The reading brain extracts syntactic information from multiple words within 50 milliseconds

To what extent do readers process multiple words in parallel? Although it is now commonly accepted that letters are processed across multiple words simultaneously, higher-order (lexical, semantic, syntactic) parallel processing remains contentious. Recent use of the flanker paradigm has revealed that the syntactic recognition of foveal target words is influenced by the syntactic congruency of parafoveal flanking words even when target and flankers are shown for only 170 ms. It has been argued, however, that such settings may allow processing of multiple words even if this were to happen on a serial one-by-one basis. To circumvent this possibility, here I have tested participants in a syntactic categorization task whereby targets and flankers were shown for only 50 ms and replaced by post-masks. Significant effects of target-flanker congruency were observed in both response times and accuracy, indicating that readers extracted syntactic information from multiple words within the very brief presentation time. The present results strongly suggest that the brain extracts higher-order linguistic information from multiple words in parallel.


Introduction
The question of whether the reading brain processes multiple words in parallel has been hotly debated in recent years (e.g., Schotter and Payne, 2019;Snell et al., 2018c;Snell andGrainger, 2019a, 2019b;White et al., 2019a).The investigation of this issue has largely proceeded along two separate lines.One of these has focused on sentence reading, and specifically, the extent to which upcoming words impact behavioral or brain measures prior to their fixation.The rationale here is that early responses to upcoming words make parallel word processing more likely to be true (e.g., Angele et al., 2013;Brothers et al., 2017;Brothers and Traxler, 2016;Kennedy and Pynte, 2005;Mirault et al., 2020;Snell et al., 2017a;Snell et al., 2017b;Snell et al., 2023a).The other line has ceded some of the ecological validity offered by sentence reading, and has instead relied on artificial tasks such as lexical or semantic decision-making about individual target words, thereby attempting to probe the reading brain's circuitry more directly (e.g., Dare and Shillcock, 2013;Snell et al., 2018b;Snell et al., 2019;Snell and Grainger, 2018;Wen et al., 2019;White et al., 2018White et al., , 2020;;White et al., 2019b).
The present study extends the latter line of inquiry.As argued by Snell and Grainger (2019a), undoubtedly during reading we aim to identify words in a serial, one-by-one fashion; but even if readers were mostly to succeed in this aim, this in itself would not evidence a strictly serial processing bottleneck per se.If the brain is truly wired to process one word at a time, then this should be obvious even in artificial settings where the brain is maximally probed on its ability to process multiple words in parallel (or, put differently, its inability to perfectly focus on single words).
The flanker paradigm lends itself perfectly for this endeavor.With this paradigm, researchers test the recognition of central target words as a function of adjacent flanker words.If the target and flankers are presented a sufficiently short amount of time, then the extent to which flankers impact responses informs the extent to which the target and flankers were processed in parallel.To appreciate this paradigm's value, let us first consider text reading.During text reading, the fact that the eyes move from left to right (in Western alphabets) inevitably imposes some seriality on word processing: an upcoming word is likely to enter the processing stream later than a fixated word simply because it enters the perceptual span later.Oculomotor behavior (which constitutes the traditional measure of interest in sentence reading research; e.g., Rayner, 1998) may not be sufficiently sensitive to the incomplete and ongoing processing of upcoming words, even if that processing is of a higher-order (lexical, semantic, syntactic) nature.In the flanker paradigm, meanwhile, the task is to categorize (either lexically, semantically or syntactically) central target words, the measures of interest being E-mail address: J.J.Snell@VU.nl.response time (RT) and accuracy.Although adjacent flanking words may well be processed to a lesser extent than the target word (e.g. because more attentional resources are dedicated to the latter), there may nonetheless be an impact on response measures to evidence flanker processing.Crucially, the flanker paradigm enables the investigation of such effects under tightly controlled circumstances: i.e., target and flankers may be presented at durations that barely allow for the recognition of a single word; (but for clever ways to control viewing times in sentence reading, see, e.g., Hohenstein et al., 2010;Pan et al., 2020).If the flankers nonetheless impact higher-order categorizations, this would provide straightforward evidence against a serial processing bottleneck.
