Glossa Psycholinguistics

1.1. The role of memory in the processing of subject-verb agreement

A link between working memory demands and the processing of agreement dependencies has been proposed by McDonald (2008a), in order to account for the fact that – compared to other morphosyntactic structures – agreement morphology is acquired later and is more error-prone in second language (L2) speakers (Goldschneider & DeKeyser, 2005; Paradis, 2005), as well as speakers with language impairments (Druks, 2006). McDonald (2008a) tested healthy adult native speakers with a grammaticality judgement task targeting several English structures. Memory load was manipulated between participants, such that half of the participants had to remember a list of seven digits during each trial, while the other half did not. Participants’ accuracy in the higher memory load condition declined more significantly for subject-verb agreement than for other structures. The results paralleled those of previous studies using a similar paradigm in both comprehension (Blackwell & Bates, 1995; Hayiou-Thomas et al., 2004; McDonald, 2006) and production (Fayol et al., 1994; Hartsuiker & Barkhuysen, 2006). The outcome was interpreted as evidence that agreement dependencies are harder to process than other syntactic dependencies, and hence are more affected by reduced working memory capacity. This was proposed to explain their later acquisition and their higher susceptibility to errors, compared to other structures.

The impact of memory on agreement processing has also been investigated by considering individual measures of working memory capacity. This approach has been recently employed in studies investigating the effects of aging on sentence comprehension. Aging is often linked to a decline in working memory capacity (Carpenter et al., 1994; Kemtes & Kemper, 1997). Using a timed grammaticality judgment task and a self-paced reading task, Reifegerste et al. (2017) investigated how this affected adult German speakers’ susceptibility to agreement attraction errors. They found that – compared to younger participants – older participants took longer to process agreement dependencies and were more likely to accept sentences with attraction errors. Among older participants, performance was modulated by working memory capacity, such that participants with lower working memory scores experienced greater processing difficulty and showed reduced accuracy compared to participants with higher scores.

The results of Reifegerste et al. (2017) were partially replicated in production. Reifegerste et al. (2020) tested agreement production in German L1 and L2 speakers across the adult lifespan, using a binary-choice sentence completion task. They found opposite effects of aging and memory capacity in the two groups of speakers. Accuracy decreased with increasing age among L1 speakers, but showed the opposite tendency among L2 speakers. Working memory capacity, instead, had no effect on the L1 group, but significantly modulated L2 speakers’ performance. Mixed results in agreement production were also reported by Bock and Cutting (1992), who found an effect of individual working memory span on accuracy in only one of the three experiments in their study. By contrast, a strong correlation between working memory and accuracy has been found by Veenstra et al. (2017) in both monolingual and bilingual children, and by Slevc and Martin (2016) in patients with brain injuries.

Following the accounts of agreement processing mentioned above, one may predict that, given two different types of agreement dependencies, the one involving a higher memory load should impose greater processing difficulties, making agreement violations harder to detect. In memory-based models of sentence comprehension (Baddeley, 1986; Gibson, 1998; Just & Carpenter, 1992; King & Just, 1991), memory load is linked to syntactic complexity. For example, sentences containing a filler-gap dependency as the result of a dislocation are thought to impose a greater load on working memory than sentences that can be parsed in a word-by-word fashion. This is often exemplified by the contrast between subject- and object-extracted relative clauses (e.g., The banker that praised the barber climbed the mountain vs. The banker that the barber praised climbed the mountain). Object-extracted relative clauses are considered harder to process, because the initial noun phrase (the banker) needs to be activated and retained in memory while the embedded clause is being processed, before being integrated with its verb phrase (climbed the mountain).

A large number of studies offer empirical evidence in support of the claim that more complex structures impose a higher load on memory (e.g., Gordon et al., 2001, 2002; Mak et al., 2002; Traxler et al., 2002, 2005). With respect to agreement computation, Franck et al. (2010) compared agreement error rates related to object relative clauses vs. complement clauses in French (*John spoke to the patients that the medicine cure vs. *John tells the patients that the medicine cure). Results showed that participants were more likely to produce agreement mistakes in the former configuration than in the latter. They interpreted the results in light of the fact that object relative clauses involve object movement, while complement clauses do not.

