Glossa Psycholinguistics

1. Introduction

Agreement has been a major focus of research in psycholinguistics, because it can illumiate how linguistic structure interacts with other aspects of cognition, such as memory access in comprehension (e.g. Dillon et al., 2013; Pearlmutter et al., 1999; Wagers et al., 2009) or planning processes in production (e.g. Bock et al., 1999; Bock et al., 2004; Eberhard et al., 2005; Franck et al., 2006; Haskell & MacDonald, 2003).

Agreement has been characterised as “the systematic covariance between a semantic or formal property of one element and a formal property of another” (Steele; 1978, p. 610), highlighting the fact that it can also be seen as part of the interface between syntax and interpretation. In this paper, we report two studies on English number agreement that investigate this interface. In the sentence The boys laugh, the (plural) number feature of the subject (the boys) matches with that of the verb (laugh). Plurality is a formal property both of the verb laugh and of the subject the boys, because these two elements are both morphosyntactically plural. It is also a semantic property of the boys, because this phrase refers to more than one individual. For nouns like boy/boys, the formal and semantic number properties align: if the noun is morphosyntactically singular, its referent is also semantically singular, and if it is morphosyntactically plural, its referent is also semantically plural. However, for collective nouns like committee, there can be a lack of alignment between morphosyntactic and semantic number: the noun committee is formally singular, but it may denote an assemblage of multiple individuals, making the referent semantically plural. If such a noun participates in subject-verb agreement, or anaphor-antecedent agreement, then the agreeing element may match either the semantic or the syntactic feature, as shown in (1):¹

In (1a,c), the agreeing element (was, itself) takes the singular form, agreeing with the morphosyntactic features of committee, while in (1b,d), plural agreement is used. We will refer to the former type of agreement as syntactic agreement and the latter as semantic agreement. We use these terms following convention,² although in the case of collective nouns, the singular agreement option could be also be argued to reflect the semantic feature, namely that committee denotes a single group. Note, however, that the value of this semantic feature correlates with morpho-syntax: multiple groups require plural marking (committees). In contrast, the fact that a committee is composed of multiple individuals is not reflected in the morphology of the noun, justifying the use of the term “semantic agreement” for cases where plural elements agree with committee in terms of this feature. In general, semantic agreement is widely attested cross-linguistically, and it can affect a variety of different linguistic features, including but not limited to number and gender (Corbett, 1979, 1983, 2000).

According to the Agreement Hierarchy (Corbett, 1979), the distribution of semantic agreement varies by dependency type cross-linguistically, with elements that are structurally closer to the agreement controller being more biased towards participating in syntactic agreement (e.g. determiner-noun agreement), and elements more structurally distant to the controller having a relatively greater possibility for semantic agreement (e.g. anaphor-antecedent or pronoun-antecedent agreement). Moreover, the hierarchy is intended to be implicational, such that if a given dependency type allows semantic agreement for a particular language, then all dependency types that are further towards the semantic side of the hierarchy will also allow semantic agreement in that language. The agreement hierarchy is illustrated in Table 1, along with examples involving English number agreement:

As shown in Table 1, determiner-noun agreement is an instance of an Attributive dependency type, and therefore occupies the syntactic end of the hierarchy. This dependency type does not allow semantic agreement in English (i.e. these committee is ungrammatical). In contrast, verb-subject and anaphor-antecedent dependencies, being instances of Predicate and Personal Pronoun dependencies respectively, do allow semantic agreement, although psycholinguistic and corpus evidence suggests that there are processing and distributional differences between these two types of dependencies, aligning with their positions on the hierarchy. One source of evidence for this comes from elicited speech production, where studies have shown that collective noun controllers lead to higher rates of plural agreement for anaphor-antecedent dependencies than they do for verb-subject dependencies (Bock et al., 2004; Bock et al., 1999). A similar asymmetry has been found in naturally occurring corpus data (Levin, 2001), and overall, the direction of this effect is what would be expected on the basis of the agreement hierarchy, with the higher rates of semantic (i.e. plural) agreement for the anaphor-antecedent dependencies aligning with these dependencies appearing further towards the semantic end of the hierarchy, relative to verb-subject dependencies. A phenomenon that may be related to this is the fact that quantificational phrases like everyone can antecede plural anaphors (e.g. Everyone praised themselves), while plural agreement on the verb is not possible, even where plural anaphor agreement is present (i.e. Everyone were praising themselves is ungrammatical).

