Abstract

It was recently proposed that lexical prediction in sentence context encompasses two qualitatively distinct prediction mechanisms: “pre-activation,” namely, activating representations stored in long-term memory, and “pre-updating,” namely, updating the sentence's representation, built online in working memory (WM), to include the predicted content [Lau, E. F., Holcomb, P. J., & Kuperberg, G. R. Dissociating N400 effects of prediction from association in single-word contexts. Journal of Cognitive Neuroscience, 25, 484–502, 2013]. The current study sought to find evidence for pre-updating and test the influence of individual differences in WM capacity on the tendency to engage in this process. Participants read strongly and weakly constraining sentences. ERPs were measured on the predictable noun as well as on the preceding verb, where the prediction is generated. Increased P600 amplitude was observed at the verb in the strongly constraining sentences, reflecting integration of the predicted upcoming argument, thus providing evidence for pre-updating. This effect was greater for participants with higher WM capacity, indicating that the tendency to engage in pre-updating is highly affected by WM capacity. The opposite effect was observed at the noun, that is, for participants with higher WM span, a greater decrease in P600 amplitude in the strongly constraining sentences was observed, indicating that the integration of a pre-updated word was easier. We discuss these results in light of previous literature and propose a plausible architecture to account for the interplay between pre-activation and pre-updating, mediating the influence of factors such as WM capacity.

INTRODUCTION

Over the past couple of decades, studies focusing on prediction processes have led to an increasingly strong consensus regarding the role of prediction in sentence processing. It is now widely assumed that, in the course of comprehending a sentence, we do not passively wait for the input and process it as it comes but rather constantly engage in some form of anticipatory processing. A classical finding demonstrating this is the decreased RTs observed for predictable as compared with unpredictable words in a sentence (e.g., Traxler & Foss, 2000; Schwanenflugel & LaCount, 1988; Schwanenflugel & Shoben, 1985; Stanovich & West, 1983; Ehrlich & Rayner, 1981; Forster, 1981). Similarly, the amplitude of the N400 ERPs component elicited by a word has been shown to inversely correlate with the word's cloze probability, meaning that N400 amplitude decreases as the word's predictability increases (e.g., Wlotko & Federmeier, 2012; DeLong, Urbach, & Kutas, 2005; Kutas & Hillyard, 1984; Kutas, Lindamood, & Hillyard, 1984). Interestingly, the amplitude of the N400 is not decreased only for predictable words. Unpredictable words that are semantically related to a predictable word elicit a smaller N400 compared with unrelated words (e.g., Thornhill & Van Petten, 2012), and this is true even for anomalous words (e.g., Federmeier & Kutas, 1999; Kutas & Hillyard, 1984; Kutas et al., 1984). These findings indicate that the amplitude of the N400 reflects an architecture that involves pre-activation, that is, that a decreased N400 likely reflects easier retrieval of words that were already activated, either due to spreading activation from the predicted word to related words or due to activation of shared features/concepts. In addition to “spreading” or shared activation, the activation level of a word, reflected in its N400 amplitude, is also affected by higher-level prediction processes. This is shown by findings of N400 sensitivity to factors that influence top–down control, such as predictive validity (e.g., Lau, Holcomb, & Kuperberg, 2013).

Pre-activation versus Pre-updating

Although the general notion of prediction is largely accepted in the psycholinguistic literature, the precise nature of the processes involved as well as the extent and ubiquity of prediction during sentence processing are still under ongoing debate. Recently, Lau et al. (2013) suggested a distinction between two qualitatively distinct mechanisms of prediction, later referred to as “pre-activation” and “pre-updating” (Kuperberg & Jaeger, 2016). “Pre-activation” refers to an increase in the activation level of knowledge stored in long-term memory, that is, the concept's representation in the lexicon, due to spreading activation as well as more controlled prediction processes. In contrast, the term “pre-updating” is used to refer to the updating of the sentence's representation built in working memory (WM) to include the predicted content.

Two different accounts were put forward regarding the relation between pre-activation and pre-updating. Although Lau et al. (2013) assume that pre-activation leads to pre-updating, Kuperberg and Jaeger (2016) assume the opposite order. To illustrate this, Kuperberg and Jaeger consider the following sentence fragment: “The day was breezy so the boy went outside to fly a….” We can hypothesize two possible processing scenarios occurring after encountering this fragment. One possibility, adopted by Lau et al. (2013), is that the partial representation of the event (<boy flies>) leads to pre-activation of the lower-level representation of “kite.” This pre-activated lexical representation is then pre-updated, that is, it enters WM to be integrated with the sentence's representation, which in turn affects higher representational levels, including updating of the event representation (to be <boy flies kite>). The other possibility, adopted by Kuperberg and Jaeger (2016), is that the partial representation of the event (<boy flies>) is first pre-updated (to be <boy flies kite>), which then causes pre-activation of the lower-level representation of “kite.” Distinguishing between these accounts is difficult; although they differ in the assumed order in which the different representational levels are updated, they ultimately lead to very similar predictions. Here, we adopt the view that pre-activation precedes pre-updating and that pre-updating can then “propagate” to higher representational levels.

An important difference between pre-activation and pre-updating is that, although the former entails priming of multiple entities, the latter entails commitment to a specific prediction, which would incur processing costs if disconfirmed. If no commitment is made about an upcoming word, then processing difficulty associated with an unexpected word should only depend on this word's activation level (correlated by hypothesis with its cloze probability). Indeed, several studies have shown that low-cloze words elicit similar N400 amplitudes regardless of whether their preceding context is strongly or weakly constraining. Crucially, however, it was also found that low-cloze words that follow strongly as opposed to weakly constraining contexts elicit a late anterior positivity, the frontal post-N400 negativity (fPNP). This component was argued to reflect an additional cost of prediction failure, exhibited only when the comprehender committed to a specific prediction (e.g., Ness & Meltzer-Asscher, 2018; Brothers, Swaab, & Traxler, 2015; Federmeier, Wlotko, De Ochoa-Dewald, & Kutas, 2007; see Van Petten & Luka, 2012, for a review).

Thus, as opposed to pre-activation, which is graded and can occur to different degrees depending on prediction strength and specificity, pre-updating is an “all or nothing” mechanism, which compels commitment. This conjecture stems from the inherent properties of the architectures of long-term memory and WM, which are vastly different. Although an immeasurable number of representations is simultaneously stored in long-term memory, the capacity of WM is highly limited (e.g., “the magical number seven” suggested by Miller, 1956 or “the magic number four” suggested by Green, 2017; Cowan, 2010), and therefore, it is unlikely that many competing predictions can simultaneously be pre-updated.1 Based on the properties of these two different memory systems, we view pre-activation as an “unavoidable” process that occurs whenever linguistic input is processed. However, only when activation of some predicted content reaches a certain threshold, this content is pre-updated; at this threshold the predicted content is, for all intents and purposes, retrieved and is therefore integrated into the sentence's representation in WM. It should be noted that we do not claim that pre-updating is a mechanism designated for prediction per se. As any sentence is incrementally processed, each word undergoes activation, retrieval, and integration. The notion of pre-updating, as we see it, merely means that retrieval can be achieved even without the need to wait for bottom–up activation when top–down activation is strong enough to reach the retrieval threshold. In this case, the retrieved content will simply move on to the following processing stages, that is, structure building, semantic integration, thematic role assignment, and so forth, similar to a word that had actually appeared in the input (but see Discussion for potential differences). This retrieval threshold will be more likely reached (before the realization of the word in the input) when the sentence being processed is more constraining (leading to a stronger prediction) and when predictive validity is high. The threshold could also potentially vary between individuals, such that for similar prediction strength, different comprehenders would be more or less likely to pre-update.

The findings mentioned above, demonstrating priming effects in RTs and in N400 amplitudes, can be explained by pre-activation alone, without the need to appeal to an additional process of pre-updating. This is also the case for other manifestations of prediction in sentence processing, such as anticipatory eye movement in the visual world paradigm (e.g., Boland, 2005; Kamide, Altmann, & Haywood, 2003; Altmann & Kamide, 1999). As all of these findings reflect activations of stored representations, without direct examination of representations built in WM, they can only provide evidence for pre-activation. Lau et al. (2013) presented findings that they claimed were more relevant to pre-updating. The authors showed that the N400 effect in word pairs was affected by predictive validity, manipulated by changing the proportion of related word pairs in the experimental context. Greater facilitation, reflected by a reduced N400 amplitude, was observed in the context that contained a larger proportion of related words, thereby encouraging prediction. This finding was suggested by the authors to indicate pre-updating, under the assumption that high predictive validity leads to pre-updating of WM representations, and the updated representation in turn affects activations, leading to a greater relatedness effect, that is, reduced N400. However, there is no reason to assume that top–down control cannot directly enhance or limit the spread of activations, without the mediation of WM representations (see, e.g., Van Berkum, 2009). Therefore, these results do not provide unambiguous evidence in favor of pre-updating. It should also be noted that the processing of word pairs may differ substantially from sentence processing, with participants adopting prediction strategies specific to the task (e.g., lexical or semantic decision to the second word in the pair).

The goal of the current study was therefore to find more direct support for pre-updating by looking for evidence for integration processes before the onset of a highly predictable word in a sentence, as well as for decreased integration demands when encountering the predicted word. We hypothesized that these effects would be reflected in the P600 ERP component, which is a measure for integration processes. The P600 component is a positive deflection in the EEG, with a posterior distribution over the scalp. This component was initially observed in response to syntactic anomalies such as violation of subcategorization constraints (e.g., Osterhout & Holcomb, 1992) and agreement errors (e.g., Hagoort, Brown, & Groothusen, 1993), as well as in “garden path” sentences, where an initial structure needs to be reanalyzed (e.g., Hagoort, Brown, & Osterhout, 1999; Osterhout, Holcomb, & Swinney, 1994; Osterhout & Holcomb, 1992). In addition, the P600 was shown to be elicited in grammatical sentences that do not involve reanalysis. Importantly, increased P600 amplitude is observed when a long-distance dependency is completed (e.g., Gouvea, Phillips, Kazanina, & Poeppel, 2010; Phillips, Kazanina, & Abada, 2005; Felser, Clahsen, & Münte, 2003; Fiebach, Schlesewsky, & Friederici, 2002; Kaan, Harris, Gibson, & Holcomb, 2000). For example, an increased P600 amplitude is measured on the verb “imitated” in a sentence such as (1a) relative to (1b) (Kaan et al., 2000). The difference between these sentences is that, at the verb in (1a), integration with the filler occurs. In contrast, the processing of the verb in (1b) does not include this additional process. Thus, there are more integration demands at this point in (1a) relative to (1b).