In the remainder of this Introduction, I will briefly review what has been learned from the flanker paradigm and comparable tasks.Subsequently I will outline the present study.

Parallel word processing in artificial tasks: the status quo
The flanker paradigm's rise to prominence in the serial-versusparallel debate started with the study of Dare and Shillcock (2013).When presenting targets for 150 ms, lexical decisions about these targets were made better when they were flanked by related letters (e.g., 'ro rock ck') than unrelated letters ('st rock ep'); a finding that has been replicated across various studies (e.g., Grainger, Mathôt, & Vitu, 2014;Snell et al., 2017a;Snell and Grainger, 2018;Snell et al., 2018d;Snell et al., 2019).Researchers now generally agree that sub-lexical orthographic information (i.e., information about letter identities and positions) is retrieved from multiple words in parallel-all the more because the socalled parafoveal-on-foveal influences have been established in natural sentence reading too (e.g., Angele et al., 2013;Dare and Shillcock, 2013;Inhoff et al., 2000;Mirault et al., 2020;Snell et al., 2017a).
It has been argued that while letter processing may occur for multiple words in parallel, higher-order processing would still occur on a strictly serial basis (e.g., Schotter and Payne, 2019).Yet, when letting participants categorize central target words as noun or verb, Snell et al. (2017b) observed that performance was better when targets were surrounded by syntactically congruent flankers than incongruent flankers, in spite of a joint 170 ms presentation time (see also Vandendaele et al., 2020).Similar flanker effects have been found in a semantic categorization task (Meade et al., 2021;Snell et al., 2018b).Thus, readers may not only extract letter information, but also syntactic and semantic information, from multiple words within a brief amount of time.
How strongly do the above findings evidence parallel word processing?That depends on whether the 170 ms presentation time as used by Snell et al. (2017bSnell et al. ( , 2018b) is short enough to prohibit serial reading of the target and flankers.Our rationale at the time was that readers typically take 150-250 ms to recognize single words (see e.g.Rayner, 1998, for a review), and that 170 ms would therefore be too brief for the flankers to impact higher-order processing under the serial assumption (for if 150 ms were spent on the target, that would leave but 20 ms to process flankers).
More recently, however, White et al. (2018White et al. ( , 2019bWhite et al. ( , 2020) ) have provided a good counter-argument.By applying a staircase procedure, they established that viewing times much briefer than 170 ms may be sufficient for the brain to achieve word recognition.Depending on the experiment and the participant's performance, viewing times ranged from 32 to 134 ms.In these studies White et al. also reported an absence of evidence for parallel word processing.Their setup was one where participants viewed two words (aligned horizontally in their 2018 and 2019b studies but vertically in their 2020 study) very briefly.Prior to stimulus onset, participants were either cued for one of the word locations, or cued for both word locations (these were referred to as singleversus dual task trials, respectively).After stimulus offset, participants were again cued for one of the words, which they had to semantically categorize.Unidirectional pre-cues always matched post-cues (i.e., they were 100% predictive).At the minimal viewing time required for decent categorization of pre-cued targets, participants performed decidedly worse in the 'dual task' trials.White et al. reasoned that accuracy was so bad in the dual-task condition that it matched the prediction of a serial model: participants could recognize only one word per trial and had to make a random guess when asked about the other.The absence of evidence for parallel processing in behavioral measures was accompanied by compelling fMRI data.These showed a significant impact on neural activity from the sub-lexical properties of both words, while lexicosemantic properties could only be traced for pre-cued targets.
