Electrophysiological correlates of Complement Coercion

Abstract

This study examined the electrophysiological correlates of complement coercion. Event-related potentials (ERPs) were measured as participants read and made acceptability judgments about plausible coerced sentences, plausible non-coerced sentences, and highly implausible animacy violated sentences (“The journalist began/wrote/astonished the article before his coffee break”). Relative to non-coerced complement nouns, the coerced nouns evoked an N400 effect. This effect was not modulated by the number of possible activities implied by the coerced nouns (e.g. began reading the article; began writing the article), and did not differ in either magnitude or scalp distribution from the N400 effect evoked by the animacy violated complement nouns. We suggest that the N400 modulation to both coerced and the animacy violated complement nouns reflected different types of mismatches between the semantic restrictions of the verb and the semantic properties of the incoming complement noun. This is consistent with models holding that a verb’s semantic argument structure is represented and stored at a distinct level from its syntactic argument structure. Unlike the coerced complement noun, the animacy violated nouns also evoked a robust P600 effect, which may have been triggered by the judgments of the highly implausible (syntactically-determined) meanings of the animacy violated propositions. No additional ERP effects were seen in the coerced sentences until the sentence-final word which, relative to sentence-final words in the non-coerced sentences, evoked a sustained anteriorly-distributed positivity. We suggest that this effect reflected delayed attempts to retrieve the specific event(s) implied by coerced complement nouns.

1. Introduction

Although it is widely acknowledged that sentences are built compositionally, there is debate over whether their meaning is determined entirely through combining individual lexical items using syntactic rule systems (strong compositionality, e.g. Montague (1970)), or whether it is possible to construct new meaning that is invisible to syntactic structure. One piece of evidence for the latter hypothesis comes from a phenomenon known as complement coercion, exemplified by the sentence, “The man began the book.” (Pustejovsky, 1995; Jackendoff, 1997). Under strong compositionality, verbs like “begin”, which semantically select for an activity, should be unable to take arguments denoting entities such as “book”. Nonetheless, we interpret such sentences as plausible. One account of this phenomenon is that eventive verbs, like “begin”, “finish”, and “enjoy”, when paired with a complement NP entity (“book”), ‘type-shift’ that complement into an event to meet the demands of the argument structure: “book” is taken to mean ‘do something with the book’ (Pustejovsky, 1995). An alternative account is that the unexpressed meaning (‘do something’) is inserted into the meaning of the sentence in order to satisfy the selectional restrictions of the verb (Jackendoff, 1997). As a result, “begin the book” is understood as “begin ‘doing something with’ the book”, for a review see Pylkkanen & McElree (2006).

It has been argued that this process of coercion should incur a processing cost. And, indeed, behavioral studies using self-paced reading (McElree, Traxler, Pickering, Seely, & Jackendoff, 2001; Traxler, Pickering, & McElree, 2002), eye-tracking (Traxler et al., 2002; Scheepers, Mohr, Keller, & Lapata, 2004; Pickering, McElree, & Traxler, 2005; Traxler, McElree, Williams, & Pickering, 2005; McElree, Frisson, & Pickering, 2006; Frisson & McElree, 2008) and speed accuracy trade-off (McElree, Pylkkänen, Pickering, & Traxler, 2006) procedures report that coerced complement NPs are harder to process than non-coerced NPs, even when they are matched for plausibility (McElree et al., 2001; Traxler et al., 2002). Importantly, these costs are not incurred on all NPs following eventive verbs, such as those denoting activities (e.g. “begin the work”), but rather appear to arise from the particular combination of an eventive verb with an entity NP (Traxler et al., 2002).

One possibility is that the increased processing costs associated with coerced versus noncoerced NPs reflect a second step of developing a full interpretation – the filling out or retrieval of the details of “do something” on the basis of real-world knowledge and context (e.g. “reading a book”, “writing a book”) (Pustejovsky, 1995; Jackendoff, 1997). This account, however, seems unlikely because such costs are still present when the activity is explicitly provided in the immediately preceding discourse context (Traxler et al., 2005).¹ Finally, these costs are unlikely to be due to the process of selecting between retrieved alternative activities (resolving ambiguity): in a recent eye-movement study, Frisson and McElree (2008) reported equal costs for processing complement NPs that were highly constrained for a single interpretation (e.g. “The student started the essay”, usually interpreted as “writing”) and for complement NPs that were less constrained (e.g. “The director started the script”, which could be interpreted as depicting “reading”, “directing, “filming”, or other actions). Similarly, Scheepers, Keller & Lapata (2008) report data from a visual world paradigm that support a serial account of coercion in which a single dominant interpretation, rather than multiple interpretations, is pursued during processing. Together, these observations have been interpreted as supporting the view that the processing cost of coercion reflects the building of a complex non-syntactic representation of the complement.

In a recent study, Pylkkänen & McElree (2007) used magneto-encephalography (MEG) to contrast activity to coerced complements, non-coerced complements and complements which violated the selection restrictions (animacy-based) of the preceding verb (e.g. “The pilot amazed the plane…”). The animacy violated complement NPs were associated with a significant MEG effect, relative to the non-coerced complements, from 300–400ms, localizing to a left temporal source. In contrast, the coerced NPs were associated with a significant anterior midline effect, relative to both the non-coerced and animacy violated complement NPs, between 350–500ms, which localized to a ventromedial prefrontal source. Since there was no difference in activity at this source between the non-coerced and animacy violated complements, the authors interpreted these observations as evidence that complement coercion engages neurocognitive processes distinct from those engaged in detecting lexical mismatch, semantic predictability or semantic implausibility. A similar anterior midline effect has been described by the same group in association with other forms of coercion where it has been interpreted as a more general neural signature of enriched composition (Brennan & Pylkkänen, 2008; Pylkkänen, Martin, McElree, & Smart, 2009).

Taken together, this series of studies provides compelling evidence that complement coercion entails a behavioral and neural cost. The present study used another technique – event-related potentials (ERPs) – to seek converging evidence and further information on the neurocognitive processes engaged during complement coercion.

In ERP studies, the component that has been most closely linked to semantic processing is the N400 – a negative-going waveform observed between approximately 300–500msec following words that are incongruous (versus congruous) with their preceding word (Rugg, 1985; Bentin, McCarthy, & Wood, 1985), sentence (Kutas & Hillyard, 1980, 1984; Hagoort, Hald, Bastiaansen, & Petersson, 2004) or discourse (Van Berkum, Hagoort, & Brown, 1999) contexts. Although it has sometimes been assumed that the N400 during sentence processing is a reflection of semantic anomaly or implausibility per se, it has been recognized for some time that an N400 effect is evoked by words that are plausible but relatively unexpected with respect to their preceding context (Kutas & Hillyard, 1984), and that the N400 amplitude is modulated by a host of factors that can influence plausibility but that can be theoretically dissociated from this construct. These include fine-grained associative relationships between individual words (Van Petten, 1993), coarser-grained categorical relationships between entities sharing common features (Federmeier & Kutas, 1999), and, in the case of verb-argument structures, selection restriction-based relationships (Friederici & Frisch, 2000), and animacy-based relationships (Weckerly & Kutas, 1999; Frisch & Schlesewsky, 2001; Paczynski & Kuperberg, 2009).

