Wait a second! Delayed impact of argument roles on on-line verb prediction

Abstract

Comprehenders can use rich contextual information to anticipate upcoming input on the fly, but recent findings suggest that salient information about argument roles may not impact verb prediction. We took advantage of the word order properties of Mandarin Chinese to examine the time course with which argument role information impacts verb prediction. We isolated the contribution of argument role information by manipulating the order of pre-verbal noun phrase arguments while holding lexical information constant, and we examined its effects on accessing the verb in long-term semantic memory by measuring the amplitude of the N400 component. Experiment 1 showed when the verb appeared immediately after its arguments, even strongly constraining argument role information failed to modulate the N400 response to the verb. An N400 effect emerged in Experiment 2 when the verb appeared at a greater delay. Experiment 3 corroborated the contrast between the first two experiments through a within-participants manipulation of the time interval between the arguments and the verb, by varying the position of an adverbial phrase. These results suggest time is a key factor governing how diverse contextual information contributes to predictions. Here argument role information is shown to impact verb prediction, but its effect is not immediate. 3

Full text

1
Wait a second! Delayed impact of argument roles on on-line verb
prediction
Wing-Yee Chow1,a, Ellen Lau2, Suiping Wang3, and Colin Phillips2
1 Department of Linguistics, University College London, Chandler House, 2 Wakefield
Street, London WC1N 1PF, United Kingdom
2 Department of Linguistics, 1401 Marie Mount Hall, University of Maryland College Park,
Maryland, 20742, United States of America
3 Department of Psychology, South China Normal University, Shipai, Guangzhou, 510631,
China
a Corresponding author.
Division of Psychology and Language Sciences
University College London
Chandler House
2 Wakefield Street
London
WC1N 1PF
United Kingdom
Telephone: +44 (0) 20 7679 4213
E-mail address: wingyee.chow@ucl.ac.uk (W.Y. Chow).
Keywords: Language comprehension; Prediction; Thematic relations; N400

2
Abstract
Comprehenders can use rich contextual information to anticipate upcoming input on the fly,
but recent findings suggest that salient information about argument roles may not impact verb
prediction. We took advantage of the word order properties of Mandarin Chinese to examine
the time course with which argument role information impacts verb prediction. We isolated the
contribution of argument role information by manipulating the order of pre-verbal noun phrase
arguments while holding lexical information constant, and we examined its effects on accessing
the verb in long-term semantic memory by measuring the amplitude of the N400 component.
Experiment 1 showed when the verb appeared immediately after its arguments, even strongly
constraining argument role information failed to modulate the N400 response to the verb. An
N400 effect emerged in Experiment 2 when the verb appeared at a greater delay. Experiment
3 corroborated the contrast between the first two experiments through a within-participants
manipulation of the time interval between the arguments and the verb, by varying the position
of an adverbial phrase. These results suggest time is a key factor governing how diverse
contextual information contributes to predictions. Here argument role information is shown to
impact verb prediction, but its effect is not immediate.

3
Introduction
Much recent work suggests that generating predictions about the future is a fundamental
principle underlying the brain’s operations (Bar, 2011; Hawkins & Blakeslee, 2004; Llinás,
2002). After an explosion of work on this topic across the last decade, predictive processes
have now been repeatedly demonstrated across domains such as visual and auditory perception
(Bar, 2007; Bendixen, Schroger, & Winkler, 2009), motor planning (Wolpert, 1997), and
language comprehension (e.g., Delong, Urbach, & Kutas, 2005; Van Berkum, Brown,
Zwitserlood, Kooijman, & Hagoort, 2005; Wicha, Moreno, & Kutas, 2004). Just as a tennis
player’s ability to anticipate a flying ball’s trajectory is critical to his/her chances of hitting the
ball, the ability to anticipate properties of upcoming input ahead of time is likely key to the
efficiency with which the brain processes information. Therefore, a better understanding of
how predictions are computed is important for understanding the brain’s capacity to process
information quickly and robustly.
However, recent work in language comprehension has illustrated a notable case in
which salient and reliable contextual information appears not to impact predictions. The
likelihood of a verb depends not only on the arguments, but also the roles they take. For
example, given two arguments ‘cop’ and ‘thief’, the likelihood of the verb ‘arrest’ can differ
greatly depending on which argument takes the agent vs. patient role. A verb like ‘arrest’ may
be likely if ‘cop’ takes an agent role and ‘thief’ takes a patient role, but it would have been
much less likely if the opposite were true. However, some evidence suggests that this
information may not be used by humans when they generate predictions online. The goal of
the current work is to determine the cause for this puzzling finding, and more broadly, their
implications for models of prediction.

4
Prediction in language comprehension
In the domain of language, previous research has provided much evidence for human
comprehenders’ ability to generate linguistic predictions on the fly (for a review see Kamide,
2008; Kutas, DeLong, & Smith, 2011). For example, listeners are more likely to direct their
eye-gaze to a picture of an edible object (e.g., a cake) when they hear the beginning of an
utterance like ‘The boy will eat…’ compared to a neutral utterance such as ‘The boy will move…’
(Altmann & Kamide, 1999). Further, much evidence has suggested that comprehenders
compute rich expectations about upcoming inputs at multiple levels of representation (syntactic:
Ilkin & Sturt, 2011; Lau, Stroud, Plesch, & Phillips, 2006; Levy, Fedorenko, Breen, & Gibson,
2012; Omaki et al., 2015; Staub & Clifton, 2006; Wicha et al., 2004; Van Berkum et al., 2005;
Yoshida, Dickey, & Sturt, 2013; lexico-semantic: Federmeier & Kutas, 1999; Kutas & Hillyard,
1984; Otten & Van Berkum, 2008; Szewczyk & Schriefers, 2013; phonological and
orthographic: Delong et al., 2005; Dikker, Rabagliati, Farmer, & Pylkkanen, 2010; Dikker,
Rabagliati, & Pylkkänen, 2009; Farmer, Yan, Bicknell, & Tanenhaus, 2015; Kim & Lai, 2012;
Laszlo & Federmeier, 2009). Here, we operationally define ‘prediction’ as the pre-activation
of stored representations before the bottom-up input is encountered, and we will make no a
priori assumptions regarding the nature of the mechanisms involved (e.g., whether they are
automatic or controlled).
Event-related brain potentials (ERPs) provide a useful tool for investigating the nature
of prediction in real time language comprehension. In particular, the amplitude of the N400, a
negative-going ERP component that starts at around 250ms and peaks at around 400ms post
stimulus onset (Kutas & Hillyard, 1980), may be used as a diagnostic of comprehenders’ lexical
semantic predictions. An N400 response is elicited by any content word (e.g., nouns, verbs,
adjectives), whether presented in isolation or in sentences (Kutas & Federmeier, 2011) and it
has been proposed to reflect the ease of accessing a word’s meaning from long-term memory

5
(Kutas & Federmeier, 2000; Lau, Phillips, & Poeppel, 2008). Crucially, its amplitude is
modulated by a word’s predictability, which is commonly operationalized as the proportion of
trials on which speakers continue the sentence context with that word in an untimed sentence
fragment completion task (also known as cloze probability; Taylor, 1953). Many have observed
an inverse relationship between the amplitude of the N400 response to a word and that word's
cloze probability (e.g., Dambacher, Kliegl, Hofmann, & Jacobs, 2006; Gunter, Stowe, &
Mulder, 1997; Kutas & Hillyard, 1984; Wlotko & Federmeier, 2012), which has been taken to
reflect facilitated lexical semantic access for more predictable words. Under this view,
reduction in the N400 response to a given word may be taken to reflect the extent to which that
word’s meaning has been pre-activated (the “prediction view” of the N400; Federmeier &
Kutas, 1999).
1
N400 amplitude measures have provided some of the primary evidence that
comprehenders can rapidly integrate various sources of contextual information to generate
predictions. Many factors that are known to affect a word’s cloze probability (e.g., negation
2
,
sentence structure, event knowledge, world knowledge, message-level representations) have
also been shown to modulate N400 amplitude (Bicknell, Elman, Hare, McRae, & Kutas, 2010;
Hagoort, Hald, Bastiaansen, & Petersson, 2004; Kos, Vosse, van den Brink, & Hagoort, 2010;
Otten, Nieuwland, & Van Berkum, 2007; Paczynski & Kuperberg, 2012; Van Berkum, 2009;
Xiang & Kuperberg, 2015). For example, listeners presented with an utterance like ‘Every
1
Federmeier and Kutas (1999) pointed out that the predictability of a word, as assessed by cloze
probability, is not always identical to the predictability of the meaning (or semantic features) of that
word. Therefore, cloze probability may be an imperfect proxy if the N400 amplitude is modulated by
the ease of accessing the meaning of a word. We return to this point in discussing the logic of the current
study.
2
Earlier studies have reported that the N400 was insensitive to negation in simple sentences (e.g., “A
robin is not a tree/bird”; Fischler, Bloom, Childers, Roucos, & Perry, 1983; Katayama, Miyata, & Yagi,
1987; Kounios & Holcomb, 1992; Lüdtke, Friedrich, De Filippis, & Kaup, 2008), but more recent work
has demonstrated that the N400 is sensitive to negation as long as it is pragmatically licensed
(Nieuwland & Kuperberg, 2008).

6
evening I drink some…’ show a smaller N400 response to the word ‘wine’ if the speaker is an
adult compared to when it is a child (Van Berkum, van den Brink, Tesink, Kos, & Hagoort,
2008).
It should be noted that the N400 has also been taken to reflect the ease with which a
word can be semantically integrated in a given context (the “integration view” of the N400;
e.g., Brown & Hagoort, 1993; Van Berkum, Hagoort & Brown, 1999). This view is most
commonly associated with the generalisation that semantically or pragmatically incongruous
words tend to elicit a larger N400 response than congruous words (e.g., Hagoort et al., 2004).
Under this view, the N400’s sensitivity to a word’s predictability does not arise from pre-
activation per se, but merely reflects the fact that words that are more predictable can be
semantically integrated more easily. While it is not the primary goal of the present study to
resolve this long-standing debate about the functional significance of the N400, we will
consider both the prediction and integration views of the N400 and discuss the present findings’
potential implications for both accounts.
Prediction in the processing of thematic relations
Previous research using different experimental techniques has demonstrated that
comprehenders can use various sources of information to compute thematic relations
predictively (Altmann & Kamide, 1999; Aoshima, Phillips, & Weinberg, 2004; Ferretti,
McRae, & Hatherell, 2001; Kamide, Altmann, & Haywood, 2003; Knoeferle, Crocker,
Scheepers, & Pickering, 2005; McRae, Hare, Elman, & Ferretti, 2005; Omaki et al., 2015). For
example, McRae et al. (2005) reported that participants were faster to name a verb when it was
preceded by a noun that was a typical event participant (e.g., agent, patient) and suggested that
comprehenders can use nouns to pre-activate the classes of events in which they typically play
a role (see also Hare, Jones, Thomson, Kelly, & McRae, 2009). In another study, Kamide et al.
(2003) used a visual world eye-tracking paradigm and found that listeners of Japanese, a verb-

7
final language, were more likely to look towards a picture of a hamburger prior to the verb
when the first argument (waitress) was marked with nominative case and the second argument
(customer) was dative-marked, compared to when the second argument was accusative-marked.
Since nominative, accusative and dative cases tend to mark arguments that serve the agent,
patient and goal roles for most verbs in Japanese, these results suggested that listeners can use
case markings on pre-verbal arguments to anticipate properties of the upcoming verb (e.g., the
number of arguments it takes). Meanwhile, other evidence suggests that comprehenders’
predictions about an upcoming verb are also sensitive to information such as the number of
pre-verbal arguments (Brusini, Brun, Brunet, & Christophe, 2015) and the animacy of likely
arguments (Altmann, 1999).
However, evidence for the impact of the arguments’ structural roles, i.e., which
argument is the subject and which is the object, on comprehenders’ verb predictions has
remained surprisingly scarce. In many languages the arguments’ structural roles are
unambiguously encoded by word order, e.g., ‘John likes Mary’ vs. ‘Mary likes John’. This
information is crucial for interpreting thematic relations in a sentence, and can drastically
change the likelihood of an upcoming verb in sentences with a verb-final word order. For
example, in a verb-final ba-construction sentence in Mandarin Chinese like (1a), the verb
‘arrest’ is much more likely when ‘cop’ is the subject and ‘thief’ is the object than if the same
pair of arguments are reversed, as in (1b).
(1) Argument role reversal in a verb-final sentence in Mandarin Chinese:
(a) jingcha ba xiaotou zhua-le…
cop BA thief arrest…
“The cop arrested the thief.”
(b) xiaotou ba jingcha zhua-le…
thief BA cop arrest…
“The thief arrested the cop.”
Since the particle ba always follows the subject and immediately precedes the direct
object, the arguments’ identity and their syntactic roles are evident even before the target verb.

