A robust finding about agreement attraction is that number attraction is typically attraction to a plural, not a singular, noun: while agreement errors like (1) are attested, errors like (5) are less frequent.

Furthermore, several studies report that the effect of a number attractor in production depends on its hierarchical (not linear) position in the syntactic representation of the sentence (Bock & Cutting, 1992; Vigliocco & Nicol, 1998). According to these findings, a plural noun in an embedded clause, like books in (6a), attracts less than a plural noun that is structurally closer to the verb, as in (6b). This happens even though the plural noun books in the two sentences is equally linearly adjacent to the verb, and even though the plural verb form is ungrammatical in both cases.

Other studies focus on the semantic number features of the attractor, testing the impact of notional number on agreement. For example, Humphreys and Bock (2005) used a sentence completion task to compare constructions like (7a), in which a collective interpretation of the noun gang was plausible, to constructions like (7b), in which only a distributive (and hence plural) interpretation was plausible. They report more plural verbs in completing (7b) than (7a) (see also Eberhard, 1997).

Vigliocco (1996) also compared singular head nouns combined with plural modifiers that are typically denoting multiple tokens (e.g., the label of the bottles, which implies multiple labels) with similar head nouns more plausibly denoting a single entity (e.g., the journey on the islands, which implies one single journey). The former triggered more plural verb forms by Italian speakers compared to the latter, although no effects of distributivity were revealed in an eye tracking reading experiment on similar materials in Spanish (Acuña-Fariña et al., 2014).

From a psycholinguistic perspective, different accounts have been proposed to explain attraction effects in production and comprehension. Since Bock et al. (2001), notional plurality plays a key role in the “Marking and Morphing” model of sentence production (see also Eberhard et al., 2005). During the marking stage, a plural noun (like cabinets in (1)) might contribute to a bias towards the notional plurality of the head noun, making plural agreement (at the morphing stage) more likely. According to cue-based retrieval models (Lewis & Vasishth, 2005, see also Smith & Vasishth, 2020), in sentences like The key to the cabinets are…, the PL feature on the verb boosts the activation of the other nouns in the sentence that share this feature (i.e., cabinets), causing interference.

Recently, Smith et al. (2018) proposed a critical revision of the “Marking and Morphing” account (cf. also Smith et al., 2021; Villata & Franck, 2020; Villata & Tabor, 2022). The account proposed by Smith et al. (2018) is directly relevant for our purposes since it builds on pseudopartitives. Under it, variability in agreement in production and comprehension is accounted for as the result of a dynamic interplay between semantics and syntax in a self-organized sentence processing (SOSP) model. According to the SOSP model, there is no sequential separation between the marking stage (which refers to notional number) and the morphing stage (which refers to morphosyntactic agreement). Rather, the system dynamically builds treelets (i.e, small pieces of syntactic structure) during incremental processing, leaving space for grammar-flouting interference and local coherence effects (cf. also Hammerley et al., 2019, for an insightful discussion about competing theories of agreement attraction).

Pseudopartitives do not really have a dedicated psycholinguistic literature. There is, on the other hand, a considerable tradition of studies on pseudopartitives in linguistics. First, these constructions are contrasted with partitive constructions. Partitive constructions (cf. (8a)) involve a quantifier (or a quantified expression) embedding a Determiner Phrase (DP, in this case, the senators) that represent the “whole” from which the head of the construct (in this case, a group) selects a part (Chierchia, 1998; Zamparelli, 2008). The embedded nominal in pseudopartitive constructions like (8b) instead is a bare NP, which is “measured” by the DP (a group (of)) (Milner, 1978; Schwarzschild, 2006; Selkirk, 1977).

Second, following the seminal work of Selkirk (1977), agreement alternations in pseudopartitives, such as the one exemplified in (2), have been explained in terms of structural ambiguity, as represented in (9). In (9a), the second NP (the senators) is embedded as a modifier of the head noun (a group) by means of the preposition of. The verb agrees with the singular head of the DP as a whole (a group), while the plural feature remains within the embedded NP. In (9b), the embedded NP senators (case marked by of) is modified by a group, yielding an interpretation in which senators are “measured” by a group. In this case, the plural feature of the second NP percolates to the higher DP, triggering plural agreement on the verb (cf. also Manzini, 2019; Manzini & Franco, 2019).