The results of White et al. are important because they make clear that parallel word processing should be tested in a more tightly controlled setting than that employed by Snell et al. (2017bSnell et al. ( , 2018b) ) and Meade et al. (2021), with considerably briefer presentation times.At the same time the studies of White et al. may be argued to comprise several shortcomings that undermine their conclusions.Let us consider these potential shortcomings below.
Firstly, in order for two words to be processed simultaneously at all, a prerequisite is that both words are situated within the processing span that typifies reading.There is no doubt that visuo-spatial attention typically encompasses multiple words (e.g., Angele et al., 2013;Dare and Shillcock, 2013;Inhoff et al., 2000), but this specifically pertains to words that are horizontally aligned.It is a well-known fact that, through the left-to-right direction of reading, the brain is trained to extend attention in the horizontal but not the vertical dimension (e.g., Rayner, 1998).Indeed, when employing a flanker paradigm with horizontally and vertically aligned flankers, Snell et al. (2018d) observed an impact from the former but not the latter.White et al.'s, 2020 study made use of vertically aligned words, which possibly made simultaneous orthographic processing of two words considerably more difficult, even the brain were in principle able to process multiple words in parallel; (several studies suffer from the same problem; see, e.g., Johnson, Palmer, Moore, & Boynton, 2023).
Meanwhile, White et al.'s, 2018 and 2019b studies did employ horizontally aligned words, but those setups may have inadvertently induced processing bottlenecks in different ways.Specifically, the target words in these studies were quite large (occupying on average 3.25 degrees of visual angle) and also quite far apart (due to being separated by a central fixation cross, there was a 5.50 • distance between target centers).Thus, in order to process both words effectively, fine-grained orthographic processing would have to occur across an area of ~8.75 • .But at such eccentricities the uptake of visual information is hampered by somewhat reduced acuity and, more importantly, increased crowding (e.g., Bouma and Legein, 1977;Yeatman and White, 2021).The span of effective vision for reading is known to extend as far as the parafovea, at approximately 6 • (e.g., Engbert et al., 2002); word recognition at greater eccentricities may be possible if letters are sufficiently large, but of course print size does not compensate indefinitely. 1Additionally, the dual task trials necessitated participants to divide attention left and right of a central fixation cross; but this central focal point may in itself have prohibited an efficient division of attention.Notably, participants were able to perform a color detection task for two stimuli at the same spatial locations; but that does not guarantee that orthographic processing could have occurred simultaneously at those two locations as well.
In light of the above, one post-hoc analysis of White et al.'s, 2020 study is quite striking.Prior to the start of each trial, the semantic category of each target word was chosen independently of the other.As such, there were trials where the target words' semantic categories were congruent with one another and trials where the categories were incongruent with one another.Analogous to the flanker studies of Snell 1 Note, researchers have often conceptualized the perceptual span for reading in terms of number of letter spaces (McConkie and Rayner, 1975) because the span, when measured in degrees of visual angle, scales with print size.Nonetheless, nobody can comfortably recognize a word in peripheral vision beyond 6 • .et al. (2017b, 2018b) and Meade et al. (2021), a better performance was observed in congruent target trials than incongruent target trials.The parallel processing account of such congruency effects is that the semantic categories of both words influence the response.On the other hand, White et al. themselves argued that this particular effect can also be explained by a serial processing model, by assuming that participants accidentally focused on the non-cued target on a portion of the trials.
With respect to these congruency effects, a problem is that White et al.only assessed accuracy and not RTs (because, e.g., responses had to be withheld while brain activity was being measured).The serial approach can indeed account for congruency effects in accuracy: namely, in incongruent trials, incorrect decisions are based on the incorrect category of the non-cued target.But there would be no reason to expect an effect in RT: the participant's response is based either on one target or the other, without any cross-influences; (note, there is one potential solution for the serial approach, but I will show in the Discussion how this is refuted by the data).Given that aforementioned flanker studies have reported congruency effects in both accuracy and RT, the question remains whether White et al. might have been able to observe congruency effects in RT too, if their experimental setups had allowed for it.If so, this would be straightforward evidence against serial processing.
