Electrophysiological Evidence for Use of the Animacy Hierarchy, but not Thematic Role Assignment, During Verb Argument Processing

Martin Paczynski; Gina R Kuperberg

doi:10.1080/01690965.2011.580143

. Author manuscript; available in PMC: 2011 Dec 21.

Published in final edited form as: Lang Cogn Process. 2011;26(9):1402–1456. doi: 10.1080/01690965.2011.580143

Electrophysiological Evidence for Use of the Animacy Hierarchy, but not Thematic Role Assignment, During Verb Argument Processing

Martin Paczynski ^*, Gina R Kuperberg ^*,^§

PMCID: PMC3244078 NIHMSID: NIHMS309730 PMID: 22199415

Abstract

Animacy is known to play an important role in language processing and production, but debate remains as to how it exerts its effects: 1) through links to syntactic ordering, 2) through inherent differences between animate and inanimate entities in their salience/lexico-semantic accessibility, 3) through links to specific thematic roles. We contrasted these three accounts in two event related potential (ERP) experiments examining the processing of direct object arguments in simple English sentences. In Experiment 1, we found a larger N400 to animate than inanimate direct object arguments assigned the Patient role, ruling out the second account. In Experiment 2 we found no difference in the N400 evoked by animate direct object arguments assigned the Patient role (prototypically inanimate) and those assigned the Experiencer role (prototypically animate), ruling out the third account. We therefore suggest that animacy may impact processing through a direct link to syntactic linear ordering, at least on post-verbal arguments in English. We also examined processing on direct object arguments that violated the animacy-based selection restriction constraints of their preceding verbs. These violations evoked a robust P600, which was not modulated by thematic role assignment or reversibility, suggesting that the so-called semantic P600 is driven by overall propositional impossibility, rather than thematic role reanalysis.

General Introduction

One of the key questions within psycholinguistics is how semantics and syntax interact to build overall sentence meaning. The animacy of a noun-phrase is known to influence both language comprehension and production, and is central to the syntax-semantic interface. However, it remains unclear how it exerts its effects. Several possibilities have been discussed. One account, explored primarily in the production literature, is that animacy acts directly by influencing the linear ordering of syntactic constituents. A second account, also mainly considered in the production literature, is that the effects of animacy arise through inherent differences in salience/lexico-semantic accessibility of animate and inanimate entities. A third possibility, explored in both production and comprehension literature, is that animacy acts by influencing the assignment or reassignment of thematic roles. We designed two experiments to distinguish between these three accounts by examining the processing of direct object arguments in active English sentences using event related potentials (ERPs).

In addition to influencing the processing of plausible sentences, animacy is also known to influence the processing of implausible sentences containing selection restriction violations. Once again, however, the mechanism by which it exerts its effects is unclear, with different models of language processing predicting different types of interactions between the animacy of an argument and the selection restrictions of a verb. The second aim of this study was to distinguish between these models by examining how animacy and thematic role reversibility influenced the processing of direct object arguments that violated the animacy-based selection restrictions of their preceding verbs.

In this General Introduction, we first review evidence that animacy can impact the linear ordering of syntactic constituents, mainly from cross-linguistic studies and studies of language production. We also consider evidence from the production literature that the effects of animacy may be driven by inherent differences between animate and inanimate entities in their salience/lexico-semantic accessibility, regardless of linear order. We then discuss evidence for an association between animacy and thematic roles, from both behavioral and electrophysiological studies of sentence comprehension. Finally, we outline the specific questions addressed in the current study.

Associations between Animacy and Linear Ordering

The animacy hierarchy, described by Silverstein (1976), as well as Dixon (1979) and Aissen (2003), refers to a preference for organizing entities along an animacy continuum, with humans ranking the highest, abstract concepts ranking lowest, with animals, plants, and inanimate objects falling in between. In the present study, we will use ‘animate’ to refer to humans and animals, and ‘inanimate’ to refer to inanimate objects. There is now fairly compelling evidence, from cross-linguistic studies and studies of language production, that the animacy hierarchy may be linked directly to the order of syntactic constituents.

Cross-Linguistic Evidence

In several languages, the animacy hierarchy is formally encoded in syntactic structure, constraining the linear ordering of arguments within a sentence. For example, in Navajo (Hale, 1972), Lakhota (Van Valin & LaPolla, 1997) and Mayan (Jakeltek, Craig, 1977; Mam-Mayan, Minkoff, 2000), animate nouns must always precede inanimate ones in transitive constructions. Thus, The rock hit the man is ungrammatical, and the utterance must instead be expressed as The man was hit by the rock. Minkoff (2000) demonstrated that such constraints are based on relative, rather than absolute animacy. Thus, The dog bit the man is ungrammatical as humans outrank animals, but The rock hit the window is grammatical as both entities are inanimate, thus neither outranks the other. In English, a similar animacy-based constraint is place on the realization of the double object ditransitive construction. Such constructions are ungrammatical when the indirect object is inanimate and direct object is animate, e.g. *The president sent England the ambassador. They require the use of the prepositional alternation, e.g. The president sent the ambassador to England. (cf. Levin, 1989). Again, such constraints are relative, as The man got the car new tires is grammatically felicitous.

Importantly, the tendency to place animate before inanimate nouns can be observed even without formal syntactic constraints on animacy ordering, as shown by several corpus studies. Examining German subordinate clauses, Kempen and Harbusch (2004) demonstrated that animate nouns tended to precede inanimate ones, independent of grammatical function. Øvrelid (2004) found that, in Norwegian, subject nouns were either higher or of equal animacy status as objects in over 97% of the transitive constructions reviewed by the author. Snider and Zaenen (2006) reported similar results in the English ‘Switchboard’ corpus, finding that animate nouns were three times more likely to appear as subject arguments than inanimate nouns, while the reverse was true for the grammatical objects. Such effects are not limited to verb arguments. Rosenbach, (2008) reported that, in English, possessive constructions vary depending on the animacy of the arguments, with the captain’s table being preferred over the table of the captain, but the man of the mountain being preferred over the mountain’s man.

Evidence from studies of Language Production

Consistent with the trends observed in language use outside of the laboratory, experimental studies of language production show that animacy influences the linear ordering of arguments within a sentence. During both spoken and written production, animate nouns are consistently placed at the sentence-initial position. In a series of studies, Bock demonstrated that, when asked to describe pictures containing an inanimate Actor and an animate Undergoer, participants employed the passive voice up to 74% of the time, compared to never when the Actor was animate (Bock 1986; Bock & Loebell, 1990, Bock, Loebell & Morey, 1992). Ferreira (1994) likewise found that, when the verb required an animate direct object argument (in canonically-ordered sentences), participants tended to use the passive voice when the arguments differed in animacy, placing the animate noun in subject position (The man was alarmed by the news), while no such preference was seen than when both arguments were animate (The vagabond alarmed the man.). Similar biases have been observed in delayed recall tasks, in which participants frequently misremember the original grammatical form of sentences, placing the animate argument before the inanimate one. This effect was found to be independent of grammatical voice in English (McDonald, Bock & Kelly, 1993) and Japanese (Tanaka et al., 2005), while in Greek, which allows relatively free argument ordering, participants often re-ordered the arguments without changing grammatical voice (Feleki & Branigan, 1997). These production biases appear to be learned at an early age (English: Dewart, 1979; Fiji: Bryne & Davison, 1985; Sesetho: Demuth, Machobane, Moloi, & Odato, 2005). Indeed, Demuth et al. (2005) found that children attained adult proficiency in using animacy hierarchy constraints in Sesetho several years prior to attaining adult proficiency in use of thematic roles.