Crucially, the adoption of one or the other syntactic analysis for these constructions also has interpretative consequences (Landman, 2004, 2016; Rothstein, 2009, 2017), which are exemplified in (10). In (10a) the only plausible interpretation is the measure one (chickpeas in the measure of a can), and both verbal agreements are available. In (10b) the only available interpretation is the container interpretation (a can is being referred to, as modified by its content, the chickpeas) which forces a singular verb.

According to Rett (2014, following Landman, 2004), the alternation of container/measure interpretations in pseudopartitives can be viewed as a case of genuine lexical ambiguity of the head nouns in pseudopartitives that trigger one of the two alternative syntactic constructions in (9). Based on two acceptability judgment studies of sentences with container-headed pseudopartitives modified by a relative clause, Duek and Brasoveanu (2015) show that both readings can be simultaneously available in the grammar.

Third, the extent to which the two alternative interpretations are available also depends on the type of head of the pseudopartitive. These are of different kinds: they can be quantificational measure phrases like a lot in (11), collective nouns, implying numerosity, like group in (8b), or containers, implying volume, like box in (12).

In the already cited work, Smith et al. (2018) tested English participants in a forced-choice selection task (Staub, 2009) and found that the probability of choosing a plural verb after a pseudopartitive depended on the type of head, with more choices of plural forms after a quantifier head noun, and more selection of singular forms after a container head. In terms of the SOSP model, when processing a sentence like (8b), the morphosyntactic and semantic features of group (i.e., syntactically being a noun, morphologically being singular, and semantically having features “related to subjecthood versus quantifierhood”) will contribute incrementally to determine the morphosyntactic agreement features of the verb. According to Smith et al., thus, what shapes the participants’ agreement preferences in the case of pseudopartitives is the semantics of the head nouns, triggering plurality, beyond their singular morphosyntactic traits.

Mazzaggio et al. (2020, Experiment 1) also reported a difference in subject-verb agreement preferences across different types of pseudopartitives in an acceptability judgment task in Italian. Participants were more likely to accept a sentence with a plural verb when the pseudopartitive’s head was a quantifier, compared to the case in which the head was a collective noun, which was most preferred with a singular verb. Moreover, variability in verb agreement preferences emerged across individuals. According to Mazzaggio et al. (2020), the observed agreement alternations in pseudopartitives, both between participants of the same language and across languages, are rooted in syntax as part of linguistic competence, and are traced back to the dual labeling options available for these constructions exemplified in (9). Specifically, the accessibility of the measure phrase reading (i.e., the interpretation in which the measure phrase modifies the lexical noun) is maximally available in the case of quantifiers.

Agreement alternations in pseudopartitives and the phenomenon of agreement attraction introduced in 1.1 and 1.2 seem to share a number of properties. First, agreement attraction is typically attraction to the plural, not to the singular. Similarly, if a plural pseudopartitive head embeds a singular NP, the plural head obligatorily triggers plural agreement, as in (13):3

Second, the acceptability of plural agreement in pseudopartitives depends on the type of head. Similarly, in agreement attraction, the semantics of the noun phrase (i.e., distributive, as in the key to the cabinets, vs. non-distributive, as in the cage with the gorillas) modulates the probability of the verb’s agreeing with the embedded NP.

In the psycholinguistic literature on agreement attraction, notional plurality has been presented as the key factor to explain most of the observed alternations in production. The more the subject DP is perceived as a plural entity, the more likely it will be marked as plural in the marking stage, and thus plural agreement will surface at the morphing stage (“Marking and Morphing” model). As Smith and colleagues (2018) point out, notional plurality is a black box that relies on the intuitive notion of notional number carried by a certain noun or noun phrase. Previous works have measured the “degree” of plurality of a given construction employing questionnaires, thus providing a post hoc explanation of notional plurality effects. Testing pseudopartitive constructions, Smith et al. (2018, pp. 1058–59) provide a more formalized account of notional plurality based on three features, namely, +/–Container, +/–Spatial configuration, and +/–Abstract N2, “placing subject NPs along the semantic feature cline related to subjecthood versus quantifierhood”.

Yet, some facts remain without explanation in accounts of agreement in terms of notional plurality. In the first place, cross-linguistic differences need to be considered. While some languages display some variability in subject-verb agreement in the case of collective nouns like police, as attested by English (14a), this is not true for all languages. For example, Italian rules out plural subject-verb agreement when the subject is a singular collective noun, as shown in (14b), which is the Italian translation of (14a). In other words, if notional number is invoked (in any form) in Italian, one needs to explain why there is no independent evidence for it in the case of notionally plural and morphologically singular collective nouns.