  • 1. 

    a. Emily wondered who the performer in the concert had imitated __ for the audience's amusement.

    b. Emily wondered whether the performer in the concert had imitated a pop star for the audience's amusement

The P600 was also found in “semantic illusion” contexts, namely syntactically sound sentences that are semantically anomalous due to thematic role reversal or thematic violations (the “Semantic P600,” see, e.g., Chow & Phillips, 2013; Kuperberg, Kreher, Sitnikova, Caplan, & Holcomb, 2007; Hoeks, Stowe, & Doedens, 2004). These and other findings have led to the suggestion that P600 amplitude reflects integration difficulty (Brouwer, Fitz, & Hoeks, 2012; Kaan et al., 2000; for further discussion regarding the functional nature of the P600, also see Chow & Phillips, 2013).2

In the current study, we adopt the suggestion that the P600 reflects some form of integration, without committing to a specific characterization of the processes reflected by the P600. In our case, it is sufficient to assume that any of the processes involved in integrating a word—syntactic structure building, dependency formation, thematic role assignment, semantic integration, and so forth—affects P600 amplitude to hypothesize that pre-updating of a predicted word, will be reflected by increased P600 amplitude at the point where pre-updating occurs, followed by decreased P600 amplitude when the already pre-updated word is encountered.

A Reanalysis of EEG Data

To test the feasibility of finding a measurable P600 difference due to pre-updating, we reanalyzed EEG data collected for a previous study (Ness & Meltzer-Asscher, 2018). The data were collected from 24 participants (14 men), native Hebrew speakers, with an average age of 25.7 years (range = 19–37 years). Eighty-four experimental sentences were used, with sentence constraints ranging from 53.6% to 100%. For the new analysis, we divided the sentences such that half were classified as high constraint and half as low constraint (example sentences are provided in Appendix A). The average constraint in the high-constraint sentences was 89.5% (range = 80–100%), and the average constraint in the low-constraint sentences was 68% (range = 50–80%). The critical word in the sentences was the verb, on which the prediction was generated. Verbs in the high-constraint and low-constraint conditions were matched on length (p = .799), frequency (p = .898, corpus: Linzen, 2009) and position in the sentence (p = .906, measured in number of words). Because of the design of the original experiment, after the critical verb the experimental sentences continued with either the predicted word, a congruent but unexpected word, or an anomalous word. Sentences were presented word-by-word in the middle of the screen for 200 msec, with a 300-msec ISI. For a more detailed description of the materials, procedure, and EEG recording, see Ness and Meltzer-Asscher (2018).

Based on the typical time window and scalp distribution of the P600, mean amplitudes over 500–800 msec from the verb onset for all centroparietal and parietal electrodes were entered into a repeated-measures ANOVA with the factors Constraint (high, low) and Electrode (nine levels: CP5, CP1, CP2, CP6, P7, P3, Pz, P4, P8). A main effect of Constraint was found, F(1, 22) = 7.05, p = .014, such that verbs in the high-constraint sentences yielded a larger P600 amplitude than those in the low-constraint sentences. Grand-averaged ERPs at a representative electrode and scalp distribution of the P600 component are provided in Figure 1.

Figure 1. 

Grand-averaged ERPs and scalp distribution for the reanalysis results.

Figure 1. 

Grand-averaged ERPs and scalp distribution for the reanalysis results.

These results provide initial evidence for pre-updating, as they indicate that more integration processes take place at the verb when it has a highly predictable argument (high constraint), relative to when there is no highly predictable argument (low constraint). A caveat to these results is that the critical word (i.e., the verb) as well as the following words were not identical between the high- and low-constraint sentences. In addition, we could not compare the ERPs to the actual predicted word between the high- and low-constraint sentences to test whether its integration is easier when pre-updating had already occurred at the verb. This was because, in the original design, the predicted word appeared in only one third of the trials (which would result in fourteen trials in each condition) and because the predicted word was presented immediately following the verb and its signal would therefore be contaminated by the P600 effect at the verb.

This Study

This study was designed to overcome the limitations of the re-analysis we conducted as well as replicate the result observed on the verb. Moreover, as pre-updating involves representations built online in WM, we hypothesized that the extent to which an individual tends to engage in pre-updating depends on their WM capacity.

In the present experiment, participants read strongly and weakly constraining sentences in Hebrew. ERPs were measured on the predictable noun phrase (NP) as well as on the preceding verb. These two critical words were separated by an additional word, the Hebrew accusative case marker, to make sure that the signal on the NP is not contaminated by effects originating at the verb. Participants' WM capacity was assessed via a reading span test.

If pre-updating indeed occurs and is more likely to occur when the prediction is stronger, a P600 effect is predicted on the verb in the strongly constraining sentences, reflecting integration of the predicted upcoming argument (the following NP). On the NP, an opposite effect is predicted, namely increased P600 in the low-constraint sentences, reflecting the benefits of pre-updating, that is, easier integration of a word that had already been pre-updated. Moreover, if participants with higher reading span exhibit a higher tendency to engage in pre-updating, then the effects should be greater for these participants.

METHODS

Participants

Participants were 37 Tel-Aviv University students (18 men), all native Hebrew speakers, with an average age of 25.4 years (range = 19–40 years). Participants were given course credit or paid 60 NIS (∼$15) for their participation. One participant was excluded from the analysis due to excessive artifacts. The experimental protocol received approval from the ethics committee in Tel Aviv University.

Materials

The materials consisted of 52 sentence pairs. Each pair included a high-constraint sentence and a low-constraint sentence (based on a cloze probability questionnaire, as detailed below). See Table 1 for an example set and Appendix B for all materials. The sentences in each pair differed at the beginning but were identical from the critical words onward. The critical words were the verb after which the sentence constraint differed between the conditions and the NP that followed it. To avoid contaminating the ERPs measured on the NP by effects stemming at the verb, these words were separated by the word “et,” the Hebrew accusative case marker. This case marker does not bear any agreement features, and it was therefore identical in all sentences and could not provide any indication of the upcoming noun. The number of words before the verb did not differ between the conditions (p = .890), nor did the length or frequency of the word before the verb (p = .163 and p = .383, respectively). Materials were divided into two lists according to a Latin square such that each participant saw only one sentence from each pair (i.e., 26 items in each condition and 52 sentences in total). Presentation order was randomized for each participant.

Table 1. 

Example Set

Example Set
Example Set

The critical words (the verb and the NP) are marked in bold. ACC = accusative case; COP = copula.

Cloze Probability Questionnaires

Two cloze probability questionnaires were conducted. Both included sentence fragments, and participants were instructed to complete each sentence with the first completion that comes to mind. In the first questionnaire, the sentences were presented truncated after the verb (i.e., the presented sentence frame included the verb). This questionnaire included 66 sentence pairs, divided into two lists such that each participant saw only one sentence from each pair. The order of presentation was randomized for each participant. One hundred two participants completed this questionnaire (average age = 25.4 years, 29 men). Based on this questionnaire, the 52 experimental sets were chosen. The average constraint in the high-constraint condition was 72.4% (range = 50–100%), meaning that, on average, 72.4% of completions provided by participants in the questionnaire were “et” (the accusative case marker), followed by the most commonly provided noun. The average constraint in the low-constraint condition was 18.8% (range = 0–50%, again, percentage reflecting “et” + the most commonly provided noun). When counting completions of the same lexical item whether it was preceded by the accusative case marker or not, the average constraint in the high-constraint condition was 80.5% and the average constraint in the low-constraint condition was 27.3%. Overall, the accusative case mark itself (with any noun) was predictable in both conditions (high, 87.0%; low, 74.8%).

In the second cloze probability questionnaire, the sentences were truncated before the verb. Because the verb is a critical word in the experiment (i.e., its ERP is of interest), this questionnaire was aimed to make sure that the verb's cloze probability did not differ between conditions, which would lead to an N400 effect. The questionnaire included the 52 experimental sets, divided into two lists such that each participant saw only one sentence from each pair. The order of presentation was randomized for each participant. This questionnaire was also completed by 102 participants, different from the ones who completed the first questionnaire (average age = 24.3, 25 men). The average cloze probability of the verbs was 0.7% and 1.1% in the high- and low-constraint conditions, respectively, with no significant difference between conditions (p = .595).

Procedure

Stimuli were presented using the E-Prime 2.0 software (Psychology Software Tools). Sentences were presented word-by-word in the middle of the screen for 250 msec, with a 350-msec ISI. A comprehension question appeared following 50% of the trials (randomly distributed). Each trial was preceded by a 1000-msec fixation point. After each trial a string of pound signs (####) appeared on the screen, and the participant pressed a button to start the next trial. Participants were encouraged to take as many breaks as needed. Before the experiment, participants completed a practice block of five trials.

Reading Span Test

To asses WM capacity, each participant also completed a reading span test.3 The test was performed following the main experiment or, for a few participants, in a separate session before their participation in the main experiment. The test's procedure is based on Daneman and Carpenter (1980), with minor differences. Participants read aloud a series of Hebrew sentences, after which they had to recall the last word of each sentence. The number of sentences in the series increased from two to six. Participants had three series in each level, and the last level at which a participant correctly recalled all words in at least two series was defined as this participant's reading span (i.e., when the participant failed to recall a word in two series in the same level the test was terminated and the participant's reading span was defined at the prior level). Two practice series (at the two-sentence level) were performed before the test, in which participants could make mistakes and ask questions.

EEG Recording and Preprocessing

The EEG was recorded using a BrainVision actiCap system with 32 Ag/AgCl scalp electrodes attached according to the 10–20 system. Two electrodes were used to monitor EOG, located at the outer canthi and the infraorbital ridge of the right and left eyes, respectively. Electrode impedances were kept below 5 kΩ for all scalp electrodes and below 15 Ω for the EOG electrodes. During recording, the EEG was referenced to Fp2 for most participants (the online reference electrode for five participants was Fp1, for technical reasons). The EEG was then rereferenced off-line (for all participants) to the average of the left and right mastoid electrodes. Data were collected at a 250-Hz sampling rate and low-pass filtered at 70 Hz. Data were then band-pass filtered between 0.1 and 30 Hz and segmented into 1200-msec epochs, including −200 to 1000 msec relative to the onset of the critical word. The 200 msec before the onset of the critical word were used for baseline correction. Trials contaminated by blinks, eye movements, excessive muscle activity, or amplifier blocking were rejected off-line before averaging and excluded from further analysis (this affected 5.36% of the trials).

EEG Data Analysis

Based on the typical time windows of the N400 and the P600, mean amplitudes over 300–500 msec and 500–800 msec, respectively, were analyzed. Electrodes were grouped based on their anteriority and laterality (anterior—left: F7, F3, Fp1, FC5, FC7, T1, C3; middle: Fz, Cz; right: F8, F4, Fp2, FC2, FC6, T7, C4; posterior—left: P7, P3, O1, CP5, CP1; middle: Pz, Oz; right: P8, P4, O2, CP2, CP6) to reduce the number of comparisons and the family-wise error rate (see Luck, 2014) while still allowing to assess the topography of the effects. Standardized reading span scores were entered to the analyses as a continuous covariate. This resulted in repeated-measures ANCOVAs with the factors Anteriority (anterior, posterior), Laterality (left, middle, right), and Constraint (high, low) and the covariate span. These analyses were conducted on time windows relative to both the verb onset and the noun onset and were followed by separate analyses for anterior and posterior sites (with the factors Laterality and Constraint and the covariate span). The Huyhn–Feldt adjustment for nonsphericity of variance was applied when the sphericity assumption was violated. In these cases, the corrected p value is reported with the original degrees of freedom.