As well as the production evidence discussed by Bock et al. (1999) and Bock et al. (2004), there is also comprehension-based evidence suggestive of the idea that the processing of verb-subject and anaphor-antecedent dependencies aligns with the agreement hierarchy. In a series of eye-tracking experiments reported by Kreiner et al. (2013), there was evidence for temporary processing difficulty for a plural marked verb in the context of a subject headed by a collective noun (e.g. the family… want), relative to appropriate control conditions (Kreiner et al., 2013, Experiments 2 and 3). However, no such evidence was evident for plural anaphors referring to a singular collective noun phrase (e.g. the family … themselves), even though reference to a non-collective singular subject did cause statistically detectable difficulty (e.g. The schoolgirl … themselves). Although this evidence is somewhat circumstantial, the findings suggest consistency with the agreement hierarchy, given that there was evidence for verb-subject dependencies to be sensitive to morpho-syntactic agreement, but the evidence was not found for anaphor-antecedent dependencies.

Bock et al. (1999) and Bock et al. (2004) suggest that the effects that they found in their production studies can be explained in terms of a sequential model of language production (see also Levelt, 1989), whereby the process of selecting a singular or plural prounoun occurs at an earlier stage of production where conceptual information (e.g. about the semantic plurality of collectives) is available, while the process of appending an agreement suffix on the verb occurs at a later stage, where grammatical information is more relevant. However, the results of Kreiner et al. (2013) suggest that the biases of the agreement hierarchy may instead be a more general feature of both production and comprehension (see Kreiner et al., 2013, for a further discussion of this point).

In this paper, we consider cases where the same agreement controller participates in two number agreement dependencies within the same sentence, as in (2) below, where the government participates in one dependency with avoid/avoids, and another one with themselves/itself:

In (2a,b), both the anaphor and the verb match in number—in (2a) both show plural (i.e. semantic) agreement, and in (2b), both show singular (i.e. syntactic) agreement. However, in (2c,d) these two elements mismatch in number. It has been claimed that there is an acceptability contrast between examples (2c) and (2d). Examples like (2c), where the verb shows plural agreement and the anaphor shows singular agreement are claimed to be less acceptable than (2d), which has the reverse configuration (Huddleston & Pullum, 2002; P. Smith, 2015, 2017), although some authors claim that both (2c) and (2d) are ungrammatical, as we will discuss in the General Discussion. Examples analogous to (2c) are also less frequently attested in corpora than those analogous to (2d) (Levin, 2001). In this paper, we will refer to the pattern of acceptability illustrated in (2a–d) as the “double mismatch effect”.

To our knowledge, the only study that has investigated double mismatches with collective nouns using experimental methods is Kreiner et al. (2013). Experiment 3 of Kreiner et al. (2013) is an eye-tracking experiment that used sentences like those in (2) among other control conditions. However, analysis of eye-movement measures in the reflexive and following region did not show significant effects indicating a mismatch cost between verb and reflexive number, whether the verb was singular and the reflexive was plural (2d), or the reverse configuration (2c). On the other hand, an ERP study reported by Molinaro et al. (2008) showed evidence of processing difficulty for a number agreement mismatch between verb and reflexive, although their study did not use collective nouns. They used sentences like (3a–d):

In ERPs measured at the reflexive (herself/themselves), Molinaro et al. (2008) reported a P600 effect for conditions (3c) and (3d), where the reflexive and verb mismatched in number, relative to the fully grammatical and matching control condition (3a). This suggests, contrary to the null result of Kreiner et al. (2013), that mismatches between reflexive and verb number can indeed elict measurable processing difficulty, at least given certain experimental methods. However, given that Molinaro et al.’s study did not use collective nouns, it is not fully relevant to the present paper. Moreover, although Molinaro et al. (2008) counterbalanced their experimental stimuli among those with a singular and a plural subject, they did not report separate analyses for the two resulting mismatch patterns (i.e. those where the verb is singular and the reflexive is plural, vs. those in the opposite configuration).

To summarize the discussion above, an intuitive acceptability contrast has been proposed between two different patterns of mismatch between verb and anaphor number in relation to a collective noun phrase (see (2c,d) above). Experimental evidence for this asymmetry has not been provided to date, although a related study has shown that mismatches between anaphor and verb number in general can lead to detectable processing difficulty (Molinaro et al., 2008) with non-collective nouns. In the following paragraphs, we consider two different ways of stating the descriptive generalization to capture the double mismatch effect. One way of stating this generalization is in terms of the positions of the two agreement targets on the agreement hierarchy:

In addition to the sentences where the agreements match, we also expect that a mismatch between the two targets can arise if it is the element to the right on the hierarchy that shows semantic agreement, and the element to the left that shows morphological agreement. (P. Smith, 2015, p. 203)

According to this form of generalization, the relative acceptability of (2d) is due to the fact that the verb avoids shows morpho-syntactic agreement, and is further towards the left (i.e. syntactic agreement) side of the agreement hierarchy, while the anaphor themselves shows semantic agreement, and is further towards the right (i.e. semantic agreement) side of the hierarchy. Conversely, (2c) is less acceptable because the two agreement targets are in the opposite configuration relative to the agreement hierarchy. In other words, the idea is that mismatching agreement is acceptable in the case that the two agreement targets show forms of agreement that coincide with their relative positions on the agreement hierarchy.