Another issue concerning the explanation based on notional number ensues from the fact that in languages like Italian, variability is attested even with respect to subject-verb gender agreement. This is shown in (15), reported by Mazzaggio et al. (2020, Experiment 2), and in examples like (16a–b),4 which again come from local newspapers.

If plural number has an obvious semantic content, it is generally assumed that gender carries no genuine semantic traits in most of the nouns in languages with grammatical gender (Vigliocco & Franck, 1999; Vigliocco et al., 1995). Nonetheless, variability in gender agreement is attested in Italian and in other gender-marked languages (cf. Acuña-Fariña et al., 2014, for Spanish; Slioussar & Malko, 2016, for Russian; Tucker et al., 2021, for Arabic).

In this study, we compare structurally ambiguous pseudopartitives in Italian with superficially similar constructions of the form NP1_[–PL] prep NP2_[PL] that only allow a single grammatical parse. Our first experimental question is whether agreement in pseudopartitives behaves differently from subject-verb agreement in classic NP1_[–PL] prep NP2_[PL] configurations like The key_[–PL] to the cabinets_[PL]. Despite the surface similarities between these two structures, only one grammatical option (i.e., singular agreement) is allowed by the grammar in the latter case, albeit plural verb agreement might result in speakers’ performance from interference effects of the plural attractor. On the contrary, variability in agreement in pseudopartitives is dictated by the availability of two possible parses in the grammar, as seen in (9), thus reflecting speakers’ competence. Interestingly, Smith et al. (2018, p. 1068) recognize that “in classical cases” of agreement attraction “there is only one major syntactic structure at play, while in pseudopartitives, there are two”. Yet their study does not extend to an experimental comparison between the two. We hypothesize a gradient difference in the acceptability of plural verbs between adnominal controls headed by a singular noun and pseudopartitives, and among pseudopartitives headed by different nouns.

To test this hypothesis, in all the experiments, we compare pseudopartitive constructions to control sentences in which a singular head with a plural modifier is followed by either a singular or a plural verb, as exemplified in (17).

Note that agreement attraction phenomena are not investigated per se, but only insofar as they provide a control for the pseudopartitive data, establishing that agreement in pseudopartitives reflects grammatical competence as opposed to a mere processing bias.

A second major comparison carried out in the present study involves the variability in agreement preferences across different kinds of pseudopartitives. Considering the two alternative parses of pseudopartitive constructions, detailed in (9), we ask whether the acceptability of singular or plural verb varies depending on the type of the pseudopartitive’s head. As we have discussed, the head determines the accessibility of a measure phrase interpretation as opposed to a container interpretation, as shown in (10a–b). Concerning this question, we compare pseudopartitives introduced by a quantifier, a collective, or a container head followed by a singular/plural verb. We predict the highest compatibility of plural verbs in the case of quantifier heads, and the lowest compatibility in the case of container heads, both in production and comprehension.

To test our two research questions, we carried out three experiments testing Italian speakers’ preferences in offline and online tasks, in comprehension, production, and reading.

In Experiment 1, we tested Italian speakers’ acceptability of pseudopartitives and surface-similar constructions followed by singular or plural verbs. In Experiment 2, we tested participants’ preferences for singular/plural verb after the above-mentioned constructions in a forced-choice task that mimics production. In Experiment 3, we recorded participants’ eye movements in reading to gather evidence for local disruption during online sentence processing, if any. In line with previous findings on coordinated subjects (Foppolo & Staub, 2020; Keung & Staub, 2018), we predicted a different pattern of eye gaze during the online processing of putatively ungrammatical sentences, in which the verb erroneously agrees with the intervening noun (as in (17)), as compared to pseudopartitive constructions in which variability within and between subjects and items has been documented, and is rooted in the grammar.

