The time-course of feature interference in agreement comprehension: Multiple mechanisms and asymmetrical attraction

Darren Tanner; Janet Nicol; Laurel Brehm

doi:10.1016/j.jml.2014.07.003

. Author manuscript; available in PMC: 2015 Oct 1.

Published in final edited form as: J Mem Lang. 2014 Oct 1;76:195–215. doi: 10.1016/j.jml.2014.07.003

The time-course of feature interference in agreement comprehension: Multiple mechanisms and asymmetrical attraction

Darren Tanner ¹, Janet Nicol ², Laurel Brehm ³

PMCID: PMC4170797 NIHMSID: NIHMS616169 PMID: 25258471

Abstract

Attraction interference in language comprehension and production may be as a result of common or different processes. In the present paper, we investigate attraction interference during language comprehension, focusing on the contexts in which interference arises and the time-course of these effects. Using evidence from event-related brain potentials (ERPs) and sentence judgment times, we show that agreement attraction in comprehension is best explained as morphosyntactic interference during memory retrieval. This stands in contrast to attraction as a message-level process involving the representation of the subject NP's number features, which is a strong contributor to attraction in production. We thus argue that the cognitive antecedents of agreement attraction in comprehension are non-identical with those of attraction in production, and moreover, that attraction in comprehension is primarily a consequence of similarity-based interference in cue-based memory retrieval processes. We suggest that mechanisms responsible for attraction during language comprehension are a subset of those involved in language production.

Keywords: agreement, morphosyntactic processing, P600, retrieval interference, language comprehension

Introduction

Successful language processing requires the integration of grammatical information across linguistic units within a sentence. One grammatical dependency that must be computed in many languages is subject-verb agreement. Agreement phenomena are of particular interest because they present some interesting paradoxes for language processing. For example, while native English speakers tend to ignore agreement information in determining participant roles in off-line language comprehension tasks (MacWhinney, Bates, & Kliegl, 1984), violations of subject-verb agreement readily trigger processing difficulty during real-time comprehension (Kreiner, Garrod, & Sturt, 2013; Osterhout & Mobley, 1995; Pearlmutter, Garnsey, & Bock, 1999). Moreover, although English agreement morphology is relatively impoverished and the rules of English agreement are relatively simple, both producers and perceivers are known to make errors in verb number marking. These errors are especially likely to occur in contexts where a sentence's subject noun phrase (NP) contains two nouns that mismatch in number features (e.g., The key to the cabinets…). In sentence production tasks, it has been shown that such cases lead to an increased number of errors in number marking on the verb (*The key to the cabinets are…), where the verb erroneously agrees with the “local” (linearly closer) plural noun (e.g., cabinets; henceforth ‘attractor’ noun) rather than singular head noun of the subject NP (e.g., key).

Importantly, these errors, known as ‘attraction’ errors, can reliably be elicited in laboratory experiments focused on language production (Bock & Cutting, 1992; Bock & Eberhard, 1993; Bock & Miller, 1991; Eberhard, 1999; Gillespie & Pearlmutter, 2011; Hartsuiker, Schriefers, Bock, & Kikstra, 2003; Solomon & Pearlmutter, 2004; Staub, 2009, 2010; Vigliocco, Butterworth, & Garrett, 1996; Vigliocco, Butterworth, & Semenza, 1995; Vigliocco & Nicol, 1998), and later research found that plural attractors can interfere with the normal processing of verb agreement during language comprehension as well (e.g., Kaan, 2002; Nicol, Forster, & Veres, 1997; Pearlmutter et al, 1999; Pearlmutter, 2000; Thornton & MacDonald, 2003; Wagers, Lau, & Phillips, 2009). Since the earlier production-based studies focused on contexts in which errors arise, many comprehension studies also included error conditions (i.e., subject-verb agreement violations), because contrasting the processing of agreement errors in comprehension and production can yield useful information about the mechanisms supporting grammatical dependencies like agreement across the two task types. Studying the conditions under which attraction errors and attraction-based interference arise can inform us not only about the types of linguistic cues that enter into the computation of agreement dependencies, but also the nature of the mechanisms responsible for these computations, grammatical encoding, and the processing of long distance dependencies more generally.

Although agreement in comprehension has received considerably less attention than in production, the findings of attraction effects in both modalities suggests that similar information sources are used when establishing agreement dependencies. An important question is whether the general susceptibility to attraction in comprehension and production also points to shared cognitive, computational, or linguistic mechanisms used to establish agreement dependencies when speaking and reading, or whether attraction interference has different sources depending on the modality. The prevailing assumption in much psycholinguistic work is that the presence of attraction in both modalities indicates that agreement has largely overlapping cognitive or linguistic bases in production and comprehension (Badecker & Kuminiak, 2007; Hartsuiker, 2006; Nicol et al, 1997; Pearlmutter et al, 1999; Severens, Jansma, & Hartsuiker, 2008; Thornton & MacDonald, 2003; Wagers et al, 2009; see also Kreiner et al, 2013), though there remains debate about the exact nature of the processes giving rise to interference effects.

A unified, modality-neutral account is an attractive possibility, not least because it appeals to notions of parsimony: similar patterns of interference in comprehension and production should be explained by assuming a common cognitive mechanism or linguistic representation. This meshes with recent studies that find similarities in syntactic processing in comprehension and production, using behavioral methods (e.g., syntactic priming, Bock, Dell, Chang, & Onishi, 2007), and neuroimaging (e.g., fMRI, Segaert, Menenti, Weber, Petersson, & Hagoort, 2012). Furthermore, all of this aligns with recent models of language processing which propose an extremely tight link between language comprehension and production mechanisms (e.g., Gennari & MacDonald, 2009; MacDonald, 2013; Pickering & Garrod, 2013).