However, the relative positions of agreement targets on the agreement hierarchy tend to correlate with both structural and linear distance to the agreement controller. For example, in terms of English number agreement, verbs tend to be closer to their subjects than pronouns or anaphors are to their antecedents (Corbett, 1979; Levin, 2001). Thus, another way to state the generalization is in terms of linear or structural distance:

We predict that any such constraint will take the form of disallowing a combination of semantic agreement of the nearer element and syntactic agreement of the further. (Corbett, 1979, p. 221)

Based on this observation, we could also formulate an alternative generalization to describe the double mismatch effect, based on the order of agreement targets. According to this generalization, the relative acceptability of (2d) would be due to the fact that the initial element (i.e. the verb) agreeing with the controller shows syntactic agreement, while the subsequent agreeing element (the anaphor) shows semantic agreement, while in (2c), the reverse order of semantic followed by syntactic agreement leads to unacceptability. Note that, according to this idea, it is only the order of singular vs. plural agreement that plays a role in explaining the double mismatch effect, and the parts of speech of the two agreement targets play no role.

These two ways of stating the descriptive generalization suggest two hypotheses regarding the acceptability of double mismatches. According to the agreement hierarchy-based account, in the case of a verb and an anaphor agreeing with the same collective noun phrase target, acceptability will be higher if the verb shows singular agreement and the anaphor shows plural agreement, relative to cases where the verb shows plural agreement and the anaphor shows singular agreement. This is because syntactic (singular) agreement of the verb and the semantic (plural) agreement of the anaphor align with the positions of the relevant two dependency types on the agreement hierarchy. In contrast, according to the order-based account, where two agreement targets agree with the same collective noun phrase that precedes both targets, acceptability will be higher if the first agreement target shows singular (i.e. syntactic) agreement and the second agreement target shows plural (i.e. semantic) agreement, relative to cases where the first target shows plural agreement and the second shows singular agreement. We will return to discuss the order-based account in more detail in Section 4.

Apart from the agreement-hierarchy based account and the order-based account, there is a third possible explanation of the pattern of acceptability in (2a–d), based on animacy or humanness, rather than number. According to this explanation, when a committee-type noun participates in plural agreement, it is usually interpreted as denoting a collection of humans, while in singular agreement the canonical interpretation would be as an inanimate entity or institution. As itself requires a non-human antecedent, the relative unacceptability of (2c) may be due to a clash in animacy requirements: avoid, being plural, biases the interpretation of committee towards a group of humans, while itself requires committee to be inanimate. In contrast, in the relatively acceptable (2d), although avoids biases committee towards an inanimate interpretation (because it is singular), there is no animacy clash because themselves can be used with either an animate or an inanimate antecedent (e.g. The authors/articles contradicted themselves). We will return to discuss this animacy-based account in 3.6, but for present purposes, it is worth noting that its predictions align completely with those of the agreement-hieararchy based account, for the English collective noun examples that we discuss here.

Below, we report two acceptability judgment experiments that are designed to tease apart the agreement hierarchy-based and the order-based accounts. In Experiment 1, we confirm the intuitive acceptability pattern illustrated in (2a–d). In Experiment 2, we reverse the relative linear order of the agreeing verb and anaphor, in order to compare the two theoretical accounts.

2. Experiment 1

2.5 Procedure

The experimental items were combined with 60 fillers, of which half were ungrammatical,⁴ and half grammatical. The ungrammatical sentences included both reflexive-antecedent and verb subject number mismatches, as well as other miscellaneous ill-formed sentences. The experiment was implemented on the PCIbex platform (Zehr & Schwartz, 2018), and took place on-line, via a web link from the prolific.co site. The PCIbex software automatically distributed experimental items into a Latin Square, such that each participant was exposed to only one condition from a given item, but across participants, observations were obtained from all conditions of all items, and each participant saw six examples from each condition. On each trial, the participant was asked to judge the acceptability of the given sentence on a 1–7 scale (7 = most acceptable), by clicking the appropriate button with the mouse, or selecting the numerical key on the keyboard (see Figure 1). One third of sentences were followed by Yes/No questions, which the participant had to answer by clicking the appropriate answer. After each sentence (or after each sentence-question pair, for stimuli that included questions), a screen was presented for 1000 msec, containing only the string “…” in the top center of the screen.

Figure 1

An example trial.

Following the experiment, the participant was routed back to the prolific.co website, and later received payment. The average time for each participant to complete the task was 14 minutes.

2.7 Results

Across all experimental and filler stimuli that included comprehension questions, mean comprehension accuracy was 94% (SE = .88).

