Cognition 178 (2018) 162–177

Contents lists available at ScienceDirect

Cognition
journal homepage: www.elsevier.com/locate/cognit

Original Articles

Who did what? A causal role for cognitive control in thematic role T
assignment during sentence comprehension
⁎ ⁎
Malathi Thothathiria, , Christine T. Asaroa, Nina S. Hsub,c, Jared M. Novickb,c,
a
Department of Speech, Language and Hearing Sciences, The George Washington University, United States
b
Department of Hearing and Speech Sciences, University of Maryland, College Park, United States
c
Program in Neuroscience and Cognitive Science, University of Maryland, College Park, United States

A R T I C LE I N FO A B S T R A C T

Keywords: Thematic role assignment – generally, figuring out who did what to whom – is a critical component of sentence
Conflict adaptation comprehension, which is influenced by both syntactic and semantic cues. Conflict between these cues can result
Sentence processing in temporary consideration of multiple incompatible interpretations during real-time sentence processing. We
Syntax-semantics conflict tested whether the resolution of syntax-semantics conflict can be expedited by the online engagement of cog-
Stroop
nitive control processes that are routinely used to regulate behavior across domains. In this study, cognitive
Eye-tracking
control deployment from a previous Stroop trial influenced eye movements during subsequent sentence com-
prehension. Specifically, when syntactic and semantic cues competed for influence on interpretation, dynamic
cognitive control engagement led to (a) fewer overall looks to a picture illustrating the competing but incorrect
interpretation (Experiment 1), or (b) steeper growth in looks to a picture illustrating the correct interpretation
(Experiment 2). Thus, prior cognitive control engagement facilitated the resolution of syntax-semantics conflict
by biasing processing towards the intended analysis. This conflict adaptation effect demonstrates a causal
connection between cognitive control and real-time thematic role assignment. Broader patterns demonstrated
that prior cognitive control engagement also modulated sentence processing irrespective of the presence of
conflict, reflecting increased integration of newly arriving cues with prior sentential content. Together, the
results suggest that cognitive control helps listeners determine correct event roles during real-time compre-
hension.

1. Introduction
An important aspect of sentence comprehension involves thematic
role assignment, namely understanding the roles that various partici-
pants play in an event. For example, when hearing a sentence such as
“The boy chased the girl”, the listener must determine that the boy was
the one who did the pursuing and the girl was the one who did the
fleeing, not the other way around. Syntactic structure is an influential
cue that can guide comprehension of who did what to whom: It helps
determine which noun phrase in the sentence is the Agent of the action
(the doer) and which is the Patient (the affected entity). However,
structure is just one cue that informs the parser. Semantic knowledge
about the sentential nouns—e.g., whether they are animate like boys
and girls—and the real-world events mentioned in the sentence—e.g.,
whether the nouns are more plausibly Agents or Patients of a particular
transitive verb—can also influence interpretation (hence, “Man bites
dog” is surprising and headline-worthy). When syntactic and semantic
cues compete for influence, there are dramatic effects on comprehen-
sion, including uncertainty about who did what to whom and the need
to resolve competing thematic role assignments (see e.g., Ferreira,
2003; Sturt, 2007). Here, we test whether cognitive control engagement
can facilitate such resolution.
Semantic constraints are reliable evidential cues to Agent- and
Patienthood that can rapidly direct comprehenders’ moment-by-mo-
ment interpretations. Consider 1a and 1b:
1a: The defendant examined by the lawyer turned out to be un-
reliable.
1b: The evidence examined by the lawyer turned out to be unreli-
able.
Both sentences have passive structures: The first noun (defendant,
evidence) is the Patient of the transitive verb “examined” rather than its
Agent, and the second noun (lawyer) is the Agent. But up until the word

⁎
Corresponding authors at: Department of Speech, Language and Hearing Sciences, The George Washington University, Washington, DC 20052, United States (M. Thothathiri).
Department of Hearing and Speech Sciences, University of Maryland, College Park, MD 20742, United States (J. Novick).
E-mail addresses: malathi@gwu.edu (M. Thothathiri), jnovick1@umd.edu (J.M. Novick).

