Participants were 53 undergraduate students. One participant was excluded from analysis, due to excessive EEG artifact. Twenty-six participants (13 female) were assigned to the low-variability stimuli and were aged 18–28 (mean = 22.3). Another 26 participants (13 female) were assigned to the high-variability stimuli and were aged 18–34 (mean = 22.1). All were right-handed native English speakers, with normal or corrected-to-normal vision, and received course credit for participating. All participants gave informed consent in accordance with the University of Colorado Institutional Review Board.

We created 70 stimulus sentences, each with a well-formed (1a) and anomalous form (1b) and each seven words long. Well-formed, plausible control sentences consisted of the following sequence: simple noun phrase (e.g., The thief), passive verb sequence (e.g., was caught), by-phrase (e.g., by the police). Anomalous sentences were derived from control sentences by replacing the determiner in the by-phrase (the) with a preposition (e.g., for). In one version of the stimuli, the anomalous preposition was always for (low-variability anomaly). In a second version of the stimuli, seven different prepositions were used to introduce the anomaly (at, of, on, for, from, over, with; high-variability anomaly), with stimulus sentences assigned to one of these prepositions.

For each stimulus set (high- and low-variability), two experimental lists were created such that half (35) of the critical stimuli within each list were assigned to the well-formed and anomalous versions, respectively, in a counter-balanced manner. For the high-variability anomaly stimuli, each of the seven anomalous prepositions occurred five times in the list. Across these stimulus sets, word length was controlled: in the high-variability stimuli, the anomalous words were two, three, or four characters long, with an average length of three, which matched the length of low-variability anomalies (“for”) and control stimuli (“the”) at the critical word position.

Stimuli were pseudorandomly interspersed among 105 filler sentences. Thirty-five of the fillers contained syntactic anomalies, containing violations of noun number agreement (e.g., “The woman bought a cars …”), pronoun agreement (e.g., “Ralph’s uncle taught herself …”), and verb-form (e.g., “The chef will chopping …”). The remaining 70 fillers were syntactically and semantically normal. Overall, 70/175 (40%) of the stimuli were anomalous.

Participants were randomly assigned to one experimental list and tested in a single session lasting about 90 min (including about 45 min of experimental preparation). Participants sat in a dimly lit, sound-attenuated booth. Stimuli were presented on an LCD monitor at a distance of 105 cm from the participant, such that words subtended approximately 2°. The participant was instructed to read normally and to try to understand the sentences. Each trial consisted of the following events: a fixation cross appeared in the center of the screen for 600 ms, after which a stimulus sentence was presented one word at a time in the center of the screen. Each word appeared on screen for 300 ms followed by a blank screen interval of 200 ms. Sentence-ending words appeared with a full stop. One-third of the sentences were pseudorandomly followed by a comprehension question about the sentence. The question appeared in the center of the screen and stayed on the screen until one of two buttons on a button box was pressed, indicating an answer of “Yes” or “No.” Participants used their thumbs to respond, with half the participants using their right thumb to answer “yes.” The remaining two-thirds of the stimulus items were followed by a prompt (“PRESS EITHER BUTTON TO CONTINUE”) appearing on the center of the screen until the YES button was pressed. A 1450 ms blank screen interval followed each trial.

Continuous EEG was recorded from 64 sintered Ag/Ag-Cl electrodes embedded in an elastic cap (Neuroscan QuikCaps) arranged according to the extended 10–20 system. EEG was amplified and digitized at 1000 Hz (Neuroscan Systems). Vertical eye movements and blinks were monitored with two electrodes placed above and below the left eye, and horizontal eye movements were monitored by electrodes placed at the outer canthi of each eye. EEG was also recorded over left and right mastoid sites. Impedances were maintained below 10 kΩ. EEG was referenced on-line to a vertex electrode and re-referenced off-line to an average reference.