Despite this, there are fundamental differences in comprehension and production, and the two necessarily use a set of processing systems that are at least partly distinct. Production begins with abstract conceptual messages; comprehension begins with sensory input (auditory, visual, or even tactile). It is therefore possible that a differing set of underlying processes may give rise to only superficially similar profiles of interference. The goal of the present paper is to investigate attraction interference during language comprehension, and in the process, gain a more detailed understanding of the relationship between the processes underlying the formation of agreement dependencies in language production and comprehension. We focus in particular on the contexts in which interference arises, as well as the time course of the interference effects. In doing so, we use a methodology which has not been extensively used to study agreement attraction, namely event-related brain potentials (ERPs).

Mechanisms of Agreement and Interference Symmetries

Despite early findings of similar patterns of interference in production and comprehension, more recent findings on attraction suggest that the two domains may differ with regard to the contexts in which attraction interference arises. This suggests that the assumption of a shared mechanism needs further scrutiny. In particular, these conflicting findings point to a larger debate about the cause of attraction itself.

One family of theories holds that attraction arises because the mental representation of the number feature of the subject NP is faulty or ambiguous (e.g., Eberhard, Cutting, & Bock, 2005; Franck, Vigliocco, & Nicol, 2002; Nicol et al, 1997; Staub, 2009, 2010; Vigliocco & Nicol, 1998). On this account, attraction arises because the abstract plural feature of an attractor noun can erroneously spread throughout the syntactic representation of the subject NP, either through upwards feature percolation (Franck et al., 2002) or via spreading activation through a hierarchically structured syntactic tree (Eberhard et al, 2005; Hartsuiker, Antón-Méndez, & van Zee, 2001). It should be noted that these theories were initially developed to account for production Fdata; however, these ideas have also been applied to comprehension (e.g., Nicol et al, 1997; Pearlmutter et al, 1999), and we do so here. Others have argued that attraction effects can have a representational locus, but from slightly different perspectives. For example, some argue that attraction arises from stages of a syntactic derivation, as conceptualized within Minimalist syntactic theory (e.g., Franck, Lassi, Frauenfelder, & Rizzi, 2006), or from statistically-based representations formed through language experience (e.g., Haskell, Thornton, & MacDonald, 2010).

A second class of approaches argues that attraction effects arise because the processing mechanisms responsible for valuing or checking verb agreement may temporarily misidentify the correct controller of agreement (Badecker & Kuminiak, 2007; Dillon, Mishler, Sloggett, & Phillips, 2013; Wagers et al, 2009; see also Martin, Nieuwland, & Carreiras, 2012). This account, couched in a content-addressable working memory architecture, does not place the cause of interference on a globally faulty representation of the subject's number feature. Instead, it holds that retrieval of the agreement controller is subject to similarity-based interference from other items in memory (see also Gordon, Hendrick, & Johnson, 2001; Gordon, Hendrick, & Levine, 2002; Van Dyke & McElree, 2006). On this account, processing verb inflections requires retrieval of possible agreement controllers represented in memory. When the subject NP contains two noun phrases, they may potentially compete for control of verb agreement, even when only one of them is syntactically licensed as an agreement controller. When the two nouns mismatch in number, competition can result in a potential mis-identification of the agreement controller.

Many aspects of agreement production can be accurately modeled within the first approach, and Eberhard and colleagues (2005) have formalized this account within a computational model known as Marking and Morphing (MM).¹ A central tenet of MM is that the number value of the subject NP is continuously valued. The model formalizes this as the S(r) value, which reflects the relative singular and plural (SAP) value of the subject NP and is the “joint product or reconciliation of the initial marking of notional number and the number specifications of morphemes that compose the phrase” (p. 539). This means that, within this model, some root NP nodes can be ‘more plural’ than others, indicated by a higher S(r) value. For example, when an NP contains a singular head and a plural attractor noun and/or refers to a notionally (semantically) plural or collective entity, the number representation of the overall subject NP will be more plural than when both nouns are singular and/or the referent of the NP is unambiguously singular. This number representation of the subject NP (established through number ‘marking’ in the model) then triggers number agreement on the verb (established through ‘morphing’ in the model) via a probabilistic process: the more plural the overall subject NP's value, the greater the probability that the agreement process will select a plural verb. Within this model, the overall S(r) value of the subject NP is the only parameter visible for number marking, and not the individual nouns within the NP. Thus, it is the relative SAP value of the whole subject NP that specifies verb number and gives rise to attraction effects.

Given a continuously valued number global representation, the MM architecture thus predicts that difficulty in processing should arise regardless of whether or not the correct verb is ultimately chosen in production, or in comprehension, whether or not the verb encountered is properly marked for number. That is, agreement attraction should be symmetrical (Wagers et al, 2009).² In production, symmetry should give rise to processing difficulty for verbs following singular head and plural attractor nouns, even when the correct noun is ultimately chosen. Recent reaction time (RT) evidence squarely supports this prediction. When participants are asked to choose the correct verb (is or are) to continue a visually-presented sentence preamble either through a forced-choice button pushing task (Staub, 2009, 2010) or through elicited oral production (Brehm & Bock, 2013; Haskell & MacDonald, 2003), they are significantly slower following plural attractor nouns embedded in an otherwise singular NP than when the attractor was singular. This slowdown occurs even when they ultimately choose the correct verb. Staub (2009) argued that these findings supported the notion that the ambiguously-valued global number representation (e.g., a ‘slightly plural’ S(r) value) gives rise to uncertainty about the correct verb form, which in turn leads to longer RTs. Thus, in addition to accurately predicting effects on attraction rates of notional number (e.g., Bock, Carreiras, & Meseguer, 2012; Brehm & Bock, 2013), morphological richness (e.g., Foote & Bock, 2012; Lorimor, Bock, Zalkind, Sheyman, & Beard, 2008), and other factors (see Bock & Middleton, 2011, for a recent review), the MM model also correctly predicts symmetrical attraction effects during language production.