In their 1992 study, Bock, Loebell and Morey found that participants were more likely to generate sentences with animate subjects after reading prime sentences with animate subjects (similarly for inanimates). They were also more likely to produce active sentences following active primes, and passive sentences following passive primes. Importantly, their choice of grammatical voice was unaffected by the animacy of the subject in the prime sentences. This lack of interaction between Animacy and Voice led the authors to conclude that, during production, animacy information may act independently of thematic roles, suggesting a more direct mapping between animacy features and syntactic arguments.

One reason that has been proposed for the ‘animate first’ bias in production is that animate entities are generally more salient, predictable or accessible than inanimate items. Generally, more things can be said about animate entities than inanimate ones (see Bock, Loebell & Morey, 1992). For example, both an athlete and car may both move fast and come from Germany, but only the former is likely to be described as humble and intelligent. According to this view, more accessible or salient items are assigned to the more prominent syntactic position of a sentence (e.g. Bock, Loebell & Morey, 1992), i.e. there is a direct mapping between increased salience of conceptual features and the first, most prominent argument position. Support for this hypothesis comes from Prat-Sala and Branigan (2000) who found that, during production, entities made more salient either through a discourse context or through an animacy manipulation showed a similar pattern of being placed first.

Associations between Animacy and Thematic Roles

Besides influencing the linear ordering of syntactic constituents, animacy has also been proposed to influence language through its associations with thematic roles. Thematic roles are a means of describing the semantic roles that a noun plays within a sentence construction. For example, in the sentence Jane broke the vase, Jane is considered an Agent, as she is the initiator of an action, while vase is a Patient, the entity that is affected by the action. In the current paper, we focus on three thematic roles, each of which has a prototypical relationship with animacy: Agent, Patient, and Experiencer¹. Animate entities are considered more ideal Agents because they can be volitional initiators of actions, while inanimate entities are considered more ideal Patients. The Experiencer role has even stronger associations with animacy than the Agent or Patient role, in that it refers to any argument that experiences an emotional or cognitive state. Thus Experiencers are necessarily animate.

Most studies of comprehension have interpreted effects of animacy as arising from its effect of facilitating the assignment or reassignment of thematic roles. This evidence comes from both behavioral studies and studies using event-related potentials (ERPs), a direct online measure of neural activity.

Behavioral studies of sentence processing

The first formal processing model that considered a role of animacy in thematic role assignment was introduced by Frazier and Rayner (1982) who discussed a thematic processing stream which relied on a small set of semantic features, such as animacy, to aid in thematic role assignment. In their original study, the authors examined the effects of animacy on the processing of reduced relative clauses (e.g. The witness examined by the lawyer…) compared to their unreduced counterparts (e.g. The witness that was examined by the lawyer…). Frazier and Rayner (1982) showed that the reading time penalties on the disambiguating ‘by’ phrase in reduced relatives were greatly reduced when the initial sentence argument was inanimate rather than animate (The evidence examined by the lawyer…). These results were replicated by Rayner, Carlson, and Frazier (1983) and by Ferreira and Clifton (1986). The authors suggested that after the initial (incorrect) syntactic parse of reduced relative clauses, the thematic processing mechanism aided in thematic role reanalysis. Specifically, because inanimate nouns are more Patient-like than Agent-like, the reassignment of the subject noun to the Patient role was easier in the inanimate-first than the animate-first sentences.

Subsequent studies suggested that animacy exerted its effects at an even earlier stage of processing (Trueswell, Tanenhaus & Garnsey, 1994; MacDonald, Pearlmutter, & Seidenberg, 1994). The authors of these studies suggested that, rather than aiding in thematic role reanalysis, animacy information was used immediately, leading readers to more readily interpret inanimate nouns as Patients during their initial parse of the sentence.

Animacy has also been shown to influence the processing of non-ambiguous non-canonical sentences. For example, Traxler, Morris and Seely (2002), in English, and Mak, Vonk and Schiefers (2002,2006) in Dutch, reported that the reading time penalty normally incurred in reading object relative clauses, e.g. The cop that the crook shot…, was eliminated when the initial argument was inanimate, e.g. The gun that the crook shot…. Likewise, Ferreira (2003) reported that participants were more accurate in assigning the correct thematic roles in simple passives (e.g. The cat/ball was chased by the dog) and object-cleft constructions (e.g. It was the cat/ball that the dog chased…) in inanimate-first than animate-first sentences (in the above examples, they more accurately indicated that the dog was the Agent of chased when the other argument was ball than when it was cat). Once again, this facilitation was interpreted as reflecting the easier assignment of inanimate than animate arguments to the Patient role, which was consistent with the role assigned to that argument by the syntactic structure.

In sum, there is clear evidence that animacy exerts an effect on sentence comprehension. While there remains some debate as to whether these effects occur during a second stage of processing (serial models, e.g. Rayner & Frazier, 1983; Rayner, Carlson, & Frazier, 1983; Ferreira & Clifton, 1986), or during a single stage of processing (constraint based models, e.g. Trueswell, Tanenhaus & Garnsey, 1994; MacDonald, Pearlmutter, & Seidenberg, 1994), within both types of architectures animacy has been suggested as exerting its effects through thematic role processing.

ERP studies of sentence processing

Evidence that animacy can influence sentence comprehension also comes from ERP studies examining both plausible sentences, where effects of animacy are seen primarily on the N400 component, and implausible selection-restriction violated sentences where effects are seen on both the N400 and P600 components.

Effects of animacy on processing plausible sentences: effects on the N400

The N400 is a negative-going waveform peaking at approximately 400msec post stimulus onset. Its amplitude is thought to reflect the semantic fit between that word’s conceptual features, the meaning of its preceding context and information stored at various grains within semantic memory (see Kutas & Federmeier, 2000 for a review). During sentence processing, a decrease in N400 amplitude has been generally interpreted as indicating that a word is easier to integrate into its preceding sentential context.

There is a growing body of work that has implicated the N400 as sensitive to thematic role processing, most often explored through manipulations of argument animacy. Several studies in English have reported a decrease in N400 amplitude to inanimate relative to animate subject nouns (Weckerly & Kutas, 1999; Kuperberg et al. 2003; Paczynski and Kuperberg, 2009; Nakano, Saron, & Swaab, 2010). Several of these authors (e.g. Weckerly & Kutas, 1999; Nakano, Saron, & Swaab, 2010) have interpreted this finding as indicating that animate arguments are more easily assigned the Agent role than inanimate nouns. Interestingly, this N400 effect is modulated by reading ability (Weckerly & Kutas, 1999) and working memory span (Nakano et al, 2010), with the effect of animacy being most pronounced for more proficient readers and those with higher working memory spans (e.g. high span readers showing a larger N400 amplitude to inanimate than animate subject nouns). Nakano et al. interpreted this as suggesting that only high span readers were able to make immediate use of noun animacy information to aid in thematic roles assignment.

In German, a verb-final language where animacy may play an even more critical role during initial thematic role assignment, Frisch and Schlesewsky (2001) found that, following an animate nominative (i.e. subject) argument, a second nominative argument evoked a smaller N400 effect (compared to correctly-marked accusative arguments) when it was inanimate than when it was animate.