https://doi.org/10.1016/j.cognition.2018.05.014
Received 19 May 2016; Received in revised form 10 May 2018; Accepted 21 May 2018
0010-0277/ © 2018 Elsevier B.V. All rights reserved.
M. Thothathiri et al.
“by”, the input is syntactically ambiguous, consistent with both the
active and passive constructions. However, a key difference between 1a
and 1b is the thematic fit of defendants and evidence as probable per-
petrators of the described action: Semantically, “defendant” is a plau-
sible Agent (it is animate) whereas “evidence” is not. Several studies
have shown that upon encountering “by the lawyer”, processing is
slower in sentences like 1a than 1b, suggesting that readers experience
temporary conflict in the case of “The defendant…” because the se-
mantic cue (animacy) competes with the syntax. By contrast, in 1b,
“evidence” is inanimate and thus an unlikely Agent but liable Patient.
Here, the semantic cue cooperates with the syntax, allowing processing
to converge earlier on the Patient interpretation (e.g., Pearlmutter &
MacDonald, 1992; Trueswell, Tanenhaus, & Garnsey, 1994). While the
relative weighting of semantic and morphosyntactic cues for compre-
hension might vary among languages (Bates & MacWhinney, 1989),
current evidence suggests that animacy can be a relevant cue for sen-
tence interpretation even in morphologically rich languages that are
different from English (e.g., Russian: Stoops, Luke, & Christianson,
2014; Turkish: Demiral, Schlesewsky, & Bornkessel-Schlesewsky,
2008).
Beyond animacy, a number of studies have shown that semantic
plausibility and world knowledge about events can affect, even shape,
online as well as offline interpretation (Altmann & Kamide, 2007;
Christianson, Hollingworth, Halliwell, & Ferreira, 2001; Christianson,
Luke, & Ferreira, 2010; Ferreira, 2003; Garnsey, Pearlmutter, Myers, &
Lotocky, 1997; Kamide, Altmann, & Haywood, 2003; Kamide,
Scheepers, & Altmann, 2003; Kim & Osterhout, 2005; Knoeferle,
Crocker, Scheepers, & Pickering, 2005; Malyutina & den Ouden, 2015;
Slattery, Sturt, Christianson, Yoshida, & Ferreira, 2013; Sturt, 2007).
For example, non-canonical sentences like passives are prone to mis-
interpretation especially when they describe implausible scenarios
(e.g., The dog was bitten by the man). Listeners not infrequently interpret
such sentences as their plausible counterparts (e.g., that the dog, and
not the man, was doing the biting) even though the syntax is un-
ambiguous (Christianson et al., 2010; Ferreira, 2003). Similarly,
garden-path sentences (e.g., While the man hunted the deer ran into the
woods) can leave behind lingering incorrect interpretations in listeners
(e.g., the deer as the Patient of the hunting action), particularly when
those interpretations are plausible (e.g., compare the sentence above
with: While the man hunted the deer paced in the zoo) (Christianson et al.,
2001; Malyutina & den Ouden, 2015; Slattery et al., 2013; Sturt, 2007).
These semantically conditioned interpretations can persist and compete
with the correct interpretation even after the syntax is disambiguated
(Sturt, 2007), or indeed even after the sentence appears to have been
syntactically reanalyzed (Slattery et al., 2013).
Convergent evidence from electrophysiological (ERP) studies also
indicates that semantic cues can lead to interpretations that compete
with the one licensed by the syntax. For example, syntactically well-
formed but semantically anomalous sentence fragments like “The hearty
meal was devouring…” can elicit a P600 rather than an N400 signal
despite unambiguous syntactic cues (Kim & Osterhout, 2005). This
pattern is consistent with the parser relying on semantic information for
thematic role assignment (meal as Patient of devouring), which triggers
syntactic revision (to “The hearty meal was devoured…”) in order to
satisfy the semantic fit (Kim & Osterhout, 2005; Kim & Sikos, 2011; see
also Kuperberg, 2007; van Herten, Chwilla, & Kolk, 2006; cf. Chow &
Phillips, 2013).
More broadly, several eye-tracking studies have shown that listeners
integrate semantic information from the incoming input with prior or
context-specific knowledge about events to rapidly assign and even
anticipate thematic roles (see e.g., Kamide, Altmann, et al., 2003;
Kamide, Scheepers, et al., 2003; Knoeferle et al., 2005). For example, in
a study by Kamide and colleagues, listeners fixated on a motorcycle in
the scene more than a carousel when they heard “The man will ride…”;
and more on the carousel when they heard “The girl will ride…”, in-
dicating that they integrated the meanings of the first noun and the verb
163
Cognition 178 (2018) 162–177
to anticipate upcoming referents (Kamide, Altmann, et al., 2003).
To summarize, sentence interpretation is guided not only by syn-
tactic analysis of the input, but also by potentially independent se-
mantic analysis (e.g., Christianson et al., 2010; Kim & Osterhout, 2005).
While these sources of information frequently lead to the same inter-
pretation, sometimes they point towards incompatible representations
of sentence meaning. What cognitive mechanics allow listeners to re-
solve competition in such cases?
1.1. Cognitive control and resolving who did what to whom
Here, we test whether cognitive control engagement facilitates lis-
teners’ ability to arrive at the correct thematic role assignment despite
competing cues. For instance, upon hearing “The fox was chas (phonetic
transcription: /tʃeıs/)…”, listeners might initially consider both the
Agent and Patient interpretations (fox as chaser and fox as fleer). When
later-arriving syntactic information (…ed by the rabbit) removes the
temporary syntactic ambiguity and clearly indicates that the fox is the
fleer (the Patient), not the pursuer (the Agent), the parser might
nevertheless continue to experience competition from the strong se-
mantic cue because foxes typically chase rabbits, not vice versa.
Resolution of such competition might rely on general cognitive control
mechanisms that are used to detect and resolve information-conflict
across domains, including recognition memory and language processing
(Novick, Trueswell, & Thompson-Schill, 2005; see also Kaan & Swaab,
2002).
Convergent evidence from behavioral, neuroimaging, and neu-
ropsychological studies broadly supports a link between cognitive
control and language (Hsu & Novick, 2016; Hussey et al., 2017; Hussey,
Ward, Christianson, & Kramer, 2015; January, Trueswell, & Thompson-
Schill, 2009; Robinson, Shallice, & Cipolotti, 2005; Schnur et al., 2009;
Snyder, Banich, & Munakata, 2011; Snyder, Banich, & Munakata, 2014;
Thothathiri, Kim, Trueswell, & Thompson-Schill, 2012; Vuong &
Martin, 2011, 2014, 2015; Ye & Zhou, 2009; inter alia). For example,
overlapping brain regions in the left ventrolateral prefrontal cortex
(hereafter left VLPFC) are activated within subjects by both canonical
cognitive control tasks (Stroop, Flanker) and language processing tasks
with competition between incompatible sentence interpretations (e.g.,
Hsu, Jaeggi, & Novick, 2017; January et al., 2009; van de Meerendonk,
Rueschemeyer, & Kolk, 2013; Ye & Zhou, 2009). Damage to the same
regions results in language deficits that are best characterized as an
impairment in resolving competition at multiple levels of representa-
tion (e.g., Novick, Kan, Trueswell, & Thompson-Schill, 2009; Robinson
et al., 2005; Schnur et al., 2009; Vuong & Martin, 2011). On the flip
side, behavioral training that targets VLPFC-supported cognitive con-
trol functions, and non-invasive stimulation of this brain region yield
improvements in the ability to correctly re-interpret garden-path sen-
tences, suggesting a cause-and-effect interplay between cognitive con-
trol and syntactic ambiguity resolution (e.g., Hussey et al., 2015, 2017;
Novick, Hussey, Teubner-Rhodes, Harbison, & Bunting, 2014). With
respect to conflict between semantic and syntactic cues in particular,
previous neuroimaging studies have shown that the left VLPFC, which
is implicated in cognitive control, is activated more during the com-
prehension of sentences containing syntax-semantics conflict like the
fox example above, compared to sentences without such conflict
(Thothathiri et al., 2012; Ye & Zhou, 2009). This provides tantalizing
correlational evidence by way of suggesting that the regions involved in
cognitive control might also be involved in resolving conflict between
thematic role assignments. However, it remains unknown if cognitive
control processes engaged by a non-sentential task can cause listeners to
suppress semantically attractive but erroneous thematic relations and
choose a less plausible but correct interpretation during sentence
comprehension.
We addressed this issue by relying on a phenomenon known as
“conflict adaptation”—where conflict detection triggers sustained cog-
nitive control, which facilitates the resolution of subsequent conflict
M. Thothathiri et al.
(Gratton, Coles, & Donchin, 1992). For example, in a Stroop task,
conflict experienced during a previous trial (e.g., “blue” in green ink)
facilitates processing of an ensuing conflict trial, compared to when the
previous trial did not contain conflict (“blue” in blue ink) (Freitas,
Bahar, Yang, & Banai, 2007; Kerns et al., 2004). Under the influential
Conflict-Monitoring Theory proposed by Botvinick and colleagues
(Botvinick, Braver, Barch, Carter, & Cohen, 2001), this effect arises
because conflict on a previous trial is detected by monitoring me-
chanisms in the anterior cingulate cortex (ACC), which signals regions
within lateral prefrontal cortex about the need for conflict resolution
(Kerns et al., 2004). The latter then implements control and helps re-
solve competition between representations by biasing activation up-
ward in regions processing task-relevant information, downward in
those processing task-irrelevant information, or both (Egner & Hirsch,
2005; King, Korb, von Cramon, & Ullsperger, 2010). Performance on a
subsequent trial containing conflict is facilitated because the cognitive
control system is already engaged and can more effectively and effi-
ciently resolve new instances of conflict. Thus, the conflict adaptation
effect measures the quantifiable behavioral savings that result from
online adjustments in cognitive control.
We tested whether the mobilization of cognitive control would fa-
cilitate listeners’ subsequent resolution of syntax-semantics conflict for
thematic role assignment. An earlier finding motivated our approach
and design. Specifically, Hsu and Novick (2016) extended the conflict
adaptation paradigm to examine how cognitive control influences
syntactic revision during language processing, by interleaving Stroop-
Congruent and -Incongruent trials (which manipulated dynamic cog-
nitive control engagement) with a spoken comprehension task invol-
ving a prepositional-phrase attachment ambiguity that was ripe for
misinterpretation (e.g., “Put the frog on the napkin onto the box”,
where listeners initially commit to the Goal interpretation of “on the
napkin” but must quickly revise following conflicting evidence from
“onto the box”. See also Novick, Thompson-Schill, & Trueswell, 2008;
Spivey, Tanenhaus, Eberhard, & Sedivy, 2002; Tanenhaus, Spivey-
Knowlton, Eberhard, & Sedivy, 1995; Trueswell, Sekerina, Hill, &
Logrip, 1999). When ambiguous ‘Put’ sentences followed incongruent
Stroop items, compared with congruent Stroop items, listeners’ eye
movements to objects in the visual world reflected greater considera-
tion of the correct goal (e.g., the box) that emerged earlier in time than
when cognitive control was relatively unengaged. Moreover, offline
hand-action errors involving the incorrect goal (e.g., an empty napkin)
decreased reliably, suggesting that cognitive control engagement fa-
cilitates syntactic revision and may even prevent comprehension errors
in healthy adults.
In the experiments reported below, we extend the investigation of
conflict adaptation to parsing conditions that create a different type of
conflict-processing demand that routinely generates comprehension
errors, namely, conflict between semantic and syntactic cues that gen-
erate divergent interpretations. Participants completed interleaved
Stroop and spoken sentence-comprehension tasks akin to the Hsu and
Novick (2016) design. The critical sentences were passive structures,
which are known to create syntax-semantics conflict when the expected
thematic roles are reversed (Ferreira, 2003). In two experiments, we
monitored listeners’ eye movements to pictures that depicted different
events, as a measure of ongoing sentence interpretation. Specifically,
we evaluated whether looks to pictures depicting the correct and in-
correct thematic role assignments varied systematically depending on
prior Stroop-trial type (incongruent vs. congruent), and hence the en-
gagement status of cognitive control.
The critical prediction is that a reliable processing benefit will
emerge for conflict sentences that follow incongruent Stroop items as
compared to congruent ones. This impact of prior cognitive control
engagement on online processing could manifest in (at least) one of two
ways. First, there may be magnitude effects, in which overall fixation
proportions to key pictures in the scene differ within critical time in-
tervals. For instance, when cognitive control is activated by a previous
164
Cognition 178 (2018) 162–177
Stroop trial versus not, there could be more looks to the implausible but
correct rabbit-chasing-fox scene, and/or correspondingly fewer looks to
the plausible but incorrect fox-chasing-rabbit scene. This would suggest
that cognitive control deployment biases interpretation towards the
correct thematic role assignment and away from the incorrect one.
Second, there may be rate effects (Barr, 2008), in which fixation pro-
portions to the critical pictures change differently over time depending
on prior cognitive control engagement. Thus, steeper growth in fixa-
tions to the correct picture or a steeper decrease in fixations to the
incorrect picture, irrespective of magnitude, would indicate that cog-
nitive control engagement accelerates convergence on the correct ana-
lysis under conditions of conflict.
The two types of analyses—magnitude and rate—can provide
complementary information. Specifically, the rate analysis can separate
effects that may begin prior to the arrival of the critical linguistic in-
formation from those that occur during the processing of that in-
formation. For example, if two experimental conditions differ, for
whatever reason, in their pattern of looks prior to the arrival of the
disambiguating information (e.g., “..ed by the rabbit”), the magnitude
analysis will reflect an ongoing summation of those prior differences
and any differences that emerge during the analysis window. In con-
trast, the rate analysis can determine if looks changed more strongly
(i.e., there was a steeper slope) in one condition than the other, which
would reflect the effect of processing the disambiguating information
independently from prior differences. Such separation could be espe-
cially important if meaningful processing differences in a critical in-
terval are masked by counteracting patterns leading up to that interval
(please see Barr (2008) for a detailed discussion with examples). The
mixed-effects models that we employ allow for testing both kinds of
effects, thereby providing a comprehensive way to evaluate the causal
interplay between cognitive control engagement and real-time lan-
guage comprehension.
2. Experiment 1
2.1. Method
2.1.1. Participants
Twenty-four right-handed native English speakers with normal or
corrected-to normal vision participated in the study (3 male; age
range = 18–25; Mage = 18.9). They provided consent under a protocol
approved by the institutional review board at The George Washington
University. Participants received compensation or course credit.
2.1.2. Materials and procedure
Participants completed two tasks: sentence comprehension and
Stroop.
Sentence comprehension. In the sentence comprehension task,
participants heard active and passive sentences and picked the
matching picture from four options. A female native English speaker
pre-recorded the sentences. Critical sentences were passive construc-
tions that were categorized as either congruent or incongruent.
Congruent sentences described a more plausible situation, i.e., the
Agent and Patient filled the expected thematic roles (e.g., The rabbit was
chased by the fox). In contrast, incongruent sentences described a rela-
tively implausible situation, i.e., the Agent and Patient violated their
expected roles in the event (e.g., The fox was chased by the rabbit). In
both conditions, the visual display contained the Target picture that
matched the sentence, the Reversal picture that depicted the reversed
scenario, and two unrelated distractors (Fig. 1). Non-critical filler trials
contained active sentences that accompanied a target picture and three
distractors. The distractors overlapped with one another and with the
target to varying degrees (no common elements, one common entity