In addition to complexity, temporary ambiguity is also thought to generate demand for additional memory resources. In the capacity constrained parsing model (Just & Carpenter, 1992), the greater difficulty of processing temporarily ambiguous, as opposed to unambiguous, sentences is claimed to depend on the cognitive effort of activating and maintaining two (or more) alternative interpretations in memory until the disambiguation point is reached. MacDonald et al. (1992) advanced empirical evidence in support of this claim by analyzing word-by-word reading times of temporarily ambiguous sentences (The soldiers warned about the danger before the midnight raid) vs. unambiguous sentences (The soldiers spoke about the danger before the midnight raid). The former sentences induced higher reading times, consistent with the additional processing cost of generating multiple representations. The same results were replicated by Long and Prat (2008). In our study, we explore the possibility that the additional effort induced by temporary ambiguity is reflected in the processing of subject-verb agreement.

Summarizing, our study tests the hypothesis that OC_litVS structures in Italian may elicit higher processing costs compared to L_ocVS structures. We base our prediction on the observation that OC_litVS structures involve a more complex parsing than L_ocVS structures, primarily due to being temporarily ambiguous with respect to the identification of the subject (see 1.3). Thus, we hypothesized that number agreement violations may be harder to process in OC_litVS as compared to L_ocVS structures. Crucially, this hypothesis would be impossible to test in sentences with a canonical subject-verb order.

1.2 The processing of agreement in sentences with non-canonical word orders

Few psycholinguistic studies – mostly on production – have examined subject-verb agreement in constructions displaying a non-canonical constituent ordering. These studies investigated whether word order affected speakers’ computation of subject-verb agreement dependencies. For example, Vigliocco and Nicol (1998) compared error rates in declarative SV sentences vs. interrogative VS sentences in English (*The helicopter for the flights are safe vs. *Are the helicopter for the flights safe?). The results showed that participants produced the same amount of agreement errors in both configurations. The authors interpreted the outcome as evidence that agreement was not sensitive to word order variations (but see Franck et al., 2006, for different results in Italian). However, the two agreement dependencies differed in terms of linear proximity: in declarative sentences, the distractor intervenes between the controller and the target, while in interrogative sentences, it does not. Thus, the potential proximity advantage in interrogative sentences may be offset by an opposing factor – namely, that agreement computation is more challenging in non-canonical interrogative sentences than in canonical declarative sentences. Capitalizing on this possibility, Santesteban et al. (2013) explored agreement production in canonical SOV vs. non-canonical OSV sentences in Basque. They found that participants’ agreement error rate increased significantly in non-canonical sentences, compared to canonical ones. The results were attributed to greater working memory demands in the OSV order, which limited the resources available for agreement computation.

More generally, previous studies have shown that non-canonical word orders are harder to process than canonical orders, likely due to greater syntactic complexity and infrequency in speech (Bader & Meng, 1999; Clahsen & Featherston, 1999; Hyönä & Hujanen, 1997; Imamura et al., 2016; Kaiser & Trueswell, 2004; Sekerina, 1997; Tamaoka et al., 2005, 2011). Moreover, non-canonical structures are licensed by particular discourse conditions, while canonical ones are more broadly available. In particular, both OSV sentences in Basque and the OVS sentences examined in our study are associated with an information structure in which the object is topicalized (see 1.3). Therefore, the presence of an object displaying features normally associated with the subject (e.g., first position; topicality) might interfere with the identification of the actual subject of the sentence (Badecker & Kumuniak, 2007; Hartsuiker et al., 2003; Thornton & McDonald, 2003). However, the study of Santesteban et al. (2013) considered non-canonical sentences in isolation, leaving open the question of whether the processing costs associated with discourse-marked structures can be mitigated by a licensing context (Bader & Meng, 2023; Brown et al., 2012; Burholt Kristensen et al., 2013, 2014; Clifton & Frazier, 2004; Erdocia et al., 2009; Hoeks et al., 2002; Kaiser & Trueswell, 2004; López-Beltrán et al., 2021; Mak et al., 2008; Roland et al., 2012; Slioussar, 2011; Stolterfoht, 2005; Torregrossa et al., 2024; Weskott et al., 2011; Yano & Koizumi, 2018).