Shared Mechanisms and (A)symmetrical Attraction

MM was designed to account solely for production data, and although MM correctly predicts a range of interference effects in language production, including findings of symmetrical attraction, it is less clear to what extent it could account for facts of agreement comprehension. Maintaining a unified account of attraction in comprehension and production would minimally require: 1) that the same linguistic factors that modulate attraction rates in production also modulate attraction incomprehension, and 2) that interference is similarly symmetrical in comprehension. With respect to the first criterion, although years of research have identified numerous linguistic factors modulating attraction in production only a few studies have investigated whether there may be parallel effects in comprehension. Some findings from comprehension have converged with those from production regarding the role of attractor-predicate plausibility, morphophonological ambiguity in number marking on head noun, and the role of semantics in number marking on verbs versus pronouns (Kreiner et al, 2013; Severens et al, 2008; Thornton & MacDonald, 2003), though findings regarding the role of hierarchical structure in agreement comprehension have been equivocal (cf. Nicol et al, 1997; Pearlmutter, 2000; see Gillespie & Pearlmutter, 2013, for evidence against heirarchical structural constraints in agreement production).

With respect to the second criterion for a shared mechanism, results are less clear. In the RT data provided by Staub (2009), the globally ambiguous number representation of the subject NP symmetrically impacted both correct and incorrect decisions about verb number during a forced choice task (see also Brehm & Bock, 2013; Haskell & MacDonald, 2003). Extending this logic to comprehension, this would predict that plural attractors embedded in PP modifiers should cause symmetrical interference, affecting the processing of both grammatical and ungrammatical verbs. Early studies investigating interference effects in comprehension supported this hypothesis, using maze-reading and sentence judgment tasks (Nicol et. al, 1997), and using self-paced reading and eye-tracking (Pearlmutter et al, 1999; Thornton & MacDonald, 2003).

More recently, Wagers and colleagues (Wagers et al, 2009) have replicated the finding that plural attractors decrease ungrammaticality effects in self-paced reading, but questioned the finding that attraction also occurs in grammatical sentences. In a series of self-paced reading studies in English they showed the attraction effect in ungrammatical sentences is quite robust, but also argued that the previously-reported attraction effects in grammatical sentences may result from spillover effects from the increased complexity of the plural noun preceding the verb, and not actually from agreement processes themselves. Wagers and colleagues concluded that plural nouns, which are morphologically and conceptually more complex than singular nouns, incur processing costs which may spill over into subsequent reading regions, and that this spillover had erroneously been interpreted as an attraction effect in grammatical sentences in previous work.³ They concluded that attraction is asymmetrical: it impacts ungrammatical sentences, making them easier to process, but it does not make grammatical sentences more difficult to process (see also Dillon et al, 2013, for eye movement evidence of asymmetrical attraction from within relative clauses). Wagers and colleagues argue that their results are consistent with the second family of theories described above, and implicate cue-based working memory retrieval mechanisms in the establishment of agreement dependencies, where similarity-based interference from items represented in memory may give rise to processing difficulty during the processing of linguistic dependencies (Gordon et al, 2001, 2002; Lewis, Vasishth, & Van Dyke, 2006; Lewis & Vasishth, 2005; Martin & McElree, 2008; McElree, Foraker, & Dyer, 2003; McElree, 2000, 2006; Van Dyke & McElree, 2011; Van Dyke, 2007).

Both Wagers et al. and Dillon et al. argue that the asymmetry in attraction effects is predicted by cue-based retrieval mechanisms in the following way. Upon encountering a verb in a context such as “The author of the speeches are…” the processing system deploys a search in working memory for possible agreement controllers. Because the plural attractor noun has a partial feature overlap with an ungrammatical plural verb, an agreement dependency can be established between the two items. In the case of grammatical verbs (e.g. “The author of the speeches is…”), a plural attractor noun does not overlap in number features with the verb and therefore is unlikely to cause interference. Moreover, it is additionally possible that retrieval mechanisms operate in tandem with predictive mechanisms, where retrievals are driven by prediction errors. In this case, retrievals would only be initiated when detected agreement features mismatch with predicted features. This would further predict attraction asymmetries, as no retrievals would be necessary in grammatical sentences.

While the evidence from Wagers and colleagues for attraction from within prepositional phrases is compelling, it is difficult to reconcile with Nicol and colleagues' (1997) findings of attraction in grammatical sentences, especially because Nicol et al. showed in a control experiment with uninfected verbs (Experiment 3) that the interference effect disappeared. It is additionally worth noting that the statistical procedure Wagers and colleagues used to control for possible spillover effects from the plural attractor noun may have removed any effects of attraction in grammatical sentences from their reaction time data. That is, by using previous words' reading times as covariates, processing difficulty at a grammatical verb following a plural noun (i.e., a true attraction effect) could be regressed out of the data, since the attraction effect would be of a similar magnitude and direction to the effect of the plural noun in the previous reading regions. Note, however, that in Experiments 2 and 3 Wagers and colleagues did not use statistical covariates and found no differences in reading times for grammatical sentences with singular and plural attractors in relative clauses with a non-intervening attractor (e.g., The musician(s) who the reviewer praise(s)…), where spillover from a plural attractor (musicians) is not a concern at the target verb (praise(s)). However, more recent findings reported by Staub (2010) suggest that interference from non-intervening attractors and interference from intervening (PP-embedded) attractors have a different cognitive source (that is, predication confusion versus true attraction). Thus, more research is needed here. Moreover, Wagers and colleagues maintain that the content-addressable retrieval mechanism they propose to explain asymmetrical attraction effects in comprehension can also be used to account for attraction effects in production (see also Badecker & Kuminiak, 2007). However, the asymmetry they report for comprehension is clearly at odds with more recent findings of symmetrical attraction in verb selection and sentence completion tasks (e.g., Brehm & Bock, 2013; Staub, 2009).