RESULTS

Accuracy

Accuracy for the comprehension questions was significantly above chance for all participants. Mean accuracy rate was 93.37% (SD = 4.83). Accuracy data were subjected to a repeated-measures ANCOVA with the factor Constraint (high, low) and the covariate span. No significant main effects or interactions were found.

EEG

Grand-averaged ERPs and scalp distributions of the components are displayed in Figures 24. The results of the different ANCOVAs are provided in Table 2.

Figure 2. 

Grand-averaged ERPs and scalp distributions of the N400 (300–500 msec) and P600 (500–800 msec) of the verb and noun for all participants.

Figure 2. 

Grand-averaged ERPs and scalp distributions of the N400 (300–500 msec) and P600 (500–800 msec) of the verb and noun for all participants.

Figure 3. 

Grand-averaged ERPs and scalp distributions of the N400 (300–500 msec) and P600 (500–800 msec) of the verb and noun for high-span participants, with a reading span of 4 or higher (M = 4.33).

Figure 3. 

Grand-averaged ERPs and scalp distributions of the N400 (300–500 msec) and P600 (500–800 msec) of the verb and noun for high-span participants, with a reading span of 4 or higher (M = 4.33).

Figure 4. 

Grand-averaged ERPs and scalp distributions of the N400 (300–500 msec) and P600 (500–800 msec) of the verb and noun for low-span participants, with a reading span of 3 or lower (M = 2.61).

Figure 4. 

Grand-averaged ERPs and scalp distributions of the N400 (300–500 msec) and P600 (500–800 msec) of the verb and noun for low-span participants, with a reading span of 3 or lower (M = 2.61).

Table 2. 

Results of ANCOVAs on EEG Data

 dfVerbNoun
N400P600N400P600
FpFpFpFp
Constraint 1, 34 0.03 .866 8.45 .006 13.41 .001 1.81 .187 
Constraint × Span 1, 34 0.03 .827 2.12 .154 0.10 .753 3.66 .064 
Anteriority × Constraint 1, 34 1.51 .227 4.31 .045 15.02 <.001 0.01 .921 
Anteriority × Constraint × Span 1, 34 0.07 .796 6.27 .017 2.08 .158 2.88 .099 
Laterality × Constraint 2, 68 0.62 .509 0.72 .491 4.82 .014 2.24 .119 
Laterality × Constraint × Span 2, 68 1.84 .175 0.48 .620 3.56 .038 6.76 .003 
Anteriority × Laterality × Constraint 2, 68 0.36 .681 0.36 .662 0.72 .486 4.75 .015 
Anteriority × Laterality × Constraint × Span 2, 68 0.20 .796 0.04 .942 0.69 .499 0.22 .772 
  
 Verb
dfN400P600
AnteriorPosteriorAnteriorPosterior
FpFpFpFp
Constraint 1, 34 0.57 .455 0.25 .623 2.81 .103 11.94 <.001 
Constraint × Span 1, 34 0.71 .792 0.00 .994 0.03 .876 5.70 .023 
Laterality × Constraint 2, 68 0.23 .796 1.08 .336 0.56 .576 0.73 .486 
Laterality × Constraint × Span 2, 68 0.73 .488 2.58 .093 0.31 .731 0.41 .668 
  
 Noun
dfN400P600
AnteriorPosteriorAnteriorPosterior
FpFpFpFp
Constraint 1, 34 2.56 .119 19.06 <.001 2.16 .151 1.18 .282 
Constraint × Span 1, 34 0.30 .590 0.72 .404 1.65 .208 4.42 .043 
Laterality × Constraint 2, 68 4.42 .043 3.10 .064 0.58 .563 6.21 .007 
Laterality × Constraint × Span 2, 68 1.36 .264 5.09 .014 4.24 .018 5.57 .011 
 dfVerbNoun
N400P600N400P600
FpFpFpFp
Constraint 1, 34 0.03 .866 8.45 .006 13.41 .001 1.81 .187 
Constraint × Span 1, 34 0.03 .827 2.12 .154 0.10 .753 3.66 .064 
Anteriority × Constraint 1, 34 1.51 .227 4.31 .045 15.02 <.001 0.01 .921 
Anteriority × Constraint × Span 1, 34 0.07 .796 6.27 .017 2.08 .158 2.88 .099 
Laterality × Constraint 2, 68 0.62 .509 0.72 .491 4.82 .014 2.24 .119 
Laterality × Constraint × Span 2, 68 1.84 .175 0.48 .620 3.56 .038 6.76 .003 
Anteriority × Laterality × Constraint 2, 68 0.36 .681 0.36 .662 0.72 .486 4.75 .015 
Anteriority × Laterality × Constraint × Span 2, 68 0.20 .796 0.04 .942 0.69 .499 0.22 .772 
  
 Verb
dfN400P600
AnteriorPosteriorAnteriorPosterior
FpFpFpFp
Constraint 1, 34 0.57 .455 0.25 .623 2.81 .103 11.94 <.001 
Constraint × Span 1, 34 0.71 .792 0.00 .994 0.03 .876 5.70 .023 
Laterality × Constraint 2, 68 0.23 .796 1.08 .336 0.56 .576 0.73 .486 
Laterality × Constraint × Span 2, 68 0.73 .488 2.58 .093 0.31 .731 0.41 .668 
  
 Noun
dfN400P600
AnteriorPosteriorAnteriorPosterior
FpFpFpFp
Constraint 1, 34 2.56 .119 19.06 <.001 2.16 .151 1.18 .282 
Constraint × Span 1, 34 0.30 .590 0.72 .404 1.65 .208 4.42 .043 
Laterality × Constraint 2, 68 4.42 .043 3.10 .064 0.58 .563 6.21 .007 
Laterality × Constraint × Span 2, 68 1.36 .264 5.09 .014 4.24 .018 5.57 .011 

Factors: anteriority (anterior, posterior), laterality (left, middle, right), constraint (high, low). Covariate: span. N400 time window: 300–500 msec. P600 time window: 500–800 msec. The table includes all results involving the manipulated factor constraint (i.e., the main effect and any interaction that includes constraint). Critical results are marked in bold.

Verb

N400.

Mean amplitudes over the 300–500 msec time window (relative to the verb onset) were entered into a repeated-measures ANCOVA with the factors Anteriority (anterior, posterior), Laterality (left, middle, right), and Constraint (high, low) and the covariate span, followed by separate ANCOVAs for anterior and posterior sites. No main effects or interactions were found.

P600.

Mean amplitudes over the 500–800 msec time window (relative to the verb onset) were entered into a repeated-measures ANCOVA with the factors Anteriority (anterior, posterior), Laterality (left, middle, right), and Constraint (high, low) and the covariate span. A significant interaction was found between Anteriority and Constraint, F(1, 34) = 4.312, p = .045, as well as between Anteriority, Constraint, and Span, F(1, 34) = 6.267, p = .017. To further explore these interactions, separate ANCOVAs were performed on posterior and anterior sites. In the ANCOVA conducted over posterior electrodes, a significant effect of Constraint was found, F(1, 34) = 11.935, p = .001, such that P600 amplitude was higher in the high-constraint condition. In addition, a significant interaction was found between Constraint and Span, F(1, 34) = 5.698, p = .023, such that the P600 effect was greater for participants with a higher span. The relation between reading span and the P600 effect at the verb is plotted in Figure 5. These effects were not found in the ANCOVA conducted over anterior electrodes.

Figure 5. 

Size of P600 effect at the verb as a function of reading span score. The size of the P600 effect is calculated as the difference in mean amplitude between the high- and low-constraint conditions over the 600–800 msec time window (relative to verb onset), in posterior electrodes. ○, single participant; ♦, average across reading span score.

Figure 5. 

Size of P600 effect at the verb as a function of reading span score. The size of the P600 effect is calculated as the difference in mean amplitude between the high- and low-constraint conditions over the 600–800 msec time window (relative to verb onset), in posterior electrodes. ○, single participant; ♦, average across reading span score.

Noun

N400.

Mean amplitudes over the 300–500 msec time window (relative to the noun onset) were entered into a repeated-measures ANCOVA with the factors Anteriority (anterior, posterior), Laterality (left, middle, right), and Constraint (high, low) and the covariate span. A significant interaction was found between Anteriority and Constraint, F(1, 34) = 15.024, p < .001, but not between Anteriority, Constraint, and Span. To follow up on the two-way interaction, separate ANCOVAs were run on posterior and anterior sites. In the ANCOVA conducted over posterior electrodes, a significant effect of Constraint was found, F(1, 34) = 19.057, p < .001, such that N400 amplitude was higher in the low-constraint condition. There was no interaction between Constraint and Span. Neither of these effects was found in the ANCOVA conducted over anterior electrodes.

P600.

Mean amplitudes over the 500–800 msec time window (relative to the noun onset) were entered into a repeated-measures ANCOVA with the factors Anteriority (anterior, posterior), Laterality (left, middle, right), and Constraint (high, low) and the covariate span. There was no significant interaction between Anteriority and Constraint, but we did observe a trend toward a three-way interaction between Anteriority, Constraint, and Span, F(1, 34) = 2.883, p = .099. Although this interaction did not reach significance, we further explored this pattern by conducting separate ANCOVAs over posterior and anterior sites, because our a priori hypothesis was about posterior effects (the P600 component). In the ANCOVA conducted over posterior electrodes, no significant effect of Constraint was found, but there was a significant interaction between Constraint and Span, F(1, 34) = 4.422, p = .043, such that, for participants with higher reading span, there was a greater decrease in P600 amplitude in the high-constraint (relative to low-constraint) sentences. This effect was not found in the ANCOVA conducted over anterior electrodes. There was also a significant three-way interaction between Laterality, Constraint, and Span, F(2, 68) = 5.57, p = .011, suggesting that the observed P600 was left-lateralized.

To summarize the results, at the verb, a significant effect of Constraint as well as an interaction between Constraint and Span were found in posterior electrodes in the 500–800 msec time window, meaning that high-constraint sentences elicited an increased P600 amplitude (relative to low-constraint ones) and that this increase was greater for participants with higher WM capacity.

At the noun, a significant effect of Constraint was found in posterior electrodes in the 300–500 msec time window, with high-constraint sentences eliciting a decreased N400 amplitude (relative to low-constraint ones). This N400 effect was not affected by WM capacity. In addition, an interaction between Constraint and Span was found in posterior electrodes in the 500–800 msec time window on the noun, meaning that, for participants with higher reading span, there was a greater decrease in P600 amplitude in the high-constraint (relative to low-constraint) sentences.

