Many languages permit flexible word orders. For instance, Japanese transitive sentences may have either Subject-Object-Verb (SOV) or Object-Subject-Verb (OSV) word orders to express the event of Hanako eating pizza, as illustrated by (1). The SOV word order is canonical while the OSV word order is non-canonical, derived by scrambling the object over the subject (Hoji, 1985; Saito, 1985). Scrambling is a type of optional movement of a constituent (Ross, 1967). It is semantically vacuous, in the sense that the basic sentence and its scrambled counterpart are truth-conditionally equivalent.

An important question in the field of psycholinguistics is how humans parse and comprehend sentences with alternative word orders, such as SOV and OSV sentences in Japanese. Numerous sentence processing studies have tackled this question, suggesting that non-canonical sentences incur a higher processing load than their canonical counterparts across different languages (e.g., Bader & Meng, 1999, 2023; Erdocia et al., 2009; Frazier & d’Arcais, 1989; Jeong et al., 2025; Kim et al., 2009; Koizumi et al., 2014; Mazuka, 2002; Miyamoto & Takahashi, 2002b; Ono et al., 2020; Rösler et al., 1998; Tamaoka et al., 2005, 2024; Weyerts et al., 2002). The observed processing asymmetry follows from the assumption that, all other things being equal, syntactically complex sentences are harder to process than their simple counterparts (e.g., Gibson, 1998; Hawkins, 2004; Koizumi, 2023; Marantz, 2005; Pritchett & Whitman, 1995). More concretely, an optional movement like scrambling introduces an additional filler–gap dependency between a moved phrase (i.e., filler) and its original position (i.e., gap); hence, non-canonical sentences are syntactically more complex and, accordingly, computationally more costly to represent than their canonical counterparts.

Other than syntactic complexity, several factors have been identified and investigated that impact the processing of flexible word order sentences (Kaiser & Trueswell, 2004; Koizumi, 2023; Miyamoto & Takahashi, 2002b; Tamaoka et al., 2005). One such factor to be investigated in the current study is information structure: how new information and old information are ordered in a sentence.1 Many studies have examined the interaction between syntax and information structure in the processing of flexible word order sentences (e.g., Bader & Meng, 2023; Jeong et al., 2025; Kahraman & Hirose, 2018; Kaiser & Trueswell, 2004; Koizumi & Imamura, 2017; Miyamoto et al., 2014; Slioussar, 2011; Yano & Koizumi, 2018, 2021; Yano et al., 2019). Although their details are different, these studies adopt the view that a non-canonical sentence is a marked option, and its use is associated with a discourse requirement that old information comes before new information in a sentence (Aissen, 1992; Birner & Ward, 2009; Kuno, 1978, 1995). For instance, Kaiser and Trueswell (2004) hypothesized that non-canonical sentences would become easier to process when they had a given–new order. Consistent with this hypothesis, their self-paced reading experiment revealed that an appropriate context alleviated the processing asymmetry between non-canonical OVS sentences and canonical SVO sentences in Finnish. Specifically, they found a significant interaction at the second NP (jänistä in (3a) vs. hiiri in (3b)) such that the reading time was longer in OVS sentences than in SVO sentences only when the first phrase was not introduced in the preceding context (2).

The findings from the other studies cited above are largely consistent with those from Kaiser and Trueswell (2004). These previous results underscore the point that the putative processing cost associated with non-canonical sentences stems not only from syntax but also from information structure.

It is important to note that previous studies have seldom found evidence that contextual support completely eliminates or overrides the processing asymmetry between canonical and non-canonical sentences (cf. Funasaki & Yano, 2025). This indicates that information structure is not the sole factor determining the processing difference under consideration.

Although most of the previous research examined the processing of object-initial non-canonical sentences (e.g., OSV sentences in Japanese), at least three studies focused on subject-initial non-canonical sentences, namely, Yasunaga et al. (2015), Yano et al. (2019), and Asami and Tomioka (2025). We review them in order below.