The N400 may reflect three-way dynamic semantic mapping or matching processes between (a) semantic relationships, stored at various grains of representation within semantic memory, (b) semantic relationships within the context prior to a critical word, (c) the semantic features of an incoming critical word (Kutas, Van Petten, & Kluender, 2006; Kutas & Federmeier, 2000; Lau, Phillips, & Poeppel, 2008). Predictions as to the nature of the incoming critical word may be generated before it has been presentd (see DeLong, Urbach, & Kutas, 2005; Van Berkum, Brown, Zwitserlood, Kooijman, & Hagoort, 2005; Federmeier, 2007) or matching processes may take place only once the critical word has been encountered (‘semantic integration’) (P.J. Holcomb, 1993; Hagoort, 2005), see Van Berkum (in press) for a recent discussion. In the case of verb-argument processing, we have contrasted these types of ‘semantic memory-based processes’ with independent, but interacting, ‘combinatorial’ stream(s) of processing which come up with full propositional representations or interpretations that may be plausible or implausible with respect to real-world knowledge (Kuperberg, 2007). Such combinatorial process(es) may be syntactically driven (for example, assigning thematic roles to arguments) or semantically driven (for example, type-shifting, coercive processing).

The present study compared the ERP responses to coerced, non-coerced and animacy violated complement NPs. Similar to the MEG study by Pylkkänen and McElree (2007), participants carried out an acceptability judgment task as they viewed these sentences. Based on this MEG study, we predicted that, relative to non-coerced NPs, coerced NPs would evoke increased activity within the N400 time window (300–500ms). Although this effect could theoretically reflect the cost of semantic combination and interpretation – the interpretative process of type-shifting the meaning of the entity NP to an event –an alternative possibility given the discussion of the N400 above, is that it might simply reflect the mismatch between the semantic properties of the eventive verb, and the semantic properties of the incoming NP (an entity). Based on previous studies demonstrating an N400 effect to object NPs that violate the selection restriction properties of their preceding verbs (Friederici & Frisch, 2000; Hoeks, Stowe, & Doedens, 2004; Kim & Osterhout, 2005; van Herten, Chwilla, & Kolk, 2006), we predicted that, relative to the non-coerced NP, the animacy violated complement NPs would also produce an N400 effect. Rather than reflecting the implausibility of the proposition produced by syntactically and thematically combining the animacy violated NP with the verb, this N400 might, once again, reflect a mismatch between the animate selection restrictions of the preceding verb and the semantic features of the direct object NP argument (an inanimate entity).

One question was whether the N400 effects evoked by the coerced and animacy violated complement NPs, each relative to non-coerced NPs, would have distinct scalp distributions. As noted above, this was the case in the MEG study. Although, the spatial resolution of ERPs is inferior to that of MEG because of its sensitivity to the effects of intervening tissues which smear the EEG patterns measured on the scalp (Geisler & Gerstein, 1961; Delucchi, Garoutte, & Aird, 1962; Cooper, Winter, Crow, & Walter, 1965), differences in the spatial distribution on the surface of the scalp of the N400 effects evoked across different experimental conditions have been well documented in other studies, (e.g Kounios & Holcomb, 1992; P. J. Holcomb, Kounios, Anderson, & West, 1999; West & Holcomb, 2002; Sitnikova, West, Kuperberg, & Holcomb, 2006). Based on the MEG findings, we therefore hypothesized that the N400 effect to coerced (vs. non-coerced) NPs would show a more anterior distribution than to that evoked by the animacy violated (vs. non-coerced) NPs.

A second question was whether the neural costs associated with processing coerced complements would be sensitive to any ambiguity in their interpretation. To examine this possibility, we followed Frisson & McElree (2008) by carrying out separate ratings that were used to categorize the coerced sentences into those with dominant interpretations (e.g. “The student started the essay…”) and those with multiple possible interpretations (e.g. “The director started the script…”). If any N400 effect evoked by the coerced (vs. non-coerced) NPs reflected a process of selecting from multiple possible activities, then it should be modulated by this parameter.

A third question was whether ERPs would be modulated across conditions in the time-window following the N400, specifically in the P600 – a centro-parietally distributed positive-going component observed between approximately 500 and 900ms. The P600 has been classically associated with syntactic anomalies and ambiguities (Osterhout & Holcomb, 1992; Hagoort, Brown, & Groothusen, 1993), but, more recently, has been described in association with certain types of semantic anomalies under particular circumstances (see Kuperberg (2007) for a review). Unlike the N400 effect, the semantic P600 effect is not usually seen to words that are simply unpredictable in plausible sentences, but primarily to words that are semantically very implausible or impossible – often violations of animacy, particularly when the verb or wider context is semantically constraining and when participants are asked to make acceptability judgments and may reflect continued combinatorial processes as participants attempt to make sense of the sentences. In the present study, we therefore, predicted that a P600 effect would be observed to the animacy violated complement NPs, but not to the coerced complements (each relative to the non-coerced complements).

Finally, in addition to examining neural activity at the point of the complement NP, we also examined ERPs to subsequent words in the sentence. Self-paced reading and eye-tracking studies have reported behavioral costs of coercion at one or two words following the complement (McElree et al., 2001; Traxler et al., 2002; Traxler et al., 2005; Frisson & McElree, 2008), but have not examined processing beyond this point. Pylkkänen and McElree’s (2007) MEG study focused on neural costs at the complement itself but did not examine past this point. We therefore aimed to determine whether the neural costs of coercion are primarily met at the complement NP or whether additional costs are incurred at subsequent words, particularly on sentence-final words where ‘wrap-up’ of sentence meaning is thought to take place.

2. Methods

2.1 Development and Pre-testing of Materials

630 sentences (210 scenarios, each with three sentence types) were developed and expanded from 70 scenario triplets originally used by Pylkkänen & McElree (2007). In this original set of 70 triplets, a sentence in a given triplet contained one of three types of verbs – non-coercive (entity-selecting), coercive (event-selecting) and object-experiencer – followed by the same inanimate critical complement NP that rendered the sentences non-coerced, coerced and animacy violated respectively (Table 1). The direct object complement was followed by between 3 and 5 words, followed by the sentence-final word. The original set was then expanded three-fold in order to counterbalance the identical critical nouns across three lists as follows: for each of the 70 original triplet scenarios, three additional sentences were created using the same verbs but new subjects and objects. This replacement of subjects and object NPs was repeated once more, yielding 9 sentences per scenario: three sentences with a given coercive verb, three with a given non-coercive verb and three with a given object-experiencer verb, but with the same subject and object NPs appearing only once with each type of verb. In half of the original scenarios used by Pylkkänen & McElree (2007) in their MEG study, the clause containing the critical verb and complement noun was embedded within a relative clause (e.g. “The staff was shocked that the journalist began the article before his coffee break…”). This varied the length of the sentences and the position at which the coercion/animacy violations were introduced, thereby introducing variety into the material. To keep the overall experimental stimuli similar the MEG study, we maintained this manipulation in the present stimulus set. However, we did not expect that it should systematically influence ERP responses to any of our main manipulations and therefore did not include this as a factor in our main series of analyses.

Table 1.

Stimuli examples and parameters for each sentence type.