DISCUSSION

In this study, we looked for evidence in support of the pre-updating process, namely, the integration of predicted material into the sentence representation in WM before its occurrence in the input. Our results showed increased P600 amplitude at the verb in the strongly constraining sentences, and this effect was greater for participants with higher WM span. We propose that this effect reflects integration processes before the onset of a highly predictable word, thus providing evidence in support of pre-updating. The benefits of pre-updating were observed on the predictable NP, where participants with higher reading span showed a greater decrease in P600 amplitude in the strongly constraining sentences (relative to the weakly constraining sentences), suggesting that the integration of an already pre-updated word was easier. These effects had the posterior distribution typical for the P600 component. It can be mentioned that a late frontal positivity would not be expected here, as this positivity is elicited by unexpected words in highly constraining sentences (e.g., Brothers et al., 2015; Federmeier et al., 2007) and is therefore thought to be related to coping with disconfirmed predictions (see further discussion of this anterior positivity below).

In addition, an N400 effect was found at the NP in the low-constraint sentences relative to the high-constraint sentences. This result replicates the very robust finding that the amplitude of the N400 is greater for words with low cloze probability relative to words with high cloze probability (e.g., Wlotko & Federmeier, 2012; DeLong et al., 2005; Kutas & Hillyard, 1984; Kutas et al., 1984). This N400 effect did not differ between the high- and low-span participants. The fact that the P600 effects were affected by WM capacity whereas the N400 effect was not, taken together with previous literature regarding the different processes reflected by these two components, corroborates the suggestion that, although both pre-activation and pre-updating are expected to occur more often in the high-constraint condition, we can see two distinct neural responses that behave differently, reflecting the two distinct processes. The P600 effects being affected by WM capacity is in line with the hypothesis that pre-updating involves WM representations while pre-activation occurs in long-term memory. These results support an architecture whereby what drives the differences due to WM capacity is not a difference in activation (as this would likely manifest as a difference in N400 amplitude) but rather a difference in the retrieval threshold, which determines the activation level sufficient for pre-updating. As mentioned in the Introduction, we see pre-updating as a process that is initiated only if the activation of the predicted content exceeds a certain threshold. The difference seen in the current study between participants with different reading spans can be interpreted as an indication that this threshold is generally lower for participants with a higher WM capacity, because pre-updating is less costly for them. Such an architecture would result in similar activations for all participants, but different likelihood to pre-update, in line with the current results.

Previous Studies on the Processing of Predictive Words

Several recent studies have investigated processing at the stage before the occurrence of a predictable word. Li, Zhang, Xia, and Swaab (2017) compared high- and low-constraint sentences in Mandarin Chinese, looking at the verb preceding a predictable or an anomalous noun. Their results did not show the P600 effect observed in our experiment, but rather a sustained anterior negativity elicited at the verb, as well as reduced beta power (19–25 Hz). There are several crucial differences between the current experiment and the experiment of Li et al. that may have led to these different results. The first and probably most influential difference is that the critical region in the Li et al. experiment consisted of a verb, followed by a classifier or an adjective, followed by a noun. Although the noun was predictable immediately after the verb, based on a sentence completion questionnaire, the classifier/adjective was not predictable. This means that in every trial the participant encountered an unpredictable word immediately following the verb. This was not the case in our study, as the verb in our sentences was always followed by the Hebrew accusative case marker “et,” a function word that was predicted based on the cloze probability questionnaire (and possibly even more predictable in the experimental context as it consistently appeared in all sentences), which was followed by the noun. This allowed participants in the current study to precisely predict the direct object of the verb (i.e., “et” + noun, forming the NP that immediately follows the verb), enabling its immediate integration. Another difference between the Li et al. experiment and ours is that, because the authors of the former experiment aimed to also test the consequences of prediction failure, the critical noun in their experiment was anomalous in half of the trials, leading to a disconfirmed prediction. It is conceivable that the proportion of disconfirmed predictions in the experimental context affects participants' tendency to commit to a prediction by pre-updating it, namely that if participants repeatedly have to endure the cost of prediction failure they would become more cautious and avoid pre-updating. Namely, the retrieval threshold that initiates pre-updating may not only vary between individuals but may also be adjusted via top–down control to adapt to different situations. In situations with a strong incentive to predict, the threshold would be lowered, leading to more frequent occurrence of pre-updating. In situations such as Li et al.'s experiment, the threshold would be raised to prevent commitment to prediction, which would likely be disconfirmed. This means that, although predicted candidates would still be pre-activated, pre-updating would be less likely to occur. These differences may explain why no P600 effect was found at the Li et al. study. Moreover, if participants in that study noticed that they consistently encounter an unpredictable word immediately after the verb but that the predictable word may still appear afterwards, then the results may reflect holding a (not-yet-integrated) prediction that is not expected to be realized immediately. This aligns well with the suggestion that the sustained anterior negativity, sometimes observed between the filler and gap sites in long-distance dependencies, is an index of WM load (e.g., Phillips et al., 2005; King & Kutas, 1995). In addition, in Li et al. the influence of WM capacity was not tested. It is possible that, because of random sampling, the participants in that study were on average with lower WM capacity than our participants, and as the effects seen in our experiment were driven by the high-span participants, this could further contribute to the difference in the results between the two studies.

Rommers, Dickson, Norton, Wlotko, and Federmeier (2017) have also looked at the stage before the occurrence of a predictable word. They performed a reanalysis of data from a study comparing high- and low-constraint sentences, ending with the most predictable word or with an unexpected word (Federmeier et al., 2007). While in the original study, ERPs of the final word were analyzed, the new analysis focused on the time–frequency domain, looking at the final word as well as a time window before its onset. Similar to Li et al., Rommers et al. also found an effect of constraint in the time–frequency domain before the predicted (or unexpected) word. However, the results of Rommers et al. show a decrease in alpha power (8–12 Hz), rather than in beta power, in the high-constraint sentences. These inconsistent results highlight the fact that in investigating predictive processes we must carefully consider factors that are likely to affect the nature of predictions, such as predictive validity within the experimental context (i.e., the proportion of trials in which predictions are disconfirmed), the immediacy of the predicted content, and so forth.

Fruchter, Linzen, Westerlund, and Marantz (2015) conducted an MEG experiment investigating processing of the adjective in adjective–NPs, when the adjective was predictive of the upcoming noun to different degrees (e.g., “stainless” is highly predictive of the noun “steel,” whereas “important” is not predictive of any particular noun). They found greater activity in the left middle temporal gyrus for highly predictive adjectives and a significant interaction in the same area between adjective predictivity and the frequency of the expected noun, such that higher noun frequency led to decreased activity when the adjective was very predictive. We believe that this interaction with the noun's frequency indicates that the effect found in that experiment is more likely to reflect pre-activation then pre-updating, as pre-activation is affected by lexical properties of the word being activated while pre-updating occurs after the word has already been retrieved (i.e., fully activated) and is therefore not expected to be affected by the lexical properties of the word.

To test whether lexical properties of the predicted words also underlie the P600 effect seen on the verb in our experiment, we performed a post hoc by-item correlation analysis between the log frequency of the predicted word (taken from the corpus of Linzen, 2009) and the average amplitude of the P600 effect on the verb (defined as the difference between the high- and low-constraint conditions, in the 500–800 msec time window, in the nine centroparietal and parietal electrodes). No significant correlation was found (Pearson's r = .076, n = 52, p = .594). We acknowledge that this is a null result, and therefore no strong conclusions can be drawn from it. However, this result is compatible with our conjecture that pre-activation, but not pre-updating, would be affected by lexical properties of the predicted word.

Individual WM Differences and the P600

In the current study, the P600 effects associated with pre-updating were greater for participants with higher reading span, indicating that participants with higher WM capacity are more likely to engage in this process. A couple of previous studies found that the P600 component was affected by individual differences in WM capacity. Nakano, Saron, and Swaab (2010) manipulated the first NP in simple subject-verb-object sentences such that the verb following it would either be plausible, implausible due to world knowledge, or in violation of thematic requirements (i.e., an inanimate subject when the verb requires an animate one). Their results showed that violation of world knowledge led to a similar N400 effect regardless of WM capacity. However, thematic violations led to distinctly different responses in the low- and high-span groups. An N400 effect (see discussion in the next subsection) was only observed in the low-span group, whereas a P600 effect was observed in the high-span group. The P600 effect in the high-span group may indicate that participants in this group made predictions regarding the upcoming verb based on the subject's properties (i.e., its inanimacy) and therefore had greater difficulty in integrating a verb requiring an animate agent.

Vos and Friederici (2003) showed that, in locally ambiguous sentences, disambiguation toward the less expected syntactic structure elicits a P600 effect at the disambiguating word only for participants with high WM capacity. This effect may indicate that high-span participants had to perform a syntactic reanalysis when the more predictable structure was incorrect because they had made structural predictions before the disambiguating word. One possible commonality between the findings in the two studies is that differences in the P600 component between high- and low-span participants tend to arise when the cause of the P600 effect is not syntactic difficulty per se (i.e., a complex or ungrammatical structure) but rather difficulty that stems from an incorrect prediction (albeit this would not necessarily be the case for all effects of WM capacity on the P600 component; see Kim, Oines, & Miyake, 2018). The cause of such differences may therefore be increased engagement in predictive processing by individuals with high WM abilities, stemming from the fact that these individuals have the available resources needed for the prediction itself, as well as sufficient resources to endure the costs incurred in the case of a contradictory input.

Individual WM Differences and the N400

Several studies have tested the influence of individual differences in WM capacity on the N400 component. For example, Van Petten, Weckerly, McIsaac, and Kutas (1997) compared the use of lexical context and sentence-level context by participants with low-, mid-, and high-WM capacity. Pairs of associated or unassociated words were embedded in congruent and incongruent sentences. Similar N400 effects for participants with low-, mid-, and high-WM capacity were observed when comparing associated and unassociated word pairs, but the contrast between unassociated word pairs in congruent and incongruent sentences revealed a significant N400 effect only in participants with mid- and high-WM capacity. These results were interpreted as indicating that WM capacity affects the degree to which purely sentence-level context is used for prediction, but not the degree to which lexical context is. A more qualitative difference in ERP responses of participants with high- and low-WM capacity was shown by Nakano and colleagues (2010). As explained in the previous subsection, in this experiment, thematic violations elicited an N400 effect only in the low-span group, whereas a P600 effect was observed in the high-span group. World knowledge violations led to similar N400 effects in both groups.

The generalization that seems to emerge from these two studies is that the amplitude of the N400 is not affected across the board by individual differences in WM capacity. Rather, specific types of information (i.e., sentence context, thematic requirements) can be differentially used for pre-activation by readers with higher or lower WM span, leading to differences in the N400 effect. In contrast, effects that stem from lexical context are not affected by WM capacity. We propose that this is consistent with the fact that in the current experiment the N400 effect did not differ between the two participant groups. This is so because, in our materials, the decrease in the amplitude of the N400 in the high-constraint sentences could have originated solely from their lexical content, facilitating retrieval of the predicted word via spreading activation. To test this proposal, we conducted a questionnaire in which participants were given lists of words, with each list containing only the content words of one experimental sentence, randomly ordered (i.e., not in the same order as in the original sentence). Participants were asked to read each list and provide the first association that comes to mind. The results of this questionnaire showed that the percentage of participants providing the experimental critical NP as an association for the word list was significantly greater in the high-constraint sentences than in the low-constraint ones (the average percentage was 3.4% and 37.6% for the low- and high-constraint sentences, respectively, t(52) = 10.52, p < .001). We do not claim that these results indicate that in our experiment only lexical context was used for prediction and sentence-level context was ignored. However, the results do indicate that an N400 effect would be expected in low-span participants even if these participants relied mostly (or only) on lexical context for prediction.