After recording, data was down-sampled to 200 Hz and filtered with a bandpass of 0.1–100 Hz. Eye-blink artifact was corrected using a subject-specific regression-based algorithm (Semlitsch et al., 1986). Any remaining voltages exceeding ±100 μV were rejected, resulting in the loss of 5 and 4% of the trials in the control and violation conditions, respectively. Data from individual channels that contained consistent artifacts within a given participant’s data were replaced by an average of the neighboring channels. ERPs were averaged in epochs of activity spanning −100 to 700 ms relative to the onset of the target stimulus, with voltages quantified relative to a 100 ms pre-stimulus window. Early ERP components (P1 and N170) were quantified for statistical analysis as peak voltages within windows of 125–145 (P1) and 170–270 ms (N170) post-stimulus-onset. These two windows were centered on the peaks of these two components in the grand-averaged data across all experimental conditions. Voltages in these time windows were averaged within four channel-groups: left-posterior (P07, PO5, P7, CB1, O1), right posterior (P08, PO6, P8, CB2, O2), left anterior (F7, F5, F3, FT7, FC5), right-anterior (F8, F6, F4, FT8, FC6). The anterior channel-groups covered electrodes at which left anterior negativities have been observed to syntactic anomalies during sentence processing (Hahne and Friederici, 1999; Lau et al., 2006). The posterior channel-groups included electrodes at which early components elicited by visual words (P1, N170) are typically maximal (Rossion et al., 2003; Maurer et al., 2005). The later N400 and P600 components, which are less “peaky” than the P1 and N170, were quantified as mean voltage within windows of 350–450 ms (N400), and 500–800 ms (P600), which were based on visual inspection of the data and were consistent with prior findings involving these components (e.g. Kim and Osterhout, 2005). Voltages within these windows were averaged within a single central-parietal channel-group (CPZ, PZ, POZ, P1, P2), where N400 and P600 effects are typically maximal (e.g., Kim and Sikos, 2011). These dependent measures (P1, N170, N400, P600) were subjected to repeated measures analyses of variance (ANOVAs). For the P1 and N170, these ANOVAs included a between-subjects factor for variability (high vs. low), and within-subjects factors for sentence-type (anomalous vs. control) and channel-group (left-posterior, right-posterior, left-anterior, right-anterior). The Greenhouse–Geisser (1959) correction for inhomogeneity of variance was applied to all repeated measures with greater than one degree of freedom in the numerator.

Both control and anomalous words elicited a positive-going occipital-temporal P1 peak at 135 ms (Figure 1D) followed by a negative-going occipital-temporal N170 peak around 200 ms after stimulus onset (Figure 1D). Inspection of individual participant’s ERPs at a representative left occipital-temporal channel (PO7), indicated that P1 and N170 peak latencies were highly consistent across participants and across levels of stimulus-variability and sentence-type, and that our analysis windows of 125–145 ms (P1) and 170–270 ms (N170) were appropriate for quantifying these components (Figure 2).

The occipital-temporal N170 appeared enhanced by anomalous stimuli, relative to controls (Figures 1B,D and 3). The P1 over left occipital-temporal channels also appeared enhanced by anomalous stimuli, relative to controls, but specifically for low-variability subjects (Figure 3) and not in the grand average data (Figure 1A). At central-parietal sites, anomalous words elicited a large positive shift, relative to controls, beginning around 450 ms and continuing beyond the end of the epoch (Figures 1C,E). Statistical analyses are reported below for the P1, N170, N400, and P600 time windows in the scalp ERP³.

Statistical tests of the back-projected, group-level ICs revealed six ICs with group and/or sentence-type effects (p < 0.01) that also had significant source estimations. Three of the six components had significant effects during the time range of early processing (∼100–300 ms), while the remaining three ICs had significant effects in the time range of the later N400 and P600 effects. There were no significant effects from the planned comparisons (i.e., interaction effects). Here we report on the network of the three components with early effects; arbitrarily named IC1, IC2, and IC3 (Figure 4).