(noun), two common entities (both nouns), the same action (verb), or a
combination of an entity and an action (noun + verb)). This variability
ensured that across all sentences, participants would be unlikely to
M. Thothathiri et al.
develop a consistent visual strategy (e.g., identifying pairs of relevant
pictures) prior to hearing the sentence. That is, participants had to both
semantically analyze the pictures and listen to the entire sentence in
order to achieve correct performance throughout. Paralleling the cri-
tical sentences, some filler sentences described likely scenarios (e.g.,
The horse carried the jockey) and others described unlikely scenarios
(e.g., The ghost painted the boy). The full set of verbal and visual stimuli
is listed in Appendix A.
Target pictures appeared equally often in one of four quadrants.
Trials began with a central fixation cross (500 ms), which was followed
by the pictures and the auditory sentence. The visual display remained
on screen for 5000 ms or until the participant responded, after which a
blank screen appeared (1000 ms). Participants pressed a number-pad
key (7, 9, 1 or 3) to select a picture.
Stroop. In the Stroop task, participants viewed words on a computer
screen and indicated the font colors of the words. Akin to the sentence
comprehension task, there were congruent and incongruent conditions
(Fig. 1). The congruent condition contained words whose meaning
matched the font color (e.g., “blue” in blue ink) whereas the incon-
gruent condition contained words whose meaning mismatched the font
color (e.g., “red” in blue ink). Six color words (“red”, “orange”,
“yellow”, “green”, “blue”, “brown”) were presented in three possible
colors (blue, green, yellow). On incongruent trials, the word stimulus
(e.g., “red”) was not a possible response. Thus, there was no conflict at
the level of responses (key presses) because there was no key that
corresponded to the word’s meaning. Instead, the trials involved con-
flict between two cognitive representations (one representation corre-
sponding to the color and one corresponding to the word meaning)
(Milham et al., 2001), similar to the conflict between two representa-
tions (i.e., interpretations) experienced during the sentence compre-
hension task (Hsu & Novick, 2016). This similarity in representational
conflict across tasks allowed us to investigate conflict adaptation from
one task to another. Each stimulus word appeared 20 times and each
165
Cognition 178 (2018) 162–177
Fig. 1. Example Stroop and sentence com-
prehension sequence. Critical Stroop-
Sentence pairs consisted of a Stroop trial
followed by a sentence trial. On sentence
comprehension trials, participants heard a
sentence and selected a picture from four
options (Target, Reversal, and two unrelated
distractors). Target picture location was
counterbalanced across trials. Note. In this
example, the picture in the lower-right
quadrant is the Target for the Congruent
sentence, but the Reversal for the
Incongruent Sentence.
target response 40 times. Trials began with a central fixation cross
(500 ms), which was followed by the color word (1000 ms or until the
participant responded), and then a blank screen (1000 ms). Participants
pressed a color-coded number-pad key (4, 5 or 6) to make a response.
Interleaved Stroop and sentence trials. Each participant completed
120 Stroop trials (60 congruent (C), 60 incongruent (I)), 40 critical
sentence trials (20 congruent (C), 20 incongruent (I)), and 40 active-
construction filler sentences that prevented a general expectation for
passives. The different trial types were pseudo-randomly interleaved.
Because we were interested in examining the effect of the previous
Stroop trial type on the subsequent sentence trial, the order of trials
included 40 critical Stroop-to-Sentence pairs (Fig. 1). The 40 pairs
comprised 10 pairs each of the following: congruent Stroop followed by
congruent sentence (CC), congruent Stroop followed by incongruent
sentence (CI), incongruent Stroop followed by congruent sentence (IC),
and incongruent Stroop followed by incongruent sentence (II). Thus,
congruent or incongruent Stroop trials were followed by congruent or
incongruent sentences equally often. Additionally, the inclusion of 40
filler sentence trials and 80 additional Stroop trials resulted in several
Stroop-to-Stroop, Sentence-to-Stroop, Stroop-to-Sentence, and Sen-
tence-to-Sentence pairs, which prevented participants from predicting
upcoming task type.
We created four lists containing a pseudorandom order of trial
types. Each participant was randomly assigned to a list. Assignment of
critical sentences to trial positions was pseudorandomized: the same
sentence appeared in different positions across lists. Any given sen-
tence, which was either congruent or incongruent, appeared after a
congruent Stroop trial in two lists and after an incongruent Stroop trial
in two other lists. Sentences and their reversals appeared in all four
possible condition combinations across lists ([CC, CI], [CC, II], [IC, CI],
[IC, II]). Reversing the order of trials in the first four lists created four
additional lists. Each participant heard sentences describing plausible
events and their reversals, which therefore described implausible
M. Thothathiri et al.
events. The two versions within a list were separated by an average of
65 trials. For each participant, the congruent version appeared before
the incongruent version for approximately half the time (9–11 times out
of 20).
We monitored eye movements using an Eyelink 1000 eye-tracker
with high spatial resolution (≤1.5°) and a sampling rate of 1000 Hz.
Practice and Stroop baseline. Before starting the experimental por-
tion of the study (i.e., the interleaved Stroop-to-sentence sequences),
participants completed 10 practice Stroop trials, which they repeated if
necessary until reaching greater than 80% accuracy. Subsequently, they
completed a baseline Stroop block (73 congruent, 72 incongruent),
which was used in an exploratory analysis to measure individual dif-
ferences (not reported here). Following the Stroop baseline, partici-
pants practiced 10 sentence comprehension trials (repeated if necessary
until reaching greater than 80% accuracy) and then 27 interleaved
Stroop and sentence trials (17 Stroop, 10 sentence). Only active con-
structions were used during practice, and none of these items appeared
in the actual experiment. Practice was intended to familiarize the par-
ticipants with the overall tasks as well as the keyboard mappings, which
were different for the two tasks (see above). Eight participants repeated
the Stroop practice, two repeated the sentence practice, and one par-
ticipant repeated both practice tasks (in all cases, only once).
2.2. Norming study
To validate our thematic-role plausibility manipulation, we con-
ducted a norming study on the critical experimental stimuli. Twenty-
three native English speakers (7 male; age range = 18–40; Mage = 25),
none of whom participated in the eye-tracking study, completed an
online survey on SurveyMonkey (http://www.surveymonkey.com).
Survey responses revealed two items for which a sentence and its re-
versed version received similar plausibility ratings (matador killed by
bull versus bull killed by matador, and dragon killed by prince versus
prince killed by dragon). Because we could not categorize these sce-
narios as either plausible or implausible, we removed them from ana-
lysis. For the remaining 18 items and their reversed versions, we con-
firmed—across three different measures—that the stimuli we
categorized as incongruent described less plausible scenarios and
should therefore generate more syntax-semantics conflict than their
corresponding congruent versions (see Appendix B for details).
2.3. Dependent measures and analysis
We analyzed Stroop accuracy and reaction time (RT). For sentence
comprehension, our primary analyses focused on online sentence pro-
cessing, as indicated by eye movements. We also report offline accuracy
and RT, with the caveat that these offline measures could reflect post-
interpretive processes such as picture selection and response verifica-
tion above and beyond the sentence interpretation processes of interest.
RT was computed relative to sentence offset.
We analyzed accuracy measures using mixed effects logistic re-
gression (glmer function with the bobyqa optimizer), and RT measures
using mixed effects linear regression (lmer function) (lme4 ver-
sion1.1.12 and lmerTest version 2.0.30 in R). For RTs, we excluded
incorrect trials and trials with values more than 2 standard deviations
from each participant’s mean. Eye-movement data are categorical, re-
quiring logistic regression. Further, these data are not temporally in-
dependent—for physiological reasons, eye movements at a given time
point are likely to constrain movements at a subsequent time point. One
way to tackle these issues is to group observations into temporal bins,
average across trials, and compute the empirical-logit for each bin. This
allows for examination of magnitude of looks and their rate of change
over time without conflating Type I error rate due to temporal depen-
dence (Barr, 2008). However, if each participant experiences each
sentence only once, as is the case here, this averaging approach pre-
cludes crossed participant and item effects (Barr, 2008). Therefore, we
166
Cognition 178 (2018) 162–177
report results below using 100-ms time bins and separate by-partici-
pants (averaged across items) and by-items (averaged across partici-
pants) analyses. We only interpret effects that are significant (p < .05)
or marginally significant (p < .1) by participants and items.
For the eye movement analyses, we organize results into the fol-
lowing two sections.
A. Overall time-course of sentence interpretation: Evidence for compe-
tition. First, we analyzed how interpretation unfolded over time across
all critical sentences and conditions, to get a sense for whether listeners
considered both competing interpretations early on (suggesting tem-
porary ambiguity) before settling on the Target. We computed a Target
Advantage score (proportion of Target looks minus proportion of
Reversal looks) that we expected to hover around zero and remain flat
while the sentence was still ambiguous (indicating no preference for
Target over Reversal), and then grow over time after the sentence was
disambiguated by morphological cues (indicating a Target preference).
Accordingly, our analyses focused on three relevant intervals. The first
interval (hereafter, “pre-disambiguation”) extended from 0 to 400 ms
after verb-onset. In this interval, incoming auditory information (e.g.,
chas… following The fox was) was consistent with both Target and
Reversal interpretations, i.e., both pictures were possible matches
(mean verb duration = 428 ms). Therefore, looks to both picture-
s—reflecting consideration of the fox as a possible chaser or a possible
fleer, for example—should increase with time and the Target Advantage
score was expected to be around zero with relatively little change until
disambiguation. The second interval (hereafter, “disambiguation”)
comprised the time period 400–800-ms from verb-onset. In this in-
terval, incoming auditory information disambiguated the syntactic
structure (e.g., “…-ed by…”). It unambiguously indicated that the
subject is the Patient of the action. Thus, in this region, we expected the
Target Advantage score to begin increasing over time. Finally, the third
interval (hereafter, “post-disambiguation”) extended between 800 and
1200 ms from verb-onset and was expected to show a significantly
positive Target Advantage (> 0) and maybe some continuing growth
over time, indicating that listeners had either converged or were con-
verging upon the correct interpretation (mean Noun2 onset = 768 ms).
We analyzed the Target Advantage score using a mixed-effects model
containing the intercept, the fixed effect of time-bin, and random in-
tercept and slope by participant/item. The time-bin factor was coded
using orthogonal polynomial coding to examine whether Target
Advantage increased or decreased with time.
B. Analyses of Target and Reversal looks: Evidence for conflict adap-
tation. The analysis in section A was expected to provide a global index
of competition between clashing sentence interpretations and the time
course of convergence towards the correct analysis, across all condi-
tions. Next, we analyzed whether prior cognitive control engagement
mitigated such competition in the form of conflict adaptation (leading
to differences between conditions). Conflict adaptation could operate
via facilitated processing of the relevant cue (here, unambiguous
syntax), suppression of the irrelevant cue (plausible but misleading
semantics), or a combination of the two. Therefore, we analyzed looks
to the Target and Reversal pictures separately to determine whether the
critical experimental conditions differed in these measures in each of
the three critical time windows (pre-disambiguation, disambiguation,
and post-disambiguation). The specific prediction for conflict adapta-
tion is as follows: For incongruent sentences (e.g., The fox was chased by
the rabbit), strong semantic attraction—e.g., plausibility of fox as Agent,
not Patient of chasing—could continue to create competition between
the Target and the Reversal interpretations even after the syntax is
disambiguated. This competition might be mitigated if previous Stroop-
conflict, as opposed to non-conflict trials, engaged cognitive control and
led to sustained, cross-task adjustments that facilitated resolution of
syntax-semantics conflict.
We conducted full mixed model analyses containing all pertinent
effects, including time bin, previous Stroop trial type, current sentence
trial type, their interactions, and random intercepts and slopes. In this
M. Thothathiri et al.
Fig. 2. Experiment 1. Looks to Target and Reversal, and the Target Advantage
score (inset) after verb-onset, collapsed across all conditions. Error bars are
95% confidence intervals. Time-bins here and elsewhere are offset by 200 ms.
analysis, a conflict adaptation effect in magnitude would manifest as a
previous-Stroop-by-current-Sentence-type interaction: Target looks
should be higher and/or Reversal looks lower for II than CI pairs, but
not for IC than CC pairs. Such a pattern would indicate that the effect of
cognitive control engagement from the previous Stroop trial was spe-
cific to conflict resolution, thereby affecting incongruent sentences
(which contain syntax-semantics conflict) and not congruent ones
(which contain no conflict). Analogously, a conflict adaptation effect in
rate would also affect incongruent but not congruent sentences.
However, it would manifest as a three-way interaction between previous
Stroop trial type, current Sentence trial type, and time bin: Target looks
should grow and/or Reversal looks should decrease more rapidly for II
than CI pairs, but not for IC than CC pairs. For both magnitude and rate
effects, we predicted that the interaction would emerge after the syntax
became unambiguous, either in the disambiguation or the post-dis-
ambiguation interval depending on how long the competition lasts (as
indicated by the results from section A). To help interpret any sig-
nificant interactions, we have included figures that show the eye
movement pattern in each critical condition. We also report estimated
marginal means for the relevant follow-up contrasts (emmeans version
1.1.3 in R). In all cases, we analyzed eye movements only for correct
sentences following correct Stroop trials. Thus, we can be confident that
any observed differences between conditions do not reflect tradeoffs in
accuracy.
2.4. Results
2.4.1. Stroop accuracy and RT
Mean accuracy on the Stroop trials interleaved with the sentence
comprehension trials was 94.10%. Accuracy did not differ significantly
between congruent and incongruent trials (p > .1). For RTs, we ex-
cluded trials with outlier values or incorrect responses (10.2%).
Analyses of the remaining trials showed that the RTs were reliably
slower in the incongruent than congruent condition (fixed effect of
167
Cognition 178 (2018) 162–177
congruency: estimate = 36.90, S.E. = 6.77, df = 69.71, t = 5.45,
p < .001. MRT for C trials = 619 ms, I trials = 657 ms). This indicated
that participants experienced conflict during the incongruent Stroop
trials, as expected.
2.4.2. Sentence comprehension accuracy and RT
Mean accuracy on the critical sentence comprehension trials was
92.92%, which is comparable to accuracy rates for similar sentences in
other paradigms (Thothathiri et al., 2012). A mixed model containing
fixed effects of previous Stroop trial type, current sentence trial type
and their interaction, and random intercept and slopes by participant
revealed no significant effects (ps > .1). Thus, listeners eventually ar-
rived at the correct interpretation (thematic role assignments) equally
across conditions. RT analysis (time to click on the correct picture) after
excluding outliers, incorrect trials, and trials following incorrect Stroop