A further possibility is that attraction in grammatical sentences exists, but is exceedingly subtle, such that the self-paced reading task used by Wagers and colleagues was insufficiently sensitive to detect it (compare the eye tracking results of Pearlmutter et al, 1999, Experiment 2). The approach we take here, namely using electrophysiological measures, provides an important source of information regarding this possibility, as ERPs have been shown to detect processing effects that go unnoticed in behavioral measures (e.g., McLaughlin, Osterhout, & Kim, 2004; Thierry & Wu, 2007; Tokowicz & MacWhinney, 2005). If it is found that attraction is asymmetrical, in contrast to findings from production-oriented experiments, this would countervail the assumption that attraction interference effects in comprehension and production have isomorphic sources. Finding symmetrical attraction in comprehension would be consistent with, though would not necessarily entail, the hypothesis that attraction in the two tasks has a shared cause. Finally, a third possibility is that attraction arises in both grammatical and ungrammatical sentences but from different sources, as has recently been shown to be the case for attraction from within prepositional phrases and non-intervening relative clauses (Staub, 2010). ERPs' multidimensional outcome measures could therefore potentially distinguish whether attraction in grammatical and ungrammatical sentences might arise from qualitatively different cognitive mechanisms.

Electrophysiology and Agreement Attraction

ERPs are direct recordings of cortical electrical activity underlying processing and have proven to be extremely useful in studying language comprehension. Importantly for our purposes, ERPs have a measurement scale on the order of milliseconds, which is fast enough to capture comprehension processes as they unfold in time, they provide multiple dimensions of outcome measures (e.g., effect timing, polarity, and scalp distribution), and have shown to be highly sensitive to aspects of morphosyntactic processing.

Most relevant to the present study is the finding that violations of morphosyntactic constraints, including subject-verb agreement violations, typically elicit an enhanced positive-going wave beginning around 500ms poststimulus which is most prominent over centroparietal scalp regions (the P600 effect: Friederici, Hahne, & Mecklinger, 1996; Kaan, Harris, Gibson, & Holcomb, 2000; Molinaro, Barber, & Carreiras, 2011; Molinaro, Vespignani, Zamparelli, & Job, 2011; Nevins, Dillon, Malhotra, & Phillips, 2007; Osterhout & Holcomb, 1992; Osterhout & Mobley, 1995; Rossi, Gugler, Hahne, & Friederici, 2005). Some studies of morphosyntactic processing have additionally elicited a negativity preceding the P600, which is sometimes strongest over left anterior portions of the scalp (the left anterior negativity, or LAN: e.g., Barber & Carreiras, 2005; Friederici et al, 1996; Hahne & Friederici, 1999; Osterhout & Holcomb, 1992; Osterhout & Mobley, 1995), though the LAN is highly inconsistent across studies (e.g., Allen, Badecker, & Osterhout, 2003; Frenck-Mestre, Osterhout, McLaughlin, & Foucart, 2008; Hagoort & Brown, 1999, 2000; Hagoort, 2003; Kim & Osterhout, 2005; Nevins et al, 2007; Osterhout, Allen, McLaughlin, & Inoue, 2002; Osterhout & Mobley, 1995; Tanner & Van Hell, 2014). We therefore focus on the P600 effect. The onset timing of the P600 has been shown to be sensitive to the relative ease in diagnosing a syntactic anomaly (Friederici, Mecklinger, Spencer, Steinhauer, & Donchin, 2001; McKinnon & Osterhout, 1996), and the amplitude of the P600 effect can be modulated by the relative perception of ungrammaticality or degree of difficulty in grammatical processing. Larger P600 magnitudes are parametrically associated with less resolvable and more salient grammatical violations, as well as increased processing difficulty in grammatical but difficult-to-process long-distance and verb-argument dependences (e.g., Kaan et al, 2000; Kaan & Swaab, 2003; Kaan, 2002; Nevins et al, 2007; Osterhout, Holcomb, & Swinney, 1994; Xiang, Dillon, & Phillips, 2009). Thus, the P600 effect onset and magnitude can reliably be used as a continuous index of syntactic processing difficulty at the brain level for both grammatical and ungrammatical sentences.

Only a few studies have used ERPs to investigate attraction-like phenomena in comprehension (see Xiang et al, 2009, for additional ERP evidence regarding intrusive licensing of negative polarity items), and as with the behavioral studies reviewed above, the results have been somewhat inconclusive regarding attraction asymmetries. Kaan (2002) investigated the impact of number interference on verb agreement in Dutch subject-object-verb sentences. Kaan's main findings were that P600 amplitudes to disagreeing verbs in clauses where the subject and object NPs' number features were both plural or mismatched were significantly smaller than in cases where both NPs were singular, suggesting that verb agreement violations in the singular-singular condition were more salient than in the other conditions. Hints of asymmetry were also observed: in the singular subject/plural object condition, grammatical verbs following plural object nouns elicited a positivity between 250-300ms poststimulus, but ungrammatical verbs showed a more typical positivity in the later time windows. Kaan suggested that participants may have initially perceived the grammatical sentences as ungrammatical given the number mismatch in the local string (i.e., plural attractor noun-singular verb), but that this misanalysis was quickly corrected.