Yasunaga et al. (2015) conducted an event-related potential (ERP) experiment in Kaqchikel (a Mayan language spoken in Guatemala). In Kaqchikel, VOS sentences are canonical while SVO sentences are non-canonical, derived by fronting the subject to sentence-initial position (García Matzar & Rodríguez Guaján, 1997). Yasunaga and colleagues found a P600 effect at the third phrase in SVO sentences (i.e., O) relative to the corresponding phrase in VOS sentences (i.e., S) during the comprehension of Kaqchikel sentences. Importantly, this effect was detected even though the experimental sentences were preceded by a non-verbal context – a picture depicting the participants involved in the sentences. This result was comparable to that of another ERP experiment in which the same sentences were presented out of the blue (Yano et al., 2017). Yasunaga and colleagues attributed the observed P600 effect to the processing of the filler–gap dependency between the sentence-initial subject and its original position after the object (i.e., S₁VOt₁). Their claim is consistent with the traditional view that this ERP component reflects the processing cost related to filler–gap dependency formation (e.g., Kaan et al., 2000; Phillips et al., 2005; Ueno & Kluender, 2003).

However, the experimental studies in Yano et al. (2019) found different patterns in Seediq (an Austronesian language spoken in Taiwan). In Seediq, VOS sentences are canonical, while SVO sentences are non-canonical, derived by preposing the subject to clause-initial position (Tsukida, 2009). Their first experiment showed that in the absence of the context, the third phrase in SVO sentences (i.e., O) elicited a larger P600 in comparison to the corresponding phrase in VOS sentences (i.e., S). Crucially, their second experiment demonstrated that a non-verbal context – a picture similar to that used in Yasunaga et al. (2015) – substantially alleviated the processing cost observed in their first experiment. Based on these results, Yano and colleagues claimed that the processing cost indexed by the P600 has to do with a discourse-related factor rather than a syntactic one, contra Yasunaga et al. (2015) (also Kaan et al., 2000; Phillips et al., 2005; Ueno & Kluender, 2003). They then concluded that the processing cost associated with non-canonical SVO sentences in Seediq is related to a discourse-related factor.

The findings from Yasunaga et al. (2015) and Yano et al. (2019) do not fit well with each other. One possible interpretation of these conflicting results is that a single picture might serve as a sufficient context for non-canonical SOV sentences in Seediq, but not in Kaqchikel, for some independent reason (we will come back to this point in 5.1). Even if this consideration is on the right track, we should be cautious about the interpretation of the results in Yano et al. (2019). It would be too hasty to conclude from their ERP results that contextual support completely eliminated the processing cost for the non-canonical SOV sentences in Seediq, because as suggested by Yano and colleagues themselves (p. 414), the P600 might only reflect a conflict between syntactic structure and information structure (see also Yano, 2019; Yano & Koizumi, 2018). If this interpretation of the P600 is correct, the findings from Yano et al. (2019) are not informative about the processing cost associated with a syntactic factor per se. Therefore, further research is necessary to see whether syntax still impacts the processing of discourse-licensed, subject-initial non-canonical sentences.

The findings from Asami and Tomioka (2025) represent a case in point. They examined the processing of inchoative sentences in Japanese that contain a nominative-marked theme subject and a postpositional causer adjunct. They found that it took less time to process causer–theme sentences like (4a) than theme–causer sentences like (4b). They interpreted this result as indicating that the former are canonical while the latter are non-canonical, derived by moving the theme subject over the causer adjunct.

Importantly, they also revealed that the processing cost for non-canonical theme–causer sentences was reduced when the sentence-initial theme subject was marked as discourse-given with the demonstrative sono ‘that’.

It was also reported that the causer–theme order was processed faster than the theme–causer order even when the fronted subject was interpreted as discourse-given. These findings indicate that both syntax and information structure play a role in the processing of subject-initial non-canonical sentences. However, there is a methodological limitation in this study. It employed a speeded forced-choice task (e.g., Asami, 2024; Chujo, 1983; Koizumi & Tamaoka, 2010; Tamaoka et al., 2005). In this task, a sentence is presented in full, and only the total time the participants take to read the whole sentence is measured. This methodology is useful for examining whether a particular word order takes longer to process than the other(s) (Tamaoka & Koizumi, 2006), but it is not very informative about the time course of sentence processing (Miyamoto & Nakamura, 2005). Therefore, the findings from Asami and Tomioka (2025) cannot tell us at which point in the sentence syntax and information structure interact during comprehension.

