Introduction

A recent finding in the reading comprehension literature is that readers are resilient to small distortions of word order within a sentence. This is evidenced in the transposed-word effect initially reported by Mirault et al. (2018). Using a speeded grammatical decision task, Mirault et al. (2018) found that participants took longer and made more errors in deciding that a transposed-word sequence (e.g., The can boy run fast) was ungrammatical compared with a control sequence where transposing any two words does not generate a correct sentence (e.g., The can boy run desk). These results indicate that participants tended to pursue a grammatical reading of the transposed-word sequences (e.g., "The can boy run fast" is understood as "The boy can run fast"), thus making ungrammatical decisions more difficult.

The presence of transposed-word effects converges nicely with recent theorizing that suggests that readers do not always construct a veridical representation of sentence structure (for a review, see Christianson, 2016). Mirault et al.’s (2018) findings demonstrated that approximate or noisy information about word order might be one characteristic of such "good-enough" sentence representations (e.g., Ferreira & Lowder, 2016). As proposed in the theoretical work of Snell et al. (2017, 2018), when multiple words are processed at the same time, word positions are flexibly encoded. That is, a word identity is associated not only with its actual position but also with neighbouring positions, which mimics the noisy letter position coding during word recognition hypothesized in certain models of orthographic processing (e.g., Gómez et al., 2008). Crucially, this noisy word position coding does not prevent the rapid construction of an elementary, approximate syntactic representation, which subsequently provides top-down feedback to word identities and guides the allocation of word identities to probable positions. The combination of positional flexibility and syntactic constraints on word position coding accounts for readers' tendency to interpret ungrammatical transposed-word sequences as being grammatical.

Motivated by Mirault et al.’s (2018) findings, a growing body of research has further investigated transposed-word effects in order to elucidate the mechanisms involved in assigning word identities to their positions in a sequence (Huang & Staub, 2021a; Liu et al., 2020, 2021; Mirault et al., 2020; Pegado et al., 2021; Pegado & Grainger, 2019, 2020, 2021; Snell & Grainger, 2019; Wen et al., 2021a, 2021b). For example, the studies by Snell and Grainger (2019) and Wen et al. (2021b) revealed two key constraints on transposed-word effects: (1) the distance separating the two transposed words (the effect was only significant with adjacent words in the Snell and Grainger study); and (2) the role of syntactic phrase boundaries (the effects were greater when transpositions occurred within a syntactic phrase relative to transpositions across a syntactic phrase in the Wen et al. study). Huang and Staub (2021a) replicated the grammatical decision results of Mirault et al. (2018) in English and found a similar pattern in a more natural reading-for-meaning experiment. Mirault et al. (2020) showed that transposed-word effects are not caused by reading the transposed words out of order, and Liu et al. (2020, 2021) reported transposed-word effects when reading a logographic script (Chinese). Finally, Pegado and Grainger (2020) found transposed-word effects in a same-different matching task (are two sequences of words composed of the same words in the same order or not?) that does not require the computation of syntactic structure. Transposed-word effects were found to be affected by the grammatical nature of the sequences to be compared (e.g., more errors were made when matching "he wants these green apples" and "HE THESE WANTS GREEN APPLES" compared with "green wants these he apples" and "GREEN THESE WANTS HE APPLES"), but only when the matching process was hard enough.

Taken together, these findings fit with a model of sentence reading according to which a certain amount of parallel word processingFootnote 1 enables the rapid association of several word identities to their positions in the sentence, followed by the rapid computation of an initial primitive sentence-level representation that then provides feedback to on-going word identification processes. Within a cascaded-interactive processing framework (McClelland & Rumelhart, 1981), transposed-word effects can be due to noise in the bottom-up association of word identities to their positions in a sequence, and also due to the fast computation of a sentence-level representation forcing a grammatical interpretation of the transposed-word sequence via top-down constraints. Alternatively, transposed-word effects can be accounted for in models of sentence reading that apply a strictly one-word-at-time serial reading by assuming that on some trials participants actually read the transposed-word sequence as a correct sentence (see Huang & Staub, 2021a, for evidence for this from eye-movement patterns). Such serial processing accounts point to re-ordering word identities during post-lexical integration process as the locus of transposed-word effects (Huang & Staub, 2021a, 2021b). In line with the serial reading interpretation, Liu et al. (2022) have shown that transposed-word effects in Chinese can be observed under conditions of rapid serial visual presentation (RSVP) of the word sequence, but only in error rates and not in response times for a grammatical decision task. Based on an extensive replication of this finding in French, Mirault et al. (2022) concluded that the fact that under conditions of serial processing the effects were only observed in error rates in their study and the Liu et al. (2022) study is most likely due to the role played by top-down constraints in forcing a re-ordering of words into a grammatically correct sequence.