Previous studies report that syntactic category violations in auditory sentence comprehension elicit an ELAN. A prominent model of sentence comprehension attributes the ELAN to fast syntactic analysis mechanisms (Friederici, 2002). Although this model focuses on auditory language processing, the same model has sometimes been associated with anterior negativities elicited by word-category violations during reading, suggesting a domain-general rapid syntactic analysis mechanism (e.g., Friederici et al., 1999). The majority of studies of syntactic category violation during reading, however, do not report ELAN effects, raising a question of how general the ELAN effect is across modalities and experimental preparations (cf. Steinhauer and Drury, 2012).

Our results indicate that the earliest ERP responses to syntactic category violations during written language processing are not ELAN effects but rather are posterior in scalp distribution and generated in visual cortex and other posterior cortical structures. Two prior MEG studies have reported similarly early effects generated in visual cortex (Dikker et al., 2009, 2010). Our findings converge with the MEG results (with some discrepancies, discussed below), and deviate from prior ERP literature, which has not previously reported such early posterior effects elicited by syntactic category anomalies. We share the conclusion of Dikker et al. (2009) that the earliest brain responses to syntactic anomaly contain major contributions from modality-specific sensory cortex, resulting in scalp distributions concentrated over occipital-temporal sites during visual sentence processing. Meanwhile, the scalp distribution of ELAN effects during auditory sentence processing is consistent with generators in auditory areas of the superior temporal lobe, although we provide no direct evidence of this last point in the current findings.

Our data demonstrate that predictive processing plays a key role in the earliest brain responses – within the initial ∼170 ms – to syntactically unexpected inputs. When the word-form at the critical word position was easier to predict (the low-variability stimuli), sensitivity to syntactic anomaly began at an earlier latency. The MEG study of Dikker et al. (2009) also concluded that predictive processing drives the earliest responses to syntactic anomaly, but did not directly examine the impact of affordances for prediction. Instead, the early latency (∼130 ms) and source localization (occipital lobe) of the effects were assumed to be incompatible with high-level syntactic analysis and were therefore attributed to top-down predictive commitments. The effect pattern we observe receives no account within the standard ELAN-based model of sentence processing. In this model, the earliest responses to syntactic category anomaly reflect a mismatch between the syntactic category of the incoming word and a rapidly computed phrase-structural grammatical representation (e.g., Friederici, 2002); the model does not explain why the application of syntactic knowledge or assignment of syntactic category should be affected by affordances for prediction.

Our data involve a combination of long-term syntactic knowledge and rapid adaptations of that knowledge in response to recent linguistic experience. The sensitivity of early brain responses (N170) to syntactic anomaly – for both high- and low-variability stimuli – presumably reflects knowledge of the grammatical regularities of the language, accumulated over a lifetime of linguistic experience. The speeding of syntactic anomaly sensitivity under conditions of reduced stimulus-variability (P1 in addition to N170 effects) is compatible with rapid learning of distributional patterns that are specific to the current experiment. We suggest here that such learning engages a general ability to adjust syntactic processing to the diverse linguistic contexts that language users encounter (e.g., specific conversational topics, and speaker-idiosyncracies, and dialects). The learning mechanisms involved here may be the same as those that mediate syntactic priming effects during language production and comprehension (Pickering and Branigan, 1998; Trueswell and Kim, 1998) and which have been associated with incremental learning of syntactic knowledge throughout a linguistic lifetime (Chang et al., 2006).

The rapid learning of local distributional contingencies may have an analog in auditory ELAN findings. In that work, anomalies were consistently introduced by the same word-initial past participle marker in German (“ge,” as in “gefuttert”). The consistency with which this morpheme marked anomaly within the stimuli raises the possibility that participants learn to predict this specific form as an anomaly introducing stimulus, speeding sensitivity to the anomaly.