Ungrammatical filler sentences received a mean rating of 2.35 (SE = 0.11), and the mean for grammatical filler sentences was 6.60 (SE = 0.05).

Means ratings per condition are displayed in Figure 2.

Figure 2

Means ratings (and standard errors) for Experiment 1.

Given that the factor of verb matching had three levels, and the matching verb was the reference level, the model output included (a) comparisons involving the matching and mismatching verb conditions, excluding the unmarked verb conditions (i.e. 4b,c,e,f), and (b) comparisons involving the matching and unmarked verb conditions, excluding the mismatching verb conditions (i.e. 4a,b,d,e). We will call (a) the main contrast set, and we will call (b) the baseline contrast set. For clarity of exposition, we will report the statistical analysis of results separately for the main and baseline contrasts. The main contrast set is used to test for the critical interaction between reflexive number and verb matching, which is predicted by both the agreement hierarchy-based account and the order-based account.

The baseline contrast set is used to test for any differences between the verb match and unmarked verb conditions, and in particular, any differences in the size of such an effect between the singular and plural reflexives (see discussion above). Such a difference in effect sizes would result in an interaction between verb (unmarked vs. match) and reflexive number, and would complicate the overall interpretation of the results.

Main contrast set: Bayesian model results for the main contrast set are summarized in Table 2:

Table 2

MAIN CONTRAST SET: Bayesian model estimates for effects involving Verb Match and Mismatch conditions (i.e. excluding unmarked verb conditions). Output shows relevant paired comparisons (X vs. Y) and interaction (X × Y).

Effect	Estimate	95% CrI	P(b < 0)
Mismatching vs. Matching verb (at plural reflexive)	–1.10	[–1.47, –0.72]	.999
Mismatching vs. Matching verb (at sing. reflexive)	–2.53	[–2.96, –2.09]	.999
Verb matching × Refl. number	–1.50	[–2.10, –0.91]	.999

The most important result from the main contrast set is the interaction of verb matching by reflexive number. There was strong evidence for this interaction (see Table 2). This interaction showed that the rating penalty for the mismatching verb relative to the matching verb was greater for the singular reflexives (4f) vs. (4e) than for the plural reflexives (4c) vs. (4b). This therefore provides strong support for the prediction of both the agreement hierarchy-based account and the order account that acceptability would be specifically reduced for the mismatching verb with the singular reflexive. Despite the difference in the size of this verb-mismatch cost, pairwise comparisons showed strong evidence for the mismatch cost at both levels of the reflexive number factor (see Table 2).

Baseline contrast set: Bayesian model results for the baseline contrast set are reported in Table 3:

Table 3

BASELINE CONTRAST SET: Bayesian model estimates for effects involving Verb Match and Unmarked conditions (i.e. excluding mismatching verb conditions). Output shows relevant paired comparisons (X vs. Y) and interaction (X × Y).

Effect	Estimate	95% CrI	P(b < 0)
Matching vs. Unmarked verb (at plural reflexive)	–0.08	[–0.42,0.26]	.68
Matching vs. Unmarked verb (at sing. reflexive)	–0.04	[–0.36,0.29]	.61
Verb (unmarked vs. match) × Refl. number	0.06	[–0.36, 0.50]	.39

In the baseline contrast set, we use the unmarked verbs as a baseline to test for the cost of a plural marked verb, relative to a singular marked verb in the same context. If a singular verb reduces acceptability, then in the plural reflexive conditions, the matching verb condition (4b) should have lower ratings than the unmarked verb condition (4a), while no such difference should be observed between the matching and unmarked verbs in the singular reflexive conditions (4d) and (4e), and this should lead to an interaction between verb (unmarked vs. match) and reflexive number. As seen in Table 3, there was no clear evidence of such an interaction, nor was there any evidence for a difference between the verb match and unmarked verb condition at either of the two levels of the reflexive number factor. Thus, the main interaction that we report in Table 2 above cannot be explained by an acceptability cost for a verb with singular morphology.

3. Experiment 2

3.5 Results

Across all experimental and filler stimuli that included comprehension questions, mean comprehension accuracy was 92% (SE = .69).

Ungrammatical filler sentences received a mean rating of 2.55 (SE = 0.12), and the mean for grammatical filler sentences was 6.54 (SE = 0.05).

Mean ratings per condition are displayed in Figure 3.

Figure 3

Means ratings (and standard errors) for Experiment 2.

Bayesian model output results are shown in Table 4:

Table 4

Bayesian model output for Experiment 2, showing (a) estimates for main effects and interactions, (b) credible intervals, and (c) the probability that the effect is greater than zero (for positive coefficients), or less than zero (for negative coefficients).