Sentence type (example)	Naturalness/ plausibility of entire sentence	LSA: at point of CN	Cloze probability of CN	Number of letters of verb	Frequency of verb
Non-Coerced
The journalist wrote the article before his coffee break.	3.8 (0.5)	0.17 (0.14)	0.14 (0.27)	6.3 (1.7)	59.4 (109.05)
Coerced
The journalist began the article before his coffee break.	3.8 (0.6)	0.14 (0.10)	0.06 (0.15)	7.0 (1.6)	102.5 (95.79)
Animacy violated
The journalist astonished the article before his coffee break.	N/A	0.12 (0.10)	---	7.8 (1.3)	15.7 (17.64)

Means are shown with standard deviations in brackets. In the examples of each sentence type, the verb is shown in italics and the critical complement NP (to which ERPs were measured) is underlined. CN: Complement Noun. LSA: Latent Semantic Analysis.

We then carried out a norming study just of this initial set of coerced and non-coerced sentences (randomized across three lists, excluding the animacy violated sentences) in order to screen out the more unnatural or implausible sentences. Twelve Tufts University undergraduates (4 for each list), who did not participate in the ERP study, and who gave written informed consent before participation, judged the likelihood that they might encounter each sentence (presented as a whole) in the real world on a scale from 1 to 5. ‘1’ indicated that the sentence did not make sense and/or it sounded unnatural. ‘5’ indicated that the sentence made sense and seemed natural. Based on these initial ratings, 30 scenarios with an average rating of less than 3 were discarded, leaving a final set of 180 scenarios.

To generate the final stimulus lists used in the ERP experiment, the appropriate animacy violated sentences for each scenario were reinserted. This yielded 540 sentences in total, counterbalanced across three lists, each list with 180 sentences, 60 of each sentence type. Across all lists, each subject and NP combination was paired with all three types of verbs (i.e. seen in all three sentence types), but within any given list, the same combination (pairing) of subject and complement NPs was not viewed with more than one type of verb (i.e. in more than one condition) except on two occasions. 84 of the 180 scenarios contained a relative clause. In each list, test sentences were pseudorandomized amongst 158 filler sentences, 50 that contained semantic incongruities. The incongruous filler sentences contained a variety of different types of incongruities ranging from animacy violations (e.g. “The whistler trained the chapstick so his lips wouldn’t chap.”), other types of selection restriction violations (e.g. “The congressman smoldered the meeting until the food ran out.”), and pragmatic real-world incongruities (e.g. “The girl smiled at the parking meter to make sure she had enough time.”). Six of the 158 fillers contained object-experiencer verbs. Thus, in total, each list contained 338 sentences and approximately 33% of these contained semantic incongruities.

This final stimulus set was further characterized in terms of several metrics and by conducting a cloze study. In the cloze study, the coercive and non-coercive sentence frames (without the critical words) were presented on a computer to 30 undergraduates at Tufts University (10 per list) who did not participate in the ERP experiment or any other rating study. Participants gave written, informed consent before participation and were asked to type in the most likely next word in the sentence.

The results of all norming and stimulus characterizations are shown in Table 1. Coercive verbs were more frequent than the non-coercive verbs (t(335) = 3.86, p < .01) that were, in turn, more frequent than the object-experiencer verbs (t(296) = 4.71, p < .01). The object-experiencer verbs were slightly longer than the coercive verbs (t(358) = 5.57, p < .01) that were longer than the non-coercive verbs (t(358) = 3.84, p < .001). A Latent Semantic Analysis (LSA, a measure of lexical co-occurrence) – calculated using pairwise comparisons of Semantic Similarity Values (SSV) on a term-by-term between each complement noun and all content words that preceded it (Landauer & Dumais, 1997; Landauer, Foltz, & Dumais, 1998) – yielded very slightly greater values in the non-coerced complement nouns (0.17) than the coerced sentences (0.14), t(358) = 2.06, p < .04. Cloze probabilities of both the coerced and non-coerced complement nouns were low (less than 15%), but greater in the non-coerced sentences (0.14) than the coerced sentences (0.06), by subjects: t(29) = 8.25, p < .01; by items: t(179) = 4.62, p <. 01. Plausibility ratings of the entire coerced and non-coerced sentences, gathered during the development of stimuli as described above, did not differ significantly, by subjects, t(11) = 0.312, p = .76 or by items, t(418) = 1.34, p = .18.

2.1.1 Subdivision of coerced sentences

Following Frisson & McElree (2008), we carried out an additional rating study to examine the precise interpretations of the activities implied by the coerced NPs in each sentence. Thirty undergraduates from Tufts University (10 per counterbalanced list) were given the coerced sentences used in the ERP experiment with a blank space in between the verb and the complement NP (e.g. “The journalist began _______ the article before his coffee break.”). Participants were asked to fill in the blank with one or two words describing the activity that best fit their interpretation of the sentence. In order to categorize the coerced sentences into those with a strongly preferred versus weakly preferred interpretation, we identified the number of unique interpretations for each sentences. Sentences in which the same interpretation was given in 80% or more of all responses were categorized as having a strongly preferred (dominant) interpretation (N=89); all others (70% or less) were categorized as having weakly preferred interpretations (N=91). We also examined three other measures identified by Frisson & McElree (2008): 1) the number of different verbs generated for the sentence; 2) the number of unique interpretations generated for the sentence; 3) the ratio of the most frequent interpretation of the sentence to the second-most frequent interpretation. These data are shown in Table 2.

Table 2.

Parameters used to subdivide the coerced sentences into those with one dominant interpretation and those with multiple possible interpretations.

Measure	Dominant Interpretation N = 89	Multiple interpretations N=91
% use of the dominant interpretation	90.1% (80%–100%)	54.3% (30%–70%)
Number of different verbs generated	3.5 (1–7)	5.56 (2–9)
Average number of different interpretations	1.82 (1–3)	3.97 (2–8)
Ratio of the most frequent interpretation to the 2nd most frequent interpretation	13:1	2.4:1

The range of values for each measure is shown in brackets.

2.2 ERP Experiment

2.2.1 Participants

26 (9 male, 17 female) undergraduates from Tufts University, aged 18 to 22 (mean: 19.6), initially participated, and 24 subjects (9 male, 15 female, mean age 19.5) were included in the final analysis (see below). All selected participants were right-handed, native American English speakers, who had not learned to speak another language fluently before the age of 5. Participants were not taking any medication, had normal or corrected-to-normal vision, no history of a reading disability or of neurological or psychiatric disorders. Written consent was obtained from all subjects before participation according to the established guidelines of Tufts University.

2.2.2 Stimulus presentation

Participants were randomly assigned to one of the three counterbalanced lists. They sat in a comfortable chair in a dimly lit room separate from the experimenter and computers. Sentences were presented word by word on a computer monitor. Each trial (one sentence) began with a fixation point (“+”) at the center of the screen for 450ms, followed by a 100ms blank screen, followed by the first word of the sentence. Each word appeared on the screen for 450ms with an interstimulus interval (ISI) of 100ms separating the words. The final word of each sentence appeared with a period and was followed by a 750 ms blank-screen interval and then a “?”. This cue remained on the screen until the participant made his/her response, at which point the next trial started. The participant’s task was to decide whether or not each sentence made sense by pressing one of two buttons on a response box with either the left or right thumb (counterbalanced across participants). Participants were instructed to wait until the “?” cue before responding. This delayed response was designed to reduce any contamination of the ERP waveform by response-sensitive components such as the P300 (Donchin & Coles, 1988). After subjects registered their responses, the word “BEGIN” was displayed until they pressed a button to begin the next trial. Each participant was given twelve practice trials at the beginning of the experiment.