8
Nevertheless, there is little evidence to date that comprehenders’ expectations about an
upcoming verb are sensitive to this information during real-time comprehension. In fact, a
number of studies in languages such as Dutch, Mandarin Chinese, Japanese and English have
reported that reversing the preverbal arguments did not modulate comprehenders’ N400
response at the verb in pairs like (1), despite the large difference in the predictability of the
verb (Chow & Phillips, 2013; Chow, Smith, Lau, & Phillips, 2016; Hoeks, Stowe, & Doedens,
2004; Kolk et al., 2003; Oishi & Sakamoto, 2010; van Herten, Chwilla, & Kolk, 2006; van
Herten, Kolk, & Chwilla, 2005; Ye & Zhou, 2008; cf. Bornkessel-Schlesewsky et al., 2011).
These findings are unexpected and constitute a clear exception to the generalization that the
amplitude of the N400 response to a word is inversely related to its offline predictability or
cloze probability.
Relatedly, there is an additional body of studies that examine the effects of role
reversals in verb-medial sentences (e.g., ‘The hearty meal was devouring/devoured …’; Kim
& Osterhout, 2005; Kuperberg, Sitnikova, Caplan, & Holcomb, 2003; Stroud & Phillips, 2012).
We suspect that there are close parallels in the mechanisms involved, but we focus here on
cases in which the arguments’ structural roles are known prior to the verb in order to study how
argument role information impact verb predictions.
It should be noted, however, that comprehenders do not simply ignore the arguments’
structural roles when they interpret these sentences. Many have shown that comprehenders can
readily detect the implausibility that results from the reversal of the arguments’ roles (e.g.,
Kolk et al., 2003). Further, ERP studies across different languages have shown that argument
role reversals elicit a larger late positivity (a P600 effect; e.g., Kim & Osterhout, 2005; Kolk et
al., 2003; Stroud & Phillips, 2012; Ye & Zhou, 2008), which has been associated with error
detection and reanalysis (Coulson, King, & Kutas, 1998; Hagoort, Brown, & Groothusen, 1993;
Hahne & Friederici, 1999; Kuperberg, 2007; Osterhout & Holcomb, 1992; van de Meerendonk,

9
Kolk, Vissers, & Chwilla, 2010). These results suggest that information about the arguments’
structural roles can immediately impact comprehenders’ interpretation, but it does not seem to
affect the cognitive processes that underlie the N400 response.
Three kinds of accounts have been proposed to explain the N400’s surprising
insensitivity to argument role reversals. First, some researchers adopted the integration view
of the N400 (e.g., Hagoort et al., 2004) and have taken these results to indicate that
comprehenders fail to detect implausibility in role-reversed sentences at least initially
3
(“Semantic Illusion” hypothesis, e.g., Bornkessel-Schlesewsky & Schlesewsky, 2008; Hoeks
et al., 2004; Kim & Osterhout, 2005; Kolk et al., 2003; Kuperberg, 2007). For example, Kolk
et al. (2003) proposed that comprehenders have a bias to compute verb-argument relations
based on what is most plausible rather than what the syntax indicates, and thus that upon
encountering the verb they compute a plausible interpretation in both the canonical and role-
reversed sentences. Under this view, argument role reversals should not modulate the N400
amplitude regardless of the impact on the verb’s predictability, since the verb can always form
a plausible thematic relation with the arguments.
On the other hand, some have focused on the N400’s sensitivity to lexical semantic
association and have attributed the lack of an N400 effect to argument role reversals to the fact
that the verb tends to be strongly associated with its arguments in both canonical and role-
reversal sentences (Brouwer, Fitz, & Hoeks, 2012). Under this view, argument role reversals
should fail to modulate the N400 only when the verb is strongly associated with the arguments,
and they should elicit an N400 effect when the verb is only weakly associated with the
arguments.
3
These accounts typically posit that such insensitivity to implausibility is temporary, as participants’
end-of-trial plausibility judgements are highly accurate across studies.

10
Lastly, with the linking hypothesis that the N400 is modulated by the extent to which a
word’s semantic representation has been pre-activated (i.e., the prediction view), we have
proposed that the N400’s insensitivity to argument role reversals may be taken to show that
verb predictions are not always sensitive to argument role information, potentially because the
impact of argument role information on online verb prediction is not immediate (Chow &
Phillips, 2013; Chow, Smith, et al., 2016). We will call this the “Slow Prediction” hypothesis.
In the case of argument role reversals (e.g., in (1)), in order for a verb (e.g., ‘arrest’) to
become differentially pre-activated and elicit an N400 effect, comprehenders must not only
pre-activate events that likely involve cops and thieves as participants, but those that
specifically involve cops as agents and thieves as patients in one case and the opposite in the
other. Therefore, even though previous findings suggest that comprehenders can quickly (i) use
case markings on pre-verbal arguments to anticipate the properties of the verb (e.g., Kamide et
al, 2003) and (ii) use nouns to pre-activate events that typically involve them as participants
(e.g., McRae et al., 2005), it is conceivable that comprehenders may require more time to
preferentially pre-activate the events that involve those arguments in specific thematic roles.
In a recent paper, we discussed different cognitive mechanisms that may be involved
in verb prediction and discussed why argument role information may not have an immediate
impact (Chow, Momma, Smith, Lau, & Phillips, 2016). We conceptualized the process of verb
prediction as a memory retrieval problem, in which comprehenders use information about the
event participants to query their event knowledge in long-term memory in order to predictively
identify which events are most likely to be described by the sentence. Further, we proposed
that our event knowledge may not be represented in such a way that can be directly queried
using linguistically encoded argument role retrieval cues, and thus additional processes may be
required for reconfiguring the cues (or the event representations themselves) in order for likely
events to be successfully identified and retrieved from memory.

11
In fact, one common characteristic among many previous studies that looked at
argument role reversals is the close proximity between the arguments and the target verb in the
experimental materials. For example, the SOA between the second argument and the target
verb was only 480 ms in Hoeks et al. (2004), 530ms in Chow, Smith, et al. (2016), 600 ms in
Chow and Phillips (2013), and 645 ms in Kolk et al. (2003) and van Herten et al. (2005, 2006).
Therefore, it is possible that comprehenders do use argument role information to predict an
upcoming verb, but they simply cannot do so quickly enough when the verb appears
immediately following the arguments. This idea is consistent with previous findings that timing
manipulations can modulate predictability effects (e.g., Dambacher et al., 2012; Ito, Corley,
Pickering, Martin, & Nieuwland, 2016; Kutas, 1993; Wlotko & Federmeier, 2015). Under this
view, the N400’s sensitivity to argument role reversals should depend on (i) the predictability
of the verb and (ii) the amount of time available for comprehenders to incorporate information
about the arguments’ roles in their verb predictions.
One piece of suggestive evidence for the Slow Prediction hypothesis comes from a
visual-world eye-tracking study by Kukona and colleagues, who examined listeners’ eye
fixations on displays containing verb-related patients (e.g., crook) and agents (e.g., policeman)
as they listened to simple English sentences (Kukona, Fang, Aicher, Chen, & Magnuson, 2011).
The authors found that listeners were nearly as likely to look at agents as patients when they
heard the beginning of active sentences like “Toby arrests the …” (Experiment 1), but they
were more likely to look at agents when they heard the beginning of passive sentences such as
“Toby was arrested by the …” (Experiment 2). With the linking hypothesis that listeners’
anticipatory eye-movements reflect their online predictions, these results may be taken to show
that listeners’ predictions are sensitive to filled vs. unfilled thematic roles in passive sentences
but not in active sentences, potentially because they had slightly more time to predict the second
argument in passive sentences than in active sentences (due to the presence of additional words

12
‘was’ and ‘by’ in passive sentences, although with spoken stimuli the duration difference may
be small). However, as the authors acknowledged, passive and active sentences differed in
other ways (e.g., the nature and number of the syntactic cues available), and anticipatory
fixations may not be fully reflective of comprehenders’ predictions as they can also be driven
by priming (e.g., Huettig & Altmann, 2005; Yee & Sedivy, 2006). Further, the fact that the
target was always presented with a small set of distractors in the visual world paradigm may
also artificially restrict (or even interfere with) comprehenders’ predictions. Therefore, even
though these results are suggestive they should not be taken as evidence for the hypothesis that
argument role information has a delayed impact on predictions.
In the current study we tested competing accounts for the N400’s insensitivity to
argument role reversals by examining the effects of argument role reversals under different
circumstances in three ERP experiments. We used the highly frequent ba-construction in
Mandarin Chinese to create canonical and role-reversed sentences (as illustrated in 1a-b above).
In Experiments 1 and 2, argument role information was manipulated by reversing the order of
the preverbal arguments in sentences in which the verb had high vs. low predictability in the
canonical condition, and we controlled the timing of the verb relative to the arguments by
manipulating between experiments the linear distance between the arguments and the verb. The
verb immediately followed the arguments in Experiment 1 (e.g., ‘cop ba thief arrest’); a
temporal phrase was placed between the arguments and the verb to increase the linear distance
between them in Experiment 2 (e.g., ‘cop ba thief (yesterday afternoon) arrest’). In Experiment
3 we manipulated the linear distance between the arguments and the verb within the same
experiment.
Experiment 1
In this experiment we examined whether argument role reversals impact the N400
response to the verb when they have a large and demonstrable impact on the verb’s cloze

13
probability. Previous studies that examined the effects of argument role reversals did not
measure cloze probability, because they typically focused on how reversals create implausible
sentences, rather than on how they change cloze probabilities. Therefore, it is unclear to what
extent the target verb’s cloze probability was modulated by argument role reversals in those
studies. Here we explicitly measured and manipulated the impact of argument role reversals on
the target verb’s cloze probability. In high-predictability sentences the target verb had high
cloze probability in the canonical control condition; in low-predictability sentences the verb
had low cloze probability in the canonical condition. In both cases the verbs always had zero
cloze probability when the arguments were reversed. Therefore, argument role reversals
resulted in a much larger cloze probability difference between the canonical and role-reversed
conditions in the high-predictability sentences (64% vs. 0%) than in the low-predictability
sentences (7% vs. 0%). Sentences were plausible in the canonical condition and implausible in
the role-reversed condition. Since each pair of canonical and role-reversed sentences contained
identical sets of words, argument role reversals effectively isolated the contribution of the
arguments’ structural roles while holding other factors such as lexical semantic association
constant.
Therefore, if the N400 is sensitive to the verb’s cloze probability and its insensitivity to
argument role reversals is attributable to a lack of cloze probability differences in previous
studies, then argument role reversals should elicit an N400 effect when there is a clear cloze
probability difference between the canonical and role-reversed conditions (i.e., in the high-
predictability sentences). Alternatively, if argument role reversals fail to elicit an N400 effect
because (i) argument role information cannot impact comprehenders’ verb predictions quickly
enough (Slow Prediction hypothesis), or (ii) comprehenders temporarily fail to detect the
implausibility in role-reversed sentences (Semantic Illusion hypothesis), then no N400 effects
should be observed in high- and low-predictability sentences alike.

14
Methods
Participants
Twenty-four students (19 female, mean age = 21 years, range 18-24 years) from South
China Normal University participated in the current study. All participants were native
speakers of Mandarin Chinese, were strongly right-handed based on the Edinburgh Handedness
Inventory (scoring 90% or above; Oldfield, 1971), and had normal or corrected-to-normal
vision and no history of neurological disorder. None of the participants in the present study
took part in more than one of the experiments. All participants gave informed consent and were
paid 20 RMB/hour for their participation.
Cloze probability norming
Cloze probability estimates were gathered from 60 student volunteers at South China
Normal University on a total of 190 pairs of subject-ba-object sentence frames (e.g., ‘cop ba
thief …’ and its role-reversed counterpart ‘thief ba cop …’). None of these participants took
part in any of the ERP experiments. The sentence frames were divided into two lists, and each
list was completed by 30 participants. Participants were asked to provide the best continuations
for the sentence frames with one or more words. Overwhelmingly participants responded with
a verb which completed the sentence. A few exceptions were found when participants extended
the second NP before supplying a verb (e.g., assistant ba manager’s speech recorded) or
preceded the verb with a preposition phrase (e.g., magician ba pigeon from a hat pulled out);
we always used the first word following the sentence fragment to compute cloze counts.
Materials
Across Experiments 1 to 3, all experimental sentences used the highly frequent SOV
ba-construction in Mandarin Chinese. Role-reversed sentences were created by reversing the
order of the pre-verbal arguments in the canonical sentences. Within each item set the canonical

15
and role-reversed sentences had an identical verb-argument triplet and differed only in the order
of the arguments. Both of the arguments were animate and the order of the arguments alone
did not induce any anomaly prior to the target verb. Further, in order to avoid sentence-final
wrap-up effects the sentences were extended beyond the target verb with words that were held
constant across conditions within each item set.
The experimental stimuli consisted of 120 pairs of sentences, each with a canonical and
reversed argument order. Argument order was manipulated in two groups of sentences (Table
1). In the high-predictability sentences the target verb had an average cloze probability of 64%
(range 41% - 97%) in the canonical condition; in the low-predictability sentences the verb had
an average cloze probability of 7% (range 3% - 21%) in the canonical condition. In both cases
the verbs always had zero cloze probability when the arguments were reversed.
4
High- and
low-predictability sentences shared the same SOV ba-construction but contained different
arguments and verbs. All sentences were plausible in the canonical conditions and implausible
in the role-reversed conditions. A complete set of the experimental materials from the present
study (Experiments 1 to 3) is available in the Supplementary Materials.
In addition, across Experiments 1 to 3, 60 plausible filler sentences from a previous
study were included to examine the effects of a standard cloze probability manipulation in the
populations studied here (a control comparison; bottom of Table 1). In each item, the same
target word was paired with two different sentence contexts, such that the target word had a
high cloze probability (M = 88%, SD = 11%) in one and a low cloze probability (M = 27%,
SD = 13%) in the other. A group of 48 participants were asked to rate the plausibility of these
sentences on a 5-point scale. Both high and low cloze sentences were rated as highly plausible
4
In Experiment 1, average sentence constraint in the canonical condition was 64% in the high-
predictability sentences and 22% in the low-predictability sentences; in the role-reversed condition,
average sentence constraint was 25% and 22% in high- and low-predictability sentences respectively.
The target verb was the most predictable word in the canonical condition in all of the high predictability
sentences and only some of the low predictability sentences.