responses (15.8%) revealed a significant effect of the previous trial type
(estimate = −112.16, S.E. = 52.78, df = 62.99, t = −2.13, p < .05)
and a marginal interaction between previous and current trial types
(estimate = 138.59, S.E. = 74.16, df = 51.35, t = 1.87, p = .07). Time
to choose the picture after sentence offset was quicker when the sen-
tences followed incongruent Stroop trials (MRT = 742 ms) than when
they followed congruent Stroop trials (MRT = 792 ms). However, this
effect was driven primarily by congruent (CC = 810 ms, IC = 678 ms,
p < .05) rather than incongruent sentences (CI = 773 ms, II = 812 ms,
p > .5).
2.4.3. Eye movements
We excluded trials that contained incorrect responses and/or fol-
lowed Stroop trials with incorrect responses (11.9%), and those that
contained track loss for more than a third of a given analysis interval
(0.8%).
2.4.3.1. A. Overall time-course of sentence interpretation: Evidence for
competition. Fig. 2 shows looks to the Target and Reversal pictures, and
the Target Advantage score (inset) collapsed across all conditions. In
the pre-disambiguation interval, Target Advantage was not significantly
different from zero (participants and items analyses ps > .4), and did
not change over time, as expected (ps > .3). In the disambiguation
interval, the score was not different from zero (ps > .6), but there was
linear growth over time that was significant in the participants analysis
(t(45.30) = 3.03, p < .01) and marginally significant in the items
analysis (t(36.24) = 1.95, p = .06). In the post-disambiguation
interval, the intercept was significantly above zero (Participants: t
(23) = 4.90, p < .001; Items: t(35.01) = 4.77, p < .001).
Additionally, there was a significant linear increase over time in the
participants analysis (t(79.60) = 2.70, p < .01) and a non-significant
effect in the items analysis (t(42.23) = 1.58, p > .1). These results
accord broadly with our predictions. They show that participants did
not have a preference for either the Active or the Passive interpretation
prior to disambiguation (and thus experienced competition), began
converging on the correct Passive interpretation in the disambiguating
interval, and arrived at the correct interpretation post-disambiguation.
2.4.3.2. B. Analyses of Target and Reversal looks: Evidence for conflict
adaptation. Here, we tested the extent to which prior cognitive control
engagement modulated the resolution of conflict between
interpretations. Fig. 3 shows the evolution of looks to the Target and
Reversal pictures, separated by each condition. The results in section A
indicated that convergence to the correct interpretation began in the
disambiguation interval. Therefore, we expected conflict adaptation
effects to primarily manifest in the disambiguation interval, and
possibly continue in the post-disambiguation interval.
In the pre-disambiguation interval, there was a significant increase
in Target looks over time (Participants: t(276) = 3.62, p < .001; Items:
t(34.13) = 3.13, p < .01) and a significant interaction between time
and previous trial type (Participants: t(276) = 2.98, p < .01; Items: t
M. Thothathiri et al.
Fig. 3. Experiment 1. Left: Proportion of Target looks in different conditions after verb-onset. Right: Proportion of Reversal looks in different conditions after verb-
onset. Error bars are within-subject 95% confidence intervals adjusted according to Morey (2008).
(170) = 2.30, p < .05). The interaction was due to reliable growth in
Target looks over time after an incongruent Stroop trial (Participants: t
(276) = 4.67, p < .001; Items: t(85.99) = 3.82, p < .001) but not a
congruent Stroop trial (Participants: t(276) = .46, p > .6; Items: t
(85.99) = 1.83, p = .07). For Reversal looks, the pre-disambiguation
interval revealed significant growth over time (Participants: t
(24.35) = 4.22. p < .001; Items: t(34.07) = 4.12, p < .001) and no
other effects. Thus, both Target and Reversal looks grew over time prior
to disambiguation, consistent with competition.
In the disambiguation interval, Target looks significantly increased
over time (Participants: t(23.63) = 2.78, p < .05; Items: t
(36.89) = 2.28, p < .05). There were no other effects. Crucially, for
Reversal looks, there was a significant previous-by-current-trial-type
interaction in the magnitude of looks that was consistent with conflict
adaptation (Participants: t(23) = −3.26, p < .01; Items: t
(34.01) = −2.43, p < .05). The proportion of Reversal looks was
significantly lower in the II compared to the CI condition (Participants:
t(26.34) = −3.28, p < .01. Items: t(54.54) = −3.46, p < .01), and
not significantly different between the IC and CC conditions
(Participants: t(29.69) = 1.13, p > .2. Items: t(54.54) = −.03,
p > .9). Additionally, there was an interaction between the linear ef-
fect of time-bin and previous trial type (Participants: t(276) = −3.34,
p < .001; Items: t(170) = −2.97, p < .01). This resulted from a re-
liable decrease in Reversal looks over time after incongruent Stroop
trials (Participants: t(276) = −5.29, p < .001; Items: t Specifically, Reversal looks were reliably lower on II than CI pairings,
(73.21) = −2.23, p < .05) but not after congruent Stroop trials
(Participants: t(276) = −.58, p > .5; Items: t(73.21) = .45, p > .6).
In sum, the disambiguation interval showed evidence for modulation by
the previous Stroop trial in Reversal but not Target looks. Specifically,
Reversal looks showed both magnitude effects that were consistent with
conflict adaptation and rate effects that suggested a broader non-con-
flict-specific influence of prior cognitive control engagement.
In the post-disambiguation interval, for Target looks, there was a
two-way interaction between time and previous trial type (Participants:
t(253) = −2.78, p < .01; Items: t(170) = −2.75, p < .01) and a
168
Cognition 178 (2018) 162–177
three-way interaction between time, previous trial type, and current
trial type (Participants: t(253) = −4.21, p < .001; Items: t
(170) = −4.17, p < .001). Splitting by each condition, there was no
reliable effect of time in the CC and IC conditions, a linear increase in
the CI condition (Participants: t(276) = 2.79, p < .01; Items: t
(204) = 3.77, p < .001), and a linear decrease in the II condition
(Participants: t(276) = −3.46, p < .001; Items: t(204) = −2.73,
p < .01).1 For Reversal looks, there was a significant three-way time
by previous trial type by current trial type interaction (Participants: t
(276) = 2.62, p < .01; Items: t(170) = 2.35, p < .05). Splitting by
each condition, there was no reliable change over time in the CC, IC and
II conditions. In the CI condition, by contrast, Reversal looks continued
to decrease over time (Participants: t(276) = −3.37, p < .001; Items: t
(68.20) = −2.58, p < .05). Overall, these patterns suggest that the
correct interpretation of incongruent sentences was delayed in the CI
but not the II condition. Unlike for congruent sentences (CC and IC), the
CI condition showed evidence of still-evolving interpretation (de-
creasing Reversal and increasing Target looks) post-disambiguation. By
contrast, the pattern for the II condition suggested that by this time
listeners had already resolved the syntax-semantics conflict and had
inhibited looks to Reversal images to a level equaling that of congruent
sentences (Fig. 3, right panel).
To summarize, a conflict adaptation effect emerged in the magni-
tude of looks to the Reversal picture in the disambiguation interval.
and not different between IC and CC pairings, suggesting that cognitive
control engagement enabled the suppression of a semantically guided
interpretation that was incompatible with syntactic cues. In the post-
disambiguation interval, the CI condition revealed a different pattern
than the other three conditions, consistent with the idea that correct
1
The linear decrease in the II condition likely reflects listeners moving away from the
screen to prepare to a keyboard response, having already arrived at the correct inter-
pretation (as indicated by the results above in the disambiguation interval).
M. Thothathiri et al.
interpretation peaked later in this condition. In contrast, the II condi-
tion did not show such delays, in accordance with conflict adaptation.
Two other results suggested a broader influence of cognitive control
engagement independent of syntax-semantics conflict. Target looks
increased over time more after incongruent than congruent Stroop trials
in the pre-disambiguation interval. Reversal effects showed a parallel
effect in the disambiguation interval (decreasing more after incon-
gruent than congruent Stroop trials). We discuss this further below.
2.5. Discussion
Figuring out who did what to whom as linguistic input unfolds is
anything but trivial, as speakers can choose to convey the same message
using a variety of structures (e.g., actives vs. passives), creating tem-
porary ambiguity in thematic role assignment. As well, subject noun
phrases can often be consistent with either role (foxes can chase and be
chased), but real-world knowledge can bias listeners’ interpretations in
particular ways (foxes are more likely to chase rabbits than vice versa).
When unambiguous syntactic cues conflict with semantic biases (e.g.,
“The fox was chased by the rabbit”), interpretation is prone to error
(Ferreira, 2003). We hypothesized that cognitive control mechanisms,
which broadly resolve conflict, may assist comprehension and prevent
communication from running aground in such cases.
The results from Experiment 1 suggest that conflicting semantic
cues can disrupt sentence interpretation but that the disruption can be
attenuated by prior cognitive control engagement. In particular, our
analyses revealed that listeners suppressed the plausible-but-incorrect
reversal interpretation earlier in the II versus the CI condition. The
magnitude of Reversal looks was significantly lower in the II than the CI
condition in the disambiguation interval. Interpretation appeared to be
delayed in the CI condition, slowly beginning to reduce alluring re-
versal interpretations only in the post-disambiguation window. The
overall conflict adaptation pattern suggests a causal influence of cog-
nitive control on thematic role assignment, especially when semantic
and syntactic cues temporarily point to different interpretations. The
results extend earlier work showing a causal connection between dy-
namic cognitive control engagement and the syntactic revision of
garden-path sentences (Hsu & Novick, 2016; Hussey et al., 2015).
In addition to the predicted effects of conflict adaptation, we ob-
served other patterns that suggested a broader modulatory function for
cognitive control. Target and Reversal looks patterned with the syn-
tactically licensed interpretation more after incongruent than con-
gruent Stroop trials. RTs for the final picture choice were also quicker
after incongruent Stroop trials. Unlike the conflict adaptation effect,
these effects were not restricted to conflict sentences. Thus, it appears
that prior cognitive control engagement could facilitate comprehension
by increasing integration of newly arriving cues with prior sentential
content, potentially independently of the type of cue or sentence. We
return to this point in the General Discussion.
Support for conflict adaptation during sentence processing comes
from our finding that looks to the Reversal picture were lower following
disambiguation in the II than the CI condition, which comprised the
same incongruent sentences but differed in the type of previous Stroop
trial. This effect could not have arisen from learned contingencies be-
tween Stroop type and sentence type because our design ensured that
there were no such contingencies. First, the experiment contained more
Stroop-to-Stroop (65) than Stroop-to-critical-sentence (40) transitions,
which meant that participants could not assume that the upcoming task
would be sentence comprehension. Second, in those cases where Stroop
trials were followed by critical sentences, the pairing between con-
gruent and incongruent trial types was counterbalanced: There was no
congruency-based contingency that predicted what type of sentence
trial would occur.
Could participants have incorrectly strategized about a relation be-
tween the two tasks anyway? Here, it is worth noting how different the
two tasks were—seeing visual stimuli and noting font color (Stroop)
169
Cognition 178 (2018) 162–177
versus hearing auditory sentences and picking a matching picture
(comprehension). The lack of overlap in task features entails that any
strategizing would have to be quite abstract and develop over the
course of the experiment. To assess this possibility, we analyzed the first
CI and first II trials for each subject, which should be the least sus-
ceptible to strategy-based explanations. Repeated-measures analysis of
Reversal looks on these trials revealed a significant effect of the pre-
vious Stroop trial type (F(1, 23) = 7.19, p < .05). The first II trial
showed fewer looks to the Reversal picture than the first CI trial. Thus,
an effect of previous Stroop-type on a current conflict sentence was
present on the earliest possible occasion that could be analyzed, in
keeping with conflict adaptation but not a strategy-accrual account.2
The design of Experiment 1 did, however, include one unintended
contingency—the auxiliary verb “was” reliably predicted a passive
structure (e.g., “The rabbit was chased by the fox”) because it was never
present in the active filler sentences (e.g., “The donkey punched the
wall”). Thus, upon hearing “was”, listeners could have reasonably an-
ticipated a passive structure. However, if employed, this strategy would
work against the conflict adaptation hypothesis. Knowing that a tem-
porarily ambiguous sentence was going to continue on as a passive
structure would mitigate differences between the critical conditions (all
of which involved passives) and not heighten them. Nevertheless, it is
possible that anticipation of a passive structure prior to disambiguation
allowed conflict adaptation effects to emerge earlier than would
otherwise be expected. To evaluate this possibility, in Experiment 2 we
altered all filler-active sentences to contain an auxiliary verb (e.g., “The
donkey was punching the wall”), thereby removing the possibility of
using “was” to predict the upcoming structure type. In the absence of
the “was” cue, the parser should default more strongly to the more
frequent Active interpretation prior to disambiguation, leading to more
prolonged competition and later convergence on the Target. Thus,
conflict adaptation effects might emerge later, during the post-dis-
ambiguation interval in Experiment 2 compared to the disambiguation
interval in Experiment 1. To preview our results, we found rate dif-
ferences between II and CI conditions that were consistent with conflict
adaptation, but now in the later, post-disambiguation interval. This
provides independent confirmation of conflict adaptation between
Stroop and sentence comprehension whose timing is consistent with
cue-dependent sentence processing.
3. Experiment 2
3.1. Method
3.1.1. Participants
Twenty-four new right-handed native English speakers with normal
or corrected-to normal vision participated (3 male; age range = 18–22;
2
A reviewer suggested that participants might use a strategy such as predicting that
incongruent Stroop trials would be followed by incongruent sentence trials (and con-
versely that congruent Stroop trials would be followed by congruent sentence trials). Such
a strategy would be counteracted and disproven by actual experience within the ex-
periment because of our counterbalanced design. Nevertheless, it could theoretically lead
to earlier convergence on the Target and/or divergence from the Reversal interpretation
on II compared to both CI and IC trials if participants expect “tricky” tasks to follow a
Stroop-Incongruent trial. Therefore, we evaluated whether there was a significant dif-
ference between the II and IC conditions in the disambiguation interval that paralleled the
reliable difference between the II and CI conditions. Reversal looks on the first II trial
were not significantly different from those on the first IC trial (F(1, 23) = 2, p > .1).
Similarly, analysis of a larger subset of trials, namely, the first half of the experiment
containing unrepeated Target-Reversal pairs also revealed a contrasting pattern: A sig-
nificant difference between II and CI (t(22.59) = −3.28, p < .01) but not between II and
IC (t(23.09) = −.78, p > .4) (Note: Items analysis was not viable for this subset of data
because half of the items appeared in only one condition). Together, these analyses show
a pattern consistent with conflict adaptation, i.e., a significant difference between II and
CI, prior to the point where learned strategies are likely to be used. Conversely, the data
do not support the idea that participants were expecting tricky sentences to follow tricky
Stroop trials (due to the lack of a reliable II-IC difference).
M. Thothathiri et al. Cognition 178 (2018) 162–177