The authors argued that this occurred because inanimate arguments were more easily assigned the thematic role of Patient than animate arguments, thus allowing a semantically coherent interpretation of the sentence. Converging evidence for the use of animacy information in guiding thematic role assignment in German, comes from two studies which reported a larger N400 to inanimate nominative arguments than animate ones following clause initial animate accusative arguments, (Frisch & Schlesewsky, 2001; Roehm, Schlesewsky, Bornkessel, Frisch, & Haider, 2004). The authors interpreted these results as indicating that, following an animate Patient, the language parse expected an ideal, i.e. animate, Agent.

This link between thematic roles, animacy and N400 modulation during verb-argument processing has been articulated within a recent model proposed by Bornkessel and Schlesewsky (2006). Within the extended Argument Dependency Model (eADM), it is proposed that several prominence hierarchies, including animacy, are used during a distinct stage of processing (compute prominence) to guide the assignment of a noun phrase to either the Actor or Undergoer role: the higher the prominence, the more likely the argument is to be interpreted as an Actor (Agent in our nomenclature). Importantly, Bornkessel-Schlesewsky and Schlesewsky (2008, p. 71) state that in English, animacy plays an important role in computing the prominence of arguments within a sentence, with mismatches between different prominence hierarchies being associated with an increase in N400 amplitude.

Effects of animacy on the processing of implausible sentences: the semantic P600 and its relationship with the N400

The P600 is a positive-going waveform that peaks between 600–1000msec post stimulus onset. It was originally described in association with syntactic processing in sentences (Osterhout & Holcomb, 1992; Hagoort, 1993). However, it is also evoked by animacy-based selection restriction violations between verbs and arguments, as well as by other types of highly semantically implausible words within sentences (reviewed by Kuperberg, 2007). For example, Kuperberg, Sitnikova, Caplan and Holcomb (2003) examined ERPs evoked by verbs in two types of implausible sentences: those that were implausible only with respect to real-world knowledge (e.g. Every morning at the boys would *plant…), and those in which the subject noun-phrase violated animacy-based selection restriction constraints of the verb (e.g. Every morning at breakfast the eggs would *eat…). Consistent with other research (e.g. Hagoort, Hald, Bastiaansen, & Petersson, 2004), real-world incongruities evoked the classic N400 effect. Surprisingly, the animacy-based selection restriction violations failed to evoke an N400 effect but rather produced a P600 effect. Similar so-called ‘semantic P600 effects’ to animacy-based selection restriction violations, sometimes in the absence of N400 effects, have also been reported by Kim and Osterhout (2005) in English and by Hoeks, Stowe and Doedens (2004) in Dutch. Although this effect has most frequently been reported on verbs, more recent studies have demonstrated a similar P600 effect on selection restriction violating direct object noun phrase (NP) arguments, at least under some experimental conditions (e.g. Nieuwland & Van Berkum, 2005, Kuperberg, Choi, Cohn, Paczynski & Jackendoff, 2010). Such semantic P600 effects may again be dependent on reading span, as Nakano et al. (2010) reported that only high span readers showed a P600 effect to animacy violations on verbs (e.g. The box *bit…), while low span readers showed an N400 effect. Importantly, animacy-based selection restriction violations are not always necessary or even sufficient to trigger a semantic P600 effect (e.g. Kolk, Chwilla, van Herten & Oor, 2003), and several other factors have been proposed that may bias towards the additional analysis reflected by this component, including sentential and discourse context and the experimental task (discussed by Kuperberg, 2007). Several frameworks have been proposed to explain why a P600 effect, often (although not always) with an attenuation of the N400 effect, is seen on animacy-based selection restriction violations.

One of the first accounts came from Herman Kolk and colleagues who proposed an Error Monitoring model (reviewed by van de Meerendonk, Kolk, Chwilla & Vissers, 2009). According to this model, the P600 is indicative of general executive error monitoring processes that can trigger a reanalysis of the entire preceding input. For semantic errors to evoke this effect, the meaning of the entire proposition must be highly implausible (see also van de Meerendonk, Kolk, Vissers, & Chwilla, 2010). Thus, the reason that animacy-based selection restriction violations evoke a P600 effect is that they are always highly implausible (eggs can never eat). According to Kolk’s account, monitoring, and thus the magnitude of the P600 effect is increased when such a highly implausible proposition conflicts with an alternative plausible, or even partially plausible interpretation (van Herten, Kolk & Chwilla, 2006). This alternative interpretation is generated by a heuristic mechanism that draws upon semantic relationships between content words (thus, although eggs can never eat, they can be eaten). The heuristically-generated alternative plausible interpretation also serves to attenuate the N400 to these selection restriction violations.

A slightly different account was proposed by Kim and Osterhout (2005) who suggested that the critical factor in triggering the P600 to these types of animacy selection restriction violations was not animacy or implausibility per se, but rather the more specific semantic ‘attraction’ between the verb and its argument(s). Semantic attraction was operationalized by the plausibility of a clause if the arguments had taken alternative thematic roles around the verb, i.e. by their semantic reversibility. Semantic attraction between verbs and arguments was also proposed to attenuate the N400 effect. Thus, the *The hearty meals were devouring… evokes a P600 effect but no N400 effect because the Agent (meals) serves as an excellent Patient for devour. In support of this hypothesis, Kim and Osterhout reported that an N400 effect but no P600 effect was evoked by verbs in non-attracted animacy-based selection restriction violation sentences like *The dusty tabletops were devouring… Note, however, that other studies have reported P600 effects to non-attracted semantic violations, including Kuperberg, Caplan, Sitnikova, Eddy, and Holcomb (2006), Kuperberg, Kreher, Sitnikova, Caplan, and Holcomb (2007), and Stroud and Phillips (2009).

Hagoort, Baggio and Willems (2009) offer a similar account to that of Kolk and colleagues as well as Kim and Osterhout. They suggest that the P600 effect is evoked when cues for a well-formed syntax are weaker than for a plausible semantic interpretation. Thus, for sentences such as For breakfast the eggs would eat… or The hearty meals were devouring…, a P600 is observed as the initial cues for semantic information being correct are very strong (the subject argument can take an alternative, plausible, role about the verb). On the other hand, for sentences such as The dusty table tops were devouring…, the subject argument and verb are only weakly related; thus the semantic cues are relatively weak compared to syntactic cues, resulting in an N400 effect.

In their eADM model, Bornkessel and Schlesewsky (2006) suggest that the semantic P600 results from increased costs at later stages of processing: a generalized-mapping step (in which numerous sources of information, including real-world knowledge, are utilized in fully integrating the argument into its syntactic structure) and/or a well-formedness evaluation/repair processing step, which evaluates the overall well-formedness of the proposition, and which may be modulated by experimental task. The generalized-mapping step is dependent on earlier phases of processing. For example, if sufficient processing costs are incurred during either the compute prominence step (see above), the generalized-mapping processing step is blocked. However, the well-formedness processing step always occurs.

Finally, in her 2007 review of studies describing the semantic P600 effect, Kuperberg noted that no single factor could account for the full range of findings, and proposed a hybrid, dynamic framework in which she suggested that, in parallel with a full, syntactically-driven combinatorial analysis, there are continuous attempts to match semantic relationships in the input against semantic associative and featural relationships, stored at various grains of representation within semantic memory, i.e. semantic memory-based processing. She argued that N400 attenuation was driven primarily by matches between the input and these stored relationships (see Kutas and Federmeier, 1999) – semantic memory-based analysis. She suggested that the P600 effect to semantic violations reflected additional analysis or reanalysis, triggered by conflict between the implausibility of the full (syntactically-determined) proposition, and the output of this semantic memory-based analysis. Thus, similar to Kolk, Kuperberg (2007) emphasizes the degree of implausibility of the syntactically-determined propositional representation as a factor that influences the amplitude of the P600 to animacy selection-restriction violations. However, unlike the accounts described above, Kuperberg (2007) also raised the possibility that the competing non-syntactic analysis may not necessarily reflect heuristic attempts to come up with alternative plausible interpretations (by switching around thematic roles). Rather, it may simply reflecting a matching or resonance between incoming semantic relationships and those that are expected and stored – the semantic memory-based analysis. Evidence for this is that neither the P600 nor the N400 evoked to verbs are necessarily modulated by the semantic-thematic reversibility of a sentence (see Kuperberg et al., 2006, 2007; Stroud & Phillips, 2009).