In sum, prior research geared at probing the brain's ability to process multiple words in parallel has produced equivocal evidence.Several recent studies have suffered from shortcomings, with the studies of Snell et al. (2017bSnell et al. ( , 2018b) ) and Meade et al. (2021) using presentation times that were too long, while the studies of White et al. (2018White et al. ( , 2019bWhite et al. ( , 2020) ) suffer from suboptimal spatial layouts and the lack of an opportunity to test the impact of target word congruency on RT.

The present study
Here I report an experiment that combines the best aspects of the methodologies of Snell et al. (2017bSnell et al. ( , 2018b) ) and Meade et al. (2021) on the one hand and White et al. (2018White et al. ( , 2019bWhite et al. ( , 2020) ) on the other.Specifically, in the spirit of the former, I tested higher-order (syntactic) recognition of central target words as a function of the target's congruency with horizontally aligned adjacent flanking words, both in accuracy and response times.Meanwhile, in the spirit of the studies of White et al., the target and flankers were replaced by post-masks after a mere 50 ms presentation time, in order to avoid any possibility that the words could be processed sequentially.Note, I chose 50 ms here because it is in the lower range of the viewing times used by White et al.; but it will be seen in due course that baseline performance was quite good in my study (>80%), meaning viewing times may be decreased even further in future investigations.
The hypotheses for this experiment are as follows: if the brain processes syntactic information for multiple words in parallel, then the syntactic congruency of task-irrelevant flanker words may impact responses.Alternatively, if the brain processes higher-order information in a strictly serial manner, then RT should be unaffected by flanker congruency.

Methods
Thirty students gave informed consent to their participation in this study for monetary compensation or course credit.All participants reported to be non-dyslexic native Dutch speakers with normal or corrected-to-normal vision.
I devised a list of 73 nouns and 73 verbs with a length of 4 or 5 letters (M = 4.33).All 146 words were used once as target word, once as congruent flanker word, and once as incongruent flanker word.Targets were paired with a congruent flanker and incongruent flanker with the constraint that the target and flankers should have no more than one letter in common with each flanker (targets had an average overlap of 0.60 and 0.41 letters with incongruent and congruent flankers, respectively).On each trial a target was flanked by a congruent or incongruent flanker on each side.Flankers were separated from the target by a single character space.As such, the entire stimulus was on average 14.97 and 14.93 character spaces long in the case of congruent and incongruent flankers, respectively.Stimuli were presented in monospaced font whereby each character space subtended 0.30 • , meaning that the entire stimulus subtended an average of 4.49 • .Thus, when assuming a central fixation, all letters were situated within the scope of parafoveal vision.
The experiment followed a 2 ⨉ 3 design with target category (noun, verb) and flanker condition (congruent, incongruent, no-flanker baseline) as factors.However, as the central test concerned effects of flanker congruency irrespective of target type, noun and verb targets were combined in all analyses.All participants saw all targets in all three flanker conditions.The total of 438 trials was presented in random order.
Fig. 1 shows the trial procedure.Vertical fixation bars right above and below the screen center were presented throughout the experiment.Each trial started with a 500 ms fixation display.Subsequently the target and flankers were shown for 50 ms, and were replaced by post-masks consisting of a hashmark ('#') at each letter location. 2Participants were instructed to categorize the target word as noun or verb as quickly and accurately as possible.They were also instructed to focus solely on the central target word and to ignore surrounding flanker words.Nouns and verbs were indicated with a right-('/') or left-handed ('Z') keyboard button press respectively.Feedback was provided with a green central fixation dot in case of a correct response, or a red central fixation dot in case of an incorrect response or having reached a time-out of 3000 ms.The 438 experimental trials were preceded by 12 practice trials, and a break was offered after every 90 trials.The entire experiment lasted approximately 20 min.