To be fair to the three studies reviewed above, it is important to emphasize that no single experimental methodology can offer a perfect picture of human sentence processing. It is essential to examine evidence from several experimental methods with different measures (i.e., a multiple-methods approach; Kaiser & Trueswell, 2004, p. 130). In light of this, the current study used a moving-window, self-paced reading task – a widely-adopted methodology to investigate cognitive processes behind online reading (Just et al., 1982) – to shed light on the time course of the interaction between syntax and information structure in the processing of subject-initial non-canonical sentences. We examined the phrase-by-phrase processing of inchoatives with a theme subject and a causer adjunct in Japanese, expanding on Asami and Tomioka (2025).

We found that the second phrase (P6) was read more quickly in causer–theme sentences than in theme–causer sentences. This result is consistent with the findings of Asami and Tomioka (2025), who reported shorter processing times for causer–theme sentences when the sentences were presented in full. Taken together with their results, the present reading-time pattern supports the view that the causer–theme order is canonical, whereas the theme–causer order is non-canonical, in inchoative constructions with a theme subject and a causer adjunct. Moreover, the observed significant difference indicates that our sentence-pair method enabled us to detect the processing asymmetry between canonical and non-canonical word orders with simple experimental stimuli.

The self-paced reading technique allowed us to examine the phrase-by-phrase processing of two word order variants of inchoative constructions. Based on the current results, we consider how non-canonical theme–causer sentences are processed under the filler–gap model (e.g., Clifton Jr & Frazier, 1989; Crain & Fodor, 1985; Frazier & Clifton Jr, 1989; Stowe, 1986); see also, for example, Miyamoto (2008) for its application to Japanese sentence processing). The previous research on non-canonical OSV sentences in Japanese (Aoshima et al., 2004; Miyamoto & Takahashi, 2002a, 2004; Nakano et al., 2002; Shibata et al., 2006) has suggested that a parser posits a gap associated with a fronted object (the filler) upon encountering the pre-verbal subject, because this is the point at which the parser detects a non-canonical structure. We assume that a similar process occurs in the theme–causer sentences examined here. Upon encountering the causer phrase, the parser infers that the sentence-initial theme has been displaced, thereby triggering the insertion of a corresponding gap. It is this gap-filling operation that can account for the observed slowdown at the second phrase in the theme–causer condition (causer) relative to the canonical causer–theme condition (theme).

The current findings shed light on the processing costs associated with different Case markers. Miyamoto and Takahashi (2002b, pp. 176–177) hypothesized that noun phrases bearing nominative Case markers take longer to process than those with other Case markers. This hypothesis is based on two considerations: first, the nominative Case often signals a clause boundary (Miyamoto, 2002); and second, it is marked, in the sense that it may be associated with a specific interpretation – for instance, the exhaustive reading of nominative-marked phrases (Kuno, 1973). Thus, the inherent properties of the nominative Case marker were thought to contribute to increased processing difficulty. However, the current findings are not compatible with this hypothesis. We observed that the nominative-marked theme was read more quickly than the postpositional causer adjunct. This result suggests that the nominative Case does not inherently lead to longer reading times (see also Tamaoka et al., 2005 for relevant discussion).

It is worth noting that we did not control for word frequency in the critical sentences, despite the well-established effect of lexical frequency on processing speed (e.g., Ashby et al., 2005). However, lexical frequency alone cannot readily account for the observed slowdown at the second phrase (P6). Crucially, no significant reading-time difference was found at the first phrase (P5). If lexical frequency were a major contributing factor, we would expect to see effects in both P5 and P6, since the same lexical items (causer and theme) were compared in opposite orders across these two positions (i.e., causer vs. theme in P5 and theme vs. causer in P6). Thus, it is reasonable to conclude that lexical frequency did not meaningfully confound the observed reading-time pattern.

Experiment 2 investigated how syntactic structure and information structure interact during the online processing of inchoative constructions in Japanese. We reiterate our four predictions: (i & ii) effects of discourse status (givenness) would be observed at the first and second phrases; (iii) an interaction between word order and information structure would emerge at the second phrase; (iv) theme–causer sentences would elicit longer reading times than causer–theme sentences at one or more phrases. In light of these predictions, we discuss the results below.