Our study tests the hypotheses of Santesteban et al. (2013) about the higher working memory demands of non-canonical word orders in agreement comprehension in Italian. To do so, it compares two postverbal subject structures that differ with respect to their degree of canonicity (see 1.3). In doing that, it takes a step forward by analysing non-canonical structures within their appropriate discourse conditions (see Experiment 3).

1.3 Postverbal subjects in transitive vs. unaccusative structures in Italian

Italian is a pro-drop language with flexible word order. The canonical order is SVO, such that the subject precedes the verb, as in (1a) and (2a). However, structures with a postverbal subject are also possible in various configurations with all verb classes (Bates et al., 1982; Maiden, 1995; Rizzi, 1982; Vincent, 1990). Here we compare transitive structures with clitic left-dislocation (OC_litVS) vs. unaccusative structures with locative inversion (L_ocVS). In OC_litVS structures, the preverbal position is filled by a topicalized direct object, which is co-referential with a resumptive clitic in the main clause (‘the dress’ in (1b)).¹ In L_ocVS structures, the preverbal position is filled by a locative PP (‘in the shop’ in (2b)). In these structures, the locative constituent is not resumed by a clitic.²

OC_litVS and L_ocVS structures differ in several ways. One is related to temporary ambiguity. In OC_litVS structures, the direct object il vestito ‘the dress’ in (1b) might be initially interpreted as the subject of the sentence, because it is in sentence-initial position and is a topic. By contrast, the locative nel negozio ‘in the shop’ in the L_ocVS structure (2b) cannot receive the same interpretation, because it is preceded by the preposition nel ‘in’. As observed by Franck et al. (2006) in French, the introduction of another potential controller for agreement might hinder agreement computations in OC_litVS structures.

Moreover, temporary ambiguity may affect the parsing of the two structures. In OC_litVS structures, the parser is forced to re-analyse the first constituent as the object upon encountering the clitic, and to establish a referential link between the two elements. Then, the dislocated object needs to be stored in memory before being integrated into the verb phrase after the verb is encountered. By contrast, in L_ocVS structures, the first element is unambiguously interpreted as a fronted adjunct before the verb is encountered. Furthermore, once the parser identifies the unaccusative verb, it can anticipate the occurrence of a postverbal subject, thus minimizing any potential surprise effect of encountering a verb before the subject. This results from the syntax of unaccusative verbs, according to which the grammatical subject is base-generated in postverbal position as an internal argument and behaves syntactically like a direct object (Belletti, 1988; Burzio, 1986; Koopman & Sportiche, 1991; Perlmutter, 1978). Based on these observations, it can be assumed that L_ocVS structures can be parsed in a more straightforward way and with a lower memory load compared to OC_litVS structures.

Another difference is related to discourse conditions. OC_litVS structures are licensed by stricter discourse conditions than L_ocVS structures, because they require the direct object to be informationally given – i.e., mentioned in the preceding discourse. For example, sentence (1b) could be an appropriate answer to a question that mentions the object, like “What about the dress?” or “Who’s observing the dress?”, as in (3). The resulting structure is a topic-focus configuration with an informationally-given object in preverbal position and an informationally-new subject in postverbal position (Arnold et al., 2000; Birner & Ward, 2009; Gundel, 1988; Vallduví, 1990; Ward & Birner, 2011).

By contrast, L_ocVS structures are less dependent on a specific information structure. A sentence like (2b) can be licensed by a context analogous to (3) – i.e., as an answer to the question “Who is entering the shop?”. Crucially, however, the syntax of unaccusative verbs also allows for them to be licensed in sentence-focus conditions – i.e, when the entire sentence is informationally new and all elements are focal (Leonetti, 2018). Thus, (2b) would also be an appropriate answer to a broad-focus question like (4).

Due to these facts, it is fair to assume that L_ocVS structures have a lesser degree of markedness compared to OC_litVS structures (see Adger, 1996; Belletti, 1988, 2018; Bentley & Cruschina, 2018; Cardinaletti, 2018; Pinto, 1994, 1997; Russi, 2014; for discussion). In both cases, the subject linearly occupies the postverbal position, but the fact that (2b) is licensed in sentence-focus conditions makes it a less marked – and more frequent – word order option. In contrast, (1b) is only licensed in narrow focus conditions and, hence, is a more marked – and less frequent – option (for frequency data about postverbal subject structures in Italian, see Bernini, 1995; Sornicola, 1994, 1995). Following Santesteban et al. (2013), we hypothesize that this asymmetry might affect the processing of agreement dependencies, with violations detected with greater difficulty in OC_litVS structures compared to L_ocVS structures.