The present study provides a test of transposed-word effects in conditions that are expected to encourage parallel word processing by using the word-in-sequence identification paradigm with rapid parallel visual presentation (RPVP) of word sequences. In the RPVP paradigm, a sequence of words is presented simultaneously for a short duration (250 ms or less) in order to minimize eye movements (Asano & Yokosawa, 2011). Prior work using this paradigm has found post-cued word identification to be more accurate when the target word is embedded in a syntactically valid context compared to a syntactically invalid context formed by changing the order of words in the correct sentence condition (Declerck et al., 2020; Snell & Grainger, 2017; Wen et al., 2019). This ‘sentence superiority effect’ is interpreted as reflecting feedback from a sentence-level representation to on-going word identification via cascaded-interactive processing.

The present study had two main aims. First, we examined whether post-cued word identification in the RPVP paradigm would be higher in transposed-word sequences (e.g., The can boy run fast) compared with an ungrammatical control sequence (e.g., The can get run fast), with the same target word tested at the same position (e.g., run) in the two conditions. Target words could appear at either position 2 or 4 in the five-word sequence and were never part of the transposed-word manipulation. Both the cascaded-interactive and the serial processing accounts of transposed-word effects, discussed above, appeal to sentence-level constraints on word-order encoding as being one major mechanism driving such effects, and therefore potentially both predict that a transposed-word manipulation will impact on target word identification.

The second goal of the present study was to test whether the support provided by a transposed-word sequence context is equivalent to the support provided by a correct sentence context. Finding a significantly smaller sentence superiority effect with transposed-word sequences compared with true sentences would provide support for our hypothesis that transposed-word effects reflect the partial activation of sentence-level structures that then provide feedback to on-going word identification processes. However, a serial "re-ordering" account of transposed-word effects should predict no difference between transposed-word sequences and grammatical sentences. Concerning this specific contrast, it is important to note that comparing transposed-word sequences and grammatically correct sentences would involve different responses in a grammatical decision task. That is another reason why the word-in-sequence identification task was used in the present study (i.e., participants are performing exactly the same task in these two conditions).

In sum, we predicted, on the basis of our cascaded-interactive account of the sentence superiority effect, that word identification would be more accurate in transposed-word sequences (e.g., the target run in the sequence: The can boy run fast) compared with ungrammatical control sequences (e.g., The can get run fast), and most accurate in grammatical sequences (e.g., The boy can run fast).

Methods

Participants

One hundred and twenty-four native English speakers (65 females; mean age = 30.75 years, SD = 9.99) were recruited online via Prolific (Palan & Schitter, 2018). Data from 16 additional participants were excluded from the analyses because of their low overall accuracy rates (< 30%, N = 12), their first language (≠ English, N = 3), or zero accuracy for targets in position 4 (N =1).

Materials and design

First, we constructed 144 grammatically correct English sentences that consisted of five words. The average word length was 4.38 letters (SD = 1.31) and the average word frequency was 5.60 (SD = 1.14) in Zipf values (van Heuven et al., 2014). For each sentence, two types of ungrammatical versions were created. First, the transposed-word condition was generated by swapping words at positions 2 and 3 (e.g., He the throws glass there) or at positions 3 and 4 (e.g., Please put jacket the here). Second, following Pegado and Grainger (2020), the control condition was generated by replacing one transposed word with a word of the same word length (e.g., He the jacket glass there/Please put throws the here). Sentences were paired so that an identical set of words was used in the two ungrammatical conditions (e.g., "He throw the glass there" paired "Please put the jacket here" to generate previous examples), thus minimising lexical-level impacts across conditions. Words at positions 2 or 4 that stayed in the same position in the grammatical and ungrammatical versions were used as the target word (e.g., the word "glass" is the target word for He throws the glass there/He the throws glass there/He the jacket glass there). Thus, the same word targets were tested in the three levels of the factor Context (grammatical, transposed-word, ungrammatical control). The design was therefore a 3 (Context) × 2 (Position) factorial. The targets consisted of 144 different words with an average word length of 4.74 letters (SD = 1.00) and an average word frequency of 5.22 (SD = 0.90) in Zipf values (van Heuven et al., 2014). Three counterbalanced lists were created to ensure that only one condition of the 144 sequences was presented in each list and all conditions (grammatical/transposed-word/control) were presented across lists. Participants were randomly assigned to one of counterbalanced lists. The complete list of stimuli is provided in the Appendix.