16
(high cloze: 4.3; low cloze: 4.1). The position of the target word was matched across conditions
in each set. As was the case in many previous studies, the target word was preceded by
completely different contexts across conditions. Therefore, all of the words in the sentence
context could have contributed to the differential predictability of the target word in these
sentences. This critically differed from the experimental sentences, in which the target verb
was preceded by sentence contexts that only differed by the order of the arguments across
conditions.
Table 1. Experimental conditions and sample materials in Experiment 1 (top) and sample materials for
the control comparison in Experiments 1 to 3 (bottom).
Experimental comparison
Low-predictability
Canonical
Role-reversed
[
老刘 把 鹦鹉
Mr. Liu BA parrot
鹦鹉 把 老刘
Parrot BA Mr. Liu
]
训练了 好一段时间 。
train quite some time.
“Mr. Liu trained the parrot for quite some time.”
vs. “The parrot trained Mr. Liu for quite some time.”
High-predictability
Canonical
Role-reversed
[
警察 把 小偷
cop BA thief
小偷 把 警察
thief BA cop
]
抓了 回警局。
arrest (and bring back) to police station.
“The cop arrested the thief (and brought him back) to the station.”
vs. “The thief arrested the cop (and brought him back) to the station.”
Control comparison
(Identical across Experiments 1-3)
Expected
牧民 从 羊身上 剪下 很多 毛…
Herdsman from sheep’s body cut-down a lot of wool…
“From the sheep the herdsman collected a lot of wool…”
Unexpected
老板 在 去年 收购 很多 毛…
Boss ZAI last year buy a lot of wool…
“Last year the boss bought a lot of wool…”
Experimental sentences were distributed across two presentation lists, such that only
one version of each item appeared in each list. Each list contained 120 experimental sentences

17
(30 per condition), 60 plausible filler sentences for the control comparison and 60 unrelated
implausible filler sentences of similar length and structural complexity, so that the overall
plausible-to-implausible ratio in each presentation list was 1:1, in line with many previous ERP
studies (e.g., Hoeks et al., 2004; Kim & Osterhout, 2005).
Procedure
Participants were comfortably seated about 100cm in front of a computer screen in a
testing room. Sentences were segmented into words (see example in (2)), which were presented
one at a time in a white font (30 pt simplified Chinese characters) on a black background at the
centre of the screen. Each sentence was preceded by a fixation cross that appeared for 500ms.
Each word appeared on the screen for 400ms, followed by 200ms of blank screen. Each word
consisted of 1 to 4 characters. The last word of each sentence was marked with a period,
followed 1000ms later by a response cue “?”. Participants were instructed to avoid eye blinks
and movements during the presentation of the sentences, and they were asked to read each
sentence attentively and to indicate whether the sentence meaning was plausible by pressing
one of two buttons. Prior to the experimental session, participants were presented with 6
practice trials with feedback to familiarize themselves with the task. The experimental session
was divided into four blocks of 60 sentences each, with short pauses in between. Including set-
up time, an experimental session lasted around 1.5 hours on average.
(2) Sentence segmentation for stimulus presentation:
/去年初， /老刘/把/鹦鹉/训练了/好一段/时间 。/
/Beginning of last year,/Mr. Liu/BA/parrot/train/(for) quite some/time./
EEG Recording
EEG was recorded continuously from 30 AgCl electrodes mounted in an electrode cap
(Electrocap International): midline: Fz, FCz, Cz, CPz, Pz, Oz; lateral: FP1/2, F3/4, F7/8, FC3/4,

18
FT7/8, C3/4, T7/8, CP3/4, TP7/8, P3/4, P7/8, and O1/2. The electro-oculogram (EOG) was
recorded at four electrode sites; vertical EOG was recorded from electrodes placed above and
below the left eye and the horizontal EOG was recorded from electrodes situated at the outer
canthus of each eye. Electrode impedances were kept below 5kΩ. The EEG and EOG
recordings were amplified (band-pass filtered at DC - 200 Hz) and digitized online at 1kHz.
Online recordings were referenced to the left mastoid. They were re-referenced to the average
of the left and right mastoids and filtered using a 0.1 - 40 Hz band-pass filter offline. This high-
pass filter value was tested in a recent study and found to not introduce artefactual effects
(Tanner, Morgan-Short, & Luck, 2015). A 15Hz low-pass filter was applied for the presentation
of grand averages only, but was not used in the analyses.
ERP Data Analysis
Event-related potentials were computed separately for each participant and each
condition for the 1000ms after the onset of the target verb relative to a 100ms baseline
preceding the target verb. ERP data in all trials (regardless of end-of-trial behavioural response)
were evaluated for EOG and other artefacts. Trials contaminated by artefacts were excluded
from the averaging procedure. This affected 7.8% of experimental trials, roughly equally
distributed across conditions (ranging between 6.3 and 9.6% across conditions).
Statistical analyses on average voltage amplitudes were conducted separately for two
time windows chosen based on previous literature and on visual inspection of the data: 350-
450 ms for the N400, and 600-800ms for the P600.
Data from the experimental sentences were analysed using two parallel approaches. As
our hypotheses centred on the effect of role-reversal on the N400 and P600, our primary
analysis examined the effect of role reversal on these ERP components within each level of
predictability. Since both N400 and P600 effects tend to be broadly distributed with a
(central-)posterior maximum, we chose a cluster of posterior electrodes for this analysis so as

19
to use the same electrodes to capture both N400 and P600 effects (see also Ito et al., 2016).
Further, in order to allow a more comprehensive evaluation of effects across the scalp, we also
conducted omnibus ANOVAs at each time window to test for main effects and interactions
with distributional factors.
For the posterior electrode cluster analyses, we averaged participants’ ERPs across 6
posterior electrodes (P3, PZ, P4, O1, OZ, and O2) and ran paired-sample t-tests to examine the
effect of role reversal within each level of predictability in each time window. Further, we used
the bootES package (Gerlanc & Kirby, 2013) to compute standardized effect sizes of these
contrasts (Pearson’s r) and their bootstrap confidence intervals (95% CI). An advantage of
using Pearson’s r to express the effect size of a contrast is that the effect is scaled to have an
absolute magnitude between 0 and 1.
For the omnibus repeated measures ANOVAs, we examined time-window average
ERPs at 18 electrodes, fully crossing predictability (high vs. low) and reversal (congruous vs.
role-reversed) with anteriority (anterior vs. central vs. posterior) and laterality (Left vs. Midline
vs. Right). We assigned topographical grouping factors to electrodes as follows: left-anterior:
F3, FC3; midline-anterior: FZ, FCZ; right-anterior: F4, FC4; left-central: C3, CP3; midline-
central: CZ, CPZ; right-central: C4, CP4; left-posterior: P3, O1; midline-posterior: PZ, OZ;
right-posterior: P4, O2. Univariate F-tests with more than one degree of freedom in the
numerator were adjusted by means of the Greenhouse-Geisser correction (Greenhouse &
Geisser, 1959). Since predictability was manipulated between different item sets (i.e., high and
low-predictability sentences were completely different), main effects of predictability are not
meaningful and we discuss effects of predictability only when they interact with the effects of
reversal. We also conducted a predictability × anteriority × laterality repeated measures
ANOVA in each time window for the filler sentences to examine the effect of cloze probability.

20
Results
Plausibility Judgments
Across both high- and low-predictability sentences, participants judged canonical
sentences to be plausible and role-reversed sentences to be implausible with an overall accuracy
of 90.6% (Low predictability, canonical = 88.3%; Low predictability, role-reversed = 90.4%;
High predictability, canonical = 90.3%; High predictability, role-reversed = 93.3%).
Event-related Potentials (Experimental comparisons)
The top half of Figure 1 shows the grand average ERPs at centro-posterior electrode
CPZ and the topographic distribution of the effects of argument role reversals at the target word
in the 350-450 ms and 600-800 ms time intervals in the low-predictability and high-
predictability conditions in Experiment 1. Results of the posterior electrode cluster analyses
and the omnibus ANOVAs are presented in Table 2 and the leftmost column of Table 3
respectively. Grand average ERPs across all 30 electrode sites are presented in Supplementary
Figures 1 and 2.
As shown in Figure 1, argument role reversals elicited a clear P600 effect with the same
topographic distribution across both high- and low-predictability sentences, indicating that
participants quickly and reliably detected that the verb was anomalous in the role-reversed
conditions. However, there was no clear effect of reversal on the N400 in either pair of
conditions, particularly in the high-predictability conditions where the cloze probability
difference was largest.
In the 350-450 ms interval, the posterior electrode cluster analyses revealed no
significant effects of reversal in either high- or low-predictability conditions. The omnibus
ANOVA revealed a three-way reversal × anteriority × laterality interaction, but further analyses
by anteriority and laterality did not reveal any significant effect.

21
In the 600-800 ms interval, the posterior electrode cluster analyses showed a larger
P600 in the role-reversed condition than in the canonical condition in high- and low-
predictability sentences alike. The omnibus ANOVA similarly revealed a main effect of
reversal, along with reversal × anteriority, reversal × laterality, and reversal × anteriority ×
laterality interactions, reflecting a larger central-posterior positivity (P600) for the role-
reversed condition in high- and low-predictability sentences alike.
Event-related Potentials (Control comparison)
The bottom half of Figure 1 shows the grand average ERPs at frontal electrode FZ and
at centro-posterior electrode CPZ and the topographic distribution of ERP effects in the control
comparison. Results of the statistical analyses are presented in the leftmost column of Table 4.
The standard cloze probability manipulation in the filler sentences elicited a clear N400
effect and a frontally distributed late positivity. Statistical analyses in the 350-450 ms time
interval revealed a significant main effect of predictability, showing that expected words elicit
a smaller N400 response than unexpected words. Interactions between predictability and the
topographic factors indicated that the effect was largest at midline central and posterior sites.
Analyses in the 600-800 ms time interval revealed a significant predictability × anteriority
interaction, due to the fact that unexpected words elicited a larger frontal positivity than
expected words.
As shown in Figure 1, the topographic distribution of this late positivity was distinct
from the posteriorly distributed P600 effect in the experimental sentences. This frontally
distributed late positivity has been observed in previous studies and has been argued to reflect
processes that are triggered when strong predictions are disconfirmed (DeLong, Quante, &
Kutas, 2014; Federmeier, Wlotko, De Ochoa-Dewald, & Kutas, 2007; Thornhill & van Petten,
2012; van Petten & Luka, 2012). We will return to discuss the relationship between this late
frontal positivity and the P600 in the General Discussion.

22
Figure 1. Top: Grand average ERPs at centro-parietal electrode CPZ and the topographic distribution
of ERP effects in the 350-450 ms and 600-800ms intervals (role-reversed minus canonical) in the low-
predictability (left) and high-predictability (right) conditions in Experiment 1, in which the target verb
appeared immediately following the second argument (600ms SOA). Argument role reversals elicited
a P600 effect in both low-predictability and high-predictability sentences, but the N400 was not
modulated by the predictability of the target verb in either condition. Bottom: Grand average ERPs at
frontal electrode FZ and at centro-posterior electrode CPZ and topographic distribution of ERP effects
in the 350-450 ms and 600-800ms intervals (expected minus unexpected) in the control comparison in
Experiment 1. A cloze probability manipulation in the filler sentences elicited a clear N400 effect
followed by a frontally distributed late positivity.

23
Table 2. Results of the posterior electrode cluster analyses in Experiment 1: t value, standardized
effect size (Pearson’s r) and 95% confidence interval for the contrast between the canonical and role-
reversed conditions in each time interval.
Table 3. Omnibus ANOVA F-values at the target verb in Experiments 1 to 3.
Table 4. Omnibus ANOVA F-values at the target word for the control comparisons in Experiments 1-
3.
Experiment 1 df t
Pearson's r95% CI
Low predictability
350-450ms 23 1.12 -0.23 [-0.55, 0.18]
600-800ms 23 3.43** 0.58 [0.24, 0.76]
High predictability
350-450ms 23 < 1 -0.07 [-0.42, 0.34]
600-800ms 23 6.4** 0.80 [0.66, 0.87]
** p < .01; * p < .05; ^ .05 < p <.1
Experiment 1 Experiment 2 Experiment 3
Omnibus ANOVA df 350-450ms 600-800ms df 350-450ms 600-800ms df 350-450ms 600-800ms
rev 1,23 < 1 18.78** 1,23 1.37 19.01** 1,23 < 1 17.95**
rev * type 1,23 < 1 < 1 1,23 1.01 < 1 1,23 < 1 < 1
rev * ant 2,46 1.24 5.97* 2,46 < 1 29.63** 2,46 < 1 13.9**
rev * lat 2,46 < 1 4.19* 2,46 < 1 < 1 2,46 < 1 2.36
rev * type * ant 2,46 < 1 1.03 2,46 < 1 < 1 2,46 < 1 < 1
rev * type * lat 2,46 < 1 1.94 2,46 < 1 < 1 2,46 5.68** 1.25
rev * ant * lat 4,92 3.66* 4.89** 4,92 1.74 1.47 4,92 < 1 1.25
rev * type * ant * lat 4,92 < 1 1.2 4,92 1.35 < 1 4,92 1.08 < 1
Factors: type = predictability (Experiments 1 & 2) or distance (Experiment 3); ant = anteriority; lat = laterality.
** p < .01; * p < .05 ; ^ .05 < p <.1
Experiment 1 Experiment 2 Experiment 3
df 350-450ms 600-800ms df 350-450ms 600-800ms df 350-450ms 600-800ms
Omnibus ANOVA
pred 1,23 20.3** 2.25 1,23 45.8** 2.09 1,23 6.93* 3.11^
pred * ant 2,46 7.21** 5.18* 2,46 9.02** 4.26* 2,46 10.44** 8.69**
pred * lat 2,46 3.56* < 1 2,46 < 1 1.68 2,46 < 1 < 1
pred * ant * lat 4,92 2.91* 1.67 4,92 3.51* 4.13** 4,92 3.07* 5**
Factors: pred = predictability; ant = anteriority; lat = laterality.
** p < .01; * p < .05 ; ^ .05 < p <.1

24
Discussion
This experiment showed that, when a verb appeared immediately following its
arguments, argument role reversals did not modulate comprehenders’ N400 response to the
verb even when they greatly impacted its offline predictability, as quantified by cloze
probability. The N400’s insensitivity to the cloze probability difference resulting from
argument role reversals is striking, especially in the context of the P600’s sensitivity to
argument role reversals as well as the N400’s sensitivity to cloze probability in the control
comparison. This finding is in line with the absence of N400 effects in previous studies that
examined the effect of pre-verbal argument role reversals across different languages (e.g.,
Chow & Phillips, 2013; Chow, Smith, et al., 2016; Hoeks et al., 2004; Kolk et al., 2003; Oishi
& Sakamoto, 2010; van Herten et al., 2006, 2005; Ye & Zhou, 2008), although the effect of
argument role reversals on the verb’s off-line predictability was quantified in only one of these
studies (Chow, Smith, et al., 2016). Importantly, the observation in the current study that role
reversals do not impact the N400, even when they have a large impact on cloze probability,
suggests that the N400’s insensitivity is not attributable to a weak experimental manipulation.
This is a clear exception to the well-established generalization that the N400 is modulated by
a word’s cloze probability.
Meanwhile, as in previous studies, participants were highly accurate in their plausibility
judgments and they showed a clear P600 effect to argument role reversals. We extended
previous findings by showing that the effect of argument role reversal on the P600 is not
modulated by the (un)predictability of the verb. This robust sensitivity suggests that
comprehenders computed an accurate interpretation of the sentences and readily detected the
implausibility resulting from argument role reversals.
Different aspects of the current results allow us to evaluate competing interpretations
of this finding. First, the participants in the current study were unlikely to be processing less

25
deeply or predictively than previous studies showing N400 cloze probability effects because
they displayed standard N400 sensitivity to a word’s cloze probability in the filler sentences.
5
Further, since the average N400 at the targets in the high-predictability sentences was clearly
larger than that elicited by the expected words in the filler sentences, we consider it unlikely
that argument role reversals failed to modulate the N400 simply because the lexical association
between the verb and the arguments was so strong as to yield a floor effect in the N400
amplitude (although we cannot conclusively rule this out).
However, the present results are still compatible with two competing hypotheses. On
the one hand, these results are compatible with the Slow Prediction hypothesis as the N400’s
insensitivity to role reversals may show that argument role information cannot impact
comprehenders’ verb predictions quickly enough. Meanwhile, they also cannot rule out the
Semantic Illusion hypothesis, according to which the N400’s insensitivity reflects a brief initial
stage when comprehenders failed to detect implausibility (e.g., Bornkessel-Schlesewsky &
Schlesewsky, 2008; Hoeks et al., 2004; Kim & Osterhout, 2005; Kolk et al., 2003; Kuperberg,
2007). We address this possibility more directly in Experiments 2 and 3.
Experiment 2
In Experiment 2 we contrasted the Slow Prediction hypothesis with the Semantic
Illusion hypothesis. To this end, we employed the same experimental design as Experiment 1
and increased the distance between the arguments and the verb. This was achieved by placing
an adverbial time expression (e.g., “shangxingqi”, last week) along with the preposition “zai”
between the second argument and the verb. As sentence fragments were presented at a fixed
5
Note that this result does not indicate that comprehenders computed predictions more quickly in the filler
sentences than in the experimental sentences. As in previous studies on the effects of cloze probability, the cloze
probability manipulation in the filler sentences did not allow precise estimation of the delay between predictive
context and predicted target because the same target words were presented in completely different sentence
contexts across conditions.