Interestingly, the differences between the structures under investigation might explain some previous findings in first and second language acquisition. As for first language acquisition – although there are no studies directly comparing the two structures – data across studies point to the fact that Italian postverbal subjects may be mastered at different time points, depending on the type of syntactic structure. In production, Italian children between 18 and 36 months of age are shown to distinguish subjects of unaccusative verbs from those of other verb classes, and to appropriately produce them in postverbal position (Lorusso et al., 2005). By contrast, the production of postverbal subjects with transitive verbs and clitic left-dislocation – despite being comparable to that of adults in terms of number of occurrences – is characterized by the frequent omission of the resumptive clitic (Cairncross & Dal Pozzo, 2022; Lorusso, 2014, 2017). In comprehension, the study of Abbot-Smith and Serratrice (2015) shows that at age 4;6, children still misinterpret postverbal subjects with transitive verbs and clitic-left dislocation. Taken together, these findings suggest that Italian children may have greater difficulty mastering postverbal subjects in transitive structures than in unaccusative structures. Similar results have been found with L2 learners and bilingual speakers (Belletti et al., 2007; Belletti & Leonini, 2004; Bettoni et al., 2009; Caloi et al., 2018; Listanti & Torregrossa, 2023, 2024; Nuzzo, 2015). Processing-based accounts of language acquisition (McDonald, 2008a) link differences in the timing of acquisition to varying working memory demands across linguistic phenomena. In this study, we test the hypothesis that such differences are reflected in varying processing difficulties in adulthood.

To summarize, OC_litVS structures in Italian differ from L_ocVS structures on a number of dimensions, including potential temporary ambiguity, higher parsing complexity, stricter discourse licensing conditions and a different timing of acquisition. Since all these differences may result from a greater working memory load in OC_litVS, we predicted that subject-verb number agreement would be harder to process in these structures compared to L_ocVS structures.

1.4 The present study

We measured adult Italian speakers’ sensitivity to subject-verb number agreement violations in sentences with postverbal subjects. We conducted three experiments to examine whether the detection of these violations was different in OC_litVS structures vs. L_ocVS structures. Experiments 1 and 2 presented target sentences in isolation, while Experiment 3 presented sentences preceded by a licensing context. The single-sentence presentation mode in Experiments 1 and 2 allowed us to focus on sentence-level processing, putting aside the sentences’ pragmatic felicity and the potential contribution of discourse-integration processes. Experiment 3 addressed whether the results of Experiments 1 and 2 extended to situations in which postverbal subjects were discourse-licensed.

2.1 Methods

2.2 Results

2.2.1 Sentences with postverbal subjects (test items)

Figure 1 shows descriptive summaries of acceptance rate and response times in items with postverbal subjects (test items). With regard to acceptance rates, participants accepted grammatical sentences more often than ungrammatical sentences, resulting in a main effect of grammaticality (β = –4.86, SE = 0.38, z = –12.74, p < .001). Moreover, participants accepted OC_litVS structures less often than L_ocVS, resulting in a main effect of structure (β = –1.63, SE = 0.21, z = –7.76, p < .001). Finally, the interaction between grammaticality and structure was not significant (β = 0.35, SE = 0.42, z = 0.82, p = .41).

Figure 1

Empirical acceptance rates and response times for sentences with postverbal subjects in Experiment 1 (test items). Error bars show bootstrapped confidence intervals computed across participants and items. Each point represents one participant.

Response times showed a two-way interaction between structure and grammaticality (β = –0.23, SE = 0.1, t = –2.28, p = .02). Inspection of Figure 1 (right panel) suggests that this was due to opposite effects of grammaticality in OC_litVS structures vs. L_ocVS structures. However, pairwise comparisons did not reveal an effect of grammaticality in either OC_litVS structures (β = 0.14, SE = 0.09, t = 1.49, p = .14) or L_ocVS structures (β = –0.08, SE = 0.06, t = –1.24, p = .20). In fact, the numeric trends suggested a stronger effect of grammaticality in transitive sentences than in unaccusative sentences. Thus, neither response times nor acceptance rates provided evidence that the detection of agreement violations was harder in OC_litVS sentences than in L_ocVS sentences.