Pre-updating and Prediction Failure

As mentioned in the Introduction, pre-updating entails some form of commitment to a prediction. However, the integration carried out during pre-updating may not be identical to the integration occurring upon actually encountering a word in the input, which includes building syntactic structure, assigning thematic roles, integrating the semantic content of the word, and so forth. If the integration of a predicted word was identical to that of an actual word, disconfirming a pre-updated prediction would lead to a bona fide reanalysis. On the other hand, if the integration of a predicted word is tentative or not as complete as the integration of an actually encountered word, then disconfirmation of such prediction, though still costly, would not entail a typical reanalysis.

Previous ERP studies have shown that a congruent-unexpected word that appears instead of a highly predictable word elicits a frontal positivity termed the “frontal post-N400 positivity” (e.g., Ness & Meltzer-Asscher, 2018; DeLong, Urbach, Groppe, & Kutas, 2011; Federmeier et al., 2007). The frontal distribution of this component is distinct from the distribution of the P600 that is commonly elicited when a syntactic reanalysis is performed (e.g., Hagoort et al., 1999; Osterhout et al., 1994), and it was suggested to reflect discourse revision (i.e., increased difficulty in updating a context after it was already wrongly updated; Brothers et al., 2015) or inhibition of the falsely predicted word (Ness & Meltzer-Asscher, 2018). The fact that a frontal positivity rather than a P600 is elicited in such cases may indicate that a pre-updated prediction has a different status than an actually encountered word, displaying tentative integration, which can more easily be undone.

An Alternative Approach—Surprisal and Entropy Reduction

Throughout the article, the hypotheses and the interpretation of the results were framed under the assumption of serial parsing; however, our results can also be conceptualized in the framework of parallel and probabilistic parsing, which relate processing difficulty throughout a sentence to complexity matrices reflecting the amount of information conveyed in each word. One common matrix of this kind is surprisal, reflecting the probability of a word given the preceding words in the sentence (Levy, 2008; Hale, 2001). Many studies have shown that greater surprisal is associated with increased processing difficulty (e.g., Frank, Otten, Galli, & Vigliocco, 2015; Smith & Levy, 2013; Boston, Hale, Kliegl, Patil, & Vasishth, 2008; Demberg & Keller, 2008). Another relevant matrix is entropy reduction. Entropy reflects the degree of uncertainty regarding what is being conveyed, and it is therefore high when many possible outcomes have similar probabilities and lower when there is a very probable outcome. As each word is encountered, it affects expectations and the probability distribution changes. The degree to which a given word reduces uncertainty (i.e., reduces entropy) represents the amount of information gained (Hale, 2003). Processing more information incurs greater processing difficulty, and indeed, studies have shown that greater entropy reduction leads to increased reading times, independent of the contribution of surprisal (Linzen & Jaeger, 2016; Frank, 2013; Wu, Bachrach, Cardenas, & Schuler, 2010; see Hale, 2016, for a review).

We can now consider our materials in light of these findings. At the verb, cloze probability did not differ between conditions, and therefore, surprisal was similar. However, although entropy was always relatively high before the verb, certainty regarding the noun was much greater in the high-constraint condition than in the low-constraint one. This means that entropy reduction at the verb was greater in the high-constraint condition, which should lead to greater processing difficulty. At the noun, cloze probability was greater in the high-constraint condition than in the low-constraint one. Therefore, surprisal was lower in the high-constraint condition, which would lead to decreased processing difficulty (relative to the low-constraint condition). Regarding entropy reduction at the noun, whereas, as explained above, entropy before the noun was lower in the high-constraint condition than in the low-constraint one, entropy following the noun was likely high in both conditions (as in both conditions there was no predicted continuation for the sentence after the noun). This means that at the noun there was no entropy reduction in any condition (entropy either increased or did not change, both considered the same in terms of processing difficulty; see Lowder, Choi, Ferreira, & Henderson, 2018; Hale, 2016). To recap, in the high-constraint (relative to the low-constraint) condition, increased processing difficulty would be predicted at the verb due to entropy reduction (surprisal being similar in both conditions), and decreased processing difficulty would be predicted at the noun due to lower surprisal (entropy reduction likely being similar in the two conditions). These predictions are in line with our results, as the increased P600 on the verb in the high-constraint condition may be taken to index greater entropy reduction and the decreased N400 and/or P600 on the noun in the high-constraint condition may be taken to reflect lower surprisal.

Additional work is needed to provide an explanation within this framework as to why the observed P600 effects depended on WM capacity. Possibly, this can be done by assuming that individuals with higher WM capacity have a larger “beam size”, namely, they consider more possible continuations, which would vary the entropy experienced by different individuals. In addition, making predictions regarding the specific ERP components that reflect surprisal and entropy reduction is not trivial, and further research is needed to account, within this framework, for why we see a P600 effect at the verb while at the noun we see both N400 and P600 effects (e.g., see Cho, Goldrick, Lewis, & Smolensky, 2018 for a mechanistic account for surprisal; such accounts can provide a means to make predictions regarding timing and neural activity).

Prediction Mechanisms in Sentence Processing

Figure 6 summarizes the proposed prediction mechanisms during sentence processing and how they are incorporated within the processing stages of a word appearing in sentence context. At every stage during sentence comprehension, multiple representations in long-term memory are pre-activated. The activation level of a word is affected by the previous context, as well as by the lexical properties of that word. Top–down control may also limit or enhance activations, mediating the influence of factors such as predictive validity, task demands, and so forth. Once the activation level of a certain word had passed a retrieval threshold, it is regarded as retrieved, that is, it can be integrated into the representation being built in WM. This threshold may differ between individuals, and it may also be adjusted by top–down control, adapting to how beneficial commitment to predictions is in a given situation (e.g., due to task demands, noisy input, predictive validity).

Figure 6. 

An outline of prediction mechanisms within the processing stages of a word.

Figure 6. 

An outline of prediction mechanisms within the processing stages of a word.

This means that top–down control can affect prediction in two ways: first, by limiting or enhancing activation levels, which enables differential influence on different representations. It is conceivable that this kind of adjustment would take place when a certain information type is more or less reliable in a given situation. For example, if when reading a fantasy book animacy requirements are often violated, then animacy may be less heavily relied upon for prediction, and therefore, words that violate an animacy requirement may be predicted and activated. Second, top–down control can affect prediction by raising or lowering the retrieval threshold, which enables a more general influence on the likelihood of any predicted content to reach retrieval before its realization in the input. This kind of adaptation will take place, for example, when predictive validity is low, meaning that predictions are often violated, and therefore, forming strong predictions is not beneficial.

If no word had passed the retrieval threshold before the realization of the next word in the input, then bottom–up activation causes retrieval of the word, and it is then integrated into the representation in WM. If a certain word did pass the threshold before the realization of the next word in the input, this word would then be pre-updated, meaning that a tentative integration would be initiated, until the pre-updated word can be matched against the input. At this point, if the input matches the pre-updated word, integration is finalized. This stage would be less demanding than the integration of a word that had not been pre-updated. If the input does not match the pre-updated word, then the falsely predicted word is inhibited to enable integration of the actual input. This mismatch between prediction and input may also be a valuable trigger for learning mechanisms, improving future predictions.

A question remains regarding the specificity of the predicted content that can be pre-updated. In the current study, the strongly constraining sentences led to a high likelihood that a specific word would appear. However, sentences can be constructed in a way that will lead to a high likelihood of occurrence not of a specific word but of a word with a specific semantic or grammatical feature (e.g., animate, human, location, liquid substance; see Szewczyk & Schriefers, 2013). It is therefore possible that such features can also be pre-updated, even in the absence of prediction of a specific word. Whether or not this is plausible depends on how knowledge is stored in long-term memory, namely whether a feature can be highly activated when no specific word is highly activated, and on how representations are built in WM, namely whether they can be partial and contain features of an upcoming word without the actual word. These questions closely relate to the suggestion by Kuperberg and Jaeger (2016) that pre-updating can occur at different representational levels. The P600 effects found in the current study provide a valuable tool for exploring this issue further.

Appendix A. Example materials from Ness and Meltzer-Asscher (2018)

graphic

graphic

Appendix B. Translated Materials

The complete list of materials, translated from Hebrew, is provided in the table below. The critical verb and NP are marked in bold. These two words were always identical between conditions (within a set) and were separated by the Hebrew accusative case marker “et”.

With regard to the translation, we note the following:

  • • 

    Some of the critical nouns in Hebrew were modified by adjectives that were omitted in the translation. In Hebrew, adjectival modifiers are postnominal (not prenominal like in English), so these adjectives did not interfere with the prediction for the noun.

  • • 

    Similarly, some of the critical nouns in Hebrew were modified by a possessor (e.g., shelo—“his”), which was omitted in the translation. Possessors are postnominal in Hebrew, unlike in English, so they did not interfere with the prediction for the noun.

  • • 

    There was no repetition of verbs between sets in the Hebrew stimuli; for example, Hebrew has separate words for “put on (clothes)”, “put on (makeup)”, and “put on (jewelry)”.

  • • 

    In Hebrew, the verb as well as the NP (determiner + noun) always consisted of one orthographic word. For example, li-r'ot “to see”; ha-nof “the view”.