Figure 1 shows the ratings’ distribution for each type of sentence, split by type of head and verb number. In control (singular, plural) conditions, the ratings are as expected: below 2 when a plural verb follows a singular subject, and when a singular verb follows a plural subject; above 6 when a plural verb follows a plural subject, and when a singular verb follows a singular subject. The presence of a plural or singular intervener between the sentential subject and the verb does not modulate this effect. In the case of pseudopartitives, singular agreement is overall preferred to plural agreement, especially when the head is a container or a collective noun (Figure 1, left panel). Interestingly, however, the ratings when a plural verb follows a pseudopartitive construction are more variable and not so low (Figure 1, right panel). As is evident in the graph, the ratings when the verb is plural vary depending on the type of head, ranging from the lowest rating when the head is a container noun (M = 2.58, SD = 1.12, Median = 2) to middle-scale ratings when the head is a quantifier noun (M = 4.67, SD = 1.28, Median = 5), with collective nouns lying in between (M = 3.53, SD = 1.03, Median = 3).

By inspecting subjects’ mean distribution of ratings (Figure 2), an interesting variability across participants is revealed in pseudopartitives followed by a plural verb, compared to a singular verb. This is most evident in the case of quantifier head nouns: when these are followed by a plural verb, peaks in the distribution are attested around grades 4 and 6 in the scale.

We implemented two mixed-effects ordinal regression models with a logit link function, using the clmm () function in the ordinal package (Christensen, 2018), which is the most suitable to treat ratings that cannot be assumed to represent equally spaced points in an interval scale. We focused on the contrast between singular/plural verbs following singular controls and the different types of pseudopartitives, thus excluding plural controls from the analyses. This choice was made in consideration of the fact that pseudopartitive constructions superficially resemble singular controls: in both constructions, a singular head noun is modified by a plural noun, and then followed by a putatively grammatical singular verb or a putatively ungrammatical plural verb. In a first model, we modeled Ratings as a function of Condition (contrasting pseudopartitives and singular controls, coded as –0.5 and +0.5, respectively) and Verb Number (contrasting plural and singular verbs, coded as –0.5 and +0.5, respectively), also considering the interaction of these factors. We also included random intercepts for participants and items (including random slopes resulted in a failure of convergence). A general significant difference is revealed between pseudopartitives and controls (Est. = –0.9662, Std. Err. = 0.1237, z-value = –7.808, p < .0001), and a significant preference for singular compared to plural verbs (Est. = 3.7082, Std. Err. = 0.0835, z-value = 44.410, p < .0001). Furthermore, a significant interaction of Condition and Verb Number emerged (Est. = 2.1014, Std.Err. = 0.1304, z-value = 16.112, p < .001): the acceptance rate of plural verbs is higher in pseudopartitives (M = 3.59, SD = 1.09) than in singular controls (M = 1.81, SD = 0.75), while singular verbs are overall accepted across conditions (M = 5.98, SD = 0.67 for pseudopartitives; M = 6.06, SD = 0.64 for singular controls).

In a second model, we set a backward difference coding schema for factor variables with 4 levels (in our case: singular, container, collective, quantifier), assuming sliding differences from level 1 (singular controls) to level 4 (quantifier-headed pseudopartitives) in their compatibility with a plural verb. This coding schema is summarized in Table 1 and was applied to all experiments.

The second model included Ratings as the dependent variable, Head Type and Verb Number as independent variables (in the contrasts defined above), Head Type by Verb Number interaction, random intercepts for participants and items, and random slopes for participants (including random slope for items resulted in convergence failure). All contrasts and interactions are fully significant and are summarized in Table 2.

These results show that the preference for singular over plural verbs (i.e., the distance in acceptability scores for plural vs. singular verbs) is significantly reduced for quantifier head nouns, and it is maximal for singular controls, with significant (and graded) differences from singular to container heads, from container heads to collectives and from collectives to quantifier heads.

This first study returned two main findings. First, pseudopartitive constructions, in which a measurement relation links the singular head with the plural (measured) noun, pattern differently from other constructions in which a singular head noun (such as the level) is modified by a plural intervener (such as of the lakes). In the latter case, the presence of a plural intervener between the subject and the verb does not affect the preference for the singular verb (which is the only grammatical option) and, most importantly, does not alter the acceptability of a plural verb, which remains extremely low overall. In the case of pseudopartitives, instead, although singular agreement is overall preferred over plural agreement, the ratings when a plural verb follows a pseudopartitive construction are less degraded than the singular head controls.

Second, also in line with previous findings, a gradient difference is observed across pseudopartitives depending on the type of head.