The Current Study

The current study had two main aims. The first was to examine the mechanism by which animacy influences the processing of direct object arguments in plausible, syntactically simple, active English sentences: 1) through links to syntactic ordering, 2) through inherent differences between animate and inanimate entities in their salience/lexico-semantic accessibility, 3) through links to specific thematic roles. In Experiment 1, we manipulated the animacy of the direct object argument while keeping the thematic role constant. This allowed us to determine whether animate nouns are generally more salient/lexico-semantically accessible, and thus easier to process than inanimate nouns, independent of syntactic position or thematic role assignment. In Experiment 2, we kept the animacy of the direct object argument constant while manipulating its thematic role assigned by the verb: the Patient role, which biases towards inanimate nouns, or the Experiencer role, which requires animate ones. This allowed us to determine whether animacy impacts processing though its relationship with thematic roles, or through biases based on the linear ordering of arguments within a sentence.

Our second aim was to examine the effects of animacy and thematic roles on the processing of direct object arguments that violated the selection restrictions (SRs) of their preceding verbs. As discussed in the preceding section, several theoretical frameworks have attempted to explain the so-called semantic P600. Each makes somewhat different predictions for when a semantic P600 is evoked, and when an N400 is attenuated. We were particularly interested in examining how thematic role reversibility modulated the P600 and N400 in selection restriction violated sentences, with the level of reversibility being varied across the two experiments.

Experiment 1

Introduction

As discussed in the General Introduction, both behavioral and ERP comprehension studies have generally attributed the effects of animacy to a mapping between animacy and specific thematic roles. In contrast, production studies provide evidence for a direct mapping between animacy features and the linear ordering of syntactic constituents, i.e. animate items are most easily mapped directly onto the initial sentence position, which is usually the subject argument in simple English sentences. However, a third possibility has been implied in some theories of language production: that animate nouns generally more accessible or salient than inanimate ones, “invariant across contexts.” (Prat-Sala and Branigan (2000, p. 169). Thus, rather than being driven by animacy hierarchy based biases, the “animate first” preference seen in production studies might reflect a more general facilitation for the processing of animate than inanimate nouns. The first aim of Experiment 1 was to contrast this third hypothesis with the first two.

In order to address this first aim, we constructed a set of plausible, canonical active English sentences with Agent-Patient verbs, i.e. ones in which the subject (always animate) was assigned the Agent role while the direct object argument was assigned the Patient role by the verb. We manipulated the animacy of the direct object argument, creating two types of plausible sentences, those with animate arguments (condition 1; see Table 1) and those with inanimate arguments (condition 2). If animacy influences processing either through a mapping of the animacy hierarchy onto linear order (less animate arguments should appear later in the sentence) or through specific thematic role entailments (inanimate arguments are more Patient-like), this would predict that the amplitude of the N400 evoked by inanimate Patient direct object arguments (…plowed the meadow…) should be smaller than that evoked by animate ones (e.g. …penalized the laborer…). If, on the other hand, animate items are generally easier to process because of their intrinsic salience or lexico-semantic accessibility, regardless of their syntactic position in a sentence, this would predict the opposite pattern of N400 modulation: inanimate direct object arguments should evoke a larger N400 than that evoked by inanimate ones.

Table 1.

Example sentences in Experiment 1, including semantic similarity value (SSV) based on Latent Semantic Analysis, and plausibility ratings.

Condition	Example	SSV	Plausibility ratings
(1) Non-violating, animate direct object Patient	At headquarters the manager interviewed the applicant for thirty minutes.	0.22	5.35
(2) Non-violating, inanimate direct object Patient	At headquarters the manager perused the application for thirty minutes.	0.21	5.54
(3) Selection restriction violating, animate direct object Patient	At headquarters the manager perused the*applicant for thirty minutes.	0.22	2.15
(4) Selection restriction violating, inanimate direct object Patient	At headquarters the manager interviewed the *application for thirty minutes.	0.20	2.24

Open in a new tab

Critical direct object arguments, to which ERP responses were measured, are bolded in the example sentences.

The second aim of this study was to examine the effects of animacy and thematic reversibility on the processing of direct object arguments in implausible sentences in which the direct object mismatched the selection restrictions of their preceding verbs (see Table 1, conditions 3 and 4). As noted in the General Introduction, animacy-based selection restriction violations on verbs (…*eggs would eat…) failed to evoke an N400 effect, despite being highly semantically anomalous (Kuperberg et al. 2003, 2006, 2007; Kim & Osterhout, 2005 Experiment 1, Hoeks et al. 2004), instead evoking a robust P600 effect. Several possibilities have been discussed for this pattern of findings.

With regard to the reduced N400 effect in these selection restriction violated sentences, one possibility is that the animacy of the argument plays a key role: Kuperberg et al. (2003) suggested that when a reader’s course-grained expectations of the animacy of an argument are violated, further, full semantic memory-based analysis, reflected by full N400 modulation, can be ‘switched off’. Kim and Osterhout (2005)’s proposal suggests that the N400 effect would be reduced only when an argument can play an alternative thematic role around a specific verb. Similarly, Kolk and colleagues (Van Herten, Chwilla, & Kolk, 2006) propose that the N400 can be attenuated when a heuristic mechanism drawing upon semantic relationships between content words comes up with an alternative plausible interpretation, e.g. when sentences are at least partially semantically reversible (see also Hagoort et al. 2009). The eADM (Bornkessel & Schlesewsky, 2006) offers two way in which animacy may affect processing is selection restriction violating sentences: 1) animate arguments assigned to the Patient role should incur a processing cost during the compute prominence step, blocking the plausibility step (similar to Kuperberg’s proposal) or 2) animate nouns, being plausible Agents, would not lead to processing costs during the initial plausibility step, as this step is non-combinatorial (similar to heuristic processing in Kolk’s proposal).

In the current experiment, all these proposals make similar predictions with regard to modulation of the N400: they all predict an attenuation of any N400 effect to violating (versus non-violating) animate direct object arguments, but the presence of an N400 effect to violating (versus non-violating) inanimate direct object arguments. Based on animacy alone, animate direct object arguments are less expected than inanimate direct object arguments; thus a full semantic memory-based analysis between the argument, verb and context is blocked, and no N400 effect would be observed, despite the violation of the selection restrictions of the preceding verb (Kuperberg et al. 2003, Kuperberg 2007, Bornkessel & Schlesewsky, 2006). Similarly, because implausibilities on animate direct objects are partially reversible (e.g. *…plowed the laborer…/…plowed by the laborer…), a plausibility heuristic (van Herten, Kolk & Chwilla, 2006, Bornkessel & Schlesewsky, 2006), a semantic-attraction mechanism (Kim & Osterhout, 2005) or multiple semantic cues (Hagoort et al. 2009) could come up with a plausible interpretation, once again reducing the N400 effect to these violating (versus non-violating) arguments. Note that this is not true for sentences with selection restriction violating inanimate direct object arguments, which are not semantically reversible (e.g. *…penalized the meadow…/*…penalized by the meadow…).