Effect	Estimate	95% Credible interval	P(b>/<0)
Reflexive number	0.24	[0.03, 0.45]	.99
Structure	–0.19	[–0.37, –0.02]	>.999
Verb matching	–0.79	[–1.04, –0.54]	>.999
Verb × Reflexive	2.06	[1.47, 2.67]	>.999
Verb × Structure	0.42	[0.09, 0.75]	.99
Reflexive × Structure	0.16	[–0.20, 0.50]	.8
Verb × Reflexive × Structure	–1.60	[–2.42, –0.80]	>.999

From the point of view of testing the theoretical predictions, the salient result was the interaction of verb matching by reflexive number, for which there was strong evidence. This interaction had a positive sign, meaning that the acceptability penalty for mismatching (vs. matching) verbs was greater for the plural reflexive than for the singular reflexive, consistent with the order-based account, and not with the agreement hierarchy-based account. We will use further contrast analyses to explore this interaction in more detail below. The analysis also showed evidence for a main effect of reflexive number (with overall higher ratings for singular reflexives than for plural reflexives), a main effect of structure (coordination structures had overall higher ratings than relative clause structures), and a main effect of verb match (matching verbs had higher ratings than mismatching verbs). As well as these effects, there was evidence for a two-way interaction of verb-matching by structure, indicating that the penalty for a mismatching verb was greater for the coordination structure than for the relative clause structure. There was no clear evidence for a two-way interaction of reflexive number by structure. Finally, there was evidence for a three-way interaction among the experimental factors, which can be interpreted as showing that the effect size of the two-way interaction of verb-matching by reflexive number was greater for the coordination structure than for the relative clause structure (see Figure 3).

The evidence for the three-way interaction justified separate analyses for the coordination and relative clause structures, as it was important to establish that the predicted interaction of verb-matching by reflexive number could be generalized across sentence types, even if the size of this effect may differ between them. The separate analyses indeed showed clear evidence of the two-way interaction for both sentence types (Coordination: b = 2.94, CrI = [2.16,3.78], P(b>0) > .999; Relative clause: b = 1.27, CrI = [0.54,1.98], P(b>0) > .999). These interactions were then further analyzed using contrast models that tested the verb-matching effect at each level of the reflexive number variable, within each sentence type. For both sentence types, there was a clear degradation in acceptability for mismatching verbs (relative to matching verbs) when the reflexive was plural (Coordination: b = 2.51, CrI = [1.97,3.06], P(b>0) > .999; Relative clause: b = 1.19, CrI = [0.81,1.57], P(b>0) > .999). However, there was no such evidence for such a verb-mismatch cost at the singular reflexive level for the relative clause structure, and for the coordination structure, if anything there was weak evidence for a cost for the matching verb (Coordination: b = –0.44, CrI = [–0.91,0.05], P(b<0) = .97; Relative Clause: b = –0.05, CrI = [–0.51,0.44], P(b<0) = .58).

4. General Discussion

Experiment 1 provided a confirmation of the double mismatch effect using controlled experimental methods, in cases where the verb precedes the anaphor. This result was compatible with either the agreement hierarchy-based account or the order-based account. Experiment 2 reversed the linear order of verb and reflexive, and showed a pattern that is more compatible with the order-based account.

Before we go on to discuss the theoretical implications, there is one aspect of the results that deserves comment, which is the fact that the overall acceptability penalty for a mismatch between verbs and anaphors was greater in Experiment 1 than it was in Experiment 2. This can be seen from the fact that in Experiment 1, both singular and plural reflexives showed strong evidence for an acceptability penalty when the number mismatched that of their respective verbs, even though the mismatch penalty was larger for singular reflexives than for plural reflexives. In contrast, in Experiment 2, there was a detectable mismatch penalty only for the plural reflexives, and no evidence of a penalty for the singular reflexives. Moreover, for the dispreferred mismatching condition (singular reflexives in Experiment 1 and plural reflexives in Experiment 2), the numerical size of the mismatch effect was larger in Experiment 1 than it was in Experiment 2 (compare Figures 2 and 3). Explanations for this difference must be speculative in the absence of further empirical studies. One possibility is that, due to the more complex sentences of Experiment 2, people engaged in more shallow processing in that experiment relative to Experiment 1, and were thus less sensitive to mismatches overall. This shallow processing explanation would be supported by a lower rate of comprehension accuracy in Experiment 2 relative to Experiment 1. However, although comprehension accuracy was numerically slightly lower in Experiment 2, the difference was small, and rates were relatively high in both experiments. This was the case even when considering only experimental stimuli (i.e. ignoring fillers), where the mean accuracy rate was 94% (SE = 0.89%) for Experiment 1 and 91% (SE = 0.88) for Experiment 2. Thus, even if the participants of Experiment 2 employed shallower processing than those of Experiment 1, this was not to such a degree as to have a large effect on comprehension accuracy.

