whose reference is not 
syntactically represented, then the constraint imposed by structured access does not 
hold.  
To a first approximation, this generalization covers existing empirical 
findings: syntactic dependencies use only syntactic information to access memory, 
and no such constraint is applied to non-syntactic anaphoric dependencies. However, 
one dependency that does not obviously fit this characterization is subject-verb 
agreement in English. One clear finding is that morphological features are used as 
retrieval cues in the construction of the dependency. This access profile is not 
obviously consistent with the statement of structured access given above. From the 
point of view of structured access, the behavior of subject-verb agreement in 
comprehension is puzzling. 
 There are two ways that the exceptional nature of subject-verb agreement 
might be understood. One possibility is that the characterization of subject-verb 
agreement as a syntactic dependency is incorrect. If this line of reasoning is correct, 
then the direct link between structured access and syntactic dependencies is 
maintained, and understanding the exceptional behavior of agreement requires further 
elaborating the correct grammatical model of agreement. An alternative possibility is 
that the direct link between structured access and syntactic dependencies is not the 
correct generalization. This would suggest that there is an extra condition on the 
 
 
301 
deployment of structured access in online processes that that subject-verb agreement 
does not meet. 
 
Agreement as uninterpreted syntax 
 
The nature of the syntactic representation of agreement has been the subject of 
an intense amount of linguistic research in recent years (Chomsky 2000; Wechsler & 
Zlatic 2000; Bhatt 2005; Corbett 2006; Badecker 2007; Bobaljik 2008; Baker 2008, 
among many others). Depending on which model of the subject-verb agreement 
dependency one assumes, agreement may not present an exception to the structured 
access hypothesis after all. Thus whether or not agreement is a true exception to the 
generalization in (5.12) turns on assumptions about its syntactic representation. 
Government & Binding models of agreement modeled the dependency as 
fundamentally similar to an anaphoric relation (e.g. Chomsky 1981), or a chain 
formed in the syntax between the agreeing element and the functional head hosting 
the agreement morphology (Rizzi 1990). On this view, it is difficult to explain the 
differences between reflexives and agreement, as they are more similar than not in 
terms of grammatical representation. However, other models of agreement model the 
dependency in quite a different manner. A range of current theories of agreement, 
together with the hypothesis of structured access, may be compatible with the 
apparent exceptional behavior of agreement.  
 For instance, agreement features may in fact by more central to the 
construction of syntactic structure than suggested by early GB models. In Minimalist 
 
 
302 
models, the agreement operation has been elevated to the status of a crucial structure-
 building operation (Chomsky 2000; Bhatt 2005; Baker 2008). This model suggests 
that in Table 5.7, agreement is exceptional in that it is the one dependency where phi-
 features are explicitly linked in the syntax between the head (the subject) and tail (T) 
of the dependency. If the dependency between the subject and the inflectional 
morphology is directly licensed by the feature match, as suggested by this model, 
then this has the effect of pushing the phi-features involved in agreement into an 
essential, structural role. On this view subject-verb agreement is exceptional in that it 
the phi-features are syntactically represented in the construction of the dependency. 
Since they are directly involved in licensing the syntax, the phi-features may be used 
to index information in subject-verb dependencies in line with (5.12).  
This stands in contrast to other syntactic dependencies in Table 5.7, where the 
role of the phi-feature in structure building is somewhat more indirect. For instance, 
no model of reflexives directly involves a syntactically represented feature match 
between the reflexive and its antecedent in the same fashion. The closest model is that 
of Reuland (2001), who proposes that the anaphor relies on the same AGREE 
operation to resolve its antecedent. However, even on this AGREE-based model of 
reflexive dependencies, the agreement relation is not directly between the antecedent 
and the reflexive; rather, the two elements are indirectly related through verbal 
morphology. Thus the phi-features in subject-verb agreement may arguably be more 
narrowly syntactic than in other dependencies, licensing the use of these 
morphological features in first-pass dependency construction in a structured access 
model. 
 
 
303 
 Alternatively, one might relegate agreement to a post-syntactic operation, as 
proposed by Bobaljik (2008). Bobaljik suggests that rather than being a fundamental 
structure-building operation, verbal agreement is actually an entirely post-syntactic 
phenomenon. If true, this would exempt agreement processes from the structured 
access hypothesis, as they would no longer represent strictly syntactic processes. 
Bobaljik?s argument is constructed on cross-linguistic observations about the 
distribution of agreement processes in language. In particular, when syntactic position 
and morphological case diverge, agreement tracks morphological case rather than 
syntactic position, almost without exception. Bobaljik casts this as an ?order of 
operations? argument: if agreement makes reference to information that is picked up 
post-syntactically (phonological expression of case), while ignoring information that 
is presumably present in the syntax (syntactic position), then this may be taken as 
evidence for the conclusion that agreement morphology is licensed post-syntactically. 
For example, he notes that agreement in Hindi is blocked whenever an NP is marked 
with an overt case marker. However, on the assumption that the case marker is added 
in a post-syntactic morphological component of the grammar, then Hindi agreement 
must be calculated post-syntactically. Compared to the standard Minimalist model 
(Chomsky 1998), this claim posits a rather different role for agreement in the 
organization of the grammar. For the hypothesis of structured access, however, either 
model has the same impact of the relationship of subject-verb agreement to the 
statement of structured access in. They both make sense of its exceptional use of phi-
 features in accessing memory, either by making phi-features directly syntactically 
 
 
304 
active (Chomsky 2000) or relegating them to an extra-syntactic computation 
(Bobaljik 2008).  
Deciding between narrowly syntactic and post-syntactic models of agreement 
is of interest for theories of online processing as well as grammatical models. 
However, it is difficult to understand the exceptional online behavior of subject-verb 
agreement only by reference to its role in a grammatical model; to do so would 
require additional linking assumptions about the relationship between the grammar 
and its online implementation, and this level of separation makes it difficult to derive 
predictions about online processing directly from the grammatical model.  
Alternatively, it may be the case that the direct link between syntactic 
dependencies and structured access suggested above is not the correct generalization 
about the range of dependencies that employ structured access.  Instead, it may be 
necessary to critically reconsider this generalization, which could prove insightful in 
constructing a theory of the use of structure in sentence processing. 
 
Structured access as an optimal access strategy  
 
 Although the predictions of the structured access hypothesis with respect to 
the processing of subject-verb agreement do depend on the exact model of subject-
 verb agreement that is assumed, reconsidering the generalization about structured 
access may generate greater understanding the exceptional status of subject-verb 
agreement in English. The initial generalization about the deployment structured 
access appears to reflect a hard constraint that provides a direct link between 
 
 
305 
grammatical constraints and online syntactic generation. However, there is no direct 
evidence for this immutability of structured access as a parsing strategy.  
 One way of understanding the origin of structured access is to note that it is 
often a good parsing strategy. This follows from the reasonable assumptions that a) 
the parsing architecture is prone to memory interference, and b) comprehenders are 
functionally motivated to choose retrieval cues that maximize parsing speed and 
accuracy. In choosing a set of optimal retrieval cues for any given dependency, there 
is functional pressure to minimize use of cues that are potentially more misleading 
than they are helpful. This could be understood as a rational strategy, wherein 
comprehenders deploy the retrieval cue set that jointly minimizes retrieval error and 
processing time. If this hypothesis is correct, than the structured access generalization 
may be as in (5.12):  
 
(5.12) Structured access: in constructing a long-distance dependency, the 
parser employs all and only the information needed to jointly minimize 
interpretive error and processing time. Structured access arises for 
interpreted, syntactically constrained dependencies.  
 
  On this view, structured access mechanisms for memory retrieval might be 
understood as reflecting good parser design. The two crucial elements that seem to be 
implicated in cases of demonstrable structured access are i) interpreted content and ii) 
structural constraints on the position of the retrieval target. Neither appears to be a 
sufficient condition on its own. (5.12) suggests this may be ?rational? parsing, but this 
notion needs to be understood with caution. What counts as ?optimal? or ?rational? 
parsing necessarily depend on the exact formulation of the objective function to be 
 
 
306 
optimized, and absent a formalization of such a function for the task of memory 
retrieval in parsing, it?s difficult to mount such an argument. Nevertheless, it seems 
reasonable to assume that the parser is seeking to maximize both speed and 
interpretive accuracy, and for interpreted, structurally constrained dependencies, 
speed and accuracy trade off against each other. Structured access can be slow, as 
seen in Chapter 4, but it is possible that the average slowdown due to structured 
access is offset by the gain in interpretive accuracy. This is not true for structurally 
unconstrained dependencies (where structured access simply slows down processing), 
or for uninterpreted dependencies (which by hypothesis have no impact on measures 
of interpretive accuracy). If this is true, then structured access may be understood as a 
good strategy for the parser to pursue for certain dependencies.  
Note that cases where the parser appears to disregard helpful information 
contained in formal gender or number cues have been used to motivate syntax-first or 
modular models of comprehension to varying degrees (see, e.g., Frazier & Clifton 
1996; De Vincenzi 1999; van Berkum et al 2000; Brysbaert & Mitchell 2000), The 
informational encapsulation suggested by these cases appears to run counter to the 
architectural commitments of highly interactive, constraint-based parsing models 
(MacDonald et al 1994). However, it has been argued previously that these data do 
not necessarily cut in favor of innately modular architectures. In their consideration of 
the ?grain size? problem for experience-based models of parsing, Mitchell and 
colleagues (1995) argued that it is a mistake to assume that the ?right? thing for a 
parser to do is to use all information it has available to it. The generalization in (5.12) 
suggests a very similar conclusion: the best strategy is not to use all possible sources 
 
 
307 
of information for all parsing decisions. The best strategy is to use only those sources 
of information that are necessary to accomplish the task at hand; in the case of 
parsing, this means recovering the intended interpretation of the sentence quickly and 
reliably.  
Mitchell, Cuetos, Corley and Brysbaert (1995) argued that structural 
abstraction in parsing routines might reflect the best strategy given a language user?s 
experience. In particular they reasoned that that for certain parsing decisions, it is 
rational for the parser to track statistics at an abstract structural level, rather than at a 
fine-grained lexical level. This is because tracking statistics at a lexical level and 
using that to guide parsing decisions (as suggested in McDonald et al 1994) risks 
leaving the comprehender with a sparse data problem. For example, there just are not 
enough instances of all lexical items in an ambiguous NP-P-NP-RC structure to get a 
solid estimate of the lexical bias for nouns in these structures. This lack of data would 
suggest that the best strategy for the parser is to back off to more robust estimates 
about the most probable attachment height over more abstract categories, for which a 
more robust set of data is available. 
Though cast in a different framework, the generalization presented in (5.12) 
may be understood as supporting a very similar argument. Retrieval-and-interference 
based models of comprehension offer another means of understanding the biases at 
work in shaping the parser?s decision metrics. For long-distance dependencies that 
rely on memory retrieval, the richer the information contained in the cue set, the more 
likely it is that some element in memory will interfere with the retrieval process. It is 
a bad strategy to use superfluous or highly correlated cues in retrieval, as they 
 
 
308 
actually have the potential to negatively impact parsing through interference and 
retrieval of non-target memories. In such a model it is advantageous for the parser to 
adopt a small number of highly distinctive cues. For a reflexive dependency, lexical 
cues such as gender, or semantic cues such as animacy, are very highly correlated 
with structural cues. By deploying these superfluous cues in retrieval, the parser 
unnecessarily risks memory interference from non-target memory elements that carry 
those gender or animacy features. These superfluous increase the risk of interference, 
with no apparent benefit in memory retrieval. If the parser is optimized to use the 
minimal, distinct set of cues for retrieval, then it will display structured access 
behavior for reflexive dependencies, and a bias for structural over lexical cues in 
parsing decisions.  
 While this approach makes sense of the general use of structured access in 
parsing, it does not yet address the puzzling behavior of subject-verb agreement in 
English. As suggested above, the key features for a dependency appear to be 
interpretive content and structural constraints on retrieval positions. If this 
generalization is correct, then the question becomes why agreement?s lack of 
interpretive content makes structured access no longer a good strategy. As suggested 
above, since agreement in English is plausibly inert with respect to any measure of 
interpretive accuracy, speed and accuracy don?t trade off against each other in the 
same way that they do for reflexive dependencies. If comprehenders are trying to 
maximize interpretive accuracy, subject-verb agreement in English may just not be 
processed in a particularly structurally sensitive way. 
 