SetConditionSentence
High Right when they entered the hotel room, Yossi drew the curtain in order to see the view of the sea 
Low While Liron was talking to Adi in the corridor, Yossi opened the door to see the view of the sea 
High The bridesmaids waited expectantly as the excited bride went to put on the dress she chose 
Low Mor lingered in the morning searching her closet since she wanted to put on the dress she chose 
High The full cup remained on the table because Yoni hurried out of the house this morning before he even finished to drink the coffee his roommate made 
Low Yoni got back from work late today, so before he went out to practice, he did not even have time to drink the coffee his roommate made 
High The room was crowded and stuffy, so I asked Amit to open the window since it was a pleasant day 
Low While Rotem was cooking, she asked Amit to open the window since it was a pleasant day 
High After Natalie had dried herself, she reached for the rack to hang the towel she had used 
Low After he had finished arranging the room, Doron asked Natalie to hang the towel she had used 
High Preparing for the festive dinner, Limor took out the plates and began to set the table which was decorated and laden with delicious food 
Low In preperation for the important meeting Limor had to return to the office to set the table which was decorated and laden with delicious food 
High The mischievous Yuval scattered and mixed all the pieces after Maor had finished assembling the puzzle that contained a thousand pieces 
Low As soon as he arrived, Meir sat down in the corner of the room to begin assembling the puzzle that contained a thousand pieces 
High Michal had already prepared the sweets and drinks for her birthday, but she had not finished baking the cake that everybody was looking forward to eat 
Low Michal has started working in a pastry shop, and of all things she mostly likes to bake the cake that everybody is looking forward to eat 
High Dana got wet when she had to walk in the rain since she had lost the umbrella on her way to work this morning 
Low Dana had to turn back since she had lost the umbrella on her way to work this morning 
10 High At the gas station, Yair bent down and connected the pump to inflate the tire because it lacked air 
10 Low Before his camping trip to the Kinneret, Yair was looking for the pump to inflate the tire because it lacked air 
11 High Since Omer collected stamps, he was careful not to destroy the stamp when he opened the envelope containing his SAT score 
11 Low Noa saw Omar passing for a moment in the living room, but he immediately went to his room and only then opened the envelope containing his SAT score 
12 High Due to all the shopping Galit did abroad, she could barely close the suitcase that was so full 
12 Low Because she was tired, Galit asked Eitan to close the suitcase that was so full 
13 High The room was very hot, so Noam looked for the remote control to turn on the air conditioner but the batteries were empty 
13 Low Noam came to his Grandma's house this afternoon to teach her how to turn on the air conditioner but the batteries were empty 
14 High After Yossi chose a wedding suit, he asked the seamstress to make a hem to shorten the pants because they were too long 
14 Low Last Sunday, Yossi asked Ronit to shorten the pants because they were too long 
15 High Nadav is about to move into an apartment that has no parking, so he intends to sell the car to save some money 
15 Low Nadav posted an ad on the bulletin board because he wanted to sell the car to save some money 
16 High Before the film even began, the kids had already finished devouring the popcorn mom had bought them 
16 Low As soon as they arrived, the hungry children sat down and began devouring the popcorn mom had bought them 
17 High While playing throw and catch, Yonatan jumped to catch the ball that his friends threw around 
17 Low In the yard this afternoon, Rexi tried to catch the ball the friends threw around 
18 High The players and the singer met for a reunion, years after they decided to break up the band because of disputes 
18 Low After the crisis, Omri decided to break up the band because of disputes 
19 High The elderly grandmother bent down to hug the grandson who was waiting to welcome her 
19 Low At the airport after her flight back, Meital went to hug the grandson who was waiting to welcome her 
20 High While walking in the street, little Itamar asked his mother if he could pet the dog that belonged to one of the neighbors 
20 Low During the tour, the children saw one of the farm workers going to pet the dog that belonged to one of the neighbors 
21 High Despite the appeal, the cruel lecturer decided to fail the student in the difficult course 
21 Low After the argument last week, there was a rumor that Yaakov was trying to fail the student in the difficult course 
22 High During the wedding ceremony, everyone began to clap as soon as the groom broke the glass which shattered into pieces 
22 Low While Michael was cleaning the house, he accidentally broke the glass which shattered into pieces 
23 High At his birthday party, Tomer blew on the cake to put out the candles and then everyone started to sing 
23 Low The friends stood waiting while Tomer went to put out the candles and then everyone started to sing 
24 High The tough boss got mad because the stairwell was dirty and so he decided to fire the maid who had been working there for years 
24 Low Because the company is in financial difficulties, next week the manager has to fire the maid who had been working there for years 
25 High Mom has taught Naama that after using the toilet you need to flush the water and wash your hands 
25 Low While she was getting ready in the morning, before living the room, Na'ama stopped for a moment in order to flush the water and wash her hands 
26 High The owner of the vineyard urged the workers to begin to harvest the grapes which had ripened recently 
26 Low Victoria went out to the flower beds in the palace courtyard, because she wanted to harvest the grapes which had ripened recently 
27 High The children brought mud home just after Orna had finished washing the floor and cleaning the whole house 
27 Low Last night after work, Orna had to wash the floor and clean the whole house 
28 High Idan was delayed when shopping at the vegetable store, because the scale wasn't working and it was impossible to weigh the vegetables at checkout, so the line grew longer 
28 Low Last Thursday, the digital scale wasn't working and it was impossible to weigh the vegetables at checkout, so the line grew longer 
29 High The minister was invited to the White House, and he even met the president at the official dinner held yesterday 
29 Low The young man was very excited when he met the president at the official dinner held yesterday 
30 High After the kids played Uno, mom asked Oren to collect the cards that were scattered on the carpet 
30 Low A few days ago, Oren began to collect the cards that were scattered on the carpet 
31 High Everyone laughed except for Yoav, because only he did not understand the joke the comedian had told 
31 Low Yoav got mad and was very hurt when he understood the joke the comedian had told 
32 High After the elderly woman finished her shopping at the supermarket, Alon offered to help her carry the bags to her house around the corner 
32 Low During the long walk, Alon got tired of carrying the bags to the house around the corner 
33 High Tamar turned her ear to mom so she would help her put on the earrings she had bought 
33 Low In honor of the festive occasion, Tamar wore makeup and chose to put on the earrings she had bought 
34 High In preparation for the fashion show, George finished to put make up on the model that was very extravagant and colorful 
34 Low Last night to have fun, Talia decided to put make up on the model that was very extravagant and colorful 
35 High Before his trip abroad, Dor went to the Interior Ministry to renew the passport but he forgot to bring a picture 
35 Low Itai stood in line for hours, until he reached the clerk to renew the passport but he forgot to bring a picture 
36 High Since Ofir isn't familiar with the library, the librarian helped him find the book he needed 
36 Low Ofir had searched for hours, but he couldn't find the book he needed 
37 High The toilet was flooded, so Guy got a plumber to open the blockage that they had for the past few days 
37 Low During the morning, Guy arrived at the apartment and went to open the blockage that they had for the past few days 
38 High Uri has a cold, so mom handed him a tissue to wipe the nose after he sneezed 
38 Low Early this morning, Uri used a paper napkin to wipe the nose after he sneezed 
39 High After the main course ended, the diners began to eat the dessert that the chef had prepared 
39 Low May was very hungry after her hard work, so even before the guests arrived she began to eat the dessert that the chef had prepared 
40 High During the acrobatics act, the acrobat stunned the audience with his impressive performances on stage 
40 Low Yesterday during the evening, Nir stunned the audience with his impressive performances on stage 
41 High Since Roy fell asleep on the bus, he missed the stop and got off farther away from home 
41 Low On Wednesday afternoon, Roy accidentally missed the stop and got off farther away from home 
42 High Since her cellphone fell and cracked, Nofar went to a repair shop to replace the screen and it was very expensive 
42 Low When Nofar walked around on the main street running errands, she needed to replace the screen and it was very expensive 
43 High When Ziv was about to leave the house, Shani asked him to throw the trash into the big garbage can near the entrance 
43 Low When Ziv was seating in his office, he missed when he tried to throw the trash into the big garbage can near the entrance 
44 High Only when he was at his house door, Moshe discover that he forgot the key when he left quickly in the morning 
44 Low After searching every compartment in his bag, Moshe realized he forgot the key when he left quickly in the morning 
45 High In the interrogation room, the police detective questioned the suspect in the investigating regarding the sting operation 
45 Low During the press conference, Ariel questioned the suspect in the investigating regarding the sting operation 
46 High In a newspaper article, the famous restaurant critic praised the restaurant that was recently opened in the city center 
46 Low When Lior spoke to Natalie, he constantly praised the restaurant that was recently opened in the city center 
47 High On Lag Ba'Omer, the boys threw some wood into the flames to strengthen the fire in which they were roasting potatoes 
47 Low While the friends sat around, Liad got up and went to strengthen the fire in which they were roasting potatoes 
48 High When the doctor came into the treatment room, he went to examine the patient that was waiting for him 
48 Low When Aviad arrived at work this morning, he first went to examine the patient that was waiting for him 
49 High The paint was dripping from the brush to the floor, when Avi painted the wall which left a big mess 
49 Low When Ohad was in kindergarten today, he painted the wall which left a big mess 
50 High Daphne jumped on a chair in the kitchen, yelling for Danny to come squash the cockroach in the corner 
50 Low Daphne walked carefully trying not to squash the cockroach in the corner 
51 High During the Treasure Hunt game, the friends tried together to solve the riddle which was very challenging 
51 Low On Monday, Lital spent the whole morning trying to solve the riddle which was very challenging 
52 High Sharon heard the bell ringing and went to open the door for the guests who came for dinner 
52 Low While Eden was in the kitchen she asked Alex to open the door for the guests who came for dinner 