Turning to the P600, the semantic-attraction framework (Kim & Osterhout, 2005) predicts that only animate selection restriction violating direct object arguments should evoke a P600 effect as only these can plausibly occupy an alternative thematic role around the verb. A similar prediction is made by Hagoort et al. (2009). The Error Monitoring framework predicts that a P600 effect would be evoked by both animate and inanimate selection restriction violating arguments because they both result in highly implausible propositions (van de Meerendork et al., 2009). However, it also predicts that the magnitude of the P600 effect will be larger to violations on animate than inanimate direct object arguments because the highly implausible selection restriction violating animate direct object arguments would conflict with the partially plausible output of the plausibility heuristic (e.g. plowed by the laborer…). On the other hand, eADM predicts a P600 effect to both animate and inanimate selection restriction violations, again because neither is well-formed, leading to increased costs during the well-formedness step. Unlike the other proposals, the eADM does not predict the well-formedness step to be affected by thematic role reversibility.

Kuperberg (2007) also predicts that both violating animate and inanimate direct object arguments should lead to a P600 effect. This is because in both cases, the full syntactic-based interpretation for both types of violations would be equally (im)plausible. This, combined with a plausibility judgment task and the presence of some semantically constraining context, should lead to conflict with matches in a semantic memory-based analysis, and thus result in continued combinatorial processes, indexed by the P600. Kuperberg (2007) outlined different accounts of how animacy might modulate the P600. If animacy functions as part of a separate combinatorial ‘thematic’ stream that assigns thematic roles purely on the basis of animacy, independently of syntax, this might lead to increased conflict and a larger P600 to the animate (versus inanimate) selection restriction violations. If, however, the main source of conflict arises simply from matches within a semantic memory-based stream, then the P600 should not be modulated by animacy-based thematic assignment within the sentences, and the amplitude of the P600 would be equal to animate and inanimate selection restriction violations. Kuperberg (2007) also stressed that the P600 is not necessarily modulated by semantic reversibility (for other evidence suggesting that the P600 is not modulated by the reversibility of thematic roles, see Kuperberg et al. 2006 and Kuperberg et al. 2007, Stroud and Phillips, 2009). A summary of all predictions is presented in Table 2.

Table 2.

Experiment 1. Predictions made by different proposals regarding the modulation of the N400 and P600 by animacy and selection restriction violations.

Non-violating Patient Arguments
Proposal	N400
Animate nouns more salient/accessible than Inanimate nouns	Animate < Inanimate
Animacy as a proxy for Thematic Role	Inanimate < Animate
Animacy Hierarchy/Linear Position	Inanimate < Animate
Violating versus Non-violating Patient Arguments
Proposal	N400	P600
Kuperberg et al.	Animate: No N400 effect due to‘switching off’ of additional semantic analysis when animacy mismatches expectations based on selection restrictions of the verb as well as either thematic role or linear position. Inanimate: N400 effect.[	Animate: P600 effect. Inanimate: P600 effect. Effects driven by implausibility/impossibility of final interpretation conflicting with some match in semantic memory-based analysis, with no effect of thematic reversibility.
Kolk et al.	Animate: No N400 effect due to heuristic mechanisms leading to partially plausible interpretation (partial thematic reversibility). Inanimate: N400 effect.	Animate: P600 effect (larger) Inanimate: P600 effect (smaller) Effects driven by reanalysis triggered by detection of highly implausible interpretation (error). Larger effect for animate arguments driven by greater difficulty in resolving detected error due to heuristic mechanisms leading to partially plausible interpretation (partial thematic reversibility).
Kim & Osterhout	Animate: No N400 effect due to semantic attraction (reversibility) between verb and argument. Inanimate: N400 effect.	Animate: P600 effect driven by semantic attraction (reversibility) between verb and argument. Inanimate: No P600 effect.
eADM	Animate: No N400 effect due to processing cost incurred during compute prominence step blocking initial plausibility step. Inanimate: N400 effect.	Animate: P600 effect. Inanimate: P600 effect. Effects driven by real-world implausibility detection during well-formedness step.
Hagoort et al.	Animate: No N400 effect due to semantic cues being stronger than syntactic cues (partial thematic reversibility). Inanimate: N400 effect.	Animate: P600 effect driven by semantic cues being stronger than syntactic cues (partial thematic reversibility). Inanimate: No P600 effect.

Open in a new tab

Methods

Development and Pre-testing of Materials

Two lists of transitive Agent-Patient verbs were generated – one list with 166 verbs that selected for animate direct object arguments (animate-selecting) and the other with 166 verbs that selected for inanimate direct objects (inanimate-selecting). Initial assessment of animacy restrictions was based on authors’ intuitions. 166 pairs of sentence-stems were then created from each of these verbs: each consisted of a short introductory context, an animate subject argument, an animate-selecting or an inanimate-selecting verb that was semantically congruous with the context and subject, and a definite article.

To confirm verb selection restriction biases, a completion experiment was carried out in which these sentence stems, counterbalanced across two lists, were presented in random order to 20 Tufts undergraduate students (10 per list) who were asked to write down the word that was the most likely continuation of the sentence stem, with explicit instructions that this word need not be the final one of the sentence. Two independent raters then judged the animacy of each participant’s response: whether the word described an animate, inanimate or ambiguous entity. Twenty-two scenarios in which more than 10% of continuations failed to conform to a priori animacy expectations were rejected, leaving 144 acceptable scenarios.

For each of these 144 pairs of sentence-stems, a plausible animate or inanimate direct object was chosen to complete each sentence and an additional 2–4 words were added so that the direct object never fell on the sentence final word. Nouns which appeared as responses for a given sentence in the completion task (described above) were not used as direct objects to minimize N400 attenuation in the plausible conditions due to high word expectancy. Each direct object appeared only once with the exception of four animate nouns and two inanimate nouns which appeared twice.

Implausible sentences with selection restriction violations were then created by swapping animate and inanimate direct objects within a pair of sentences, i.e. for a given sentence pair, the plausible animate direct object following an animate-selecting verb became the implausible direct object following the inanimate-selecting verb, and vice versa. This generated 144 sets of sentences, each with four conditions: (1) Normal with animate direct objects (following animate-selecting verbs); (2) Normal with inanimate direct objects (following inanimate-selecting verbs); (3) Selection restriction violating with animate direct objects (following inanimate-selecting verbs), and (4) Selection restriction violating with inanimate direct objects (following animate-selecting verbs). Each set of four sentences is henceforth referred to as a ‘scenario’. Example sentences for one scenario are shown in Table 1. A larger sample of example stimuli is available for download at the following URL: http://kuperberglab.nmr.mgh.harvard.edu/materials.htm

The frequency (Kucera & Francis, 1967) of the animate direct objects was not significantly different from that of the inanimate direct objects, (44.48 vs. 49.88, per million), t(143) = .68, p = 0.41. The animate direct objects were, on average, one letter longer than the inanimate direct objects (7.5 vs. 6.6), t(358) = 5.89, p < 0.001, and the mean concreteness (Pavio, Yuille & Madigan, 1968; Coltheart, 1981) of the animate direct objects was slightly, though significantly, less than that of the inanimate direct objects, (510.8 vs. 525.1), t(132) = 2.56, p < 0.05.