 
309 
A secondary question is how comprehenders come to distinguish the access 
strategies employed for interpreted structural dependencies (such as reflexives) and 
uninterpreted or nonstructural dependencies (subject-verb agreement or cross-
 sentential anaphora). It may be that these are universal principles of parsing that could 
be made to follow from the role of the dependencies in the grammar. Another 
possibility is that comprehenders develop the distinction through a process of cue 
optimization. English agreement carries a low functional load, marking only a subset 
of person / number distinctions in a subset of verbal paradigms. The somewhat 
restricted nature of English agreement might make agreement interference benign 
from the point of view of constructing an interpretation from a syntactic parse. The 
generalization in (5.12) suggests that this is the key reason that agreement in English 
does not employ structured access. If a cue optimization process is at work, then in 
the absence of clear interpretive effects, agreement in English might not generate the 
same sort of error signal necessary to cause the parser to adopt a structured access 
strategy. In general, structured access is slower than feature-based direct access. 
There is no pressure to minimize interference as there is no risk of generating an 
incorrect interpretation, and so the English parser may be optimized to compute 
agreement as quickly as possible even if the occasional retrieval error arises.  This 
conjecture is consistent with empirical findings about the interpretive reflex of 
agreement attraction in English. It has been argued that incorrect agreement 
computation does not adversely impact comprehenders? interpretations (Lau, Wagers, 
Stroud & Phillips 2008).  
 
 
310 
If the exceptional behavior of English agreement is due to its low functional 
or interpretive load, then cross-linguistic investigation should reveal that a greater 
functional load for agreement causes a narrower, more accurate set of retrieval cues to 
be used in constructing agreement dependencies. Structured access should be evident 
in agreement for languages where there is functional pressure to minimize agreement 
interference. Preliminary investigation suggests that this prediction is borne out 
crosslinguistically. Lorimor, Bock, Zalkind, Sheyman & Beard (2008) investigated 
attraction errors in Russian, a free word order, pro-drop language. They argued that 
the rich morphology in Russian aids comprehenders in filtering out agreement errors, 
a point that was also argued by Badecker & Kuminiak for Slovak (2007). A cross-
 linguistic comparison of gender attraction effects reveals that attraction is inversely 
correlated with functional load, as shown in Figure 5.6. Languages that may rely on 
verbal agreement as a cue to interpretation, due to scrambling or argument dropping 
properties, show a diminished rate of attraction; languages with all but moribund 
gender systems, such as Dutch, are at the complete opposite end of the spectrum, and 
they show a high rate of gender agreement interference.  
 
 
311 
 
Figure 5.6: Rate of occurrence of gender attraction errors across languages. Figure 
from Lorimor, Bock, Zalkind & Sheyman (2008). 
 
If comprehenders optimize their retrieval cue sets to increase accuracy in 
response to disruptive interference, these patterns of cross-linguistic variation are 
expected. On this view agreement interference should be all but absent in languages 
like Hindi, which has an agreement system that alternately agrees with the subject or 
the object. The variation in the target position, the system?s dependence on overt case 
cues, and the scrambling, pro-drop nature of Hindi makes a narrow, entirely 
structured access procedure for constructing agreement dependencies a smart choice 
for accurately constructing the agreement dependency. Native speaker intuition 
suggests that no agreement attraction obtains for Hindi, across a range of structures 
and feature configurations (Rajesh Bhatt, Shravan Vasishth, p.c.). In contrast, 
agreement attraction for English speakers is intuitively evident. If further 
investigation upholds this intuition about attraction in Hindi, then a optimized-cue 
account of structured access is supported.  
 
 
312 
The crucial generalization that appears to emerge from the body of evidence 
thus far is that structured access appears to be engaged for dependencies that are 
structurally constrained and interpreted. The accounts for the apparent exceptional 
nature of agreement processing: although it is structurally constrained, it is not 
interpreted and so does not engage structured access. I suggested that this 
generalization might be understood as reflecting an optimal parsing strategy: because 
interpretive accuracy matters for online parsing, there may be pressure to minimize 
interference through structured search. However, at this point the source of this 
strategy remains unclear: it may be a general principle relating different sorts of 
grammatical rules to online processing routines, or it may reflect an experience-based 
cue optimization process. Further research on the relationship between interpretive 
load of agreement and its interference profile is necessary before drawing any firm 
conclusions.  
  
Structured access and Mandarin anaphors 
 
 The hypothesis of structured access as an adaptive strategy endorsed here 
weakens the direct link between grammatical processes and online parsing procedures 
suggested in the original statement of structured access. Morphological or semantic 
constraints can either be satisfied in first-pass structure generation (e.g. English 
subject-verb agreement) or after the construction of a dependency (e.g. English 
reflexives). If features are classified in one way or another based on an extra-
 grammatical decision principle, then it is not clear that the processing evidence 
 
 
313 
presented here necessarily entails that ziji and ta-ziji are qualitatively different 
anaphors, one (broadly speaking) semantic or pragmatic, and the other syntactic.  
The contrast between ziji and ta-ziji observed in Experiments 8 and 9 instead 
means that whatever the grammatical distinctions between the two there might be, the 
optimal operationalized content of ta-ziji?s structure-building cues should include the 
notion of animacy. If, as Bergeton?s (2007) account suggests, ta-ziji is more likely to 
be understood as a contrastive, discourse-bound anaphor, then using non-structural, 
discourse-based cues in the retrieval cue set is a very reasonable strategy. On this 
view, the relevant constraint is not necessarily stated in terms of syntactic position. If 
a structured access strategy were employed, very low accuracy would result. 
Specifically, the under-sampling of the possible antecedents that would come from 
structurally narrowing the search space would lead to very low recall (though 
precision could in principle be quite high).  
The contrastive pronoun model of ta-ziji was supported by the acceptability 
judgment task, where it was seen that multiple animate noun phrases for ta-ziji 
actually led to a decrement in acceptability relative to conditions with only a single 
antecedent. This may reflect the fact that contrastive pronouns such as ta-ziji are less 
felicitous in an out-of-the-blue discourse context with two possible antecedent 
antecedents. A similar contrast obtains in emphatic reflexive usage in English: 
 
(5.14)  a. John told Mary that she herself had failed the exam. 
  b. #John told Bill that he himself had failed the exam. 
 
 
 
314 
However, even accounts that maintain that ta-ziji is entirely bound in the 
syntax may be compatible with the feature-based direct access view. For example, 
Pan (1998) explicitly argues that contrastive ta-ziji is a distinct element from bound 
ta-ziji. He notes that bound ta-ziji can take long-distance antecedents just in case there 
are no local, animate antecedents, and that the relevant constraints are best understood 
as tree-geometric in nature. He proposes that the unique feature of ta-ziji is that its 
binding domain is computed relative to the phrase-structurally closest, most 
prominent (on Pan?s account, this means animate and human) antecedent. On this 
view, it seems also very likely that the optimal access procedure would include a 
notion of prominence in the memory retrieval cues. If this is the relevant statement of 
the constraints, then it seems that good parsing accuracy could be achieved by 
retrieving on a mixture of structural and prominence-based cues. On this view, though 
ta-ziji is a syntactically licensed anaphor, the range of structures in which it is 
licensed is determined by the semantic feature content of its antecedents. This is 
unlike reflexives or agreement, where the feature content and the structural 
restrictions are orthogonal. For this reason, in this model of ta-ziji, the structure and 
feature content are not in any sense redundant, and so leveraging both sources of 
information to compute its reference is an effective strategy.      
 
The footprint of structured access 
 
In the absence of a formal model of the cue optimization process, it is difficult 
to rigorously evaluate the claim of structured access as an adaptive strategy based on 
 
 
315 
the data given above. For now, the level of explanation for various patterns of access 
for phenomena such as English subject-verb agreement, Dutch RC attachment, and 
Mandarin ta-ziji is somewhat superficial. I have offered plausible explanations for 
why they pattern the way they do, but these depend on assumptions about the 
distribution of these dependencies, and the utility of various access strategies (i.e. the 
specific objective function to be optimized). A more thorough account for any given 
dependency involves investigating the level of accuracy (precision and recall) 
obtained by using any given combination of retrieval cues to complete a long-distance 
dependency. Presumably, any claims of ?optimality? about the best cue set for a given 
dependency should be derived from statistics of use, accuracy of retrieval, and a cost 
function for incorrect retrievals. In the interest of providing clear predictions, 
formally developing this insight is an important goal for future research. 
In (5.12), I offered a generalization that appears to describe the dependencies 
that engage structured access for memory retrieval: structurally constrained, 
interpreted dependencies. This generalization generates predictions about other 
dependencies where one might expect to find evidence of a structured access strategy. 
As suggested above and elsewhere (Wagers 2008; Lau 2009), it is likely that 
predictive strategies may dominate for dependencies that can be reliably anticipated 
ahead of the putative memory access site. This feature makes it difficult to draw clear 
conclusions about the nature of retrieval in these dependencies (Wagers 2008; Martin 
& McElree 2008, 2009). A productive strategy is to search for evidence of structured 
access for structurally constrained, retrospective dependencies. Bound variable 
anaphors of all sorts provide an excellent case in point: they are subject to significant 
 
 
316 
structural constraints, and are not generally predictable ahead of the integration site 
(the variable). Structured access and search is a good strategy in this situation. In 
order to get an accurate parse, and to minimize disruptive interference, it is worth 
more narrowly searching licit syntactic positions.  
There are a number of other dependencies that I have not considered that 
might provide further evidence for structured access in comprehension. One example 
is wh-in-situ dependencies, which could be understood as the retrospective 
counterpart of wh-movement dependencies in languages like English. Wh-in-situ 
occurs in languages like Mandarin Chinese, who leave wh-elements in their base 
position (i.e. Mary saw who?) rather than move them to a scope position. On most 
models the in-situ wh-dependency involves a long-distance relationship between the 
site of the interrogative operator that determines the scope of the wh-element, and the 
position where the wh-element is thematically integrated. One way to model this 
dependency is with covert LF movement (Huang 1982; Cheng 2009), where the wh-
 element is moved covertly to occupy the scope position. Alternatively, it may be 
modeled as unselective binding (Aoun & Li 1993; Tsai 1994), where the wh-element 
is converted to a variable in situ and bound by a wh-operator with the appropriate 
scope. On either account, the position of the scope operator must be retrieved at the 
point of processing the in-situ wh-element. In this sense, both accounts model wh-in-
 situ as a retrospective dependency between the wh-element and the structural position 
where wh-scope is marked. It has been shown that in Chinese (Xiang, Dillon & 
Wagers 2010) and Japanese (Sprouse, Fukuda, Ono & Kluender 2011) that 
processing wh-in-situ is costlier for embedded wh-elements than it is for local ones. 
 