2.2.2 Sentences with preverbal subjects (control items)

Figure 2 shows descriptive summaries of acceptance rates and response times in the sentences with preverbal subjects (control items). As expected, acceptance rates only showed a main effect of grammaticality, with grammatical sentences being accepted more often than ungrammatical sentences (β = –6.39, SE = 0.49, z = –13.07, p < .001). Response times only showed a significant effect of grammaticality, with faster responses for ungrammatical sentences than for grammatical sentences (β = 0.61, SE = 0.05, t = 11.54, p < .001).

Figure 2

Empirical acceptance rates and response times for sentences with preverbal subjects in Experiment 1 (control items). Error bars show bootstrapped confidence intervals computed across participants and items. Each point represents one participant.

2.3 Discussion

We hypothesized that agreement violations with postverbal subjects would be harder to detect in OC_litVS structures than in L_ocVS structures. The results of Experiment 1 failed to support this hypothesis, as acceptance rates did not show a significant interaction between grammaticality and structure. Participants clearly detected ungrammatical sentences, and there was also evidence that OC_litVS structures were accepted less often than L_ocVS structures. But, crucially, there was no evidence that the detection of violations occurred less quickly or less often in OC_litVS sentences than in L_ocVS sentences. An interaction between grammaticality and structure did appear in response times, but the pairwise comparisons did not reach significance.

The lack of a significant interaction between structure and grammaticality in the test items is unlikely to result from lexical differences between the transitive and unaccusative verbs used. This is because the control items, which featured the same verbs but with preverbal definite subjects, showed the predicted main effect of grammaticality in the absence of any effects of structure or interactions. This suggests that, irrespective of structure, participants uniformly rejected ungrammatical sentences with preverbal subjects and were faster to judge them, compared to grammatical sentences. The latter outcome is in line with previous studies showing that if the unacceptability of a structure is evident, participants are faster to reject it than to accept its grammatical counterpart (De Vogelaer et al., 2020; Lago et al., 2019).

While the results of Experiment 1 fail to support an effect of type of structure in the detection of agreement violations with postverbal subjects, it could be argued that the judgment task only provided a coarse test of this hypothesis. This is because responses were collected only at the end of the sentence, rather than when postverbal subjects were encountered. Specifically, it is possible that the processing of postverbal subjects was initially harder in OC_litVS structures, but that participants had overcome this difficulty by the end of the sentence, when their judgments were collected. To address this possibility, Experiment 2 used a self-paced reading design to measure processing difficulty time-locked to the appearance of postverbal subjects.

3.1 Methods

3.2 Results

3.2.1 Sentences with postverbal subjects (test items)

Figure 3 shows a descriptive summary of the reading times in items with postverbal subjects (test items). The results in the combined critical and spillover regions showed a main effect of grammaticality, with longer reading times in ungrammatical sentences than in grammatical sentences: (β = –0.09, SE = 0.02, t = –5.05, p < .001). The predicted interaction between structure and grammaticality failed to reach significance (β = 0.04, SE = 0.02, t = 1.56, p = .12).

Figure 3

Participants’ reading patterns in sentences with postverbal subjects in Experiment 2 (test items). Error bars show +/– 1 standard errors averaged across participants and items.

3.2.2 Sentences with preverbal subjects (control items)

Figure 4 shows a descriptive summary of the reading times in items with preverbal subjects (control items). The results showed a main effect of grammaticality, with longer reading times in ungrammatical sentences than in grammatical sentences (β = –0.18, SE = 0.02, t = –9.2, p < .001). There was also a main effect of region, with longer reading times in the spillover region than in the critical region (β = –0.08, SE = 0.02, t = –5.1, p < .001).⁶ Finally, there was an interaction between grammaticality and region (β = –0.17, SE = 0.03, t = –6.2, p < .001). The interaction was due to a larger grammaticality effect in the spillover region (β = –0.26, SE = 0.02, t = –13.22, p < .001) than in the critical region (β = –0.09, SE = 0.02, t = –4.44, p < .001). Crucially, no significant interaction between structure and grammaticality was found.