The first phrase was read more slowly when it was marked as discourse-given than when it was not. This finding aligns with our first prediction that the discourse effect would emerge at this phrase. We speculate that the observed slowdown reflects the pragmatic inference required when a noun phrase marked with the demonstrative sono ‘that’ is encountered. Specifically, such a phrase signals that its referent is discourse-given. Because no explicit discourse context was provided in the experiment, this inference may have imposed an additional cognitive load, resulting in longer reading times.

A main effect of information structure was also observed at the second phrase, but in the opposite direction: reading times were longer in the given-second condition than in the given-first condition. This pattern confirms our second prediction that the discourse effect would also emerge at the second phrase, where the given-marked phrase appears in the given-second condition. We interpret the observed slowdown similarly to that detected at the first phrase: the demonstrative in the absence of explicit contextual support likely triggered a pragmatic inference about discourse givenness, which incurred additional cognitive cost.

Aside from this effect, no significant reading-time difference was observed at the second phrase. This result does not support our third prediction that this would be the point at which an interaction between word order and information structure would emerge. One possible explanation for this unexpected finding is that the processing time at this phrase was confounded or masked by the additional pragmatic inference triggered by the demonstrative sono, as discussed above.

Crucially, an interaction between word order and information structure was observed at the verb: reading times were significantly longer for the non-canonical theme–causer order than for the canonical causer–theme order when the second phrase was marked as discourse-given. When the first phrase was discourse-given, however, we found no evidence for such a reading-time difference. This finding is at least partially consistent with our third prediction. It also aligns with previous studies of object-initial non-canonical sentences using the self-paced reading paradigm. For example, Kaiser and Trueswell (2004) found that the reading-time difference between canonical SVO and non-canonical OVS sentences in Finnish was not significant at the second noun phrase when the sentence-initial phrase represented old information. Similarly, Koizumi and Imamura (2017) reported that the processing costs associated with non-canonical OSV sentences in Japanese were greatly reduced when the fronted object was mentioned in prior context.

It is important to note that the timing of the interaction observed in the current study differs from that reported by Koizumi and Imamura (2017). In our experiment, the interaction emerged at the verb, whereas in Koizumi and Imamura’s study, it was detected at the second phrase immediately before the verb. We conjecture that in the present study, the interaction may have initially occurred in the second (preverbal) phrase but was obscured by the additional cognitive cost associated with the pragmatic inference triggered by the demonstrative; still, the interaction effect may have persisted and become visible at the verb as a spill-over effect. Unlike our experiment, Koizumi and Imamura included a context sentence to explicitly establish whether a particular phrase represented given information without additional elements like the demonstrative sono ‘that’. We reason that, in contrast to our design, this contextual support minimized the cognitive cost associated with the additional pragmatic inference discussed above. As a result, Koizumi and Imamura were able to detect the interaction at the preverbal phrase.

Finally, theme–causer sentences were read more slowly than causer–theme sentences at the sentence-final phrase (i.e., rasii ‘I hear’).10 This slowdown is consistent with our fourth prediction that the non-canonical theme–causer order would take longer to process than the canonical causer–theme order at one or more points in the sentence. As in Experiment 1, we attribute this slowdown to the formation of a filler–gap dependency between the fronted theme and its corresponding gap. What distinguishes the current finding from that of Experiment 1 is the position at which the slowdown occurred: in Experiment 1, it appeared at the second phrase, whereas in Experiment 2, it emerged at the final phrase. We explain this difference as follows. In Experiment 2, the slowdown due to the filler–gap dependency formation likely originated at the second phrase but was masked by the additional discourse-related processing triggered by the demonstrative. Nonetheless, the associated processing cost persisted, extending into the following phrases. This spill-over effect was detected in Experiment 2 but not in Experiment 1, presumably because the extra discourse-related process depleted cognitive resources, thereby making the cost related to the non-canonical sentences more evident in a subsequent phrase. Although the precise mental mechanisms underlying this effect remain to be determined, our interpretation is supported by prior findings that reading-time slowdowns for non-canonical sentences tend to become more evident if complex sentences are used as experimental stimuli, as introduced in 2 (see also Miyamoto, 2008, pp. 232–234 for relevant discussion).