26
rate (600ms SOA), this effectively delayed the presentation of the verb relative to its arguments
by 1200 ms. As in Experiment 1, each pair of canonical and role-reversed sentences contained
identical sets of words and differed only in the order of the pre-verbal arguments.
Under the Slow Prediction hypothesis, if information about the arguments’ structural
roles can impact comprehenders' verb prediction within the extended time interval, then
argument role reversals should elicit an N400 effect in the high-predictability conditions in this
experiment. Even though argument role reversals also make the sentence implausible in the
low-predictability conditions, we do not expect to observe an N400 effect there because role
reversals does not give rise to a clear difference in the verb’s predictability low-predictability
sentences. Alternatively, if comprehenders simply do not incorporate information about the
arguments’ structural roles into their verb prediction (or if they fail to do so within the extended
time interval), then the N400 should remain insensitive to argument role reversals. Lastly, since
both the high- and low-predictability conditions compared plausible and implausible sentences,
the Semantic Illusion hypothesis would predict the same pattern of results in the high- and low-
predictability conditions in this experiment. Either both should show no N400 effects (if the
heuristics is in play) or both should show an N400 effect (if the more careful algorithm has had
time to compute).
Methods
Participants
Twenty-four students (19 female, mean age = 19.9 years, range 18-24 years) from South
China Normal University participated in the current study. Informed consent was obtained
from all participants. All were right-handed, native Mandarin Chinese speakers with normal or
corrected-to-normal vision and were paid for their participation.

27
Materials
The current experiment shared the same predictability × reversal design as Experiment
1. The experimental materials were adapted directly from those in Experiment 1 to increase the
linear distance between the target verb and its arguments. A temporal adverbial phrase (e.g.,
“zai shangxingqi”, last week) always appeared between the second argument and the verb (e.g.,
cop BA thief ZAI last week arrest…), thereby increasing the SOA between them from 600ms
to 1800ms. We favoured this over an overall increase in the SOA because a small increase in
the overall SOA (e.g., to 700ms) might not provide sufficient time for argument role
information to impact comprehenders’ verb predictions before the verb appears, whereas a
larger SOA increase (e.g., to 1000ms) would force participants to read unnaturally slowly and
could place excessive demands on their working memory. By contrast, introducing a filler
phrase between the second argument and the verb can easily add hundreds of milliseconds to
the SOA between them without causing participants to read any more slowly.
The choice of filler phrases was based on two considerations. On the one hand, the
naturalness of this word order is sensitive to the “heaviness” (or length) of the adverbial phrase,
such that sentences like “Jingcha ba xiaotou zai zuotian-xiawu zhua-le…” (cop BA thief ZAI
yesterday afternoon arrest…) would sound more natural than "Jingcha ba xiaotou zai zuotian
zhua-le" (cop BA thief ZAI yesterday arrest…). On the other hand, introducing a long and
complex adverbial might prevent comprehenders from further processing the main clause
during the added time interval. Therefore, we chose to use simple time adverbials that are 4 to
5 characters long here. All sentences were checked by two native Mandarin-speaking
informants to ensure their naturalness prior to the study, and plausibility rating data from 12
speakers who did not participate in any of the present experiments showed that canonical
sentences were rated high in plausibility in both short-distance (6.4 out of 7) and long-distance
(5.7 out of 7) conditions. Although temporal phrases more commonly appear at the beginning

28
of a sentence or following the subject in Mandarin Chinese, it is important to note that
comparisons were always made between canonical and role-reversed sentences with identical
temporal phrase placement, and so any differences between conditions are unlikely to be due
to constraints on adverb placement.
In order to verify that the adverbial phrases did not make the target verbs more
predictable, we obtained cloze probability estimates for the target verbs in these adapted
sentences from 60 additional students from South China Normal University who did not
participate in any of the ERP experiments. The cloze probabilities for the Experiment 2 items
closely resembled the cloze probabilities for Experiment 1 items. In Experiment 2 the target
verb had an average cloze probability of 57.1% (range = 20% to 100%) and 8.7% (range = 0%
to 40%)
6
in the canonical conditions in the high- and low-predictability sentences respectively.
In both cases the verbs’ cloze probability was close to zero when the arguments were reversed
(high-predictability sentences: 1.9%; low predictability sentences: 0.4%). This confirms that
the temporal adverbials did not themselves affect the cloze probability of the target verbs.
7
As in Experiment 1, the experimental sentences were distributed in two presentation
lists of 240 items. Each list contained 120 experimental sentences (30 per condition) and the
same set of 120 filler sentences used in Experiment 1.
6
Since these sentences were adapted from those used in Experiment 1, some target words in the low
predictability sentences (which had very low cloze probabilities in Experiment 1) may have a zero cloze
probability simply due to sampling error.
7
Sentence constraint were also highly similar between items in Experiments 1 and 2. In Experiment 2,
average sentence constraint (cloze probability of the most likely continuation) in the canonical condition
was 58% in the high-predictability sentences and 21% in the low-predictability sentences; in the role-
reversed condition, average sentence constraint was 26% and 22% in high-predictability and low-
predictability sentences respectively.

29
Procedure, EEG Recording and ERP Data Analysis
The procedures for the experiment, EEG recording and data analysis were identical to
Experiment 1. A total of 11.5% of experimental trials, roughly equally distributed across
conditions (9% - 13.6%), were excluded from the averaging procedure due to artefacts.
Results
Plausibility Judgments
Participants reliably judged canonical sentences to be plausible and the role-reversed
sentences to be implausible with an overall accuracy of 90.9% (Low predictability, canonical
= 88.0%; Low predictability, role-reversed = 89.5%; High predictability, canonical = 90.7%;
High predictability, role-reversed = 95.1%).
Event-related Potentials (Experimental comparisons)
The top half of Figure 2 shows the grand average ERPs at centro-parietal electrode CPZ
and the topographic distribution of the ERP effects in the 350-450 ms and 600-800 ms time
intervals in the low-predictability and high-predictability conditions in Experiment 2. Results
of the posterior cluster analyses and the omnibus ANOVAs are presented in Table 5 and the
middle column of Table 3 respectively. Grand average ERPs across all 30 electrode sites are
presented in Supplementary Figures 3 and 4.
As in Experiment 1, a clear P600 effect to argument role reversals was observed in both
high- and low-predictability conditions, indicating that participants detected the anomaly in the
role-reversed sentences. However, in contrast to Experiment 1, argument role reversals in the
high-predictability conditions appeared to also elicit an N400 effect.
In the 350-450 ms interval, the posterior electrode cluster analyses showed that
argument role reversals elicited a significant negativity (an N400 effect) in the high-
predictability conditions but not in the low-predictability conditions. The main effect of

30
reversal and the interaction between reversal and predictability failed to reach statistical
significance in the omnibus ANOVA, likely because the size of the effect was not very large
even in the high-predictability conditions (~0.7µV; r = 0.42).
In the 600-800 ms interval, the posterior electrode cluster analyses revealed a larger
P600 in the role-reversed condition than in the canonical condition in high- and low-
predictability sentences alike. This is in line with the results of the omnibus ANOVA, which
revealed a main effect of reversal and a reversal × anteriority interaction.
Event-related Potentials (Control comparison)
The bottom half of Figure 2 shows the grand average ERPs and the topographic
distribution of ERP effects in the control comparison. Results of the statistical analyses are
presented in the middle column of Table 4.
As in Experiment 1, the cloze probability manipulation in the filler sentences elicited a
clear N400 effect and a frontally distributed late positivity. Statistical analyses in the 350-450
ms time interval revealed a main effect of predictability and predictability × anteriority and
predictability × anteriority × laterality interactions, due to the fact that expected words elicited
a smaller N400 than unexpected words. Analyses in the 600-800 ms time interval revealed
predictability × anteriority and predictability × anteriority × laterality interactions, due to the
fact that unexpected words elicited a larger frontal positivity than expected words.

31
Figure 2. Top: Grand average ERPs at centro-posterior electrode CPZ and topographic distribution of
ERP effects in the 350-450 ms and 600-800ms intervals (role-reversed minus canonical) in the low-
predictability (left) and high-predictability (right) conditions in Experiment 2, in which the target verb
appeared 1800ms after the onset of the second argument. Argument role reversals elicited a P600 effect
in the low-predictability conditions, and an N400 effect followed by a P600 effect in the high-
predictability conditions. Bottom: Grand average ERPs at frontal electrode FZ and at centro-posterior
electrode CPZ and topographic distribution of ERP effects in the 350-450 ms and 600-800ms intervals
(expected minus unexpected) in the control comparison in Experiment 2. A cloze probability
manipulation in the filler sentences elicited a clear N400 effect followed by a frontally distributed late
positivity.

32
Table 5. Results of the posterior electrode cluster analyses in Experiment 2: t value, standardized effect
size (Pearson’s r) and 95% confidence interval for the contrast between the canonical and role-reversed
conditions in each time interval.
Discussion
In Experiment 2 we observed evidence that delaying the time between arguments and
verbs by inserting a neutral adverbial phrase resulted in a significant argument role reversal
N400 effect in the high-predictability conditions; this was not observed in the low-
predictability conditions. Other aspects of the results remained the same as in Experiment 1.
First, as in Experiment 1 argument role reversals were also readily detected and they elicited a
clear P600 effect that did not differ between high- and low-predictability sentences. Second,
the cloze probability manipulation in the filler sentences once again elicited a clear N400 effect
followed by a late frontal positivity.
The most important finding in this experiment is the presence of an N400 effect in the
high-predictability conditions but not in the low-predictability conditions. If these results are
reliable, they would be incompatible with Semantic Illusion hypothesis, according to which
comprehenders fail to detect implausibility in role-reversed sentences upon encountering the
verb because they assign argument roles according to what is most plausible (e.g., Bornkessel-
Schlesewsky & Schlesewsky, 2008; Hoeks et al., 2004; Kim & Osterhout, 2005; Kolk et al.,
2003; Kuperberg, 2007). This is because the verb could always form a plausible thematic
relation with the arguments in the experimental sentences regardless of its predictability, and
thus if heuristics is at work then comprehenders should be able to compute a plausible
Experiment 2 df t
Pearson's r95% CI
Low predictability
350-450ms 23 < 1 0.01 [-0.42, 0.38]
600-800ms 23 4.53** 0.69 [0.48, 0.81]
High predictability
350-450ms 23 2.24* -0.42 [-0.65, 0.06]
600-800ms 23 4.5** 0.68 [0.44, 0.82]
** p < .01; * p < .05; ^ .05 < p <.1

33
interpretation and show no N400 effect in both conditions, just as in Experiment 1. It is possible
that participants were sensitive to differences in how strongly (or weakly) constraining the
sentence contexts were prior to the verb, but such early detection should only facilitate
detection of the implausibility introduced by the verb and would not explain why the N400
effect emerged only when the verb was delayed. Further, these results would also be at odds
with the hypothesis that comprehenders simply do not use argument role information in
generating verb predictions. Instead, they would be in line with the Slow Prediction hypothesis
and suggest that the N400 was insensitive to role-reversals in previous studies because
comprehenders did not have sufficient time to incorporate information about the arguments’
structural roles into their verb predictions.
Thus far our interpretation of these results is tentative. Even though the adverbial phrase
was added in this experiment to increase the amount of time available for predictive
computations and it did not affect either the cloze probability of the verbs or the constraint of
the sentences, a within-participants comparison would allow a more direct test of the Slow
Prediction hypothesis. Therefore, we manipulated the distance between the arguments and the
verb within the same items and a single group of participants in Experiment 3.
Experiment 3
In this experiment we more directly tested the Slow Prediction hypothesis by combining
aspects of Experiments 1 and 2 in a single design. As illustrated in Figure 3, in this experiment
we varied the position of the adverbial time expression (e.g., “shangxingqi”, last week) within
the same set of materials. In the short-distance conditions, the time expression appeared at the
beginning of the sentence and the SOA between the second argument and the verb was 600 ms,
as in Experiment 1. Meanwhile, in the long-distance conditions, the time expression, along
with the preposition “zai,” appeared between the second argument and the target verb and the
SOA between them became 1800 ms, as in Experiment 2. Since the sentences in the short- and

34
long-distance conditions were synonymous and differed only in the location of the temporal
phrase, they should provide qualitatively similar contextual information for computing verb
predictions. Under the Slow Prediction hypothesis, if argument role information can impact
comprehenders' predictions for the verb within the extended time interval between the
arguments and the verb, then the N400 should become sensitive to argument role reversals
when the verb is further away from its arguments, i.e., in the long-distance conditions. Further,
this experiment also provides another test of the Semantic Illusion hypothesis. Since the verb
can form a plausible thematic relation with the arguments regardless of the linear distance
between them, this hypothesis predicts no N400 effects to argument role reversals in either
short- or long-distance conditions.
Figure 3. Distance manipulation in Experiment 3. Linear distance between the arguments and the verb
was manipulated by placing a temporal phrase (e.g., last week) either at the beginning of the sentence
(short-distance; 600ms SOA between the second argument and the verb) or between the direct object
and the verb (long-distance; 1800ms SOA).
As the goal of Experiment 3 was to examine how time impacts the effect of predictive
argument role information, we only included sentences in which argument role reversals had a
clear impact on the verb’s cloze probability, similar to the high-predictability conditions in
Experiments 1 and 2. However, in order to reduce the likelihood of a floor effect on the N400,
we tried to avoid strong lexical semantic associations between the verb and its arguments. To
our knowledge no lexical association norms (or resources like Latent Semantic Analysis,