These facts speak to the experimental questions outlined above: (i) beyond surface similarities, pseudopartitives are more subject to variability in preferences with respect to other constructions that are superficially similar but do not involve a measure phrase construction and for which strict rules of subject-verb agreement apply; (ii) the acceptability of pseudopartitives followed by plural verbs does not seem to depend on notional plurality per se, as all these constructions are notionally plural in terms of the set they denote. Instead, the variability in participants’ judgments seems to depend on the semantic features of the head noun of the pseudopartitive construction (cf. Smith et al., 2018): in the case of container and collective heads, a clear preference emerges for singular verb forms, and this is greater for container heads; this preference is significantly reduced in the case of quantifier headed nouns. Moreover, the acceptance rate of plural verbs gradually increases from singular controls (for which plural verbs are ungrammatical), to containers (for which it remains low), to collectives (for which it improves), to quantifiers (for which the plural verb is not perceived as much degraded, with mean judgments above 4.5).

Figure 3 (left panel) shows the proportion of singular/plural verb choices. As it is evident, the control conditions patterned as expected: there is a clear preference for a singular verb (94.7%) after a complex NP in which the head is singular and the modifier contains a plural noun (e.g., the level of the lakes, cf. (17)), and there is a clear preference for a plural verb (92.5%) after a complex NP in which the head is plural and the modifier contains a singular noun (e.g., the pilots of the aircraft). With pseudopartitives, the preference for a singular verb, albeit predominant (86.8%) overall, varies gradually depending on the type of head: plural verbs are selected 24.2% of the time after quantifier head nouns, 9.5% of the time after collective nouns and 5.9% of the time after container nouns.

For RT analyses, we excluded RT below 150 ms and above 1000 ms (51 trials were excluded on this basis, less than 1% of the data). Figure 3 (right panel) plots the mean RT for the selection of singular/plural verbs in the different conditions. The plot is only indicative, since few datapoints are included in the bars for plural verbs following pseudopartitives (except for quantifier heads) and singular controls, and for singular verbs following plural controls.

One trial in the singular control condition was excluded due to a typo in the sentence. As done in Experiment 1, we excluded plural controls from the analyses. To analyze the proportion of plural verb forms selected (compared to singular), we ran a first logistic mixed model regression (by means of the glmer function in lmerTest package) including the selected Verb Form (singular or plural) as the dependent variable, Condition as the independent variable (comparing pseudopartitives to singular controls, coded as –0.5 and +0.5 respectively, as done before), random intercepts for participants and items and random slopes for participants (including random slopes for items resulted in convergence failure). A significant difference was revealed, with more plural verbs selected in pseudopartitives than in singular controls (Est. = –2.0729, Std. Err. = 0.5916, z-value = –3.504, p = 0.0005). In a second model, we set contrasts as summarized in Table 1 to compare all the levels of the Head Type variable. The output of the second model is provided in Table 3.

The second model shows that the selection of a plural verb increases gradually from containers to collectives (as revealed by the significant second contrast), and from collectives to containers (as revealed by the significant third contrast).

To analyze RT data, we ran two mixed regression models with RTs (log-transformed) as the dependent variable, Verb Chosen and Condition/Head Type as the independent variables, and their interaction, and random intercepts for participants and items (including random slopes for subjects or items resulted in convergence failure). The first model revealed a significant difference in RTs, showing that decisions in singular controls were faster than in pseudopartitives (1^st contrast: Est = –0.0769, Std. Err. = 0.0336, t-value = –2.291, p = 0.0239). The second model, in which Head Type was considered as one of the predictors setting the sliding difference contrasts as defined above, revealed only a significant difference in the first contrast (Est. = 0.1104, Std. Err. = 0.04079, t-value = 2.707, p-value = 0.0079), and a marginal difference in the second contrast (p = 0.0768), but no other significant results. These results, though, are to be taken with caution due to the small number of datapoints in some of the conditions.

This second study returns two main findings, which align with those reported for Experiment 1. First, pseudopartitive constructions, in which a measurement relation links a singular head with a plural (measured) noun, pattern differently from singular controls in which a singular head noun is modified by a plural intervener. In line with the findings of Experiment 1, the presence of a plural intervener between the sentential subject and the verb does not affect the preference for the singular verb (which is the only grammatical option) when the sentential subject is singular. In the case of pseudopartitives, instead, although a singular verb is selected in most cases, the degree to which this choice is made depends on the type of head of the pseudopartitive: the selection of a plural verb is significantly higher when the head noun is a quantifier, and it is the lowest when the head noun is a container, with collectives again lying inbetween. The analysis of RTs also confirms that participants are faster in the case of singular controls compared to pseudopartitives. This is compatible with the idea that only one parse is available for singular controls, while alternative parses can be generated for pseudopartitives, with a different relative probability depending on the head noun.