It is worth noting that Kaiser and Trueswell (2004) also found a processing difference between canonical SVO and non-canonical OVS sentences in Finnish, even when a felicitous discourse context was provided. In particular, they observed a slowdown at the verb in non-canonical OVS sentences relative to canonical VOS sentences, despite the presence of a supportive context. They attributed this effect to predictive processing: upon encountering the verb, the parser anticipates an upcoming phrase (which is assumed to represent new information), and it is this predictive process that incurs a processing cost at the verb in contextually supported OVS sentences. Evidence consistent with this hypothesis was reported in their second experiment, which used eye-tracking methodology. However, this line of reasoning does not readily account for the slowdown observed at the final phrase in the current experiment. Due to the verb-final structure of Japanese, the parser would have already processed the new-information phrase prior to reaching the final element (rasii ‘I hear’). Thus, the observed slowdown cannot easily be attributed to predictive processing. Instead, we suggest that the spill-over effect of filler–gap dependency formation better explains the result. Further research is required to determine to what extent syntactic factors are at work along with the predictive process in the processing of discourse-licensed non-canonical OVS sentences in Finnish.

Overall, the results of Experiment 2 suggest that discourse factors play a crucial role in the processing of subject-initial non-canonical sentences, as is the case with the processing of object-initial ones. The current results give further credence to the view that non-canonical word orders are associated with discourse requirements (Aissen, 1992; Birner & Ward, 2009; Kuno, 1978, 1995). However, it is important to emphasize that this does not mean that syntactic factors become irrelevant once the discourse requirement is satisfied. This is evidenced by the observed slowdown at the final phrase in the Theme_given–Causer condition relative to the Causer_given–Theme condition. The observed persistence of the processing advantage for canonical sentences is consistent with what previous sentence processing studies have found (e.g., Bader & Meng, 2023; Imamura et al., 2016; Jeong et al., 2025; Kaiser & Trueswell, 2004; Koizumi & Imamura, 2017; Miyamoto et al., 2014). We thus conclude that both syntax and information structure are relevant to the processing of non-canonical word orders.

Before moving on to the next section, we note that the current experiment employed pairs of simple sentences in each trial; hence, we expected high accuracy on the secondary forced-choice task, and consequently, no significant accuracy differences among conditions because of a ceiling effect. However, the causer–theme order yielded significantly higher accuracy than the theme–causer order in the given-first condition. Although it is unexpected, the observed significant difference is consistent with our conclusion we drew from the reading-time data: the canonical word order enjoys a processing advantage over its non-canonical counterpart, even when information structure is taken into account. In contrast, in the given-second condition, no significant difference was observed between the causer–theme and theme–causer orders. We must leave it open why we saw a significant difference between the two word orders only in the given-first condition.

The results of Experiments 1 and 2 showed that both syntactic structure and information structure have an impact on the processing of subject-initial non-canonical sentences. Crucially, processing costs associated with syntactic structure were detected even when discourse-related factors were controlled. This finding aligns with prior research on the processing of object-initial non-canonical constructions (e.g., Bader & Meng, 2023; Kahraman & Hirose, 2018; Kaiser & Trueswell, 2004; Koizumi & Imamura, 2017; Miyamoto et al., 2014; Slioussar, 2011). In light of these results, we now turn to reconciling the contrasting findings of two ERP studies, Yasunaga et al. (2015) and Yano et al. (2019), which investigated the processing of canonical VOS and non-canonical SVO sentences in Kaqchikel and Seediq, respectively.

As introduced in 1.2, using the P600 component as an index of processing cost, Yasunaga et al. (2015) reported a clear processing advantage for canonical VOS sentences over non-canonical SVO sentences in Kaqchikel even when a non-verbal context was provided. In contrast, Yano et al. (2019) found that a similar contextual manipulation significantly reduced the processing cost for non-canonical SVO sentences relative to canonical OVS sentences in Seediq. There are two different possible ways to reconcile the seemingly contradicting ERP studies so that their results are still compatible with our findings.

Yano et al. (2019) proposed that the P600 reflects processing costs arising from a conflict between syntactic structure and information structure (see also Yano, 2019; Yano & Koizumi, 2018). If their interpretation of the P600 is to be adopted, the processing cost associated with non-canonical SVO sentences may still be present in Seediq even when a supportive non-verbal context is present, as the P600 may not be sensitive to this type of cost. Therefore, the findings of Yano et al. (2019) are not incompatible with our conclusion that syntactic processing costs exist even when discourse requirements are satisfied.