Mage = 19.1). They provided consent under a protocol approved by the

institutional review board at The George Washington University.
Participants received compensation or course credit.

3.1.2. Materials and procedure

All aspects of the stimuli and procedure were identical to
Experiment 1 except for the following. The sentence comprehension
stimuli were re-recorded by a female native English speaker: the critical
passive sentences were of the form “The < noun1 > was < verb > -ed
by the < noun2 > .” (as before) and the filler active sentences were of
the form “The < noun1 > was < verb > -ing the < noun2 > .” Thus,
unlike Experiment 1, the auxiliary verb “was” appeared equally often in
active and passive structures, eliminating any contingency between its
presence and the eventual structure.

3.2. Dependent measures and analysis

The dependent measures and analyses were identical to Experiment

1.

3.3. Results

3.3.1. Stroop accuracy and RT

Mean accuracy on the Stroop trials was 93.72%. Congruent and
incongruent trials did not differ (estimate = −.29, S.E. = .19,
z = −1.53, p > .1. Maccuracy for C trials = 94.39%, I trials = 93.04%).
For RTs, we excluded trials with outlier values or incorrect responses
(10.9%). Analysis of the remaining trials showed a significant con-
gruency effect (estimate = 39.82, S.E. = 6.70, df = 22.36, t = 5.94, Fig. 4. Experiment 2. Looks to Target and Reversal, and the Target Advantage
p < .001. MRT for C trials = 623 ms, I trials = 664 ms), consistent with score (inset) after verb-onset, collapsed across all conditions. Error bars are
participants experiencing conflict during incongruent Stroop trials. 95% confidence intervals.

3.3.2. Sentence comprehension accuracy and RT
Mean accuracy on the critical sentence comprehension trials was
90.10%. Mixed model analyses revealed no significant effects of pre-
vious or current trial type (ps > .1). Thus, listeners arrived at the
correct interpretation equally across conditions. RT analyses after ex-
cluding outliers, incorrect trials, and trials following incorrect Stroop
responses (18.4%) also did not reveal any differences between condi-
tions (ps > .1).
3.3.3. Eye movements
As before, we excluded trials that contained incorrect responses
and/or followed Stroop trials with incorrect responses (14.4%), and
those that contained track loss for more than a third of a given analysis
interval (2.9%).
3.3.3.1. A. Overall time-course of sentence interpretation: Evidence for
competition. Fig. 4 shows the evolution of Target and Reversal looks and
the Target Advantage (inset) collapsed across all conditions. As can be
seen, in the pre-disambiguation interval, Target Advantage was not
significantly different from zero (ps > .2) and did not change over time
(ps > .2). A similar pattern was found in the disambiguation interval
(intercept not significantly above zero and non-significant change over
time, all ps > .1). Both intervals showed growing or persistent looks to
the Reversal picture, suggestive of interference from an active structure
interpretation. In the post-disambiguation interval, the Target
Advantage was significantly positive (Participants: t(24.3) = 6.10,
p < .001; Items: t(35.02) = 3.06, p < .01) and grew reliably over
time (Participants: t(102.70) = 3.56, p < .001; Items: t(36.81) = 3.96,
p < .001). Together, these results suggest that competition from the
interpretation corresponding to an active structure was more potent
and longer lasting in Experiment 2, as predicted. Unlike Experiment 1,
there was no indication of a growth in Target Advantage in the
disambiguation interval. Instead, convergence to the Target over the
Reversal, reflecting a passive structure interpretation, emerged only in
the post-disambiguation interval. Increased competition from the active
structure and later emergence of a Target Advantage compared to
Experiment 1 is consistent with our design change. Because “was” did
not predict a passive structure in the present experiment, the active
structure was likely a stronger competitor until later in the sentence,
forestalling the choice of the correct passive structure until after
disambiguation. Therefore, we expect any conflict adaptation effect
(section B) to emerge later in this experiment, during the post-
disambiguation interval.
3.3.3.2. B. Analyses of Target and Reversal looks: Evidence for conflict
adaptation. Fig. 5 shows the change in Target and Reversal looks over
time in each of the four conditions. As in Experiment 1, we analyzed
Target and Reversal looks in the three relevant time intervals.
In the pre-disambiguation interval, for Target looks, there were no
effects reliable by either participants or items.3 Analysis of Reversal
looks revealed a reliable increase over time and no other effects (Par-
ticipants: t(322) = 3.77, p < .001; Items: t(204) = 6.80, p < .001).
This pattern (significant growth for Reversal but not Target) is con-
sistent with a preference for the Active interpretation (which corre-
sponds to the Reversal picture).
In the disambiguation interval, there was a significant interaction
between the linear increase in Target looks over time and the previous
trial type (Participants: t(299) = 2.74, p < .01; Items: t
(170.00) = 3.31, p < .01). The interaction resulted from a reliable
increase in Target looks over time after incongruent Stroop trials
(Participants: t(276) = 4.69, p < .001; Items: t(68.06) = 2.40,
p < .05) but not congruent Stroop trials (Participants: t(276) = −.63,
p > .5; Items: t(68.06) = −.20, p > .8). There were no other effects.
3
Because visual inspection suggested a possible difference between the CI and II
conditions in the pre-disambiguation interval, we tested and verified that this difference
was not reliable for either Target or Reversal looks.

170
M. Thothathiri et al. Cognition 178 (2018) 162–177

Fig. 5. Experiment 2. Left: Proportion of Target looks in different conditions after verb-onset. Right: Proportion of Reversal looks in different conditions after verb-
onset. Error bars are within-subject 95% confidence intervals adjusted according to Morey (2008).

For Reversal looks, there was a significant time by previous trial type previous Stroop conflict trial on the subsequent sentence trial. In par-
interaction (Participants: t(253) = −2.29, p < .05; Items: t ticular, in the post-disambiguation interval (i.e., the interval where
(170) = −3.32, p < .01). Splitting by previous trial type, the linear interference from the Active interpretation was resolved; see section A),
trend over time was not reliable after incongruent Stroop trials the II condition showed an increase in Target looks and the CI condition
(Participants: t(78.27) = −2.15, p < .05; Items: t(204) = −.04, did not, suggesting facilitation from encountering conflict on a previous
p > .9) or congruent Stroop trials (Participants: t(78.27) = .79, Stroop trial. Additionally, in the disambiguation interval, a prior Stroop
p > .4; Items: t(204) = 3.53, p < .001). conflict trial led to growth in Target looks for both congruent and in-
In the post-disambiguation interval, there was a significant three- congruent sentences, reflecting a non-specific modulatory effect.
way time bin by previous trial type by current trial type interaction for
Target looks (Participants: t(276) = 2.27, p < .05; Items: t
3.4. Discussion
(170) = 2.33, p < .05)—a rate effect that is consistent with conflict
adaptation. For congruent sentences, there was a significant increase
Experiment 2 eliminated the contingency between “was” and a
over time in the CC condition (Participants: t(276) = 2.35, p < .05;
passive structure, which was present in the previous experiment. The
Items: t(204) = 2.68, p < .01) and a non-reliable tendency in the same
only change was to the auditory stimuli, specifically to the structure of
direction in the IC condition (Participants: t(276) = .84, p > .3; Items:
the filler sentences; the visual stimuli were the same as in Experiment 1.
t(204) = .95, p > .3). As can be seen in Fig. 5 (left panel), for incon-
An important consequence of the change was that at the point of en-
gruent sentences, there was an increase in Target looks over time in the
countering the verb on critical items, listeners could not predict whe-
II condition that was marginally significant in the participants analysis
ther the unfolding sentence was going to be an active or a passive.
and significant in the items analysis (Participants: t(276) = 1.89,
Under these conditions, listeners experienced more persistent inter-
p = .06; Items: t(204) = 2.85, p < .01). In contrast, there was no
ference from the active (canonical agent-first) interpretation when
effect in the CI condition (with a tendency, if anything, for decreasing
processing the critical passive sentences, as compared to Experiment 1.
Target looks) (Participants: t(276) = −2.53, p < .05; Items: t
Convergence towards the Target (passive) over the Reversal (active)
(204) = .48, p > .6). In sum, Target looks showed a deviant pattern in
interpretation did not emerge until well after disambiguation, in the
the CI condition, suggesting difficulty in arriving at the correct inter-
post-disambiguation interval. Because no changes were made to visual
pretation when semantic cues conflicted with the syntax. Conversely,
stimuli across experiments, later arrival at the target interpretation in
the II condition showed a similar pattern to the two congruent sentence
Experiment 2 compared to Experiment 1 suggests that the results arose
conditions, indicating that participants converged on the Target inter-
from the availability (or non-availability) of statistical, linguistic cues to
pretation over time despite the conflicting semantic cues, when in-
structure rather than extra-linguistic factors (like visual cues alone).
congruent sentences followed an incongruent Stroop trial. This is con-
Looking more specifically at whether the prior Stroop trial type
sistent with conflict adaptation. For Reversal looks, there was a reliable
influenced sentence interpretation in different conditions, we found
decrease over time (Participants: t(23.34) = −5.37, p < .001; Items: t
evidence consistent with conflict adaptation. Specifically, looks to the
(34.13) = −3.57, p < .01) but no significant effects or interactions
Target grew over time in the II condition post-disambiguation, but
with previous or current trial type.
showed a deviant pattern in the CI condition. Of note, the II condition,
To summarize, Target (but not Reversal) looks showed effects of a
which contained syntax-semantics conflict (incongruent sentences),