 
317 
Sprouse and colleagues argue that this is due to a backward structured search. 
However, using SAT, Xiang et al (2010) found no evidence of structured search in 
long-distance wh-in-situ dependencies. In their data, the difficulty associated with 
long-distance wh-construals was reflected in decreased accuracy, not decreased speed 
or dynamics parameters. 
However, languages such as Chinese or Japanese, for which wh-in-situ is the 
primary method of constructing wh-dependencies, might not provide the best test case 
for structured access in wh-in-situ due to comprehenders simply optimizing their 
processing to access the matrix [Spec,CP] in all cases where interrogative force is 
required. In light of this, French is interesting in that it permits a mixture of wh-
 fronting and wh-in-situ strategies for question construction (Oiry 2011). Oiry (2011) 
notes that there are conflicting reports on the acceptability of long-distance 
(embedded) wh-in-situ questions such as (5.14), with some authors claiming they are 
not grammatical and others claiming they are (see discussion in Oiry 2011). Native 
speaker informants appear to have conflicting judgments about these sentences in out-
 of-the-blue contexts, though they appear to be accepted more often than not. 
 
(5.14) ?? Tu penses que le policier a fait quoi? 
 You think that the policeman has done what 
 ?What do you think the policeman did?? 
 
One possibility is that the mixture of strategies in French has led to a situation 
where comprehenders have not fully optimized the access procedure for wh-words in 
embedded clauses. If they simply searched for [spec,CP] to construct the appropriate 
scope for the wh-word, then they would have perfect performance in parsing matrix 
 
 
318 
wh-in-situ questions, but in examples like (5.14), there would be disruptive 
interference from the embedded [spec,CP] position. This leads to the possibility that 
the mixed pattern of results that has been reported on sentences like these reflects 
interference-related processing difficulty due to multiple potential landing sites for 
the wh-element. This suggestion remains highly speculative, but the contrast between 
French and Chinese-type wh-in-situ could shed light on the structured nature of wh-
 in-situ processing, as well as the role for cue optimization in this process.  
Another interesting area of investigation is the attachment of relative clauses 
to complex NPs. Cuetos and Mitchell (1988) originally suggested that languages vary 
arbitrarily with respect to their preferences in attaching relative clauses in ambiguous 
NP-PP-NP environments such as the servant of the actress who was on the balcony. 
They showed that while English speakers reliably interpret the RC as modifying the 
actress (in accordance with late closure, Frazier 1978), Spanish speakers preferred to 
attach the RC to the higher of the two NPs in Spanish. This cross-linguistic different 
behavior was later modeled as a sort of parsing non-determinism (Frazier & Clifton 
1996, 1997; Kamide & Mitchell 1997), whereby comprehenders are able to delay full 
commitment to a parse of the relative clause?s attachment pending extra 
disambiguating information. Frazier and Clifton (1996) noted that semantic and 
discourse factors, such as focus accent or referentiality, impacted attachment 
preferences.  
 An alternative possibility is suggested by the structured access hypothesis. It 
may simply be that the retrieval cues in these environments reflect only the fact that 
in a certain context, the RC must be attached to an NP in a certain structural 
 
 
319 
environment (e.g., a thematic domain, Frazier & Clifton 1996). This opens up the 
possibility that non-determinism impacts the attachment process at the level of 
memory retrieval, which is by hypothesis stochastic and indirectly impacted by a 
number of processes that modulate the relative activation of the two NPs in memory. 
On this account, the reason that comprehenders prefer to attach to the focused NP is 
that is the accent conveys an activation advantage, which translates into a greater 
probability of retrieval at the point of constructing the NP-RC dependency.  
Interestingly, there seems to be some evidence that ?general? memory 
dynamics are at work in selecting the site of attachment for RCs. Gibson, Pearlmutter, 
Canseco-Gonzalez & Hickok (1996) showed that in sentences where an RC can be 
attached to one of three NP positions in a noun phrase (the key to the house on the 
hill), the most recent and most distant NPs are preferred over the middle NP position. 
This is compatible with the serial position effects noted in recall, where it is known 
that the first and last elements are more accurately recalled due to effects of primacy 
and recency (Murdock 1962). Gibson and colleagues argue that there are problems 
with applying the simple list recall model to the RC data. This is in part due to 
different preferences for primacy and recency across constructions within Spanish; 
for VP-level attachments, such as temporal adverbials like yesterday, comprehenders 
prefer to attach low, whereas for NP-level RC attachments, they prefer to attach high. 
It is unclear if this is a real challenge, as it is unclear the processes involved in the 
two cases are the same. They furthermore note that it is hard to reconcile this view 
with the between-languages differences. However, if this can be tied to independently 
observable differences in how Spanish and English speakers distribute their attention 
 
 
320 
across a complex noun phrase structure, then a memory dynamics account of this 
ambiguity could actually provide a principled way of determining which languages 
will show high or low attachment preferences. This is a speculative possibility, and 
further work is necessary to determine whether or not RC attachment preferences 
track activation in memory.  
 
Conclusion  
 
 In this chapter, I have attempted to sharpen the hypothesis of structured access 
in comprehension. By comparing similar but distinct anaphors in Mandarin Chinese, I 
showed that structured access is not a property of interpretive dependencies. Instead, 
the correct generalization appears to be that structured access is engaged for 
dependencies that are both interpreted and structurally constrained. I suggested 
instead that this may reflect an adaptive strategy that reflects a pressure to minimize 
interference error for interpreted dependencies. By adopting the minimal set of 
distinctive cues necessary to arrive at a grammatical interpretation of a sentence, 
comprehenders minimize interference and guard against misinterpretation. On this 
account, agreement does not use structured access mechanisms in retrieval because it 
is an uninterpreted dependency, and thus does not impact interpretive accuracy (at 
least in English). 
 It has long been argued that abstract parsing principles might reflect 
functional pressures from the memory architecture of the parser (Frazier 1978; 
Frazier & Fodor 1978; Mitchell et al 1995), and the arguments provided here make a 
 
 
321 
similar point. If comprehension is prone to interference, a good strategy is to 
minimize the information used to access memory in the interest of minimizing 
distraction from irrelevant memory elements. Interference effects have been of 
intense research interest in recent years (Gordon et al 2001; Lewis & Vasishth 2005; 
Vasishth et al 2008), and the arguments provided here suggest that they may in fact 
provide another functional pressure for the parser to construct structure with reference 
only to abstract structural categories and positions. Thus structured access might 
reflect a simple pressure for the parser to get it right as often as possible, and in the 
shortest amount of time possible.   
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
322 
Chapter 6: Conclusion  
 
 
 
 
 
 
 
 
 
 
 
 
 
  
 The main claim of this thesis was that for certain long-distance dependencies, 
comprehenders engage a purely structural mechanism for accessing linguistic 
working memory. This conclusion is compatible with various architectures: in the 
context of a content-addressable memory architecture, for example, structured access 
could be implemented by only utilizing structural information as retrieval cues. More 
generally, structured access is a claim about the type of information that 
comprehenders use to assemble long-distance dependencies online. In this thesis I 
provided computational and experimental experiments in support of this claim, 
demonstrating that comprehenders selectively attend to structural information over 
morphological or semantic information while accessing linguistic memory in the 
course of sentence comprehension.  
 At the outset, one of the major questions of interest was the relationship 
between the grammar and the parser. There have been a number of important 
challenges to the idea that fully grammatically elaborated structure is constructed 
online, and the challenge from constraints in working memory was the focus of this 
work. I have argued that despite some apparent limitations of the memory 
 
 
323 
architecture of sentence comprehension, the parser is nonetheless able to effectively 
target and retrieve particular syntactic positions in order to construct long-distance 
dependencies online. More generally, however, the arguments I have presented here 
form a bounding condition on the grammatical fidelity of the structures constructed 
by the parser: they must be elaborated sufficiently for the parser to access particular 
elements in a grammatically sophisticated manner, even if the parser does not always 
make full use of that structure. If true, this fidelity poses a challenge for models that 
assume a more indirect or heuristic relationship between online parsing processes and 
grammatical knowledge (Townsend & Bever 2001; Ferreira et al 2002; Ferreira & 
Patson 2007). 
 There were two specific arguments that I presented to make the case for a 
narrowly structural access mechanism in linguistic comprehension. I review these 
first, then turn to a brief consideration of the scope of structural access, and 
implications for the memory architecture of language comprehension. 
 
No consideration of illicit antecedents: Experiments 1-5 
 
 The first empirical argument for structured access focused on the processing 
of reflexive anaphors in English. An experimental investigation of the processing 
profile of English reflexives suggests that the initial candidate antecedent set 
considered by the parser is narrower than expected if a feature-based access 
mechanism is used; only the local subject is retrieved, regardless of the feature 
content of other NPs in the sentence. This profile was contrasted with English 
 
 
324 
agreement, and the impact of feature-matched, inaccessible nouns for both 
dependencies was investigated in three eye-tracking experiments (Experiments 1-3). 
In immediate online processing, agreement dependencies retrieved inaccessible noun 
phrases that had the correct morphological features, building illicit agreement 
dependencies and causing illusion of grammaticality effects in behavioral measures. 
In contrast, reflexive anaphors showed a qualitatively different access profile: they 
were only sensitive to a feature match with the local c-commanding noun phrase.  
 This provides the first empirical argument for structured access: for certain 
dependencies, there is no consideration of inaccessible noun phrases during memory 
access. Possible objections to this conclusion were addressed with computational 
modeling in Experiments 4-5, which supported the claim that reflexives build the 
dependency with their antecedent using only structural information. 
  