35
Landauer & Dumais, 1997) are currently available in Mandarin Chinese, so we took high cloze
probability values as a proxy for strong lexical associations and excluded sentences in which
the verb had a very high cloze probability (> 70%) in the canonical condition.
Methods
Participants
Twenty-four students (22 female, mean age = 22 years, range 19-28 years) from South
China Normal University participated in the current study. Informed consent was obtained from
all participants. All were right-handed, native Mandarin Chinese speakers with normal or
corrected-to-normal vision and were paid for their participation.
Materials
The stimuli consisted of 120 sets of sentences in the ba-construction in Mandarin
Chinese. Argument order was manipulated in sentences in which the verb was closer vs. further
away from the arguments, fully crossing distance (short vs. long distance) and reversal
(canonical vs. role-reversed). Since the adverbial time expression (e.g., “shangxingqi”, last
week) appeared either at the start of the sentence (the short-distance conditions), or between
the direct object and the verb along with the preposition “zai” (the long-distance conditions),
the sentence contained the same content words before the verb in all four conditions, and the
conditions differed only in the order of presentation. This effectively manipulated the linear
distance, and thereby the amount of time, between the arguments and the verb without
qualitatively changing the contextual information for computing verb predictions.
With the exception of items in which the verb had a cloze probability greater than 70%
in the canonical condition, the stimuli largely overlapped with those in the high-predictability
conditions in Experiments 1 and 2. As in Experiment 2, we gathered additional cloze
probability norming data in order to verify that argument role reversals had a clear impact on

36
the verb’s cloze probability in short- and long-distance conditions alike and that placing the
adverbial phrase after the arguments did not make the target verbs any more predictable. These
were confirmed by the cloze data, which showed that there was a clear cloze probability
difference between canonical and role-reversed sentences in both short-distance and long-
distance conditions. In short-distance sentences, the target verb has an average target cloze
probability of 37.4% (range = 16.7% - 69%) in the canonical condition and always had 0%
cloze in the role-reversed condition. In the long-distance sentences, average cloze probability
was 31.0% (range = 0 - 86.7%) in the canonical condition and 0.4% (range = 0 - 13.3%) in the
role-reversed condition.
8
Crucially, the difference in the verb’s cloze probability between
canonical and role-reversed sentences was smaller (t(119) = 4.09, p < 0.01), not larger, in the
long-distance conditions than in the short-distance conditions. In other words, if this cloze
probability difference between the short- and long-distance conditions is meaningful, it actually
creates a bias against our prediction that role reversals should have a bigger impact on the N400
in the long-distance conditions.
9
Experimental sentences were distributed across four presentation lists, such that only
one version of each item appeared in each list. Each list contained 120 experimental sentences
(30 per condition) and the same set of 120 filler sentences used in Experiments 1 and 2. The
sentences were presented in four blocks of 60 sentences each, and the order of the blocks was
randomized across participants.
8
Since the materials were developed by first selecting the short-distance sentences with a suitable cloze
probability profile and then adapting those sentences to create the long-distance versions, in some cases
the target word had a lower cloze probability in the long-distance condition simply because a synonym
(or a closely related word) was more frequent in the cloze responses in the long-distance sentences.
9
As in Experiments 1 and 2, sentence constraint was not affected by the distance manipulation in
Experiment 3. In Experiment 3, average sentence constraint in the canonical condition was 37% and
35% in the short- and long-distance sentences respectively; in the role-reversed condition, average
sentence constraint was 27% in both short- and long-distance sentences. The target word was the most
predictable word in the canonical condition in the majority of the items.

37
Procedure, EEG Recording and ERP Data Analysis
The procedures for the experiment, EEG recording and data analysis were identical to
those used in Experiments 1 and 2, with distance taking the place of predictability in the
analyses. Although distance was manipulated within the same item sets in the current
experiment, its effects are nonetheless not meaningful unless they interact with the effects of
reversal. Therefore, only effects involving reversal are discussed in the main text. A total of
6.6% of experimental trials, roughly equally distributed across conditions (5% - 8.3%), were
excluded from the averaging procedure due to artefacts.
Results
Plausibility Judgments
Participants reliably judged canonical sentences to be plausible and the role-reversed
sentences to be implausible with an overall accuracy of 89.5% (Short-distance, canonical =
93.4%; Short-distance, role-reversed = 87.5%; Long-distance, canonical = 89.9%; Long-
distance, role-reversed = 86.7%).
Event-related Potentials (Experimental comparisons)
The top half of Figure 4 shows the grand average ERPs at CPZ and the topographic
distribution of the effects of argument role reversals at the target word in the 350-450 ms and
600-800 ms time intervals in the short-distance and long-distance conditions in Experiment 3.
Results of the posterior electrode cluster analyses and the omnibus ANOVA are presented in
Table 6 and the rightmost column of Table 3 respectively. Grand average ERPs across all 30
electrode sites are presented in Supplementary Figures 5 and 6.
In Experiment 3, an N400 effect of role-reversal was elicited in the long-distance
conditions but not in the short-distance conditions. In the 350-450 ms interval, the posterior
electrode cluster analyses showed a significant N400 effect in the long-distance conditions but

38
not in the short-distance conditions, as predicted by the hypothesis that argument role
information has a delayed impact on comprehenders’ verb predictions. This is in line with the
results of the omnibus ANOVA, which revealed a reversal × distance × laterality interaction,
reflecting a larger negativity over central and right posterior electrodes in the long-distance
condition than in the short-distance condition.
In the 600-800 ms interval, the posterior cluster analyses showed that role reversal
elicited a significant P600 effect in long- and short-distance conditions alike. This is consistent
with the omnibus ANOVA, which revealed a main effect of reversal, a reversal × anteriority
interaction, but no interactions involving distance and reversal.
Event-related Potentials (Control comparison)
Grand average ERPs and the topographic distribution of ERP effects in the control
comparison are presented on the bottom half of Figure 4. Results of the statistical analyses are
presented in the rightmost column of Table 4.
As in Experiments 1 and 2, the standard cloze probability manipulation in the filler
sentences elicited a clear N400 effect and a frontally distributed late positivity. Statistical
analyses in the 350-450 ms time interval revealed a main effect of predictability and
predictability × anteriority and predictability × anteriority × laterality interactions, showing
that expected words elicited a smaller central posterior negativity (N400) than unexpected
words. Analyses in the 600-800 ms time interval revealed predictability × anteriority and
predictability × anteriority × laterality interactions, showing that unexpected words elicited a
larger frontal positivity than expected words.

39
Figure 4. Top: Grand average ERPs at centro-posterior electrode CPZ and topographic distribution of
ERP effects in the 350-450ms and 600-800ms intervals (role-reversed minus canonical) in the short-
distance (left) and long-distance (right) conditions in Experiment 3. Argument role reversals elicited a
P600 effect in the short-distance conditions, and an N400 effect followed by a P600 effect in the long-
distance conditions. The N400’s sensitivity to the target verb’s offline predictability critically depended
on the time elapsed between the arguments and the verb. When the verb appears immediately after the
arguments, expected and unexpected verbs elicit no N400 difference. When an adverbial phrase
appeared between the arguments and the verb, the N400 is reduced for the expected verb compared to
the unexpected verb. Bottom: Grand average ERPs at frontal electrode FZ and at centro-posterior
electrode CPZ and topographic distribution of ERP effects in the 350-450ms and 600-800ms intervals
(expected minus unexpected) in the control comparison in Experiment 3. A cloze probability
manipulation in the filler sentences elicited a clear N400 effect followed by a frontally distributed late
positivity.

40
Table 6. Results of the posterior electrode cluster analyses in Experiment 3: t value, standardized
effect size (Pearson’s r) and 95% confidence interval for the contrast between the canonical and role-
reversed conditions in each time interval.
Discussion
Figure 5 presents a summary of the main results of Experiments 1 to 3. The results of
Experiment 3 demonstrated that argument role reversals impacted the amplitude of the N400
at the target verb when it was further away from the arguments (in the long-distance conditions)
but not when it immediately followed the arguments (in the short-distance conditions),
replicating key findings of Experiments 1 and 2. Other aspects of the results were also
consistent with those of Experiments 1 and 2. First, argument role reversals were readily
detected in both short- and long-distance conditions and elicited a robust P600 effect in both
short- and long-distance conditions. Second, the cloze probability manipulation in the filler
sentences once again elicited a clear N400 effect followed by a late frontal positivity.
Experiment 3 df t
Pearson's r95% CI
Short distance
350-450ms 23 < 1 -0.14 [-0.51, 0.31]
600-800ms 23 3.6** 0.60 [0.21, 0.80]
Long distance
350-450ms 23 2.24* -0.42 [-0.68, -0.005]
600-800ms 23 4.21** 0.66 [0.32, 0.82]
** p < .01; * p < .05; ^ .05 < p <.1

41
Figure 5. The left and centre columns show average ERPs at centro-posterior electrode CPZ to the target
verb (350-450ms highlighted) in Experiments 1-3. The right column shows, from top to bottom, the
topographic distribution of the difference between role-reversed and canonical conditions during the
N400 time interval (350-450ms) in the high-predictability conditions in Experiments 1 and 2, and in
the long-distance conditions in Experiment 3. Argument role reversals elicited a P600 effect in all cases,
likely reflecting the detection of implausibility. Meanwhile, argument role reversals elicited a
significant N400 effect at the verb only when they resulted in a clear difference in the verb’s offline
predictability and when the target verb was further away from the arguments.
The N400’s sensitivity to argument role reversals in the long-distance conditions stands
in sharp contrast with its insensitivity in the short-distance conditions and in many previous
studies. Since the sentences in the short- and long-distance conditions were synonymous and
differed only by the location of the temporal phrase, the change in the N400’s sensitivity is not
attributable to changes in the content of the context per se. Further, since the distance

42
manipulation did not affect the cloze probability difference between the canonical and role-
reversed sentences, the observed difference between the short- and long-distance conditions
cannot be attributed to a change in the target verb’s cloze probability either. Instead, it is likely
due to the additional time elapsed between the second argument and the verb (600ms vs.
1800ms SOA) that was introduced by the distance manipulation. We take the presence of an
N400 effect in the long-distance conditions to indicate that access to the target verb’s meaning
was facilitated in the canonical sentences relative to the role-reversed sentences when
additional time elapsed between the arguments and the verb. Further, we argue that the fact
that the N400 effect only emerged in the long-distance conditions constitutes supporting
evidence for the Slow Prediction Hypothesis, according to which information about the
arguments’ roles has a delayed impact on comprehenders’ verb predictions.
General Discussion
In the current study, we investigated the possibility that even salient and unambiguous
information about preverbal arguments’ structural roles may have a delayed impact on online
verb predictions (the Slow Prediction hypothesis). We manipulated the linear distance between
the arguments and the verb to examine how it may affect comprehenders’ response to the verb
in canonical vs. role-reversed sentences in a series of ERP experiments.
Across all of the current experiments, the implausibility created by reversal of pre-
verbal arguments was readily detected and it consistently elicited a robust P600 effect at the
verb. These results are in line with previous findings (Chow & Phillips, 2013; Chow, Smith, et
al., 2016; Hoeks et al., 2004; Kolk et al., 2003; van Herten et al., 2006, 2005), and show that
comprehenders reliably use argument role information to interpret thematic relations upon
encountering the verb.
Meanwhile, argument role reversals elicited an N400 effect only when two conditions
were met: argument reversal had a strong impact on the verb’s cloze probability and the verb

43
was further away from the arguments. In Experiment 1 we extended previous findings by
showing that argument role reversals did not modulate the N400 at the verb even when they
resulted in a large cloze probability difference between conditions. In Experiments 2 and 3 we
found that argument role reversals elicited an N400 effect when (i) there was a clear cloze
probability differences between conditions and when (ii) the linear distance and, effectively,
the time interval between the arguments and the verb was widened. Note, however, that the
N400 did not simply become sensitive to argument role reversals when the verb was further
away from its arguments: argument role reversals did not elicit an N400 effect in the low-
predictability sentences even with the added distance. This strongly suggests that the N400
effect that emerged in the long-distance conditions reflected sensitivity to the verb’s offline
predictability. These distance effects are also in line with findings in a recent study in Japanese
that directly manipulated SOA in simple SV and OV sentences (Momma et al., 2015). Taken
together, these findings have two main implications.
Firstly, the present findings are incompatible with the Semantic Illusion hypothesis,
according to which comprehenders are initially blind to the implausibility in role-reversed
sentences when they first encounter the verb (e.g., Bornkessel-Schlesewsky & Schlesewsky,
2008; Hoeks et al., 2004; Kim & Osterhout, 2005; Kolk et al., 2003; Kuperberg, 2007). This
hypothesis, which assumes the integration view of the N400, would expect the N400 to be
insensitive to argument role reversals regardless of the distance between the verb and the
arguments because they could form a plausible thematic relation in all cases. However, the
observation of an N400 effect in the high-predictability conditions in Experiment 2 and in the
long-distance condition in Experiment 3 shows that argument role reversals can elicit an N400
effect even when the verb can form a plausible thematic relation with the arguments.
On the other hand, these findings are compatible with the prediction view of the N400
and provide support for the Slow Prediction hypothesis, according to which argument role