Procedure

All participants provided their informed consent before the online experiment started. The presentation of the stimuli was controlled by LabVanced (Finger et al., 2017). A unique random trial order was generated for each participant. Each trial began with two vertical fixation bars presented for 500 ms at the screen centre. Next, a sequence of five words was presented for 250 ms. We increased stimulus duration compared with the 200-ms presentation duration used in our previous studies (Declerck et al., 2020; Snell & Grainger, 2017; Wen et al., 2019) given that here we tested five-word sequences as opposed to the four-word sequences tested in our prior work. Then, a sequence of hash marks was presented at all prior letter locations, together with an underline at the target location as the post-cue (see Fig. 1). Participants were instructed to focus their eyes on the space between the fixation bars and to report the target at the post-cued location. They could take as long as needed to type in their response. The inter-trial interval was set at 500 ms. Prior to the experiment, six practice trials were used to familiarise the participants with the procedure.

Fig. 1
figure 1

Illustration of the sequence of events in the post-cued partial report Rapid Parallel Visual Presentation (RPVP) procedure

Full size image

Data analysis

A response was coded as correct only if it was an exact match of the target. Using the lme4 (Bates et al., 2015) and lmerTest (Kuznetsova et al., 2017) packages in R (R Core Team, 2021), the accuracy data were analysed with a logistic mixed-effects model (Jaeger, 2008) using sum contrasts. Participants and items were included as random effects (Baayen et al., 2008), and by-participant and by-item random slopes were also included (Barr et al., 2013). The main effects (Context/Position) were obtained from the Type II Wald χ2 test using the car package (Fox & Weisberg, 2019). Pairwise comparisons were conducted using Tukey's adjustment to control the familywise error rate.

Results

Condition means are shown in Fig. 2. Average identification accuracy in the grammatical, transposed-word, and control conditions was 62.2%, 58.8%, and 55.3% respectively. The analysis using mixed-effects modelling revealed a main effect of Context, χ2 (2) = 47.9058, p < .001. Planned pairwise comparisons showed that identification accuracy rates for words presented in the grammatical condition were higher than in the control condition (β = 0.424, SE = 0.0619, z = 6.845, p <.0001), a standard sentence superiority effect. Crucially, accuracy was also significantly higher in the grammatical condition compared with the transposed-word condition (β = 0.216, SE = 0.0537, z = 4.019, p = .0002), and significantly higher in the transposed-word condition compared with the control condition (β = 0.208, SE = 0.0591, z = 3.516, p = .0013). Although Position did influence identification accuracy (χ2 (1) = 24.6982, p < .001), with higher accuracy for words in position 2 (66.9%) than in position 4 (50.3%), it did not interact with Context (χ2 (2) < 2, p > .50).

Fig. 2
figure 2

Mean identification accuracy rates with 95% confidence intervals (Cousineau, 2005) at the two target positions in the control, transposed-word (TW) and grammatical context conditions

Full size image

Discussion

The present study aimed to investigate whether a transposed-word effect could be obtained in a word-in-sequence identification experiment using the RPVP technique combined with post-cued identification. That is, contrary to all prior observations of transposed-word effects, in the present study participants only had to identify one word. Finding a transposed-word effect in the present study would therefore provide evidence that such effects are at least partly driven by the transposed-word sequences activating the corresponding correct sentence structure, which then constrains on-going word identification processes. To investigate this, we compared identification accuracy of a target word in five-word sequences that could be (1) a grammatically correct sentence (e.g., The boy can run fast), (2) a transposed-word sequence (e.g., The can boy run fast), or (3) an ungrammatical control sequence (e.g., The can get run fast), with the same target word at the same position (here the word run) tested in the three conditions. Target words could appear at either position 2 or 4 and were never part of the transposed-word manipulation.

Our first main finding is that post-cued word identification accuracy was higher in transposed-word sequences than in control sequences, which demonstrates for the first time a transposed-word effect in word-in-sequence identification. This finding fits with the general hypothesis that transposed-word sequences provide bottom-up support for the corresponding grammatical base sentence from which they are derived, and the sentence-level representation of the base sentence then constrains processing of the target word in the transposed-word sequence. Both parallel and serial accounts of word sequence processing appeal to sentence-level constraints as a key mechanism in driving transposed-word effects.