Severens and colleagues (Severens et al, 2008) included an agreement attraction manipulation in a study on the role of morphophonology in Dutch agreement processing. They found that, whereas disagreeing verbs following complex subjects containing PP modifiers with mismatching number features (i. e., singular head/plural attractor) showed a P600, disagreeing verbs following complex subjects with matching number features (i.e., singular head/singular attractor) showed an unexpected N400-like negativity. The authors attributed this unexpected N400 to the differences in the depth of processing needed in the match and mismatch conditions. They speculate that participants may use ‘shallow processing’ strategies in the relatively simpler match condition, and may use a deeper processing strategy in the more complex mismatch condition (Kuperberg, 2007; see also Ferreira, Bailey, & Ferraro, 2002; Ferreira, 2003). Importantly, neither the Kaan (2002) nor the Severens et al. (2008) studies directly compared brain responses for the singular and plural attractor conditions in grammatical sentences, so they provide no electrophysiological evidence regarding the symmetry of attraction in comprehension.

More recent work has provided tentative support for asymmetrical attraction effects. However, some of the findings have implicated ERP effects other than the canonical P600. Martin and colleagues (Martin et al, 2012) investigated gender attraction effects in Spanish NP ellipsis using ERPs. Sentences contained two NPs, which either matched or mismatched in gender, as well as a gapped NP indicated by the determiner otro/a, which either correctly agreed with the first NP, or incorrectly agreed with the, second NP (e.g., Marta se compró la camiseta que estaba al lado de la falda/del vestido y Miren cogió otra/*otro […] para salir de fiesta, “Marta bought the t-shirt_Fem that was next to the skirt_Fem/dress_Masc and Miren took another_Fem/*Masc to go to the party”). Results showed that gender mismatches at the gapping site elicited a sustained, widespread negativity, rather than a P600, that largely impacted grammatical rather than ungrammatical sentences. Symmetrical agreement emerged only in a late time window.

Shen, Staub, and Sanders (2013) recorded ERPs to agreement violations, including attraction violations, in auditorily-presented spoken English discourses. They found a typical LAN-P600 pattern of effects for simple agreement violations (i.e., those without an intervening attractor noun). Relevant to the current discussion, in complex NPs they found no differences when comparing ERP responses to grammatical verbs following singular and plural attractors, suggesting no increased processing difficulty in these contexts, in line with the asymmetrical attraction findings reported by Wagers and colleagues. However, other effects of number and grammaticality they reported were atypical of most ERP studies of morphosyntactic processing (either auditory or visual), as they largely involved posterior negativities to ungrammaticalities. Given the differences in materials and procedure relative to other studies of agreement attraction, it is unclear how well these results might generalize to the broader agreement attraction literature.

Here we present data from three experiments (two ERP, one behavioral) on agreement processing in the comprehension of English sentences. Given the conflicting findings regarding the contexts in which attraction interference exerts itself in comprehension, our primary goals were to identify the contexts in which attraction interference occurs in English agreement as well as the time course of the interference. Establishing asymmetric effects of attraction during comprehension would indicate that, counter to some current assumptions, attraction has differing sources in comprehension and production.

Experiment 1

In Experiment 1 we investigated classic agreement attraction effects in English using ERPs. We tested interference effects from plural attractor nouns in NPs with singular head nouns, as singular attractor nouns with plural head nouns have been shown to exert very little interference in both production and comprehension (Bock & Miller, 1991; Eberhard, 1997; Nicol et al., 1997; Pearlmutter et al, 1999; Thornton & MacDonald, 2003). In line with previous research, we expected that subject-verb agreement violations would elicit robust P600 (and possibly LAN) effects. Of particular importance to the theoretical questions posed here is whether interference would be apparent primarily in grammatical sentences (Martin et al, 2012), primarily in ungrammatical sentences (Wagers et al, 2009), or both (Pearlmutter et al, 1999). As P600 effects can reflect processing difficulty in both grammatical and ungrammatical sentences, attraction interference should manifest itself in ungrammatical sentences as a reduction in P600 amplitude, whereas attraction in grammatical sentences should manifest itself as an increase in P600 amplitude, relative to grammatical sentences with singular attractors.

Method

Participants

Our participants included 24 native English speakers. Participants were all strongly right-handed, as assessed by an abridged version of the Edinburgh Handedness Inventory (Oldfield, 1971), and had normal or corrected-to-normal vision. Data from one participant were excluded because a language history questionnaire revealed significant early exposure to a non-English language with agreement patterns substantially different from English (Mandarin Chinese); all other participants reported no significant knowledge of or exposure to languages other than English and were functionally monolingual. Two additional participants were excluded due to an excessive amount of blink artifact in the raw EEG. This left 21 participants in the final analyses (13 male; mean age: 21.2 years, range: 18-32). Participants either volunteered or received course credit for taking part.

Materials

120 sentence quadruplets were constructed to serve as critical materials for this experiment in a 2×2 factorial design. The factor grammaticality (grammatical, ungrammatical) was crossed with attractor number (singular, plural), resulting in four versions of each sentence frame (see Table 1). Critical words were short auxiliary verbs (is/are, was/were, has/have) which either agreed with the singular head noun of the subject NP (grammatical verbs) or disagreed (ungrammatical verbs); the number of the attractor noun embedded within a PP subject modifier was also manipulated as either singular or plural. Some preambles were adapted from published stimulus materials investigating agreement attraction (Bock & Miller, 1991; Staub, 2009; Wagers et al, 2009); others were newly constructed. The four versions of each sentence were distributed across four experimental lists, such that each participant saw only one version of each sentence and there were 30 critical trials in each condition per list.

Table 1. Example experimental sentences from Experiment 1.