One-hundred-and-twenty-eight pairs of normal and semantically implausible filler sentences were then created. In each normal filler sentence, an animate subject was followed by a transitive main verb, followed by a direct object argument and then by an adjunct argument that contained a noun that was related to the preceding sentence context, e.g. The hooligan hid his identity with a mask and sun glasses. Anomalous versions were created by pseudorandomly substituting critical noun on the adjuncts from other fillers such that they were somewhat semantically associated but were highly implausible, e.g. The hooligan hid his identity with a *pistol and sun glasses.

Four counterbalanced lists were created, each containing 144 experimental scenarios (36 sentences in each of the four sentence types) and 128 filler sentences (64 normal and 64 implausible). A given scenario was presented only once within a list and, across lists, each scenario appeared in all four conditions. The order of scenarios was pseudo-randomized within each list such that no more than two sentences in a row were of the same condition.

Plausibility ratings and LSA Values

To gather plausibility ratings, all sentences (experimental and fillers) were presented, counterbalanced across the four lists, up to the point of the critical noun to 16 undergraduate participants (four per list) who did not participate in any other portion of the study. Three periods after the critical nouns were used to indicate that the sentences could continue after the critical noun. Participants were told that they were seeing ‘beginnings of sentences’ and were asked to give ratings from 1 through 7, with 1 indicating that the sentence fragment described something that would be very unlikely to occur in the real world and 7 indicating that the sentence fragment described something that would be very likely to occur in the real world. Several examples were given but subjects were told to go with their first instincts and that there were no right or wrong answers.

The mean plausibility ratings of the four types of sentences are shown in Table 1. A 2×2 (Animacy X Violation) ANOVA revealed no main effect Animacy on either the subjects, F(1,15)=1.17, p=0.29, nor items, F(1,143)=0.40, p=0.53, analysis. However, as expected, a significant main effect of Violation was found on both subjects, F(1,15)= 305.2, p < 0.0001, and items, F(1,143)= 1525.1, p < 0.0001, analysis. There was no interaction between Animacy and Violation on the subjects analysis, F(1,15)=2.96, p=0.11, but the interaction approached significance on the items analysis, F(1,143)=2.88, p=0.08. Follow-up pairwise comparisons indicated that this was due to normal inanimate sentences being slightly, although not significantly, more plausible than normal animate sentences, (F(1,143)=2.52, p=0.12), but there being no difference in plausibility between violated animate and inanimate violated sentences (F(1,143)<1, p=0.438). Filler items showed a main effect of Violation on both subjects, F(1,15)=700.4, p <0.0001, and items, F(1,127)= 564.6, p <0.0001, analysis.

To confirm authors’ intuitions that critical direct object arguments were semantically related to the preceding context we calculated semantic similarity values (SSVs) using Latent Semantic Analysis (LSA), (Landauer and Dumais 1997; Landauer et al. 1998; available on the internet at http://lsa.colorado.edu), for all four conditions, using term-to-term comparison between the critical word and its preceding context. Scenarios for which SSVs could not be obtained (14) were removed from analysis. The mean SSA values of the four types of sentences are shown in Table 1. These values were all at or above 0.2, similar to the SSV between ‘mop’ and ‘broom’ (0.21). Although very similar to each other (0.20 – 0.22), a 2(Animacy) X 2(Violation) ANOVA indicated a main effect of Violation, F(1,129)=23.43, p < 0.001, as well as a significant Animacy X Violation interaction, F(1,129)=6.33, p <0.05. Follow-up pairwise comparisons indicated that the mean SSV value for the inanimate violating direct objects was significantly smaller than for the non-violating inanimate direct objects, F(1,129)=20.42, p<0.001; the mean SSV value for the animate violating direct objects was also smaller than for the non-violating animate direct objects but this difference only approached significance, F(1,129)=2.91, p=0.09. There was no significant difference in SSV between plausible animate and inanimate direct object arguments, F(1,129)<1, p=0.54, or implausible animate and inanimate arguments, F(1,129)=1.52, p=0.22.

ERP Experiment

Participants

Twenty-four undergraduate native English speakers from Tufts University were recruited. Participants had normal or corrected-to-normal vision, were not taking psychoactive medications, had no learning disability, no history of neurological or psychiatric disorders and had not learned languages other than English before the age of 5. All were right handed as assessed through a modified version of the Edinburgh handedness inventory (Oldfield, 1971). Written consent was obtained from all subjects before participation according to the established guidelines of Tufts University. Participants were paid for their participation.

Stimulus presentation

Each participant was randomly assigned to one of the four lists (six participants per list). Participants sat comfortably in a dimly lit room, separate from the experimenter. Stimuli were presented on a video monitor. Each trial was preceded by the word “READY” and participants pressed a button to initiate the trial. Each trial began with a fixation-cross displayed in the center of the screen for 500ms followed by 100ms blank screen. The sentence was then presented one word at a time. Each word was presented for 450ms with 100ms blank screen interstimulus interval, with the exception of the sentence-final word which was followed by a blank screen for 750ms before a “?” that cued participants to press one of two buttons on a response pad, depending on whether they judged the sentence to make sense or not. Participants were asked to refrain from blinking or moving during the trial. Stimuli were presented in eight blocks of 40 trials, with short breaks between blocks. Participants viewed twelve practice trials prior to the start of the experiment.

Electrophysiological Recording

Twenty-nine active tin electrodes were held in place on the scalp by an elastic cap (Electro-Cap International, Inc., Eaton, OH), see Figure 1. Electrodes were also placed below the left eye and at the outer canthus of the right eye to monitor vertical and horizontal eye movements, and on the left and right mastoids. Impedance was kept below 2.5 kΩfor all scalp and mastoid electrode sites and below 10 k for the two eye channels. The EEG signal was amplified by an Isolated Bioelectric Amplifier System Model HandW-32/BA (SA Instrumentation Co., San Diego, CA) with a bandpass of 0.01 to 40 Hz and was continuously sampled at 200 Hz by an analogue-to-digital converter. The stimuli and behavioral responses were simultaneously monitored with a digitizing computer.

Electrode montage. Analyses of variance were conducted at midline, medial and lateral electrode columns shown (see Methods).

Data Analysis

ERPs were formed by off-line averaging of artifact-free trials, time-locked to the onset of critical words in each sentence. Only trials that were correctly classified were averaged. Each trial was baselined to a 100ms pre-stimulus onset. However, when a significant effect of experimental manipulation was found within the 0–100msec time window (noted in the Results and likely due to artifact), trials were rebaselined to 0–100ms post-stimulus onset, in order to reduce the possibility that any later effects were driven by artifact. We report results in three time windows of interest: 350–500ms, 500–700ms, and 700–900ms. We also carried out analyses across the 100–250ms time window to determine if any early effects could explain any main effects or interactions within these later time windows. Modulation of average ERPs within each of these time windows was examined using analyses of variance (ANOVAs) for repeated measures at each of three electrode columns (see Figure 1). The Midline column had five levels of electrode sites along the anterior-posterior distribution (AP Distribution) of the scalp (FPz, Fz, Cz, Pz, Oz), the Medial column had three levels for AP Distribution and two levels for Hemisphere (FC1/FC2, C3/C4, CP1/CP2), and the Lateral column had four levels for AP Distribution and two levels for Hemisphere (F3/F4, FC5/FC6, CP5/CP6, P3/P4). At each electrode column, we carried out 2 (Animacy) x 2 (Violation) ANOVAs, with AP Distribution and Hemisphere (for the medial and lateral columns) as additional within-subject factors. These were followed up with planned simple effects ANOVAs, contrasting each sentence type with one another. Significance was set at alpha equal to 0.05. A Geisser-Greenhouse correction was applied to all repeated measures with more than one degree of freedom, for which original degrees of freedom and corrected probability levels are reported.