2.2.3 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 5 kΩ for all scalp electrode sites, 2.5 kΩ for mastoid electrode sites and below 10 kΩ 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 by a digitizing computer.

Figure 1.

Electrode montage. Simple effects ANOVAs based on a priori hypotheses were conducted at each of the three columns shown (midline, medial and lateral).

2.2.4 Data Analysis

Accuracy was computed as the percentage of correct responses. A correct response was a judgment of acceptable for the non-coerced and coerced sentences and unacceptable for the animacy violated sentences.

Averaged ERPs, time-locked to target words, were formed off-line from trials free of ocular and muscular artifact and were quantified by calculating the mean amplitude (relative to a 100 ms prestimulus baseline) in time windows of interest. Because of our a priori hypotheses, we proceeded straight to planned pair-wise comparisons between conditions of interest (coerced vs. non-coerced vs. animacy violated). We conducted ANOVAs at a midline column, containing 5 electrode sites, and two lateral columns, each containing 3 (medial column) or 4 (lateral column) electrodes (see Figure 1). Within-subject factors were Sentence Type, Anterior Posterior (AP) Distribution (with the number of levels corresponding to electrode sites along the AP axis) and, for the lateral analyses, Hemisphere (2 levels).

Our series of ANOVAs yielded statistical information about differences in the distribution of effects along the AP axis of the scalp and across the two hemispheres. Main effects and interactions involving Sentence Type, which were of most theoretical interest, were followed up using appropriate simple effects ANOVAs. The N400 at the complement noun was quantified from 300–500ms and the P600 was quantified between 600–900ms (to avoid overlap with the N400 effect). For each of the three words following the complement noun, 400–600ms time-windows were used for analyses. For the sentence-final word, a 300–700ms time-window was used for analyses.

In all these ANOVAs, the Geisser-Greenhouse correction was used in cases with more than one degree of freedom in the numerator (Greenhouse & Geisser, 1959) to protect against Type 1 error resulting from violations of sphericity. In these cases, we report the original degrees of freedom with the corrected p value. In all analyses, a significance level of alpha = .05 was used as, in all cases, we were testing a priori hypotheses. Linearly interpolated voltage maps showing the scalp distribution of differences in ERPs elicited by critical words between the three conditions within the time windows of interest were produced by the EEGLab program (MatLab).

3. Results

3.1 Behavioral Data

One participant was excluded on the basis of a clear behavioral response bias. One other participant was excluded due to ERP artifact. Of the 24 remaining participants, accuracy on the acceptability judgment task was high: coerced sentences and non-coerced sentences were correctly identified as acceptable on 91.3 % (SD=5.4) and 92.1% (SD=4.1) of trials, respectively. Animacy violated sentences were correctly identified as unacceptable on 95.8% (SD=5.4) of trials. Accuracy judgments significantly differed between sentence types, F(2,46) = 8.78, p < .01) due to more accurate judgments to the animacy violated sentences than to both the coerced and non-coerced sentences (p < .001 and p < .01 for pairwise comparisons, respectively)². There was no significant difference in accuracy between the coerced and non-coerced sentences (p > 0.1).

3.2 ERP Data

Across the 24 participants included in the analysis, approximately 11% of the critical trials were rejected due to artifact. All ERP analyses reported are based on correctly answered trials. However, analyses were repeated including all responses and yielded qualitatively similar findings.

3.2.1 ERPs on the Complement Noun

Grand average ERPs elicited by the complement nouns for all sentence types at selected electrode sites are presented in Figure 2. There were no significant differences in the N1-P2 complex over the first 250ms after the onset of the critical word across conditions (no main effects or interactions involving sentence type, ps > 0.05).

Figure 2.

Grand averaged waveforms to complement NPs in all three sentence types. Voltage maps comparing ERPs evoked by the complement noun between 300–500 ms – the N400 effect (left) and between 600–900 ms – the P600 effect (right). Note that, to best illustrate the full scalp distribution of the ERP effects, the scale used for the voltage maps of the N400 effects (left) was half that used for the voltage maps of the P600 effect (right). Grand averaged waveforms to a cloze-matched dataset demonstrate similar results and are available at http://www.nmr.mgh.harvard.edu/kuperberglab/publications/materials/ComplementCoercion_suppl_figures.pdf.

3.2.1.1 The N400

A significantly more negative N400 was observed to both the coerced and animacy violated complement nouns than non-coerced nouns (Table 3). The amplitude of the N400 to the coerced and animacy violated complement nouns did not differ significantly from each other (Table 3, Figure 2). N400 effects to both coerced and violated (relative to non-coerced) complement nouns were fairly widely distributed across the scalp (no interactions between sentence type and AP distribution).³ An analysis that included a subset of the main stimulus set in which LSA and cloze probability at the point of the complement noun were all fully matched between the coerced and non-coerced sentence types revealed a similar set of findings (see note to Table 3 and http://www.nmr.mgh.harvard.edu/kuperberglab/publications/materials/ComplementCoercion_suppl_figures.pdf). A second analysis that included the presence or absence of a relative clause (in just under 50% of stimuli) confirmed that this did not interact significantly with sentence type at any electrode column, all ps > 0.05.

Table 3.

ANOVAs comparing ERPs to complement nouns across the N400 (300–500ms) and the P600 (600–900ms) time windows.

		N400: 300–500 ms		P600: 600–900 ms
	Effect	F (degrees of freedom)	P value	F (degrees of freedom)	P value
A. Coerced vs. Non-Coerced
Midline	ST	5.71 (1, 23)	0.025	0.09 (1, 23)	0.763
	ST × AP	1.02 (4, 92)	0.362	2.32 (4, 92)	0.114
Medial	ST	5.85 (1, 23)	0.024	0.44 (1, 23)	0.514
	ST × H	0.64 (1, 23)	0.433	0.87 (1, 23)	0.361
	ST × AP	0.71 (2, 46)	0.492	0.55 (2, 46)	0.566
Lateral	ST	3.84 (1, 23)	0.062	0.38 (1, 23)	0.545
	ST × H	0.97 (1, 23)	0.335	1.35 (1, 23)	0.257
	ST × AP	0.31 (3, 69)	0.651	1.70 (2, 69)	0.195
B. Animacy violated vs. Non-Coerced
Midline	ST	8.72 (1, 23)	0.007	4.37 (1, 23)	0.048
	ST × AP	2.29 (4, 92)	0.105	7.15 (4, 92)	0.004
Medial	ST	9.23 (1, 23)	0.006	4.35 (1, 23)	0.048
	ST × H	13.61 (1, 23)	0.001	5.98 (1, 23)	0.023
	ST × AP	1.24 (2, 46)	0.294	7.05 (2, 46)	0.009
Lateral	ST	7.82 (1, 23)	0.010	4.64 (1, 23)	0.042
	ST × H	3.19 (1, 23)	0.087	0.25 (1, 23)	0.619
	ST × AP	0.36 (2, 69)	0.656	14.02 (2, 69)	0.0002
C. Coerced vs. Animacy violated
Midline	ST	2.41 (1, 23)	0.135	4.49 (1, 23)	0.045
	ST × AP	0.61 (4, 92)	0.562	5.82 (4, 92)	0.008
Medial	ST	2.41 (1, 23)	0.135	5.66 (1, 23)	0.026
	ST × H	2.37 (1, 23)	0.138	0.25 (1, 23)	0.625
	ST × AP	0.65 (2, 46)	0.493	2.85 (2, 46)	0.075
Lateral	ST	1.59 (1, 23)	0.219	6.07 (1, 23)	0.022
	ST × H	0.54 (1, 23)	0.470	0.25 (1, 23)	0.624
	ST × AP	0.40 (2, 69)	0.647	7.74 (2, 69)	0.002