171
M. Thothathiri et al.
showed a similar pattern to the conditions containing no-conflict
(congruent sentences). This suggests that prior cognitive control en-
gagement could help eliminate any interpretive disadvantage for in-
congruent sentences. In contrast, the same incongruent sentences, when
preceded by a congruent Stroop trial, showed no growth over time in
choosing the target interpretation, suggesting that syntax-semantics
conflict can disrupt processing.
While Experiment 1 showed a magnitude effect, Experiment 2
showed a rate effect that was consistent with conflict adaptation. Rate
effects (over time) can be just as illuminating as magnitude effects
(overall proportion of looks to a picture) regarding the influence of
cognitive control on online processing. In particular, here they reveal
how cognitive control engagement “pushes around” interpretation
processes as linguistic input unfolds, independently of other antici-
patory effects. That is, no matter the starting point of looking propor-
tions in different conditions (which could differ simply by chance or
some other unrelated factor), how the looks increase or decrease over
time can also be quite telling (e.g., see Barr (2008)).
Why did we not replicate the magnitude effect from Experiment 1,
but instead observed a rate effect? We can only speculate based on the
observed data. As seen in Fig. 5 (left panel), Target looks in the CI
condition (white squares) are above those in the II condition (black
squares) already at verb onset, namely prior to disambiguation. The
reason for this early nominal difference is unclear, but it obviously
cannot be associated with the arrival of syntax-semantics conflict be-
cause that hasn’t happened yet – not until disambiguation. Importantly
however, following disambiguation Target looks grew in the II but not
in the CI condition (in Fig. 5, starting with the disambiguating region
and continuing into the post-disambiguating region, the black squares
(II) increase dramatically but the white squares (CI) remain flat). This
growth may not surpass the random pre-disambiguation bias to lead to
an overall magnitude effect, but it is critically in response to the arrival
of the linguistic information of interest.
We also found a broad modulatory effect from a previous incon-
gruent Stroop trial that parallels a similar effect in Experiment 1. In the
disambiguation interval, Target looks grew more reliably when the
sentences followed an incongruent than a congruent Stroop trial. This
effect was observed regardless of sentence conflict. Thus, it seems that
prior cognitive control engagement might not only influence conflict
resolution but could also affect subject-verb integration during sentence
processing more generally (e.g., by facilitating the processing of in-
coming linguistic information; See General Discussion for elaboration).
4. General discussion
Thematic role assignment relies on both syntactic and semantic
cues; when these cues are pitted against each other, listeners experience
difficulty determining who did what to whom. We hypothesized that
cognitive control could facilitate conflict resolution under such condi-
tions. Previous studies have demonstrated causal effects of cognitive
control on language comprehension during garden-path sentences that
necessitate syntactic revision (Hsu & Novick, 2016; Hussey et al., 2017;
Novick et al., 2014; see also Hussey et al., 2015). The current study
employed simple passive sentences, and showed that cognitive control
engagement influenced thematic role assignment for sentences con-
taining syntax-semantics conflict.
Across all sentence types, listeners ultimately arrived at the correct
interpretation. Convergence on the Target interpretation, however, was
delayed during online processing for incongruent sentences compared
to congruent sentences in the CI but not the II condition. In other words,
semantic cues pushed away from the target interpretation for sentences
describing implausible scenarios (the CI condition), but encountering
Stroop-conflict trials just prior to those sentences successfully coun-
teracted the delay (the II condition). These effects arose in magnitude
(Experiment 1) or in rate (Experiment 2), providing evidence for the
facilitative and accelerating effect of cognitive control engagement on
172
Cognition 178 (2018) 162–177
resolving competition during language comprehension.
A significant effect of prior cognitive control engagement on real-
time measures of sentence interpretation, but not on final accuracy, is
consistent with the Conflict Monitoring Theory described in the
Introduction. Under this account, the detection of conflict triggers
mental adjustments to override irrelevant representations and/or focus
on task-appropriate goals. The resulting sustained cognitive control
activity readies the system for new instances of conflict (Botvinick
et al., 2001; see also Miller & Cohen, 2001; for discussions on how this
applies to language, see Novick et al., 2005; Nozari, Mirman, &
Thompson-Schill, 2014; Nozari & Thompson-Schill, 2015). In the cur-
rent study, we extended the same idea to show a causal interplay be-
tween cognitive control engagement on a previous Stroop trial and
conflict resolution during a subsequent sentence trial. In the II condi-
tion, a Stroop-conflict (incongruent) trial would have activated the
cognitive control system, thereby enabling relatively quicker conflict
resolution during sentence comprehension due to sustained cognitive
control engagement. By contrast, online resolution would be slower in
the CI condition because the conflict engendered by the incoming input
would have to be detected and resolved anew. However, even in the CI
condition, the parser, given sufficient time, can engage the necessary
conflict resolution mechanisms and arrive at the correct final inter-
pretation. In this context, it is worth noting that the instructions given
to participants in the current study did not emphasize response speed.
Participants were given as long as five seconds to respond. Future
studies could investigate whether conflict adaptation effects on com-
prehension accuracy emerge in cases of syntax-semantics conflict if
listeners are under greater time pressure. Indeed, there is evidence for
greater comprehension accuracy following cognitive control engage-
ment when acting out interpretations of garden-path sentences: prior
Stroop conflict resulted in reliably fewer errors, as measured by lis-
teners’ hand-actions when they carried out temporarily ambiguous in-
structions (Hsu & Novick, 2016).
Like accuracy, RT to pick the correct picture is an offline measure
that might not be sensitive to conflict adaptation effects during online
sentence processing. This is especially true in the present case because
the experiments interleaved two tasks with different sets of response
keys for Stroop and sentence comprehension trials. On the sentence
trials, participants had to choose between similar pictures and map the
quadrant location of the chosen picture to the correct response option
on a keyboard. Under these conditions, RTs are likely to reflect post-
interpretive processes such as picture selection and response verifica-
tion (and not just online conflict resolution). Therefore, we report the
RT measures for completeness, but do not draw any definitive conclu-
sions from these analyses.
Selection of the correct interpretation over a competing one could
occur via facilitation of the former and/or inhibition of the latter. In
fact, some have argued that, unlike the direct inhibition observed in
some subcortical systems, choice in the neocortex mostly occurs via
competitive interactions that modulate the relative weights of task-re-
levant and task-irrelevant information (Munakata et al., 2011). Thus,
facilitation and suppression are two sides of the same coin that are hard
to disentangle in the context of complex cortical (and, by extension,
cognitive) functions. In the current study, we observed adaptation ef-
fects that sometimes manifest as suppression of Reversal looks (Ex-
periment 1) and sometimes as promotion of Target looks (Experiment
2). However, for reasons stated above, this does not necessarily imply
that these effects arose from distinct mechanisms. Future studies could
attempt to disentangle facilitation- versus suppression-relevant me-
chanisms using alternate experimental designs (e.g., by including ad-
ditional distractors that temporarily match the correct and the incorrect
interpretations).
Different researchers have proposed different taxonomies of cogni-
tive control functions (e.g., see Harnishfeger, 1995; Nigg, 2000). One
common dimension is whether control occurs at the level of a motor
response or at the level of cognitive representations. We take the effects
M. Thothathiri et al. Cognition 178 (2018) 162–177

reported here to fall clearly in the latter category. We used eye move-
ments as an indirect measure of competition between ongoing sentence
interpretations, which are cognitive representations. The critical effects
emerged while participants were still listening to the sentences, well
before they planned and made keyboard responses. Previous research
suggests that resistance to interference, especially interference from
contents in working memory, is a separate ability from response in-
hibition (Friedman & Miyake, 2004). Considerable evidence has also
linked the ability to resolve interference to the VLPFC (see e.g., Jonides
& Nee, 2006 for a review). Thus, one coherent explanation of our results
is that prefrontal cortex engagement triggered by a previous Stroop trial
enabled listeners to more effectively resist interference from a se-
mantically plausible interpretation when incoming input pointed to-
wards an alternative meaning.
Before closing, it is worth discussing the results that suggest a
broader, non-conflict-specific influence of cognitive control on sentence
processing. In both experiments, we found that encountering conflict on
a previous Stroop trial might have modulated the rate at which listeners
integrated information during online sentence comprehension, in-
dependently of whether a sentence contained conflict. This is consistent
with evidence suggesting a role for the left VLPFC in semantic in-
tegration during comprehension (see e.g., Hagoort, 2005 for a synth-
esis). For example, patients with damage to this region are known to
have general conflict resolution deficits that, during language proces-
sing, manifest as a failure to revise syntactic misinterpretations (Novick
et al., 2009; Vuong & Martin, 2015). But they are also slower than
patients with more posterior brain damage in using verb semantics to
predict upcoming referents even when there is no representational
conflict (e.g., ruling out verb-incompatible distractors and narrowing
down the referent to “apple” when hearing “She will eat the…”). This
pattern suggests that a functional VLPFC might be necessary for in-
tegrating semantic information (Nozari, Mirman, & Thompson-Schill,
2016). In the current study, it is possible that the activation of the
VLPFC by a previous Stroop-conflict trial influenced the rate at which
listeners integrated the subject and the disambiguated verb while in-
terpreting sentences moment by moment. How this purported integra-
tion function should be reconciled with the conflict resolution function
of the VLPFC remains an open question (see Nozari, Mirman, et al.,
2016 for discussion and a compelling unifying framework). Interest-
ingly though, individual differences in cognitive control abilities appear
to modulate the extent to which listeners are misled by locally attrac-
tive but irrelevant linguistic information in real time (Nozari,
Trueswell, & Thompson-Schill, 2016); thus, there is a tight interplay
between integration and cognitive control functions subserved by
common regions within the left VLPFC (Nozari, Mirman, et al., 2016;
Nozari, Trueswell, 2016). Future work using an interleaved paradigm
such as ours should investigate how ramping up prefrontal activity
(e.g., on a prior Stroop trial) modulates both integration and conflict
resolution performance during online sentence processing in a cause-
and-effect fashion.
To conclude, language comprehension is incremental and multi-
determined: Phonological, lexical, syntactic, and semantic cues must be
integrated in real-time to understand the meaning of an utterance.
Though these various cues frequently conspire to inform correct pro-
cessing decisions, sometimes they compete for influence on inter-
pretation, resulting in prolonged uncertainty or even misinterpretation.
The results of this study show that cue integration can benefit from the
online engagement of cognitive control processes, which can prevent
delays in accurately comprehending who did what to whom.

Appendix A. Verbal and visual stimuli used in Experiment 1

Sentence Sentence Picture 1 scenario Picture 2 scenario Picture 3 scenario Picture 4 scenario
type

Critical The soldier was shot by the old matador patted cow duck followed old woman shot soldier soldier shot old woman
woman ducklings
Critical The child was licked by the dog child licked dog dog licked child patient fed cat prince killed troll
Critical The bull was killed by the matador duck followed bull killed matador matador killed bull paramedic carried baby
ducklings
Critical The dog was bitten by the man man bit dog fox watched squirrel dog bit man FBI agent handcuffed
woman
Critical The fox was chased by the rabbit rabbit chased fox man pushed sniper fox chased rabbit matador killed cow
Critical The woman was bitten by the cat tackled ball doctor treated dog mosquito bit woman woman bit mosquito
mosquito
Critical The donkey was kicked by the man nurse fed cat man kicked donkey donkey kicked man mom tickled husband
Critical The rabbit was chased by the fox rabbit chased fox stewardess pushed dog licked treat fox chased rabbit
pilot
Critical The man was handcuffed by the FBI agent handcuffed jockey tickled horse man handcuffed FBI horse carried woman
FBI agent man agent
Critical The ghost was scared by the boy ghost scared boy boy scared ghost lion pushed lioness sniper painted man
Critical The patient was treated by the mosquito bit man dog licked treat doctor treated patient patient treated doctor
doctor
Critical The lion was licked by the cub man painted donkey cub licked lion patient tickled doctor lion licked cub
Critical The monster was scared by the girl victim shot girl scared monster monster scared girl cop handcuffed
paramedic protestor
Critical The child was fed by the mom donkey punched wall cop pushed protestor mom fed child child fed mom
Critical The sniper was shot by the man monster scared nurse tickled doctor sniper shot man man shot sniper
puppy
Critical The doctor was treated by the man painted donkey doctor treated boy scared girl patient treated doctor
patient patient
Critical The jockey was kicked by the monster tickled girl horse kicked jockey matador painted jockey kicked horse
horse stadium
Critical The matador was killed by the bull bull killed matador man served waitress matador killed bull horse kicked sheep