It is possible that the difference in the size of the mismatch effect in Experiment 2 relative to Experiment 1 is related to the difference in structural distance between the verb and the reflexive in Experiment 1 relative to Experiment 2. In Experiment 1, the anaphor appeared in the verb phrase local to the number-marked auxiliary verb. In Experiment 2, however, the anaphor appeared in a separate verb phrase from the one that was local to the number-marked auxiliary. It may be the case that number mismatches between two agreement targets sharing a controller are less acceptable when the two targets are local to each other than when they are not.

We now move on to consider the interpretation of the main results of this paper. The first question is whether the pattern of results that we report could be captured at the competence level. A competence-based explanation would mean that certain types of agreement mismatch could be generated by the competence grammar while other types could not, and that this distinction would be reflected in acceptability ratings. In order to consider this question, we will discuss some of the proposals that have been made about collective noun agreement in the theoretical linguistics literature. We will see that the theories can be divided into those that allow agreement mismatches of the type that we examine in this paper and those that do not. Of those that allow agreement mismatches, none of the proposals that we discuss can account for the full range of acceptability contrasts reported in this paper between different types of mismatches. We then move on to sketch a tentative processing-based explanation based on incremental comprehension.

In Head-Driven Phrase Structure Grammar, henceforth HPSG (Pollard & Sag, 1994), agreement features, including those for number, are attributes of referential indices, described by the authors as “abstract objects that function in discourse to keep track of the entities that are being talked about” (Pollard & Sag, 1994, p. 67). The object denoted by a collective noun like government may be assigned either a singular or a plural index, depending on whether the referent is interpreted as a non-aggregate entity, or as an aggregate of multiple entities, and this leads to singular or plural agreement respectively. If this aspect of the interpretation changes, a referent may be assigned a new index, so that, for example, the index assigned to the entity denoted by government may change from singular to plural. This can account for cases where the number agreement changes from one sentence to the next (e.g. The government was unpopular. They failed to produce good policies.) However, Pollard and Sag (1994) predict that the types of within-sentence double mismatch that we consider in this paper (e.g. (4c,f)) should be ungrammatical. Consider these examples repeated below:

In HPSG (Pollard & Sag, 1994), an anaphor like themselves in (4c), or itself in (4f) is co-indexed with the antecedent the government, via structure sharing. As indices bear number features, this index should therefore be plural in (4c) and singular in (4f). However, the process of subject-verb agreement requires that the same index be singular in (4c) and plural in (4f), resulting in a clash of features and thus ungrammaticality in both cases. Thus, HPSG does not allow either type of mismatch considered here to be generated by the grammar. The proposal of Pollard and Sag (1994) therefore does not capture the critical constrasts that we report in this paper, simply because no type of mismatch is generated by the competence grammar.

The second account that we consider is a proposal by den Dikken (2001). According to this proposal, a phrase like the government may be headed by a null plural pronominal element. When this pronominal element is present, it participates in dependencies with agreement targets, such as verbs or anaphors, resulting in plural agreement. In contrast, when the pronominal element is not present, agreement takes place directly with the (morphologically singular) collective phrase, and singular agreement is the result. Thus, if there are two agreement targets that form dependencies with a single collective controller within the same sentence, then agreement may either be singular (pronominal absent) or plural (pronominal present), but not both. Thus, den Dikken (2001) predicts both mismatching examples (4c) and (4f) to be ungrammatical, so again, this proposal does not allow for the acceptability contrast between these two to be captured at the competence level.

We now consider theoretical approaches which allow for double mismatches to be generated within the competence grammar. Sauerland and Elbourne (2002) achieve this by assuming that number agreement in English involves two different categories of features, either of which can take a singular or a plural value. The number feature reflects whether there is one entity or more than one entity under discussion, and the singular or plural value for this feature will correlate with the morpho-syntactic marking on the noun. For example, government corresponds to a single entity and governments to multiple such entities. The mereology feature reflects whether or not the entity is being interpreted as consisting of more than one member, and the value of this feature is not reflected in the morphology of the noun. For example, government may take either a singular or a plural value for the mereology feature, depending on its interpretation. Double mismatches are then handled by allowing, for example, verb-subject agreement to occur via the number feature and anaphor-antecedent agreement to occur via the mereology feature, or vice versa. Thus, Sauerland and Elbourne’s approach allows both (4c) and (4f) to be generated by the competence grammar. Therefore, the proposal does not capture the acceptability difference between (4c) and (4f) at the competence level.