The question remains, however, why Kaqchikel and Seediq showed a different P600 pattern. There are at least three potential causes: (i) the non-verbal context used to facilitate pragmatic support, (ii) the syntactic structure of the SVO word order, and (iii) information structure that supports or encourages subject movement. As for the non-verbal contexts in the two studies, the difference is minimal. Both studies employed line drawings depicting events involving two individuals, the subject referent and the object referent, in different colors. It is important to note that the kind of drawing used in the experiments makes both the subject and the object referents discourse-given, which will be a crucial part of the explanation later.

Let us now turn to the syntactic structures of VOS and SVO sentences in Kaqchikel and Seediq. The following are two sample stimuli tested in Yasunaga et al. (2015) and Yano et al. (2019). In these examples, the subjects are marked in boldface. One notable difference between the two languages concerns the morphological marking of the subject in SVO sentences. In Kaqchikel, the fronted subject in SVO order appears morphologically identical to its in-situ counterpart in VOS order. In contrast, in Seediq, the moved subject is marked with the particle o on its right edge, while the in-situ subject bears the nominative Case marker ka on its left.

This morphological difference raises the possibility that subject movement in SVO sentences may differ qualitatively between the two languages. Particularly relevant is the analysis of Seediq by Tang (2011), who treats the particle o as the topic marker. The availability of an overt topic-marking strategy can affect the kind of information structure that is suitable for the subject initial word order. Our speculation is as follows. In Kaqchikel, the SVO order is supported by a context where the subject referent is discourse-given and the object referent is discourse-new. In Seediq, on the other hand, the context that supports the SVO order makes reference only to fronted subjects: being topical, the referent of a fronted subject with o should be discourse-given whereas the object referent can be either discourse-new or -given. This asymmetry between the two languages can account for the reported contrast with respect to the P600. Recall that the drawings used in the experiments portrayed both the subject and the object referents as discourse-given. Such a context is only partially congruent with SVO in Kaqchikel. The subject satisfies the condition, but the object does not, and consequently, a P600 effect is expected, as detected in Yasunaga et al. (2015). There is no such incongruity in Seediq, which should lead to significant attenuation of the P600, as observed in Yano et al. (2019).

We also see another possibility that is coupled with the more traditional view of the P600 that this ERP component indexes cognitive load associated with filler–gap dependency formation (e.g., Kaan et al., 2000; Phillips et al., 2005; Ueno & Kluender, 2003). Under this view, the contrast between Kaqchikel and Seediq may come down to the presence vs. absence of a filler–gap dependency in SVO sentences. If a fronted subject with o in Seediq is a topic phrase, it is conceivable that a topical subject is base-generated at the sentence-initial position without undergoing any movement, as has been argued for wa-marked thematic topic phrases in Japanese (Saito, 1985). This means that the subject-initial stimuli in Yano et al. (2019) may not necessarily involve a filler–gap dependency, which would account for the significant reduction of the P600 in their studies. It can be maintained, on the other hand, that the subject-initial order in Kaqchikel is a result of movement, and the presence of the P600 effect observed in Yasunaga et al. (2015) is straightforwardly interpreted as the effect of a filler–gap dependency that cannot be completely eliminated by their contextual manipulation. The findings in our studies are compatible with this possibility.

To sum up, we reviewed the connection between our studies and the two previous ERP studies of subject-initial non-canonical sentences. Since our knowledge of the grammar of Kaqchikel and Seediq is very limited, our interpretation of their findings remains speculative. However, there is more than one way to accommodate the seemingly contradictory results of the previous studies in such a way that they are compatible with our conclusion that the processing cost associated with a non-canonical structure exists even with matching information structure.

We now turn to the nature of processing costs associated with non-canonical sentences. In Experiment 2, the given–new information structure did not fully eliminate the processing difference between canonical causer–theme and non-canonical theme–causer sentences. The distinction between canonical and non-canonical orders remained evident in the reading times at the sentence-final phrase (rasii ‘I hear’) (and in accuracy in the secondary forced-choice task). These findings suggest that satisfying discourse-related constraints does not necessarily eliminate processing asymmetries tied to syntactic configuration.