44
information cannot impact comprehenders’ verb predictions immediately. The effect of
distance suggests that argument role information can inform comprehenders’ verb predictions,
but that this requires sufficient time between the arguments and the verb. Further, the
observation that the N400 became sensitive to argument role reversals only when there was a
clear difference in the verb’s cloze probability between conditions suggests that it is changes
in comprehenders’ predictions that underlies the N400’s re-emerged sensitivity.
The present study extends previous eye-tracking findings by Kukona et al. (2011), who
found that listeners’ anticipatory eye-movements were sensitive to filled vs. unfilled thematic
roles in passive sentences but not in active sentences. First, whereas Kukona and colleagues
compared passive and active sentences that differed in multiple ways, in the present study we
directly manipulated the distance between the arguments and the verb in active sentences by
only changing the position of the temporal adverbial. This allowed us to keep the rest of the
sentence constant and examine the effect of timing more directly. Meanwhile, instead of
presenting only a small set of objects on the screen and tracking listeners’ looks to these objects
as they process the linguistic stimuli, in the present study we measured comprehenders’ ERP
response as they read. This allowed us to examine comprehenders’ response to the linguistic
stimuli without artificially restricting their options. Further, while the extent to which listeners’
anticipatory fixations are reflective of comprehenders’ predictions vs. other cognitive processes
remains debatable, there is much evidence that the N400 ERP response is modulated by the
ease of accessing a word’s meaning from long-term memory (Kutas & Federmeier, 2000; Lau,
Phillips, & Poeppel, 2008) and its well-attested sensitivity to a word’s predictability makes it
an excellent measure of the extent to which a word’s meaning is pre-activated.
More broadly, our observation that comprehenders’ verb predictions are initially not
sensitive to information about the arguments’ structural roles calls into question a simplifying
assumption that is common among existing models of prediction, namely, that all relevant

45
information can impact prediction immediately (e.g., Altmann & Mirković, 2009; Demberg &
Keller, 2008a; Elman, 1990; Hale, 2001; Levy, 2008; Smith & Levy, 2013). Instead, they
provide empirical support for Laszlo and Federmeier’s (2009) proposal that “quantitative shifts
in the timing of processing can potentially lead to qualitative differences in what particular
facets of semantics come to be linked up with a given input” (p.32) and highlight the
importance of timing considerations in the study of prediction (cf. Dambacher et al., 2012; Ito
et al., 2016; Wlotko & Federmeier, 2015). Further, the widened time interval in Experiments 2
and 3 (1800ms between the onset of the second argument and the verb) places an upper limit
on the time required for argument role information to impact verb predictions. By isolating the
time course of a specific component of predictive processing, these findings open up a host of
questions about how different sources of information are integrated to feed linguistic prediction
in real time.
Note, however, that the present proposal is distinct from ‘good-enough’ models of
language comprehension (e.g., Christianson, Hollingworth, Halliwell, & Ferreira, 2001;
Ferreira, Bailey, & Ferraro, 2002), or any model which posits that detailed syntactic
information may be ignored during comprehension (e.g., Townsend & Bever, 2001). As we
argued above, comprehenders’ accurate plausibility judgments and the presence of a robust
P600 effect across all experiments suggest that they reliably use argument role information for
interpreting the sentences in real time. Meanwhile, the N400 results suggest that it is
specifically the comprehenders’ predictions that are not immediately sensitive to argument role
information. Under this account, comprehenders immediately and reliably use argument role
information to interpret thematic relations when the verb appears, but they may not use this
information quickly enough to predict the upcoming verb before it appears in the input.

46
The role of lexical semantic association
Just as listeners’ eye-movements are sensitive to semantic relationships between the
linguistic stimuli and the visual objects (e.g., Huettig & Altmann, 2005; Yee & Sedivy, 2006),
the amplitude of the N400 response to a word is reduced when it is preceded by semantically
related words (for a review see Ledoux, Camblin, Swaab, & Gordon, 2006). Therefore, it is
important to consider how lexical semantic association between the arguments and the verb
might have contributed to the results observed here. In particular, one might attribute the
absence of N400 effects in the short-distance conditions and in previous studies to strong
lexical semantic associations between the arguments and verbs (Brouwer et al., 2012). If such
associations had rendered the N400 at the verb at floor, then we would not be able to observe
any N400 effect at the verb, even if argument role information did have an immediate impact
on verb predictions. According to this explanation, the N400’s sensitivity could have re-
emerged in the long-distance conditions not because of the amount of time available for
prediction, but because the effect of lexical association from the arguments decayed during the
adverbial phrase (e.g., Ratcliff, Hockley, & McKoon, 1985). This is an interesting alternative
possibility that warrants further investigation in future work. However, we note that the average
N400 at the target words in the experimental sentences was consistently much larger than that
elicited by the expected words in the filler sentences across all three experiments, and for this
reason we consider it unlikely that the N400 response to the verb in the short-distance
conditions was at floor.
10
10
Since the target words in the filler sentences were not matched in any way with those in the
experimental sentences, such a comparison may be suggestive but must be interpreted with caution.

47
Late frontal positivity and the P600
Although a late positivity was observed in experimental and control comparisons across
Experiments 1 to 3, these effects had clearly different topographic distributions. In the
experimental sentences, where the role-reversal manipulation could affect both predictability
and plausibility, target verbs elicited a larger late positivity in the role-reversed condition than
in the canonical condition, regardless of the predictability of the verb (Experiments 1 and 2) or
the distance between the arguments and the verb (Experiment 3). In all cases this late positivity
was most pronounced at parietal scalp sites, which is typical of a P600 effect. Meanwhile in
the filler sentences, where predictability was modulated but plausibility was held constant, the
late positivity elicited by unexpected words was frontally distributed across all three
experiments. The different topographies of these effects suggest that they might have arisen
from different brain regions and might reflect different functional processes.
While a proposal by Kuperberg (2013) has linked these late positivities to different
kinds of prediction error, a meta-analysis by Van Petten and Luka (2012) suggests that only
the frontal positivity is associated with prediction. They observed that late positivities to
semantically incongruous words tend to be posteriorly distributed and that late positivities to
unexpected but nonetheless semantically congruous words are often frontally distributed (Van
Petten & Luka, 2012). In the current study, argument role reversals resulted in implausible
experimental sentences, whereas all of the sentences in the control comparison were plausible.
The observation of a P600 effect in the experimental comparisons but not in the control
comparisons is in line with Van Petten and Luka’s (2012) observation and suggests that the
P600 is not modulated by a word’s predictability per se, but by the plausibility or congruence
of the sentence. Therefore, the P600 effect likely reflects processes that are triggered in case of
comprehension failures (e.g., upon detecting implausibility or grammatical anomalies in the
input), and these processes might involve re-analyses (e.g., Friederici, 1995), context updating

48
(Coulson et al., 1998), and/or error corrections in a noisy channel model (e.g., Gibson, Bergen,
& Piantadosi, 2013; Kim & Sikos, 2011).
Meanwhile, results in the control comparisons are consistent with previous reports that
unexpected but congruent words elicited a larger N400 and a frontal positivity relative to
strongly expected words (e.g., Delong, Urbach, Groppe, & Kutas, 2011; Federmeier et al., 2007;
Thornhill & Van Petten, 2012). Previous research has proposed that the frontal positivity
reflects the suppression or inhibition of a prediction for a different word or concept (Federmeier
et al., 2007; Kutas, 1993) and/or learning signal from which to update future predictions
(Federmeier, Kutas, & Schul, 2010). However, existing results have remained mixed with
regard to whether the positivity becomes larger when a strong prediction is violated
(Federmeier et al., 2007), or if the amplitude of the frontal positivity is reduced when a strongly
expected word is encountered (Thornhill & van Petten, 2012). Therefore, future research will
be needed to determine exactly how this component is related to predictive processes.
Other cases of apparent “prediction failure”
The present study demonstrates that even prominent and unambiguous contextual
information such as word order may fail to immediately impact predictions even in healthy
young adult native speakers. This reinforces an important but often overlooked distinction
between a word’s offline predictability (cloze probability) and its online predictability (Chow,
Smith, et al., 2016). Since offline predictability measures obtained in language corpora and
offline cloze tasks do not capture the temporal aspects of linguistic prediction, models of
language comprehension that are based on offline predictability measures would fail to capture
cases in which some contextual information has a non-immediate impact on comprehenders’
online predictions.
Further, it highlights the fact that a prediction can facilitate processing only if it
precedes the predicted event, and as such careful considerations about timing are crucial for

49
differentiating between genuine failures to engage predictive mechanisms and cases in which
specific sources of contextual information fail to impact predictions quickly enough to facilitate
bottom-up processing. For example, recent studies have reported that predictive processing
effects are absent under certain circumstances (e.g., Chwilla, Virgillito, & Vissers, 2011) and
in different speaker groups such as children with low vocabulary scores (e.g., Borovsky, Elman,
& Fernald, 2012; Mani & Huettig, 2012), older adults (DeLong, Groppe, Urbach, & Kutas,
2012; Federmeier & Kutas, 2005; Federmeier, McLennan, De Ochoa, & Kutas, 2002; Wlotko,
Federmeier, & Kutas, 2012), second language learners (e.g., Grüter, Lew-Williams, & Fernald,
2012; Kaan, 2014; Martin et al., 2013), illiterate adults (Mishra, Singh, Pandey, & Huettig,
2012), and patients with schizophrenia (e.g., Ford & Mathalon, 2012; Kuperberg, 2010). While
such findings may suggest that certain speaker groups cannot (or do not) engage predictive
processing, it is also possible that these speakers do in fact anticipate upcoming inputs during
comprehension, but that some of the computations involved are still incomplete when the
relevant input arises.
To take a more vivid analogy, imagine throwing a ball to a child. He might fail to catch
it because he lacks the ability to predict where it is going, but it could also be because he simply
cannot compute those predictions quickly enough in order to move to the predicted position
before the ball gets there. Therefore, to examine potential causes for apparent prediction
failures we must isolate sub-components of predictive computations and examine how different
sources of contextual information impact comprehenders’ prediction as the input unfolds in
real time. For example, future research can examine whether second language speakers are
more or less likely to show predictive effects depending on the type of contextual information
involved (e.g., grammatical vs. pragmatic cues) and the amount of the time they are given to
process it (e.g., Ito, Martin & Nieuwland, 2016).

50
Towards an explicit model of predictive processing
We believe that the current experimental paradigm can be applied more broadly to
examine the time course of predictive computations. This paradigm has three key components.
First, it requires isolating the source of contextual information under investigation. Second, it
manipulates the timing at which this piece of information becomes available relative to a target
word. Third, it requires a dependent measure that is modulated by the output of predictive
processing. The time course with which the contextual information of interest impacts
comprehenders’ prediction can then be inferred by examining this dependent measure’s
sensitivity at different time points. By controlling the strength and timing of specific predictive
cues, this paradigm can be applied more broadly to examine the time course with which
different information sources impact predictive processing and how it might be affected in
different populations.
For example, in a recent study in English, we expanded this paradigm to examine the
possibility that lexical semantic information about the arguments can impact comprehenders’
verb predictions more quickly than information about their structural roles (Chow, Smith, et
al., 2016). We manipulated the verb’s cloze probability by either (i) substituting one of the pre-
verbal arguments (e.g., “The superintendent overheard which tenant/realtor the landlord had
evicted…”) or (ii) reversing the order of the arguments (e.g., “The restaurant owner forgot
which customer/waitress the waitress/customer had served …”). Crucially, even though these
manipulations had the same impact on the verb’s cloze probability, an N400 effect was
observed in the argument substitution condition only. This suggests the arguments’ lexical
meaning may have a more immediate impact on verb predictions than their structural roles,
and, more generally, that different sources of information may impact linguistic predictions on
different time scales.

51
Why might information about argument roles have a delayed impact on verb predictions
compared to lexical semantic information about the arguments? Here we briefly consider three
different ways in which the impact of argument role information on the process of retrieving
likely events from memory may be delayed. For a more detailed discussion, see Chow, Momma,
et al. (2016).
One possibility is that the delay originates at the start of the memory retrieval process,
that is, when retrieval cues are issued. In languages such as Mandarin Chinese and English,
arguments’ structural roles are determined in large part by word order, and comprehenders may
be able to access the arguments’ lexical representation (which is independent of word order)
before they can determine the arguments’ structural roles. In turn, a delay could arise if
comprehenders start to retrieve from memory events that involve those arguments as
participants even before their structural roles are identified. Under this view, even a slight delay
in when the relevant retrieval cues are issued could have a cascading effect through the retrieval
process and result in a much bigger delay in their impact on comprehenders’ verb predictions.
However, this account might seem less plausible in light of recent studies in Japanese using
case-marking as a cue to argument role (Momma et al., 2015); although it seems likely that
case information could be processed rapidly, they similarly report evidence that argument role
information cannot impact comprehenders’ verb prediction within 800ms from argument onset.
Another possibility is that the delay may lie in the dynamics of the competition among
predictive cues (Kuperberg, 2016) and/or partially matching candidates (Lewis, Vasishth, &
Van Dyke, 2006). Under one version of this type of account, argument role information may
impact verb predictions more quickly when it is more prominent or strongly primed by the
preceding context. For example, if a preceding question asked “Who was doing what?”,
focusing the argument role information, readers might be faster to process this information and
incorporate it into predictions in the subsequent sentence.

52
The third possibility is that argument role information cannot be used to query event
knowledge directly, due to the way in which event knowledge is encoded in memory. Under
this view, there may be information in the linguistic input that constrains which events are
likely to be discussed, but that cannot be directly used as a cue for retrieving those events from
memory. For example, while arguments are associated with grammatical roles such as subject
and object in the linguistic input, event representations naturally do not encode event
participants in those terms. Therefore, at a minimum comprehenders would have to use an
argument and its grammatical role to identify its likely thematic role (e.g., animate subjects are
relatively likely to be agents) before they can query events that take this argument in a specific
role. Furthermore, it remains controversial whether sentence or verb meanings make reference
to abstract thematic roles such as agent and patient, or equally whether such roles are
fundamental primitives of non-linguistic event representations (see Williams, 2015 for
discussion). As such, when arguments are encountered prior to the verb, a more indirect
mechanism is likely required to use grammatical or thematic role information to retrieve events
from memory.
In Chow, Momma, et al. (2016), we discussed the possibility that using argument role
information for verb prediction might involve an indirect mechanism, in which arguments (or
event participants) are initially used to retrieve events (and the verbs associated with those
events), irrespective of their roles (e.g., both serving and tipping events are likely to involve
customers and waitresses as participants), and that the set of candidate verbs is then evaluated
to pick out those that match the specific roles of the arguments (e.g., a tipping event is likely
given the customer-as-agent and waitress-as-patient cues but a serving event is not). Further
research will be needed to investigate how event knowledge is encoded in long-term memory
and the way(s) in which such stored representations may be accessed during language
comprehension.