No immediate access to distant but accessible antecedents: Experiments 6-9 
 
 Experiments 6-9 advanced the second major empirical argument for structured 
access. In cases where there are multiple nouns that could contain an antecedent for 
an anaphor, the initial candidate antecedent set is not narrowed by the feature content 
of those NPs. Thus even in the presence of structurally accessible, feature-matched 
antecedents for the Chinese reflexive ziji, the parser instead initially retrieves local, 
feature-mismatched antecedents. The ziji findings provide a complementary piece of 
evidence for structured access. To account for the fact that the local subject is 
considered first, memory access must make reference to only structural information, 
 
 
325 
rather than incorporating semantic or morphological feature constraints immediately. 
Thus results from the processing of both Chinese and English reflexives provide an 
argument in favor of a uniquely structural access mechanism in processing reflexive 
dependencies.  
  Experiments 8 and 9 directly contrasted the processing of two distinct 
Chinese anaphors: ziji and ta-ziji. There were two main conclusions from this 
comparison. First, it was shown that it is not interpretive content alone that leads to 
structured access; the anaphor ta-ziji immediately accesses structurally distant 
antecedents based on their semantic features. In addition, this comparison provided 
experimental evidence that it is not linear position in the string that causes structured 
access effects, as ziji and ta-ziji occupied the same linear position in the string,   
 
Interpreting interference effects  
 
 In Chapter 3 a consideration of the computational basis of the interference 
model here led to the conclusion that a demonstration of facilitatory interference is 
necessary to conclude that the parser has constructed illicit structure. Due to the 
rational basis of the ACT-R model, it was seen that this is true of any model that links 
probability of a structure or retrieval with behavioral indices such as reaction time. 
One way of understanding this result is that the consideration of grammatically 
unlicensed parses leads to a situation of ?spurious ambiguity?, where, due to 
interference, the parser is jointly considering grammatical and ungrammatical parses 
alike. Facilitation occurs in this situation for the same reason that ambiguity eases 
 
 
326 
processes in truly ambiguous situations: given multiple options and a ?race? to adopt 
the fastest option to be constructed, the observed reaction time distribution represents 
a mixture of the fastest processing times from either option, showing facilitation 
effects on average. 
  Given this finding, I argued that the evidence for online consideration of 
ungrammatical parses is sparser than is generally assumed. Of the range of 
interference results that have been reported, there are only two dependencies that 
reliably show facilitatory interference: subject-verb agreement and negative polarity 
item dependencies. The interpretation of NPI interference as the result of partial-
 matching facilitation has been questioned (Xiang et al 2009), leaving subject-verb 
agreement (primarily in English) as the parade case of ungrammatical structure 
construction online. Most online effects that have been attributed to interference are 
inhibitory, which does not provide online evidence for consideration of illicit 
structures. Instead, much of the experimental evidence that has been taken to indicate 
incorrect structure generation comes from offline measures of interpretation, which is 
a more indirect measure.  
 
Structured access and the architecture of comprehension 
 
I argued at several points that the claim of structured access is compatible 
with a variety of architectures. In Chapter 3 I presented an implemented 
computational model of structured access in a content-addressable memory 
architecture, and in Chapter 5 I argued that structured access might be understood as 
 
 
327 
an optimal adaptation to memory interference in this framework. The need to provide 
a grammatically faithful parse provides a functional pressure for abstract structural 
cues, as they are on average the most reliable and effective cues to memory access or 
structure generation; similar arguments about the functional advantage to abstraction 
in structure generation were presented by Mitchell and colleagues (1995).   
However, an alternative possibility is that structured access reflects the 
deployment of a qualitatively different memory access mechanism. That is, it may 
reflect the deployment of a memory indexing system that indexes memories only 
according to their position, serially traversing all positions to find the target memory 
for any given retrieval operation. This possibility is similar to the observation that 
there are two types of retrieval mechanisms that are deployed for general memory 
tasks, depending on the nature of the information that needs to be recalled (McElree 
& Dosher 1993).  
If one assumes that structural cues can be defined to target arbitrarily precise 
syntactic positions (an assumption that is far from obvious), the predictions of a 
content-addressable architecture that uses narrowly structural search cues and a serial 
architecture that literally traverses a parse tree node-by-node align in almost all cases. 
There are a number of possibilities for deciding between these possibilities, however.  
 
Structured access as an optimal adaptation 
 
 One argument for narrowly structural search cues depended on the assumption 
of a content-addressable architecture. I suggested in Chapter 5 that they could 
 
 
328 
represent an optimal behavioral policy for a parser that a) is prone to interference but 
b) nonetheless needs to recover only grammatically licensed parses. On the 
assumption that the cues used to access memory in a content-addressable architecture 
are subject to tuning or optimization, I suggested that the parser should only adopt 
abstract structural cues for certain dependencies, such as English reflexives and 
Mandarin ziji.  
 In order to make a stronger case for this, this informal reasoning should be 
complemented by formal modeling. This would more firmly demonstrate that the 
narrow use of syntactic cues is the best course of action to take, given the constraints 
on the parser. The question of the optimal behavioral policy to maximize reward is a 
general problem that has been intensely studied in computer science under the name 
reinforcement learning (see, e.g. Sutton & Barto 1998). These techniques can be used 
as an informal proof by simulation that a given behavioral policy is optimal given a 
space of possible actions. If the reasoning in Chapter 5 is correct, then such an 
analysis should show that the optimal policy is to ignore the correlated but potentially 
disruptive information contained in morphological cues. 
 A formal demonstration of this argument using existing computational tools 
would provide a proof of principle that would significantly bolster the informal 
arguments I have provided here. This would provide support to the content-
 addressable architecture implementation of structured access by demonstrating that 
the arguments are in fact sound. More generally, however, it would suggest that 
adopting a content-addressable architecture provides a mean of explaining the origin 
of structured access in the parser.  
 
 
329 
 
Blocking effects 
 
 Another potentially illuminating source of evidence about the memory 
architecture of the parser is the existence of blocking effects. The term ?blocking 
effect? is often used to refer to a particular class of constraints on Mandarin 
reflexives, but the crucial ?blocking? nature of these constraints is apparent in a range 
of other dependencies. To demonstrate the usefulness (and generality) of these effects 
for the present discussion, let us use ?blocking effect? to refer to any instance in which 
a long-distance dependency between two elements is disrupted by an element along 
the path between the two elements. Examples include the blocking of wh-movement 
by non-bridge verbs (Erteschik-Shir 2006) or definite noun phrases (Chomsky 1973), 
definite island effects in NPI licensing (Ladusaw 1979), quantifier intervention 
effects in NPI licensing (Beck 1996), and person blocking effects in Mandarin 
reflexives (Huang & Li 2009). 
 These effects have the potential to be informative with respect to the question 
of memory architecture. In retrieval-driven content-addressable architectures, 
dependencies are modeled are direct relations between two constituents in a parse. 
Interference in this relation is only expected to occur from other constituents that 
share feature content with either the probe or the target memory. For the blocking 
effects mentioned above, however, the dependency is disrupted by an intervening 
element that is not feature-matched to either the probe or the target. For example, 
non-bridge verbs such as lisp do not contribute interfering [+wh] features, so it is 
 
 
330 
unclear on what dimension they would interfere with the construction of the wh-
 dependency. 
 Given the nature of these constraints, an important empirical question is 
whether or not comprehenders are immediately sensitive to blocking elements in 
constructing a long-distance dependency. If so, then this would provide evidence for a 
serial architecture. This is because in order to be immediately sensitive to these 
blocking constraints, comprehenders must be consider syntactic positions that are not 
directly involved in constructing the dependency. In a serial architecture, this can be 
made to easily follow from the requirement that the parse tree be traversed node-by-
 node. In a content-addressable architecture, on the other hand, only constituents that 
have interfering content should be able to impact early dependency building 
processes. Thus, blocking effects such as the quantifier intervention effect should not 
be immediately evident to comprehenders. In a content-addressable architecture, it 
seems that these structures must be generated and then filtered out at a later stage of 
comprehension. 
 
Negative constraints 
 
 In addition to blocking constraints, an investigation of negative constraints on 
parsing operations has the potential to be informative about the architectural source of 
structured access. Negative constraints are restrictions on which syntactic positions 
cannot be considered for participation in a long-distance dependency. Principle B of 
 
 
331 
the binding theory (Chomsky 1981) is the most prominent example: the antecedent 
for a pronoun can be any NP that is not in its local binding domain.  
 Although positive syntactic constraints can arguably be accurately stated in a 
content-addressable implementation of structured access, it is not clear that such 
negative constraints can. In these architectures, the retrieval cues consist of a positive 
set of features that are matched against the content of items in memory; it is not clear 
that negative constraints are well-formed in such a system. An intuitive way of seeing 
this is to consider the effect of the instructing someone not to direct their attention to 
a prominent event; a more likely outcome is that this will increase the chances that 
they will in fact attend to the event.  
 Thus the control structure needed to accurately and quickly exclude positions 
from consideration is likely to require an alternative architecture. For this reason, the 
degree of accuracy in the online implementation of Principle B effects has the 
potential to be highly informative. There are mixed empirical results concerning how 
quickly comprehenders can exclude the local subject position from consideration, 
with some studies suggesting that there is no consideration of the local subject for 
free pronouns (Nicol 1988; Clifton, Kennison & Albrecht 1997; Chow, Lewis, Lee & 
Phillips 2011), and others suggesting that the local subject is in fact considered 
(Badecker & Straub 2002; Kennison 2003).  
 If comprehenders are able quickly exclude the local subject from 
consideration, then it is possible to mount an argument for the deployment of a more 
structured architecture to account for the immediate application of this negative 
constraint. Otherwise, if the parser is ultimately found to be unable to exclude the 
 
 
332 
local subject position from consideration in Principle B configurations, then support 
for a serial, complementary memory architecture is weakened; instead, the inability to 
quickly implement a negative syntactic constraint would point to a more general use 
of a content-addressable memory architecture. 
 
Conclusion 
 
 The link that has been made in recent years between working memory 
architectures and parsing has generated much insight into the fine computational 
properties of representation and memory during sentence comprehension. The present 
thesis contributes to this line of research; in particular, I have presented several 
experiments that demonstrate that despite the apparent widespread existence of 
interference effects, the parser can effectively target particular syntactic positions for 
retrieval. Recent advances in understanding of the memory architecture of the parser 
do not entail that the parser is architecturally constrained to entertain ungrammatical 
parses. On the contrary, an investigation into to the scope of structured access raises 
the possibility that narrowly syntactic memory access actually arises as a functional 
adaptation to exactly this sort of noisy cognitive architecture.  
 
 
 
 
 
 
 
 
333 
Appendix A: Retrieval schedules for models in Chapter 3 
 
 For each model simulation, a schedule of constituent creation times and a 
schedule of hypothesized retrievals was constructed. The time t at which a given 
constituent was created was estimated from the empirical reading times in 
Experiment 1; this time reflects the total amount of time that readers spent reading 
material to the left of the relevant constituent (i.e., average cumulative regression path 
duration). Retrievals tied to the processing of any given constituent, such as the 
retrieval of a subject upon processing a verb, occurred 200ms after the creation of the 
constituent that triggered the retrieval. The sole exception was the reflexive retrieval, 
which occurred at 300ms post processing to allow for extra time to attach the 
reflexive as the object of the verb. 
 The differences between the agreement and reflexive conditions were 
modeled only as differences in the feature makeup of the parse constituents and the 
feature specifications of the retrievals.  
 
Agreement conditions: 
 
[The executive]DP1 who [oversaw] VP1 [the middle manager(s)] DP2 [definitely] ADV [was] 
BE .. 
 
 DP1 VP1 DP2 Adv BE 
t 0 590 1116 1906 2330 
Gender masc - masc - - 
Number sing sing sing/pl - - 
Category DP VP DP ADV T 
Role [spec,T] [head,T] [comp,V] [adj,T] [head,T] 
Local + - - + + 
Table A.1: Constituent creation times and feature makeup for agreement conditions. 
 
 
 R1:VP1 R2:DP2 R3:BE 
t 790 1316 2530 
Gender - - - 
Number - - sing/pl 
Category DP VP DP 
Role [spec,T] [head,T] [head,T] 
Local - - + 
Table A.2: Schedule of retrievals and cue sets. R1 = attachment of DP1 to VP1; R2 = 
attachment of DP2 to VP1; R3 = critical retrieval to attach DP1 to BE. 
 