One potential concern about our experiment is that it did not include a context sentence to explicitly manipulate information structure, which could explain why the processing asymmetry between canonical and non-canonical sentences did not disappear. However, we emphasize that in most reading-time studies, a felicitous context rarely eliminates the processing advantage for canonical word order altogether (e.g., Bader & Meng, 2023; Imamura et al., 2016; Kaiser & Trueswell, 2004; Koizumi & Imamura, 2017; Miyamoto et al., 2014). Moreover, these studies consistently showed that non-canonical sentences seldom become significantly easier to process than their canonical counterparts, even when discourse conditions were carefully controlled. Taken together, these findings suggest that the persistence of the processing asymmetry in our study is not solely attributable to the absence of a context sentence.

In relation to this point, and setting aside the reading-time literature, it is worth highlighting the recent work by Funasaki and Yano (2025) using pupil size or pupillometry as an index of cognitive load. Using auditory stimuli, Funasaki and Yano found that non-canonical OSV sentences in Japanese were, at least temporarily, easier to process than canonical SOV sentences when used as a response to subject-wh questions, which required the subject to be a focus. Based on these findings, they argued that contextually marking the displaced phrase as discourse-given is not sufficient to eliminate the processing asymmetry between canonical SOV and non-canonical OSV sentences. They further suggested that the felicitous use of the OSV order requires not only that the fronted object be discourse-given or topicalized, but also that the pre-verbal subject be interpreted as a focus. We find this interpretation compelling and expand on it as follows: the persistent difficulty in eliminating or overriding the processing asymmetry between canonical and non-canonical word orders suggests that the processing cost associated with non-canonical structures is substantial. This cost may only be canceled out or overridden when multiple discourse conditions are properly satisfied, such as givenness and focus alignment. In other words, the primary source of the processing difficulty may be syntactic in nature, as argued by Koizumi (2023) and others.

Needless to say, the present study has not explored in detail other potential sources of processing cost, including but not limited to anticipatory processes, frequency of certain word orders, and (implicit) prosody (e.g., Kaiser & Trueswell, 2004; Koizumi, 2023; Miyamoto & Nakamura, 2005; Miyamoto & Takahashi, 2002b; Tamaoka et al., 2005). These factors are not mutually exclusive and may jointly contribute to processing difficulty during comprehension. Further research is needed to clarify the respective roles of these sources in shaping real-time processing. In addition, while the current investigation has focused on Japanese, it is important to consider how the present claims generalize to other languages with flexible word order, such as German, Finnish, and Russian (see Kaiser, 2014, 527–533 for a short overview of early studies on these languages and relevant cross-linguistic discussion).

Finally, we highlight our methodological contribution to the examination of phrase-by-phrase processing of mono-clausal, simple sentences. As discussed in 2, earlier self-paced reading research often failed to detect processing asymmetries between canonical and non-canonical sentences (Mazuka, 2002; Miyamoto & Takahashi, 2002a; Nakayama, 1995; Tamaoka et al., 2003; Yamashita, 1997). Consequently, several researchers have questioned the suitability of the self-paced reading technique for investigating the online processing of non-canonical structures, particularly when sentences are syntactically simple (Mazuka, 2002; Miyamoto & Nakamura, 2005; Tamaoka & Koizumi, 2006; Witzel & Witzel, 2016). One plausible explanation for the earlier null results is that participants developed a fixed rhythm for pressing the button, which led to minimal variance in recorded reading times (Tamaoka & Koizumi, 2006; Tamaoka et al., 2003). Another possibility is that the experimental sentences themselves were too simple to reveal subtle processing-time differences between canonical and non-canonical structures (Miyamoto, 2008). The present study addressed these concerns by introducing the sentence-pair method, in which each trial consisted of an unrelated filler sentence followed by a target sentence. This method allowed us to detect reliable reading-time differences even with mono-clausal, simple sentences. We interpret this as evidence that the inclusion of an unrelated filler sentence helps to disrupt rhythmic button-pressing and add sufficient complexity to the task, enabling the self-paced reading task to reveal processing costs that might otherwise go undetected.