53
Conclusion
The current study investigates the possibility that even salient information about
preverbal arguments’ roles may have a delayed impact on online verb predictions. By
investigating how the timing of argument role information impacts comprehenders’ online verb
predictions, we show that argument role information does contribute to verb prediction, but the
predictive computations involved may be relatively slow. These results highlight the
importance of timing considerations in studying how diverse contextual information
contributes to predictions.
Acknowledgements
This work was supported in part by a National Science Foundation grant (BCS-0848554)
to Colin Phillips, and by a Key Project of National Social Science Foundation of China
(15AZD048) and a Key Project of National Natural Science Foundation of Guangdong
Province, China (2014A030311016) to Suiping Wang. We thank Nan Li, Lan Chen, Jie Li,
Qiong Sun, Wenjia Zhang, Luodi Yu, Jian Wang, Zhizhou Deng, Dongxia Sun, Xiayan Zhu,
Yangling Cui and Gangyi Feng for invaluable help in material creation and data collection.
References
Altmann, G. T. M. (1999). Thematic role assignment in context. Journal of Memory and
Language, 41(1), 124–145. http://doi.org/10.1006/jmla.1999.2640
Altmann, G. T. M., & Kamide, Y. (1999). Incremental interpretation at verbs: Restricting the
domain of subsequent reference. Cognition, 73(3), 247–264.
http://doi.org/10.1016/S0010-0277(99)00059-1
Altmann, G. T. M., & Mirković, J. (2009). Incrementality and prediction in human sentence
processing. Cognitive Science, 33(4), 583–609. http://doi.org/10.1111/j.1551-
6709.2009.01022.x
Aoshima, S., Phillips, C., & Weinberg, A. (2004). Processing filler-gap dependencies in a
head-final language. Journal of Memory and Language, 51(1), 23–54.
http://doi.org/10.1016/j.jml.2004.03.001
Bar, M. (2007). The proactive brain: using analogies and associations to generate predictions.
Trends in Cognitive Sciences, 11(7), 280–289. http://doi.org/10.1016/j.tics.2007.05.005
Bar, M. (Ed.). (2011). Predictions in the brain: Using our past to generate a future. Oxford
University Press.
Bendixen, A., Schroger, E., & Winkler, I. (2009). I Heard That Coming: Event-Related

54
Potential Evidence for Stimulus-Driven Prediction in the Auditory System. Journal of
Neuroscience, 29(26), 8447–8451. http://doi.org/10.1523/JNEUROSCI.1493-09.2009
Bicknell, K., Elman, J. L., Hare, M., McRae, K., & Kutas, M. (2010). Effects of event
knowledge in processing verbal arguments. Journal of Memory and Language, 63(4),
489–505. http://doi.org/10.1016/j.jml.2010.08.004
Bornkessel-Schlesewsky, I., Kretzschmar, F., Tune, S., Wang, L., Genç, S., Philipp, M., …
Schlesewsky, M. (2011). Think globally: Cross-linguistic variation in
electrophysiological activity during sentence comprehension. Brain and Language,
117(3), 133–152. http://doi.org/10.1016/j.bandl.2010.09.010
Bornkessel-Schlesewsky, I., & Schlesewsky, M. (2008). An alternative perspective on
“semantic P600” effects in language comprehension. Brain Research Reviews, 59(1),
55–73. http://doi.org/10.1016/j.brainresrev.2008.05.003
Borovsky, A., Elman, J. L., & Fernald, A. (2012). Knowing a lot for one’s age: Vocabulary
skill and not age is associated with anticipatory incremental sentence interpretation in
children and adults. Journal of Experimental Child Psychology, 112(4), 417–436.
http://doi.org/10.1016/j.jecp.2012.01.005
Brouwer, H., Fitz, H., & Hoeks, J. C. J. (2012). Getting real about Semantic Illusions:
Rethinking the functional role of the P600 in language comprehension. Brain Research,
1446, 127–143. http://doi.org/10.1016/j.brainres.2012.01.055
Brusini, P., Brun, M., Brunet, I., & Christophe, A. (2015). Listeners Exploit Syntactic
Structure On-Line to Restrict Their Lexical Search to a Subclass of Verbs. Frontiers in
Psychology, 6(December), 1841. http://doi.org/10.3389/fpsyg.2015.01841
Chow, W.-Y., Momma, S., Smith, C., Lau, E., & Phillips, C. (2016). Prediction as Memory
Retrieval : Timing and Mechanisms. Language, Cognition and Neuroscience, 31(5),
617-627. doi:10.1080/23273798.2016.1160135
Chow, W.-Y., & Phillips, C. (2013). No semantic illusions in the Semantic P600
phenomenon: ERP evidence from Mandarin Chinese. Brain Research, 1506, 76–93.
http://doi.org/10.1016/j.brainres.2013.02.016
Chow, W.-Y., Smith, C., Lau, E., & Phillips, C. (2016). A “ bag-of-arguments ” mechanism
for initial verb predictions. Language, Cognition and Neuroscience, 31(5), 577-596.
doi:10.1080/23273798.2015.1066832
Christianson, K., Hollingworth, A., Halliwell, J. F., & Ferreira, F. (2001). Thematic roles
assigned along the garden path linger. Cognitive Psychology, 42(4), 368–407.
http://doi.org/10.1006/cogp.2001.0752
Chwilla, D. J., Virgillito, D., & Vissers, C. T. W. M. (2011). The relationship of language
and emotion: N400 support for an embodied view of language comprehension. Journal
of Cognitive Neuroscience, 23(9), 2400–2414. http://doi.org/10.1162/jocn.2010.21578
Coulson, S., King, J. W., & Kutas, M. (1998). Expect the Unexpected: Event-related Brain
Response to Morphosyntactic Violations. Language and Cognitive Processes, 13(1), 21–
58. http://doi.org/10.1080/016909698386582
Dambacher, M., Dimigen, O., Braun, M., Wille, K., Jacobs, A. M., & Kliegl, R. (2012).
Stimulus onset asynchrony and the timeline of word recognition: Event-related
potentials during sentence reading. Neuropsychologia, 50(8), 1852–1870.
http://doi.org/10.1016/j.neuropsychologia.2012.04.011
Dambacher, M., Kliegl, R., Hofmann, M., & Jacobs, A. M. (2006). Frequency and
predictability effects on event-related potentials during reading. Brain Research,
1084(1), 89–103. http://doi.org/10.1016/j.brainres.2006.02.010
DeLong, K. A., Groppe, D. M., Urbach, T. P., & Kutas, M. (2012). Thinking ahead or not?
Natural aging and anticipation during reading. Brain and Language, 121(3), 226–239.
http://doi.org/10.1016/j.bandl.2012.02.006

55
DeLong, K. A., Quante, L., & Kutas, M. (2014). Predictability, plausibility, and two late ERP
positivities during written sentence comprehension. Neuropsychologia, 61(1), 150–162.
http://doi.org/10.1016/j.neuropsychologia.2014.06.016
Delong, K. A., Urbach, T. P., Groppe, D. M., & Kutas, M. (2011). Overlapping dual ERP
responses to low cloze probability sentence continuations. Psychophysiology, 48(9),
1203–1207. http://doi.org/10.1111/j.1469-8986.2011.01199.x
Delong, K. A., Urbach, T. P., & Kutas, M. (2005). Probabilistic word pre-activation during
language comprehension inferred from electrical brain activity. Nature Neuroscience,
8(8), 1117–1121. http://doi.org/10.1038/nn1504
Demberg, V., & Keller, F. (2008). Data from eye-tracking corpora as evidence for theories of
syntactic processing complexity. Cognition, 109(2), 193–210.
http://doi.org/10.1016/j.cognition.2008.07.008
Dikker, S., Rabagliati, H., Farmer, T. a., & Pylkkanen, L. (2010). Early Occipital Sensitivity
to Syntactic Category Is Based on Form Typicality. Psychological Science, 21(5), 629–
634. http://doi.org/10.1177/0956797610367751
Dikker, S., Rabagliati, H., & Pylkkänen, L. (2009). Sensitivity to syntax in visual cortex.
Cognition, 110(3), 293–321. http://doi.org/10.1016/j.cognition.2008.09.008
Elman, J. L. (1990). Finding structure in time. Cognitive Science, 14(2), 179–211.
http://doi.org/10.1207/s15516709cog1402_1
Farmer, T. A., Yan, S., Bicknell, K., & Tanenhaus, M. K. (2015). Form-To-Expectation
Matching Effects on First-Pass Eye Movement Measures During Reading. Journal of
Experimental Psychology: Human Perception and Performance, 41(4), 958–976.
http://doi.org/10.1037/xhp0000054
Federmeier, K. D., & Kutas, M. (1999). Right words and left words: Electrophysiological
evidence for hemispheric differences in meaning processing. Cognitive Brain Research,
8(3), 373–392. http://doi.org/10.1016/S0926-6410(99)00036-1
Federmeier, K. D., & Kutas, M. (2005). Aging in context: Age-related changes in context use
during language comprehension. Psychophysiology, 42(2), 133–141.
http://doi.org/10.1111/j.1469-8986.2005.00274.x
Federmeier, K. D., Kutas, M., & Schul, R. (2010). Age-related and individual differences in
the use of prediction during language comprehension. Brain and Language, 115(3),
149–161. http://doi.org/10.1016/j.bandl.2010.07.006
Federmeier, K. D., McLennan, D. B., De Ochoa, E., & Kutas, M. (2002). The impact of
semantic memory organization and sentence context information on spoken language
processing by younger and older adults: An ERP study. Psychophysiology, 39(2), 133–
146. http://doi.org/10.1017/S0048577202001373
Federmeier, K. D., Wlotko, E. W., De Ochoa-Dewald, E., & Kutas, M. (2007). Multiple
effects of sentential constraint on word processing. Brain Research, 1146(1), 75–84.
http://doi.org/10.1016/j.brainres.2006.06.101
Ferreira, F., Bailey, K. G. D., & Ferraro, V. (2002). Good-Enough Representations in
Language Comprehension. Current Directions in Psychological Science, 11(1), 11–15.
Ferretti, T. R., McRae, K., & Hatherell, A. (2001). Integrating verbs, situation schemas, and
thematic role concepts. Journal of Memory and Language, 44(4), 516–547.
http://doi.org/10.1006/jmla.2000.2728
Ford, J. M., & Mathalon, D. H. (2012). Anticipating the future: Automatic prediction failures
in schizophrenia. International Journal of Psychophysiology, 83(2), 232–239.
http://doi.org/10.1016/j.ijpsycho.2011.09.004
Friederici, A. D. (1995). The time course of syntactic activation during language processing:
a model based on neuropsychological and neurophysiological data. Brain and Language.
http://doi.org/10.1006/brln.1995.1048

56
Gerlanc, D., & Kirby, K. N. (2013). bootES. Retrieved from http://cran.r-
project.org/web/packages/bootES/index.html
Gibson, E., Bergen, L., & Piantadosi, S. T. (2013). Rational integration of noisy evidence and
prior semantic expectations in sentence interpretation. Proceedings of the National
Academy of Sciences of the United States of America, 110(20), 8051–6.
http://doi.org/10.1073/pnas.1216438110
Greenhouse, S., & Geisser, S. (1959). On methods in the analysis of profile data.
Psychometrika, 24(2), 95–112. http://doi.org/10.1007/BF02289823
Gronlund, S. D., Edwards, M. B., & Ohrt, D. D. (1997). Comparison of the retrieval of item
versus spatial position information. Journal of Experimental Psychology. Learning,
Memory, and Cognition, 23(5), 1261–74. http://doi.org/10.1037//0278-7393.23.5.1261
Grüter, T., Lew-Williams, C., & Fernald, A. (2012). Grammatical gender in L2: A production
or a real-time processing problem? Second Language Research, 28(2), 191–215.
http://doi.org/10.1177/0267658312437990
Gunter, T. C., Stowe, L. A., & Mulder, G. (1997). When syntax meets semantics.
Psychophysiology, 34, 660–676.
Hagoort, P., Brown, C. M., & Groothusen, J. (1993). The Syntactic Positive Shift (SPS) as an
ERP measure of Syntactic Processing. Language and Cognitive Processes, 8(4), 439–
483.
Hagoort, P., Hald, L., Bastiaansen, M., & Petersson, K. M. (2004). Integration of Word
Meaning and World Knowledge in Language Comprehension. Science, 438(2004).
http://doi.org/10.1126/science.1095455
Hahne, A., & Friederici, A. D. (1999). Electrophysiological Evidence for Two Steps in
Syntactic Analysis : Early Automatic and Late Controlled Processes. Journal of
Cognitive Neuroscience, 11(2), 194–205.
Hale, J. (2001). A probabilistic earley parser as a psycholinguistic model. Proceedings of
NAACL 2001, 2, 159–166. http://doi.org/10.3115/1073336.1073357
Hare, M., Jones, M., Thomson, C., Kelly, S., & McRae, K. (2009). Activating event
knowledge. Cognition, 111(2), 151–167. http://doi.org/10.1016/j.cognition.2009.01.009
Hawkins, J., & Blakeslee, S. (2004). On intelligence. Times Books.
Hoeks, J. C. J., Stowe, L. A., & Doedens, G. (2004). Seeing words in context: The interaction
of lexical and sentence level information during reading. Cognitive Brain Research,
19(1), 59–73. http://doi.org/10.1016/j.cogbrainres.2003.10.022
Huettig, F., & Altmann, G. T. M. (2005). Word meaning and the control of eye fixation:
Semantic competitor effects and the visual world paradigm. Cognition, 96(1), 23–32.
http://doi.org/10.1016/j.cognition.2004.10.003
Ilkin, Z., & Sturt, P. (2011). Active prediction of syntactic information during sentence
processing. Dialogue and Discourse, 2(1), 35–58. http://doi.org/10.5087/dad.2011.103
Ito, A., Corley, M., Pickering, M. J., Martin, A. E., & Nieuwland, M. S. (2016). Predicting
form and meaning: Evidence from brain potentials. Journal of Memory and
Language, 86, 157-171. http://doi.org/10.1016/j.jml.2015.10.007
Ito, A., Martin, A. E., & Nieuwland, M. S. (2016). On predicting form and meaning in a
second language. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 43(4), 635-652. http://doi.org/10.1037/xlm0000315
Kaan, E. (2014). Predictive sentence processing in L2 and L1: What is different? Linguistic
Approaches to Bilingualism, 4(2), 257–282. http://doi.org/10.1075/lah-4-2.05kaa
Kamide, Y., Altmann, G. T. M., & 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. http://doi.org/10.1016/S0749-596X(03)00023-
8