By means of the different experimental technique employed in Experiment 2, we corroborate previous findings in showing that (i) pseudopartitives are more subject to variability in verb preference compared to other constructions that are superficially similar but for which grammatical rules of subject-verb agreement apply; (ii) speakers’ preferences for a singular or plural verb after a pseudopartitive vary depending on the type of head.

No participant was excluded due to inaccurate answers to the comprehension questions (accuracy was above 92% overall; the lowest individual participants’ accuracy was 85%). Eye data were first processed in Dataviewer (SR Research) performing a 4-stage cleaning process as described in EyeLink Data Viewer User’s Manual Document (Version 3.1.97) to remove short fixations (shorter than 80 ms in step 1, shorter than 40 ms in step 2 and shorter than 140 ms in steps 3 and 4), as well as fixation larger than 800 ms. We also removed all fixations falling outside any of the interest areas (5730 overall); individual eye fixations that were less than 80 ms in duration and within one grade of a previous or subsequent fixation were incorporated into their neighboring fixation (969 overall). For regression analyses, trials were excluded if a word was skipped during first pass reading of the verb region. On this basis, 260 trials were removed, corresponding to less than 1% of all datapoints. For first pass reading, trials were excluded if a later region was visited before the first fixation entered the verb region in the sentence, leading to an exclusion of 82 trials. No participant was excluded due to data loss above the threshold set at 20%. In addition, 4 trials were excluded due to a typo in the sentence.

For each sentence, five regions of interest were created: R1 corresponds to the lead-in adverbial or propositional phrase; R2 corresponds to the first NP (head) of the sentential subject; R3 corresponds to the second NP (embedded modifier); R4 corresponds to the verb (aux + past participle) and R5 to the final phrase. To exemplify, sentence (21) above was segmented as in Table 4:

We considered four standard eye movement measures for analyses. First pass reading time, which is the sum of all eye fixation durations on the region during participants’ first reading of the sentence, i.e., before leaving the region to the left or right. The variable that was used for this analysis is IA FIRST RUN DWELL TIME in the interest area report generated by Dataviewer, removing all trials in which later regions were visited before entering the verb region, as detailed above. Go past time, which is the sum of all fixation durations beginning with the first on the region, also including any regressive re-reading of earlier material and any re-reading of the critical region itself. The variable that was used for this analysis is IA REGRESSION PATH DURATION. Total reading time, which is the sum of the duration across all fixations that fell in the current interest area during the trial. The variable that was used for this analysis is IA DWELL TIME. Regression probability, which is the probability of a regressive eye movement during first pass reading, i.e., the probability of a leftward saccade out of the region, rather than a forward saccade. The variable that was used for this analysis is IA REGRESSION OUT.

Given that the sentences differed in length, all analyses of reading times include the length of the strings in characters as a covariate in the analyses.

All analyses focus on R4, which includes the number marked auxiliary be/have and the past participle. This corresponds to the region in which a possible mismatch in subject-verb agreement surfaces. The different eye tracking measures recorded in this region for controls and pseudopartitives are plotted in Figure 4.

First, controls pattern as expected: signals of disruption, across all measures, are detected when a plural verb follows a singular subject and when a singular verb follows a plural subject, despite the presence of a plural/singular modifier. Similarly, some evidence for disruption is visible in some of the measures when a plural verb follows a pseudopartitive. This confirms an overall preference for singular agreement in the case of pseudopartitives, although the size and the pattern of the effects in the case of pseudopartitives followed by a plural verb seem to differ from the pattern observed in the case of singular sentences. Also, a difference seems to emerge among different types of pseudopartitives.

We explored these effects in statistical analyses. We ran two models for each of the eye tracking measures specified above, setting the contrasts for the critical variables (Condition/Head Type and Verb Number) as detailed for Experiment 1. All p-values reported in the outputs of the models are based on the Satterthwaite approximation to the denominator degrees of freedom, as implemented in the lmerTest package (Kuznetsova et al., 2017; Luke, 2017). As in the previous experiments, we modeled contrasts to compare singular controls to pseudopartitives in the first model, and sliding different contrasts in the second model, as detailed in 2.2.