 
 
 
 
 
334 
 
 
 
Reflexive conditions: 
 
[The executive]DP1 who [oversaw] VP1 [the middle manager(s)] DP2 [definitely] ADV [hurt] 
VP2 [himself] REFL ? 
 
 DP1 VP1 DP2 Adv VP2 REFL 
t 0 580 1091 1850 2285 2624 
Gender masc - masc - - - 
Number sing sing sing/pl - sing/pl - 
Category DP VP DP ADV VP REFL 
Role [spec,T] [head,T] [comp,V] [adj,T] [head,T] [comp,V] 
Local + - - + + + 
Table A.3: Constituent creation times and feature makeup for agreement conditions. 
 
 
 
Table A.4: Schedule of retrievals and cue sets for the structured access reflexive 
model. R1 = attachment of DP1 to VP1; R2 = attachment of DP2 to VP1; R3 = 
attachment of DP1 to VP2; R4 = attachment of REFL to VP2; R5 = critical retrieval 
of REFL?s antecedent. 
 
 
 
Table A.5: Schedule of retrievals and cue sets for the feature-based access reflexive 
model. R1 = attachment of DP1 to VP1; R2 = attachment of DP2 to VP1; R3 = 
attachment of DP1 to VP2; R4 = attachment of REFL to VP2; R5 = critical retrieval 
of REFL?s antecedent. 
 
 
 
 R1:VP1 R2:DP2 R3:VP2 R4:REFL R5:REFL 
t 780 1316 2530 2824 2924 
Gender - - - - - 
Number - - - - - 
Category DP VP DP VP DP 
Role [spec,T] [head,T] [head,T] [head,T] [spec,T] 
Local - - + + + 
 R1:VP1 R2:DP2 R3:VP2 R4:REFL R5:REFL 
t 780 1316 2530 2824 2924 
Gender - - - - masc/- 
Number - - - - sing/pl 
Category DP VP DP VP DP 
Role [spec,T] [head,T] [head,T] [head,T] [spec,T] 
Local - - + + + 
 
 
335 
 
References 
 
Alcocer, P., & Phillips, C. (2009). A cross-language reversal in illusory licensing. 
Poster presented at the 22nd Annual Meeting of the CUNY Conference on 
Human Sentence Processing, Davis, CA: March 26-28. 
Anand, P. (2006). De de se. Doctoral dissertation, Massachusetts Institute of 
Technology.  
Anderson, J. R. (1974). Retrieval of propositional information from long-term 
memory. Cognitive Psychology, 5, 451-474. 
Anderson, J. R. (1989). A rational analysis of human memory. In Roediger III, H.,  
and Craik, F., (eds.) Varieties of Memory and Consciousness: Essays in 
Honor of Endel Tulving: 195-210. Erlbaum, Hillsdale, NJ. 
Anderson, J. R. (1990). The Adaptive Character of Thought. Erlbaum, Hillsdale, NJ. 
Anderson, J. R. & Milson, R. (1989). Human Memory: An Adaptive Perspective. 
Psychological Review, 96, 703-719. 
Anderson, J. R., & Schooler, L. (1991). Reflections of the environment in memory. 
Psychological Science, 2, 396-408. 
Anderson, J. R., & Lebiere, C. (1998). The atomic components of thought. Erlbaum, 
Mahwah, NJ. 
Anderson, J. R. & Reder, L. (1999). The fan effect: New results and new theories. 
Journal of Experimental Psychology: General, 128, 186-197. 
Ant?n-Mendez, I., Nicol, J., & Garrett, M. (2002). The relation between gender and 
number agreement processing. Syntax, 5, 1-25. 
Aoun, J., & Li, Y-H A. (1993). Wh-elements in situ: Syntax or LF? Linguistic 
Inquiry, 24, 199-238. 
Baayen, H. (2008). Analyzing Linguistic Data: A Practical Introduction to Statistics 
Using R. Cambridge University Press, Cambridge, UK.  
Baayen, H., Davidson, D., & Bates, D. (2008). Mixed-effects modeling with crossed 
random effects for subjects and items. Journal of Memory and Language, 59, 
390-412. 
Badecker, W. A feature principle for partial agreement. Lingua, 117, 1541-1565. 
Badecker, W., & Straub K. (2002). The processing role of structural constraints on 
the interpretation of pronouns and anaphors. Journal of Experimental 
Psychology: Learning, Memory and Cognition, 28, 748-769. 
Badecker, W., & Kuminiak, F. (2007). Morphology, agreement and working memory 
retrieval in sentence production: Evidence from gender and case in Slovak. 
Journal of Memory and Language, 56, 65-85. 
 
 
336 
Baker, C-L. (1995). Contrast, discourse prominence, and intensification, with special 
reference to locally-free reflexives in British English. Language, 71, 63-101. 
Baker, M. (2008). The Syntax of Agreement and Concord. Cambridge University 
Press, Cambridge, UK. 
Baron-Cohen, S., Wheelwright, S., Skinner, R., Martin, J., & Clubley, E. (2001). The 
Autism-Spectrum Quotient (AQ): Evidence from Asperger Syndrome/high-
 functioning Autism, males and females, scientists and mathematicians. 
Journal of Autism and Developmental Disorders, 31, 5-17. 
Beck, S. (1996). Quantified structures as barriers for LF-movement. Natural 
Language Semantics, 4, 1-56. 
Bergeton, U. (2007). The independence of binding and intensification. Doctoral 
dissertation, University of Southern California.  
Berwick, R., & Weinberg, A. (1984). The Grammatical Basis of Linguistic 
Performance. MIT Press, Cambridge.  
Bever, T. (1970). The cognitive basis for linguistic structures. In Hayes, R. (ed.), 
Cognition and Language. Wiley & Sons, New York. 
Bhatt, R. (2005). Long-distance agreement in Hindi-Urdu. Natural Language and 
Linguistic Theory, 23, 757-807. 
Bobaljik, J. (2008). Where?s phi? Agreement as a post-syntactic operation. In 
Harbour, D., Adger, D., & B?jar, S. (eds), Phi-Theory: Phi Features Across 
Interfaces and Modules. Oxford University Press, Oxford, UK.  
Bock, K. & Miller, C. (1991). Broken Agreement. Cognitive Psychology, 23, 45-93. 
Bock, K., Nicol, J., & Cutting, J. C. (1999). The ties that bind: Creating number 
agreement in speech. Journal of Memory and Language, 40, 330-346.  
Bock, K., Eberhard, K., & Cutting, J. C. (2004). Producing number agreement: How 
pronouns equal verbs. Journal of Memory and Language, 51, 251-278. 
Bock., K., Butterfield, S., Cutler, A., Cutting, J. C., Eberhard, K., & Humphreys, K. 
Number agreement in British and American English: Disagreeing to agree 
collectively. Language, 82, 64-113. 
Bornkessel, I., McElree, B., Schlesewsky, M., & Friederici, A. D. (2004). Multi-
 dimensional contributions to garden path strength: dissociating phrase 
structure from relational structure. Journal of Memory and Language, 51, 
495?522. 
Bornkessel, I., & Schlesewsky, M. (2006). The extended argument dependency 
model: A neurocognitive approach to sentence comprehension across 
languages. Psychological Review, 113, 787-821.  
Brysbaert, M., & Mitchell, D. (2000). The failure to use gender information in 
parsing: A comment on van Berkum, Brown & Hagoort (1999). Journal of 
Psycholinguistic Research, 29, 453-455. 
B?ring, D. (2005). Binding Theory. Cambridge University Press, Cambridge. 
 
 
337 
Burkhardt, P. (2005). The Syntax-Discourse Interface: Representing and Interpreting 
Dependency. John Benjamins, Amsterdam. 
Caplan, D., & Waters, G. (1998). Verbal working memory and sentence 
comprehension. Brain and Behavioral Sciences, 22, 77-126. 
Carrasco, M., Giordano, A. M., & McElree, B. (2006). Attention speeds processing 
across eccentricity: Feature and conjunction searches. Vision Research, 46, 
2028?2040.  
Chandler, J. P. (1969). Subroutine STEPIT?finds local minimum of a smooth 
function of several parameters. Behavioral Science, 14, 81?82. 
Cheng, L-S L. (2009). Wh-in-situ, from the 1980s to now. Language and Linguistics 
Compass, 3, 767-791. 
Chomsky, N. (1973). Conditions on Transformations. In Kiparsky, P., & Peters, S. 
(eds), A Festschrift for Morris Halle. Mouton, The Hague. 
Chomsky, N. (1977). Essays on Form and Interpretation. Elsevier, North-Holland.  
Chomsky, N. (1981). Lectures on government and binding. Mouton de Gruyter: 
Berlin. 
Chomsky, N. (1986). Knowledge of language: its nature, origins and use. New York. 
Chomsky, N., & Lasnik, H. 1993. The theory of principles and parameters. In Jacobs, 
J., Von Stechow, A., & Sterneveld, W., (eds.) Syntax: An International 
Handbook of Contemporary Research, Vol. 1. Mouton de Gruyter, Berlin. 
Chomsky, N. (2000). Minimalist Inquiries: The Framework. In Martin, R., Michaels, 
D., & Uriagereka, J. (eds), Step by Step: Essays in Minimalist Syntax in Honor 
of Howard Lasnik. MIT Press, Cambridge, MA.  
Chow, W-Y., Lewis, W., Lee, S., & Phillips, C. (2011). Immediate structural 
constraints on pronoun antecedent retrieval. Poster presented at the 24nd 
Annual Meeting of the CUNY Conference on Human Sentence Processing, 
Stanford, CA: March 24-26. 
Clifton, C., Kennison, S., & Albrecht, J. (1997). Reading the words her, his, him: 
Implications for parsing principles based on frequency and on structure. 
Journal of Memory and Language, 36, 276-292. 
Clifton, C., Frazier, L., & Deevy, P. (1999). Feature manipulation in sentence 
comprehension. Rivista di Linguistica, 11, 11-39. 
Cole, P., Hermon, G., & Sung, L-M. (1990). Principles and parameters of long-
 distance reflexives. Linguistic Inquiry, 21, 1-22.  
Cole, P., & Sung, L-M. (1994). Head movement and long-distance reflexives. 
Linguistic Inquiry, 25, 355-406. 
Cole, P., Hermon, G., & Lee, C-L. (2001). Grammatical and discourse conditions on 
long-distance reflexives in two Chinese dialects. In Cole, P., Hermon, G., & 
Huang, J. (eds), Long Distance Reflexives. Academic Press, New York. 
 