ST – Main effect of Sentence Type

ST × H – Sentence Type by Hemisphere interaction

ST × AP – Sentence Type by Anterior Posterior Distribution

Note: A subanalysis performed on a cloze-matched dataset produced comparable results, with significant differences in N400 modulation between the coerced and non-coerced NPs at medial and lateral columns, ps < 0.05, but no significant differences in the N400 evoked by coerced and animacy violated NPs, ps > 0.1). Please refer to the supplementary figure demonstrating these effects at: http://www.nmr.mgh.harvard.edu/kuperberglab/publications/materials/ComplementCoercion_suppl_figures.pdf

3.2.1.2 600–900: The P600

The P600 was larger to animacy violated than to both non-coerced complement nouns (Table 3B, Figure 2 right bottom) and coerced complement nouns (Table 3C). This P600 effect was generally more positive posteriorly than anteriorly and had a slight left-lateralized distribution (main effects of sentence type and/or sentence type by AP distribution or hemisphere interactions at all columns). In contrast, there was no difference in the amplitude of the P600 evoked by coerced and non-coerced complement nouns (no significant main effects or interactions involving sentence type in any column, Table 3, Figure 2, right top).

3.2.2 ERPs to words following the Complement NP

In comparing the coerced and non-coerced sentences, the waveforms evoked by each of the three words that followed the complement nouns did not diverge from one another (Figure 3, Table 4). In contrast, the positivity evoked by the animacy violated complement nouns, relative to the other two conditions, remained evident at the first word following the complement noun (Figure 3, Table 4, CN+1). At the second and third words following the animacy violated complement noun, however, the polarity of this effect reversed such that the waveforms to these words were more negative than in the coerced and non-coerced sentences (Figure 3, Table 4, CN+2, CN+3).

Figure 3.

Grand averaged waveforms at one, two and three words after the complement noun (CN), comparing all three sentence types.

Table 4.

ANOVAs comparing ERPs to 1, 2, and 3 words following the complement noun (400–600 ms time window).

		CN+1: 400–600 ms		CN+2: 400–600 ms		CN+3: 400–600 ms
	Effect	F (degrees of freedom)	P value	F (degrees of freedom)	P value	F (degrees of freedom)	P value
A. Coerced vs. Non-Coerced
Midline	ST	0.74 (1, 23)	0.399	1.89 (1, 23)	0.182	0.32 (1, 23)	0.575
	ST × AP	3.18 (4, 92)	0.052	0.54 (4, 92)	0.593	0.68 (4, 92)	0.680
Medial	ST	0.37 (1, 23)	0.548	1.64 (1, 23)	0.213	0.17 (1, 23)	0.686
	ST × H	0.14 (1, 23)	0.715	0.95 (1, 23)	0.340	0.02 (1, 23)	0.900
	ST × AP	2.16 (2, 46)	0.136	1.98 (2, 46)	0.160	0.06 (2, 46)	0.912
Lateral	ST	0.48 (1, 23)	0.494	1.28 (1, 23)	0.270	0.24 (1, 23)	0.627
	ST × H	0.77 (1, 23)	0.389	3.60 (1, 23)	0.071	0.39 (1, 23)	0.541
	ST × AP	2.84 (3, 69)	0.084	0.94 (3, 69)	0.405	0.71 (3, 69)	0.581
B. Animacy violated vs. Non-Coerced
Midline	ST	1.03 (1, 23)	0.321	3.96 (1, 23)	0.059	5.20 (1, 23)	0.032
	ST × AP	7.89 (4, 92)	0.002	3.13 (4, 92)	0.064	1.83 (4, 92)	0.164
Medial	ST	4.16 (1, 23)	0.053	4.51 (1, 23)	0.045	4.34 (1, 23)	0.049
	ST × H	3.99 (1, 23)	0.058	0.95 (1, 23)	0.340	0.08 (1, 23)	0.784
	ST × AP	10.13 (2, 46)	0.0003	5.68 (2, 46)	0.018	0.74 (2, 46)	0.435
Lateral	ST	2.24 (1, 23)	0.148	2.34 (1, 23)	0.140	3.75 (1, 23)	0.065
	ST × H	4.65 (1, 23)	0.042	0.95 (1, 23)	0.341	1.90 (1, 23)	0.182
	ST × AP	6.93 (3, 69)	0.005	4.08 (3, 69)	0.034	0.07 (3, 69)	0.902
C. Coerced vs. Animacy violated
Midline	ST	2.60 (1, 23)	0.121	6.64 (1, 23)	0.017	9.34 (1, 23)	0.006
	ST × AP	5.38 (4, 92)	0.006	5.11 (4, 92)	0.011	1.41 (4, 92)	0.254
Medial	ST	4.29 (1, 23)	0.049	6.99 (1, 23)	0.014	10.84 (1, 23)	0.003
	ST × H	3.19 (1, 23)	0.087	4.65 (1, 23)	0.042	0.02 (1, 23)	0.883
	ST × AP	6.08 (2, 46)	0.008	8.68 (2, 46)	0.001	0.46 (2, 46)	0.611
Lateral	ST	2.98 (1, 23)	0.098	4.26 (1, 23)	0.050	8.70 (1, 23)	0.007
	ST × H	2.10 (1, 23)	0.161	10.86 (1, 23)	0.003	1.02 (1, 23)	0.322
	ST × AP	0.90 (3, 69)	0.380	8.84 (3, 69)	0.001	0.69 (3, 69)	0.470

ST – Main effect of Sentence Type

ST × H – Sentence Type by Hemisphere interaction

ST × AP – Sentence Type by Anterior Posterior Distribution

Note: All analyses presented in this table were repeated with the waveforms time-locked to the complement NP itself (rather than to prestimulus baselines of each successive word). These analyses revealed a similar overall pattern of findings.

At the sentence-final word (SFW), the coerced sentences evoked a more positive (less negative) waveform than the non-coerced sentences at anterior sites and the animacy violated sentences evoked a more negative waveform than the non-coerced sentences at posterior sites, see Figure 4, Table 5.

Figure 4.

Grand-averaged waveforms and voltage maps to the sentence-final word comparing ERPs to (A) Coerced vs. Non-coerced sentence types and (B) Animacy violated vs. Non-coerced sentence types. Note that the scale used for the voltage maps in this figure is different from that used to illustrate the ERP effects at the point of the critical noun in Figure 2.

Table 5.

ANOVAs comparing ERPs to the sentence-final word for the 300–700 ms time window.