SetConditionSentence
High Right when they entered the hotel room, Yossi drew the curtain in order to see the view of the sea 
Low While Liron was talking to Adi in the corridor, Yossi opened the door to see the view of the sea 
High The bridesmaids waited expectantly as the excited bride went to put on the dress she chose 
Low Mor lingered in the morning searching her closet since she wanted to put on the dress she chose 
High The full cup remained on the table because Yoni hurried out of the house this morning before he even finished to drink the coffee his roommate made 
Low Yoni got back from work late today, so before he went out to practice, he did not even have time to drink the coffee his roommate made 
High The room was crowded and stuffy, so I asked Amit to open the window since it was a pleasant day 
Low While Rotem was cooking, she asked Amit to open the window since it was a pleasant day 
High After Natalie had dried herself, she reached for the rack to hang the towel she had used 
Low After he had finished arranging the room, Doron asked Natalie to hang the towel she had used 
High Preparing for the festive dinner, Limor took out the plates and began to set the table which was decorated and laden with delicious food 
Low In preperation for the important meeting Limor had to return to the office to set the table which was decorated and laden with delicious food 
High The mischievous Yuval scattered and mixed all the pieces after Maor had finished assembling the puzzle that contained a thousand pieces 
Low As soon as he arrived, Meir sat down in the corner of the room to begin assembling the puzzle that contained a thousand pieces 
High Michal had already prepared the sweets and drinks for her birthday, but she had not finished baking the cake that everybody was looking forward to eat 
Low Michal has started working in a pastry shop, and of all things she mostly likes to bake the cake that everybody is looking forward to eat 
High Dana got wet when she had to walk in the rain since she had lost the umbrella on her way to work this morning 
Low Dana had to turn back since she had lost the umbrella on her way to work this morning 
10 High At the gas station, Yair bent down and connected the pump to inflate the tire because it lacked air 
10 Low Before his camping trip to the Kinneret, Yair was looking for the pump to inflate the tire because it lacked air 
11 High Since Omer collected stamps, he was careful not to destroy the stamp when he opened the envelope containing his SAT score 
11 Low Noa saw Omar passing for a moment in the living room, but he immediately went to his room and only then opened the envelope containing his SAT score 
12 High Due to all the shopping Galit did abroad, she could barely close the suitcase that was so full 
12 Low Because she was tired, Galit asked Eitan to close the suitcase that was so full 
13 High The room was very hot, so Noam looked for the remote control to turn on the air conditioner but the batteries were empty 
13 Low Noam came to his Grandma's house this afternoon to teach her how to turn on the air conditioner but the batteries were empty 
14 High After Yossi chose a wedding suit, he asked the seamstress to make a hem to shorten the pants because they were too long 
14 Low Last Sunday, Yossi asked Ronit to shorten the pants because they were too long 
15 High Nadav is about to move into an apartment that has no parking, so he intends to sell the car to save some money 
15 Low Nadav posted an ad on the bulletin board because he wanted to sell the car to save some money 
16 High Before the film even began, the kids had already finished devouring the popcorn mom had bought them 
16 Low As soon as they arrived, the hungry children sat down and began devouring the popcorn mom had bought them 
17 High While playing throw and catch, Yonatan jumped to catch the ball that his friends threw around 
17 Low In the yard this afternoon, Rexi tried to catch the ball the friends threw around 
18 High The players and the singer met for a reunion, years after they decided to break up the band because of disputes 
18 Low After the crisis, Omri decided to break up the band because of disputes 
19 High The elderly grandmother bent down to hug the grandson who was waiting to welcome her 
19 Low At the airport after her flight back, Meital went to hug the grandson who was waiting to welcome her 
20 High While walking in the street, little Itamar asked his mother if he could pet the dog that belonged to one of the neighbors 
20 Low During the tour, the children saw one of the farm workers going to pet the dog that belonged to one of the neighbors 
21 High Despite the appeal, the cruel lecturer decided to fail the student in the difficult course 
21 Low After the argument last week, there was a rumor that Yaakov was trying to fail the student in the difficult course 
22 High During the wedding ceremony, everyone began to clap as soon as the groom broke the glass which shattered into pieces 
22 Low While Michael was cleaning the house, he accidentally broke the glass which shattered into pieces 
23 High At his birthday party, Tomer blew on the cake to put out the candles and then everyone started to sing 
23 Low The friends stood waiting while Tomer went to put out the candles and then everyone started to sing 
24 High The tough boss got mad because the stairwell was dirty and so he decided to fire the maid who had been working there for years 
24 Low Because the company is in financial difficulties, next week the manager has to fire the maid who had been working there for years 
25 High Mom has taught Naama that after using the toilet you need to flush the water and wash your hands 
25 Low While she was getting ready in the morning, before living the room, Na'ama stopped for a moment in order to flush the water and wash her hands 
26 High The owner of the vineyard urged the workers to begin to harvest the grapes which had ripened recently 
26 Low Victoria went out to the flower beds in the palace courtyard, because she wanted to harvest the grapes which had ripened recently 
27 High The children brought mud home just after Orna had finished washing the floor and cleaning the whole house 
27 Low Last night after work, Orna had to wash the floor and clean the whole house 
28 High Idan was delayed when shopping at the vegetable store, because the scale wasn't working and it was impossible to weigh the vegetables at checkout, so the line grew longer 
28 Low Last Thursday, the digital scale wasn't working and it was impossible to weigh the vegetables at checkout, so the line grew longer 
29 High The minister was invited to the White House, and he even met the president at the official dinner held yesterday 
29 Low The young man was very excited when he met the president at the official dinner held yesterday 
30 High After the kids played Uno, mom asked Oren to collect the cards that were scattered on the carpet 
30 Low A few days ago, Oren began to collect the cards that were scattered on the carpet 
31 High Everyone laughed except for Yoav, because only he did not understand the joke the comedian had told 
31 Low Yoav got mad and was very hurt when he understood the joke the comedian had told 
32 High After the elderly woman finished her shopping at the supermarket, Alon offered to help her carry the bags to her house around the corner 
32 Low During the long walk, Alon got tired of carrying the bags to the house around the corner 
33 High Tamar turned her ear to mom so she would help her put on the earrings she had bought 
33 Low In honor of the festive occasion, Tamar wore makeup and chose to put on the earrings she had bought 
34 High In preparation for the fashion show, George finished to put make up on the model that was very extravagant and colorful 
34 Low Last night to have fun, Talia decided to put make up on the model that was very extravagant and colorful 
35 High Before his trip abroad, Dor went to the Interior Ministry to renew the passport but he forgot to bring a picture 
35 Low Itai stood in line for hours, until he reached the clerk to renew the passport but he forgot to bring a picture 
36 High Since Ofir isn't familiar with the library, the librarian helped him find the book he needed 
36 Low Ofir had searched for hours, but he couldn't find the book he needed 
37 High The toilet was flooded, so Guy got a plumber to open the blockage that they had for the past few days 
37 Low During the morning, Guy arrived at the apartment and went to open the blockage that they had for the past few days 
38 High Uri has a cold, so mom handed him a tissue to wipe the nose after he sneezed 
38 Low Early this morning, Uri used a paper napkin to wipe the nose after he sneezed 
39 High After the main course ended, the diners began to eat the dessert that the chef had prepared 
39 Low May was very hungry after her hard work, so even before the guests arrived she began to eat the dessert that the chef had prepared 
40 High During the acrobatics act, the acrobat stunned the audience with his impressive performances on stage 
40 Low Yesterday during the evening, Nir stunned the audience with his impressive performances on stage 
41 High Since Roy fell asleep on the bus, he missed the stop and got off farther away from home 
41 Low On Wednesday afternoon, Roy accidentally missed the stop and got off farther away from home 
42 High Since her cellphone fell and cracked, Nofar went to a repair shop to replace the screen and it was very expensive 
42 Low When Nofar walked around on the main street running errands, she needed to replace the screen and it was very expensive 
43 High When Ziv was about to leave the house, Shani asked him to throw the trash into the big garbage can near the entrance 
43 Low When Ziv was seating in his office, he missed when he tried to throw the trash into the big garbage can near the entrance 
44 High Only when he was at his house door, Moshe discover that he forgot the key when he left quickly in the morning 
44 Low After searching every compartment in his bag, Moshe realized he forgot the key when he left quickly in the morning 
45 High In the interrogation room, the police detective questioned the suspect in the investigating regarding the sting operation 
45 Low During the press conference, Ariel questioned the suspect in the investigating regarding the sting operation 
46 High In a newspaper article, the famous restaurant critic praised the restaurant that was recently opened in the city center 
46 Low When Lior spoke to Natalie, he constantly praised the restaurant that was recently opened in the city center 
47 High On Lag Ba'Omer, the boys threw some wood into the flames to strengthen the fire in which they were roasting potatoes 
47 Low While the friends sat around, Liad got up and went to strengthen the fire in which they were roasting potatoes 
48 High When the doctor came into the treatment room, he went to examine the patient that was waiting for him 
48 Low When Aviad arrived at work this morning, he first went to examine the patient that was waiting for him 
49 High The paint was dripping from the brush to the floor, when Avi painted the wall which left a big mess 
49 Low When Ohad was in kindergarten today, he painted the wall which left a big mess 
50 High Daphne jumped on a chair in the kitchen, yelling for Danny to come squash the cockroach in the corner 
50 Low Daphne walked carefully trying not to squash the cockroach in the corner 
51 High During the Treasure Hunt game, the friends tried together to solve the riddle which was very challenging 
51 Low On Monday, Lital spent the whole morning trying to solve the riddle which was very challenging 
52 High Sharon heard the bell ringing and went to open the door for the guests who came for dinner 
52 Low While Eden was in the kitchen she asked Alex to open the door for the guests who came for dinner 