 
338 
Corbett, G. (2006). Agreement. Cambridge University Press, Cambridge, UK. 
Coulson, S., King, J., & Kutas, M. (1998). Expect the unexpected: Event-related brain 
responses to morphosyntactic violations. Language and Cognitive Processes, 
13, 21-58. 
Cuetos, F., & Mitchell, D. (1988). Cross-linguistic differences in parsing: Restrictions 
on the use of the Late Closure strategy in Spanish. Cognition, 30, 73-105. 
De Vincenzi, M. (1999). Differences between the morphology of gender and number: 
Evidence from establishing coreferences. Journal of Psycholinguistic 
Research, 28, 537-553. 
Den Dikken, M. (2001). ?Pluringulars,? pronouns, and quirky agreement. The 
Linguistic Review, 18, 19-41. 
Dosher, B. A. (1984). Discriminating pre-experimental (semantic) information from 
learned (episodic) associations: A speed?accuracy study. Cognitive 
Psychology, 16, 519-555.  
Dosher, B. A., & Rosedale, G. (1991). Judgments of semantic and episodic 
relatedness: Common time-course and failure of segregation. Journal of 
Memory and Language, 30, 125-160. 
Drenhaus, H., Frisch, S., & Saddy, D. (2005). Processing negative polarity items: 
When negation comes through the backdoor. In: Kepser, S. and Reis, M. 
(eds.), Linguistic Evidence-Empirical, Theoretical, and Computational 
Perspectives. Mouton de Gruyter, Berlin. 
Eberhard, K., Cutting, J., & Bock, K. (2005). Making syntax of sense: number 
agreement in sentence production. Psychological Review, 112, 531-559. 
Erteschik-Shir, N. (2006). Bridge phenomena. In Everaert, M., & van Riemsdijk, H. 
(eds), The Blackwell Companion to Syntax. Blackwell/Wiley, New York. 
Fedorenko, E., Babyonyshev, M., & Gibson, E. (2004). The nature of case 
interference in online sentence processing in Russian. NELS 34 Conference 
Proceedings. 
Ferreira, F., Bailey, K, & Ferraro, V. (2002). Good-enough representations in 
language comprehension. Current Directions in Psychological Science, 11, 
11-15. 
Ferreira, F., & Patson, N. (2007). The ?good-enough? approach to language 
comprehension. Language and Linguistics Compass, 1, 71-83. 
Fiengo, R., & May, R. (1994). Indices and identity. MIT Press, Cambridge, MA. 
Fodor, J.A., Bever, T., & Garrett, J. (1974). The psychology of language: an 
introduction to psycholinguistics and generative grammar. McGraw-Hill, 
New York. 
Foote, R., & Bock, K. (2011). The role of morphology in subject-verb number 
agreement: A comparison of Mexican and Dominican Spanish. Language and 
Cognitive Processes, forthcoming.  
 
 
339 
Foraker, S., & McElree, B. (2007). The role of prominence in pronoun resolution: 
Availability versus accessibility. Journal of Memory and Language, 56, 357-
 383. 
Franck, J., Vigliocco, G., & Nicol, J. (2002). Subject-verb agreement errors in French 
and English: The role of syntactic hierarchy. Language and Cognitive 
Processes, 17, 371-404. 
Frazier, L. (1978). On comprehending sentences: Syntactic parsing strategies. 
Doctoral dissertation, University of Connecticut. 
Frazier, L. (1987). Syntactic processing: Evidence from Dutch. Natural Language 
and Linguistic Theory, 5, 519-560. 
Frazier, L. (1998). Getting there (slowly). Journal of Psycholinguistic Research, 27, 
123-146. 
Frazier, L., & Fodor, J. D. (1978). The sausage machine: a new two-stage parsing 
model. Cognition, 6, 291-325. 
Frazier, L., Clifton, C., & Randall, J. (1983). Filling gaps: decision principles and 
structure in sentence comprehension. Cognition, 13, 187-222. 
Frazier, L., & Clifton, C. (1989). Successive cyclicity in the grammar and the parser. 
Language and Cognitive Processes, 4, 93-126. 
Frazier, L., Flores-d?Arcais, G. (1989). Filler-driven parsing: A study of gap filling in 
Dutch. Journal of Memory and Language, 28, 331-344. 
Frazier, L., & Clifton, C. (1996). Construal. MIT Press, Cambridge. 
Frazier, L., & Clifton, C. (1997). Construal: Overview, motivation and some new 
evidence. Journal of Psycholinguistic Research, 26, 277-295.  
Friederici, A. (1995). The time course of syntactic activation during sentence 
processing: A model based on neuropsychological and neurophysiological 
data. Brain and Language, 50, 259-281. 
Friederici, A., Pfeifer, E., & Hahne, A. (1993). Event-related brain potentials during 
natural speech processing: Effects of semantic, morphological and syntactic 
violations. Brain Research: Cognitive Brain Research, 1, 183?192. 
Friederici, A. D., Hahne, A., & Saddy, D. (2002). Distinct neurophysiological 
patterns reflecting aspects of syntactic complexity and syntactic repair. 
Journal of Psycholinguistic Research, 31, 45-63. 
Gallistel, R. (2009). The importance of proving the null. Psychological Review, 116, 
439-453. 
Gallistel, R., & King, A. (2009). Memory and the computational brain: Why cognitive 
science will transform neuroscience. Blackwell/Wiley, New York.  
Gao, L., Liu, Z., & Huang, Y.. (2005). Who is ziji? An experimental research on 
Binding Principle. Linguistic Science, 2, 39?50. 
Gelman, A., & Hill, J. (2005). Data analysis using regression and multilevel / 
 
 
340 
hierarchical models. Cambridge University Press, Cambridge, UK.  
Gernsbacher, M. (1989). Mechanisms that improve referential access. Cognition, 32, 
99-156. 
Gibson, E., Pearlmutter, N., Canseco-Gonzalez, E., & Hickok, G. (1996). Recency 
preference in the human sentence processing mechanism. Cognition, 59, 23-
 59. 
Gibson, E., & Warren, T. (2004). Reading time evidence for intermediate linguistic 
structure in long-distance dependencies. Syntax, 7, 55-78. 
Gillespie, M., & Pearlmutter, N. (2011). Hierarchy and scope of planning in subject-
 verb agreement production. Cognition, 118, 377-397. 
Gillund, G., & Shiffrin, R. M. (1984). A retrieval model for both recognition and 
recall. Psychological Review, 91, 1?65. 
Gordon, P.C., Hendrick, R., & Johnson, M. (2001). Memory interference during 
sentence processing. Psychological Science, 13, 425-430. 
Gordon, P.C., Hendrick, R., & Levine, W. (2002). Memory-load interference in 
syntactic processing. Journal of Experimental Psychology: Learning, Memory 
& Cognition, 27, 1411-1423. 
Gordon, P.C., Hendrick, R., & Johnson, M. (2004). Effects of noun phrase type on 
sentence complexity. Journal of Memory and Language, 51, 97-114. 
Gordon, P.C., Hendrick, R., Johnson, M., & Lee, Y. (2006). Similarity-based 
interference during language comprehension: Evidence from eye tracking 
during reading. Journal of Experimental Psychology: Learning, Memory and 
Cognition, 32, 1304-1321.  
Gronlund, S. D., Edwards, M. B., & Ohrt, D. D. (1997). Comparison of the retrieval 
of item versus spatial position information. Journal of Experimental 
Psychology: Learning, Memory, & Cognition, 23, 1261?1274. 
Greene, S.B., McKoon, G., & Ratcliff, R. (1992). Pronoun resolution and discourse 
models. Journal of Experimental Psychology; Learning, Memory and 
Cognition, 18, 266-283. 
Hagoort, P. (2003). How the brain solves the binding problem for language: A 
neurocomputational model of syntactic processing. Neuroimage, 20, 18-29. 
Hagoort, P., Brown, C. M., & Groothusen, J. (1993). The syntactic positive shift 
(SPS) as an ERP measure of syntactic processing. Language and Cognitive 
Processes, 8, 439?483. 
Hagoort, P., Wassenaar, M. E. D., & Brown, C. M. (2003). Syntax-related ERP 
effects in Dutch. Cognitive Brain Research, 16, 38-50. 
Hahne, A., & Friederici, A. D. (1999). Electrophysiological evidence for two steps in 
syntactic analysis: Early automatic and late controlled processes. Journal of 
Cognitive Neuroscience, 11, 194-205. 
Hankamer, J., & Sag, I. (1976). Deep and surface anaphora. Linguistic Inquiry,7, 391-
 
 
341 
426 
Harris, T., Wexler, K., & Holcomb, P. (2000). An ERP investigation of binding and 
coreference. Brain and Language, 75, 313?346. 
Hartsuiker, R., Schriefers, H., Bock, K., & Kikstra, G. (2003). Morphophonological 
influences on the construction of subject-verb agreement. Memory & 
Cognition, 31, 1316-1326. 
H?ussler, J., & Bader, M. (2009). Agreement checking and number attraction in 
sentence comprehension: Insights from German relative clauses. Travaux du 
cercle linguistique de Prague 7. 
Hopf, J-M., Bayer, J., Bader, M., & Meng, M. (1998). Event-related brain potentials 
and case information in syntactic ambiguities. Journal of Cognitive 
Neuroscience, 10, 264-280. 
Hopfield, J.J. (1982). Neural networks and physical with emergent collective 
computational abilities. Proceedings of the National Academy of Sciences, 79,  
2554-2558. 
Hornstein, N. (2007). Pronouns in a minimalist setting. In Corver, N., & Nunes, J. 
(eds.), The Copy Theory of Movement. John Benjamins, Philadelphia.  
Huang, C-T J. (1982). Logical Relations in Chinese and the Theory of Grammar. 
Doctoral Dissertation, Massachusetts Institute of Technology. 
Huang, C-T J., & Tang, C-C, J. (1991). The local nature of the long-distance reflexive 
in Chinese. In Koster, J., & Reuland, E. (eds.), Long Distance Anaphora. 
Cambridge University Press, Cambridge, UK. 
Huang, C-T J., & Liu, L. (2001). Logophoricity, attitudes, and ziji at the interface. In: 
Cole, P., Huang, C-T J., and Hermon, G. (eds.), Long Distance Reflexives, 
Syntax and Semantics 33, Academic Press, New York, 141-195. 
Huang, C-T. J., Cole, P., & Hermon, G. (2006). Long-Distance Reflexives: an East 
Asian Perspective. In: Everaert, M., van Riemsdijk, H., Goedemans, R., and 
Hollebrandse, B (eds.) The Blackwell Companion to Syntax Volume III, pp. 
21-84. Location: Blackwell. 
Huang, C-T J., Li, Y-H A., & Li, Y. (2009). The Syntax of Chinese. Cambridge 
University Press, Cambridge, UK.  
Huang, Y. (2000). Anaphora: A Cross-linguistic Study. Oxford University Press, 
Oxford, UK.  
Jackendoff, R. (1972). Semantic Interpretation in Generative Grammar. MIT Press, 
Cambridge, MA. 
Jaeger, F. (2008). Categorical data analysis: Away from ANOVAs (transformation or 
not) and towards logit mixed models. Journal of Memory and Language, 59, 
434-446. 
Jarvella, R. (1970)  Effects of syntax on running memory span for connected 
discourse. Psychonomic Science, 19, 235-236. 
 