173
M. Thothathiri et al. Cognition 178 (2018) 162–177

Critical The cub was licked by the lion lion licked cub cub licked lion man carried waitress cop pushed protestor
Critical The man was shot by the sniper nurse tickled patient sniper shot man man shot sniper paramedic patted cat
Critical The prince was killed by the dragon killed prince prince killed dragon waitress tickled man child fed man
dragon
Critical The patient was fed by the nurse nurse fed patient rabbit bit fox lion licked bone patient fed nurse
Critical The horse was kicked by the sniper shot zebra horse kicked jockey jockey kicked horse FBI agent tackled
jockey woman
Critical The victim was treated by the paramedic treated matador patted cow FBI agent tackled victim treated
paramedic victim woman paramedic
Critical The cat was chased by the mouse cat chased mouse mouse chased cat man painted dog sniper shot zebra
Critical The mosquito was bitten by the boy scared girl soldier tackled man woman bit mosquito mosquito bit woman
woman
Critical The boy was scared by the ghost sniper painted man boy scared ghost ghost scared boy soldier shot man
Critical The dragon was killed by the woman painted fox chased squirrel dragon killed prince prince killed dragon
prince mosquito
Critical The nurse was fed by the patient patient fed nurse fox watched squirrel sniper painted rabbit nurse fed patient
Critical The mom was fed by the child child fed mom fox chased squirrel donkey punched wall mom fed child
Critical The dog was licked by the child dog licked child paramedic treated nurse tickled patient child licked dog
cat
Critical The cop was handcuffed by the old woman pushed donkey kicked wall robber handcuffed cop cop handcuffed robber
robber soldier
Critical The man was kicked by the donkey old woman pushed man kicked donkey donkey kicked man child carried dog
soldier
Critical The old woman was shot by the old woman shot stewardess punched mosquito bit man soldier shot old woman
soldier soldier pilot
Critical The robber was handcuffed by the robber handcuffed cop handcuffed stewardess punched boy scared girl
cop cop robber pilot
Critical The FBI agent was handcuffed by nurse tickled doctor man handcuffed FBI monster tickled girl FBI agent handcuffed
the man agent man
Critical The mouse was chased by the cat ghost tickled girl mouse chased cat passenger scared cat chased mouse
stewardess
Critical The girl was scared by the monster patient tickled doctor nurse fed cat girl scared monster monster scared girl
Critical The paramedic was treated by the nurse tickled doctor jockey tickled horse paramedic treated victim treated
victim victim paramedic
Critical The man was bitten by the dog girl scared ghost duck fed ducklings man bit dog dog bit man
Filler The nurse painted the patient nurse painted patient matador painted bull nurse fed patient nurse fed cat
Filler The matador killed the cow matador killed cow matador painted matador patted cow waitress tickled man
stadium
Filler The stewardess pushed the stewardess pushed prince scolded troll stewardess pushed cop pushed robber
passenger pilot passenger
Filler The waitress painted the man waitress painted man sniper shot elk man painted FBI agent man served waitress
Filler The nurse tickled the patient nurse patted cat horse kicked sheep nurse tickled patient nurse tickled doctor
Filler The soldier shot the man soldier shot man soldier tackled man soldier tickled old girl pushed monster
woman
Filler The sniper tickled the man sniper shot man man painted donkey passenger scared sniper tickled man
stewardess
Filler The lion pushed the cub lion pushed lioness donkey punched wall lion pushed cub lion hugged bone
Filler The matador painted the bull monster tickled girl matador painted bull matador killed cow matador painted
stadium
Filler The donkey carried the man bull carried matador donkey carried donkey carried man lion pushed cub
squirrel
Filler The fox chased the squirrel man painted dog fox chased squirrel fox watched squirrel cat chased ball
Filler The waitress tickled the man waitress tickled man robber pushed cop waitress painted man sniper tickled man
Filler The boy scared the girl stewardess punched boy scared ghost child carried dog boy scared girl
pilot
Filler The paramedic carried the victim paramedic carried donkey punched wall paramedic carried paramedic treated cat
baby victim
Filler The soldier pushed the old woman soldier carried prince scolded troll soldier pushed old sniper tickled man
woman woman
Filler The cat chased the ball duck followed nurse patted cat cat chased mouse cat chased ball
ducklings
Filler The dog licked the treat monster scolded dog licked treat sniper painted man dog punched treat
puppy
Filler The monster scared the puppy lion pushed lioness monster scolded ghost tickled monster monster scared puppy
puppy

174
M. Thothathiri et al. Cognition 178 (2018) 162–177

Filler The donkey kicked the wall donkey carried donkey punched wall donkey kicked wall nurse painted patient
squirrel
Filler The dragon carried the prince dragon carried prince dragon killed prince prince scolded troll doctor watched dog
Filler The nurse fed the cat man painted donkey nurse fed cat nurse patted cat nurse tickled patient
Filler The ghost painted the boy ghost scared boy sniper tickled man ghost painted boy girl scared ghost
Filler The horse carried the jockey horse carried woman horse kicked sheep ghost tickled girl horse carried jockey
Filler The monster carried the girl monster scared monster tickled girl cop photographed Monster carried girl
puppy protestor
Filler The cop handcuffed the protestor cop handcuffed lion hugged bone cop handcuffed FBI agent handcuffed
robber protestor woman
Filler The rabbit bit the fox donkey carried man rabbit bit fox fox watched squirrel rabbit chased fox
Filler The paramedic treated the cat paramedic treated paramedic carried paramedic patted cat mom tickled husband
cat baby
Filler The dog bit the bone dog bit man dog bit bone duck fed ducklings lion licked bone
Filler The prince killed the troll prince killed troll dragon carried prince lion pushed lioness prince killed dragon
Filler The mom tickled the baby mom carried child cop handcuffed mom tickled husband mom tickled baby
protestor
Filler The cop pushed the robber nurse painted patient cop pushed robber cop pushed protestor soldier pushed old
woman
Filler The man pushed the sniper man shot sniper man pushed sniper sniper shot elk lion pushed cub
Filler The sniper shot the elk mom tickled child soldier carried sniper shot zebra sniper shot elk
woman
Filler The woman painted the mosquito woman painted woman bit mosquito mosquito bit man sniper shot zebra
mosquito
Filler The doctor treated the dog waitress painted chef doctor treated dog doctor treated patient dog licked treat
Filler The FBI agent handcuffed the mouse bit cat FBI agent tackled woman painted FBI agent handcuffed
woman woman mosquito woman
Filler The patient fed the cat patient fed cat mouse bit cat patient tickled doctor patient fed nurse
Filler The soldier tickled the old woman soldier pushed old soldier tickled old paramedic treated cat soldier shot man
woman woman
Filler The lion licked the bone lion licked cub lion pushed cub passenger scared lion licked bone
stewardess
Filler The horse kicked the sheep horse kicked jockey mom tickled baby horse kicked sheep horse carried woman

Appendix B. Norming results

The online survey contained 3 sections: plausibility ratings, Agent/Patient ratings, and noun visibility ratings. Each section collected responses on
a 6-point Likert-type scale (−5, −3, −1, 1, 3, 5). For plausibility ratings, participants were asked to read each critical sentence and answer: “How
likely is the scenario described by the sentence?” Response options ranged from highly unlikely (−5) to highly likely (+5). For Agent/Patient
ratings, participants were asked: “Which scenario is more likely: Noun verb-ed someone/an(other) animal or Noun was verb-ed by someone/an
(other) animal?” Response options ranged from sentence 1 being highly more likely than sentence 2 (−5) to the reverse (+5). Thus, negative
numbers denoted a preference for the noun as Agent of the action, and positive numbers denoted a Patient preference. For noun visibility ratings,
participants saw a noun and the target and reversal pictures that contained the corresponding entity and were asked to: “Rate whether the entity
listed below is more visible in the left or right picture.” Response options ranged from highly more visible in the left picture (−5) to highly more
visible in the right picture (+5).
Plausibility ratings revealed two items (matador killed by bull versus bull killed by matador, and dragon killed by prince versus prince killed by
dragon) that received similar ratings (difference < 1). Because we could not categorize these scenarios as plausible or implausible, we removed
them from analysis. The remaining 18 scenarios and their reversed versions were clearly separable (Mdifference between ratings for plausible and
implausible scenarios = 7.90; t(17) = 16.39, p < .001). Crucially, participants gave significantly higher plausibility ratings to scenarios we cate-
gorized as congruent than the reversed scenarios we categorized as incongruent.
The Agent/Patient ratings further confirmed that subjects of incongruent sentences were more likely to be Agents of the actions described by the
verbs, thereby leading to conflict in a passive structure, because the first noun phrase is actually the Patient in these constructions (MAgent
rating = −3.57, t(17) = −11.33, p < .001). Conversely, subjects of congruent sentences were judged more likely to be Patients of the actions
(MPatient rating = 2.92, t(17) = 7.74, p < .001), indicating that the continuation of the sentence as a passive structure matched semantic expecta-
tions.
Noun visibility ratings were collected outside a sentence context to evaluate purely visual confounds. However, these ratings also showed
plausibility effects. Participants rated nouns that were the subjects of incongruent sentences as more visible when they were Agents, irrespective of
whether the picture was on the left or right (MAgent picture preference = 2.71, t(17) = −9.91, p < .001). Conversely, subjects of congruent sentences
were rated as more visible when they were Patients (MPatient picture preference = 1.38, t(17) = 4.25, p < .001). That is, participants rated pictures
depicting congruent scenarios as more visible, which should yield conflict when the sentences described the reversed incongruent scenario.
In sum, the entire survey supported our assumption that incongruent sentence stimuli described less plausible scenarios and should therefore
generate more syntax-semantics conflict than their corresponding congruent versions.