57
Kim, A., & Lai, V. (2012). Rapid interactions between lexical semantic and word form
analysis during word recognition in context: evidence from ERPs. Journal of Cognitive
Neuroscience, 24(5), 1104–12. http://doi.org/10.1162/jocn_a_00148
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. http://doi.org/10.1016/j.jml.2004.10.002
Kim, A., & Sikos, L. (2011). Conflict and surrender during sentence processing: An ERP
study of syntax-semantics interaction. Brain and Language, 118(1-2), 15–22.
http://doi.org/10.1016/j.bandl.2011.03.002
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.
http://doi.org/10.1016/j.cognition.2004.03.002
Kolk, H. H. J., Chwilla, D. J., van Herten, M., & Oor, P. J. W. (2003). Structure and limited
capacity in verbal working memory: A study with event-related potentials. Brain and
Language, 85(1), 1–36. http://doi.org/10.1016/S0093-934X(02)00548-5
Kos, M., Vosse, T., van den Brink, D., & Hagoort, P. (2010). About edible restaurants:
Conflicts between syntax and semantics as revealed by ERPs. Frontiers in Psychology,
1(DEC), 1–11. http://doi.org/10.3389/fpsyg.2010.00222
Kukona, A., Fang, S. Y., Aicher, K. A., Chen, H., & Magnuson, J. S. (2011). The time course
of anticipatory constraint integration. Cognition, 119(1), 23–42.
http://doi.org/10.1016/j.cognition.2010.12.002
Kuperberg, G. R. (2007). Neural mechanisms of language comprehension: Challenges to
syntax. Brain Research, 1146(1), 23–49. http://doi.org/10.1016/j.brainres.2006.12.063
Kuperberg, G. R. (2010). Language in Schizophrenia Part 1: An Introduction. Linguistics and
Language Compass, 4(8), 576–589. http://doi.org/10.1111/j.1749-818X.2010.00216.x
Kuperberg, G. R. (2013). The Proactive Comprehender: What Event-Related Potentials tell
us about the dynamics of reading comprehension. Gina. In B. Miller, L. Cutting, & P.
McCardle (Eds.), Unraveling the Behavioral, Neurobiological, and Genetic Components
of Reading Comprehension (pp. 176–19). Baltimore: Paul Brookes Publishing.
Kuperberg, G. R. (2016). Separate streams or probabilistic inference? What the N400 can tell
us about the comprehension of events. Language, Cognition and Neuroscience.
http://doi.org/10.1080/23273798.2015.1130233
Kuperberg, G. R., Sitnikova, T., Caplan, D. N., & Holcomb, P. J. (2003).
Electrophysiological distinctions in processing conceptual relationships within simple
sentences. Cognitive Brain Research, 17(1), 117–129. http://doi.org/10.1016/S0926-
6410(03)00086-7
Kutas, M. (1993). In the company of other words: Electrophysiological evidence for single-
word and sentence context effects. Language and Cognitive Processes, 8(4), 533–572.
http://doi.org/10.1080/01690969308407587
Kutas, M., & Federmeier, K. D. (2000). Electropsysiology reveals semantic memory use in
language comprehension. Trends in Cognitive Science, 12(12), 463–470.
http://doi.org/10.1016/S1364-6613(00)01560-6
Kutas, M., & Federmeier, K. D. (2011). Thirty years and counting: finding meaning in the
N400 component of the event-related brain potential (ERP). Annual Review of
Psychology, 62(August), 621–47. http://doi.org/10.1146/annurev.psych.093008.131123
Kutas, M., & Hillyard, S. A. (1980). Reading senseless sentences: brain potentials reflect
semantic incongruity. Science, 207(4427), 203–205.
http://doi.org/10.1126/science.7350657
Kutas, M., & Hillyard, S. A. (1984). Brain potentials reflect word expectancy and semantic

58
association during reading. Nature, 307, 161–163.
Laszlo, S., & Federmeier, K. D. (2009). A beautiful day in the neighborhood: An event-
related potential study of lexical relationships and prediction in context. Journal of
Memory and Language, 61(3), 326–338. http://doi.org/10.1016/j.jml.2009.06.004
Lau, E. F., Phillips, C., & Poeppel, D. (2008). A cortical network for semantics:
(de)constructing the N400. Nature Reviews Neuroscience, 9(12), 920–933.
http://doi.org/Doi 10.1038/Nrn2532
Ledoux, K., Camblin, C. C., Swaab, T. Y., & Gordon, P. C. (2006). Reading Words in
Discourse: The Modulation of Lexical Priming Effects by Message-Level Context.
Behavioral and Cognitive Neuroscience Reviews, 5(3), 107–127.
http://doi.org/10.1177/1534582306289573
Levy, R. P. (2008). Expectation-based syntactic comprehension. Cognition, 106(3), 1126–
1177. http://doi.org/10.1016/j.cognition.2007.05.006
Levy, R. P., Fedorenko, E., Breen, M., & Gibson, E. (2012). The processing of extraposed
structures in English. Cognition, 122(1), 12–36.
http://doi.org/10.1016/j.cognition.2011.07.012
Lewis, R. L., Vasishth, S., & Van Dyke, J. A. (2006). Computational principles of working
memory in sentence comprehension. Trends in Cognitive Sciences, 10(10), 447–454.
http://doi.org/10.1016/j.tics.2006.08.007
Llinás, R. (2002). I of the vortex: from neurons to self. Cambridge, MA: MIT Press.
Mani, N., & Huettig, F. (2012). Prediction during language processing is a piece of cake—
But only for skilled producers. Journal of Experimental Psychology: Human Perception
and Performance, 38(4), 843–847. http://doi.org/10.1037/a0029284
Martin, C. D., Thierry, G., Kuipers, J. R., Boutonnet, B., Foucart, A., & Costa, A. (2013).
Bilinguals reading in their second language do not predict upcoming words as native
readers do. Journal of Memory and Language, 69(4), 574–588.
http://doi.org/10.1016/j.jml.2013.08.001
McElree, B., & Dosher, B. A. (1993). Serial recovery processes in the recovery of order
information. Journal of Experimental Psychology: General, 122(3), 291–315.
http://doi.org/10.1037/0096- 3445.122.3.291
McRae, K., Hare, M., Elman, J. L., & Ferretti, T. R. (2005). A basis for generating
expectancies for verbs from nouns. Memory & Cognition, 33(7), 1174–1184.
http://doi.org/10.3758/BF03193221
Mishra, R. K., Singh, N., Pandey, A., & Huettig, F. (2012). Spoken language-mediated
anticipatory eye- movements are modulated by reading ability - Evidence from Indian
low and high literates. Journal of Eye Movement Research, 5(1), 1–10.
Momma, S., Sakai, H., & Phillips, C. (2015). Give me several hundred more milliseconds:
Temporal dynamics of verb prediction. Talk given at the 28th annual CUNY Conference
on Human Sentence Processing, Los Angeles, CA.
Oishi, H., & Sakamoto, T. (2010). Immediate interaction between syntactic and semantic
outputs: evidence from event-related potentials in Japanese sentence processing. In
Poster presented at the 22nd annual CUNY Human Sentence Processing Conference.
Davis, CA.
Oldfield, R. C. (1971). The assessment and analysis of handedness: The Edinburgh inventory.
Neuropsychologia. http://doi.org/10.1016/0028-3932(71)90067-4
Omaki, A., Lau, E., White, I. D., Dakan, M. L., Apple, A., & Phillips, C. (2015). Hyper-
active gap filling. Frontiers in Psychology, 6(MAR), 1–30.
http://doi.org/10.3389/fpsyg.2015.00384
Osterhout, L., & Holcomb, P. J. (1992). Event-related brain potentials elicited by syntactic
anomaly. Journal of Memory and Language, 31(6), 785–806.

59
http://doi.org/10.1016/0749-596X(92)90039-Z
Otten, M., Nieuwland, M. S., & Van Berkum, J. J. A. (2007). Great expectations: specific
lexical anticipation influences the processing of spoken language. BMC Neuroscience, 8,
89. http://doi.org/10.1186/1471-2202-8-89
Otten, M., & Van Berkum, J. J. A. (2008). Discourse-Based Word Anticipation During
Language Processing: Prediction or Priming? Discourse Processes, 45(6), 464–496.
http://doi.org/10.1080/01638530802356463
Paczynski, M., & Kuperberg, G. R. (2012). Multiple influences of semantic memory on
sentence processing: Distinct effects of semantic relatedness on violations of real-world
event/state knowledge and animacy selection restrictions. Journal of Memory and
Language, 67, 426–448. http://doi.org/10.1016/j.jml.2012.07.003
Ratcliff, R., Hockley, W., & McKoon, G. (1985). Components of activation: repetition and
priming effects in lexical decision and recognition. Journal of Experimental Psychology.
General, 114(4), 435–50. http://doi.org/10.1037//0096-3445.114.4.435
Smith, N. J., & Levy, R. P. (2013). The effect of word predictability on reading time is
logarithmic. Cognition, 128(3), 302–319. http://doi.org/10.1016/j.cognition.2013.02.013
Stroud, C., & Phillips, C. (2012). Examining the evidence for an independent semantic
analyzer: An ERP study in Spanish. Brain and Language, 120(2), 108–126.
http://doi.org/10.1016/j.bandl.2011.02.001
Szewczyk, J. M., & Schriefers, H. (2013). Prediction in language comprehension beyond
specific words: An ERP study on sentence comprehension in Polish. Journal of Memory
and Language, 68(4), 297–314. http://doi.org/10.1016/j.jml.2012.12.002
Tanner, D., Morgan-Short, K., & Luck, S. J. (2015). How inappropriate high-pass filters can
produce artifactual effects and incorrect conclusions in ERP studies of language and
cognition. Psychophysiology, 52, 997–1009. http://doi.org/10.1111/psyp.12437
Taylor, W. L. (1953). Cloze procedure: A new tool for measuring readability. Journalism
Quarterly, 30(4), 415.
Thornhill, D. E., & van Petten, C. (2012). Lexical versus conceptual anticipation during
sentence processing: Frontal positivity and N400 ERP components. International
Journal of Psychophysiology, 83(3), 382–392.
http://doi.org/10.1016/j.ijpsycho.2011.12.007
Townsend, D. J., & Bever, T. G. (2001). Sentence comprehension: The integration of habits
and rules. Cambridge, MA: MIT Press.
Van Berkum, J. J. A. (2009). The neuropragmatics of “simple” utterance comprehension: An
ERP review. In U. Sauerland & K. Yatsushiro (Eds.), Semantic and pragmatics: From
experiment to theory (pp. 276–316). Palgrave.
Van Berkum, J. J. A., Brown, C. M., Zwitserlood, P., Kooijman, V., & Hagoort, P. (2005).
Anticipating upcoming words in discourse: evidence from ERPs and reading times.
Journal of Experimental Psychology. Learning, Memory, and Cognition, 31(3), 443–
467. http://doi.org/10.1037/0278-7393.31.3.443
Van Berkum, J. J. A., van den Brink, D., Tesink, C. M. J. Y., Kos, M., & Hagoort, P. (2008).
The neural integration of speaker and message. Journal of Cognitive Neuroscience, 20,
580–591. http://doi.org/10.1162/jocn.2008.20054
van de Meerendonk, N., Kolk, H. H. J., Vissers, C. T. W. M., & Chwilla, D. J. (2010).
Monitoring in language perception: mild and strong conflicts elicit different ERP
patterns. Journal of Cognitive Neuroscience, 22(1), 67–82.
http://doi.org/10.1162/jocn.2008.21170
van Herten, M., Chwilla, D. J., & Kolk, H. H. J. (2006). When heuristics clash with parsing
routines: ERP evidence for conflict monitoring in sentence perception. Journal of
Cognitive Neuroscience, 18(7), 1181–97. http://doi.org/10.1162/jocn.2006.18.7.1181

60
van Herten, M., Kolk, H. H. J., & Chwilla, D. J. (2005). An ERP study of P600 effects
elicited by semantic anomalies. Cognitive Brain Research, 22(2), 241–255.
http://doi.org/10.1016/j.cogbrainres.2004.09.002
van Petten, C., & Luka, B. J. (2012). Prediction during language comprehension: Benefits,
costs, and ERP components. International Journal of Psychophysiology, 83(2), 176–
190. http://doi.org/10.1016/j.ijpsycho.2011.09.015
Wicha, N. Y. Y., Moreno, E. M., & Kutas, M. (2004). Anticipating words and their gender:
an event-related brain potential study of semantic integration, gender expectancy, and
gender agreement in Spanish sentence reading. Journal of Cognitive Neuroscience,
16(7), 1272–88. http://doi.org/10.1162/0898929041920487
Williams, A. (2015). Arguments in Syntax and Semantics. Cambridge University Press.
Wlotko, E. W., & Federmeier, K. D. (2012). So that’s what you meant! Event-related
potentials reveal multiple aspects of context use during construction of message-level
meaning. NeuroImage, 62(1), 356–366.
http://doi.org/10.1016/j.neuroimage.2012.04.054
Wlotko, E. W., & Federmeier, K. D. (2015). Time for prediction? The effect of presentation
rate on predictive sentence comprehension during word-by-word reading. Cortex, 68,
20–32. http://doi.org/10.1016/j.cortex.2015.03.014
Wlotko, E. W., Federmeier, K. D., & Kutas, M. (2012). To predict or not to predict: age-
related differences in the use of sentential context. Psychology and Aging, 27(4), 975–
88. http://doi.org/10.1037/a0029206
Wolpert, D. M. (1997). Computational approaches to motor control. Trends in Cognitive
Sciences, 1(6), 1–5. http://doi.org/10.1007/s13398-014-0173-7.2
Xiang, M., & Kuperberg, G. R. (2015). Reversing expectations during discourse
comprehension. Language, Cognition and Neuroscience, 30(6), 648–672.
http://doi.org/10.1080/23273798.2014.995679
Ye, Z., & Zhou, X. (2008). Involvement of cognitive control in sentence comprehension:
Evidence from ERPs. Brain Research, 1203, 103–115.
http://doi.org/10.1016/j.brainres.2008.01.090
Yee, E., & Sedivy, J. C. (2006). Eye Movements to Pictures Reveal Transient Semantic
Activation During Spoken Word Recognition. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 32(1), 1–14. http://doi.org/10.1037/0278-7393.32.1.1
Yoshida, M., Dickey, M. W., & Sturt, P. (2013). Predictive processing of syntactic structure:
Sluicing and ellipsis in real-time sentence processing. Language and Cognitive
Processes, 28(3), 272–302. http://doi.org/10.1080/01690965.2011.622905