	SFW: 300–700ms
	Effect	F (degrees of freedom)	P value
A. Coerced vs. Non-Coerced
Midline	ST	3.60 (1, 23)	0.071
	ST × AP	5.95 (4, 92)	0.003
Medial	ST	3.93 (1, 23)	0.060
	ST × H	0.64 (1, 23)	0.433
	ST × AP	5.72 (2, 46)	0.013
Lateral	ST	3.27 (1, 23)	0.083
	ST × H	3.18 (1, 23)	0.088
	ST × AP	3.40 (3, 69)	0.057
B. Animacy violated vs. Non-Coerced
Midline	ST	4.13 (1, 23)	0.054
	ST × AP	6.49 (4, 92)	0.002
Medial	ST	3.75 (1, 23)	0.065
	ST × H	4.84 (1, 23)	0.038
	ST × AP	6.73 (2, 46)	0.009
Lateral	ST	2.52 (1, 23)	0.126
	ST × H	5.96 (1, 23)	0.023
	ST × AP	11.39 (3, 69)	0.0001
C. Coerced vs. Animacy violated
Midline	ST	16.78 (1, 23)	0.0004
	ST × AP	3.93 (4, 92)	0.028
Medial	ST	20.47 (1, 23)	0.0002
	ST × H	7.98 (1,23)	0.010
	ST × AP	1.72 (2, 46)	0.0002
Lateral	ST	12.80 (1, 23)	0.002
	ST × H	14.12 (1, 23)	0.001
	ST × AP	4.61 (3, 69)	0.017

ST – Main effect of Sentence Type

ST × H – Sentence Type by Hemisphere interaction

ST × AP – Sentence Type by Anterior Posterior Distribution

3.2.3 ERPs in the coerced sentences: effects of interpretational ambiguity

ERPs were separately averaged in the coerced sentences with dominant (n=30 per list on average) and with multiple interpretations (n=30 per list on average), see Table 2 for parameters of this subdivision. After artifact rejection, there remained, on average, 24 trials in each of these two conditions. As shown in Figure 5, there appeared to be no divergence at all in the waveforms evoked by these two types of sentences either at the complement noun or at the sentence-final word. This was confirmed by ANOVAs conducted between 300–500ms following the onset of complement nouns and between 300–700ms following the onset of SFWs which showed no significant main effects or interactions involving sentence type (all Fs < 2.52 and all ps > 0.12).

Figure 5.

Grand-averaged waveforms comparing ERPs of coerced sentences with strongly preferred interpretations and multiple interpretations at the point of the critical noun (left) and the sentence-final word (right).

4. Discussion

This study aimed to examine the electrophysiological correlates of processing coerced complement NPs that violated the semantic structural specifications of their preceding verbs. NPs denoting entities (e.g. “book”) that were preceded by verbs that selected for complements denoting activities (e.g. “began”) evoked a larger N400 than when the same NPs were preceded by entity-selecting verbs (e.g. “wrote”). An N400 effect of the same magnitude was evoked by entity NPs that violated the animacy selection restrictions of their preceding verbs (e.g. “pleased the book”). Unlike the coerced NPs, the animacy violated NPs were highly implausible and also evoked a robust later positivity – a P600 effect.

The neural response across the three conditions also differed as the sentences unfolded word by word after the complement NP. On the word following the animacy violated complement NP, the positivity effect was still present, but, on the subsequent word, the waveform flipped to a posteriorly-distributed negativity effect (relative to both other conditions) that continued up to and including the sentence-final word. In contrast, there was no divergence in the waveform to words following the coerced and non-coerced complement NPs until the end of the sentence: relative to the sentence-final words of the non-coerced sentences, the sentence-final words of the coerced sentences produced a prolonged anteriorly-distributed positivity effect.

In the following discussion, we consider each of these effects in relation to previous studies examining complement coercion, and in relation to what we know more generally from ERP studies about their functional significance.

4.1 Modulation of the N400 on the complement NP

Our demonstration of increased neural costs to entity NPs following event-selecting verbs is consistent with the series of reading time studies reviewed in the Introduction that also report processing costs in association with complement coercion (McElree et al., 2001; Traxler et al., 2002; Scheepers et al., 2004; Pickering et al., 2005; Traxler et al., 2005; McElree, Pylkkänen et al., 2006; McElree, Frisson et al., 2006; Frisson & McElree, 2008). Our finding of neural modulation primarily between 300–500ms is also consistent with Pylkkänen and McElree’s (2007) MEG study, which reported neuromagnetic modulation between 350–500ms in this contrast (although, as discussed further below, the ERP and MEG effects differed in their scalp distribution). Finally, these findings are consistent with a very recent study examining ERP correlates of complement coercion using similar stimuli to those used here and which also found an N400 effect to coerced (versus non-coerced) complement NPs (Baggio, Choma, van Lambalgen, & Hagoort, in press).

As in previous studies, these costs are unlikely to be fully accounted for by systematic differences between the coerced and non-coerced NPs in their cloze probabilities or their semantic co-occurrences with their preceding content words (LSA values) (McElree et al., 2001; Scheepers et al., 2004). In the present study, although these values did differ slightly between the coerced and non-coerced conditions, the differences were small (much less than those between the animacy violated and non-coerced NPs), and the N400 effect remained significant when we reanalyzed our data using a subset of the stimuli in which these factors were fully matched.

In previous self-paced reading and eye-movement studies, the processing cost on coerced complements has often been interpreted within the theoretical framework proposed by Pustejovsky (1995), i.e. as reflecting the semantic work of type-shifting the complement from an entity to an activity (e.g. “book” to “reading a book”) so as to reach a plausible interpretation of the event (McElree et al., 2001; Scheepers et al., 2004; Pickering et al., 2005; Traxler et al., 2005; McElree, Pylkkänen et al., 2006; Frisson & McElree, 2008). Here, we suggest a slightly different interpretation: that, rather than indexing the work of type-shifting, the N400 to the coerced complement reflected the mismatch between the semantic properties of the verb and those of the complement.

On this account, verbs such as “begin” and “finish” are stored in association with their particular semantic argument structures – their selection for events rather than entities. When an argument that matches this semantic argument structure is encountered, processing is facilitated, leading to an attenuation of the N400, compared to when arguments are encountered that mismatch this argument structure. This attenuation might result from a ‘preactivation’ of eventive semantic frames, leading to the ‘prediction’ of the upcoming argument as an activity rather than an entity (see DeLong et al., 2005; Van Berkum et al., 2005; Federmeier, 2007 for evidence that the N400 can reflect the result of such predictive processing), or it might result from facilitation after the presentation of the complement (‘semantic integration’) (P.J. Holcomb, 1993; Hagoort, 2005).

The mismatch between the verb and complement may have been associated with implicit attempts to retrieve information from memory (inferencing) and this may have also contributed to N400 modulation (see Baggio et al. (in press) for a related interpretation). We suggest that such implicit memory-based inferencing was fairly course-grained and limited to retrieving a general event-schema (e.g. “begin ‘doing something with’ the book”) (see Jackendoff (2002; 1997) for a theoretical account) rather than retrieving and selecting the specific event or events implied by a particular combination of verb and complement (e.g. “begin ‘writing/reading’ the book”). Consistent with this idea, the amplitude of the N400 to complements in sentences such as “The student started the essay…”, where there was only one dominant interpretation (started writing the essay), did not differ from that of the N400 to complements in sentences such as “The director started the script…” where there were many possible interpretations (e.g. reading the script; marking the script; examining the script, etc).⁴ Like Frisson et al. (2008), we take this as evidence that the cost of processing coerced complements does not reflect the cost of selecting between alternative specific interpretations.