Grammaticality	Attractor Number	Sentence
Grammatical	Singular	The chemist with the test tube is conducting an experiment.
Ungrammatical	Singular	The chemist with the test tube are conducting an experiment.
Grammatical	Plural	The chemist with the test tubes is conducting an experiment.
Ungrammatical	Plural	The chemist with the test tubes are conducting an experiment.

Open in a new tab

Note: The critical word for ERP averaging is underlined.

Each list also contained 120 filler sentences. Half of the filler sentences were, grammatical and half contained a violation of determiner-noun agreement or numerical quantifier-noun agreement (e.g., I read one/*two good book last summer). Each list thus contained 240 sentences, half of which contained an ungrammaticality. Each list was pseudo-randomized such that no more than three trials from any single condition occurred in succession and no more than three grammatical or ungrammatical trials occurred in succession. Lists were divided into 3 blocks of 80 sentences each.

Procedure

Participants were tested in a single session lasting approximately eighty-five minutes (including about thirty minutes of experimental preparation). Each participant was randomly assigned to one of the stimulus lists and was seated in a comfortable recliner in front of a CRT monitor. Participants were instructed to relax and minimize movements while reading and to read each sentence as normally as possible. Each trial consisted of the following events: each sentence was preceded by a blank screen for 1000ms, followed by a fixation cross, followed by a stimulus sentence, presented one word at a time. The fixation cross and each word appeared on the screen for 450ms followed by a 200ms blank screen between words. Sentence-ending words were followed by a full stop, followed by a “yes” or “no” prompt, instructing participants to give an acceptability judgment for the sentence. Participants were instructed to respond “yes” if they felt that it was a well-formed English sentence or “no” if they felt something was wrong with the sentence. The “yes” response hand was counterbalanced across participants. Participants were put under no time pressure to make their judgments.

Data acquisition and analysis

Continuous EEG was recorded from 19 tin electrodes attached to an elastic cap (Electro-cap International) in accordance with the 10-20 system (Jasper, 1958). Eye movements and blinks were monitored by two electrodes, one placed beneath the left eye and one placed to the right of the right eye. All electrodes were referenced to an electrode placed over the left mastoid and were amplified with a bandpass of 0.01-40 Hz (-3db cutoff) by an SA Instruments bioamplifier system. EEG was recorded from an additional electrode placed on the right mastoid to identify if there were any experimental effects detectable over the mastoids; no such effects were found. Impedances at scalp and mastoid electrodes were held below 5 kΩ and below 15 kΩ at eye electrodes.

Continuous analog-to-digital conversion of the EEG and stimulus trigger codes was performed at a sampling frequency of 200Hz. ERPs, time-locked to the onset of the critical word, were averaged off-line for each participant at each electrode site in each condition. Grand average waveforms were created by averaging over participants. Trials characterized by eye blinks, excessive muscle artifact, or amplifier blocking were not included in the averages; 4.6% of trials overall were removed due to artifacts and the number of rejected trials did not differ across conditions.

Behavioral data was quantified both as proportion judged correctly in each of the four conditions in Table 1 and as d-prime scores for the each of the singular and plural attractor conditions to test decreases in sensitivity to the grammaticality manipulation from attractor interference. Proportions were arcsine transformed and analyzed using a 2×2 repeated-measures ANOVA with factors grammaticality (grammatical, ungrammatical) and match (match, mismatch); d-prime data were analyzed using a one-factor repeated measures ANOVA to show whether overall sensitivity to the agreement violations decreased in the match versus mismatch conditions. ERP components of interest were quantified by computer as mean voltage within a window of activity. In-accordance with previous literature and visual inspection of the data, the following time windows were chosen: 150-300ms (P2), 300-500ms (LAN/N400), and 500-800ms (P600), relative to a 200ms prestimulus baseline. Differences between conditions were analyzed using a repeated-measure ANOVA with two levels of grammaticality (grammatical, ungrammatical) and two levels of attractor number (singular, plural). Additionally, to investigate the scalp distribution of the relevant effects, data from midline (Fz, Cz, Pz), medial (right hemisphere: Fp2, F4, C4, P4, O2; left hemisphere: Fp1, F3, C3, P3, O1), and lateral (right hemisphere: F8, T8, P8; left hemisphere: F7, T7, P7) were treated separately. ANOVAs on midline electrodes included electrode as an additional within-subjects factor (3 levels), ANOVAs on medial-lateral electrodes included hemisphere (2 levels) and electrode pair (5 levels) as additional within-subjects factors, and ANOVAs over lateral-lateral electrodes included hemisphere (2 levels) and electrode pair (3 levels) as additional within-subjects factors. The Greenhouse-Geisser correction for inhomogeneity of variance was applied to all repeated measures with greater than one degree of freedom in the numerator. In such cases, the corrected p-value is reported.

Results

Behavioral results

Mean proportion of sentences judged correctly as acceptable or unacceptable and d-prime scores are presented in Table 2. A repeated measures ANOVA on arcsine-transformed proportions judged correctly showed no main effect of grammaticality (p > .27). There was a main effect of attractor number, F(1, 20) = 17.899, MSE = 0.016, p < .001, and an interaction which did not reach significance, F(1, 20) = 2.454, p = .133. Because the theoretical questions of interest crucially hinge on whether there is an asymmetry in attraction effects between grammatical and ungrammatical) sentences, we computed planned pair-wise tests comparing the effects of attractor number within both grammatical and ungrammatical sentences. These effects showed no reliable effect of attractor number within grammatical sentences, p > .05, but a reliable effect in ungrammatical sentences, t(20) = 3.719, p = .001. Analysis of d-prime scores showed that sensitivity to grammatical violations decreased significantly in the presence of a plural attractor, as there was a reliable effect of attractor number, F(1, 20) = 18.143, MSE = .326, p < .001. In sum, behavioral accuracy showed a reliable attraction effect, as sensitivity to subject-verb agreement grammaticality was significantly poorer in the plural attractor condition. However, this effect was asymmetrical and driven primarily by a reduction in sensitivity to ungrammatical verbs following plural attractor nouns and not by a perception of ungrammaticality for grammatical verbs following plural attractors.