Results

Behavioral Results

On average, participants’ acceptability classifications during the ERP experiment matched a priori classifications 90% of the time, see Table 3. There was no main effect of Violation or Animacy, nor a Violation X Animacy interaction, all Fs < 2.79, p > 0.12.

Table 3.

Experiment 1. Percentage of trials for each type of sentence in which participants’ acceptability judgments during the ERP experiment matched the originally defined conditions.

Condition	Mean (SD) % Accuracy
(1) Non-violating, animate direct object Patient	85.3 (8.9)
(2) Non-violating, inanimate direct object Patient	89.3 (6.3)
(3) Selection restriction violating, animate direct object Patient	90.1 (17.7)
(4) Selection restriction violating, inanimate direct object Patient	93.4 (10.6)

Open in a new tab

ERP Results: The Critical Direct Object Argument

Early effects

In the 0–100ms interval, there was a significant Violation X AP-Distribution interaction at the midline, F(4,92)=6.80, p<0.01, medial, F(2,46)=10.60, p <0.01, and lateral, F(3,69)=10.28, p<0.01, electrode columns. The effects were due to a significantly greater positivity over anterior electrode sites to selection restriction violating direct object arguments, compared to non-violating ones. Because these effects appear to begin prior to stimulus onset, they may have been driven by artifact. Subsequent analyses were therefore carried out using a post-stimulus 0–100ms baseline. Using the post-stimulus baseline, in the 100–250ms time window (the P2), there were no significant main effects of either Violation or Animacy at any electrode columns, all Fs < 1. Similarly, the Violation X AP-Distribution, Animacy X AP-Distribution, and Violation X Animacy, interactions were all non-significant, all Fs < 1.22.

350–500ms

Within the 350–500ms time window, there was a main effect of Violation at the midline (F(1,23) = 12.11, p < 0.01), medial (F(1,23)= 15.76, p < 0.001) and lateral (F(1,23)= 13.22, p < 0.01) electrode columns, due to selection restriction violating direct objects evoking larger N400s than non-violating direct objects. This effect was larger at posterior than anterior midline sites, as reflected by a Violation X AP-distribution interaction at the Midline electrode column, F(4,92)= 3.50, p < 0.0001. There were no main effects of Animacy or interactions between Animacy and spatial variables, with the exception of an Animacy X Hemisphere interaction at the Lateral column (follow-ups comparing ERPs animate and inanimate direct objects at each level of hemisphere, however, failed to reveal significant effects.)

Of most interest, a Violation X Animacy interaction was significant at the Midline (F(1,23)= 5.76, p < 0.01) and Medial (F(1,23)= 4.62, p < 0.05) columns, and approached significance at the Lateral column (F(1,23)= 3.40, p = 0.08). See Figure 2 for a comparison of all four conditions along the Midline electrode column. Planned pairwise comparisons were carried out to further examine the source of this interaction. The inanimate violating direct objects evoked a significantly larger N400 than inanimate normal direct objects at the midline (F(1,23)=16.64, p < 0.001), medial (F(1,23)=18.91, p < 0.001) and lateral (F(1,23)=16.87, p < 0.001) columns, see Figure 3B, but there was no significant effect of Violation on the animate direct objects (all Fs <1.48, ps > 0.24), see Figure 3A. Additionally, selection restriction violating inanimate arguments evoked significantly a larger N400 than selection restriction violating animate arguments at the midline (F(1,23)=9.88, p<0.001), medial (F(1,23)=7.12, p<0.05), and lateral (F(1,23)=6.88, p<0.05) columns. Finally, while there was no significant main effect of Animacy between normal inanimate and animate direct objects (Fs < 1.943, ps > 0.15), there was a significant Animacy by AP Distribution effect along the midline, F(4,92)=3.35, p < 0.05, due to normal animate direct object arguments evoking a larger N400 amplitude at centro-posterior sites, than to normal inanimate direct object arguments.

Experiment 1. ERPs evoked by direct object arguments in all four experimental conditions along the midline. Solid black line indicates normal animate nouns, dotted blue line indicates normal inanimate nouns, dashed red line indicates selection violating animate nouns, and dash-dotted green line indicates selection restriction violating inanimate nouns. The plots are shown using a 0–100ms post-stimulus baseline (consistent with how the statistics were carried out). The modulation across the four conditions, however, was the same using a prestimulus −25ms–0ms baseline, see http://www.nmr.mgh.harvard.edu/kuperberglab/materials.htm.

Experiment 1. ERPs evoked by direct object arguments that were either animate (Figure 3A) or inanimate (Figure 3B). ERPs to normal non-violating direct object arguments are shown in a solid line, while those to selection restriction violating animate noun arguments are shown in a dotted line. Voltage maps show average voltage difference between normal and selection restriction violating arguments in the N400 (left) and P600 (right) time windows

500–700ms

There were no main effects of Violation, Animacy or interactions involving Violation or Animacy, with the exception of a main effect of Animacy that approached significance at the Lateral column (F(1,23)=3.84, p=0.06).

700–900ms

There were no main effects or interactions involving Animacy. However, the main effect of Violation was significant at the midline (F(1,23)=15.92, p < 0.001), medial (F(1,23)=14.47, p<0.001), and lateral (F(1,23)=14.26, p<0.001), due to a P600 effect to the violating (vs. normal) direct objects, see Figures 2 and 3. This effect had a more posterior than anterior distribution, as reflected by interactions between Violation and AP Distribution that reached significant at the Medial (F(1,23) = 8.74, p < 0.01) and Lateral (F(1,23) = 10.96, p < 0.01) columns and approached significance at the Midline (F(1,23) = 3.04, p = 0.06) column.

Discussion

Experiment 1 examined ERPs to direct object NPs that were either animate or inanimate, and that either violated or did not violate the selection restrictions of their preceding verbs. Within the N400 time window, we observed a significant Animacy by Violation interaction. Pair-wise comparisons indicated that this interaction arose because of a significant N400 effect to selection restriction violating (versus non-violating) inanimate, but not animate, direct object arguments. In the plausible sentences, the N400 (at posterior sites) was slightly larger to the animate than inanimate direct objects, whereas in the violating sentences, it was larger to the inanimate than animate direct objects. Within the P600 time window we observed only a main effect of Violation: selection restriction violating (versus non-violating) direct objects evoked a P600 effect, regardless of animacy.

Before discussing these findings in more detail, we consider two alternative explanations for the pattern of findings within the N400 time window. The first is that it was primarily driven by differences between conditions in the semantic relatedness across content words, rather than animacy per se. This seems unlikely. Although there were differences in the semantic relatedness between the critical word and its preceding contexts (as indexed by average SSVs on an LSA) across the four conditions, these differences were very small and, as is clear from Table 1, the pattern of SSVs across the four conditions did not pattern with the modulation of the N400. Specifically, the non-violating inanimate arguments were less semantically related to their contexts than the non-violating or selection restriction violating animate arguments, predicting the largest N400 in this condition. In fact, the smallest N400 was seen to these inanimate non-violated arguments.