The final theoretical account that we consider is the proposal by P. Smith (2015, 2017). According to Smith’s proposal, a number feature may be split into two halves, of which one (which Smith calls iF) is interpreted by semantics, and the other (called uF) is interpreted by morphology. A collective noun like government may include a singular value for uF and a plural value for iF. Double mismatches are then handled by different agreement targets interacting with different parts of the feature, bearing different values, for example, a verb agreeing with the uF and an anaphor with iF. Unlike the theoretical accounts mentioned above, Smith’s proposal does predict a difference in acceptability between (4c) and (4f), with (4c) being predicted to be more acceptable, as confirmed by the results of Experiment 1 above. Smith’s explanation of this asymmetry is based on the principle of valuation economy, which requires that, if two targets agree with the same controller in the same domain, then either both need to agree via uF or both need to agree via iF. The double mismatch in (4c) is the result of agreement occuring in two different domains, allowing uF to be used in one domain and iF in another, without violating valuation economy. Specifically, the (singular) verb-subject agreement occurs via uF post-syntactically (in the branch to the P(honetic) F(orm) component of grammar, in Smith’s derivational Minimalist-based system), and the (plural) anaphor-antecedent agreement occurs via iF, within the syntactic component. We now explain why the mismatch in (4f) is ruled out in this system. First, it is assumed that anaphor-antecedent agreement has to occur within the syntactic component, to allow the dependency to be interpreted semantically, while verb-subject agreement may occur either within the syntactic component or post-syntactically (i.e. in the PF branch of the grammar). Given that the anaphor-antecedent agreement in (4f) is singular, it must have been derived via uF, and, being an anaphor-antecedent dependency, it must have occurred within the syntactic component of the grammar. However, for the verb-subject agreement to be plural, it must have occurred via iF, but in order to satisfy valuation economy, this agreement cannot take place within the syntactic component, because a uF agreement has already taken place there, involving the same controller. But there is no possibility for this agreement to take place in the PF branch of the grammar, because, according to the theory iFs are visible only to agreement within the syntactic component.

Given the discussion above, it is clear how the proposal of P. Smith (2015, 2017) can be used to explain the acceptability asymmetry between between (4c) and (4f) that we found in Experiment 1. However, the proposal makes the incorrect prediction for Experiment 2, where the linear order of the agreeing verb and anaphor is reversed. This is because the relevant distinctions are based on the types of agreement targets (i.e. verb vs. anaphor), rather than on linear order. In fact P. Smith (2015) discusses two examples that are similar in all relevant respects to our mismatch conditions in Experiment 2, with the following predicted grammaticality judgements (P. Smith, 2015, p. 232):⁹

In P. Smith (2015), (8b) is predicted to have lower acceptability than (8a), for analogous reasons that (4f) is predicted to be less acceptable than (4c). However, the results of Experiment 2 show that in fact, sentences like (8b) are judged to be more acceptable than those like (8a).

To summarize, of the four theoretical proposals that we have discussed so far, none can capture our results in terms of grammaticality at the competence level. How, then, might the results be explained? One way to think about this is to consider how morpho-syntactic and semantic agreement processes might unfold in real time during sentence comprehension. In a study on German collective nouns, Schweppe (2013) showed that people’s preferences for pronoun-antecedent agreement tended to shift from syntactic to semantic agreement as the distance from the agreement controller increased. For example, in her Experiment 2 (self-paced reading), Schweppe (2013) examined sentences like the following:

In the design, Schweppe (2013) manipulated whether the pronoun referring to a collective noun was singular (Es) or plural (Sie). She also manipulated whether or not an intervening clause (Daran hatte sich nichts geändert ‘Nothing had changed’) intervened between the sentence containing the pronoun and the sentence containing the collective noun, resulting in long distance (intervening sentence) and short distance (no intervening sentence) conditions. On the word following the verb in the critical pronoun sentence, pronoun number and distance interacted, such that, for the singular pronoun, reading times were longer in the long distance than the short distance condition, whereas no such difference was found for the plural pronouns (with a numerical trend in the opposite direction). In a production experiment in the same paper, using a “fill in the gap” task, Schweppe (2013) also found that the proportion of singular pronouns (relative to plurals) reduced in the long distance condition relative to the short distance condition. These findings are compatible with corpus results showing that rates of plural agreement for pronouns referring to collective nouns in English increase with distance (Levin, 2001). More generally, this pattern of results may be related to the faster decay of syntactic details of a sentence (such as morpho-syntactic agreement features) relative to the representation of its meaning (Sachs, 1967).

In the present study, across the two experiments, the conditions that led to the lowest relative acceptability rates were those in which an initial semantic (i.e. plural) agreement dependency was followed by a later syntactic (i.e. singular) agreement dependency, with the same agreement controller. This suggests that, once a collective agreement controller has been implicated in a semantic agreement process in incremental processing, it becomes difficult for later processes to access morpho-syntactic agreement for the same agreement controller. Conversely, the move from syntactic to semantic agreement, even within the same sentence, is more natural, and fits with the general pattern established for the semantic agreement preference to increase with distance across sentences (e.g. Schweppe, 2013).