We acknowledge that the sentence-pair method is not without limitations.11 One potential concern is that presenting two unrelated sentences in succession is somewhat unnatural, as readers do not typically encounter such a sequence in everyday language use. We were aware of this issue during the experimental design phase and addressed it by having participants complete twelve practice trials, which familiarized them with the atypical format prior to the main experiment (see 3.1.3). Notably, none of the participants reported difficulty with the sentence-pair method itself. A related concern is that the presentation of two unrelated sentences might impede parsing and comprehension. However, this issue does not appear to apply to our experiments because participants’ performance on the secondary forced-choice task was consistently high, as evidenced by higher than 90% accuracy.

In light of these potential concerns, we do not claim that the sentence-pair method is the best approach for all sentence-processing studies. Like any experimental technique, it has its strengths and limitations and should be employed selectively, depending on the research goals and the nature of the target constructions. As discussed later, there are cases in which this method is particularly useful. On the other hand, when the target constructions are sufficiently complex, such as relative clauses, there is no need to adopt our method. The same holds for designs in which a context sentence is followed by a target sentence, as in Miyamoto et al. (2014) and Koizumi and Imamura (2017). In such cases, concerns about failing to detect reading-time differences generally do not arise.

One might suggest that alternative experimental methods (e.g., eye-tracking, maze, and ERP techniques) should be used instead of the self-paced reading paradigm (see Kaiser, 2013, for a review of key experimental paradigms). The use of such methods is indeed valuable and often provides richer data. However, we maintain that the self-paced reading paradigm remains a useful and productive tool, particularly due to its relatively low financial and technical demands; thus, the availability of more advanced paradigms does not diminish the value of our sentence-pair method within the context of self-paced reading studies. Rather, we see our method as a practical enhancement to this widely accessible approach.

Before concluding, let us illustrate how the sentence-pair method can offer new insights into the online processing of structurally simple sentences. As an example, we consider ditransitive constructions in Japanese. These constructions involve a dative-marked indirect object (DAT) and an accusative-marked direct object (ACC), and allow at least two grammatical word orders: DAT–ACC and ACC–DAT. We provide relevant examples below:

Using a speeded forced-choice task (referred to as a plausibility judgment task in their terminology), Koizumi and Tamaoka (2004) found that DAT–ACC sentences were processed more quickly than ACC–DAT sentences. Based on this result, they concluded that the correct theoretical analysis of Japanese ditransitives aligns with the proposal of Hoji (1985), in which the dative-marked object originates hierarchically above and linearly precedes the accusative-marked object (cf. Matsuoka, 2003; Miyagawa, 1997).

However, Miyamoto and Nakamura (2005) responded critically to Koizumi and Tamaoka (2004), arguing that their theoretical conclusion was premature. They noted that the experimental task in Koizumi and Tamaoka’s study only measured the total time to read the entire sentence, offering coarse temporal resolution and leaving unclear when the processing cost was incurred. Miyamoto and Nakamura proposed that the results could be compatible with syntactic theories other than that of Hoji (1985), depending on how sentence processing unfolds incrementally (see also Miyagawa, 2012, pp. 113–116 for a response to Koizumi & Tamaoka, 2004). Their critique was rich in proposals for future research aimed at probing the nature of the processing differences observed by Koizumi and Tamaoka. However, those proposals have yet to be fully implemented, presumably due to the widespread belief that the self-paced reading paradigm lacks the sensitivity to detect cognitive costs in the processing of structurally simple sentences (see especially Koizumi & Tamaoka, 2006 for a direct response to Miyamoto & Nakamura, 2005). Crucially, our sentence-pair method refutes this belief by demonstrating that the self-paced reading technique, when suitably modified, can detect such subtle processing costs. It is therefore timely to adopt and adapt Miyamoto and Nakamura’s ideas to revisit the conclusion drawn by Koizumi and Tamaoka. Needless to say, theoretical analyses of Japanese ditransitives have continued to evolve since the debate between Koizumi and Tamaoka, on the one hand, and Miyamoto and Nakamura, on the other (see, for instance, Asami & Bruening, 2025). Future research should take these theoretical developments into account to formulate a detailed mapping between theories of competence and language processing.

The potential of our sentence-pair method extends beyond the specific case discussed above. We hope that this approach will prove valuable for future research, particularly when processing costs in structurally simple sentences are otherwise difficult to detect with standard methods.