A second alternative possibility is that the results were driven by inherent differences in the selection restrictions of animate-selecting and inanimate-selecting verbs. Selection restrictions, beyond those of animacy, may have led to different degrees of match/mismatch with the semantic features of the direct objects. Inanimate-selecting verbs not only select for inanimate direct objects but will also tend to select for more fine-grained semantic features. For example, the verb sip will not only select for an inanimate entity such as brandy, but will also select for more specific features such is a beverage. In contrast, an animate selecting verb such as thank will tend to select only for animate direct objects, without selecting for further fine-grained features. Thus, during processing, encountering an inanimate-selecting verb, particularly after a fairly constraining context, might have allowed for a more accurate prediction of the direct object. This account can explain why the N400 was smaller to non-violating inanimate than animate direct object arguments: inanimate direct objects are more likely to have matched all the selection restrictions of their preceding verb. However, it does not easily explain why there was an enhanced N400 to selection restriction violating inanimate, relative to violating animate direct objects. Encountering an animate direct object after an inanimate-selecting verb would violate more selection restrictions than encountering an inanimate direct object after an animate-selecting verb. This might have a larger N400 to animate than inanimate violating direct objects – the opposite pattern of findings to those observed here.

Effects of animacy on the processing of plausible sentences

As noted in the introduction, the first aim of Experiment 1 was to contrast the hypothesis that animate nouns are more salient/lexico-semantically accessible independent of syntactic position or whether animacy impacts processing through a mapping of the animacy hierarchy onto linear order or specific thematic role entailments. Our results clearly rule out this first alternative. Unlike previous studies, which have demonstrated a larger N400 amplitude to inanimate subject arguments (Weckerly & Kutas, 1999; Kuperberg et al. 2003; Paczynski and Kuperberg, 2009; Nakano et al., 2010), in the current experiment we found an effect in the opposite direction: inanimate direct object arguments evoked an attenuated N400 compare to animate ones. This result is incompatible with the proposal that animate nouns are generally more salient/lexico-semantically accessible. Instead, our results indicate that semantic processing of inanimate direct objects was facilitated relative to animate direct objects. This facilitation may have been driven either by animacy hierarchy based biases on linear order, or through a link between animacy and thematic role. As noted in the introduction, Experiment 1 cannot distinguish between these two possibilities.

In previous studies examining the effect of animacy on subject NPs, any effects on the N400 have primarily been observed in good comprehenders (Weckerly and Kutas, 1999; Nakano et al., 2010). Although we did not test reading spans in our participants, it is possible that, as Nakano et al. suggests, only high span readers are able to use animacy information online. Additionally, Nakano et al. reported that the N400 effect to inanimate versus animate subject NPs had an anterior distribution, similar to the distribution of the analogous effect reported by Paczynski and Kuperberg (2009), and, for poor comprehenders by Weckerly and Kutas (1999). This contrasts with the results of the present experiment that showed that the N400 effect to non-violated animate versus inanimate direct objects was largest over central sites. However, the distribution of the effect found in the present study is in keeping Weckerly and Kutas’ (1999) finding that, for good comprehenders, the animacy N400 effect on the subject was centrally distributed, a finding similar to Kuperberg et al. (2003). Future studies will be necessary to determine the functional significance of these differences in scalp distribution across experiments.

Effects of animacy on processing selection-restriction violated sentences

As summarized above, we saw an interaction between Animacy and Violation on the N400, but only a main effect of Violation on the P600. As discussed in the Introduction, different accounts have placed different emphases on the triggers of the P600 effect, and its relationship to the N400 effect in semantically anomalous sentences that violate verb-argument structure (see Table 2). Below we discuss the present findings in relation to these frameworks.

The findings are somewhat consistent with Kolk’s Monitoring hypothesis that, as discussed in the Introduction, predicted the attenuated N400 effect we observed to the violating (versus non-violating) animate direct argument NPs, but not to the violating (versus non-violating) inanimate direct object arguments. This framework also predicted the P600 effect we observed to both animate and inanimate selection restriction violating arguments as they are both highly implausible². However, according to the Monitoring hypothesis, this P600 effect should be larger to reversible animate selection restriction violations than to irreversible inanimate selection restriction violations. While visual inspection of the waveform suggests a possible trend in this direction, the interaction between Animacy and Violation did not approach significance (all Fs < 1). Nonetheless, it is possible that the lack of a significant interaction between Animacy and Violation in this experiment was due to the fact that thematic role reversibility was only partial: although the selection restriction violating animate direct object arguments could have plausibly acted as Agents for their preceding verbs (e.g. *…plowed the laborer…/… plowed by the laborer…), the subjects of these verbs could not plausibly be assigned the Patient role (e.g. * …the farmer was plowed by the laborer…). We investigate this possibility in Experiment 2.

The pattern of findings is inconsistent with both the ‘semantic-attraction’ hypothesis (Kim & Osterhout, 2005), as well as the proposal by Hagoort and colleagues (2009). Although both these frameworks predicted the attenuated N400 effect and P600 effect we observed to semantically-attracted animate violating (versus non-violating) direct object arguments, they failed to predict the P600 effect seen to the non-semantically-attracted inanimate violating (versus non-violating) arguments.

The findings are consistent with the eADM framework. With regards to the attenuation of the N400 effect for selection restriction violating animate arguments, the eADM offers to possible explanations. First, animate arguments (high prominence) assigned Patient role (low prominence) could have incur processing cost during the compute prominence step, thus blocking initial evaluation of plausibility. Alternatively, because the plausibility processing step within the eADM is non-combinatorial, no processing costs were incurred as the violating animate arguments could serve as plausible Agents of the preceding verb. In contrast, violating (versus non-violating) inanimate arguments did incur costs during initial plausibility evaluation, as these could not play a plausible alternative role about the verb, leading to the N400 effect to these types of violation. According to the eADM, the P600 effect to both inanimate and animate selection restriction violations reflected increased costs during the later well-formedness step.

Finally, the findings are consistent with Kuperberg (2003, 2007). Similar to Kuperberg et al. (2003), we suggest that mismatches between animacy-based expectations and the input ‘switched off’ a full, more detailed (fine-grained) semantic memory-based analysis between verb and argument, leading to the reduced N400 effect to selection restriction violating (versus non-violating) animate direct object arguments. In the present study, selection restriction violating animate arguments mismatched expectations that inanimate arguments should be encountered in the direct object position. This switched off a more detailed semantic memory-based analysis between the selection restrictions of the verb and the semantic features of the argument, and thus the N400 effect to the violating (versus non-violating) animate arguments was attenuated (this attenuation may have been further driven by the semantic associations between the context and the critical word). We elaborate further on this idea in the General Discussion.

Kuperberg (2007) also suggested that the P600 effect is triggered by the conflict between the propositionally impossible arguments (determined by a fully combinatorial analysis with a syntactic assignment of thematic roles) and a match between the context and stored information within semantic memory (semantic memory-based processing). This conflict is most likely to occur given a semantically constrained context and when participants explicitly judge plausibility (Kuperberg, 2007). In the present study, we suggest that the P600 effect to both types of violations was triggered by the conflict between the propositionally impossible direct objects (it is impossible to plow a laborer or to penalize a meadow) and the relatively close semantic association between these arguments and their preceding content words in the context. Importantly, Kuperberg argued that the semantic P600 effect is not directly driven by the reversibility of thematic roles: in previous studies (Kuperberg et al., 2006, 2007), selection restriction violations in non-reversible sentences (e.g. The trumpets would *courtesy…) evoked P600 effects of similar amplitude as those occurring in reversible sentences (e.g. The eggs would *eat…). This idea is supported by the present findings: the amplitude of the P600 effect to the violating (versus non-violating) direct objects was not modulated by their semantic or thematic reversibility.