175
M. Thothathiri et al.
Appendix C. Supplementary material
Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.cognition.2018.05.014.
References
Altmann, G. T., & Kamide, Y. (2007). The real-time mediation of visual attention by
language and world knowledge: Linking anticipatory (and other) eye movements to
linguistic processing. Journal of Memory and Language, 57(4), 502–518.
Barr, D. J. (2008). Analyzing ‘visual world’ eyetracking data using multilevel logistic
regression. Journal of Memory and Language, 59(4), 457–474.
Bates, E., & MacWhinney, B. (1989). Functionalism and the competition model. In B.
MacWhinney, & E. Bates (Eds.). The crosslinguistic study of sentence processing (pp. 3–
73). Cambridge: Cambridge University Press.
Botvinick, M. M., Braver, T. S., Barch, D. M., Carter, C. S., & Cohen, J. D. (2001). Conflict
monitoring and cognitive control. Psychological Review, 108(3), 624–652.
Chow, W. Y., & Phillips, C. (2013). No semantic illusions in the “Semantic P600” phe-
nomenon: ERP evidence from Mandarin Chinese. Brain Research, 1506, 76–93.
Christianson, K., Hollingworth, A., Halliwell, J. F., & Ferreira, F. (2001). Thematic roles
assigned along the garden path linger. Cognitive Psychology, 42(4), 368–407.
Christianson, K., Luke, S. G., & Ferreira, F. (2010). Effects of plausibility on structural
priming. Journal of Experimental Psychology: Learning, Memory, and Cognition, 36(2),
538.
Demiral, Ş. B., Schlesewsky, M., & Bornkessel-Schlesewsky, I. (2008). On the universality
of language comprehension strategies: Evidence from Turkish. Cognition, 106(1),
484–500.
Egner, T., & Hirsch, J. (2005). Cognitive control mechanisms resolve conflict through
cortical amplification of task-relevant information. Nature Neuroscience, 8(12),
1784–1790.
Ferreira, F. (2003). The misinterpretation of noncanonical sentences. Cognitive Psychology,
47(2), 164–203.
Freitas, A. L., Bahar, M., Yang, S., & Banai, R. (2007). Contextual adjustments in cognitive
control across tasks. Psychological Science, 18(12), 1040–1043.
Friedman, N. P., & Miyake, A. (2004). The relations among inhibition and interference
control functions: A latent-variable analysis. Journal of Experimental Psychology:
General, 133(1), 101–135.
Garnsey, S. M., Pearlmutter, N. J., Myers, E., & Lotocky, M. A. (1997). The contributions
of verb bias and plausibility to the comprehension of temporarily ambiguous sen-
tences. Journal of Memory and Language, 37(1), 58–93.
Gratton, G., Coles, M. G., & Donchin, E. (1992). Optimizing the use of information:
Strategic control of activation of responses. Journal of Experimental Psychology:
General, 121(4), 480–506.
Hagoort, P. (2005). On Broca, brain, and binding: A new framework. Trends in Cognitive
Sciences, 9(9), 416–423.
Harnishfeger, K. K. (1995). The development of cognitive inhibition: Theories, defini-
tions, and research evidence. In F. N. Dempster, & C. J. Brainerd (Eds.). Interference
and inhibition in cognition (pp. 175–204). New York: Academic Press.
Hsu, N. S., Jaeggi, S., & Novick, J. M. (2017). A common neural hub resolves syntactic
and non-syntactic conflict through cooperation with task-specific networks. Brain and
Language, 166, 63–77.
Hsu, N. S., & Novick, J. M. (2016). Dynamic engagement of cognitive control modulates
recovery from misinterpretation during real-time language processing. Psychological
Science, 27, 572–582. http://dx.doi.org/10.1177/0956797615625223.
Hussey, E. K., Harbison, J. I., Teubner-Rhodes, S. E., Mishler, A., Velnoskey, K., & Novick,
J. M. (2017). Memory and language improvements following cognitive control
training. Journal of Experimental Psychology: Learning, Memory, and Cognition, 43(1),
23–58.
Hussey, E. K., Ward, N., Christianson, K., & Kramer, A. F. (2015). Language and memory
improvements following tDCS of left lateral prefrontal cortex. PLoS ONE, 10(11),
e0141417.
January, D., Trueswell, J. C., & Thompson-Schill, S. L. (2009). Co-localization of Stroop
and syntactic ambiguity resolution in Broca's area: Implications for the neural basis of
sentence processing. Journal of Cognitive Neuroscience, 21(12), 2434–2444.
Jonides, J., & Nee, D. E. (2006). Brain mechanisms of proactive interference in working
memory. Neuroscience, 139(1), 181–193.
Kaan, E., & Swaab, T. Y. (2002). The brain circuitry of syntactic comprehension. Trends in
Cognitive Sciences, 6(8), 350–356.
Kamide, Y., Altmann, G. T., & Haywood, S. L. (2003). The time-course of prediction in
incremental sentence processing: Evidence from anticipatory eye movements. Journal
of Memory and Language, 49(1), 133–156.
Kamide, Y., Scheepers, C., & Altmann, G. T. (2003). Integration of syntactic and semantic
information in predictive processing: Cross-linguistic evidence from German and
English. Journal of Psycholinguistic Research, 32(1), 37–55.
Kerns, J. G., Cohen, J. D., MacDonald, A. W., Cho, R. Y., Stenger, V. A., & Carter, C. S.
(2004). Anterior cingulate conflict monitoring and adjustments in control. Science,
303(5660), 1023–1026.
Kim, A., & Osterhout, L. (2005). The independence of combinatory semantic processing:
Evidence from event-related potentials. Journal of Memory and Language, 52(2),
205–225.
Kim, A., & Sikos, L. (2011). Conflict and surrender during sentence processing: An ERP
study of syntax-semantics interaction. Brain and Language, 118(1), 15–22.
King, J. A., Korb, F. M., von Cramon, D. Y., & Ullsperger, M. (2010). Post-error behavioral
176
Cognition 178 (2018) 162–177
adjustments are facilitated by activation and suppression of task-relevant and task-
irrelevant information processing. Journal of Neuroscience, 30(38), 12759–12769.
Knoeferle, P., Crocker, M. W., Scheepers, C., & Pickering, M. J. (2005). The influence of
the immediate visual context on incremental thematic role-assignment: Evidence
from eye-movements in depicted events. Cognition, 95(1), 95–127.
Kuperberg, G. R. (2007). Neural mechanisms of language comprehension: Challenges to
syntax. Brain Research, 1146, 23–49.
Malyutina, S., & den Ouden, D. (2015). What is it that lingers? Garden-path (mis) in-
terpretations in younger and older adults. The Quarterly Journal of Experimental
Psychology, 69(5), 880–906.
Milham, M., Banich, M., Webb, A., Barad, V., Cohen, N., Wszalek, T., & Kramer, A.
(2001). The relative involvement of anterior cingulate and prefrontal cortex in at-
tentional control depends on nature of conflict. Cognitive Brain Research, 12(3),
467–473.
Miller, E. K., & Cohen, J. D. (2001). An integrative theory of pre-frontal cortex function.
Annual Review of Neuroscience, 24, 167–202.
Morey, R. D. (2008). Confidence intervals from normalized data: A correction to
Cousineau (2005). Tutorial in Quantitative Methods for Psychology, 4(2), 61–64.
Munakata, Y., Herd, S. A., Chatham, C. H., Depue, B. E., Banich, M. T., & O’Reilly, R. C.
(2011). A unified framework for inhibitory control. Trends in Cognitive Sciences,
15(10), 453–459.
Nigg, J. T. (2000). On inhibition/disinhibition in developmental psychopathology: Views
from cognitive and personality psychology and a working inhibition taxonomy.
Psychological Bulletin, 126(2), 220–246.
Novick, J. M., Hussey, E., Teubner-Rhodes, S., Harbison, J. I., & Bunting, M. F. (2014).
Clearing the garden-path: Improving sentence processing through cognitive control
training. Language, Cognition and Neuroscience, 29(2), 186–217.
Novick, J. M., Kan, I. P., Trueswell, J. C., & Thompson-Schill, S. L. (2009). A case for
conflict across multiple domains: Memory and language impairments following da-
mage to ventrolateral prefrontal cortex. Cognitive Neuropsychology, 26(6), 527–567.
Novick, J. M., Thompson-Schill, S. L., & Trueswell, J. C. (2008). Putting lexical constraints
in context into the visual-world paradigm. Cognition, 107(3), 850–903.
Novick, J. M., Trueswell, J. C., & Thompson-Schill, S. L. (2005). Cognitive control and
parsing: Reexamining the role of Broca’s area in sentence comprehension. Cognitive,
Affective, & Behavioral Neuroscience, 5(3), 263–281.
Nozari, N., Mirman, D., & Thompson-Schill, S. L. (2014). The role of the left ventrolateral
prefrontal cortex in online sentence processing. Frontiers in Psychology. http://dx.doi.
org/10.3389/conf.fpsyg.2014.64.00012.
Nozari, N., Mirman, D., & Thompson-Schill, S. L. (2016). The ventrolateral prefrontal
cortex facilitates processing of sentential context to locate referents. Brain and
Language, 157–158, 1–13.
Nozari, N., & Thompson-Schill, S. L. (2015). Left ventrolateral prefrontal cortex in pro-
cessing of words and sentences. In G. Hickok, & S. L. Small (Eds.). The neurobiology of
language. Academic Press.
Nozari, N., Trueswell, J. C., & Thompson-Schill, S. L. (2016). The interplay of local at-
traction, context and domain-general cognitive control in activation and suppression
of semantic distractors during sentence comprehension. Psychonomic Bulletin &
Review. http://dx.doi.org/10.3758/s13423-016-1068-8.
Pearlmutter, N. J., & MacDonald, M. C. (1992). Plausibility and syntactic ambiguity re-
solution. In Proceedings of the 14th annual conference of the Cognitive Science Society
(pp. 521–542).
Robinson, G., Shallice, T., & Cipolotti, L. (2005). A failure of high level verbal response
selection in progressive dynamic aphasia. Cognitive Neuropsychology, 22(6), 661–694.
Schnur, T. T., Schwartz, M. F., Kimberg, D. Y., Hirshorn, E., Coslett, H. B., & Thompson-
Schill, S. L. (2009). Localizing interference during naming: Convergent neuroimaging
and neuropsychological evidence for the function of Broca's area. Proceedings of the
National Academy of Sciences of the United States of America, 106(1), 322–327. http://
dx.doi.org/10.1073/pnas.0805874106.
Slattery, T. J., Sturt, P., Christianson, K., Yoshida, M., & Ferreira, F. (2013). Lingering
misinterpretations of garden path sentences arise from competing syntactic re-
presentations. Journal of Memory and Language, 69(2), 104–120.
Snyder, H. R., Banich, M. T., & Munakata, Y. (2011). Choosing our words: Retrieval and
selection processes recruit shared neural substrates in left ventrolateral prefrontal
cortex. Journal of Cognitive Neuroscience, 23(11), 3470–3482.
Snyder, H. R., Banich, M. T., & Munakata, Y. (2014). All competition is not alike: Neural
mechanisms for resolving underdetermined and prepotent competition. Journal of
Cognitive Neuroscience, 26(11), 2608–2623.
Spivey, M. J., Tanenhaus, M. K., Eberhard, K. M., & Sedivy, J. C. (2002). Eye movements
and spoken language comprehension: Effects of visual context on syntactic ambiguity
resolution. Cognitive Psychology, 45(4), 447–481.
Stoops, A., Luke, S. G., & Christianson, K. (2014). Animacy information outweighs mor-
phological cues in Russian. Language, Cognition and Neuroscience, 29(5), 584–604.
Sturt, P. (2007). Semantic re-interpretation and garden path recovery. Cognition, 105(2),
477–488.
Tanenhaus, M. K., Spivey-Knowlton, M. J., Eberhard, K. M., & Sedivy, J. C. (1995).
Integration of visual and linguistic information in spoken language comprehension.
Science, 1632–1634.
Thothathiri, M., Kim, A., Trueswell, J. C., & Thompson-Schill, S. L. (2012). Parametric
effects of syntactic–semantic conflict in Broca’s area during sentence processing.
M. Thothathiri et al.
Brain and Language, 120(3), 259–264.
Trueswell, J. C., Sekerina, I., Hill, N. M., & Logrip, M. L. (1999). The kindergarten-path
effect: Studying on-line sentence processing in young children. Cognition, 73(2),
89–134.
Trueswell, J. C., Tanenhaus, M. K., & Garnsey, S. M. (1994). Semantic influences on
parsing: Use of thematic role information in syntactic ambiguity resolution. Journal of
Memory and Language, 33(3), 285–318.
van de Meerendonk, N., Rueschemeyer, S.-A., & Kolk, H. H. J. (2013). Language com-
prehension interrupted: Both language errors and word degradation activate Broca’s
area. Brain and Language, 126(3), 291–301.
van Herten, M., Chwilla, D. J., & Kolk, H. H. (2006). When heuristics clash with parsing
177
Cognition 178 (2018) 162–177
routines: ERP evidence for conflict monitoring in sentence perception. Journal of
Cognitive Neuroscience, 18(7), 1181–1197.
Vuong, L. C., & Martin, R. C. (2011). LIFG-based attentional control and the resolution of
lexical ambiguities in sentence context. Brain & Language, 116(1), 22–32.
Vuong, L. C., & Martin, R. C. (2014). Domain-specific executive control and the revision of
misinterpretations in sentence comprehension. Language, Cognition and Neuroscience,
29(3), 312–325.
Vuong, L., & Martin, R. (2015). The role of LIFG-based executive control in sentence
comprehension. Cognitive Neuropsychology, 32(5), 243–265.
Ye, Z., & Zhou, X. (2009). Conflict control during sentence comprehension: fMRI evi-
dence. Neuroimage, 48(1), 280–290.