The amplitude of the N400 effect evoked by coerced complement nouns did not differ from that evoked by the animacy violated (vs. non-coerced) nouns. We suggest that the N400 effect produced by the animacy violated complements also reflected semantic memory-based processes: matching between the requirements of the verb and the properties of the complement and possibly attempts to retrieve additional information from semantic memory. Of course, the type of mismatch between the coerced and non-coerced complements and between the animacy violated and non-coerced complements differed. In the case of the coerced complements, the mismatch was between the semantic eventive restrictions of the verb and the entity argument and any implicit retrieval of an event schema resulted in a plausible interpretation. In the case of the animacy violations, the mismatch was between the strict animacy-based restrictions of the object-experiencer verbs and any attempts to retrieve additional information failed to result in a plausible representation. However, these differences made little difference to the amplitude of the N400.

The account outlined above makes two related assumptions. The first is that, linguistically, a verb’s (or class of verbs’) semantic argument structure is represented at a distinct level from its syntactic argument structure and that a verb’s semantic structural constraints are ‘invisible’ to the syntax. This deviates from a fairly standard view that the selection restrictions and thematic constraints of a verb are both closely linked to its syntactic argument structure (Chomsky, 1981). In our view, syntactic and semantic structures are represented independently of one another (Jackendoff, 1997, 2002; Culicover & Jackendoff, 2005).⁵ The second assumption is that semantic memory-based processes of matching and retrieval, reflected by the N400, are at least partially independent of processes that (syntactically) assign thematic roles to generate full propositions and that assess the plausibility of such propositions (Kuperberg, 2007). This view is supported by observations that the amplitude of the N400 does not necessary pattern with degree of implausibility of a word within a sentence (Kuperberg et al., 2003; Geyer, Holcomb, Kuperberg, & Perlmutter, 2006; Van de Meerendonk, Kolk, Vissers, & Chwilla, in press). This was very clear in the present study: there was no significant difference between its amplitude to complement nouns in plausible coerced sentences and highly implausible animacy violated sentences (see also Baggio et al., in press).

Of note, a semantic mismatch and retrieval account of complement coercion was considered by Traxler et al. (2005), but rejected mainly because the pattern of eye-movement findings observed to coerced (vs. non-coerced) complement nouns differs from the pattern of eye-movement findings that others have described to outright semantic violations: whereas Traxler and others have consistently shown that the costs of coercion were confined mainly to complement NP itself (Scheepers et al., 2004; Traxler et al., 2005; Pickering et al.), highly implausible sentences with outright selection restriction violations are often associated with additional downstream effects past the violated word (Rayner, Warren, Juhasz, & Liversedge, 2004; Warren & McConnell, 2007). Our current ERP data help reconcile these observations. They suggest that initially the coerced and violated sentences were treated similarly (both evoked an N400 effect that may reflect semantic mismatch and memory-based retrieval). Only in the animacy violated sentences, however, did the syntactic assignment of thematic roles lead to the generation of a highly implausible proposition. As discussed below, we suggest that this implausibility triggered the P600 effect which continued downstream as a positivity to several words past the critical word. These downstream late positivity effects may map on to the downstream eye-movement effects previously seen in association with severe semantic implausibilities.

The N400 effects evoked by the coerced and animacy violated complement NP were not only similar in amplitude but, relative to the non-violated NPs, they both had similar widespread scalp distributions (although the voltage map in Figure 2 suggests that the N400 effect to the animacy violated complement nouns may have been more widespread than that to the coerced complement nouns, this difference was not statistically significant). These observations differ from the pattern of data described in the MEG study by Pylkkänen & McElree (2007) which used similar stimuli and the same acceptability judgment task. In that study, the effect to the coerced (vs. non-coerced) complement nouns was more anteriorly distributed (localizing to an anterior midline field) while that to the animacy violated (vs. non-coerced) complement nouns had a more posterior distribution (localizing to temporal sources). The reasons for this discrepancy are unclear but it is important to note that MEG and ERP measures at the surface of the scalp can be differentially sensitive to a given underlying neural source: for example, the contribution of radially oriented sources, prominent in EEG, are weak in MEG (Baule & McFee, 1965; Sharon, Hamalainen, Tootell, Halgren, & Belliveau, 2007).⁶

Because of the poor spatial resolution of ERPs, and because the N400 is likely to composed of multiple underlying neural generators interacting over the same time-scale (McCarthy, Nobre, Bentin, & Spencer, 1995; Halgren, Baudena, Heit, Clarke, Marinkovic, Chauvel et al., 1994; Halgren, Baudena, Heit, Clarke, Marinkovic, & Clarke, 1994; Dale et al., 2000; Halgren et al., 2002; Marinkovic et al., 2003), we cannot deduce that these two contrasts necessarily engage identical neurocognitive systems. Future studies combining MEG and ERP methods (Sharon et al., 2007) will be able to shed further light on differences and similarities between the neurocognitive mechanisms engaged to coerced versus animacy violated complement nouns.

4.2 Effects following the N400 effect

In our study, the ERP response that did clearly distinguish between the coerced and the animacy violated complement NPs was the P600: relative to non-coerced complements, the animacy violated complements evoked a robust P600 effect, but the coerced complements failed to evoke this effect. This finding is also discrepant with that of the MEG study using similar stimuli and the same task and which did not report any modulation within the 500–900ms time window to the animacy violated NPs. Again, the reasons for this difference between the ERP and MEG findings are unclear, but it does accord with others’ observations that Late Positivity ERPs following the N400 component can sometimes be invisible to MEG (Ellen Lau, personal communication). Indeed, MEG is less sensitive than ERPs to the classic oddball P300 effect (Eulitz, Eulitz, & Elbert, 1997; Okada, Kaufman, & Williamson, 1983; Simpson et al., 1995), which shares some functional commonalities (Coulson, King, & Kutas, 1998), although also some differences (Osterhout & Hagoort, 1999), with the P600.

The presence of a P600 ERP effect to the animacy violations in the present study is, however, consistent with a growing ERP literature documenting a P600 effect not only to syntactic violations where it was classically associated (Osterhout & Holcomb, 1992), but also to clear semantic implausibilities/impossibilities (reviewed by Kuperberg, 2007). These include animacy violations falling on verbs (Kuperberg et al., 2003; Hoeks et al., 2004; Kim & Osterhout, 2005; Kuperberg, Caplan et al., 2006; Kuperberg et al., 2007), as well as animacy violations falling on arguments following verbs. These include inanimate NPs following animate-selecting object-experiencer verbs (Paczynski & Kuperberg, 2009) (as in the present study), inanimate arguments following animate-selecting agent-patient verbs (Nieuwland & Van Berkum, 2005; Paczynski & Kuperberg, 2009), and animate arguments following inanimate-selecting agent-patient verbs (Paczynski & Kuperberg, 2009). There is also some evidence that a P600 can be evoked by other types of selection restriction violations (Geyer et al., 2006) and other types of severe implausibilities (Van de Meerendonk et al., in press).