Table 2.

Means and standard errors for behavioral judgments in Experiment 1. Proportion judged correctly and d-prime scores. Standard errors are reported in parentheses.

	Grammatical	Ungrammatical	d-prime
Sg. Attractor	.92 (.02)	.90 (.04)	3.23 (.23)
Pl. Attractor	.90 (.02)	.79 (.05)	2.48 (.26)

Open in a new tab

ERP results

Grand mean ERP waveforms comparing brain responses to grammatical and ungrammatical verbs in the singular attractor condition are presented in Figure 1, and for the plural attractor condition in Figure 2. Waveforms comparing all four conditions at central midline electrode Cz are presented in Figure 3. In these and all subsequent waveforms, the general shapes of the waveforms were consistent with previous data using visually presented language stimuli (e.g., Osterhout & Holcomb, 1992). Visual inspection of the waveform showed that, relative to grammatical verbs, brain responses to ungrammatical verbs were more positive-going beginning around 450ms poststimulus, though the magnitude of the positivity was much larger in the singular attractor condition than the plural attractor condition. In both the singular and plural attractor conditions there appeared to be a small left-central negativity elicited by ungrammatical verbs beginning around 300ms, and in the singular attractor condition, ungrammatical verbs showed an additional early positivity between 150-300ms poststimulus. Statistical analysis confirmed all of these observations.

Grand mean ERP waveforms from Experiment 1 for all 19 scalp electrodes in the grammatical (solid line) and ungrammatical (dashed line) singular attractor conditions. Onset of the verb is indicated by the vertical calibration bar; each tick mark represents 100ms of time. Negative voltage is plotted up. ERPs were filtered with a 15 Hz low-pass filter for plotting purposes only in these and all subsequent waveforms.

Grand mean ERP waveforms from Experiment 1 in the grammatical (solid line) and ungrammatical (dashed line) plural attractor conditions. Onset of the verb is indicated by the vertical calibration bar; each tick mark represents 100ms of time. Negative voltage is plotted up.

Grand mean ERP waveforms for all four experimental conditions in Experiment 1 at midline vertex electrode Cz. Onset of the verb is indicated by the vertical calibration bar; each tick mark represents 100ms of time. Negative voltage is plotted up.

In the 150-300ms time window, there were no main effects of grammaticality or attractor number over any scalp region; however, there was a significant interaction between grammaticality and attractor number over midline electrodes, F(1, 20) = 6.614, MSE = 4.447, p = .018, and which neared significance over lateral electrodes, F(1, 20) = 3.843, MSE = 3.350, p = .064. Follow-up tests over midline sites showed that this interaction was driven by an increased positivity to ungrammatical verbs in the singular attractor condition, F(1, 20) = 11.776, MSE = 3.349, p = .003, which was not present in the plural attractor condition (F < 1). In the 300-500ms time window, the only reliable effect of interest was a significant grammaticality by electrode by hemisphere interaction over lateral electrodes, F(2, 40) = 4.165, MSE = .538, p = .041. This effect indicates a small negativity over left hemisphere sites and a small positivity over right hemisphere sites, particularly over electrodes T7/T8 (cf. Tanner & Van Hell, 2014).

Between 500-800ms there was a large, widespread main effect of grammaticality, indicating a P600 effect to ungrammatical verbs [midline: F(1, 20) = 23.747, MSE = 14.339, p < .001; medial: F(1, 20) = 16.017, MSE = 28.607, p < .001; lateral: F(1, 20) = 12.454, MSE = 10.361, p = .002], which was largest over posterior sites [grammaticality × electrode interaction, midline: F(2, 40) = 9.095, MSE = 3.027, p = .004; medial: F(4, 80) = 12.127, MSE = 11.691, p < .001; lateral: F(2, 40) = 14,634, MSE = 2.336, p < .001), and which also was dominant over right posterior lateral sites [grammaticality × electrode × hemisphere interaction: F(2, 40) = 6.339, MSE = .758, p = .040].

Importantly, there was also a grammaticality by attractor number interaction [midline: F(1, 20) = 14.081, MSE = 5.099, p = .001; medial: F(1, 20) = 7.794, MSE =13.576, p = .011; lateral: F(1, 20) = 6.193, MSE = 7.141, p = .022]. We investigated this interaction by computing pair-wise ANOVAs between all four conditions over midline electrodes, since this is where the interaction was strongest. This analysis showed that the effect of grammaticality was highly reliable in the singular attractor condition, F(1, 20) =41.263, MSE = 8.785, p < .001, but much less so in the plural attractor condition, F(1, 20)= 4.675, MSE = 10.652, p = .043. A comparison of brain responses in the two ungrammatical conditions revealed a reliable main effect of attractor number, F(1, 20) =10.216, MSE = 10.514, p = .005, indicating that the positivity to ungrammatical verbs was significantly smaller following plural attractor nouns. However, when comparing brain responses in the two grammatical conditions, there were no detectable differences, F(1, 20) = .517.