Although our stimuli were designed to violate constraints on syntactic category, the mechanisms underlying our results are likely not dedicated to syntactic analysis specifically. A number of studies of semantic processing in context report effects on posterior, visual processing ERP components, with early latencies resembling the effects here, suggesting rapid semantic influences on early stages of visual word recognition (Sereno et al., 2003; Dambacher et al., 2009; Kim and Lai, 2012).

Our findings raise a question of why effects similar to our own have not been more widely reported, given the number of previous studies of syntactic anomaly during reading. We suggest that, relative to previous work, the current and related research combine some simple but potent methodological and experimental advances, which improve sensitivity to effects of high-level variables (e.g., syntactic anomaly) on short-latency ERP effects. First, the current study explicitly focuses on early latency ERPs at occipital-temporal sites, where early visual processing effects are most pronounced (Maurer et al., 2005). Many studies have not examined such effects, in part due to an absence of dense sampling over occipital-temporal channels in older recording apparatus and in part due to a priori hypotheses that higher-level linguistic variables will mainly affect the later N400 or P600 components, as has been observed in a large number of findings involving these two components (e.g., Kutas and Hillyard, 1980; Osterhout and Holcomb, 1992). Although a number of recent studies of visual word recognition have focused on early visual processing ERP components (e.g., Bentin et al., 1999), these studies have often presented single words without context, sometimes using shallow tasks (e.g., passive reading). Embedding words in sentence contexts, as we did here, may engage high-level processing and anticipatory commitments more so than isolated word recognition tasks, enhancing sensitivity to high-level processing, such as syntactic analysis. Finally, uncontrolled variability in low-level features such as word length can strongly modulate the earliest brain responses in a way that is much less visible at later components, and obscuring effects of higher-level variables, such as semantic status (Hauk and Pulvermüller, 2004; see also Penolazzi et al., 2007). Therefore, careful control over word length, as implemented here, may be critical for achieving sensitivity to high-level processing effects.

It is not clear how to reconcile the early posterior effects observed here with the minority of studies reporting that syntactic anomalies during reading elicit early anterior negativities resembling the auditory processing ELAN (e.g., Neville et al., 1991). We note here that interpretation of some of these ELAN-like effects during reading is complicated by the possibility that they contain differences between conditions in baseline activity prior to the critical word (Steinhauer and Drury, 2012). The most widely cited example of an ELAN effect during reading is elicited by comparisons like The scientist criticized Max’s proof of the theorem. vs. The scientist criticized Max’s of proof the theorem (Neville et al., 1991). Here, the critical stimulus word of occurs at different sentence positions (one word later in the control than the anomaly condition) and is also preceded by different word-types (a noun vs. a possessor preceded by a noun). Differences between experimental conditions prior to the critical word may affect the baseline calculation, resulting in differences in the ELAN time window that do not reflect the brain’s response to the critical word. In fact, given that all studies reporting ELAN effects to syntactic anomaly involve differences in pre-critical context material (cf. Steinhauer and Drury, 2012), such results must be interpreted with caution.

Independent components analysis extracted three components that accounted for significant portions of variance within the scalp-recorded ERP and also showed significant effects of either sentence-type (anomaly vs. well-formed) or predictability (high- vs. low-variability) within the initial 200 ms of processing at the target word position. Distributed source localization for these three components estimated generators in the left occipital-temporal cortex (middle occipital gyrus), right posterior occipital lobe (cuneus), bilateral medial parietal lobe (posterior cingulate gyrus), respectively. Three additional sources – in the right superior temporal sulcus, right middle temporal gyrus, and left planum temporale – also responded to stimulus-variability and/or sentence-type but at longer latencies, in the time window of the N400 and P600 ERP components; these sources are discussed in a separate report (Gilley and Kim, in preparation).

The three early active sources responded differently to syntactic anomaly and to stimulus-variability (predictability). The occipital-temporal source, but not the other two sources, showed increased activity and a delayed peak to syntactic anomaly, relative to control words, during the initial 200 ms of processing. This sensitivity to anomaly, furthermore, occurred earlier under conditions of low stimulus-variability. These functional properties are consistent with a role in the response to syntactically unexpected words, which is modulated by predictive processing.