On the basis of a review of this ‘semantic P600 effect’, Kuperberg (2007) suggested that this component reflects a continued combinatorial analysis (or reanalysis) that is triggered by a highly implausible or unlicensed proposition generated by a full interpretative combinatorial analysis, and that is particularly likely to be triggered in the presence of a conflict with a semantic memory-based analysis. This detection of conflict and reanalysis may draw upon more general executive functions, constituting an online ‘monitoring’ process (see Kolk & Chwilla, 2007). Several factors, acting in combination, can bias towards increased conflict, including a strong semantic constraint of the context, a very implausible (as opposed a semi-implausible) final representation of meaning, and the performance of a plausibility judgment task (see Kuperberg (2007) for further discussion). In the present study, we suggest that the combination of triggers of this effect were the semantic constraint imposed by the object-experiencer verbs, the highly implausible resulting proposition, together with the requirement to make explicit acceptability judgments.⁷

4.3 Downstream ERP effects after the complement NP

The positivity evoked by the animacy violated complement NP remained evident from 400–600ms after the onset of the subsequent word. After this, however, the waveform flipped such that, relative to the other sentence types, a sustained negativity effect was seen on all words up until and including the sentence-final word of the animacy violated sentences. There have been several previous reports of prolonged negativity effects on sentence-final words following mid-sentence anomalies (both semantic and syntactic) (Osterhout & Holcomb, 1992, 1993; Hagoort et al., 1993; Hagoort & Brown, 2000; Hagoort, 2003; Ditman, Holcomb, & Kuperberg, 2007), and its functional significance is debated. One suggestion has been that it reflects an ongoing difficulty in semantic integration, i.e. a prolongation of the N400 or the result of multiple N400s (Osterhout & Holcomb, 1992). An alternative possibility is that it reflects a lack of processing relative to the non-violated sentences. We are currently attempting to distinguish between these two accounts by combining ERPs with self-paced reading (Ditman et al., 2007) and using experimental tasks that vary in their requirements for participants to read until the end of the sentences.

Unlike the pattern of ERPs following the animacy violated complements, the waveforms evoked by the three words following the coerced complement NPs did not diverge from those following the non-coerced complements. At the point of the sentence-final word, however, the waveforms did differ significantly: sentence-final words in the coerced sentences evoked a robust anteriorly-distributed sustained positivity effect relative to the sentence-final words of the non-coerced sentences. The functional significance of this effect is unclear (none of the previous behavioral or neural studies of complement coercion have examined effects at the point of the sentence-final word). One possibility is that it reflects a frontally-mediated active attempt to retrieve a specific unstated event (or possible set of events) in the coerced sentences to form a discourse-level representation. This interpretation links it to other types of frontal positivities that have been reported in various other situations where new information must be retrieved to build a coherent mental model (Friederici, Hahne, & Saddy, 2002; Kaan & Swaab, 2003; Coulson & Williams, 2005; Dwivedi, Philips, Lague-Beauvais, & Baum, 2006; Filik, Sanford, & Leuthold, 2008). For example, in a study by Dwivedi et al. (2006), a sustained frontal positivity was evoked to words such as “ends” in scenarios like “John is considering writing a novel. It ends quite abruptly”, relative to “John is reading a novel. It ends quite abruptly”. Here, the reader must infer that John wrote the novel to make full sense of the meaning. Similarly, in a recent study by Filik et al. (2008), a frontal positivity was evoked by the pronoun “she” versus “they” in scenarios such as “The in-flight meal I got was more impressive than usual. In fact, she/they courteously presented the food as well” where, again, the reader must make an inference that the in-flight meal was presented by a female. There are, of course, important differences between the types of stimuli used in these previous studies and those employed in the present investigation, and future studies will determine whether anterior positivities can, in fact, be linked to these types of inferential processes. What is clear from the present dataset is that the sentence-final anterior positivity in the coerced sentences did not reflect the resolution of ambiguity, as has been hypothesized for frontal positivities observed in other situations, e.g. Kaan & Swaab (2003); similar to the earlier N400 effect evoked at the point of the complement NP, this sentence-final frontal positivity effect was not modulated by the dominance or number of possible interpretations of the coerced sentences. (Retrieval and selection processes can be neuroanatomically dissociated and there is fMRI evidence that they are mediated by distinct regions within the inferior frontal cortex (Wagner, Pare-Blagoev, Clark, & Poldrack, 2001; Thompson-Schill, D'Esposito, & Kan, 1999).

4.4 Conclusions

In sum, we have demonstrated a widespread N400 effect to entity complement NPs following verbs that selected for activities rather than entities. These findings are consistent with previous behavioral and MEG evidence indicating that the processing system registers such discrepancies between the semantic structure of verbs and arguments, even though such violations are invisible to the syntax and do not lead to an implausible interpretation. In the present study, the amplitude of this N400 effect was very similar to that evoked by complements that violated the animacy-based selection restrictions of their preceding verbs. We have suggested that, in both cases, N400 modulation might reflect the registration of a mismatch between the semantics of the verb (whether this be its selection restrictions for events or features) and the semantic properties of the incoming complement, and possibly implicit attempts to retrieve relevant information from semantic memory to ‘fill in’ such mismatches. We also suggest that a delayed sustained anterior positivity on the sentence-final words of coerced sentences may reflect delayed more explicit efforts to retrieve the specific unstated activit(ies) implied by the verb-argument combination.

The interpretation of the N400 to the coerced complements outlined in this article is based on a growing literature suggesting that the modulation of this component is driven by semantic memory-based processes at several different levels and grains of representation. In this study, we have suggested that the N400 to coerced complements was modulated by a mismatch between the aspectual semantic properties of the verb and the argument. However, it is unlikely that this specific type of mismatch between a verb and argument is the only trigger to ‘coercion’ or other types of inferencing in all situations. For example, within sentences, there is some evidence that coercion on complements can occur in the absence of verb-argument semantic mismatch (Frisson, McElree, & Thyparampil, 2005), and within discourse, online inferences can be generated even when semantic associative relationships between individual words are held constant (Kuperberg, Lakshmanan, Caplan, & Holcomb, 2006; Paczynski, Ditman, Okano, & Kuperberg, 2007). We use many different types of stored information to comprehend language online; the N400 is known to be sensitive to categorial feature-based relationships (Federmeier & Kutas, 1999) including selection featural restrictions (Friederici & Frisch, 2000), animacy-based relationships (Weckerly & Kutas, 1999; Frisch & Schlesewsky, 2001; Paczynski & Kuperberg, 2009), associative-based relationships (Van Petten, 1993) including those based on real-world expectations (Hagoort et al., 2004; Kuperberg et al., 2003), and pragmatic relationships (Nieuwland & Kuperberg, 2008; Van Berkum, Van den Brink, Tesink, Kos, & Hagoort, 2008). Mismatches between language input and any of these levels of representations could, in theory, be associated with attempts to retrieve unstated meaning which may, in some cases, lead to plausible representations. Future studies using complementary ERP, fMRI and MEG methodologies will be necessary to examine the full range of triggers and neural mechanisms engaged to retrieve unstated meaning in order to make full sense of language.