 
342 
Jarvella, R, & Pisoni, S. (1971). Relation between syntactic and perceptual units in 
speech processing. Journal of the Acoustical Society of America, 48, 84-88.  
Jin, Z.H., 2003. Verb restraint function to ziji long-distance binding. Chinese 
Language Learning, 4, 9?12. 
Johnson, K. (2001). What VP ellipsis can do, what it can?t, but not why. In Baltin, 
M., & Collins, C. (eds), The Handbook of Contemporary Syntactic Theory. 
Blackwell/Wiley, New York. 
Jonides, J., Lewis, R.L., Nee, D.E., Lustig, C.A., Berman, M.G., & Moore, K.S. 
(2008). The mind and brain of short-term memory. Annual Review of 
Psychology, 59, 193-224. 
Jurafsky, D. (1996). A probabilistic model of lexical and syntactic access and 
disambiguation. Cognitive Science, 20, 137-194. 
Kaan, E. & Swaab, T. (2003). Repair, revision, and complexity in syntactic analysis: 
An electrophysiological differentiation. Journal of Cognitive Neuroscience, 
15, 98-110. 
Kaiser, E. (2006). The quest for a referent: A cross-linguistic look at reference 
resolution. Doctoral dissertation, University of Pennsylvania.  
Kaiser, E., Runner, J., Sussman, R., & Tanenhaus, M. (2009). Structural and semantic 
constraints on the resolution of pronouns and reflexives. Cognition, 112, 55-
 80. 
Kamide, Y., & Mitchell, D. (1997). Relative clause attachment: Nondeterminism in 
Japanese parsing. Journal of Psycholinguistic Research, 26, 247-254. 
Kayne, R. (1989). Notes on English agreement. Central Institute of English and 
Foreign Languages Bulletin, 1, 46-67. 
Kennsion, S. (2003). Comprehending the pronouns her, him and his: Implications for 
theories of referential processing. Journal of Memory and Language, 49, 335-
 352. 
Kennsion, S., & Trofe, J. (2003). Comprehending pronouns: A role for word-specific 
gender stereotype information. Journal of Psycholinguistic Research, 32, 355-
 378. 
Kim, A., & Osterhout, L. (2005). The independence of combinatory semantic 
processing: Evidence from event-related potentials. Journal of Memory and 
Language, 52, 205-225. 
Kimball, J. & Aissen, J. (1971). I think, you think, he think. Linguistic Inquiry, 2, 
241-246. 
Kimball, J. (1973). Seven principles of surface structure parsing in natural language. 
Cognition, 2, 15-47.  
King, J., & Kutas, M. (2005). Who did what and when? Using world- and clause-
 level ERPs to monitor working memory usage in reading. Journal of 
Cognitive Neuroscience, 7, 376-395.  
 
 
343 
Kluender, R., & Kutas, M. (1993). Bridging the gap: Evidence from ERPs on the 
processing of unbounded dependencies. Journal of Cognitive Neuroscience, 5, 
196-214.  
Kohonen, T. (1980). Content-addressable memories. Springer Verlag, Berlin. 
Kolk, H., Chwilla, D., van Herten, M., & Oor, P. (2003). Structure and limited 
capacity in verbal working memory: A study with event-related potentials. 
Brain and Language, 85, 1-36. 
K?nig, E., & Siemund, P. (1999). Intensifiers and reflexives: A typological 
perspective. In Frajzyngier, Z., & Curl, T. (eds), Reflexives: Forms and 
functions. John Benjamin, Amsterdam.  
Kratzer, A. (2009). Making a pronoun: fake indexicals as windows into the properties 
of pronouns. Linguistic Inquiry, 40, 187-237.  
Kuno, S. (1987). Functional syntax: anaphora, discourse and empathy. University of 
Chicago Press, Chicago.  
Kuperberg, G. (2007). Neural mechanisms of language comprehension: Challenges to 
syntax. Brain Research, 1146, 23-49.  
Kutas, M., & Hillyard, S. A. (1980). Reading senseless sentences: Brain potentials 
reflect semantic incongruity. Science, 207, 203?205. 
Kutas, M. and Federmeier, K. D. (2000). Electrophysiology reveals semantic memory 
use in language comprehension. Trends in Cognitive Science, 4, 463-470. 
Ladusaw, W. A. (1979). Negative polarity items as inherent scope relations. Doctoral 
dissertation, University of Texas, Austin. 
Lago, S., Alcocer, P., & Phillips, C. (2011). Agreement attraction in Spanish: 
Immediate vs. delayed sensitivity. Poster presented at the 24nd Annual 
Meeting of the CUNY Conference on Human Sentence Processing, Stanford, 
CA: March 24-26. 
Lasnik, H. (1976). Remarks on coreference. Linguistic Analysis, 2, 1-22. 
Lau, E. (2010). The predictive nature of language comprehension. Doctoral 
dissertation, University of Maryland. 
Lau, E. F., Stroud, C., Plesch, S., & Phillips, C. (2006). The role of structural 
prediction in rapid syntactic analysis. Brain & Language, 98, 74-88. 
Lau, E. F., Phillips, C., & Poeppel, D. (2008). A cortical network for semantics: 
(de)constructing the N400. Nature Reviews Neuroscience, 9, 920-933. 
Lees, R., & Klima, E. (1963). Rules for English pronominalization. Language, 39, 
17-28. 
Levelt, W. (1974). Formal grammars in linguistics and psycholinguistics. John 
Benjamins, Philadelphia.  
Levy, R. (2008). Expectation-based syntactic comprehension. Cognition, 106, 1126-
 1177. 
 
 
344 
Lewis, R. (1996). Interference in short-term memory: the magical number two (or 
three) in Sentence Processing. Journal of Psycholinguistic Research, 25,93-
 115. 
Lewis, R., & Vasishth, S. (2005). An activation-based model of sentence processing 
as skilled memory retrieval. Cognitive Science, 29(3), 375-419  
Lewis, R., Vasishth, S., & Van Dyke, J (2006). Computational principles of working 
memory in sentence comprehension. Trends in Cognitive Science, 10, 447-
 454. 
Li, X & Zhou, X. (2010). Who is ziji? ERP responses to the Chinese reflexive 
pronound during sentence comprehension. Brain Research, 1331, 96-104. 
Lidz, J., & Idsardi, W. (1999). Chains and phono-logical form. University of 
Pennsylvania Working Papers in Lingusitics, 5, 119-125. 
Liu, Z. (2009). The cognitive process of Chinese reflexive processing. Journal of 
Chinese  Linguistics, 37, 1-27. 
Li, C., & Smith, P. (2009). Comparing time-accuracy curves: beyond goodness-of-fit 
measures. Psychonomic Bulletin & Review, 16, 190-203. 
L?drup, H. (2009). Animacy and long distance binding in Norwegian. Nordic Journal 
of Linguistics, 32, 111-136. 
Logacev, P. & Vasishth, S. (to appear). Morphological ambiguity and working 
memory. In Lamers, M., & de Swart, P., (eds.), Case, Word Order and 
Prominence. 
Lorimor, H., Bock, K., Zalkind, E., Sheyman, A., & Beard R. (2008). Agreement and 
attraction in Russian. Language and Cognitive Processes, 23, 769-799. 
MacDonald, M., & MacWhinney, B. (1990). Measuring inhibition and facilitation 
from pronouns. Journal of Memory and Language, 29, 469-492. 
MacDonald, M., Pearlmutter, N., & Seidenberg, M. (1994). Lexical nature of 
ambiguity resolution. Psychological Review,101, 676-703.  
Martin, A. E. & McElree, B. (2008). A content-addressable pointer mechanism 
underlies comprehension of verb-phrase ellipsis. Journal of Memory and 
Language, 58, 879-906. 
Martin, A. E. & McElree, B. (2009). Memory operations that support sentence 
comprehension: evidence from verb-phrase ellipsis. Journal of Experimental 
Psychology: Learning, Memory and Cognition, 35, 1231-1239. 
Marcus, M. P. (1980). Theory of Syntactic Recognition for Natural Languages. MIT 
Press, Cambridge.  
Marr, D. (1982). Vision: A Computational Investigation into the Human 
Representation and Processing of Visual Information. Freeman, New York. 
McCloskey, J. (2001). The morphosyntax of WH-extraction in Irish. Journal of 
Linguistics, 37, 67-100. 
 
 
345 
McElree, B. (1990). The time course of recency discrimination: A speed accuracy 
analysis. Doctoral dissertation, Columbia University. 
McElree, B. (1993). The locus of lexical preference effects in sentence 
comprehension: A time-course analysis.Journal of Memory and 
Language, 32, 536-571.  
McElree, B. (1998). Attended and non-attended states in working memory: Accessing 
categorized structures. Journal of Memory & Language, 38, 225-252. 
McElree, B. (2000). Sentence comprehension is mediated by content-addressable 
memory structures. Journal of Psycholinguistic Research, 29, 111-123. 
McElree, B., & Dosher, B. (1989). Serial position and set size in short-term memory: 
time course of recognition. Journal of Experimental Psychology: General, 18, 
346-373. 
McElree, B., & Dosher, B. (1993). Serial retrieval processes in the recovery of order 
information. Journal of Experimental Psychology: General, 122, 291-315. 
McElree, B., & Griffith, T. (1998). Structural and lexical constraints on filling gaps 
during sentence processing: A time-course analysis. Journal of Experimental 
Psychology: Learning, Memory, and Cognition, 24, 432-460. 
McElree, B., Foraker, S. & Dyer, L. (2003). Memory structures that subserve 
sentence comprehension. Journal of Memory and Language, 48, 67-91.  
McRoy, S.W., & Hirst, G. (1990). Race-based parsing and syntactic disambiguation. 
Cognitive Psychology, 14, 313-353. 
Miller, G., & Chomsky, N. (1963). Finitary models of language users. In Luce, R.D., 
Bush, R.R., & Galanter, E. (eds), Handbook of Mathematical Psychology, 
Volume II. John Wiley, New York.  
Mitchell, D., Cuetos, F., Corley, M., & Brysbaert, M. (1995). Exposure-based models 
of human parsing: Evidence for the use of coarse-grained (nonlexical) 
statistical records. Journal of Psycholinguistic Research, 24, 469-488. 
Mudock, B.B. Jr. (1962). The serial position effect of free recall. Journal of 
Experimental Psychology, 64, 482-488. 
Nairne, J. (1988). A framework for interpreting recency effects in immediate serial 
recall. Memory & Cognition, 16, 343-352. 
Nairne, J. (1990). A feature model of immediate memory. Memory & Cognition, 18, 
251-269. 
Neville, H. J., Nicol, J., Barss, A., Forster, K. I., & Garrett, M. F. (1991). 
Syntactically based sentence processing classes: Evidence from event-related 
brain potentials. Journal of Cognitive Neuroscience, 3, 151?165. 
Nevins, A., Dillon, B., Malhotra, S., & Phillips, C. (2007). The role of feature-
 number and feature-type in processing Hindi verb agreement violations. Brain 
Research, 1164, 81-94. 
 