Figure 4

Participants’ reading patterns in sentences with preverbal subjects in Experiment 2 (control items). Error bars show +/– 1 standard errors averaged across participants and items.

4. Experiment 3: Self-paced reading with context

4.2 Results

4.2.1 Sentences with postverbal subjects (test items)

Figure 5 shows a descriptive summary of the reading times in sentences with postverbal subjects (test items). There was a main effect of grammaticality (β = –0.08, SE = 0.02, t = –3.28, p < .005), indicating longer reading times in ungrammatical conditions. There was also a main effect of structure (β = 0.1, SE = 0.03, t = 3.49, p < .001), indicating longer reading times in OC_litVS structures than in L_ocVS structures. There was also a main effect of region, indicating longer reading times in the spillover region vs. the critical region (β = –0.08, SE = 0.02, t = –3.87, p < .001).

Figure 5

Participants’ reading patterns in sentences with postverbal subjects in Experiment 3 (test items). Error bars show +/– 1 standard errors averaged across participants and items.

As in previous experiments, the interaction between grammaticality and structure was not significant (β = –0.04, SE = 0.05, t = –0.73, p = .47). However, visual inspection of Figure 5 suggested a larger effect of grammaticality in the critical region for L_ocVS structures as opposed to OC_litVS structures. We followed up on this observation by running an exploratory analysis and found that the effect of grammaticality was significant in L_ocVS structures (β = –0.14, SE = 0.04, t = –3.47, p < .001), but not in OC_litVS structures (β = –0.06, SE = 0.04, t = –1.4, p = .16).

4.2.2 Sentences with preverbal subjects (control items)

Figure 6 shows a descriptive summary of the reading times in sentences with preverbal subjects (control items). There was a main effect of grammaticality (β = –0.15, SE = 0.03, t = –5.17, p < .001), indicating longer reading times in ungrammatical conditions than in grammatical conditions, and a main effect of region (β = –0.27, SE = 0.02, t = –12.37, p < .001) indicating longer reading times in the spillover region than in the critical region. There was also an interaction between grammaticality and region (β = –0.24, SE = 0.04, t = –5.64, p < .001). Pairwise comparisons indicated that the effect of grammaticality was significant in the spillover region (β = –0.27, SE = 0.03, t = –9.09, p < .001), but not in the critical region (β = –0.04, SE = 0.03, t = –1.17, p = .24).

Figure 6

Participants’ reading patterns in sentences with preverbal subjects in Experiment 3 (control items). Error bars show +/– 1 standard errors averaged across participants and items.

4.3 Discussion

The results of Experiment 3 replicated those of Experiment 2, with a few exceptions. First, OC_litVS structures were read significantly more slowly than L_ocVS structures. This effect was surprising, because we expected a licensing context to facilitate the reading of OC_litVS structures. Second, pairwise comparisons at the critical subject region revealed a significant effect of grammaticality in L_ocVS structures, but not in OC_litVS structures. This second result is consistent with the hypothesis that agreement violations might be harder to detect in OC_litVS structures, due to these structures being more complex and, thus, eliciting higher memory load. However, the pairwise comparisons were not supported by an interaction between grammaticality and structure, which is the appropriate statistical test for our hypothesis. Due to the lack of a significant interaction and the fact that Experiment 3 had fewer participants than Experiment 2 (and, thus, less statistical power), we do not view the results as strong support for the hypothesis that the processing of agreement in sentences with postverbal subjects (within an appropriate context) is modulated by the type of structure involved.

As in Experiment 2, an unexpected result was that the grammatical and ungrammatical conditions seemed to differ at the verb region in sentences with postverbal subjects, this time with unaccusative verbs. This is unexpected, because the sentence only became ungrammatical at the subject. However, the difference seems to disappear at the pre-critical region, so we think it is unlikely to affect our comparisons at the critical subject region.

5.1 No evidence that the type of structure affects the detection of agreement violations with postverbal subjects

Our results failed to support the prediction that agreement violations would be harder to detect in OC_litVS structures than in L_ocVS structures. Experiment 1 failed to show the predicted interaction between structure and grammaticality in acceptance rates, and the response times actually showed an interaction in the opposite direction. Experiments 2 and 3 also failed to show the predicted interaction in the reading times of the postverbal subject and the spillover region.