All three early active sources exhibited increased activity for high-variability relative to low-variability stimulus words, during the initial 200 ms of processing. The occipital-temporal source additionally peaked later for high-variability than low-variability stimuli. These effects of stimulus-variability occurred not only at the critical word but also at the preceding and following words, during the same early time window. This pattern of effects is consistent with prediction-related processing, which is upregulated by the difficulty of predicting input in the high-variability stimuli relative to the low-variability stimuli. This prediction-related processing appears to be sustained across multiple words within the sentence.

The contributions of the three cortical regions implicated in our findings can be discussed in the context of their experimental responsiveness and existing understanding of these brain structures. The left occipital-temporal generator showed increased activity as prediction became more difficult (high-variability stimuli) and increased activity for syntactically anomalous stimuli relative to controls, suggesting that it is involved in prediction and the evaluation of those predictions. This generator is part of higher-order visual cortex, and may be well-suited to representing visual word-form features. Anticipatory commitments may pre-activate representations within the left occipital-temporal cortex, resulting in lateral-inhibitory conflict when the bottom-up input does not match the pre-activated pattern. The occipital-temporal representations involved here may be related to those implicated in a number of fMRI and patient studies of visual word processing (Dehaene et al., 2005; Price and Devlin, 2011). ERP studies report effects of word recognition and face processing on the occipital-temporal N170 component, which have sometimes been associated with activity in occipital-temporal cortex, although only some of these findings include source localization results (e.g., Rossion et al., 2003). It should be noted that fMRI BOLD effects have overlapped most across studies in the fusiform gyrus, ventral to the source localizations described here. It is possible that the discrepancy between our source localization and these fusiform activations reflects error in sLORETA source estimation and, perhaps to a lesser extent, inaccuracies in fMRI BOLD localization. Alternatively, our sentence stimuli may impose different functional demands from the word recognition tasks that are most common in fMRI studies, resulting in recruitment of related but distinct regions of the ventral visual stream. Understanding potential differences in the contributions of lateral vs. more ventral regions of occipital-temporal cortex to language processing is an issue for future research. However, consistent with any further illumination of this issue is a core conclusion that predictive commitments about visual word recognition during sentence comprehension may be imposed on and evaluated in higher-order ventral visual cortex, in the left hemisphere.

Posterior occipital cortex showed increased activity for high-variability stimuli, consistent with upregulation by prediction demands, but did not show early sensitivity to syntactic anomaly (Figure 4, IC2). We suggest that perceptual representations in posterior occipital cortex undergo preparatory pre-activation, like those in occipital-temporal cortex – reflected in the effect of stimulus-variability. However, the representations in posterior occipital cortex encode lower-level visual features, which do not systematically discriminate expected from unexpected words in the situations examined here. Previous MEG studies indicate that posterior occipital cortex is sensitive to basic visual distinctions such as foveal position and spatial extent at ∼100 ms (e.g., Tarkiainen et al., 1999). Such low-level features were highly controlled in the current experiment (e.g., control and anomaly stimuli were matched for average length).

Whereas our results indicate that the earliest sensitivity to syntactic category violations occurs in lateral occipital cortex, Dikker et al.’s (2009, 2010) MEG studies reported sensitivity at a similar latency localized to posterior occipital cortex, medial to our posterior occipital source. These different source localizations might reflect distinct response properties of posterior and downstream ventral visual system areas, such as lateral occipital-temporal cortex. In the Dikker et al. (2009) results, anomaly was correlated with word length, and sensitivity to this low-level feature may underlie the effects (Dikker et al., 2009). In the findings of Dikker et al. (2010), category-violating words elicited M100 responses whose amplitudes correlated with the degree to which their phonological properties were atypical for the licensed syntactic category – as quantified by a statistical analysis of the relationship between phonological features and syntactic categories (Farmer et al., 2006). This effect was interpreted as evidence that predictions about grammatical category result in predictions about low-level visual features in posterior occipital cortex, leading to sensitivity to category violations. The types of visual cortical representations that underlie such sensitivity remain poorly understood, and it is not clear what distinguishes such sensitivity from the sensitivity to word-forms that we associated here with lateral occipital cortex. Future work could systematically manipulate the correlation between syntactic anomaly and course-grained word-shape features like length or phonological word-class typicality and investigate the relative involvement of posterior occipital vs. occipital-temporal cortex⁵.