 
346 
Nicol, J. (1988). Coreference processing during sentence comprehension. Doctoral 
dissertation, MIT.  
Nicol, J., & Swinney, D. (1989). The role of structure in coreference assignment 
during sentence comprehension. Journal of Psycholinguistic Research, 18, 5-
 19. 
Oberauer, K., & Kliegl, R. (2006). A formal model of capacity limits in working 
memory. Journal of Memory and Language, 55, 601-626. 
Oiry, M. (2011). A case of true optionality: Wh-in-situ patterns like long movement in 
French. In Roussou, A., & Vlachos, C., (eds)  Linguistic Analysis: Optionality 
in wh-movement,  
Osterhout, L., & Holcomb, P. (1992). Event-related potentials elicited by syntactic 
anomaly. Journal of Memory and Language, 6, 785-806. 
Osterhout, L., Holcomb, P. J., & Swinney, D. A. (1994). Brain potentials elicited by 
garden path sentences: Evidence of the application of verb information during 
parsing. Journal of Experimental Psychology: Learning, Memory and 
Cognition, 20, 786?803. 
Osterhout, L., Bersick, M., & McLaughlin, J. (1997). Brain potentials reflect 
violations of gender stereotypes. Memory & Cognition, 25, 273?285. 
?ztekin, I., & McElree, B. (2010). Relationship between measures of working 
memory capacity and the time course of short-term memory retrieval and 
interference resolution. Journal of Experimental Psychology: Learning, 
Memory & Cognition, 36, 383-397.  
Pan, H. (1998). Closeness, prominence, and binding theory. Natural Language and 
Linguistic Theory, 16, 771-815. 
Pan, H. (2000). Why the blocking effect? In Cole, P., Huang, C-T J., & Hermon, G 
(eds), Syntax and Semantics Vol. 33: Long Distance Reflexives. Academic 
Press, New  York.  
Parker, D., & Phillips, C. (2011). Illusory negative polarity item licensing is selective. 
Poster presented at the 24nd Annual Meeting of the CUNY Conference on 
Human Sentence Processing, Stanford, CA: March 24-26. 
Pearlmutter, N., & MacDonald, M. (1995). Individual differences and probabilistic 
constraints in syntactic ambiguity resolution. Journal of Memory and 
Language, 34, 521-542.   
Pearlmutter, N., Garnsey, S., & Bock, K. (1999). Agreement processes in sentence 
comprehension. Journal of Memory and Language, 41, 427-456. 
Phillips, C., Wagers, M., & Lau, E. F. (2010). Grammatical illusions and selective 
fallibility in real-time language comprehension. In: J. Runner (ed.) 
Experiments at the Interference, Syntax & Semantics, vol. 37, Bingley, UK: 
Emerald Publications. 
 
 
347 
Pica, P. (1986). On the nature of the reflexivization cycle. In McDonough, J., & 
Plunkett, B. (eds), Proceedings of the 17th Annual Meeting of the North East 
Linguistic Society. 
Pinheiro, J., & Bates, D. (2000). Mixed-effects models in S and S-PLUS. Springer 
Verlag, New York. 
Pollard, C., & Sag, I. (1992). Anaphors in English and the scope of binding theory. 
Linguistic Inquiry, 23, 261-303. 
Pritchett, B. (1993). Grammatical Competence and Parsing Performance. University 
of Chicago Press, Chicago. 
Pulverm?ller, F., Shtyrov, Y., Hasting, A., & Carlyon, R. (2008). Syntax as a 
reflexive: Neurophysiological evidence for early automaticity of grammatical 
processing. Brain and Language, 104, 244-253.  
Ratcliff, R. (1980). A note on modeling accumulation of information when the rate of 
accumulation changes over time. Journal of Mathematical Psychology, 21, 
178-184. 
Ratcliff, R. (1983). Methods for dealing with reaction time outliers. Psychological 
Bulletin, 114, 510-532.  
Ratcliff, R., & McKoon, G. (2008). Passive parallel automatic minimalist processing. 
In  Engel, C., & Singer, W., (Eds.), Better than Conscious? Decision Making, 
the Human Mind, and Implications for Institutions. MIT Press, Cambridge. 
Reed, A. V. (1976). The time course of recognition in human memory. Memory and 
Cognition, 4, 16?30. 
Reinhart, T. (1976). The Syntactic Domain of Anaphora. Doctoral dissertation, 
Massachusetts Institute of Technology. 
Reinhard, T., & Reuland, E. (1993). Reflexivity. Linguistic Inquiry, 24, 657-720.  
Reuland, E. (2001). Primitives of binding. Linguistic Inquiry, 32, 439-492. 
Ricker, T., Aubuchon, A., & Cowen, N. (2010). Working memory. Wiley 
Interdisciplinary Reviews: Cognitive Science, 1, 573-585. 
Rizzi, L. (1990). On the anaphor-agreement effect. Rivista di Linguistica, 2, 27-42. 
Ross, J. (1967). Constraints on variables in syntax. Doctoral dissertation, 
Massachusetts Institute of Technology. 
Runner, J., Sussman, R., & Tanenhaus, M. (2003). Assignment of reference to 
reflexives and pronouns in picture noun phrases: evidence from eye 
movements. Cognition, 89, 1-13. 
Sag, I. (1976). Deletion and logical form. Doctoral dissertation, Massachusetts 
Institute of Technology. 
Sag, I., & Hankamer, J. (1984). Toward a theory of anaphoric processing. Linguistics 
and Philosophy, 7, 325-345. 
Sells, P. (1987). Aspects of logophoricity. Linguistic Inquiry, 18, 445-479. 
 
 
348 
Sprouse, J., Fukuda, S., Ono, H., & Kluender, R. (2011). Reverse island effects and 
the backward search for a licensor in multiple wh-questions. Syntax, 14, 179-
 203. 
Staub, A. (2010). Response time distributional evidence for distinct varieties of 
number attraction. Cognition, 114, 447-454. 
Sternberg, S. (1969). Memory scanning: Mental processes revealed by reaction-time 
experiments. American Scientist, 57, 421-457. 
 
Sturt, P. (2003a). The time-course of the application of binding constraints in 
reference resolution. Journal of Memory and Language, 48, 542-562. 
Sturt, P. (2003b). A new look at the syntax-discourse interference: The use of binding 
principles in sentence processing. Journal of Psycholinguistic Research, 32, 
125-139. 
Sturt, P., Pickering, M.J., & Crocker, M.W. (2000). Search strategies in syntactic 
reanalysis. Journal of Psycholinguistic Research, 29, 183-194. 
Suckow, K., Vasishth, S., & Lewis, R. (2005). Interference and memory overload 
during parsing. In Proceedings of AMLaP 2005, Ghent, Belgium, September 
2005. Ghent University. 
Sutton, R, & Barto, A. (1998). Introduction to Reinforcement Learning. MIT Press, 
Cambridge, MA.  
Tabor, W., Galantucci, B., & Richardson, D. (2004). Effects of merely local syntactic 
coherence on sentence processing. Journal of Memory and Language, 50, 
355-370. 
Tang, C-C J. (1989). Chinese reflexives. Natural Language and Linguistic Theory,7, 
93-121. 
Townsend, D. & Bever, T. (2001). Sentence comprehension: the integration of habits 
and rules. MIT Press, Cambridge.   
Traxler, M., Pickering, M., & Clifton, C. (1998). Adjunct attachment is not a form of 
lexical ambiguity resolution. Journal of Memory and Language, 39, 558-592. 
Tsai, W-T D. (1994). On nominal islands and LF extraction in Chinese. Natural 
Language and Linguistic Theory, 12, 121-175. 
van Berkum, J., Brown, C., & Hagoort, P. (1999a). When does gender constrain 
parsing? Evidence from ERPs. Journal of Psycholinguistic Research, 28, 555-
 566. 
van Berkum, J., Brown, C., & Hagoort, P. (1999b). Early referential context effects in 
sentence processing: Evidence from event-related brain potentials. Journal of 
Memory and Language, 41, 147-182. 
Van Dyke, J. A. (2007). Interference effects from grammatically unavailable 
constituents during sentence processing. Journal of Experimental Psychology: 
Learning, Memory, and Cognition, 33, 407-430. 
 
 
349 
Van Dyke, J. A. & McElree, B. (2006). Retrieval interference in sentence processing. 
Journal of Memory and Language, 55, 157-166. 
Van Gompel, R., Traxler, M., & Pickering, M. (2001). Reanalysis in sentence 
processing: Evidence against current constraint-based and two-stage Models. 
Journal of Memory and Language, 45, 225-258. 
Vasishth, S., Br?ssow, S., Lewis, R. & Drenhaus, H. (2008). Processing Polarity: 
How the ungrammatical intrudes on the grammatical. Cognitive Science, 32, 
685-712. 
Vigliocco, G., & Nicol J. (1998). Separting hierarchical relations and word order in 
language production: is proximity concord syntactic or linear? Cognition, 68, 
13-29. 
Vigliocco, G., & Franck, J. (2001). When sex affects syntax: Contextual influences in 
sentence production. Journal of Memory and Language, 45, 368-390. 
Vigliocco, G., & Hartsuiker, R. (2002). The interplay of meaning, sound, and syntax 
in language production. Psychological Bulletin, 128, 442?472. 
Wagers, M. (2008). The structure of memory meets memory for structure in linguistic 
comprehension. Doctoral dissertation, University of Maryland, College Park. 
Wagers, M., & McElree, B. (2009). Focal attention and the timing of memory 
retrieval in language comprehension. Talk given at the Architectures and 
Mechanisms for Language Processing Conference, September, Barcelona.  
Wagers, M., Lau, E. F., & Phillips, C. (2009). Agreement attraction in 
comprehension: representations and processes. Journal of Memory and 
Language, 61, 206-237. 
Wanner, E., & Maratsos, M. (1978). An ATN approach to comprehension. In Halle, 
M., Besnan, J., & Miller, G.A. (eds.), Linguistic Theory and Psychological 
Reality. MIT Press, Cambridge. 
Wasow, T. (1972). Anaphoric relations in English. Doctoral dissertation, 
Massachusetts Institute of Technology.  
Wechsler, S., & Zlatic, L. (2000). A theory of agreement and its application to Serbo-
 Croatian. Language, 76, 799-832.  
Wickens, T. (2001). Elementary Signal Detection Theory. Oxford University Press, 
New York. 
Wickelgren, W. (1977). Speed-accuracy tradeoff and information processing 
dynamics. Acta Psychologica, 41, 67-85.  
Wickelgren, W. A., Corbett, A., & Dosher, B. (1980). Priming and retrieval from 
short-term memory: A speed-accuracy tradeoff analysis. Journal of Verbal 
Learning and Verbal Behavior, 19, 387-404. 
Xiang, M., Dillon, B., & Phillips, C. (2006). Testing the strength of the spurious 
licensing effect for negative polarity items. Talk given at the 19th Annual 
CUNY Conference on Human Sentence Processing, New York, NY: March. 
 
 
350 
Xiang, M., Dillon, B., & Phillips, C. (2009). Illusory licensing across dependency 
types: ERP evidence. Brain and Language, 108, 40-55. 
Xiang, M., Dillon, B., & Wagers, M. (2010). Processing wh-movement dependencies 
in a language without wh-movement. Poster presented at the 23rd Annual 
Meeting of the CUNY Conference on Human Sentence Processing, New 
York, NY: March. 
Xue, P., Pollard, C., & Sag, I. A new perspective on Chinese ziji. Proceedings of the 
13th Annual Meeting of the West Coast Conference on Formal Linguistics, 
432-447. 
Yngve, V. (1960). A model and an hypothesis for language structure. Proceedings of 
the American Philosophical Society, 104, 444-466. 
Zribi-Hertz, A. (1995). Emphatic or reflexive? On the endophoric character of French 
lui-meme and similar complex pronouns. Journal of Linguistics, 31, 333-374.