Discussion

The main results of Experiment 1 are as follows. Plural attractors showed a reliable interference effect in both participants' judgments of the sentences and in the pattern of brain responses elicited by verbs during reading. Importantly, the attraction effect was asymmetrical: plural attractors ameliorated effects of ungrammaticality, whereas they did not cause grammatical sentences to be perceived as correspondingly less grammatical. For behavioral judgments, this manifested itself as a reduction in accuracy for ungrammatical verbs following plural attractors, with no such reduction in accuracy for grammatical singular verbs following plural attractors. For participants' brain responses, effects of ungrammaticality emerged as early as the 150-300ms time window, but only following singular attractors; effects of ungrammaticality following plural attractors were less reliable and did not appear until the 500-800ms time window, suggesting these agreement violations were detected more slowly and processed less robustly in that condition. Indeed, the P600 for ungrammatical verbs was reliably smaller in the plural attractor than singular attractor condition. However, the plural attractor had no effect on grammatical verbs: no brainwave effect of processing difficulty emerged as a P600 or any other ERP component.

Overall these results are consistent with the notion that attraction interference in comprehension is asymmetrical with respect to grammaticality. Moreover, our results also indicate that this asymmetrical interference manifested itself very quickly after onset of the verb: the early effect of ungrammaticality on the P2 peak between 150-300ms was absent following plural attractors. Note that we did not replicate the early positivity for grammatical verbs following plural attractors reported by Kaan (2002), nor did we replicate the pattern of agreement attraction reported by Martin et al. (2012). Kaan suggested that grammatical singular verbs following plural attractor nouns may initially be perceived as ungrammatical, but that participants quickly overcome this initial misparse. In the current experiment both visual inspection of the waveforms as well as statistical analyses showed no evidence of an early misparse in grammatical sentences, though there was a clear early effect of ungrammaticality, modulated by agreement attraction. Martin and colleagues showed that gender agreement violations involving NP ellipsis elicited a broadly distributed, sustained negativity, but that attraction impacted primarily grammatical sentences. We return to this point in the general discussion.

The pattern of results reported here is inconsistent with the MM model of agreement processing. If features of attractor nouns can erroneously value the subject NP as ‘slightly more plural’ (in the terms of the MM model), it is predicted that ungrammatical verbs will be perceived as ‘slightly grammatical’ and that grammatical verbs will be perceived as ‘slightly ungrammatical’ to the same extent. However, neither the judgment data nor ERP results from this experiment support that view for comprehension, suggesting that – assuming the MM is the right characterization for production - the processing mechanisms which give rise to attraction interference in comprehension and production are at least partly distinct.

One possible objection to this claim is that the time course of activation of the attractor noun's plural feature may differ in production and comprehension. For example, work by Nicol (1995), and more recently by Pearlmutter and colleagues (Gillespie & Pearlmutter, 2011, 2013; Solomon & Pearlmutter, 2004), argues that the timing of activation of the head and attractor nouns can impact rates of agreement attraction, where plural attractor nouns that are activated simultaneously with a singular head noun will cause more interference than those that are activated later. During speech planning, conceptual representations of the subject NP's referent(s) and associated numerosities are activated prior to morpho syntactic marking of number on the verb according to the MM model. On this approach, a plural attractor's number feature will have adequate time to infiltrate the representation of the subject NP number and lead to interference in verb. marking during production. However, in comprehension, information is added incrementally to a conceptual representation of the sentence as the language input unfolds over time. A reader or hearer will have no knowledge of an upcoming noun's number feature until that noun is eventually encountered, and incorporating a plural number feature into the syntactic or conceptual representation of the sentence will necessarily take time. Thus, it may have been the case that, given the linear adjacency of the attractor and verb, participants did not have time to update the representation of the subject NP's number before encountering the verb, such that there was no increase in P600 amplitude for grammatical verbs following plural attractors.

Some evidence from this experiment argues against this objection. Principally, if attraction is driven primarily by the overall representation of the subject NP number (e.g., the S(r) value, within the MM model), and a plural attractor noun had not had adequate time to alter participants' representations of subject NP number, there should have been no amelioration effect from plural attractors on ungrammatical verbs in the ERP record. The fact that plural attractors reduced positivities to ungrammatical verbs as early as 150ms after onset of the verb suggests that the processes responsible for checking number agreement are sensitive to attraction interference from the earliest stages of anomaly detection. Nonetheless, there maybe multiple mechanisms that can give rise to attraction in comprehension, each operating on a different time course. For example, retrieval interference, which would trigger an illusion of grammaticality like that reported here, may operate extremely rapidly. Revaluation of the subject number feature, which would trigger an illusion of ungrammaticality, may require more time (see also Brehm & Bock, 2013) and show up only in late measures of processing. Our judgment data do not support this possibility, however. While there was an overall decrease in accuracy following a plural attractor noun in participants' acceptability judgments, this was driven primarily by a decrease in accuracy in the ungrammatical condition. That is, attraction effects were similarly asymmetrical in both the ERP and off-line judgments. These offline judgments represent a late, global assessment of sentence-level acceptability and should include representational information about subject NP numerosity. Nonetheless, allowing more time between the attractor noun and verb might facilitate revaluation of subject NP number as more plural, allowing interference to manifest in grammatical sentences. We investigate this possibility in Experiment 2.

Experiment 2

In this experiment we tested the hypothesis that revaluation of the subject NP's number feature following a plural attractor noun requires time, meaning that creating greater linear (and therefore temporal) separation between the attractor noun and verb would give rise to a more symmetrical pattern of attraction. Specifically, we linearly separated the attractor noun and verb with an adverb (cf. Wagers et al, 2009), thereby increasing the amount of time available for a plural attractor noun's number feature to impact the representation of the subject NP's number. If number representation is a crucial contributor to attraction interference in comprehension, as predicted by the MM model for production, we should see attraction effects for both grammatical and ungrammatical verbs in this experiment. Alternately, if agreement dependencies during comprehension are established primarily via cue-based search in working memory (e.g., McElree, 2006; Van Dyke & McElree, 2011; Van Dyke, 2007), we predict a pattern similar to that seen in Experiment 1.