Medial parietal cortex also showed increased activity for high-variability stimuli, indicating that this area is part of the network of prediction-related language processing structures. Medial parietal cortex has been activated in previous fMRI studies of syntactic processing (Kuperberg et al., 2003) and narrative processing (Yarkoni et al., 2008).

However, the contribution of medial parietal cortex to language understanding remains little explored. Critically, this area is not a primary or even secondary sensory area (Parvizi et al., 2006). fMRI studies have recently implicated medial parietal lobe, including posterior cingulate and surrounding retrosplenial cortex, in the processing of contextual associations (Bar and Aminoff, 2003; Bar et al., 2007) which mediate predictive cognitive commitments (Bar, 2007). The medial parietal cortex is interconnected with both medial temporal lobe (MTL) and thalamus (Parvizi et al., 2006), and has been associated functionally with “assigning mnemonic associations to sensory input” (Vogt et al., 1992). We suggest that the medial parietal lobe’s known connectivity and functional properties render it well-situated to modulate sensory processing based on predictions generated from MTL mnemonic representations of syntactic and semantic knowledge. Within a linguistic processing context, medial parietal cortex may serve as a hub for implementing perceptual predictions about linguistic forms based on recent context.

The conclusions here posit rapid recurrent interactions between visual cortex and higher-order representations encoding syntactic knowledge – within ∼200 ms post-stimulus onset. Such conclusions contrast with a number of previous ERP and MEG studies suggesting that prior to ∼200 ms, the visual system is still in the early stages of a feedforward sequence of low-level feature extraction. For instance prior studies reported that brain responses before ∼200 ms distinguish alphabetic character strings from non-alphabetic stimuli but do not distinguish among alphabetic stimuli on the basis of semantic properties or lexical status (Nobre et al., 1994; Bentin et al., 1999; Tarkiainen et al., 1999; Pylkkanen et al., 2002; Mariol et al., 2008; Solomyak and Marantz, 2009). These findings might indicate that the visual system has extracted visual features that respond to word-like stimuli but has not yet accessed higher levels of representation. In the context of such findings, it is worth addressing whether rapid recurrence is physiologically viable and why we observe it here, where other studies do not.

We note here that a fast timeline for feedforward and recurrent information flow within the system is compatible with a substantial body of physiological data, even if the prior language processing ERP literature has not pointed clearly to this conclusion. Human ERP studies find that occipital cortex responds to visual stimuli by 56 ms and that frontal cortex is active by 80 ms (Foxe and Simpson, 2002). Monkey intracranial recordings show that feedforward information flow from V1 to the highest levels of the ventral visual system (inferotemporal cortex, IT) occurs in ∼23 ms (Schroeder et al., 1998, 2001) and that robust selectivity for complex stimuli (e.g., faces) occurs at latencies of ∼100 ms (Rolls and Tovee, 1994). A number of studies indicate that transmission time for information flowing along a single synaptic distance is 10–15 ms, both between and within cortical regions (Tovee, 1994). Thus, the latency of our anomaly related effects includes sufficient time for feedforward information flow from V1 forward to multi-modal integration areas of the brain and even recurrent interaction among lower- and higher-order brain regions. Once information has arrived at higher-order areas, numerous feedback projections, which outnumber feedforward projections in the brain (Felleman and Van Essen, 1991), provide routes for recurrent information flow.