Results
Of the thirty participants, three performed poorly with accuracies of 56.4%, 55.7% and 13.7% respectively.After excluding these participants, the remaining 27 participants yielded a total of 3942 measurements per condition (noun and verb targets combined), which is well beyond the recommended 1200 measurements for sufficient statistical power (Brysbaert and Stevens, 2018).Prior to the analysis of RTs, incorrectly answered trials (~23%) were excluded.Both RTs and accuracy were analyzed after excluding trials beyond 2.5 SD from the mean (~3%).Noun and verb targets were jointly analyzed with linear mixed-effects models (LMMs) with Flanker condition (no-flanker baseline, congruent, incongruent) as fixed factor and items and participants as random effects.For RTs I report, b-values, standard errors (SEs) and tvalues.Accuracy was analyzed with generalized LMMs, whereby I report z-values instead of t-values.In all analyses, values | t | and | z | > 1.96 were taken to indicate significance.
One might argue that the congruency effect was driven by participants who were so proficient that they could process two words serially within 50 ms.If this were the case, then congruency effects should be modulated by subjective task difficulty.To investigate this, in a post-hoc analysis I divided participants into two groups on the basis of their overall accuracy.Flanker congruency influenced RTs in the better scoring group (N = 14; b = 19.66,SE = 8.34, t = 2.36) but not in the worse scoring group (N = 13; b = 4.07, SE = 9.95, t = 0.41).On the other hand, the impact of flanker congruency on accuracy appeared greater in the weaker group (b = 0.65 (Δ ≈ 10%), SE = 0.07, z = 8.89) than in the stronger group (b = 0.62 (Δ ≈ 9%), SE = 0.09, z = 6.86).Yet, none of these group differences were supported by significant interactions between overall accuracy and flanker condition (for RTs, b = 136.69,SE = 81.13,t = 1.69; for accuracy, b = 0.62, SE = 0.73, z = 0.86).Similar patterns were seen when ranking participants by overall speed, or by both speed and accuracy (inverse efficiency scores: RT/p(correct)).

Discussion
Here I have investigated the hotly debated issue of whether the reading brain is wired to process one or multiple words simultaneously.In previous research, higher-order (e.g.syntactic, semantic) categorizations for target words were made better if those target words were surrounded by congruent words than incongruent words (e.g., Meade et al., 2021;Snell et al., 2017bSnell et al., , 2018b)).These findings were taken as evidence for parallel word processing with the assumption that the 170 ms presentation time as employed in those studies would be sufficiently short to preclude sequential processing of the target and flankers.However, White et al. (2019aWhite et al. ( , 2019bWhite et al. ( , 2020) ) have convincingly argued that tests of parallel word processing warrant a more tightly controlled setting.
Following the recommendations of White et al., in the present experiment stimuli were shown for only 50 ms 3 and replaced by post-masks to prohibit sequential word processing in sensory memory.In line with previous flanker experiments, I tested readers' ability to syntactically categorize central target words as a function of the target's congruency with adjacent flanking words.
The results are quite clear-cut.In spite of the 50 ms presentation time, response times (RTs) and accuracy were impacted by the syntactic congruency of the flankers.I also observed that flanker presence per se was detrimental, as performance was best in the baseline condition without flankers.This particular pattern appears to suggests that, in spite of the fact that the flankers are task-irrelevant and to be ignored, the brain cannot prevent itself from having some attentional resources directed away from the target.The results are also in line with a recent flanker study by Vandendaele & Grainger (2023) employing 50 ms stimulus durations and evidencing higher-order flanker processing through lexical inhibition effects.