The second main finding of the present study was the higher word identification accuracy in correct sentences compared with transposed-word sequences, another novel finding that could not be attested with the grammatical decision task given that different responses ("yes" vs. "no") are associated with the grammatical decisions made to these two types of sequence. We predicted this pattern on the basis of our cascaded-interactive account of the sentence superiority effect (Declerck et al., 2020; Snell & Grainger, 2017; Wen et al., 2019). That is, given the positional mismatch between words in the correct sentence representation and the transposed-word sequence, we predicted that bottom-up support for sentence-level representations would be reduced in transposed-word sequences compared with true sentences, resulting in less feedback and less accurate word identification. We would further argue that the difference observed between transposed-word sequences and correct sentences is evidence against a "re-ordering" account, according to which a transposed-word sequence is mistakenly processed as a correct sequence (Huang & Staub, 2021a, 2021b).Footnote 2

Concerning the main effect of target position, our results are in line with prior research that has consistently reported highest identification accuracy at position 2 within 4-word sequences (e.g., 58.2% for position 1, 79.2% for position 2, 61.3% for position 3 and 62.2% for position 4 in Wen et al., 2019), hence strongly suggesting that participants were not performing a left-to-right serial processing of words in the sequence, otherwise highest accuracy should have been obtained at position 1. We nevertheless acknowledge that the higher performance at position 2 compared with position 4 in the present study, and compared with position 3 in prior studies, merits further examination in future research.

One further means to test the cascaded-interactive account proposed here would be to manipulate the location of the target word relative to the transposed words (i.e., the target is one of the transposed words or not). In the present study, the target word was never one of the transposed words. We predict that when the target is one of the transposed-words (e.g., the target "glass" in the sequence "he throws glass the there"), then positional noise in the feedback process should diminish the difference between transposed-word sequences and the ungrammatical controls (i.e., a smaller transposed-word effect). In addition, future research could also examine whether transposed-word effects can be observed with syntactically valid but semantically odd sentences (e.g., Angry water flies quietly) in the same-different matching task (see Massol et al., 2021, for a demonstration of a sentence superiority effect with semantically anomalous sentences). The existence of transposed-word effects in this case would speak against the post-lexical integration account proposed by Huang and Staub (2021a, 2021b) since semantic anomalies should diminish the role played by sentence-level constraints in re-ordering transposed words.

Finally, the present study also provides a further demonstration of the utility of the RPVP paradigm as a tool to investigate reading comprehension. The brief simultaneous presentation of horizontally aligned words is intended to capture the kind of processing that might occur across multiple words during a single fixation in natural reading. We nevertheless acknowledge that in the absence of eye-movement recordings, we cannot be absolutely sure that our participants were indeed fixating the central fixation point as per instructions, and not moving their eyes during stimulus presentation. It will be therefore important for future studies to combine eye-movement recordings with the RPVP paradigm in order to monitor participants' eye fixation location and eye movements. Furthermore, we admit that the present study did not control for viewing angle since the display size (the size of stimuli and screens) varied across participants and participants' viewing distance was also unknown. Although the lack of control over the display size and viewing distance is a longstanding limitation for web-based experiments (Angele et al., 2022; Grootswagers, 2020), means to address such limitations have only recently been investigated (Brascamp, 2021; Li et al., 2020). It is therefore recommended that future online studies using the RPVP paradigm adopt recently developed methods to better estimate viewing angle, given that changes in viewing angle might impact on the serial versus parallel processing of word sequences. We note, nevertheless, that everyday reading is characterized by variations in text size and viewing distance, and so such variations in the present study might actually be a better reflection of reading out of the lab. Regardless of such variations, the present study replicated a reliable lab-based result: the sentence superiority effect (Declerck et al., 2020; Snell & Grainger, 2017; Wen et al., 2019). Thus, we reason that variations in viewing angle likely did not impact our main findings (see Angele et al., 2022, for a similar reasoning for masked-priming effects in online vs. in-lab experiments).

To conclude, the present study found a transposed-word effect expressed as a higher word identification accuracy in transposed-word sequences compared with control sequences. We also observed that word identification accuracy was greater in correct sentences compared with transposed-word sequences. We suggest that the word identification advantage in transposed-word sequences is driven by facilitatory feedback to on-going word identification processes from partially activated sentence-level representations within the framework of a cascaded-interactive theory of word identification and sentence comprehension.