The only evidence consistent with our prediction was in the pairwise comparisons of Experiment 3, which revealed a significant effect of grammaticality at the point of the postverbal subject in L_ocVS structures, but not in OC_litVS structures. But the pairwise comparisons cannot be taken as conclusive, given that they were not replicated in Experiment 2 (which had a larger participant sample) and that they were not supported by a significant interaction – which is the appropriate statistical diagnostic of our prediction (for discussion of the need for interactions to support claims about asymmetric effects, see Gelman & Stern, 2006; Nieuwenhuis et al., 2011). Therefore, the combined results of the three experiments did not yield evidence of a modulatory effect of the type of structure on the detection of agreement with postverbal subjects.

Our findings do not align with those of Santesteban et al. (2013) about agreement production in OSV and SOV sentences in Basque. However, it is important to note that Santesteban et al. (2013) compared canonical sentences with non-canonical ones. By contrast – although one of the two structures is considered less marked than the other – the subject phrase in both structures in our study occupied a position different from the one considered canonical in Italian (i.e., the preverbal position). Therefore, it is possible that the effect of a higher memory load associated with OC_litVS structures was masked by the overall difficulty of processing agreement in postverbal subject configurations. Some evidence in support of this explanation comes from the comparison of agreement processing patterns in sentences with postverbal subjects vs. preverbal subjects. In the judgment task (Experiment 1), the average acceptance rate of sentences with agreement violations was lower with preverbal subjects than with postverbal subjects in both structure types: transitives: 3% vs. 4.5%; unaccusatives: 3.9% vs. 15%. In the reading tasks (Experiments 2 and 3), sentences with violations elicited longer reading times as soon as the incorrect agreement marking was encountered – consistent with processing disruption. However, the mean reading disruption was higher with preverbal subjects than with postverbal subjects (Experiment 2: 67 ms vs. 28 ms; Experiment 3: 49 ms vs. 22 ms). These findings suggest that – regardless of syntactic structure – Italian adults’ ability to detect agreement violations is less robust in non-canonical postverbal subject structures than in canonical preverbal subject structures.

The results with preverbal subjects replicate the high sensitivity to agreement violations previously demonstrated in behavioral measures and brain responses in Italian (Biondo et al., 2018; De Vincenzi et al., 2003; Mancini et al., 2014) and other languages (Lago et al., 2021; Mancini et al., 2011; Pearlmutter et al., 1999; Reifegerste et al., 2017; Severens et al., 2008), and they show that our participants were performing the experiments attentively.

One explanation for our results is that some of the differences between the two structures under investigation were not strong enough to affect the processing of agreement. With respect to the syntactic differences, it should be noted that some Italian dialects allow for the use of left-dislocated locative PPs without a resumptive clitic (Benincà & Poletto, 2007; Ledgeway, 2009; Loporcaro, 2009). Such structures are superficially identical to the L_ocVS structures used in our experiments. Nevertheless, they involve a topical left-dislocated constituent like the OC_litVS structures. Hence, it is possible that some of the participants interpreted L_ocVS structures as left dislocations and not as locative inversions, especially in Experiment 3, where the context sentence makes the preverbal constituent topical. Future research could address this possibility by comparing the processing of preverbal locatives with and without a resumptive clitic. Finally, regarding the temporary ambiguity in OC_litVS structures, it is important to note that all our stimuli featured inanimate objects, which may have decreased the probability of their being interpreted as subjects, thereby reducing overall ambiguity.

5.2 General effects of type of structure on the processing of postverbal subjects

While there was no evidence that the type of structure differentially affected the detection of agreement violations, some of our findings do suggest that the OC_litVS and L_ocVS structures elicit differential processing effort. In Experiment 1, we found a main effect of structure in the acceptance rates of sentences with postverbal subjects. The outcome was due to the fact that OC_litVS structures were accepted significantly less often than L_ocVS structures. Interestingly, the difference emerged only in the grammatical sentences, while ungrammatical sentences were identified accurately regardless of structure type. We speculate that this might be explained by the fact that in Experiment 1, the sentences were presented in isolation. L_ocVS structures are felicitous in sentence-focus conditions, while OC_litVS structures are only licensed if the preverbal object is informationally-given. Therefore, OC_litVS structures are more sensitive to contextual constraints and, hence, should be less acceptable in isolation. Thus, the lower acceptability of grammatical sentences may be explained by the fact that the discourse requirements for the use of OC_litVS structures were not met in Experiment 1.