Can the present findings be explained under the assumption that the brain processes words in a strictly serial fashion?In an earlier study, White et al. (2020) accounted for an effect of adjacent word congruency by assuming that participants (erroneously) focused on the non-target word on a portion of the trials, thereby producing an incorrect response if the non-target was incongruent with the target.In the present study, however, I observed an effect of flanker congruency not only in accuracy but also in RTs.Here the account of White et al. does not hold: even if participants indeed responded to the flanker rather than the target, then seriality necessitates that participants were not processing the target.Consequentially there is no reason as to why flanker-target incongruency would delay the response.
However, let us entertain one potential solution for the serial perspective. 4Assume again that participants focused (solely) on a flanker on a portion of trials.The analysis of RTs only includes correctly answered trials.This effectively means, in the case of incongruent trials, that we are including those trials where the flanker was incorrectly categorized (e.g., the target is a verb, the flanker is a noun, the participant focuses on the flanker and categorizes it as a verb, hence the trial is marked as "correct").Of the congruent trials, meanwhile, we are only including trials where the flanker was correctly categorized.Assume also that correct responses are typically slower than incorrect responses  3 I should acknowledge that applying a staircase procedure to bring presentation times in line with individual reading speeds-as did White et al.-would be an even safer way to prohibit sequential processing of target and flankers.On the other hand, I observed no significant modulation of congruency effects by overall performance here. 4For this clever serial account of RT effects I must credit Alex White.
(this is certainly true in the present study: in the baseline condition, correct and incorrect responses were made after 742 and 837 ms, respectively).As such, might the congruency effect on RTs be explained by a serial model after all?
I reckon the answer is still no.If the above were true, then incongruent trials marked as "correct" should have slower responses than incongruent trials marked as "incorrect", given that these would be respectively incorrect versus correct flanker judgments.But an inspection of the data refutes this: the former were much quicker than the latter, at 831 ms and 895 ms, respectively.These data warrant a due focus on the target in a good majority of trials, but that would leave little room for the emergence of flanker effects under serial processing.Future modelling efforts may bolster these mathematical intuitions (see e.g., Snell, n.d., in press, for flanker task simulations).
As I see it, the present results are most naturally explained by a parallel processing framework.It is evident that syntactic information was extracted from multiple words within 50 ms.If the brain were a serial word processor, then attention must have somehow shifted entirely from the target to a flanker (or vice versa) within those 50 ms.But it has long been understood that visuo-spatial attention is not such a high-speed switching mechanism (e.g., Duncan et al., 1994), and in fact, covert attentional shifts away from the fovea are estimated to take ~140 ms (Carlson et al., 2006).
On a final note, I have argued previously that conscious awareness is not key in the serial versus parallel processing debate (see Snell and Grainger, 2019b).What I mean by this is that, contrary to what some researchers have implied (e.g., Schotter and Payne, 2019;White et al., 2019a), it does not matter whether recognition-if recognition were a punctual event at all-occurs for multiple words simultaneously.Limits imposed by visual acuity and attention virtually guarantee that simultaneously viewed words are processed at different strengths, implying asynchronous recognition.It is therefore interesting to realize that, if I had directly probed flanker word recognition, the data would likely have provided 'evidence' for the serial approach; for without conscious recognition, flanker identification would be at chance-level.Some previous 'evidence' for serial processing is similarly based on null-effects (e. g., Angele et al., 2013;Brothers et al., 2017); but we should not forget to adhere to the principle that absence of evidence is not evidence of absence.This is precisely why the reliance on target-flanker congruency effects is so important: it allows us to evidence flanker processing irrespective of whether the flanker identities reached conscious awareness.What solely matters for the serial-versus-parallel debate is the fact that higher-order information was extracted rapidly and in parallel from multiple words here.Serial word processing is an oculomotor strategy that may well be successful most of the time; but it is not intrinsic to the neural circuitry underlying reading.

Fig. 1 .
Fig. 1.Schematic overview of a trial.The size of stimuli relative to the display is enlarged to aid visibility in this example.

Fig. 2 .
Fig. 2. Average response times and accuracies in all three flanker conditions.Error bars indicate standard errors.