Statistical analyses of the reading time measures (first pass, go past, and total reading times) and regression probability were carried out using mixed-effects linear and logistic regression models, respectively. In all models, the fixed effect’s structure includes Length of the region (centered), Trial order, Condition/Head Type, and Verb Number, as well as the interaction of Condition/Head Type and Verb Number. Random intercepts for participants and items were always included (adding random slopes resulted in a failure of convergence in some of the models; for consistency, we never included them).

Apart from the significant effects of region Length and Verb Number (and, in some cases, of Trial order), which were expected and consistent across most measures, the most theoretically relevant results are discussed below.

With respect to the first model, which compares singular controls to pseudopartitives, the most relevant results are the following (the output of all models is available in the R script in the OSF repository). Total time measures reveal a significant interaction of Condition and Verb Number (Est. = –0.1109, Std. Err. = 0.0296, t-value = –3.748, p = 0.0002), also evident in Figure 4: the overall time spent on the verb region is much greater for singular controls when the verb is plural, compared to when it is singular, while the difference between reading measures for singular and plural verbs is reduced in pseudopartitives. The probability of regressions is significantly higher in singular controls compared to pseudopartitives (Est. = 0.5409, Std. Err. = 0.1461, z-value = 3.701, p = 0.0002), independently of verb number. In addition, the interaction with verb number suggests that the disruption observed in the case of plural verbs tends to be significantly greater for singular controls compared to pseudopartitives (p = 0.073).

The pattern observed in the second model reveals other significant differences among different types of sentences, albeit some interactions and comparisons did not reach full significance. The output of the second model for all measures is reported in Table 5.

The first contrast is between singular controls and the container-headed pseudopartitives. Across the different measures, there is some indication of a difference between the two sentence types: in first pass and regression probability, there is a tendency towards a difference, with shorter reading times, but a higher probability of regressions, in singular controls; in total reading time, the interaction with Verb Number is fully significant, consistent with bigger disruption effects in the case of singular controls when the verb is plural, compared to when it is singular: such a difference is reduced in the case of container-headed pseudopartitives. The second contrast is between container-headed and collective-headed pseudopartitives. Our prediction was that the availability of a measure-phrase interpretation is bigger for the latter compared to the former. A difference clearly emerges in regression and go past measures, and a tendency towards significance is revealed in total reading time: these show that the time spent in the verb region is overall shorter for collective heads than container heads, and the probability of regression is lower in the former compared to the latter. The third contrast is between collective-headed and quantifier-headed pseudopartitives. Our prediction was that the availability of a measure phrase interpretation is maximal in the latter, consistent with higher acceptability of plural agreement. Although this contrast is never significant across any of the measures, the interaction with verb number is fully significant in regressions: in the case of quantifier heads, the plural verb triggers less regressive eye movements, compared to singular verbs, while an opposite trend is found for collective heads. This result is compatible with the hypothesis that the availability of a measure phrase interpretation, compatible with plural agreement, is higher in the case of quantifier-headed pseudopartitives.

This third study provides online measures of the processing of pseudopartitive constructions during reading. What emerges, once again, is that these structures display a different behavior from superficially similar, but syntactically different, constructions. This even though, in all cases, a singular head noun is followed by a plural intervener and the singular verb is the putatively grammatical option.

First, the results suggest that readers experience greater difficulty when encountering a plural verb after a singular subject followed by a plural modifier, compared to pseudopartitive constructions. In the latter, the disruption in reading when encountering a plural verb is significantly reduced. This suggests that pseudopartitive constructions are parsed differently from singular controls. Second, although there is a putative overall preference for singular agreement in the case of pseudopartitives, some difference is again found across different head types when followed by a plural verb: reduced disruption effects emerged in collective compared to container-headed pseudopartitives, and reverse effects in the case of quantifier heads, for which the accessibility of the measure interpretation is maximal.

In general, the eye tracking data are consistent with the optionality of singular verb agreement in the case of pseudopartitives, for which two alternative structures are available in the grammar. As for the fact that some of the contrasts did not return fully significant effects, this might be due to the fact that, as discussed above, disruption effects tend to be milder, if not absent, when variable agreement is allowed in the grammar, compared to ungrammatical cases. This is what emerges in our study as well: greater disruption effects are revealed for singular controls followed by a plural verb: only in this case does the grammar allow for a single parse in which the verb must be singular, despite the plural intervener.