Meanwhile, OC_litVS structures elicited significantly longer reading times than L_ocVS structures when preceded by a licensing context (Experiment 3), but not when the context was absent (Experiment 2). The outcome is at odds with the expectation that the processing of informationally-marked sentences is facilitated by appropriate discourse conditions. Instead, our results point to the possibility that the integration of contextual information is associated with greater processing effort in comparison to marked sentences without context. From this standpoint, the longer reading times associated with OC_litVS structures in Experiment 3 might indicate that participants spent more time integrating the target sentences with the preceding context to construct an appropriate interpretation. A possible cause may be that the establishment of a reference for the clitic pronoun in the target sentence was initially unsuccessful: participants may have first failed to find a referent in the context sentence before realizing that it is co-referential with the topicalized object. This hypothesis requires further investigation.

Taken together, our results suggest that, independently of their grammaticality, OC_litVS structures require an appropriate context to be comprehended and accepted (Experiment 1), but that the presence of a context may increase their processing cost online (Experiment 3). In this respect, our results appear to be in line with previous studies showing that a supportive context helps global comprehension of non-canonical structures, but it does not necessarily affect online measures, such as reading times (Bornkessel & Schlesewsky, 2006; Burholt Kristensen et al., 2014).

5.3 Limitations and future directions

While our results do not support the hypothesis that the type of structure – OC_litVS vs. L_ocVS – modulates the processing of agreement with postverbal subjects, some limitations should be acknowledged before this conclusion can be taken at face value. First, we cannot rule out the possibility that evidence of such sensitivity would have emerged if other processing paradigms had been used, such as brain-related potentials or reading eye-tracking (Dillon et al., 2013; Jäger et al., 2020; Pearlmutter et al., 1999; Schlesewsky et al., 2003; Yano & Koizumi, 2018; Zawiszewski & Friederici, 2009). For example, previous reading eye-tracking studies have reported that regressive eye movements are a key indicator of sensitivity to agreement violations (Dillon et al., 2013; Jäger et al., 2020; Pearlmutter et al., 1999). Since the self-paced reading paradigm does not allow readers to regress to previous material, our task may have lacked the sensitivity to detect differences between structures with transitive and unaccusative verbs.

Second, we may have lacked the statistical power to detect agreement processing differences between OC_litVS and L_ocVS structures. Recent recommendations suggest that samples of around 110 participants are needed to detect a 2×2 medium-size interaction in within-subject designs (Brysbaert, 2019). Of our three experiments, only Experiment 2 met this threshold (200 participants), while Experiments 1 and 3 had fewer participants (56 and 96, respectively).

To address this possibility, we collapsed the two self-paced reading experiments (Experiments 2 and 3) and conducted a statistical analysis on the combined data (Appendix S4). The combined dataset included 296 participants and 1.776 observations per condition, which is considered adequate to detect a 2×2 interaction with an effect size identified as medium (Brysbaert & Stevens, 2018). In line with the single-experiment analyses, our combined analysis did not yield a significant two-way interaction between grammaticality and structure in the critical regions of interest. Thus, although we cannot rule out the possibility that our effect of interest might emerge with a larger number of observations, it seems reasonable to conclude that if such an effect exists, it is relatively small, requiring more than 296 participants to be detected in a self-paced reading task like ours.

In sum, although our main prediction was not borne out, some of our results suggest that transitive and unaccusative structures with postverbal subjects in Italian may elicit a differential processing cost. Specifically, the processing difficulty associated with one structure vs. the other might not affect agreement morphosyntactic computations, but it might disrupt other processes, such as discourse integration and thematic role assignment. If this were the case, measures directly targeting the interpretation of OC_litVS vs. L_ocVS structures may be more successful than agreement processing in detecting processing differences between them. We believe that this is a productive avenue for future research.