Reprint requests should be sent to Tal Ness, Sagol School of Neuroscience, Tel Aviv University, Webb Building, Room 413, Ramat Aviv, Tel Aviv 69978, Israel, or via e-mail: talness@mail.tau.ac.il.

Notes

1. 

For discussion of a different approach, see An Alternative Approach—Surprisal and Entropy Reduction section.

2. 

The P600, characterized by a posterior scalp distribution, should be distinguished from the fPNP described above, indexing the costs of failed prediction.

3. 

Reading span is a limited measure of WM capacity, as it can also be affected by differences in language proficiency. Participants in the current study were likely rather homogenous in their language proficiency because they were all monolingual Hebrew-speaking university students, which would allow reading span scores to reflect WM capacity relatively well. However, future studies can disentangle WM capacity from language proficiency, possibly by using a factor analytic approach (e.g., Kim et al., 2018).

REFERENCES

Altmann
,
G. T.
, &
Kamide
,
Y.
(
1999
).
Incremental interpretation at verbs: Restricting the domain of subsequent reference
.
Cognition
,
73
,
247
264
.
Boland
,
J. E.
(
2005
).
Visual arguments
.
Cognition
,
95
,
237
274
.
Boston
,
M. F.
,
Hale
,
J.
,
Kliegl
,
R.
,
Patil
,
U.
, &
Vasishth
,
S.
(
2008
).
Parsing costs as predictors of reading difficulty: An evaluation using the Potsdam Sentence Corpus
.
Journal of Eye Movement Research
,
2
,
1
12
.
Brothers
,
T.
,
Swaab
,
T. Y.
, &
Traxler
,
M. J.
(
2015
).
Effects of prediction and contextual support on lexical processing: Prediction takes precedence
.
Cognition
,
136
,
135
149
.
Brouwer
,
H.
,
Fitz
,
H.
, &
Hoeks
,
J.
(
2012
).
Getting real about semantic illusions: Rethinking the functional role of the P600 in language comprehension
.
Brain Research
,
1446
,
127
143
.
Cho
,
P. W.
,
Goldrick
,
M.
,
Lewis
,
R. L.
, &
Smolensky
,
P.
(
2018
).
Dynamic encoding of structural uncertainty in gradient symbols
. In
Proceedings of the 8th Workshop on Cognitive Modeling and Computational Linguistics (CMCL 2018)
(pp.
19
28
).
Chow
,
W. Y.
, &
Phillips
,
C.
(
2013
).
No semantic illusions in the “Semantic P600” phenomenon: ERP evidence from Mandarin Chinese
.
Brain Research
,
1506
,
76
93
.
Cowan
,
N.
(
2010
).
The magical mystery four: How is working memory capacity limited, and why?
Current Directions in Psychological Science
,
19
,
51
57
.
Daneman
,
M.
, &
Carpenter
,
P. A.
(
1980
).
Individual differences in working memory and reading
.
Journal of Verbal Learning and Verbal Behavior
,
19
,
450
466
.
DeLong
,
K. A.
,
Urbach
,
T. P.
,
Groppe
,
D. M.
, &
Kutas
,
M.
(
2011
).
Overlapping dual ERP responses to low cloze probability sentence continuations
.
Psychophysiology
,
48
,
1203
1207
.
DeLong
,
K. A.
,
Urbach
,
T. P.
, &
Kutas
,
M.
(
2005
).
Probabilistic word pre-activation during language comprehension inferred from electrical brain activity
.
Nature Neuroscience
,
8
,
1117
1121
.
Demberg
,
V.
, &
Keller
,
F.
(
2008
).
Data from eye-tracking corpora as evidence for theories of syntactic processing complexity
.
Cognition
,
109
,
193
210
.
Ehrlich
,
S. F.
, &
Rayner
,
K.
(
1981
).
Contextual effects on word perception and eye movements during reading
.
Journal of Verbal Learning and Verbal Behavior
,
20
,
641
655
.
Federmeier
,
K. D.
, &
Kutas
,
M.
(
1999
).
A rose by any other name: Long-term memory structure and sentence processing
.
Journal of Memory and Language
,
41
,
469
495
.
Federmeier
,
K. D.
,
Wlotko
,
E. W.
,
De Ochoa-Dewald
,
E.
, &
Kutas
,
M.
(
2007
).
Multiple effects of sentential constraint on word processing
.
Brain Research
,
1146
,
75
84
.
Felser
,
C.
,
Clahsen
,
H.
, &
Münte
,
T. F.
(
2003
).
Storage and integration in the processing of filler-gap dependencies: An ERP study of topicalization and wh-movement in German
.
Brain and Language
,
87
,
345
354
.
Fiebach
,
C. J.
,
Schlesewsky
,
M.
, &
Friederici
,
A. D.
(
2002
).
Separating syntactic memory costs and syntactic integration costs during parsing: The processing of German wh-questions
.
Journal of Memory and Language
,
47
,
250
272
.
Forster
,
K. I.
(
1981
).
Priming and the effects of sentence and lexical contexts on naming time: Evidence for autonomous lexical processing
.
Quarterly Journal of Experimental Psychology
,
33
,
465
495
.
Frank
,
S. L.
(
2013
).
Uncertainty reduction as a measure of cognitive load in sentence comprehension
.
Topics in Cognitive Science
,
5
,
475
494
.
Frank
,
S. L.
,
Otten
,
L. J.
,
Galli
,
G.
, &
Vigliocco
,
G.
(
2015
).
The ERP response to the amount of information conveyed by words in sentences
.
Brain and Language
,
140
,
1
11
.
Fruchter
,
J.
,
Linzen
,
T.
,
Westerlund
,
M.
, &
Marantz
,
A.
(
2015
).
Lexical pre-activation in basic linguistic phrases
.
Journal of Cognitive Neuroscience
,
27
,
1912
1935
.
Gouvea
,
A. C.
,
Phillips
,
C.
,
Kazanina
,
N.
, &
Poeppel
,
D.
(
2010
).
The linguistic processes underlying the P600
.
Language and Cognitive Processes
,
25
,
149
188
.
Green
,
C.
(
2017
).
Usage-based linguistics and the magic number four
.
Cognitive Linguistics
,
28
,
209
237
.
Hagoort
,
P.
,
Brown
,
C. M.
, &
Groothusen
,
J.
(
1993
).
The syntactic positive shift (SPS) as an ERP measure of syntactic processing
.
Language and Cognitive Processes
,
8
,
439
483
.
Hagoort
,
P.
,
Brown
,
C. M.
, &
Osterhout
,
L.
(
1999
).
The neurocognition of syntactic processing
. In
C. M.
Brown
&
P.
Hagoort
(Eds.),
The neurocognition of language
(pp.
273
317
).
Oxford
:
Oxford University Press
.
Hale
,
J.
(
2001, June
).
A probabilistic Earley parser as a psycholinguistic model
. In
Proceedings of the second meeting of the North American Chapter of the Association for Computational Linguistics on Language technologies
(pp.
1
8
).
Stroudsburg, PA
:
Association for Computational Linguistics
.
Hale
,
J.
(
2003
).
The information conveyed by words in sentences
.
Journal of Psycholinguistic Research
,
32
,
101
123
.
Hale
,
J.
(
2016
).
Information-theoretical Complexity Metrics
.
Language and Linguistics Compass
,
10
,
397
412
.
Hoeks
,
J. C.
,
Stowe
,
L. A.
, &
Doedens
,
G.
(
2004
).
Seeing words in context: The interaction of lexical and sentence level information during reading
.
Cognitive Brain Research
,
19
,
59
73
.
Kaan
,
E.
,
Harris
,
A.
,
Gibson
,
E.
, &
Holcomb
,
P.
(
2000
).
The P600 as an index of syntactic integration difficulty
.
Language and Cognitive Processes
,
15
,
159
201
.
Kamide
,
Y.
,
Altmann
,
G. T.
, &
Haywood
,
S. L.
(
2003
).
The time-course of prediction in incremental sentence processing: Evidence from anticipatory eye movements
.
Journal of Memory and Language
,
49
,
133
156
.
Kim
,
A. E.
,
Oines
,
L.
, &
Miyake
,
A.
(
2018
).
Individual differences in verbal working memory underlie a tradeoff between semantic and structural processing difficulty during language comprehension: An ERP investigation
.
Journal of Experimental Psychology: Learning, Memory, and Cognition
,
44
,
406
420
.
King
,
J. W.
, &
Kutas
,
M.
(
1995
).
Who did what and when? Using word- and clause-level ERPs to monitor working memory usage in reading
.
Journal of Cognitive Neuroscience
,
7
,
376
395
.
Kuperberg
,
G. R.
, &
Jaeger
,
T. F.
(
2016
).
What do we mean by prediction in language comprehension?
Language, Cognition and Neuroscience
,
31
,
32
59
.
Kuperberg
,
G. R.
,
Kreher
,
D. A.
,
Sitnikova
,
T.
,
Caplan
,
D. N.
, &
Holcomb
,
P. J.
(
2007
).
The role of animacy and thematic relationships in processing active English sentences: Evidence from event-related potentials
.
Brain and Language
,
100
,
223
237
.
Kutas
,
M.
, &
Hillyard
,
S. A.
(
1984
).
Brain potentials during reading reflect word expectancy and semantic association
.
Nature
,
307
,
161
163
.
Kutas
,
M.
,
Lindamood
,
T. E.
, &
Hillyard
,
S. A.
(
1984
).
Word expectancy and event-related brain potentials during sentence processing
. In
S.
Kornblum
&
J.
Requin
(Eds.),
Preparatory states and processes
(pp.
217
237
).
Hillsdale, NJ
:
Erlbaum
.
Lau
,
E. F.
,
Holcomb
,
P. J.
, &
Kuperberg
,
G. R.
(
2013
).
Dissociating N400 effects of prediction from association in single-word contexts
.
Journal of Cognitive Neuroscience
,
25
,
484
502
.
Levy
,
R.
(
2008
).
Expectation-based syntactic comprehension
.
Cognition
,
106
,
1126
1177
.
Li
,
X.
,
Zhang
,
Y.
,
Xia
,
J.
, &
Swaab
,
T. Y.
(
2017
).
Internal mechanisms underlying anticipatory language processing: Evidence from event-related-potentials and neural oscillations
.
Neuropsychologia
,
102
,
70
81
.
Linzen
,
T.
(
2009
).
Corpus of blog postings collected from the Israblog website
.
Tel Aviv, Israel
:
Tel Aviv University
.
Linzen
,
T.
, &
Jaeger
,
T. F.
(
2016
).
Uncertainty and expectation in sentence processing: Evidence from subcategorization distributions
.
Cognitive Science
,
40
,
1382
1411
.
Lowder
,
M. W.
,
Choi
,
W.
,
Ferreira
,
F.
, &
Henderson
,
J. M.
(
2018
).
Lexical predictability during natural reading: Effects of surprisal and entropy reduction
.
Cognitive Science
,
42(Suppl. 4)
,
1166
1183
.
Luck
,
S. J.
(
2014
).
An introduction to the event-related potential technique
.
Cambridge, MA
:
MIT Press
.
Miller
,
G. A.
(
1956
).
The magical number seven plus or minus two: Some limits on our capacity for processing information
.
Psychological Review
,
63
,
81
97
.
Nakano
,
H.
,
Saron
,
C.
, &
Swaab
,
T. Y.
(
2010
).
Speech and span: Working memory capacity impacts the use of animacy but not of world knowledge during spoken sentence comprehension
.
Journal of Cognitive Neuroscience
,
22
,
2886
2898
.
Ness
,
T.
, &
Meltzer-Asscher
,
A.
(
2018
).
Lexical inhibition due to failed prediction: Behavioral evidence and ERP correlates
.
Journal of Experimental Psychology: Learning, Memory, and Cognition
,
44
,
1269
1285
.
Osterhout
,
L.
, &
Holcomb
,
P. J.
(
1992
).
Event-related brain potentials elicited by syntactic anomaly
.
Journal of Memory and Language
,
31
,
785
806
.
Osterhout
,
L.
,
Holcomb
,
P. J.
, &
Swinney
,
D. A.
(
1994
).
Brain potentials elicited by garden-path sentences: Evidence of the application of verb information during parsing
.
Journal of Experimental Psychology: Learning, Memory, and Cognition
,
20
,
786
803
.
Phillips
,
C.
,
Kazanina
,
N.
, &
Abada
,
S. H.
(
2005
).
ERP effects of the processing of syntactic long-distance dependencies
.
Cognitive Brain Research
,
22
,
407
428
.
Rommers
,
J.
,
Dickson
,
D. S.
,
Norton
,
J. J.
,
Wlotko
,
E. W.
, &
Federmeier
,
K. D.
(
2017
).
Alpha and theta band dynamics related to sentential constraint and word expectancy
.
Language, Cognition and Neuroscience
,
32
,
576
589
.
Schwanenflugel
,
P. J.
, &
LaCount
,
K. L.
(
1988
).
Semantic relatedness and the scope of facilitation for upcoming words in sentences
.
Journal of Experimental Psychology: Learning, Memory, and Cognition
,
14
,
344
354
.
Schwanenflugel
,
P. J.
, &
Shoben
,
E. J.
(
1985
).
The influence of sentence constraint on the scope of facilitation for upcoming words
.
Journal of Memory and Language
,
24
,
232
252
.
Smith
,
N. J.
, &
Levy
,
R.
(
2013
).
The effect of word predictability on reading time is logarithmic
.
Cognition
,
128
,
302
319
.
Stanovich
,
K. E.
, &
West
,
R. F.
(
1983
).
On priming by a sentence context
.
Journal of Experimental Psychology: General
,
112
,
1
36
.
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
,
297
314
.
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
,
382
392
.
Traxler
,
M. J.
, &
Foss
,
D. J.
(
2000
).
Effects of sentence constraint on priming in natural language comprehension
.
Journal of Experimental Psychology: Learning, Memory, and Cognition
,
26
,
1266
1282
.
Van Berkum
,
J. J.
(
2009
).
The neuropragmatics of ‘simple’ utterance comprehension: An ERP review
. In
U.
Sauerland
&
K.
Yatsushiro
(Eds.),
Semantics and pragmatics: From experiment to theory
(pp.
276
316
).
Basingstoke, UK
:
Palgrave Macmillan
.
Van Petten
,
C.
, &
Luka
,
B. J.
(
2012
).
Prediction during language comprehension: Benefits, costs, and ERP components
.
International Journal of Psychophysiology
,
83
,
176
190
.
Van Petten
,
C.
,
Weckerly
,
J.
,
McIsaac
,
H. K.
, &
Kutas
,
M.
(
1997
).
Working memory capacity dissociates lexical and sentential context effects
.
Psychological Science
,
8
,
238
242
.
Vos
,
S. H.
, &
Friederici
,
A. D.
(
2003
).
Intersentential syntactic context effects on comprehension: The role of working memory
.
Cognitive Brain Research
,
16
,
111
122
.
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
,
356
366
.
Wu
,
S.
,
Bachrach
,
A.
,
Cardenas
,
C.
, &
Schuler
,
W.
(
2010, July
).
Complexity metrics in an incremental right-corner parser
. In
Proceedings of the 48th annual meeting of the Association for Computational Linguistics
(pp.
1189
1198
).
Stroudsburg, PA
:
Association for Computational Linguistics
.