Accounts that might explain why morpho-syntactic agreement becomes less accessible once semantic agreement has taken place will inevitably be speculative pending further research. One possibility is that the discourse representation of the referent of a collective noun like government may often be underspecified in terms of aspects of its interpretation—for example, whether it is interpreted as a single entity or as a collection of individuals. Further, a singular agreement target (e.g. the verb was) may be processed without adding interpretative detail to this underspecified representation, matching the collective noun phrase purely at the morpho-syntactic level. In contrast, according to this proposal, a plural agreement target, being the marked case, would be more likely to lead to an update in the discourse representation, signalling that the referent should be interpreted as a collection of individuals. Once this further specification has been made, there would be a preference for subsequent agreements involving the same agreement controller to match this semantic commitment. This would be consistent with the results reported here, since the relatively acceptable agreement mismatch cases were those in which a singular agreement target was followed by a plural agreement target. The initial singular agreement may have had a minimal effect on the semantic representation of the collective noun phrase, while the later plural agreement may have added extra detail. In contrast, the relatively unacceptable mismatch cases were those in which an initial plural agreement was followed by a subsequent singular agreement. According to this proposal, the initial plural agreement would lead to an update in the discourse representation, which would then clash with the later singular agreement. Assuming that such a clash affects acceptability ratings, this could explain the results.

Alternatively, it may be possible to overlay processing mechanisms on some of the theoretical linguistic accounts that we discussed above. For example, both den Dikken (2001) and Pollard and Sag (1994) require the verb and anaphor to match in number marking, for the types of sentences that we examine in this paper, at least at the competence level. However, if we assume (a) that these competence grammars can be combined with incremental parsers, and (b) that the representation of the collective noun phrase may change during the processing of a single sentence, then mismatches in number could be captured. For example, in the case of den Dikken (2001), whether the committee participates in plural or singular agreement depends on whether or not an empty plural pronominal is adjoined to this phrase, while in the case of Pollard and Sag (1994), this depends on whether the referent is assigned a singular or a plural index. During the left-to-right processing of the sentence, this aspect of the representation could be updated. For example, in the case of den Dikken (2001), the plural empty pronominal could be inserted, presumably reflecting an update in the interpretation of the referent. This would mean that one could observe singular agreement at the first agreement target (e.g. the verb in Experiment 1) if this element is processed before the insertion of the pronominal, and subsequent plural agreement at the second agreement target (e.g. the anaphor in Experiment 1), if it is processed after the insertion, leading to a mismatch in number between verb and anaphor. In order to capture the results of this paper, one would need further assumptions in order to capture the asymmetry; for example, that it is more difficult to delete the empty pronominal than it is to insert it.

Above, we have discussed competence-based proposals, and how they may potentially be combined with performance-based mechanisms. However, instead of seeing competence and performance as two factors that have to be reconciled, another approach would be to consider models that do not cleanly make such a distinction. Self-Organized Sentence Processing (SOSP) models (G. Smith et al., 2018, 2021; Villata & Tabor, 2022) may be suitable for examining agreement mismatches. In these models, small fragments of structure known as treelets combine together to form larger structures, based on graded satisfaction of requirements between head and dependent treelets. G. Smith et al. (2018) show how such models can be applied to the investigation of notional number agreement in pseudo-partitive constructions like (10):

In their simulation of (10), G. Smith et al. (2018) allowed semantic features to influence whether the model would reach a stable state where either the singular box or the plural oranges was represented as the head of the phrase a box of oranges, thus predicting whether is or are would be chosen as the auxiliary.

The types of sentences examined in the current paper are similar in some ways to the pseudo-partitives examined by G. Smith et al. (2018), in that a noun phrase may participate in either singular or plural agreement, depending on semantic features. A successful model of double mismatches would be one where the model can reach a stable state with one dependency showing plural agreement while the other shows singular agreement, and then it would be possible to test the preferences for one type of mismatch over another.

5. Conclusion

We have described two experiments that examine the double mismatch effect in number agreement for morphologically singular collective noun phrases in English. The results show relatively high acceptability for sentences in which a collective noun controls singular agreement with an initial target, followed by plural agreement with a later target, while the opposite configuration results in lower acceptability. Given that both possible orderings of verb and anaphor were represented across the two experiments, this provides support for the order-based account as described above, suggesting that incremental processing will need to play an important role in any theoretical account. Finally, note that we are not claiming that relative distance between controllers and targets can provide a complete explanation for the semantic vs. syntactic agreement biases of the agreement hierarchy, or that the parts of speech of the agreement targets play no role. However, based on the results reported here, we argue that linear order plays a major role in explaining the double mismatch effect.

Competing interests

The author has no competing interests to declare.